diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
new file mode 100644
index 0000000..314eb58
--- /dev/null
+++ b/.github/workflows/ci.yml
@@ -0,0 +1,138 @@
+# Memscribe CI (whitepaper §8.9).
+#
+# Memscribe is deterministic and zero-LLM by construction, so CI is a hard gate,
+# not a smoke test: the same input bytes must always produce the same nodes, the
+# tree must be clippy- and rustfmt-clean, the dependency set must satisfy the
+# license/advisory policy, and the crate must keep building on its MSRV.
+#
+# The toolchain is pinned to match rust-toolchain.toml (1.96.0). The fuzz job is
+# best-effort: cargo-fuzz needs a nightly compiler, so it is allowed to fail
+# without failing the workflow.
+
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+  workflow_dispatch:
+
+# Cancel superseded runs on the same ref to save CI minutes.
+concurrency:
+  group: ci-${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+
+env:
+  CARGO_TERM_COLOR: always
+  RUST_BACKTRACE: 1
+  # Pinned toolchain — keep in lockstep with rust-toolchain.toml.
+  RUST_PINNED: 1.96.0
+  # Resilience against transient crates.io download blips (SSL EOFs, flaky
+  # mirrors): retry network ops aggressively and fetch the index over the
+  # sparse protocol with the git CLI, which recovers from partial transfers
+  # better than the built-in downloader.
+  CARGO_NET_RETRY: 10
+  CARGO_NET_GIT_FETCH_WITH_CLI: "true"
+  CARGO_REGISTRIES_CRATES_IO_PROTOCOL: sparse
+
+jobs:
+  # 1. The deterministic test suite: unit + golden + conformance + property.
+  test:
+    name: test (workspace, all-features)
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Install pinned toolchain
+        uses: dtolnay/rust-toolchain@master
+        with:
+          toolchain: 1.96.0
+      - uses: Swatinem/rust-cache@v2
+      - name: cargo test
+        run: cargo test --workspace --all-features --locked
+
+  # 2. Lints as errors. No warning escapes review.
+  clippy:
+    name: clippy (-D warnings)
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Install pinned toolchain
+        uses: dtolnay/rust-toolchain@master
+        with:
+          toolchain: 1.96.0
+          components: clippy
+      - uses: Swatinem/rust-cache@v2
+      - name: cargo clippy
+        run: cargo clippy --workspace --all-targets --all-features -- -D warnings
+
+  # 3. Formatting. The output is byte-stable, so the source should be too.
+  fmt:
+    name: rustfmt (--check)
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Install pinned toolchain
+        uses: dtolnay/rust-toolchain@master
+        with:
+          toolchain: 1.96.0
+          components: rustfmt
+      - name: cargo fmt --check
+        run: cargo fmt --all --check
+
+  # 4. License + advisory gate (deny.toml).
+  deny:
+    name: cargo-deny
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: cargo-deny check
+        uses: EmbarkStudios/cargo-deny-action@v2
+        with:
+          command: check
+          arguments: --all-features
+
+  # 5. MSRV — the crate must build on its declared minimum (1.96).
+  #    Build + check only: tests pin newer dev-deps and run under `test`.
+  msrv:
+    name: MSRV (1.96)
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Install MSRV toolchain
+        uses: dtolnay/rust-toolchain@master
+        with:
+          toolchain: 1.96.0
+      - uses: Swatinem/rust-cache@v2
+        with:
+          key: msrv
+      - name: cargo check (MSRV)
+        run: cargo check --workspace --all-features --locked
+
+  # 6. cargo-fuzz smoke build. Best-effort: cargo-fuzz needs nightly, and the
+  #    fuzz/ targets may not be wired yet — never fail the workflow on this.
+  fuzz:
+    name: cargo-fuzz smoke build (best-effort)
+    runs-on: ubuntu-latest
+    continue-on-error: true
+    steps:
+      - uses: actions/checkout@v4
+      - name: Install nightly toolchain
+        uses: dtolnay/rust-toolchain@master
+        with:
+          toolchain: nightly
+      - uses: Swatinem/rust-cache@v2
+        with:
+          key: fuzz
+      - name: Install cargo-fuzz
+        run: cargo install cargo-fuzz --locked
+      - name: Build fuzz targets (no run)
+        working-directory: fuzz
+        run: |
+          if [ -f Cargo.toml ]; then
+            cargo +nightly fuzz build
+          else
+            echo "fuzz/ has no Cargo.toml yet — nothing to build (best-effort job)."
+          fi
diff --git a/ARCHITECTURE.md b/ARCHITECTURE.md
new file mode 100644
index 0000000..8fc4770
--- /dev/null
+++ b/ARCHITECTURE.md
@@ -0,0 +1,195 @@
+# Memscribe architecture
+
+Memscribe is a **deterministic, zero-LLM** pipeline that turns the transcript
+logs AI coding agents already write into typed nodes the downstream
+inference-and-governance layer (**MemCortex**) can consume. No model is ever
+called: capture is reading and parsing, never summarizing. The output is an exact
+function of the input, which is what makes the whole module golden-file,
+property, and fuzz testable.
+
+It is the bottom of a three-layer stack — **Memtrace** uses **MemCortex**, and
+**MemCortex** uses **Memscribe**. The dependency direction is strictly one-way:
+each layer depends only on the one below it, and `memscribe-core` depends on
+nothing else in the workspace. Memscribe never calls upward.
+
+---
+
+## The pipeline
+
+A single, linear, deterministic pipeline. Each stage is a trait, so it can be
+tested in isolation and swapped. Everything between **Source** and **Sink** is a
+pure, synchronous function of the event stream.
+
+```
+ Source                Adapter           Gate        Segmenter      Binder        NodePrep        Sink
+ (memscribe-io)        (memscribe-       (core)      (core)         (core)        (core)          (memscribe-sink)
+                        adapters)
+ tail JSONL        →   parse one     →   admit?  →   arc / turn  →  decision  →   assemble    →   NDJSON / SQLite
+ hook stdin            RawRecord →       commitment  spans;         ↔ edit,       PreparedNode    / MemDB
+ OTLP receiver         CaptureEvent[]    markers     elevate gated  PROV          stream
+                       (version-                     turns; seed    (t_use
+                        tolerant)                     decisions;    ≤ t_gen)
+                                                      collect edits
+   RawRecord               CaptureEvent      markers    Segmentation   BindingEdge   PreparedNode    (consumer)
+   (bytes + provenance)    (normalized)                                              stream
+```
+
+- **`Source → Adapter`** produces the normalized `CaptureEvent` stream — the
+  system of record. This is the only stage that touches tool-specific formats.
+- **`Gate → Segmenter → Binder → NodePrep`** transform that stream into
+  `PreparedNode`s. Pure and synchronous given the events.
+- An optional **redaction** pass runs over the prepared nodes before the sink.
+- **`Sink`** writes the nodes out. It is the single seam that decouples
+  Memscribe from MemDB.
+
+The orchestration lives in `memscribe-core::pipeline::DefaultPipeline`:
+
+```rust
+let nodes = DefaultPipeline::new()                 // redaction ON by default
+    .run_records(adapter.as_ref(), &records);      // parse → prepare → redact
+// or stream straight to a sink:
+let n = DefaultPipeline::new()
+    .run_to_sink(adapter.as_ref(), &records, &mut sink)?;
+```
+
+`DefaultPipeline::prepare_events(&events)` is the **pure** core: its output is an
+exact function of `events`. `without_redaction()` turns the redactor off (golden
+tests assert on verbatim content), and `with_gate(..)` / `with_redactor(..)`
+swap in config-driven stages.
+
+---
+
+## Crate responsibilities
+
+| Crate | Responsibility |
+|-------|----------------|
+| `memscribe-core` | The frozen contract: the event model, the prepared-node output types, the `TranscriptAdapter` and `Sink` traits, and the deterministic pipeline (`gate` → `segmenter` → `binder` → `nodeprep`) plus the `redact` pass. Depends on nothing in the workspace. |
+| `memscribe-adapters` | Per-tool parsers behind feature flags. Each implements `TranscriptAdapter`. The `registry` assembles the enabled set (`all_adapters`) and resolves one by `SourceKind` (`adapter_for`). |
+| `memscribe-io` | Generic sources: a notify-based file tailer (offset resume), a hook server, and an OTLP receiver. Turns raw bytes into `RawRecord`s. |
+| `memscribe-sink` | Concrete `Sink`s: `NdjsonSink` (canonical default), `SqliteSink` (feature `sqlite`), and `MemDbSink` (feature `memdb`, off by default). |
+| `memscribe-cli` | The `memscribe` binary: `watch` / `hook` / `parse` / `replay` / `verify` / `redact`. |
+| `memscribe-testkit` | The harness: `parse_events` / `prepare_nodes`, the invariant checks, golden-fixture loaders, and the cross-tool conformance scenario catalog. |
+
+---
+
+## The contract types
+
+All of these live in `memscribe-core` and are re-exported from its crate root.
+**Do not change their behavior or public shape** — the test suite and every
+consumer depend on exact output.
+
+### Input: the normalized event model (`model.rs`)
+
+`CaptureEvent` is the system of record produced by adapters. Every field is
+copied verbatim from the source; none is generated by Memscribe.
+
+```rust
+pub struct CaptureEvent {
+    pub schema_version: u16,        // SCHEMA_VERSION; consumers gate on this
+    pub source: SourceKind,         // which tool produced it
+    pub session_id: String,         // tool-native session/thread id
+    pub seq: u64,                   // monotonic per-session, from file order
+    pub event_id: String,           // tool-native id, or blake3(content) fallback
+    pub parent_id: Option<String>,  // DAG link where the tool provides one
+    pub timestamp: OffsetDateTime,  // RFC3339, verbatim
+    pub project: ProjectRef,        // cwd / repo_root / git, from session start
+    pub kind: EventKind,            // the payload
+    pub provenance: SourceLocation, // pointer back into the source bytes
+}
+```
+
+`EventKind` is the payload enum. `EventKind::Unknown` is **load-bearing**: an
+unrecognized record type or a new field is preserved verbatim and flagged,
+never discarded — that is how the stream stays lossless across tool-version
+churn.
+
+| `EventKind` variant | Meaning |
+|---------------------|---------|
+| `SessionStart` | cwd, git ref, model, tool version |
+| `UserTurn` | a user message (flattened text + structured `Part`s) |
+| `AssistantTurn` | an assistant message (text, thinking, model, usage, parts) |
+| `ToolCall` | a tool invocation (`call_id`, name, raw args) |
+| `ToolResult` | a tool result (`call_id`, `ok`, raw output) |
+| `FileEdit` | a normalized `Diff` (from Edit/Write/apply_patch/replace) |
+| `Compaction` | model-side history compaction — flagged, never stored as truth |
+| `Rewind` | a user rewind back to an earlier event |
+| `SessionEnd` | the session ended |
+| `Unknown` | an unrecognized record, preserved verbatim and flagged |
+
+`SourceKind` enumerates the nine tools plus `Unknown`; `SourceKind::parse` maps
+CLI/`--as` slugs (tolerant of aliases such as `claude` / `claude-code`).
+
+### Output: the prepared-node stream (`node.rs`)
+
+`PreparedNode` is the typed data a consumer ingests. It is a tagged enum:
+
+| `PreparedNode` variant | Payload | Meaning |
+|------------------------|---------|---------|
+| `Conversation` | `ConversationSpan` | a gated, verbatim dialogue span with the markers that fired |
+| `Decision` | `DecisionRecord` | a deterministically-parsed decision (IBIS / QOC / MADR / Kruchten shape) |
+| `Episode` | `CodeEpisode` | a code edit episode: path, `Diff`, git ref, deterministic `episode_id` |
+| `Binding` | `BindingEdge` | a decision/conversation → episode edge carrying a `ProvRecord` |
+
+### Epistemic honesty: `FactStatus`
+
+Every node and edge carries a `FactStatus`. **Memscribe only ever emits the
+first two**; the latter two are *flags* for a downstream inference layer —
+values Memscribe never computes by guessing. This is the property that keeps the
+module zero-LLM and its output golden-testable.
+
+| `FactStatus` | Who sets it |
+|--------------|-------------|
+| `Observed` | Memscribe — verbatim from the source |
+| `DeterministicallyDerived` | Memscribe — a pure function of observed data |
+| `StatisticallyRanked` | downstream — a statistical measure |
+| `LlmHypothesis` | downstream — an LLM hypothesis; Memscribe only *flags* it |
+
+`ProvRecord` records `used(session, decision)` + `wasGeneratedBy(diff, session)`
+with the temporal invariant `t_use ≤ t_gen` (`ProvRecord::is_temporally_valid`).
+
+---
+
+## How to add a new adapter
+
+Adapters are the volatile part — every tool's format churns — so adding one is a
+well-trodden, five-step path. The contract: a parser is **version-tolerant**
+(it pattern-matches on the fields it needs and routes anything unrecognized to
+`EventKind::Unknown`) and **must never panic**.
+
+1. **Add a `SourceKind` variant** (`memscribe-core/src/model.rs`). Wire its
+   stable snake_case slug into `SourceKind::as_str` and into `SourceKind::parse`
+   (include any aliases). This is the one allowed touch of `memscribe-core` for
+   a new tool — coordinate it, since the frozen contract is shared.
+
+2. **Add the adapter module** (`memscribe-adapters/src/<tool>.rs`) behind a
+   `#[cfg(feature = "<tool>")]` and a matching entry in the crate's `[features]`
+   table. Implement `TranscriptAdapter`:
+   - `source_kind()` — return your `SourceKind`.
+   - `discover(&DiscoverCfg)` — locate live & historical transcripts. Honor the
+     per-tool override key in `DiscoverCfg.overrides` (e.g. `CLAUDE_CONFIG_DIR`,
+     `CODEX_HOME`) and fall back to `cfg.home_dir()`. Return handles in a
+     deterministic (sorted) order.
+   - `parse(&RawRecord, &mut ParseCtx)` — turn ONE record into zero or more
+     `CaptureEvent`s. Use `ParseCtx::alloc_seq` for the monotonic `seq`,
+     `ParseCtx::first_seen` for dedup, and `ParseCtx::project_or_default` for the
+     project binding. Never panic; route unknowns to `EventKind::Unknown`.
+   - `schema_fingerprint(&RawRecord)` — return a `SchemaVariant` so the corpus
+     and runtime can version-gate the parser.
+
+3. **Register it** (`memscribe-adapters/src/registry.rs`). Add the cfg-gated
+   `push` in `all_adapters()` and the cfg-gated arm in `adapter_for()`.
+
+4. **Add fixtures** under `fixtures/<tool>/<version>/<scenario>.jsonl` for the
+   canonical scenarios in `memscribe-testkit::scenarios::SCENARIOS`, and bless
+   the expected outputs under `fixtures-expected/<tool>/<version>/` (see
+   [CONTRIBUTING.md](./CONTRIBUTING.md) for the capture → golden → bless flow).
+
+5. **Add tests.** Unit-test the parser; run the shared invariant checks from
+   `memscribe-testkit::invariants` (`check_monotonic_seq`, `check_lossless`,
+   `check_unique_event_ids`, `check_determinism`); and add a `cargo-fuzz` target
+   so the never-panic contract is enforced. Verify in isolation:
+   `cargo test -p memscribe-adapters --test <your_file_stem>`.
+
+The conformance suite then asserts your tool normalizes the canonical scenarios
+to the **same shape** as every other tool — that cross-tool equivalence is the
+point of the thin-waist event model.
diff --git a/CHANGELOG.md b/CHANGELOG.md
new file mode 100644
index 0000000..48e00ee
--- /dev/null
+++ b/CHANGELOG.md
@@ -0,0 +1,24 @@
+# Changelog
+
+All notable changes to Memscribe are documented here. The format follows
+[Keep a Changelog](https://keepachangelog.com/), and the project adheres to
+[Semantic Versioning](https://semver.org/). The event schema additionally
+carries its own `schema_version` so the consumer layer (MemCortex) can refuse
+or adapt to an incompatible event schema independently of the crate version.
+
+## [Unreleased]
+
+### Added
+- **M1 — Core contract.** The frozen thin-waist: `CaptureEvent` / `EventKind`
+  normalized event model, `PreparedNode` output contract with `FactStatus`,
+  the `TranscriptAdapter` and `Sink` traits, and the deterministic pipeline
+  (Gate → Segmenter → Binder → NodePrep).
+- **Adapters.** Claude Code, Codex CLI, Gemini CLI, OTel GenAI, plus
+  VS Code / Copilot / Cursor / Windsurf / Zed, each version-tolerant and
+  routing unknowns to `EventKind::Unknown`.
+- **Sinks.** NDJSON (canonical default), SQLite, and a feature-gated MemDB sink.
+- **IO sources.** notify-based file tailer with persisted byte-offset resume,
+  hook server, and an optional OTLP receiver.
+- **CLI.** `watch`, `hook`, `parse`, `replay`, `verify`, `redact`.
+- **Testkit.** Golden-file harness, cross-tool conformance suite, synthetic
+  generators, property tests, and fuzz targets.
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
new file mode 100644
index 0000000..2f28d61
--- /dev/null
+++ b/CONTRIBUTING.md
@@ -0,0 +1,173 @@
+# Contributing to Memscribe
+
+Thanks for helping. Memscribe is the **deterministic, zero-LLM** data layer
+beneath MemCortex (which Memtrace builds on). The bar for a change is unusual: the output must be
+an *exact function* of the input, so the test suite is the contract. Read
+[ARCHITECTURE.md](./ARCHITECTURE.md) first — it explains the pipeline and the
+contract types you must not break.
+
+---
+
+## Build & test
+
+The toolchain is pinned in `rust-toolchain.toml` (**1.96.0**); `rustup` picks it
+up automatically. The MSRV is **1.96**.
+
+```console
+# The whole gate, the way CI runs it:
+cargo test  --workspace --all-features            # unit + golden + conformance + property
+cargo clippy --workspace --all-targets --all-features -- -D warnings
+cargo fmt   --all --check
+cargo deny  check                                 # license + advisory policy (deny.toml)
+```
+
+When you are iterating on a single crate or a single integration test, build it
+in **isolation** so you do not compile peers' in-progress test files:
+
+```console
+cargo test -p memscribe-adapters --test <your_file_stem>
+cargo test -p memscribe-core
+```
+
+Do **not** run `cargo fmt` to reformat in a PR that you want reviewed cleanly —
+write rustfmt-clean code and let `cargo fmt --all --check` verify it.
+
+---
+
+## The rules that make Memscribe Memscribe
+
+These are not style preferences; they are the invariants the test suite (and the
+downstream consumer) rely on. A change that violates one is a bug.
+
+### 1. Determinism
+
+The same input bytes must always produce byte-identical output. No clocks, no
+randomness, no hash-map iteration order leaking into output, no filesystem walk
+order leaking into discovery (sort it). The property test
+`invariants::check_determinism` runs two parses and asserts the serialized
+events are identical — keep it green.
+
+### 2. No LLM, ever
+
+Capture is reading and parsing, never summarizing or inferring. Memscribe emits
+nodes only with `FactStatus::Observed` (verbatim) or
+`FactStatus::DeterministicallyDerived` (a pure function of observed data).
+Anything that would require inference — fine-grained decision typing, concept
+naming, statistical ranking — is **flagged** (`StatisticallyRanked`,
+`LlmHypothesis`) for a downstream layer to compute, never guessed here. If you
+find yourself reaching for a heuristic that "usually" gets it right, stop: that
+belongs downstream.
+
+### 3. Losslessness
+
+Every non-blank source record maps to at least one event. Unrecognized records
+and new fields are preserved verbatim and routed to `EventKind::Unknown` /
+`Part::Other`, never dropped. `invariants::check_lossless` enforces the lower
+bound.
+
+### 4. Monotonic, unique, idempotent
+
+`seq` is strictly increasing within a session and matches file order
+(`check_monotonic_seq`). A record is deduplicated once on its tool-native
+`event_id` (`check_unique_event_ids`), so re-ingesting the same input is
+idempotent.
+
+### 5. Never panic
+
+A parser must never panic on any input — malformed, truncated, adversarial, or
+from a tool version it has never seen. Every adapter parser has a `cargo-fuzz`
+target that asserts this. Use the version-tolerant pattern: match the fields you
+need, route the rest to `Unknown`.
+
+---
+
+## The adapter version-tolerance contract
+
+Tool transcript formats churn constantly; that is precisely why the adapters are
+the open, community-versioned part of Memscribe. An adapter must:
+
+- **Pattern-match only the fields it needs** and route anything unrecognized to
+  `EventKind::Unknown` (with the raw record preserved) rather than failing the
+  stream. A new field in a record you understand must not break parsing.
+- **Never panic** on any input (see rule 5).
+- **Fingerprint** its input via `schema_fingerprint` → `SchemaVariant`, so the
+  corpus and runtime can version-gate the parser (e.g. `claude_code/2.1`,
+  `codex/rollout-v2`). When a tool ships an incompatible format, add a new
+  variant and a fixture under that version — do not silently widen the old one.
+- **Honor `DiscoverCfg`** in `discover`: read the per-tool override key from
+  `DiscoverCfg.overrides` (e.g. `CLAUDE_CONFIG_DIR`, `CODEX_HOME`), fall back to
+  `cfg.home_dir()`, and return handles in a **sorted** (deterministic) order.
+
+The payoff is the conformance suite: the same canonical scenario, captured from
+any tool, must normalize to the **same shape**. See ARCHITECTURE.md, "How to add
+a new adapter," for the full five-step path.
+
+---
+
+## The fixture-corpus workflow: capture → golden → bless
+
+Memscribe is tested the way a compiler is — fixtures in, exact expected output.
+The corpus lives in two trees:
+
+```text
+fixtures/<tool>/<version>/<scenario>.jsonl                  # input transcript
+fixtures-expected/<tool>/<version>/<scenario>.events.json   # expected CaptureEvent[]
+fixtures-expected/<tool>/<version>/<scenario>.nodes.json    # expected PreparedNode[]
+```
+
+The canonical scenario slugs are defined once in
+`memscribe-testkit::scenarios::SCENARIOS` (e.g.
+`happy_path_decision_then_edits`, `rejected_alternative`, `ban`,
+`tool_failure`, …). Every tool's fixtures should cover them so the cross-tool
+conformance suite can assert equivalence.
+
+**1. Capture.** Get a real (or hand-authored, minimal, redacted) transcript for
+the scenario and drop it at `fixtures/<tool>/<version>/<scenario>.jsonl`. Real
+captures must be scrubbed of secrets and personal paths first — run them through
+the redactor and eyeball the result. For a live session you can snapshot:
+
+```console
+memscribe verify --capture     # (planned) snapshot a live session into a new fixture
+```
+
+**2. Golden.** Generate the expected output and inspect it by hand. The harness
+parses with the adapter and runs the pipeline with **redaction off** so the
+golden asserts on verbatim content:
+
+```console
+# Eyeball the normalized events and prepared nodes for a fixture:
+memscribe parse fixtures/<tool>/<version>/<scenario>.jsonl --as <tool> --no-redact
+```
+
+Confirm it satisfies the invariants and the scenario's stated expectation
+(e.g. for `ban`, the `DecisionRecord.is_ban` flag is `true`).
+
+**3. Bless.** Write the reviewed output to the `fixtures-expected/` paths
+(`*.events.json`, `*.nodes.json`). The golden tests use `insta`; accept a
+reviewed snapshot with:
+
+```console
+cargo insta review          # interactively accept/reject changed snapshots
+# or, after eyeballing the diff:
+cargo insta accept
+```
+
+**Never bless a snapshot you have not read.** A blessed golden is a claim about
+exactly what the deterministic pipeline produces; an unreviewed `accept` turns a
+real regression into a permanent "expected" value. If a golden changes
+unexpectedly, that is the suite doing its job — find out *why* before you
+re-bless.
+
+---
+
+## Submitting a change
+
+- Keep the change scoped; touch only the crate(s) you own. `memscribe-core` is a
+  frozen contract — changing a public type or its output ripples through every
+  consumer and every golden, so coordinate those separately.
+- Make sure all five hard gates above pass locally before you open the PR.
+- Add or update fixtures + tests for any behavior change. A behavior change with
+  no golden delta is a red flag.
+
+By contributing you agree your work is dual-licensed under **MIT OR Apache-2.0**,
+matching the project.
diff --git a/Cargo.lock b/Cargo.lock
new file mode 100644
index 0000000..1313272
--- /dev/null
+++ b/Cargo.lock
@@ -0,0 +1,1691 @@
+# This file is automatically @generated by Cargo.
+# It is not intended for manual editing.
+version = 4
+
+[[package]]
+name = "ahash"
+version = "0.8.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5a15f179cd60c4584b8a8c596927aadc462e27f2ca70c04e0071964a73ba7a75"
+dependencies = [
+ "cfg-if",
+ "once_cell",
+ "version_check",
+ "zerocopy",
+]
+
+[[package]]
+name = "aho-corasick"
+version = "1.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ddd31a130427c27518df266943a5308ed92d4b226cc639f5a8f1002816174301"
+dependencies = [
+ "memchr",
+]
+
+[[package]]
+name = "anstream"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "824a212faf96e9acacdbd09febd34438f8f711fb84e09a8916013cd7815ca28d"
+dependencies = [
+ "anstyle",
+ "anstyle-parse",
+ "anstyle-query",
+ "anstyle-wincon",
+ "colorchoice",
+ "is_terminal_polyfill",
+ "utf8parse",
+]
+
+[[package]]
+name = "anstyle"
+version = "1.0.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "940b3a0ca603d1eade50a4846a2afffd5ef57a9feac2c0e2ec2e14f9ead76000"
+
+[[package]]
+name = "anstyle-parse"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "52ce7f38b242319f7cabaa6813055467063ecdc9d355bbb4ce0c68908cd8130e"
+dependencies = [
+ "utf8parse",
+]
+
+[[package]]
+name = "anstyle-query"
+version = "1.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "40c48f72fd53cd289104fc64099abca73db4166ad86ea0b4341abe65af83dadc"
+dependencies = [
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "anstyle-wincon"
+version = "3.0.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "291e6a250ff86cd4a820112fb8898808a366d8f9f58ce16d1f538353ad55747d"
+dependencies = [
+ "anstyle",
+ "once_cell_polyfill",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "anyhow"
+version = "1.0.102"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7f202df86484c868dbad7eaa557ef785d5c66295e41b460ef922eca0723b842c"
+
+[[package]]
+name = "arrayref"
+version = "0.3.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "76a2e8124351fda1ef8aaaa3bbd7ebbcb486bbcd4225aca0aa0d84bb2db8fecb"
+
+[[package]]
+name = "arrayvec"
+version = "0.7.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f02882884d3e1bc524fb12c79f107f6ad0e1cfd498c536ffb494301740995dfe"
+
+[[package]]
+name = "assert_cmd"
+version = "2.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2aa3a22042e45de04255c7bf3626e239f450200fd0493c1e382263544b20aea6"
+dependencies = [
+ "anstyle",
+ "bstr",
+ "libc",
+ "predicates",
+ "predicates-core",
+ "predicates-tree",
+ "wait-timeout",
+]
+
+[[package]]
+name = "autocfg"
+version = "1.5.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f2032f911046de80f0a198e0901378627c33f59ea0ac00e363d481118bd70a53"
+
+[[package]]
+name = "bit-set"
+version = "0.8.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "08807e080ed7f9d5433fa9b275196cfc35414f66a0c79d864dc51a0d825231a3"
+dependencies = [
+ "bit-vec",
+]
+
+[[package]]
+name = "bit-vec"
+version = "0.8.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5e764a1d40d510daf35e07be9eb06e75770908c27d411ee6c92109c9840eaaf7"
+
+[[package]]
+name = "bitflags"
+version = "1.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bef38d45163c2f1dde094a7dfd33ccf595c92905c8f8f4fdc18d06fb1037718a"
+
+[[package]]
+name = "bitflags"
+version = "2.13.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b4388bee8683e3d04af747c73422af53102d2bd24d9eadb6cbc100baef4b43f8"
+
+[[package]]
+name = "blake3"
+version = "1.8.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0aa83c34e62843d924f905e0f5c866eb1dd6545fc4d719e803d9ba6030371fce"
+dependencies = [
+ "arrayref",
+ "arrayvec",
+ "cc",
+ "cfg-if",
+ "constant_time_eq",
+ "cpufeatures",
+]
+
+[[package]]
+name = "block2"
+version = "0.6.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cdeb9d870516001442e364c5220d3574d2da8dc765554b4a617230d33fa58ef5"
+dependencies = [
+ "objc2",
+]
+
+[[package]]
+name = "bstr"
+version = "1.12.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "63044e1ae8e69f3b5a92c736ca6269b8d12fa7efe39bf34ddb06d102cf0e2cab"
+dependencies = [
+ "memchr",
+ "regex-automata",
+ "serde",
+]
+
+[[package]]
+name = "bytes"
+version = "1.12.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8ae3f5d315924270530207e2a68396c3cc547f6dca3fbdca317cfb1a51edb593"
+
+[[package]]
+name = "cc"
+version = "1.2.65"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e228eec9be7c17ccb640b59b36a5cd805ea2a564a4c5e162c2f659fea30d3b96"
+dependencies = [
+ "find-msvc-tools",
+ "jobserver",
+ "libc",
+ "shlex",
+]
+
+[[package]]
+name = "cfg-if"
+version = "1.0.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9330f8b2ff13f34540b44e946ef35111825727b38d33286ef986142615121801"
+
+[[package]]
+name = "cfg_aliases"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "613afe47fcd5fac7ccf1db93babcb082c5994d996f20b8b159f2ad1658eb5724"
+
+[[package]]
+name = "clap"
+version = "4.6.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1ddb117e43bbf7dacf0a4190fef4d345b9bad68dfc649cb349e7d17d28428e51"
+dependencies = [
+ "clap_builder",
+ "clap_derive",
+]
+
+[[package]]
+name = "clap_builder"
+version = "4.6.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "714a53001bf66416adb0e2ef5ac857140e7dc3a0c48fb28b2f10762fc4b5069f"
+dependencies = [
+ "anstream",
+ "anstyle",
+ "clap_lex",
+ "strsim",
+]
+
+[[package]]
+name = "clap_derive"
+version = "4.6.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f2ce8604710f6733aa641a2b3731eaa1e8b3d9973d5e3565da11800813f997a9"
+dependencies = [
+ "heck",
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "clap_lex"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c8d4a3bb8b1e0c1050499d1815f5ab16d04f0959b233085fb31653fbfc9d98f9"
+
+[[package]]
+name = "colorchoice"
+version = "1.0.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1d07550c9036bf2ae0c684c4297d503f838287c83c53686d05370d0e139ae570"
+
+[[package]]
+name = "console"
+version = "0.16.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d64e8af5551369d19cf50138de61f1c42074ab970f74e99be916646777f8fc87"
+dependencies = [
+ "encode_unicode",
+ "libc",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "constant_time_eq"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3d52eff69cd5e647efe296129160853a42795992097e8af39800e1060caeea9b"
+
+[[package]]
+name = "cpufeatures"
+version = "0.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8b2a41393f66f16b0823bb79094d54ac5fbd34ab292ddafb9a0456ac9f87d201"
+dependencies = [
+ "libc",
+]
+
+[[package]]
+name = "crossbeam-channel"
+version = "0.5.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "82b8f8f868b36967f9606790d1903570de9ceaf870a7bf9fbbd3016d636a2cb2"
+dependencies = [
+ "crossbeam-utils",
+]
+
+[[package]]
+name = "crossbeam-utils"
+version = "0.8.21"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d0a5c400df2834b80a4c3327b3aad3a4c4cd4de0629063962b03235697506a28"
+
+[[package]]
+name = "ctrlc"
+version = "3.5.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e0b1fab2ae45819af2d0731d60f2afe17227ebb1a1538a236da84c93e9a60162"
+dependencies = [
+ "dispatch2",
+ "nix",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "deranged"
+version = "0.5.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7cd812cc2bc1d69d4764bd80df88b4317eaef9e773c75226407d9bc0876b211c"
+dependencies = [
+ "serde_core",
+]
+
+[[package]]
+name = "difflib"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6184e33543162437515c2e2b48714794e37845ec9851711914eec9d308f6ebe8"
+
+[[package]]
+name = "dispatch2"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1e0e367e4e7da84520dedcac1901e4da967309406d1e51017ae1abfb97adbd38"
+dependencies = [
+ "bitflags 2.13.0",
+ "block2",
+ "libc",
+ "objc2",
+]
+
+[[package]]
+name = "encode_unicode"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "34aa73646ffb006b8f5147f3dc182bd4bcb190227ce861fc4a4844bf8e3cb2c0"
+
+[[package]]
+name = "equivalent"
+version = "1.0.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "877a4ace8713b0bcf2a4e7eec82529c029f1d0619886d18145fea96c3ffe5c0f"
+
+[[package]]
+name = "errno"
+version = "0.3.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
+dependencies = [
+ "libc",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "fallible-iterator"
+version = "0.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2acce4a10f12dc2fb14a218589d4f1f62ef011b2d0cc4b3cb1bba8e94da14649"
+
+[[package]]
+name = "fallible-streaming-iterator"
+version = "0.1.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7360491ce676a36bf9bb3c56c1aa791658183a54d2744120f27285738d90465a"
+
+[[package]]
+name = "fastrand"
+version = "2.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9f1f227452a390804cdb637b74a86990f2a7d7ba4b7d5693aac9b4dd6defd8d6"
+
+[[package]]
+name = "file-id"
+version = "0.2.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e1fc6a637b6dc58414714eddd9170ff187ecb0933d4c7024d1abbd23a3cc26e9"
+dependencies = [
+ "windows-sys 0.60.2",
+]
+
+[[package]]
+name = "filetime"
+version = "0.2.29"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5c287a33c7f0a620c38e641e7f60827713987b3c0f26e8ddc9462cc69cf75759"
+dependencies = [
+ "cfg-if",
+ "libc",
+]
+
+[[package]]
+name = "find-msvc-tools"
+version = "0.1.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5baebc0774151f905a1a2cc41989300b1e6fbb29aff0ceffa1064fdd3088d582"
+
+[[package]]
+name = "float-cmp"
+version = "0.10.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b09cf3155332e944990140d967ff5eceb70df778b34f77d8075db46e4704e6d8"
+dependencies = [
+ "num-traits",
+]
+
+[[package]]
+name = "fnv"
+version = "1.0.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3f9eec918d3f24069decb9af1554cad7c880e2da24a9afd88aca000531ab82c1"
+
+[[package]]
+name = "fsevent-sys"
+version = "4.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "76ee7a02da4d231650c7cea31349b889be2f45ddb3ef3032d2ec8185f6313fd2"
+dependencies = [
+ "libc",
+]
+
+[[package]]
+name = "getrandom"
+version = "0.3.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "899def5c37c4fd7b2664648c28120ecec138e4d395b459e5ca34f9cce2dd77fd"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "r-efi 5.3.0",
+ "wasip2",
+]
+
+[[package]]
+name = "getrandom"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "300e883d756b2e4ec94e02791f39b04b522276138852cfc41d9fb7e904106099"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "r-efi 6.0.0",
+]
+
+[[package]]
+name = "hashbrown"
+version = "0.14.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e5274423e17b7c9fc20b6e7e208532f9b19825d82dfd615708b70edd83df41f1"
+dependencies = [
+ "ahash",
+]
+
+[[package]]
+name = "hashbrown"
+version = "0.17.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ed5909b6e89a2db4456e54cd5f673791d7eca6732202bbf2a9cc504fe2f9b84a"
+
+[[package]]
+name = "hashlink"
+version = "0.9.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6ba4ff7128dee98c7dc9794b6a411377e1404dba1c97deb8d1a55297bd25d8af"
+dependencies = [
+ "hashbrown 0.14.5",
+]
+
+[[package]]
+name = "heck"
+version = "0.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2304e00983f87ffb38b55b444b5e3b60a884b5d30c0fca7d82fe33449bbe55ea"
+
+[[package]]
+name = "indexmap"
+version = "2.14.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d466e9454f08e4a911e14806c24e16fba1b4c121d1ea474396f396069cf949d9"
+dependencies = [
+ "equivalent",
+ "hashbrown 0.17.1",
+]
+
+[[package]]
+name = "inotify"
+version = "0.9.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f8069d3ec154eb856955c1c0fbffefbf5f3c40a104ec912d4797314c1801abff"
+dependencies = [
+ "bitflags 1.3.2",
+ "inotify-sys",
+ "libc",
+]
+
+[[package]]
+name = "inotify-sys"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e05c02b5e89bff3b946cedeca278abc628fe811e604f027c45a8aa3cf793d0eb"
+dependencies = [
+ "libc",
+]
+
+[[package]]
+name = "insta"
+version = "1.48.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "86f0f8fee8c926415c58d6ae43a08523a26faccb2323f5e6b644fe7dd4ef6b82"
+dependencies = [
+ "console",
+ "once_cell",
+ "serde",
+ "similar",
+ "tempfile",
+]
+
+[[package]]
+name = "is_terminal_polyfill"
+version = "1.70.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a6cb138bb79a146c1bd460005623e142ef0181e3d0219cb493e02f7d08a35695"
+
+[[package]]
+name = "itoa"
+version = "1.0.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682"
+
+[[package]]
+name = "jobserver"
+version = "0.1.34"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9afb3de4395d6b3e67a780b6de64b51c978ecf11cb9a462c66be7d4ca9039d33"
+dependencies = [
+ "getrandom 0.3.4",
+ "libc",
+]
+
+[[package]]
+name = "kqueue"
+version = "1.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "273c0752728918e0ac4976f2b275b6fefb9ecd400585dec929419f3844cd87b5"
+dependencies = [
+ "kqueue-sys",
+ "libc",
+]
+
+[[package]]
+name = "kqueue-sys"
+version = "1.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "07293a4e297ac234359b510362495713f75ea345d5307140414f20c69ffeb087"
+dependencies = [
+ "bitflags 2.13.0",
+ "libc",
+]
+
+[[package]]
+name = "lazy_static"
+version = "1.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bbd2bcb4c963f2ddae06a2efc7e9f3591312473c50c6685e1f298068316e66fe"
+
+[[package]]
+name = "libc"
+version = "0.2.186"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "68ab91017fe16c622486840e4c83c9a37afeff978bd239b5293d61ece587de66"
+
+[[package]]
+name = "libsqlite3-sys"
+version = "0.30.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2e99fb7a497b1e3339bc746195567ed8d3e24945ecd636e3619d20b9de9e9149"
+dependencies = [
+ "cc",
+ "pkg-config",
+ "vcpkg",
+]
+
+[[package]]
+name = "linux-raw-sys"
+version = "0.12.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "32a66949e030da00e8c7d4434b251670a91556f4144941d37452769c25d58a53"
+
+[[package]]
+name = "lock_api"
+version = "0.4.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "224399e74b87b5f3557511d98dff8b14089b3dadafcab6bb93eab67d3aace965"
+dependencies = [
+ "scopeguard",
+]
+
+[[package]]
+name = "log"
+version = "0.4.33"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0ceec5bc11778974d1bcb055b18002eba7f4b3518b6a0081b3af5f21666da9ad"
+
+[[package]]
+name = "matchers"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d1525a2a28c7f4fa0fc98bb91ae755d1e2d1505079e05539e35bc876b5d65ae9"
+dependencies = [
+ "regex-automata",
+]
+
+[[package]]
+name = "memchr"
+version = "2.8.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "88904434abc2901f197fe8cc55f0445e7ded921dba5911dad2e2b39b48e663c4"
+
+[[package]]
+name = "memscribe-adapters"
+version = "0.1.0"
+dependencies = [
+ "blake3",
+ "memscribe-core",
+ "proptest",
+ "regex",
+ "serde",
+ "serde_json",
+ "time",
+ "walkdir",
+ "zstd",
+]
+
+[[package]]
+name = "memscribe-cli"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "assert_cmd",
+ "clap",
+ "ctrlc",
+ "memscribe-adapters",
+ "memscribe-core",
+ "memscribe-io",
+ "memscribe-sink",
+ "memscribe-testkit",
+ "predicates",
+ "serde",
+ "serde_json",
+ "tempfile",
+ "toml",
+ "tracing",
+ "tracing-subscriber",
+ "zstd",
+]
+
+[[package]]
+name = "memscribe-core"
+version = "0.1.0"
+dependencies = [
+ "blake3",
+ "proptest",
+ "regex",
+ "serde",
+ "serde_json",
+ "thiserror",
+ "time",
+]
+
+[[package]]
+name = "memscribe-io"
+version = "0.1.0"
+dependencies = [
+ "memscribe-core",
+ "notify",
+ "notify-debouncer-full",
+ "rusqlite",
+ "serde",
+ "serde_json",
+ "tempfile",
+ "time",
+ "tokio",
+ "tracing",
+ "walkdir",
+ "zstd",
+]
+
+[[package]]
+name = "memscribe-sink"
+version = "0.1.0"
+dependencies = [
+ "memscribe-core",
+ "rusqlite",
+ "serde",
+ "serde_json",
+ "tempfile",
+ "time",
+]
+
+[[package]]
+name = "memscribe-testkit"
+version = "0.1.0"
+dependencies = [
+ "insta",
+ "memscribe-adapters",
+ "memscribe-core",
+ "memscribe-io",
+ "memscribe-sink",
+ "proptest",
+ "serde",
+ "serde_json",
+]
+
+[[package]]
+name = "mio"
+version = "0.8.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a4a650543ca06a924e8b371db273b2756685faae30f8487da1b56505a8f78b0c"
+dependencies = [
+ "libc",
+ "log",
+ "wasi",
+ "windows-sys 0.48.0",
+]
+
+[[package]]
+name = "nix"
+version = "0.31.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cf20d2fde8ff38632c426f1165ed7436270b44f199fc55284c38276f9db47c3d"
+dependencies = [
+ "bitflags 2.13.0",
+ "cfg-if",
+ "cfg_aliases",
+ "libc",
+]
+
+[[package]]
+name = "normalize-line-endings"
+version = "0.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "61807f77802ff30975e01f4f071c8ba10c022052f98b3294119f3e615d13e5be"
+
+[[package]]
+name = "notify"
+version = "6.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6205bd8bb1e454ad2e27422015fb5e4f2bcc7e08fa8f27058670d208324a4d2d"
+dependencies = [
+ "bitflags 2.13.0",
+ "crossbeam-channel",
+ "filetime",
+ "fsevent-sys",
+ "inotify",
+ "kqueue",
+ "libc",
+ "log",
+ "mio",
+ "walkdir",
+ "windows-sys 0.48.0",
+]
+
+[[package]]
+name = "notify-debouncer-full"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fb7fd166739789c9ff169e654dc1501373db9d80a4c3f972817c8a4d7cf8f34e"
+dependencies = [
+ "crossbeam-channel",
+ "file-id",
+ "log",
+ "notify",
+ "parking_lot",
+ "walkdir",
+]
+
+[[package]]
+name = "nu-ansi-term"
+version = "0.50.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7957b9740744892f114936ab4a57b3f487491bbeafaf8083688b16841a4240e5"
+dependencies = [
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "num-conv"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "521739c6d2bac4aa25192232afe6841231376b2b26d4d9fae5ecf8ca5772e441"
+
+[[package]]
+name = "num-traits"
+version = "0.2.19"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "071dfc062690e90b734c0b2273ce72ad0ffa95f0c74596bc250dcfd960262841"
+dependencies = [
+ "autocfg",
+]
+
+[[package]]
+name = "objc2"
+version = "0.6.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3a12a8ed07aefc768292f076dc3ac8c48f3781c8f2d5851dd3d98950e8c5a89f"
+dependencies = [
+ "objc2-encode",
+]
+
+[[package]]
+name = "objc2-encode"
+version = "4.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ef25abbcd74fb2609453eb695bd2f860d389e457f67dc17cafc8b8cbc89d0c33"
+
+[[package]]
+name = "once_cell"
+version = "1.21.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50"
+
+[[package]]
+name = "once_cell_polyfill"
+version = "1.70.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
+
+[[package]]
+name = "parking_lot"
+version = "0.12.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "93857453250e3077bd71ff98b6a65ea6621a19bb0f559a85248955ac12c45a1a"
+dependencies = [
+ "lock_api",
+ "parking_lot_core",
+]
+
+[[package]]
+name = "parking_lot_core"
+version = "0.9.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2621685985a2ebf1c516881c026032ac7deafcda1a2c9b7850dc81e3dfcb64c1"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "redox_syscall",
+ "smallvec",
+ "windows-link",
+]
+
+[[package]]
+name = "pin-project-lite"
+version = "0.2.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a89322df9ebe1c1578d689c92318e070967d1042b512afbe49518723f4e6d5cd"
+
+[[package]]
+name = "pkg-config"
+version = "0.3.33"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "19f132c84eca552bf34cab8ec81f1c1dcc229b811638f9d283dceabe58c5569e"
+
+[[package]]
+name = "powerfmt"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "439ee305def115ba05938db6eb1644ff94165c5ab5e9420d1c1bcedbba909391"
+
+[[package]]
+name = "ppv-lite86"
+version = "0.2.21"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "85eae3c4ed2f50dcfe72643da4befc30deadb458a9b590d720cde2f2b1e97da9"
+dependencies = [
+ "zerocopy",
+]
+
+[[package]]
+name = "predicates"
+version = "3.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ada8f2932f28a27ee7b70dd6c1c39ea0675c55a36879ab92f3a715eaa1e63cfe"
+dependencies = [
+ "anstyle",
+ "difflib",
+ "float-cmp",
+ "normalize-line-endings",
+ "predicates-core",
+ "regex",
+]
+
+[[package]]
+name = "predicates-core"
+version = "1.0.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cad38746f3166b4031b1a0d39ad9f954dd291e7854fcc0eed52ee41a0b50d144"
+
+[[package]]
+name = "predicates-tree"
+version = "1.0.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d0de1b847b39c8131db0467e9df1ff60e6d0562ab8e9a16e568ad0fdb372e2f2"
+dependencies = [
+ "predicates-core",
+ "termtree",
+]
+
+[[package]]
+name = "proc-macro2"
+version = "1.0.106"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8fd00f0bb2e90d81d1044c2b32617f68fcb9fa3bb7640c23e9c748e53fb30934"
+dependencies = [
+ "unicode-ident",
+]
+
+[[package]]
+name = "proptest"
+version = "1.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4b45fcc2344c680f5025fe57779faef368840d0bd1f42f216291f0dc4ace4744"
+dependencies = [
+ "bit-set",
+ "bit-vec",
+ "bitflags 2.13.0",
+ "num-traits",
+ "rand",
+ "rand_chacha",
+ "rand_xorshift",
+ "regex-syntax",
+ "rusty-fork",
+ "tempfile",
+ "unarray",
+]
+
+[[package]]
+name = "quick-error"
+version = "1.2.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a1d01941d82fa2ab50be1e79e6714289dd7cde78eba4c074bc5a4374f650dfe0"
+
+[[package]]
+name = "quote"
+version = "1.0.46"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dfbc457d0c7a0759a614551b11a6409e5951f6c7537be1f1b7682b9ae9230368"
+dependencies = [
+ "proc-macro2",
+]
+
+[[package]]
+name = "r-efi"
+version = "5.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "69cdb34c158ceb288df11e18b4bd39de994f6657d83847bdffdbd7f346754b0f"
+
+[[package]]
+name = "r-efi"
+version = "6.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f8dcc9c7d52a811697d2151c701e0d08956f92b0e24136cf4cf27b57a6a0d9bf"
+
+[[package]]
+name = "rand"
+version = "0.9.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "44c5af06bb1b7d3216d91932aed5265164bf384dc89cd6ba05cf59a35f5f76ea"
+dependencies = [
+ "rand_chacha",
+ "rand_core",
+]
+
+[[package]]
+name = "rand_chacha"
+version = "0.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d3022b5f1df60f26e1ffddd6c66e8aa15de382ae63b3a0c1bfc0e4d3e3f325cb"
+dependencies = [
+ "ppv-lite86",
+ "rand_core",
+]
+
+[[package]]
+name = "rand_core"
+version = "0.9.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "76afc826de14238e6e8c374ddcc1fa19e374fd8dd986b0d2af0d02377261d83c"
+dependencies = [
+ "getrandom 0.3.4",
+]
+
+[[package]]
+name = "rand_xorshift"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "513962919efc330f829edb2535844d1b912b0fbe2ca165d613e4e8788bb05a5a"
+dependencies = [
+ "rand_core",
+]
+
+[[package]]
+name = "redox_syscall"
+version = "0.5.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ed2bf2547551a7053d6fdfafda3f938979645c44812fbfcda098faae3f1a362d"
+dependencies = [
+ "bitflags 2.13.0",
+]
+
+[[package]]
+name = "regex"
+version = "1.12.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f1292b7759ae1cb9ec195452d1390a074f0cd8541ab7a5a8c31cd6db45d4a6ba"
+dependencies = [
+ "aho-corasick",
+ "memchr",
+ "regex-automata",
+ "regex-syntax",
+]
+
+[[package]]
+name = "regex-automata"
+version = "0.4.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6e1dd4122fc1595e8162618945476892eefca7b88c52820e74af6262213cae8f"
+dependencies = [
+ "aho-corasick",
+ "memchr",
+ "regex-syntax",
+]
+
+[[package]]
+name = "regex-syntax"
+version = "0.8.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d6f6ff9a378485b298a5286656da665ba74413d36db0979633275d2e708145d4"
+
+[[package]]
+name = "rusqlite"
+version = "0.32.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7753b721174eb8ff87a9a0e799e2d7bc3749323e773db92e0984debb00019d6e"
+dependencies = [
+ "bitflags 2.13.0",
+ "fallible-iterator",
+ "fallible-streaming-iterator",
+ "hashlink",
+ "libsqlite3-sys",
+ "smallvec",
+]
+
+[[package]]
+name = "rustix"
+version = "1.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b6fe4565b9518b83ef4f91bb47ce29620ca828bd32cb7e408f0062e9930ba190"
+dependencies = [
+ "bitflags 2.13.0",
+ "errno",
+ "libc",
+ "linux-raw-sys",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "rusty-fork"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cc6bf79ff24e648f6da1f8d1f011e9cac26491b619e6b9280f2b47f1774e6ee2"
+dependencies = [
+ "fnv",
+ "quick-error",
+ "tempfile",
+ "wait-timeout",
+]
+
+[[package]]
+name = "same-file"
+version = "1.0.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "93fc1dc3aaa9bfed95e02e6eadabb4baf7e3078b0bd1b4d7b6b0b68378900502"
+dependencies = [
+ "winapi-util",
+]
+
+[[package]]
+name = "scopeguard"
+version = "1.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "94143f37725109f92c262ed2cf5e59bce7498c01bcc1502d7b9afe439a4e9f49"
+
+[[package]]
+name = "serde"
+version = "1.0.228"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9a8e94ea7f378bd32cbbd37198a4a91436180c5bb472411e48b5ec2e2124ae9e"
+dependencies = [
+ "serde_core",
+ "serde_derive",
+]
+
+[[package]]
+name = "serde_core"
+version = "1.0.228"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "41d385c7d4ca58e59fc732af25c3983b67ac852c1a25000afe1175de458b67ad"
+dependencies = [
+ "serde_derive",
+]
+
+[[package]]
+name = "serde_derive"
+version = "1.0.228"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d540f220d3187173da220f885ab66608367b6574e925011a9353e4badda91d79"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "serde_json"
+version = "1.0.150"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e8014e44b4736ed0538adeecded0fce2a272f22dc9578a7eb6b2d9993c74cfb9"
+dependencies = [
+ "indexmap",
+ "itoa",
+ "memchr",
+ "serde",
+ "serde_core",
+ "zmij",
+]
+
+[[package]]
+name = "serde_spanned"
+version = "0.6.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bf41e0cfaf7226dca15e8197172c295a782857fcb97fad1808a166870dee75a3"
+dependencies = [
+ "serde",
+]
+
+[[package]]
+name = "sharded-slab"
+version = "0.1.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f40ca3c46823713e0d4209592e8d6e826aa57e928f09752619fc696c499637f6"
+dependencies = [
+ "lazy_static",
+]
+
+[[package]]
+name = "shlex"
+version = "2.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f8fadd59c855ef2080decdef8ff161eb6661b86933c9d82e5ba29dc602a55aba"
+
+[[package]]
+name = "similar"
+version = "2.7.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bbbb5d9659141646ae647b42fe094daf6c6192d1620870b449d9557f748b2daa"
+
+[[package]]
+name = "smallvec"
+version = "1.15.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8ed6a63f02c8539c91a8685a86f4099661ba3da017932f6ebbea6de3f0fa7c90"
+
+[[package]]
+name = "strsim"
+version = "0.11.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7da8b5736845d9f2fcb837ea5d9e2628564b3b043a70948a3f0b778838c5fb4f"
+
+[[package]]
+name = "syn"
+version = "2.0.118"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1b9ae57f904213ebb649ce6895b8a66c66f0203b9319718f69a5612a065b1422"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "unicode-ident",
+]
+
+[[package]]
+name = "tempfile"
+version = "3.27.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "32497e9a4c7b38532efcdebeef879707aa9f794296a4f0244f6f69e9bc8574bd"
+dependencies = [
+ "fastrand",
+ "getrandom 0.4.3",
+ "once_cell",
+ "rustix",
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "termtree"
+version = "0.5.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8f50febec83f5ee1df3015341d8bd429f2d1cc62bcba7ea2076759d315084683"
+
+[[package]]
+name = "thiserror"
+version = "2.0.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4288b5bcbc7920c07a1149a35cf9590a2aa808e0bc1eafaade0b80947865fbc4"
+dependencies = [
+ "thiserror-impl",
+]
+
+[[package]]
+name = "thiserror-impl"
+version = "2.0.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ebc4ee7f67670e9b64d05fa4253e753e016c6c95ff35b89b7941d6b856dec1d5"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "thread_local"
+version = "1.1.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f60246a4944f24f6e018aa17cdeffb7818b76356965d03b07d6a9886e8962185"
+dependencies = [
+ "cfg-if",
+]
+
+[[package]]
+name = "time"
+version = "0.3.51"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "85c17d80feb7334b40c484e45ed1a5273dfd8bfda537c3be2e74a06a6686f327"
+dependencies = [
+ "deranged",
+ "num-conv",
+ "powerfmt",
+ "serde_core",
+ "time-core",
+ "time-macros",
+]
+
+[[package]]
+name = "time-core"
+version = "0.1.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9e1c906769ad99c88eaa54e728060edef082f8e358ff32030cb7c7d315e81109"
+
+[[package]]
+name = "time-macros"
+version = "0.2.30"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dcef1a61bdb119096e153208ec5cbec23944ce8bca13be5c7f60c634f7403935"
+dependencies = [
+ "num-conv",
+ "time-core",
+]
+
+[[package]]
+name = "tokio"
+version = "1.52.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8fc7f01b389ac15039e4dc9531aa973a135d7a4135281b12d7c1bc79fd57fffe"
+dependencies = [
+ "bytes",
+ "pin-project-lite",
+ "tokio-macros",
+]
+
+[[package]]
+name = "tokio-macros"
+version = "2.7.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "385a6cb71ab9ab790c5fe8d67f1645e6c450a7ce006a33de03daa956cf70a496"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "toml"
+version = "0.8.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dc1beb996b9d83529a9e75c17a1686767d148d70663143c7854d8b4a09ced362"
+dependencies = [
+ "serde",
+ "serde_spanned",
+ "toml_datetime",
+ "toml_edit",
+]
+
+[[package]]
+name = "toml_datetime"
+version = "0.6.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "22cddaf88f4fbc13c51aebbf5f8eceb5c7c5a9da2ac40a13519eb5b0a0e8f11c"
+dependencies = [
+ "serde",
+]
+
+[[package]]
+name = "toml_edit"
+version = "0.22.27"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "41fe8c660ae4257887cf66394862d21dbca4a6ddd26f04a3560410406a2f819a"
+dependencies = [
+ "indexmap",
+ "serde",
+ "serde_spanned",
+ "toml_datetime",
+ "toml_write",
+ "winnow",
+]
+
+[[package]]
+name = "toml_write"
+version = "0.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5d99f8c9a7727884afe522e9bd5edbfc91a3312b36a77b5fb8926e4c31a41801"
+
+[[package]]
+name = "tracing"
+version = "0.1.44"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "63e71662fa4b2a2c3a26f570f037eb95bb1f85397f3cd8076caed2f026a6d100"
+dependencies = [
+ "pin-project-lite",
+ "tracing-attributes",
+ "tracing-core",
+]
+
+[[package]]
+name = "tracing-attributes"
+version = "0.1.31"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7490cfa5ec963746568740651ac6781f701c9c5ea257c58e057f3ba8cf69e8da"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "tracing-core"
+version = "0.1.36"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "db97caf9d906fbde555dd62fa95ddba9eecfd14cb388e4f491a66d74cd5fb79a"
+dependencies = [
+ "once_cell",
+ "valuable",
+]
+
+[[package]]
+name = "tracing-log"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ee855f1f400bd0e5c02d150ae5de3840039a3f54b025156404e34c23c03f47c3"
+dependencies = [
+ "log",
+ "once_cell",
+ "tracing-core",
+]
+
+[[package]]
+name = "tracing-subscriber"
+version = "0.3.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cb7f578e5945fb242538965c2d0b04418d38ec25c79d160cd279bf0731c8d319"
+dependencies = [
+ "matchers",
+ "nu-ansi-term",
+ "once_cell",
+ "regex-automata",
+ "sharded-slab",
+ "smallvec",
+ "thread_local",
+ "tracing",
+ "tracing-core",
+ "tracing-log",
+]
+
+[[package]]
+name = "unarray"
+version = "0.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "eaea85b334db583fe3274d12b4cd1880032beab409c0d774be044d4480ab9a94"
+
+[[package]]
+name = "unicode-ident"
+version = "1.0.24"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e6e4313cd5fcd3dad5cafa179702e2b244f760991f45397d14d4ebf38247da75"
+
+[[package]]
+name = "utf8parse"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821"
+
+[[package]]
+name = "valuable"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ba73ea9cf16a25df0c8caa16c51acb937d5712a8429db78a3ee29d5dcacd3a65"
+
+[[package]]
+name = "vcpkg"
+version = "0.2.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "accd4ea62f7bb7a82fe23066fb0957d48ef677f6eeb8215f372f52e48bb32426"
+
+[[package]]
+name = "version_check"
+version = "0.9.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0b928f33d975fc6ad9f86c8f283853ad26bdd5b10b7f1542aa2fa15e2289105a"
+
+[[package]]
+name = "wait-timeout"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09ac3b126d3914f9849036f826e054cbabdc8519970b8998ddaf3b5bd3c65f11"
+dependencies = [
+ "libc",
+]
+
+[[package]]
+name = "walkdir"
+version = "2.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "29790946404f91d9c5d06f9874efddea1dc06c5efe94541a7d6863108e3a5e4b"
+dependencies = [
+ "same-file",
+ "winapi-util",
+]
+
+[[package]]
+name = "wasi"
+version = "0.11.1+wasi-snapshot-preview1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ccf3ec651a847eb01de73ccad15eb7d99f80485de043efb2f370cd654f4ea44b"
+
+[[package]]
+name = "wasip2"
+version = "1.0.4+wasi-0.2.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b67efb37e106e55ce722a510d6b5f9c17f083e5fc79afc2badeb12cc313d9487"
+dependencies = [
+ "wit-bindgen",
+]
+
+[[package]]
+name = "winapi-util"
+version = "0.1.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
+dependencies = [
+ "windows-sys 0.61.2",
+]
+
+[[package]]
+name = "windows-link"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f0805222e57f7521d6a62e36fa9163bc891acd422f971defe97d64e70d0a4fe5"
+
+[[package]]
+name = "windows-sys"
+version = "0.48.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "677d2418bec65e3338edb076e806bc1ec15693c5d0104683f2efe857f61056a9"
+dependencies = [
+ "windows-targets 0.48.5",
+]
+
+[[package]]
+name = "windows-sys"
+version = "0.60.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f2f500e4d28234f72040990ec9d39e3a6b950f9f22d3dba18416c35882612bcb"
+dependencies = [
+ "windows-targets 0.53.5",
+]
+
+[[package]]
+name = "windows-sys"
+version = "0.61.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ae137229bcbd6cdf0f7b80a31df61766145077ddf49416a728b02cb3921ff3fc"
+dependencies = [
+ "windows-link",
+]
+
+[[package]]
+name = "windows-targets"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9a2fa6e2155d7247be68c096456083145c183cbbbc2764150dda45a87197940c"
+dependencies = [
+ "windows_aarch64_gnullvm 0.48.5",
+ "windows_aarch64_msvc 0.48.5",
+ "windows_i686_gnu 0.48.5",
+ "windows_i686_msvc 0.48.5",
+ "windows_x86_64_gnu 0.48.5",
+ "windows_x86_64_gnullvm 0.48.5",
+ "windows_x86_64_msvc 0.48.5",
+]
+
+[[package]]
+name = "windows-targets"
+version = "0.53.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4945f9f551b88e0d65f3db0bc25c33b8acea4d9e41163edf90dcd0b19f9069f3"
+dependencies = [
+ "windows-link",
+ "windows_aarch64_gnullvm 0.53.1",
+ "windows_aarch64_msvc 0.53.1",
+ "windows_i686_gnu 0.53.1",
+ "windows_i686_gnullvm",
+ "windows_i686_msvc 0.53.1",
+ "windows_x86_64_gnu 0.53.1",
+ "windows_x86_64_gnullvm 0.53.1",
+ "windows_x86_64_msvc 0.53.1",
+]
+
+[[package]]
+name = "windows_aarch64_gnullvm"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2b38e32f0abccf9987a4e3079dfb67dcd799fb61361e53e2882c3cbaf0d905d8"
+
+[[package]]
+name = "windows_aarch64_gnullvm"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a9d8416fa8b42f5c947f8482c43e7d89e73a173cead56d044f6a56104a6d1b53"
+
+[[package]]
+name = "windows_aarch64_msvc"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dc35310971f3b2dbbf3f0690a219f40e2d9afcf64f9ab7cc1be722937c26b4bc"
+
+[[package]]
+name = "windows_aarch64_msvc"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b9d782e804c2f632e395708e99a94275910eb9100b2114651e04744e9b125006"
+
+[[package]]
+name = "windows_i686_gnu"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a75915e7def60c94dcef72200b9a8e58e5091744960da64ec734a6c6e9b3743e"
+
+[[package]]
+name = "windows_i686_gnu"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "960e6da069d81e09becb0ca57a65220ddff016ff2d6af6a223cf372a506593a3"
+
+[[package]]
+name = "windows_i686_gnullvm"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fa7359d10048f68ab8b09fa71c3daccfb0e9b559aed648a8f95469c27057180c"
+
+[[package]]
+name = "windows_i686_msvc"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8f55c233f70c4b27f66c523580f78f1004e8b5a8b659e05a4eb49d4166cca406"
+
+[[package]]
+name = "windows_i686_msvc"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1e7ac75179f18232fe9c285163565a57ef8d3c89254a30685b57d83a38d326c2"
+
+[[package]]
+name = "windows_x86_64_gnu"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "53d40abd2583d23e4718fddf1ebec84dbff8381c07cae67ff7768bbf19c6718e"
+
+[[package]]
+name = "windows_x86_64_gnu"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9c3842cdd74a865a8066ab39c8a7a473c0778a3f29370b5fd6b4b9aa7df4a499"
+
+[[package]]
+name = "windows_x86_64_gnullvm"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0b7b52767868a23d5bab768e390dc5f5c55825b6d30b86c844ff2dc7414044cc"
+
+[[package]]
+name = "windows_x86_64_gnullvm"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0ffa179e2d07eee8ad8f57493436566c7cc30ac536a3379fdf008f47f6bb7ae1"
+
+[[package]]
+name = "windows_x86_64_msvc"
+version = "0.48.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ed94fce61571a4006852b7389a063ab983c02eb1bb37b47f8272ce92d06d9538"
+
+[[package]]
+name = "windows_x86_64_msvc"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d6bbff5f0aada427a1e5a6da5f1f98158182f26556f345ac9e04d36d0ebed650"
+
+[[package]]
+name = "winnow"
+version = "0.7.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "df79d97927682d2fd8adb29682d1140b343be4ac0f08fd68b7765d9c059d3945"
+dependencies = [
+ "memchr",
+]
+
+[[package]]
+name = "wit-bindgen"
+version = "0.57.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1ebf944e87a7c253233ad6766e082e3cd714b5d03812acc24c318f549614536e"
+
+[[package]]
+name = "zerocopy"
+version = "0.8.52"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ce1022995ff5ff5d841ad7d994facc23098cd40152f2c1d11cd607c6f530653f"
+dependencies = [
+ "zerocopy-derive",
+]
+
+[[package]]
+name = "zerocopy-derive"
+version = "0.8.52"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1ae7f38b72ec2a254e2b87ef277cf2cd4fb97cbebf944faa6f33354da0867930"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "zmij"
+version = "1.0.21"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b8848ee67ecc8aedbaf3e4122217aff892639231befc6a1b58d29fff4c2cabaa"
+
+[[package]]
+name = "zstd"
+version = "0.13.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e91ee311a569c327171651566e07972200e76fcfe2242a4fa446149a3881c08a"
+dependencies = [
+ "zstd-safe",
+]
+
+[[package]]
+name = "zstd-safe"
+version = "7.2.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8f49c4d5f0abb602a93fb8736af2a4f4dd9512e36f7f570d66e65ff867ed3b9d"
+dependencies = [
+ "zstd-sys",
+]
+
+[[package]]
+name = "zstd-sys"
+version = "2.0.16+zstd.1.5.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "91e19ebc2adc8f83e43039e79776e3fda8ca919132d68a1fed6a5faca2683748"
+dependencies = [
+ "cc",
+ "pkg-config",
+]
diff --git a/Cargo.toml b/Cargo.toml
new file mode 100644
index 0000000..af12e73
--- /dev/null
+++ b/Cargo.toml
@@ -0,0 +1,74 @@
+# Memscribe — a self-contained, open-source Rust workspace for deterministic,
+# zero-LLM conversation capture from AI coding agents.
+#
+# The dependency direction is strictly one-way: Memtrace depends on Memscribe,
+# never the reverse. See the implementation whitepaper, §2 and §9.
+
+[workspace]
+resolver = "2"
+members = [
+    "crates/memscribe-core",
+    "crates/memscribe-adapters",
+    "crates/memscribe-io",
+    "crates/memscribe-sink",
+    "crates/memscribe-cli",
+    "crates/memscribe-testkit",
+]
+# The cargo-fuzz crate is a standalone, workspace-EXCLUDED member: it depends on
+# libfuzzer-sys (a nightly-only sanitizer runtime) and must not be pulled into a
+# plain `cargo build`/`cargo test` of the workspace. See fuzz/README.md.
+exclude = ["fuzz"]
+
+[workspace.package]
+version = "0.1.0"
+edition = "2021"
+rust-version = "1.96"
+license = "MIT OR Apache-2.0"
+repository = "https://github.com/Alex793x/Memscribe"
+authors = ["Memrack / Syncable"]
+description = "Deterministic, zero-LLM transcript capture and node preparation for AI coding agents."
+
+# Shared dependency versions. Pinned to major lines; exact versions resolve in
+# Cargo.lock for reproducible builds (see the whitepaper's recommended stack).
+[workspace.dependencies]
+serde = { version = "1", features = ["derive"] }
+serde_json = { version = "1", features = ["preserve_order"] }
+time = { version = "0.3", features = ["serde", "serde-well-known", "formatting", "parsing", "macros"] }
+thiserror = "2"
+blake3 = "1"
+regex = "1"
+tracing = "0.1"
+tracing-subscriber = { version = "0.3", features = ["env-filter"] }
+tokio = { version = "1", features = ["rt", "rt-multi-thread", "macros", "fs", "io-util", "sync", "time"] }
+futures = "0.3"
+notify = "6"
+notify-debouncer-full = "0.3"
+zstd = "0.13"
+rusqlite = { version = "0.32", features = ["bundled"] }
+clap = { version = "4", features = ["derive"] }
+anyhow = "1"
+similar = "2"
+walkdir = "2"
+# dev / test
+proptest = "1"
+insta = { version = "1", features = ["json"] }
+assert_cmd = "2"
+predicates = "3"
+tempfile = "3"
+
+# Workspace-wide member crates depend on each other through these handles.
+memscribe-core = { path = "crates/memscribe-core", version = "0.1.0" }
+memscribe-adapters = { path = "crates/memscribe-adapters", version = "0.1.0" }
+memscribe-io = { path = "crates/memscribe-io", version = "0.1.0" }
+memscribe-sink = { path = "crates/memscribe-sink", version = "0.1.0" }
+memscribe-testkit = { path = "crates/memscribe-testkit", version = "0.1.0" }
+
+[workspace.lints.rust]
+unsafe_code = "forbid"
+
+[workspace.lints.clippy]
+# Keep the deterministic core honest without drowning contributors in pedantry.
+all = "warn"
+
+[profile.release]
+lto = "thin"
diff --git a/LICENSE b/LICENSE-APACHE
similarity index 100%
rename from LICENSE
rename to LICENSE-APACHE
diff --git a/LICENSE-MIT b/LICENSE-MIT
new file mode 100644
index 0000000..969d061
--- /dev/null
+++ b/LICENSE-MIT
@@ -0,0 +1,17 @@
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/README.md b/README.md
new file mode 100644
index 0000000..0f7aca7
--- /dev/null
+++ b/README.md
@@ -0,0 +1,190 @@
+<div align="center">
+
+<img src="assets/memscribe-logo.svg" alt="Memscribe" width="132" height="132" />
+
+# Memscribe
+
+**Deterministic, zero-LLM conversation capture for AI coding agents.**
+
+Memscribe tails the transcript logs your AI coding agents already write — Claude Code, Codex, Gemini, Cursor, Windsurf, Zed, VS Code / Copilot, and any OpenTelemetry-instrumented agent — and prepares them into typed, queryable nodes. No model calls. Same bytes in, same nodes out, every time.
+
+[![CI](https://github.com/Alex793x/Memscribe/actions/workflows/ci.yml/badge.svg)](https://github.com/Alex793x/Memscribe/actions/workflows/ci.yml)
+[![License: MIT OR Apache-2.0](https://img.shields.io/badge/license-MIT%20OR%20Apache--2.0-blue.svg)](#license)
+[![Rust 1.96+](https://img.shields.io/badge/rust-1.96%2B-orange.svg)](rust-toolchain.toml)
+[![Tests](https://img.shields.io/badge/tests-260%20passing-success.svg)](#testing)
+[![Zero-LLM](https://img.shields.io/badge/LLM%20calls-0-1a1a2e.svg)](#why-deterministic-matters)
+
+</div>
+
+---
+
+## What it is
+
+A coding agent is a stream of decisions and edits — *"use Postgres instead of MySQL,"* followed by the diffs that implement it. That stream is gold for memory, audit, analytics, and replay, but it's buried in five different churning log formats. **Memscribe is the boring, deterministic half of a memory system:** it reads those logs, normalizes them, and emits typed nodes — and because it never calls a model, its output is an exact function of its input.
+
+That single property is the whole point. It makes capture **golden-file, property, and fuzz testable**, so the day a tool changes its format, the test suite fails loudly instead of silently corrupting your memory.
+
+### Where Memscribe sits
+
+Memscribe is the foundation of a three-layer stack. Each layer uses the one below it, and the dependency only ever points **downward**:
+
+| Layer | Role | Calls a model? |
+|:------|:-----|:--------------:|
+| [**Memtrace**](https://github.com/syncable-dev/memtrace) | The product — a code-intelligence graph with agent memory | — |
+| **MemCortex** | Inference & governance — the judgment calls on top of the captured data | yes |
+| **Memscribe** | Deterministic capture — normalizes transcripts into typed nodes *(this repo)* | **no** |
+
+Memtrace builds on MemCortex; MemCortex builds on Memscribe. Because Memscribe sits at the bottom, depends on nothing above it, and never calls a model, the boundary between the layers is a single stable data type — which is exactly what keeps this layer small, auditable, and exhaustively testable.
+
+## The pipeline
+
+One linear, deterministic pipeline. Every stage is a trait, so it can be tested in isolation and swapped.
+
+```
+  Source (memscribe-io)          Adapter (memscribe-adapters)
+  tail JSONL / hook stdin   ─►   parse_line ─► CaptureEvent[]
+  / OTLP receiver                (version-tolerant)
+         │  RawRecord(bytes + provenance)        │  normalized events
+         ▼                                       ▼
+  Gate ─► Segmenter ─► Binder ─► NodePrep   ─►   Sink (memscribe-sink)
+  admit?   arc / turn  decision   Prepared        MemDB · ndjson · sqlite
+  markers  spans       ↔ edit     Node
+```
+
+`Source → Adapter` produces a normalized `CaptureEvent` stream — the system of record. `Gate → Segmenter → Binder → NodePrep` turn that into `PreparedNode`s. The `Sink` writes them. Everything between Source and Sink is pure and synchronous given the event stream, which is what makes the whole thing golden-testable end to end.
+
+## Quick start
+
+```bash
+# Parse a transcript to NDJSON (the workhorse — great for trying it out)
+cargo run -p memscribe-cli -- parse ~/.claude/projects/<slug>/<session>.jsonl --as claude_code
+
+# Tail your agents live and write prepared nodes to a local SQLite store
+cargo run -p memscribe-cli -- watch --tools claude,codex,gemini --sink sqlite --out memory.db
+
+# See exactly what the redaction pass would strip from a file
+cargo run -p memscribe-cli -- redact session.jsonl
+```
+
+Every tool's transcript normalizes to the **same shape**. Here a Claude Code decision-and-edits session becomes four kinds of node:
+
+```jsonc
+// memscribe parse fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl --as claude_code
+{"node":"conversation","text":"Let's use Postgres instead of MySQL for the orders service.",
+ "markers":[{"rule_id":"decision_verb.use",...},{"rule_id":"rejection.instead_of",...}], "fact_status":"observed"}
+{"node":"decision","epitome":"Let's use Postgres instead of MySQL ...",
+ "considered_options":[{"text":"MySQL","chosen":false},{"text":"Postgres","chosen":true}],"is_ban":false}
+{"node":"episode","path":"src/db/config.rs","diff":{"added_lines":1,"removed_lines":1,...}}
+{"node":"binding","relation":"produced","prov":{"t_use":"...10:00:00Z","t_gen":"...10:00:03Z"},
+ "fact_status":"deterministically_derived","correlation":{...}}
+```
+
+## Supported tools
+
+Nine version-tolerant adapters, each behind a Cargo feature flag. Parsers pattern-match the fields they need and route anything unrecognized to `Unknown` — they never panic and never drop a record.
+
+| Tool | Transcript source | Status |
+|:-----|:------------------|:-------|
+| **Claude Code** | `~/.claude/projects/<slug>/<session>.jsonl` (append-only JSONL, DAG via `parentUuid`) | ✅ native |
+| **Codex CLI** | `~/.codex/sessions/.../rollout-*.jsonl[.zst]` (`apply_patch` V4A diffs, transparent zstd) | ✅ native |
+| **Gemini CLI** | `~/.gemini/tmp/<hash>/chats/session-*.jsonl` (`$set` / `$rewindTo` control lines) | ✅ native |
+| **OpenTelemetry** | OTLP / GenAI semconv records — the universal fallback for any instrumented agent | ✅ native |
+| **Cursor** · **Windsurf** · **Zed** · **VS Code** · **Copilot** | exported chat JSON (desktop stores are SQLite/undocumented — export-based, per the whitepaper) | ✅ export-shape |
+
+All five **CLI/OTel** scenarios and the cross-tool conformance suite prove these adapters are interchangeable behind the contract.
+
+## Usable with MemDB — and fully usable without it
+
+The seam is the `Sink` trait. Nothing in the pipeline knows what a sink does with a node:
+
+```rust
+pub trait Sink: Send {
+    fn emit(&mut self, node: &PreparedNode) -> Result<(), SinkError>;
+    fn flush(&mut self) -> Result<(), SinkError>;
+}
+```
+
+| Sink | Feature | Use |
+|:-----|:--------|:----|
+| `NdjsonSink` | default | One JSON node per line — the canonical, audit-friendly default. |
+| `SqliteSink` | default | A queryable local store with zero external services. |
+| `MemDbSink` | `--features memdb` | Writes nodes into MemDB with bi-temporal headers, for Memtrace. **Off by default.** |
+
+Remove the `memdb` feature and Memscribe is a complete, auditable, local capture tool. See [`crates/memscribe-sink/MEMDB.md`](crates/memscribe-sink/MEMDB.md) for the integration design.
+
+## The output contract
+
+Memscribe only ever emits nodes with `Observed` or `DeterministicallyDerived` fact-status. Anything that would require inference (fine-grained decision typing, statistical ranking) is **flagged for a downstream layer, never guessed.**
+
+| Node | Meaning | Fact status |
+|:-----|:--------|:------------|
+| `Conversation` | A gated, verbatim dialogue span with the commitment markers that fired | `Observed` |
+| `Decision` | Parsed deterministically (IBIS/QOC/MADR/Kruchten): epitome, options, `is_ban` | `Observed` |
+| `Episode` | The edit(s): path, diff, git sha | `DeterministicallyDerived` |
+| `Binding` | decision → episode, with PROV (`t_use ≤ t_gen`) + correlation tuple | `DeterministicallyDerived` |
+
+The **commitment-marker gate** (a config-driven, unit-tested rule table over decision verbs, rejections, bans, and imperatives) is the gate-before-store that the production audits showed is the difference between a working memory and a 97.8%-junk one.
+
+## Why deterministic matters
+
+| | Memscribe | LLM-based capture |
+|:--|:--|:--|
+| Output is a function of input | ✅ exact | ❌ varies run to run |
+| Golden / property / fuzz testable | ✅ | ❌ |
+| Cost per session | **$0.00** | API tokens |
+| Reads your prompts & secrets | locally, redacted, auditable | sent to a model |
+| Fails when a format changes | loudly (a test) | silently (bad data) |
+
+## Testing
+
+Because the pipeline is zero-LLM, it's tested the way a compiler is — fixtures in, exact expected output. The test corpus is a first-class deliverable.
+
+- **Golden-file / snapshot** tests per tool, version, and scenario (`insta`)
+- **Cross-tool conformance** — all 9 §8.2 scenarios (happy path, rejected alternative, ban, interleaved arcs, multi-edit, tool failure, rewind/compaction, subagent, no-marker) must normalize to the same shape regardless of tool
+- **Property tests** (`proptest`): determinism, idempotency, monotonic seq, losslessness, gate purity, offset resumption
+- **Fuzzing** (`cargo-fuzz`): one target per adapter — never panic, never loop, skip-and-continue
+- **Redaction & privacy**, **crash/resume**, and a **cross-version corpus**
+
+```bash
+cargo test --workspace --all-features          # 260 tests
+cargo clippy --workspace --all-targets --all-features -- -D warnings
+cargo deny check
+```
+
+## CLI
+
+| Command | What it does |
+|:--------|:-------------|
+| `memscribe watch [--tools …] [--sink …] [--out …]` | The steady-state capture daemon: tail transcripts, write nodes. |
+| `memscribe parse <file> [--as <tool>]` | One-shot parse a transcript to NDJSON (the workhorse for tests/debugging). |
+| `memscribe replay <file>` | Re-run preparation over a historical session. |
+| `memscribe verify [--capture <file> --as <tool>]` | Run the conformance summary; `--capture` snapshots a live session into a fixture. |
+| `memscribe redact <file> [--no-content]` | Show what the redaction pass would strip. |
+| `memscribe hook` | The hook handler agents invoke (reads stdin, records, returns immediately). |
+
+Configure per-tool path overrides, a custom commitment-marker table, redaction patterns, and the sink target in `memscribe.toml` — see [`memscribe.example.toml`](memscribe.example.toml).
+
+## Workspace layout
+
+| Crate | Responsibility |
+|:------|:---------------|
+| [`memscribe-core`](crates/memscribe-core) | The contract: model, traits, gate, segmenter, binder, node-prep, redaction. Depends on nothing in the workspace. |
+| [`memscribe-adapters`](crates/memscribe-adapters) | The 9 per-tool parsers, behind feature flags. |
+| [`memscribe-io`](crates/memscribe-io) | Sources: file reader, crash-safe offset tailer, live notify watcher, hook handler, OTLP receiver. |
+| [`memscribe-sink`](crates/memscribe-sink) | NDJSON, SQLite, and the feature-gated MemDB sink. |
+| [`memscribe-cli`](crates/memscribe-cli) | The `memscribe` binary. |
+| [`memscribe-testkit`](crates/memscribe-testkit) | Golden harness, conformance suite, synthetic generators, invariant checks. |
+
+See [`ARCHITECTURE.md`](ARCHITECTURE.md) for the deep dive and [`CONTRIBUTING.md`](CONTRIBUTING.md) to add an adapter.
+
+## Requirements
+
+- **Rust ≥ 1.96** (pinned in [`rust-toolchain.toml`](rust-toolchain.toml))
+- **Git** — for repo/branch binding on episodes
+- No network in the core path; the optional OTLP receiver binds to loopback only.
+
+## License
+
+Dual-licensed under either of [MIT](LICENSE-MIT) or [Apache-2.0](LICENSE-APACHE), at your option.
+
+<div align="center"><sub>Built by Memrack / Syncable · the deterministic half of the memory system.</sub></div>
diff --git a/SECURITY.md b/SECURITY.md
new file mode 100644
index 0000000..cb5aae9
--- /dev/null
+++ b/SECURITY.md
@@ -0,0 +1,27 @@
+# Security & Privacy
+
+Memscribe reads files that contain prompts, source code, and potentially
+secrets (API keys, tokens, `.env` contents). It is designed to be safe for
+security-conscious teams to run:
+
+- **Local-only, no network in the core path.** The Source → Adapter → Gate →
+  Segmenter → Binder → NodePrep → Sink pipeline performs no network I/O. The
+  optional OTLP receiver binds to loopback only. The optional MemDB sink is the
+  only component that talks to another process, and it is feature-gated off by
+  default.
+- **Redaction on by default.** Known secret patterns (API keys, bearer tokens,
+  `.env` assignments, private-key blocks) are stripped before the Sink. See
+  `memscribe redact <file>` to preview exactly what would be removed.
+- **`--no-content` mode.** Stores structure only (event kinds, spans, diffs
+  stats) with all verbatim text elided — for the most sensitive environments.
+- **Honors tool suppression switches.** When a tool exposes a privacy switch
+  (e.g. `CLAUDE_CODE_SKIP_PROMPT_HISTORY`), Memscribe respects it and does not
+  capture the suppressed content.
+- **Auditable.** Because the default sink is NDJSON and every node carries a
+  `SourceLocation` provenance pointer, you can audit exactly what was captured
+  and trace any node back to the byte range it came from.
+
+## Reporting a vulnerability
+
+Please report security issues privately to the maintainers via the repository's
+security advisory channel rather than a public issue.
diff --git a/assets/memscribe-logo.svg b/assets/memscribe-logo.svg
new file mode 100644
index 0000000..d4ce547
--- /dev/null
+++ b/assets/memscribe-logo.svg
@@ -0,0 +1,29 @@
+<svg width="480" height="480" viewBox="0 0 120 120" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Memscribe">
+  <!-- outer ring — matches the Memtrace mark -->
+  <circle cx="60" cy="60" r="54" fill="none" stroke="#0A1628" stroke-width="4"/>
+  <!-- faint inner family ring (shared with MemDB / MemFleet) -->
+  <circle cx="60" cy="60" r="45.5" fill="none" stroke="#00D4B8" stroke-width="1" opacity="0.28" stroke-dasharray="2 6"/>
+
+  <!-- fountain pen -->
+  <!-- barrel -->
+  <line x1="60" y1="29" x2="60" y2="61" stroke="#00D4B8" stroke-width="13" stroke-linecap="round"/>
+  <!-- barrel highlight -->
+  <line x1="56.4" y1="33" x2="56.4" y2="57" stroke="#5FE9D6" stroke-width="2.2" stroke-linecap="round" opacity="0.75"/>
+  <!-- cap band -->
+  <line x1="53.6" y1="38" x2="66.4" y2="38" stroke="#0A1628" stroke-width="3.4"/>
+  <!-- clip -->
+  <path d="M 64.8 33 L 64.8 45" fill="none" stroke="#0A1628" stroke-width="2" stroke-linecap="round" opacity="0.85"/>
+  <!-- nib -->
+  <path d="M 52.4 61 L 67.6 61 L 60 90 Z" fill="#00D4B8"/>
+  <!-- nib slit -->
+  <line x1="60" y1="69" x2="60" y2="86" stroke="#0A1628" stroke-width="2.6" stroke-linecap="round"/>
+  <!-- breather hole / navy accent dot -->
+  <circle cx="60" cy="69.5" r="2.7" fill="#0A1628"/>
+
+  <!-- written transcript line -->
+  <path d="M 41 98 Q 60 93.5 79 98" fill="none" stroke="#00D4B8" stroke-width="3" stroke-linecap="round"/>
+  <!-- captured nodes -->
+  <circle cx="46" cy="96.4" r="2.7" fill="#0A1628"/>
+  <circle cx="60" cy="94.8" r="2.7" fill="#0A1628"/>
+  <circle cx="74" cy="96.4" r="2.7" fill="#0A1628"/>
+</svg>
diff --git a/crates/memscribe-adapters/Cargo.toml b/crates/memscribe-adapters/Cargo.toml
new file mode 100644
index 0000000..825fc60
--- /dev/null
+++ b/crates/memscribe-adapters/Cargo.toml
@@ -0,0 +1,47 @@
+[package]
+name = "memscribe-adapters"
+version.workspace = true
+edition.workspace = true
+rust-version.workspace = true
+license.workspace = true
+repository.workspace = true
+description = "Per-tool transcript adapters for Memscribe (Claude Code, Codex, Gemini, OTel, Cursor, Windsurf, Zed, VS Code, Copilot)."
+
+[lints]
+workspace = true
+
+[features]
+default = [
+    "claude_code",
+    "codex",
+    "gemini",
+    "otel",
+    "cursor",
+    "windsurf",
+    "zed",
+    "vscode",
+    "copilot",
+]
+claude_code = []
+codex = ["dep:zstd"]
+gemini = []
+otel = []
+cursor = []
+windsurf = []
+zed = []
+vscode = []
+copilot = []
+
+[dependencies]
+memscribe-core = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+time = { workspace = true }
+regex = { workspace = true }
+blake3 = { workspace = true }
+walkdir = { workspace = true }
+zstd = { workspace = true, optional = true }
+
+[dev-dependencies]
+proptest = { workspace = true }
+serde_json = { workspace = true }
diff --git a/crates/memscribe-adapters/src/claude_code.rs b/crates/memscribe-adapters/src/claude_code.rs
new file mode 100644
index 0000000..3374e48
--- /dev/null
+++ b/crates/memscribe-adapters/src/claude_code.rs
@@ -0,0 +1,844 @@
+//! Claude Code adapter.
+//!
+//! Transcripts: `~/.claude/projects/<slug>/<session>.jsonl`, append-only JSONL,
+//! DAG via `parentUuid`. Dialogue: `type:user`/`assistant` records;
+//! `message.content[]` text/thinking/tool_use/tool_result; `model`, `usage`.
+//! Edits: `toolUseResult.structuredPatch` (Edit/Write/MultiEdit) → old/new/
+//! unified; `file-history-snapshot` baseline. Quirks: dedup by `uuid`; a
+//! session's summary may live in another file (join by `leafUuid`); subagents
+//! are separate `isSidechain` files.
+//!
+//! One on-disk record can normalize to several events: an assistant record with
+//! a `tool_use` block emits an [`EventKind::AssistantTurn`] plus an
+//! [`EventKind::ToolCall`]; a user record with a `tool_result` block emits an
+//! [`EventKind::UserTurn`] plus an [`EventKind::ToolResult`]; an edit record's
+//! top-level `toolUseResult.structuredPatch` emits an [`EventKind::FileEdit`].
+//! Every event for one record shares that record's `uuid` lineage; secondary
+//! events derive a deterministic id from `uuid` + a stable discriminator so they
+//! never collide. The whole record is deduplicated once, on its `uuid`.
+
+use crate::util;
+use memscribe_core::{
+    content_id, CaptureEvent, Diff, DiscoverCfg, EventKind, GitRef, ParseCtx, ParseError, Part,
+    ProjectRef, RawRecord, SchemaVariant, SourceKind, TranscriptAdapter, TranscriptHandle, Usage,
+};
+use std::path::PathBuf;
+
+const SOURCE: SourceKind = SourceKind::ClaudeCode;
+
+/// Adapter for Anthropic Claude Code transcripts.
+#[derive(Debug, Default, Clone, Copy)]
+pub struct ClaudeCodeAdapter;
+
+impl TranscriptAdapter for ClaudeCodeAdapter {
+    fn source_kind(&self) -> SourceKind {
+        SOURCE
+    }
+
+    fn discover(&self, cfg: &DiscoverCfg) -> Vec<TranscriptHandle> {
+        discover_transcripts(cfg)
+    }
+
+    fn parse(&self, raw: &RawRecord, ctx: &mut ParseCtx) -> Result<Vec<CaptureEvent>, ParseError> {
+        // A blank line or non-JSON line carries nothing; stay lossless via the
+        // shared stub (which routes non-JSON to Unknown and skips blanks).
+        let Some(value) = util::parse_json_line(raw) else {
+            return util::stub_parse(SOURCE, raw, ctx);
+        };
+        Ok(parse_record(raw, ctx, &value))
+    }
+
+    fn schema_fingerprint(&self, sample: &RawRecord) -> SchemaVariant {
+        fingerprint(sample)
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Discovery
+// ---------------------------------------------------------------------------
+
+/// Discover `<config>/projects/<slug>/<session>.jsonl` transcripts. The config
+/// dir is `CLAUDE_CONFIG_DIR` (override) else `<home>/.claude`.
+fn discover_transcripts(cfg: &DiscoverCfg) -> Vec<TranscriptHandle> {
+    let base = cfg
+        .overrides
+        .get("CLAUDE_CONFIG_DIR")
+        .cloned()
+        .unwrap_or_else(|| cfg.home_dir().join(".claude"));
+    let projects = base.join("projects");
+
+    let mut out = Vec::new();
+    for entry in walkdir::WalkDir::new(&projects)
+        .into_iter()
+        .filter_map(Result::ok)
+    {
+        let path = entry.path();
+        if !path.is_file() {
+            continue;
+        }
+        if path.extension().and_then(|e| e.to_str()) != Some("jsonl") {
+            continue;
+        }
+        let session_hint = path
+            .file_stem()
+            .and_then(|s| s.to_str())
+            .map(str::to_string);
+        out.push(TranscriptHandle {
+            path: path.to_path_buf(),
+            source: SOURCE,
+            session_hint,
+            compressed: false,
+        });
+    }
+    // Deterministic order regardless of filesystem walk order.
+    out.sort_by(|a, b| a.path.cmp(&b.path));
+    out
+}
+
+// ---------------------------------------------------------------------------
+// Fingerprinting
+// ---------------------------------------------------------------------------
+
+/// Fingerprint a sample record. Claude Code records are JSON objects carrying a
+/// `type` plus `uuid`/`message`/`parentUuid` shape and (modern) a top-level
+/// `version`.
+fn fingerprint(sample: &RawRecord) -> SchemaVariant {
+    let Some(value) = util::parse_json_line(sample) else {
+        return SchemaVariant::unknown(SOURCE);
+    };
+    let has_type = value.get("type").and_then(|v| v.as_str()).is_some();
+    let looks_claude = value.get("uuid").is_some()
+        || value.get("parentUuid").is_some()
+        || value.get("sessionId").is_some()
+        || value.get("message").is_some();
+    if !has_type || !looks_claude {
+        return SchemaVariant::unknown(SOURCE);
+    }
+    // The `2.x` line stamps a top-level `version`; older lines do not.
+    let variant = match value.get("version").and_then(|v| v.as_str()) {
+        Some(v) if v.starts_with("2.") => "claude_code/2.0",
+        Some(_) => "claude_code/1.x",
+        None => "claude_code/unknown",
+    };
+    SchemaVariant::certain(SOURCE, variant)
+}
+
+// ---------------------------------------------------------------------------
+// Parsing
+// ---------------------------------------------------------------------------
+
+/// Parse one record into zero or more normalized events.
+fn parse_record(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    value: &serde_json::Value,
+) -> Vec<CaptureEvent> {
+    // Learn the session id the first time we see it (subagents keep their own
+    // distinct sessionId — we never merge sidechains into the parent session).
+    if ctx.session_id.is_none() {
+        if let Some(sid) = value.get("sessionId").and_then(|v| v.as_str()) {
+            ctx.session_id = Some(sid.to_string());
+        }
+    }
+
+    // The record's native id and DAG parent.
+    let uuid = value
+        .get("uuid")
+        .and_then(|v| v.as_str())
+        .map(str::to_string)
+        .unwrap_or_else(|| content_id(&raw.bytes));
+    let parent_uuid = value
+        .get("parentUuid")
+        .and_then(|v| v.as_str())
+        .map(str::to_string);
+
+    // Idempotency: a repeated record (same uuid) yields nothing on the replay.
+    if !ctx.first_seen(&uuid) {
+        return Vec::new();
+    }
+
+    let ts = util::ts_from(value, &["timestamp", "time", "ts"]);
+    let rec_type = value.get("type").and_then(|v| v.as_str()).unwrap_or("");
+
+    // The very first record opens the session: capture cwd / git / model /
+    // version as a SessionStart, then continue parsing the same record's body.
+    let mut events: Vec<CaptureEvent> = Vec::new();
+    let is_session_start = ctx.project.is_none() && session_startable(rec_type);
+    if is_session_start {
+        let project = project_from(value);
+        ctx.project = Some(project.clone());
+        let model = string_field(value, "model").or_else(|| {
+            value
+                .get("message")
+                .and_then(|m| m.get("model"))
+                .and_then(|v| v.as_str())
+                .map(str::to_string)
+        });
+        let tool_version = string_field(value, "version");
+        events.push(util::mk_event(
+            SOURCE,
+            ctx,
+            raw,
+            session_start_id(&uuid),
+            parent_uuid.clone(),
+            ts,
+            EventKind::SessionStart {
+                cwd: project.cwd.clone(),
+                git: project.git.clone(),
+                model,
+                tool_version,
+            },
+        ));
+    }
+
+    match rec_type {
+        "user" => parse_turn(raw, ctx, value, &uuid, parent_uuid, ts, false, &mut events),
+        "assistant" => parse_turn(raw, ctx, value, &uuid, parent_uuid, ts, true, &mut events),
+        "summary" | "file-history-snapshot" | "system" => {
+            // Recognized container records we do not normalize into a turn:
+            // keep them lossless as Unknown (a summary explicitly maps to
+            // Unknown per the format spec).
+            events.push(util::unknown_event(SOURCE, ctx, raw, value.clone()));
+        }
+        _ => {
+            events.push(util::unknown_event(SOURCE, ctx, raw, value.clone()));
+        }
+    }
+
+    events
+}
+
+/// Whether a record type can open a session (only the dialogue records carry the
+/// cwd/git/version we bind the project from).
+fn session_startable(rec_type: &str) -> bool {
+    matches!(rec_type, "user" | "assistant" | "system")
+}
+
+/// Parse a `user`/`assistant` record body into a turn plus any embedded
+/// tool_use / tool_result / file-edit events.
+#[allow(clippy::too_many_arguments)]
+fn parse_turn(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    value: &serde_json::Value,
+    uuid: &str,
+    parent_uuid: Option<String>,
+    ts: memscribe_core::Timestamp,
+    is_assistant: bool,
+    events: &mut Vec<CaptureEvent>,
+) {
+    let message = value.get("message");
+    let blocks = message.and_then(|m| m.get("content"));
+
+    let mut text = String::new();
+    let mut thinking = String::new();
+    let mut parts: Vec<Part> = Vec::new();
+    // Tool calls / results discovered inside the content blocks; emitted after
+    // the turn so the turn always precedes its embedded tool events.
+    let mut tool_calls: Vec<(String, String, serde_json::Value)> = Vec::new();
+    let mut tool_results: Vec<(String, bool, serde_json::Value)> = Vec::new();
+
+    match blocks {
+        // content as a plain string.
+        Some(serde_json::Value::String(s)) => {
+            push_text(&mut text, s);
+            parts.push(Part::Text { text: s.clone() });
+        }
+        // content as an array of typed blocks.
+        Some(serde_json::Value::Array(arr)) => {
+            for block in arr {
+                let btype = block.get("type").and_then(|v| v.as_str()).unwrap_or("");
+                match btype {
+                    "text" => {
+                        let t = block.get("text").and_then(|v| v.as_str()).unwrap_or("");
+                        push_text(&mut text, t);
+                        parts.push(Part::Text {
+                            text: t.to_string(),
+                        });
+                    }
+                    "thinking" => {
+                        let t = block.get("thinking").and_then(|v| v.as_str()).unwrap_or("");
+                        push_text(&mut thinking, t);
+                        parts.push(Part::Thinking {
+                            text: t.to_string(),
+                        });
+                    }
+                    "tool_use" => {
+                        let call_id = block
+                            .get("id")
+                            .and_then(|v| v.as_str())
+                            .unwrap_or("")
+                            .to_string();
+                        let name = block
+                            .get("name")
+                            .and_then(|v| v.as_str())
+                            .unwrap_or("")
+                            .to_string();
+                        let args = block
+                            .get("input")
+                            .cloned()
+                            .unwrap_or(serde_json::Value::Null);
+                        parts.push(Part::ToolUse {
+                            call_id: call_id.clone(),
+                            name: name.clone(),
+                            args: args.clone(),
+                        });
+                        tool_calls.push((call_id, name, args));
+                    }
+                    "tool_result" => {
+                        let call_id = block
+                            .get("tool_use_id")
+                            .and_then(|v| v.as_str())
+                            .unwrap_or("")
+                            .to_string();
+                        let is_error = block
+                            .get("is_error")
+                            .and_then(serde_json::Value::as_bool)
+                            .unwrap_or(false);
+                        let output = block
+                            .get("content")
+                            .cloned()
+                            .unwrap_or(serde_json::Value::Null);
+                        parts.push(Part::ToolResult {
+                            call_id: call_id.clone(),
+                            output: output.clone(),
+                        });
+                        tool_results.push((call_id, !is_error, output));
+                    }
+                    "image" => {
+                        let media_type = block
+                            .get("source")
+                            .and_then(|s| s.get("media_type"))
+                            .and_then(|v| v.as_str())
+                            .map(str::to_string);
+                        parts.push(Part::Image { media_type });
+                    }
+                    _ => {
+                        parts.push(Part::Other { raw: block.clone() });
+                    }
+                }
+            }
+        }
+        _ => {}
+    }
+
+    // The turn event itself, carrying the record's native uuid.
+    if is_assistant {
+        let model = message
+            .and_then(|m| m.get("model"))
+            .and_then(|v| v.as_str())
+            .map(str::to_string)
+            .or_else(|| string_field(value, "model"));
+        let usage = message.and_then(|m| m.get("usage")).map(usage_from);
+        events.push(util::mk_event(
+            SOURCE,
+            ctx,
+            raw,
+            uuid.to_string(),
+            parent_uuid.clone(),
+            ts,
+            EventKind::AssistantTurn {
+                text,
+                thinking: if thinking.is_empty() {
+                    None
+                } else {
+                    Some(thinking)
+                },
+                model,
+                usage,
+                parts,
+            },
+        ));
+    } else {
+        events.push(util::mk_event(
+            SOURCE,
+            ctx,
+            raw,
+            uuid.to_string(),
+            parent_uuid.clone(),
+            ts,
+            EventKind::UserTurn { text, parts },
+        ));
+    }
+
+    // Embedded tool calls (assistant) — one ToolCall per tool_use block.
+    for (idx, (call_id, name, args)) in tool_calls.into_iter().enumerate() {
+        events.push(util::mk_event(
+            SOURCE,
+            ctx,
+            raw,
+            derived_id(uuid, "tool_call", idx),
+            Some(uuid.to_string()),
+            ts,
+            EventKind::ToolCall {
+                call_id,
+                name,
+                args,
+            },
+        ));
+    }
+
+    // Embedded tool results (user) — one ToolResult per tool_result block.
+    for (idx, (call_id, ok, output)) in tool_results.into_iter().enumerate() {
+        events.push(util::mk_event(
+            SOURCE,
+            ctx,
+            raw,
+            derived_id(uuid, "tool_result", idx),
+            Some(uuid.to_string()),
+            ts,
+            EventKind::ToolResult {
+                call_id,
+                ok,
+                output,
+            },
+        ));
+    }
+
+    // A top-level `toolUseResult` with a structured patch is a file edit.
+    if let Some(edit) = parse_file_edit(value) {
+        events.push(util::mk_event(
+            SOURCE,
+            ctx,
+            raw,
+            derived_id(uuid, "file_edit", 0),
+            Some(uuid.to_string()),
+            ts,
+            edit,
+        ));
+    }
+}
+
+/// Build a [`EventKind::FileEdit`] from a record's top-level `toolUseResult`,
+/// when it carries a `structuredPatch`. Returns `None` otherwise.
+fn parse_file_edit(value: &serde_json::Value) -> Option<EventKind> {
+    let tur = value.get("toolUseResult")?;
+    let patch = tur.get("structuredPatch")?.as_array()?;
+
+    let path = tur
+        .get("filePath")
+        .and_then(|v| v.as_str())
+        .map(PathBuf::from)
+        .unwrap_or_default();
+    let old = tur
+        .get("oldString")
+        .and_then(|v| v.as_str())
+        .map(str::to_string);
+    let new = tur
+        .get("newString")
+        .and_then(|v| v.as_str())
+        .map(str::to_string);
+
+    let mut unified_lines: Vec<String> = Vec::new();
+    let mut added: u32 = 0;
+    let mut removed: u32 = 0;
+    for hunk in patch {
+        if let Some(lines) = hunk.get("lines").and_then(|v| v.as_array()) {
+            for line in lines {
+                if let Some(s) = line.as_str() {
+                    if let Some(first) = s.chars().next() {
+                        if first == '+' {
+                            added += 1;
+                        } else if first == '-' {
+                            removed += 1;
+                        }
+                    }
+                    unified_lines.push(s.to_string());
+                }
+            }
+        }
+    }
+    let unified = if unified_lines.is_empty() {
+        None
+    } else {
+        Some(unified_lines.join("\n"))
+    };
+
+    // The originating tool call id. An edit's structuredPatch arrives on the
+    // tool_result record, so the call id is the content block's `tool_use_id`
+    // (a `tool_result`) — or, when the patch is colocated with the call, the
+    // `tool_use` block's `id`. Either block type resolves the same call.
+    let call_id = value
+        .get("message")
+        .and_then(|m| m.get("content"))
+        .and_then(|c| c.as_array())
+        .and_then(|arr| {
+            arr.iter()
+                .find_map(|b| match b.get("type").and_then(|v| v.as_str()) {
+                    Some("tool_result") => b
+                        .get("tool_use_id")
+                        .and_then(|v| v.as_str())
+                        .map(str::to_string),
+                    Some("tool_use") => b.get("id").and_then(|v| v.as_str()).map(str::to_string),
+                    _ => None,
+                })
+        });
+
+    Some(EventKind::FileEdit {
+        call_id,
+        diff: Diff {
+            path,
+            old,
+            new,
+            unified,
+            added_lines: added,
+            removed_lines: removed,
+        },
+    })
+}
+
+// ---------------------------------------------------------------------------
+// Small deterministic helpers
+// ---------------------------------------------------------------------------
+
+/// Build the project binding from a session-opening record's cwd / git fields.
+fn project_from(value: &serde_json::Value) -> ProjectRef {
+    let cwd = value
+        .get("cwd")
+        .and_then(|v| v.as_str())
+        .map(PathBuf::from)
+        .unwrap_or_else(|| PathBuf::from("."));
+    let branch = value
+        .get("gitBranch")
+        .and_then(|v| v.as_str())
+        .filter(|s| !s.is_empty())
+        .map(str::to_string);
+    let sha = value
+        .get("gitSha")
+        .and_then(|v| v.as_str())
+        .map(str::to_string)
+        .unwrap_or_default();
+    let git = if branch.is_some() || !sha.is_empty() {
+        Some(GitRef { sha, branch })
+    } else {
+        None
+    };
+    ProjectRef {
+        cwd,
+        repo_root: None,
+        git,
+    }
+}
+
+/// Read a top-level string field.
+fn string_field(value: &serde_json::Value, key: &str) -> Option<String> {
+    value.get(key).and_then(|v| v.as_str()).map(str::to_string)
+}
+
+/// Build a [`Usage`] from a Claude Code `usage` object.
+fn usage_from(u: &serde_json::Value) -> Usage {
+    let g = |k: &str| u.get(k).and_then(serde_json::Value::as_u64);
+    Usage {
+        input_tokens: g("input_tokens"),
+        output_tokens: g("output_tokens"),
+        cache_read_tokens: g("cache_read_input_tokens"),
+        cache_creation_tokens: g("cache_creation_input_tokens"),
+    }
+}
+
+/// Append `s` to a text accumulator, separating multiple text blocks by a single
+/// newline so concatenation stays readable and deterministic.
+fn push_text(acc: &mut String, s: &str) {
+    if s.is_empty() {
+        return;
+    }
+    if !acc.is_empty() {
+        acc.push('\n');
+    }
+    acc.push_str(s);
+}
+
+/// A deterministic event id for the SessionStart synthesized from a record.
+fn session_start_id(uuid: &str) -> String {
+    content_id(format!("claude_code/session_start/{uuid}").as_bytes())
+}
+
+/// A deterministic, collision-free id for a secondary event derived from a
+/// record's `uuid` (the primary turn keeps the bare `uuid`).
+fn derived_id(uuid: &str, kind: &str, idx: usize) -> String {
+    content_id(format!("claude_code/{kind}/{idx}/{uuid}").as_bytes())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use memscribe_core::SourceLocation;
+
+    fn raw(s: &str) -> RawRecord {
+        RawRecord::from_line(s, SourceLocation::new("session.jsonl", 0, 1))
+    }
+
+    fn parse_all(lines: &[&str]) -> Vec<CaptureEvent> {
+        let adapter = ClaudeCodeAdapter;
+        let mut ctx = ParseCtx::new();
+        let mut out = Vec::new();
+        for line in lines {
+            let evs = adapter.parse(&raw(line), &mut ctx).expect("never errors");
+            out.extend(evs);
+        }
+        out
+    }
+
+    fn tags(evs: &[CaptureEvent]) -> Vec<&'static str> {
+        evs.iter().map(|e| e.kind.tag()).collect()
+    }
+
+    // --- TDD: the normalized sequence for a small dialogue --------------------
+
+    #[test]
+    fn first_user_record_yields_session_start_then_user_turn() {
+        let line = r#"{"type":"user","uuid":"u1","parentUuid":null,"timestamp":"2026-06-22T10:00:00Z","sessionId":"s1","cwd":"/repo","gitBranch":"main","version":"2.0.1","message":{"role":"user","content":"Let's use Postgres instead of MySQL."}}"#;
+        let evs = parse_all(&[line]);
+        assert_eq!(tags(&evs), vec!["session_start", "user_turn"]);
+        // Session + project were learned from the first record.
+        assert_eq!(evs[1].session_id, "s1");
+        match &evs[0].kind {
+            EventKind::SessionStart {
+                cwd,
+                git,
+                tool_version,
+                ..
+            } => {
+                assert_eq!(cwd.to_str(), Some("/repo"));
+                assert_eq!(git.as_ref().and_then(|g| g.branch.as_deref()), Some("main"));
+                assert_eq!(tool_version.as_deref(), Some("2.0.1"));
+            }
+            other => panic!("expected session_start, got {other:?}"),
+        }
+        match &evs[1].kind {
+            EventKind::UserTurn { text, .. } => {
+                assert_eq!(text, "Let's use Postgres instead of MySQL.");
+            }
+            other => panic!("expected user_turn, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn assistant_with_tool_use_yields_turn_then_tool_call() {
+        let session = r#"{"type":"user","uuid":"u1","parentUuid":null,"timestamp":"2026-06-22T10:00:00Z","sessionId":"s1","cwd":"/repo","gitBranch":"main","version":"2.0.1","message":{"role":"user","content":"go"}}"#;
+        let asst = r#"{"type":"assistant","uuid":"a1","parentUuid":"u1","timestamp":"2026-06-22T10:00:01Z","sessionId":"s1","message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":10,"output_tokens":5,"cache_read_input_tokens":2,"cache_creation_input_tokens":1},"content":[{"type":"text","text":"Editing now."},{"type":"tool_use","id":"call_1","name":"Edit","input":{"file_path":"/repo/a.rs"}}]}}"#;
+        let evs = parse_all(&[session, asst]);
+        assert_eq!(
+            tags(&evs),
+            vec!["session_start", "user_turn", "assistant_turn", "tool_call"]
+        );
+        match &evs[2].kind {
+            EventKind::AssistantTurn {
+                text, model, usage, ..
+            } => {
+                assert_eq!(text, "Editing now.");
+                assert_eq!(model.as_deref(), Some("claude-opus-4-8"));
+                let u = usage.as_ref().expect("usage present");
+                assert_eq!(u.input_tokens, Some(10));
+                assert_eq!(u.cache_read_tokens, Some(2));
+                assert_eq!(u.cache_creation_tokens, Some(1));
+            }
+            other => panic!("expected assistant_turn, got {other:?}"),
+        }
+        match &evs[3].kind {
+            EventKind::ToolCall { call_id, name, .. } => {
+                assert_eq!(call_id, "call_1");
+                assert_eq!(name, "Edit");
+            }
+            other => panic!("expected tool_call, got {other:?}"),
+        }
+        // The secondary tool_call carries a distinct, deterministic id.
+        assert_ne!(evs[2].event_id, evs[3].event_id);
+        assert_eq!(evs[3].parent_id.as_deref(), Some("a1"));
+    }
+
+    #[test]
+    fn tool_result_record_yields_user_turn_then_tool_result() {
+        let session = r#"{"type":"user","uuid":"u1","parentUuid":null,"timestamp":"2026-06-22T10:00:00Z","sessionId":"s1","cwd":"/repo","version":"2.0.1","message":{"role":"user","content":"go"}}"#;
+        let res = r#"{"type":"user","uuid":"u2","parentUuid":"a1","timestamp":"2026-06-22T10:00:02Z","sessionId":"s1","message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"call_1","content":"ok","is_error":false}]}}"#;
+        let evs = parse_all(&[session, res]);
+        // The session record emits session_start + user_turn; the tool_result
+        // record emits its own (empty) user_turn carrier + the tool_result.
+        assert_eq!(
+            tags(&evs),
+            vec!["session_start", "user_turn", "user_turn", "tool_result"]
+        );
+        let res_ev = evs.iter().find(|e| e.kind.tag() == "tool_result").unwrap();
+        match &res_ev.kind {
+            EventKind::ToolResult { call_id, ok, .. } => {
+                assert_eq!(call_id, "call_1");
+                assert!(*ok);
+            }
+            other => panic!("expected tool_result, got {other:?}"),
+        }
+    }
+
+    // --- TDD: a decision + an edit produces UserTurn then FileEdit ------------
+
+    #[test]
+    fn decision_then_edit_yields_user_turn_then_file_edit() {
+        let decision = r#"{"type":"user","uuid":"u1","parentUuid":null,"timestamp":"2026-06-22T10:00:00Z","sessionId":"s1","cwd":"/repo","gitBranch":"main","version":"2.0.1","message":{"role":"user","content":"Let's use Postgres instead of MySQL."}}"#;
+        let edit = r#"{"type":"assistant","uuid":"a1","parentUuid":"u1","timestamp":"2026-06-22T10:00:01Z","sessionId":"s1","message":{"role":"assistant","model":"claude-opus-4-8","content":[{"type":"tool_use","id":"call_1","name":"Edit","input":{"file_path":"/repo/db.rs"}}]},"toolUseResult":{"filePath":"/repo/db.rs","oldString":"mysql","newString":"postgres","structuredPatch":[{"oldStart":1,"oldLines":1,"newStart":1,"newLines":1,"lines":["-mysql","+postgres"]}]}}"#;
+        let evs = parse_all(&[decision, edit]);
+        let t = tags(&evs);
+        // user_turn appears (the decision), and a file_edit appears.
+        assert!(t.contains(&"user_turn"), "tags: {t:?}");
+        assert!(t.contains(&"file_edit"), "tags: {t:?}");
+        // The user_turn precedes the file_edit.
+        let ut = t.iter().position(|x| *x == "user_turn").unwrap();
+        let fe = t.iter().position(|x| *x == "file_edit").unwrap();
+        assert!(ut < fe, "user_turn must precede file_edit: {t:?}");
+
+        let edit_ev = evs.iter().find(|e| e.kind.tag() == "file_edit").unwrap();
+        match &edit_ev.kind {
+            EventKind::FileEdit { call_id, diff } => {
+                assert_eq!(call_id.as_deref(), Some("call_1"));
+                assert_eq!(diff.path.to_str(), Some("/repo/db.rs"));
+                assert_eq!(diff.added_lines, 1);
+                assert_eq!(diff.removed_lines, 1);
+                assert_eq!(diff.old.as_deref(), Some("mysql"));
+                assert_eq!(diff.new.as_deref(), Some("postgres"));
+                assert_eq!(diff.unified.as_deref(), Some("-mysql\n+postgres"));
+            }
+            other => panic!("expected file_edit, got {other:?}"),
+        }
+    }
+
+    // --- TDD: tool failure → FileEdit references a failing result -------------
+
+    #[test]
+    fn failed_edit_keeps_call_id_so_segmenter_can_drop_it() {
+        // The FileEdit must carry the call_id, and a sibling tool_result with
+        // ok=false must exist, so the segmenter drops the episode.
+        let session = r#"{"type":"user","uuid":"u1","parentUuid":null,"timestamp":"2026-06-22T10:00:00Z","sessionId":"s1","cwd":"/repo","version":"2.0.1","message":{"role":"user","content":"edit"}}"#;
+        let edit = r#"{"type":"assistant","uuid":"a1","parentUuid":"u1","timestamp":"2026-06-22T10:00:01Z","sessionId":"s1","message":{"role":"assistant","content":[{"type":"tool_use","id":"call_x","name":"Edit","input":{}}]},"toolUseResult":{"filePath":"/repo/x.rs","oldString":"a","newString":"b","structuredPatch":[{"oldStart":1,"oldLines":1,"newStart":1,"newLines":1,"lines":["-a","+b"]}]}}"#;
+        let fail = r#"{"type":"user","uuid":"u2","parentUuid":"a1","timestamp":"2026-06-22T10:00:02Z","sessionId":"s1","message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"call_x","content":"boom","is_error":true}]}}"#;
+        let evs = parse_all(&[session, edit, fail]);
+
+        let edit_ev = evs.iter().find(|e| e.kind.tag() == "file_edit").unwrap();
+        let edit_call = match &edit_ev.kind {
+            EventKind::FileEdit { call_id, .. } => call_id.clone(),
+            _ => unreachable!(),
+        };
+        assert_eq!(edit_call.as_deref(), Some("call_x"));
+
+        let res_ev = evs.iter().find(|e| e.kind.tag() == "tool_result").unwrap();
+        match &res_ev.kind {
+            EventKind::ToolResult { call_id, ok, .. } => {
+                assert_eq!(call_id, "call_x");
+                assert!(!ok, "the failing edit's result must be ok=false");
+            }
+            _ => unreachable!(),
+        }
+    }
+
+    // --- TDD: never panics on garbage ----------------------------------------
+
+    #[test]
+    fn garbage_input_never_panics_and_is_lossless() {
+        let adapter = ClaudeCodeAdapter;
+        let mut ctx = ParseCtx::new();
+        // Not JSON at all.
+        let e1 = adapter.parse(&raw("}{ not json"), &mut ctx).unwrap();
+        assert_eq!(e1.len(), 1);
+        assert_eq!(e1[0].kind.tag(), "unknown");
+        // Valid JSON, unrecognized shape.
+        let e2 = adapter
+            .parse(&raw(r#"{"hello":"world"}"#), &mut ctx)
+            .unwrap();
+        assert_eq!(e2.len(), 1);
+        assert_eq!(e2[0].kind.tag(), "unknown");
+        // Blank line yields nothing.
+        let e3 = adapter.parse(&raw("   "), &mut ctx).unwrap();
+        assert!(e3.is_empty());
+        // A record with a non-string/array content does not panic.
+        let e4 = adapter
+            .parse(
+                &raw(r#"{"type":"user","uuid":"z","message":{"content":42}}"#),
+                &mut ctx,
+            )
+            .unwrap();
+        assert!(!e4.is_empty());
+    }
+
+    // --- TDD: dedup / idempotency on a repeated record -----------------------
+
+    #[test]
+    fn repeated_uuid_is_deduped_to_empty() {
+        let adapter = ClaudeCodeAdapter;
+        let mut ctx = ParseCtx::new();
+        let line = r#"{"type":"user","uuid":"dup","parentUuid":null,"timestamp":"2026-06-22T10:00:00Z","sessionId":"s1","cwd":"/repo","version":"2.0.1","message":{"role":"user","content":"hi"}}"#;
+        let first = adapter.parse(&raw(line), &mut ctx).unwrap();
+        assert!(!first.is_empty());
+        let second = adapter.parse(&raw(line), &mut ctx).unwrap();
+        assert!(second.is_empty(), "a repeated uuid must yield nothing");
+    }
+
+    // --- summary → Unknown ----------------------------------------------------
+
+    #[test]
+    fn summary_record_is_unknown() {
+        let line = r#"{"type":"summary","summary":"A recap","leafUuid":"x"}"#;
+        let evs = parse_all(&[line]);
+        assert_eq!(tags(&evs), vec!["unknown"]);
+    }
+
+    // --- ban turn surfaces as a UserTurn (gate runs downstream) ---------------
+
+    #[test]
+    fn ban_turn_is_a_user_turn() {
+        let line = r#"{"type":"user","uuid":"b1","parentUuid":null,"timestamp":"2026-06-22T10:00:00Z","sessionId":"s1","cwd":"/repo","version":"2.0.1","message":{"role":"user","content":"We will never add a dependency on left-pad."}}"#;
+        let evs = parse_all(&[line]);
+        assert_eq!(tags(&evs), vec!["session_start", "user_turn"]);
+        match &evs[1].kind {
+            EventKind::UserTurn { text, .. } => {
+                assert!(text.contains("never add a dependency"));
+            }
+            other => panic!("expected user_turn, got {other:?}"),
+        }
+    }
+
+    // --- determinism ----------------------------------------------------------
+
+    #[test]
+    fn parsing_is_deterministic() {
+        let line = r#"{"type":"user","uuid":"u1","parentUuid":null,"timestamp":"2026-06-22T10:00:00Z","sessionId":"s1","cwd":"/repo","gitBranch":"main","version":"2.0.1","message":{"role":"user","content":"Let's use Postgres."}}"#;
+        let a = parse_all(&[line]);
+        let b = parse_all(&[line]);
+        assert_eq!(a, b);
+    }
+
+    // --- fingerprint ----------------------------------------------------------
+
+    #[test]
+    fn fingerprint_recognizes_claude_2x() {
+        let adapter = ClaudeCodeAdapter;
+        let v = adapter.schema_fingerprint(&raw(
+            r#"{"type":"user","uuid":"u1","sessionId":"s1","version":"2.0.1","message":{"content":"hi"}}"#,
+        ));
+        assert_eq!(v.source, SourceKind::ClaudeCode);
+        assert_eq!(v.variant, "claude_code/2.0");
+        assert_eq!(v.confidence, 100);
+    }
+
+    #[test]
+    fn fingerprint_unknown_for_foreign_json() {
+        let adapter = ClaudeCodeAdapter;
+        let v = adapter.schema_fingerprint(&raw(r#"{"foo":"bar"}"#));
+        assert_eq!(v.confidence, 0);
+    }
+
+    // An edit's structuredPatch arrives on the tool_result record, so the
+    // FileEdit must resolve its call_id from the `tool_result` block's
+    // `tool_use_id` — this is what lets the segmenter drop a failed edit.
+    #[test]
+    fn file_edit_on_result_record_resolves_call_id_from_tool_result_block() {
+        let session = r#"{"type":"user","uuid":"u1","parentUuid":null,"timestamp":"2026-06-22T10:00:00Z","sessionId":"s1","cwd":"/repo","version":"2.0.1","message":{"role":"user","content":"go"}}"#;
+        // The structuredPatch is colocated with the tool_result block (no
+        // tool_use block on this record), exactly as Claude Code writes it.
+        let result = r#"{"type":"user","uuid":"u2","parentUuid":"a1","timestamp":"2026-06-22T10:00:02Z","sessionId":"s1","message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"call_z","content":"ok","is_error":false}]},"toolUseResult":{"filePath":"/repo/z.rs","oldString":"a","newString":"b","structuredPatch":[{"oldStart":1,"oldLines":1,"newStart":1,"newLines":1,"lines":["-a","+b"]}]}}"#;
+        let evs = parse_all(&[session, result]);
+        let edit = evs.iter().find(|e| e.kind.tag() == "file_edit").unwrap();
+        match &edit.kind {
+            EventKind::FileEdit { call_id, .. } => {
+                assert_eq!(call_id.as_deref(), Some("call_z"));
+            }
+            other => panic!("expected file_edit, got {other:?}"),
+        }
+    }
+}
diff --git a/crates/memscribe-adapters/src/codex.rs b/crates/memscribe-adapters/src/codex.rs
new file mode 100644
index 0000000..ca54f95
--- /dev/null
+++ b/crates/memscribe-adapters/src/codex.rs
@@ -0,0 +1,1054 @@
+//! Codex CLI adapter.
+//!
+//! Transcripts: `~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl[.zst]`,
+//! `{timestamp,type,payload}` JSONL. Dialogue: `response_item.message`
+//! (role user/assistant), `function_call`/`function_call_output` paired by
+//! `call_id`; `session_meta.git`. Edits: `function_call` `name=apply_patch` →
+//! V4A diff in `arguments` → parse to `Diff`. Quirks: handle `.jsonl.zst`; the
+//! protocol enum ≠ wire format (build to wire data); `history.jsonl` ≠ rollouts;
+//! files may be `0644` (secrets).
+//!
+//! The io reader decompresses `.zst` before records reach `parse`, so this
+//! module only ever sees plain JSON lines. Every record maps to zero or more
+//! [`CaptureEvent`]s; any shape we do not recognize is routed to
+//! [`memscribe_core::EventKind::Unknown`] via [`util::unknown_event`] so the
+//! stream stays lossless.
+
+use crate::util;
+use memscribe_core::{
+    content_id, CaptureEvent, Diff, DiscoverCfg, EventKind, GitRef, ParseCtx, ParseError, Part,
+    ProjectRef, RawRecord, SchemaVariant, SourceKind, TranscriptAdapter, TranscriptHandle,
+};
+use std::path::PathBuf;
+
+/// Adapter for OpenAI Codex CLI transcripts.
+#[derive(Debug, Default, Clone, Copy)]
+pub struct CodexAdapter;
+
+impl TranscriptAdapter for CodexAdapter {
+    fn source_kind(&self) -> SourceKind {
+        SourceKind::Codex
+    }
+
+    fn discover(&self, cfg: &DiscoverCfg) -> Vec<TranscriptHandle> {
+        discover_rollouts(cfg)
+    }
+
+    fn parse(&self, raw: &RawRecord, ctx: &mut ParseCtx) -> Result<Vec<CaptureEvent>, ParseError> {
+        // Blank lines carry nothing; skip them (the io layer may hand us trailers).
+        let Some(value) = util::parse_json_line(raw) else {
+            return Ok(Vec::new());
+        };
+        Ok(parse_record(raw, ctx, &value))
+    }
+
+    fn schema_fingerprint(&self, sample: &RawRecord) -> SchemaVariant {
+        match util::parse_json_line(sample) {
+            Some(v) if is_codex_record(&v) => {
+                SchemaVariant::certain(SourceKind::Codex, "codex/rollout-v2")
+            }
+            _ => SchemaVariant::unknown(SourceKind::Codex),
+        }
+    }
+}
+
+/// A record is recognizably a Codex rollout line if it carries a top-level
+/// `type` and `payload`. We keep this lenient so version churn still fingerprints.
+fn is_codex_record(value: &serde_json::Value) -> bool {
+    value.get("type").and_then(|t| t.as_str()).is_some() && value.get("payload").is_some()
+}
+
+/// Parse one decoded record into zero or more events. Never panics.
+fn parse_record(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    value: &serde_json::Value,
+) -> Vec<CaptureEvent> {
+    let rec_type = value.get("type").and_then(|t| t.as_str());
+    let payload = value.get("payload");
+
+    match (rec_type, payload) {
+        (Some("session_meta"), Some(p)) => parse_session_meta(raw, ctx, value, p),
+        (Some("response_item"), Some(p)) => parse_response_item(raw, ctx, value, p),
+        // `event_msg` and `turn_context` are valid but carry no normalized
+        // payload of their own — preserve them losslessly as Unknown.
+        _ => vec![util::unknown_event(
+            SourceKind::Codex,
+            ctx,
+            raw,
+            value.clone(),
+        )],
+    }
+}
+
+/// `session_meta` → [`EventKind::SessionStart`]. Sets `ctx.session_id` and
+/// `ctx.project` so later records inherit the binding.
+fn parse_session_meta(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    value: &serde_json::Value,
+    payload: &serde_json::Value,
+) -> Vec<CaptureEvent> {
+    // Learn the session id (used by mk_event for every subsequent event).
+    if let Some(id) = payload.get("id").and_then(|v| v.as_str()) {
+        ctx.session_id = Some(id.to_string());
+    }
+
+    let cwd = payload
+        .get("cwd")
+        .and_then(|v| v.as_str())
+        .map(PathBuf::from)
+        .unwrap_or_else(|| PathBuf::from("."));
+
+    let git = parse_git(payload.get("git"));
+
+    // Populate the project binding from the session-start record.
+    ctx.project = Some(ProjectRef {
+        cwd: cwd.clone(),
+        repo_root: None,
+        git: git.clone(),
+    });
+
+    // Codex labels its version under either `cli_version` or `originator`.
+    let tool_version = payload
+        .get("cli_version")
+        .and_then(|v| v.as_str())
+        .or_else(|| payload.get("originator").and_then(|v| v.as_str()))
+        .map(str::to_string);
+
+    let model = payload
+        .get("model")
+        .and_then(|v| v.as_str())
+        .map(str::to_string);
+
+    let ts = util::ts_from(value, &["timestamp", "time", "ts"]);
+    let event_id = session_event_id(ctx, raw);
+    if !ctx.first_seen(&event_id) {
+        return Vec::new();
+    }
+
+    let kind = EventKind::SessionStart {
+        cwd,
+        git,
+        model,
+        tool_version,
+    };
+    vec![util::mk_event(
+        SourceKind::Codex,
+        ctx,
+        raw,
+        event_id,
+        None,
+        ts,
+        kind,
+    )]
+}
+
+/// Build a stable event id for the session-start record: prefer the session id,
+/// else a content hash, so dedup/idempotency holds for repeated meta lines.
+fn session_event_id(ctx: &ParseCtx, raw: &RawRecord) -> String {
+    if let Some(id) = ctx.session_id.as_deref() {
+        return format!("session_meta:{id}");
+    }
+    content_id(&raw.bytes)
+}
+
+/// Parse a `git` object `{sha, branch}` into a [`GitRef`].
+fn parse_git(git: Option<&serde_json::Value>) -> Option<GitRef> {
+    let g = git?;
+    let sha = g.get("sha").and_then(|v| v.as_str())?.to_string();
+    let branch = g.get("branch").and_then(|v| v.as_str()).map(str::to_string);
+    Some(GitRef { sha, branch })
+}
+
+/// A `response_item` payload has its own `type`. Dispatch on it.
+fn parse_response_item(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    value: &serde_json::Value,
+    payload: &serde_json::Value,
+) -> Vec<CaptureEvent> {
+    let item_type = payload.get("type").and_then(|t| t.as_str());
+    let ts = util::ts_from(value, &["timestamp", "time", "ts"]);
+
+    match item_type {
+        Some("message") => parse_message(raw, ctx, payload, ts),
+        Some("function_call") => parse_function_call(raw, ctx, payload, ts),
+        Some("function_call_output") => parse_function_call_output(raw, ctx, payload, ts),
+        // `reasoning` and any other item type carry no first-class mapping —
+        // keep them lossless as Unknown rather than guessing.
+        _ => vec![util::unknown_event(
+            SourceKind::Codex,
+            ctx,
+            raw,
+            value.clone(),
+        )],
+    }
+}
+
+/// `message` → [`EventKind::UserTurn`] / [`EventKind::AssistantTurn`].
+fn parse_message(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    payload: &serde_json::Value,
+    ts: memscribe_core::Timestamp,
+) -> Vec<CaptureEvent> {
+    let role = payload.get("role").and_then(|v| v.as_str()).unwrap_or("");
+    let (text, parts) = flatten_content(payload.get("content"));
+
+    let event_id = item_event_id(payload, raw);
+    if !ctx.first_seen(&event_id) {
+        return Vec::new();
+    }
+
+    let kind = match role {
+        "user" => EventKind::UserTurn { text, parts },
+        "assistant" => EventKind::AssistantTurn {
+            text,
+            thinking: None,
+            model: payload
+                .get("model")
+                .and_then(|v| v.as_str())
+                .map(str::to_string),
+            usage: None,
+            parts,
+        },
+        // A message with an unexpected role: keep it lossless.
+        _ => {
+            return vec![util::unknown_event(
+                SourceKind::Codex,
+                ctx,
+                raw,
+                payload.clone(),
+            )];
+        }
+    };
+    vec![util::mk_event(
+        SourceKind::Codex,
+        ctx,
+        raw,
+        event_id,
+        None,
+        ts,
+        kind,
+    )]
+}
+
+/// Flatten a `content` array of `{type:input_text|output_text, text}` parts into
+/// a joined text blob and the structured [`Part`] list.
+fn flatten_content(content: Option<&serde_json::Value>) -> (String, Vec<Part>) {
+    let mut text = String::new();
+    let mut parts: Vec<Part> = Vec::new();
+    let Some(items) = content.and_then(|c| c.as_array()) else {
+        return (text, parts);
+    };
+    for item in items {
+        let ptype = item.get("type").and_then(|v| v.as_str()).unwrap_or("");
+        match ptype {
+            "input_text" | "output_text" | "text" => {
+                if let Some(t) = item.get("text").and_then(|v| v.as_str()) {
+                    if !text.is_empty() {
+                        text.push('\n');
+                    }
+                    text.push_str(t);
+                    parts.push(Part::Text {
+                        text: t.to_string(),
+                    });
+                }
+            }
+            "input_image" | "image" | "output_image" => {
+                parts.push(Part::Image {
+                    media_type: item
+                        .get("media_type")
+                        .or_else(|| item.get("image_url"))
+                        .and_then(|v| v.as_str())
+                        .map(str::to_string),
+                });
+            }
+            _ => parts.push(Part::Other { raw: item.clone() }),
+        }
+    }
+    (text, parts)
+}
+
+/// `function_call` → [`EventKind::ToolCall`], plus one
+/// [`EventKind::FileEdit`] per file section when the call is `apply_patch`.
+fn parse_function_call(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    payload: &serde_json::Value,
+    ts: memscribe_core::Timestamp,
+) -> Vec<CaptureEvent> {
+    let name = payload
+        .get("name")
+        .and_then(|v| v.as_str())
+        .unwrap_or("")
+        .to_string();
+    let call_id = payload
+        .get("call_id")
+        .and_then(|v| v.as_str())
+        .map(str::to_string);
+
+    // `arguments` is a JSON *string* on the wire; parse it, falling back to a
+    // string value when it is not valid JSON (still lossless).
+    let args = parse_arguments(payload.get("arguments"));
+
+    let event_id = call_id
+        .clone()
+        .unwrap_or_else(|| item_event_id(payload, raw));
+    if !ctx.first_seen(&event_id) {
+        return Vec::new();
+    }
+
+    // Remember the call name so a later `function_call_output` can pair with it.
+    if let Some(cid) = &call_id {
+        ctx.call_names.insert(cid.clone(), name.clone());
+    }
+
+    let mut events = Vec::new();
+    events.push(util::mk_event(
+        SourceKind::Codex,
+        ctx,
+        raw,
+        event_id.clone(),
+        None,
+        ts,
+        EventKind::ToolCall {
+            call_id: call_id.clone().unwrap_or_default(),
+            name: name.clone(),
+            args: args.clone(),
+        },
+    ));
+
+    // apply_patch also yields one FileEdit per file section in the V4A patch.
+    if name == "apply_patch" {
+        if let Some(patch) = extract_patch_text(&args) {
+            for section in parse_v4a_patch(&patch) {
+                // A unique, deterministic id per FileEdit so dedup does not
+                // collapse multiple edits from one call.
+                let edit_id =
+                    content_id(format!("{}:edit:{}", event_id, section.path.display()).as_bytes());
+                if !ctx.first_seen(&edit_id) {
+                    continue;
+                }
+                events.push(util::mk_event(
+                    SourceKind::Codex,
+                    ctx,
+                    raw,
+                    edit_id,
+                    call_id.clone(),
+                    ts,
+                    EventKind::FileEdit {
+                        call_id: call_id.clone(),
+                        diff: section.into_diff(),
+                    },
+                ));
+            }
+        }
+    }
+
+    events
+}
+
+/// `function_call_output` → [`EventKind::ToolResult`]. `ok` is false when the
+/// output indicates an error.
+fn parse_function_call_output(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    payload: &serde_json::Value,
+    ts: memscribe_core::Timestamp,
+) -> Vec<CaptureEvent> {
+    let call_id = payload
+        .get("call_id")
+        .and_then(|v| v.as_str())
+        .map(str::to_string)
+        .unwrap_or_default();
+
+    let output = payload
+        .get("output")
+        .cloned()
+        .unwrap_or(serde_json::Value::Null);
+    let ok = output_is_ok(&output);
+
+    let event_id = if call_id.is_empty() {
+        item_event_id(payload, raw)
+    } else {
+        format!("{call_id}:output")
+    };
+    if !ctx.first_seen(&event_id) {
+        return Vec::new();
+    }
+
+    // Record the result outcome so downstream pairing (and the segmenter) can
+    // drop edits from a failed call.
+    if !call_id.is_empty() {
+        ctx.call_ok.insert(call_id.clone(), ok);
+    }
+
+    vec![util::mk_event(
+        SourceKind::Codex,
+        ctx,
+        raw,
+        event_id,
+        Some(call_id.clone()),
+        ts,
+        EventKind::ToolResult {
+            call_id,
+            ok,
+            output,
+        },
+    )]
+}
+
+/// Decide whether a tool output indicates success. Codex outputs are sometimes a
+/// bare string, sometimes an object `{output, metadata:{exit_code}}` or carry a
+/// `success` flag. We treat an explicit failure signal as not-ok; otherwise ok.
+fn output_is_ok(output: &serde_json::Value) -> bool {
+    match output {
+        serde_json::Value::String(s) => !string_indicates_error(s),
+        serde_json::Value::Object(map) => {
+            // Explicit booleans win.
+            if let Some(b) = map.get("success").and_then(|v| v.as_bool()) {
+                return b;
+            }
+            if let Some(b) = map.get("ok").and_then(|v| v.as_bool()) {
+                return b;
+            }
+            // A non-zero exit code is a failure.
+            if let Some(code) = map
+                .get("exit_code")
+                .or_else(|| map.get("exitCode"))
+                .and_then(serde_json::Value::as_i64)
+            {
+                return code == 0;
+            }
+            if let Some(code) = map
+                .get("metadata")
+                .and_then(|m| m.get("exit_code").or_else(|| m.get("exitCode")))
+                .and_then(serde_json::Value::as_i64)
+            {
+                return code == 0;
+            }
+            // Otherwise sniff the textual output for an error signature.
+            if let Some(s) = map.get("output").and_then(|v| v.as_str()) {
+                return !string_indicates_error(s);
+            }
+            true
+        }
+        // Null / numbers / arrays: assume ok (no negative signal).
+        _ => true,
+    }
+}
+
+/// Heuristic, deterministic error detection for a free-text tool output.
+fn string_indicates_error(s: &str) -> bool {
+    let lower = s.to_ascii_lowercase();
+    lower.contains("error")
+        || lower.contains("failed")
+        || lower.contains("failure")
+        || lower.contains("traceback")
+        || lower.contains("exception")
+        || lower.contains("not found")
+        || lower.contains("no such file")
+        || lower.contains("patch does not apply")
+        || lower.contains("could not apply")
+}
+
+/// `arguments` arrives as a JSON-encoded string. Parse it; if it is not a JSON
+/// string (or not valid JSON), preserve whatever value was there verbatim.
+fn parse_arguments(arguments: Option<&serde_json::Value>) -> serde_json::Value {
+    match arguments {
+        Some(serde_json::Value::String(s)) => {
+            serde_json::from_str(s).unwrap_or_else(|_| serde_json::Value::String(s.clone()))
+        }
+        Some(other) => other.clone(),
+        None => serde_json::Value::Null,
+    }
+}
+
+/// Pull the V4A patch text out of parsed `apply_patch` arguments. Codex stores it
+/// under `input` or `patch`; tolerate a bare string too.
+fn extract_patch_text(args: &serde_json::Value) -> Option<String> {
+    match args {
+        serde_json::Value::Object(map) => map
+            .get("input")
+            .or_else(|| map.get("patch"))
+            .and_then(|v| v.as_str())
+            .map(str::to_string),
+        serde_json::Value::String(s) => Some(s.clone()),
+        _ => None,
+    }
+}
+
+/// One file section parsed out of a V4A patch.
+struct PatchSection {
+    path: PathBuf,
+    body: String,
+    added: u32,
+    removed: u32,
+}
+
+impl PatchSection {
+    fn into_diff(self) -> Diff {
+        Diff {
+            path: self.path,
+            old: None,
+            new: None,
+            unified: Some(self.body),
+            added_lines: self.added,
+            removed_lines: self.removed,
+        }
+    }
+}
+
+/// Parse a V4A patch string (the `*** Begin Patch` / `*** End Patch` envelope)
+/// into one [`PatchSection`] per file. Deterministic, allocation-only, and
+/// panic-free: it indexes nothing and never unwraps.
+fn parse_v4a_patch(patch: &str) -> Vec<PatchSection> {
+    let mut sections: Vec<PatchSection> = Vec::new();
+    let mut current: Option<PatchSection> = None;
+
+    for line in patch.lines() {
+        if let Some(path) = section_header(line) {
+            if let Some(sec) = current.take() {
+                sections.push(sec);
+            }
+            current = Some(PatchSection {
+                path: PathBuf::from(path),
+                body: String::new(),
+                added: 0,
+                removed: 0,
+            });
+            continue;
+        }
+
+        // The envelope markers themselves are not part of any section body.
+        if line.starts_with("*** Begin Patch") || line.starts_with("*** End Patch") {
+            continue;
+        }
+
+        if let Some(sec) = current.as_mut() {
+            // Count added/removed lines. A leading '+'/'-' marks the change; '@@'
+            // and context (leading space) lines are body but not counted.
+            if let Some(first) = line.as_bytes().first() {
+                match first {
+                    b'+' => sec.added += 1,
+                    b'-' => sec.removed += 1,
+                    _ => {}
+                }
+            }
+            if !sec.body.is_empty() {
+                sec.body.push('\n');
+            }
+            sec.body.push_str(line);
+        }
+    }
+
+    if let Some(sec) = current.take() {
+        sections.push(sec);
+    }
+    sections
+}
+
+/// If `line` is a V4A file header (`*** Update/Add/Delete File: <path>`), return
+/// the path. Otherwise `None`.
+fn section_header(line: &str) -> Option<&str> {
+    for prefix in ["*** Update File: ", "*** Add File: ", "*** Delete File: "] {
+        if let Some(rest) = line.strip_prefix(prefix) {
+            return Some(rest.trim());
+        }
+    }
+    None
+}
+
+/// A stable event id for a `response_item` that carries no native id: prefer an
+/// explicit `id`, else a content hash of the raw record bytes.
+fn item_event_id(payload: &serde_json::Value, raw: &RawRecord) -> String {
+    payload
+        .get("id")
+        .and_then(|v| v.as_str())
+        .map(str::to_string)
+        .unwrap_or_else(|| content_id(&raw.bytes))
+}
+
+/// Discover Codex rollout transcripts under `~/.codex/sessions/**/rollout-*`.
+/// `history.jsonl` is the prompt history file, not a rollout, so it is skipped.
+fn discover_rollouts(cfg: &DiscoverCfg) -> Vec<TranscriptHandle> {
+    let root = codex_sessions_root(cfg);
+    if !root.exists() {
+        return Vec::new();
+    }
+
+    let mut handles: Vec<TranscriptHandle> = Vec::new();
+    for entry in walkdir::WalkDir::new(&root)
+        .follow_links(false)
+        .into_iter()
+        .filter_map(Result::ok)
+    {
+        if !entry.file_type().is_file() {
+            continue;
+        }
+        let path = entry.path();
+        let Some(name) = path.file_name().and_then(|n| n.to_str()) else {
+            continue;
+        };
+        if name == "history.jsonl" {
+            continue; // prompt history, not a rollout
+        }
+        if !name.starts_with("rollout-") {
+            continue;
+        }
+        let compressed = name.ends_with(".zst");
+        let is_jsonl = name.ends_with(".jsonl") || name.ends_with(".jsonl.zst");
+        if !is_jsonl {
+            continue;
+        }
+        handles.push(TranscriptHandle {
+            path: path.to_path_buf(),
+            source: SourceKind::Codex,
+            session_hint: session_hint_from_name(name),
+            compressed,
+        });
+    }
+
+    // Deterministic order regardless of filesystem walk order.
+    handles.sort_by(|a, b| a.path.cmp(&b.path));
+    handles
+}
+
+/// The `~/.codex/sessions` root, honoring a `CODEX_HOME` override.
+fn codex_sessions_root(cfg: &DiscoverCfg) -> PathBuf {
+    if let Some(p) = cfg.overrides.get("CODEX_HOME") {
+        return p.join("sessions");
+    }
+    cfg.home_dir().join(".codex").join("sessions")
+}
+
+/// Derive a session-id hint from a `rollout-<...>.jsonl[.zst]` filename, if one
+/// is embedded after the `rollout-` prefix.
+fn session_hint_from_name(name: &str) -> Option<String> {
+    let stem = name
+        .strip_suffix(".jsonl.zst")
+        .or_else(|| name.strip_suffix(".jsonl"))
+        .unwrap_or(name);
+    let rest = stem.strip_prefix("rollout-")?;
+    if rest.is_empty() {
+        None
+    } else {
+        Some(rest.to_string())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use memscribe_core::SourceLocation;
+
+    fn raw(s: &str) -> RawRecord {
+        RawRecord::from_line(s, SourceLocation::new("rollout-test.jsonl", 0, 1))
+    }
+
+    /// Parse a whole multi-line transcript through one shared context, the way
+    /// the pipeline does. Returns the flat event stream.
+    fn parse_all(lines: &[&str]) -> Vec<CaptureEvent> {
+        let adapter = CodexAdapter;
+        let mut ctx = ParseCtx::new();
+        let mut out = Vec::new();
+        for (i, line) in lines.iter().enumerate() {
+            let r = RawRecord::from_line(
+                line,
+                SourceLocation::new("rollout-test.jsonl", i as u64, (i + 1) as u64),
+            );
+            out.extend(adapter.parse(&r, &mut ctx).expect("never errors"));
+        }
+        out
+    }
+
+    fn tags(events: &[CaptureEvent]) -> Vec<&'static str> {
+        events.iter().map(|e| e.kind.tag()).collect()
+    }
+
+    const META: &str = r#"{"timestamp":"2026-06-22T10:00:00Z","type":"session_meta","payload":{"id":"sess-abc","cwd":"/home/u/proj","git":{"sha":"deadbeef","branch":"main"},"cli_version":"0.5.0"}}"#;
+
+    #[test]
+    fn session_meta_maps_to_session_start_with_project() {
+        let events = parse_all(&[META]);
+        assert_eq!(tags(&events), vec!["session_start"]);
+        let ev = &events[0];
+        assert_eq!(ev.session_id, "sess-abc");
+        match &ev.kind {
+            EventKind::SessionStart {
+                cwd,
+                git,
+                tool_version,
+                ..
+            } => {
+                assert_eq!(cwd, &PathBuf::from("/home/u/proj"));
+                let g = git.as_ref().expect("git present");
+                assert_eq!(g.sha, "deadbeef");
+                assert_eq!(g.branch.as_deref(), Some("main"));
+                assert_eq!(tool_version.as_deref(), Some("0.5.0"));
+            }
+            other => panic!("expected SessionStart, got {other:?}"),
+        }
+        // Project binding is stamped from session_meta.
+        assert_eq!(ev.project.cwd, PathBuf::from("/home/u/proj"));
+        assert!(ev.project.git.is_some());
+    }
+
+    #[test]
+    fn user_and_assistant_messages_map_to_turns() {
+        let user = r#"{"timestamp":"2026-06-22T10:00:01Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"Let's use Postgres instead of MySQL."}]}}"#;
+        let asst = r#"{"timestamp":"2026-06-22T10:00:02Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Sounds good."}]}}"#;
+        let events = parse_all(&[META, user, asst]);
+        assert_eq!(
+            tags(&events),
+            vec!["session_start", "user_turn", "assistant_turn"]
+        );
+        match &events[1].kind {
+            EventKind::UserTurn { text, .. } => {
+                assert_eq!(text, "Let's use Postgres instead of MySQL.");
+            }
+            other => panic!("expected UserTurn, got {other:?}"),
+        }
+        // The user-turn inherits the session id learned from session_meta.
+        assert_eq!(events[1].session_id, "sess-abc");
+    }
+
+    #[test]
+    fn function_call_and_output_pair_by_call_id() {
+        let call = r#"{"timestamp":"2026-06-22T10:00:03Z","type":"response_item","payload":{"type":"function_call","name":"shell","arguments":"{\"command\":[\"ls\"]}","call_id":"call-1"}}"#;
+        let out = r#"{"timestamp":"2026-06-22T10:00:04Z","type":"response_item","payload":{"type":"function_call_output","call_id":"call-1","output":"file1\nfile2"}}"#;
+        let events = parse_all(&[META, call, out]);
+        assert_eq!(
+            tags(&events),
+            vec!["session_start", "tool_call", "tool_result"]
+        );
+        match &events[1].kind {
+            EventKind::ToolCall {
+                call_id,
+                name,
+                args,
+            } => {
+                assert_eq!(call_id, "call-1");
+                assert_eq!(name, "shell");
+                // arguments string was parsed into JSON.
+                assert_eq!(args["command"][0], "ls");
+            }
+            other => panic!("expected ToolCall, got {other:?}"),
+        }
+        match &events[2].kind {
+            EventKind::ToolResult { call_id, ok, .. } => {
+                assert_eq!(call_id, "call-1");
+                assert!(*ok, "plain output should be ok");
+            }
+            other => panic!("expected ToolResult, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn apply_patch_emits_tool_call_then_one_file_edit_per_section() {
+        let patch = "*** Begin Patch\n*** Update File: src/db.rs\n@@\n-let url = \"mysql://...\";\n+let url = \"postgres://...\";\n*** Add File: src/pg.rs\n+pub fn connect() {}\n*** End Patch";
+        let args = serde_json::json!({ "input": patch }).to_string();
+        let call = serde_json::json!({
+            "timestamp": "2026-06-22T10:00:05Z",
+            "type": "response_item",
+            "payload": {
+                "type": "function_call",
+                "name": "apply_patch",
+                "arguments": args,
+                "call_id": "call-edit"
+            }
+        })
+        .to_string();
+        let events = parse_all(&[META, &call]);
+        assert_eq!(
+            tags(&events),
+            vec!["session_start", "tool_call", "file_edit", "file_edit"]
+        );
+        // First edit: Update File with one add + one remove.
+        match &events[2].kind {
+            EventKind::FileEdit { call_id, diff } => {
+                assert_eq!(call_id.as_deref(), Some("call-edit"));
+                assert_eq!(diff.path, PathBuf::from("src/db.rs"));
+                assert_eq!(diff.added_lines, 1);
+                assert_eq!(diff.removed_lines, 1);
+                assert!(diff.unified.as_deref().unwrap().contains("postgres"));
+            }
+            other => panic!("expected FileEdit, got {other:?}"),
+        }
+        match &events[3].kind {
+            EventKind::FileEdit { diff, .. } => {
+                assert_eq!(diff.path, PathBuf::from("src/pg.rs"));
+                assert_eq!(diff.added_lines, 1);
+                assert_eq!(diff.removed_lines, 0);
+            }
+            other => panic!("expected FileEdit, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn decision_then_edit_yields_user_turn_then_file_edit() {
+        let user = r#"{"timestamp":"2026-06-22T10:00:01Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"Let's use Postgres instead of MySQL."}]}}"#;
+        let patch = "*** Begin Patch\n*** Update File: src/db.rs\n+let url = \"postgres://...\";\n*** End Patch";
+        let args = serde_json::json!({ "patch": patch }).to_string();
+        let call = serde_json::json!({
+            "type": "response_item",
+            "payload": {"type":"function_call","name":"apply_patch","arguments":args,"call_id":"c1"}
+        })
+        .to_string();
+        let events = parse_all(&[META, user, &call]);
+        let t = tags(&events);
+        // The decision (UserTurn) precedes the FileEdit in stream order.
+        let user_idx = t.iter().position(|x| *x == "user_turn").unwrap();
+        let edit_idx = t.iter().position(|x| *x == "file_edit").unwrap();
+        assert!(
+            user_idx < edit_idx,
+            "user turn must precede file edit: {t:?}"
+        );
+    }
+
+    #[test]
+    fn failed_function_call_output_marks_not_ok() {
+        let out = r#"{"type":"response_item","payload":{"type":"function_call_output","call_id":"c9","output":"error: patch does not apply"}}"#;
+        let events = parse_all(&[META, out]);
+        match &events[1].kind {
+            EventKind::ToolResult { ok, .. } => assert!(!*ok, "error output must be not-ok"),
+            other => panic!("expected ToolResult, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn exit_code_object_output_failure_is_not_ok() {
+        let out = r#"{"type":"response_item","payload":{"type":"function_call_output","call_id":"c8","output":{"output":"done","metadata":{"exit_code":1}}}}"#;
+        let events = parse_all(&[META, out]);
+        match &events[1].kind {
+            EventKind::ToolResult { ok, .. } => assert!(!*ok, "exit_code 1 must be not-ok"),
+            other => panic!("expected ToolResult, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn unrecognized_record_routes_to_unknown_losslessly() {
+        let weird =
+            r#"{"timestamp":"2026-06-22T10:00:09Z","type":"turn_context","payload":{"foo":"bar"}}"#;
+        let events = parse_all(&[weird]);
+        assert_eq!(tags(&events), vec!["unknown"]);
+        match &events[0].kind {
+            EventKind::Unknown { raw_type, raw } => {
+                assert_eq!(raw_type, "turn_context");
+                assert_eq!(raw["payload"]["foo"], "bar");
+            }
+            other => panic!("expected Unknown, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn reasoning_item_routes_to_unknown() {
+        let reasoning =
+            r#"{"type":"response_item","payload":{"type":"reasoning","summary":"thinking..."}}"#;
+        let events = parse_all(&[META, reasoning]);
+        assert_eq!(tags(&events), vec!["session_start", "unknown"]);
+    }
+
+    #[test]
+    fn garbage_input_never_panics() {
+        // Invalid JSON, partial JSON, empty, non-record JSON, bare scalar.
+        let inputs = [
+            "not json at all",
+            "{",
+            "",
+            "   ",
+            "[1,2,3]",
+            "42",
+            r#"{"type":"session_meta"}"#, // missing payload
+            r#"{"payload":{"id":"x"}}"#,  // missing type
+            r#"{"type":"response_item","payload":{}}"#, // item with no type
+            r#"{"type":"response_item","payload":{"type":"function_call","name":"apply_patch","arguments":"not-json"}}"#,
+        ];
+        let adapter = CodexAdapter;
+        let mut ctx = ParseCtx::new();
+        for s in inputs {
+            // Must not panic; result is fine either way.
+            let _ = adapter.parse(&raw(s), &mut ctx);
+        }
+    }
+
+    #[test]
+    fn repeated_record_is_deduped() {
+        // Same session_meta twice → only one SessionStart.
+        let events = parse_all(&[META, META]);
+        assert_eq!(tags(&events), vec!["session_start"]);
+
+        // Same function_call (same call_id) twice → only one ToolCall.
+        let call = r#"{"type":"response_item","payload":{"type":"function_call","name":"shell","arguments":"{}","call_id":"dup-1"}}"#;
+        let events = parse_all(&[META, call, call]);
+        assert_eq!(tags(&events), vec!["session_start", "tool_call"]);
+    }
+
+    #[test]
+    fn malformed_patch_is_panic_free_and_emits_only_tool_call() {
+        // apply_patch whose arguments are not valid JSON: no FileEdit, just the
+        // ToolCall, and the args are preserved as a string.
+        let call = r#"{"type":"response_item","payload":{"type":"function_call","name":"apply_patch","arguments":"*** Begin Patch (truncated","call_id":"cx"}}"#;
+        let events = parse_all(&[META, call]);
+        assert_eq!(tags(&events), vec!["session_start", "tool_call"]);
+        match &events[1].kind {
+            EventKind::ToolCall { args, .. } => {
+                assert!(args.is_string(), "non-JSON arguments preserved as string");
+            }
+            other => panic!("expected ToolCall, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn schema_fingerprint_recognizes_codex_records() {
+        let adapter = CodexAdapter;
+        let fp = adapter.schema_fingerprint(&raw(META));
+        assert_eq!(fp.source, SourceKind::Codex);
+        assert_eq!(fp.confidence, 100);
+        assert_eq!(fp.variant, "codex/rollout-v2");
+
+        let fp2 = adapter.schema_fingerprint(&raw("not a codex record"));
+        assert_eq!(fp2.confidence, 0);
+    }
+
+    /// Load and parse a fixture from `fixtures/codex/v2/<name>` (workspace root).
+    fn parse_fixture(name: &str) -> Vec<CaptureEvent> {
+        // CARGO_MANIFEST_DIR is .../crates/memscribe-adapters; fixtures live two
+        // levels up at the workspace root.
+        let path = PathBuf::from(env!("CARGO_MANIFEST_DIR"))
+            .join("..")
+            .join("..")
+            .join("fixtures")
+            .join("codex")
+            .join("v2")
+            .join(name);
+        let body = std::fs::read_to_string(&path)
+            .unwrap_or_else(|e| panic!("read fixture {}: {e}", path.display()));
+        let adapter = CodexAdapter;
+        let mut ctx = ParseCtx::new();
+        let mut out = Vec::new();
+        for (i, line) in body.lines().enumerate() {
+            let r = RawRecord::from_line(
+                line,
+                SourceLocation::new(path.clone(), i as u64, (i + 1) as u64),
+            );
+            out.extend(adapter.parse(&r, &mut ctx).expect("never errors"));
+        }
+        out
+    }
+
+    #[test]
+    fn fixture_happy_path_decision_then_edits() {
+        let events = parse_fixture("happy_path_decision_then_edits.jsonl");
+        let t = tags(&events);
+        // session_start, user decision, assistant, tool_call(apply_patch),
+        // two file_edits, tool_result, assistant. (reasoning/event_msg → unknown)
+        assert_eq!(t.iter().filter(|x| **x == "session_start").count(), 1);
+        assert_eq!(t.iter().filter(|x| **x == "user_turn").count(), 1);
+        assert_eq!(t.iter().filter(|x| **x == "file_edit").count(), 2, "{t:?}");
+        // The decision precedes both edits.
+        let user_idx = t.iter().position(|x| *x == "user_turn").unwrap();
+        let first_edit = t.iter().position(|x| *x == "file_edit").unwrap();
+        assert!(user_idx < first_edit, "decision must precede edits: {t:?}");
+        // The successful edit's tool result is ok.
+        let res = events
+            .iter()
+            .find_map(|e| match &e.kind {
+                EventKind::ToolResult { ok, call_id, .. } if call_id == "call_apply_patch_001" => {
+                    Some(*ok)
+                }
+                _ => None,
+            })
+            .expect("tool result present");
+        assert!(res, "successful patch result must be ok");
+        // Edits carry the originating call_id so the segmenter can pair them.
+        for e in &events {
+            if let EventKind::FileEdit { call_id, .. } = &e.kind {
+                assert_eq!(call_id.as_deref(), Some("call_apply_patch_001"));
+            }
+        }
+    }
+
+    #[test]
+    fn fixture_rejected_alternative_has_decision_no_edits() {
+        let events = parse_fixture("rejected_alternative.jsonl");
+        let t = tags(&events);
+        assert_eq!(t.iter().filter(|x| **x == "user_turn").count(), 1);
+        assert_eq!(t.iter().filter(|x| **x == "file_edit").count(), 0, "{t:?}");
+    }
+
+    #[test]
+    fn fixture_ban_has_decision_no_edits() {
+        let events = parse_fixture("ban.jsonl");
+        let t = tags(&events);
+        assert_eq!(t.iter().filter(|x| **x == "user_turn").count(), 1);
+        assert_eq!(t.iter().filter(|x| **x == "file_edit").count(), 0, "{t:?}");
+        // The ban text is preserved verbatim on the user turn.
+        let txt = events.iter().find_map(|e| match &e.kind {
+            EventKind::UserTurn { text, .. } => Some(text.clone()),
+            _ => None,
+        });
+        assert!(txt.unwrap().contains("never add a dependency"));
+    }
+
+    #[test]
+    fn fixture_tool_failure_marks_edit_result_not_ok() {
+        let events = parse_fixture("tool_failure.jsonl");
+        // The edit IS emitted (losslessly) ...
+        let edit = events.iter().find_map(|e| match &e.kind {
+            EventKind::FileEdit { call_id, .. } => call_id.clone(),
+            _ => None,
+        });
+        assert_eq!(edit.as_deref(), Some("call_apply_patch_fail_001"));
+        // ... but its paired tool result is NOT ok, so the segmenter will drop
+        // the Episode (verified there; here we lock the not-ok signal).
+        let ok = events
+            .iter()
+            .find_map(|e| match &e.kind {
+                EventKind::ToolResult { ok, call_id, .. }
+                    if call_id == "call_apply_patch_fail_001" =>
+                {
+                    Some(*ok)
+                }
+                _ => None,
+            })
+            .expect("tool result present");
+        assert!(!ok, "failed patch result must be not-ok");
+    }
+
+    #[test]
+    fn discover_finds_rollouts_and_skips_history() {
+        // Build a fake $CODEX_HOME tree under a temp dir.
+        let base = std::env::temp_dir().join(format!("codex-disc-{}", std::process::id()));
+        let day = base.join("sessions").join("2026").join("06").join("22");
+        std::fs::create_dir_all(&day).expect("mkdir");
+        std::fs::write(day.join("rollout-2026-06-22T10-00-00-sess.jsonl"), b"{}").unwrap();
+        std::fs::write(day.join("rollout-cold.jsonl.zst"), b"{}").unwrap();
+        std::fs::write(base.join("sessions").join("history.jsonl"), b"{}").unwrap();
+        std::fs::write(day.join("notes.txt"), b"x").unwrap();
+
+        let mut overrides = std::collections::HashMap::new();
+        overrides.insert("CODEX_HOME".to_string(), base.clone());
+        let cfg = DiscoverCfg {
+            overrides,
+            ..Default::default()
+        };
+        let handles = discover_rollouts(&cfg);
+
+        // Two rollouts found; history.jsonl and notes.txt excluded.
+        assert_eq!(handles.len(), 2, "handles: {handles:?}");
+        assert!(handles.iter().all(|h| h.source == SourceKind::Codex));
+        assert!(handles.iter().any(|h| h.compressed));
+        assert!(handles.iter().all(|h| h
+            .path
+            .file_name()
+            .unwrap()
+            .to_str()
+            .unwrap()
+            .starts_with("rollout-")));
+
+        std::fs::remove_dir_all(&base).ok();
+    }
+}
diff --git a/crates/memscribe-adapters/src/copilot.rs b/crates/memscribe-adapters/src/copilot.rs
new file mode 100644
index 0000000..f827e50
--- /dev/null
+++ b/crates/memscribe-adapters/src/copilot.rs
@@ -0,0 +1,797 @@
+//! GitHub Copilot adapter.
+//!
+//! Covers the GitHub Copilot CLI / Copilot chat export shape (distinct from the
+//! VS Code Copilot Chat `workspaceStorage` handled by the `vscode` adapter).
+//! Copilot's live store is an undocumented SQLite/`workspaceStorage` blob, so for
+//! the initial deterministic model this adapter parses an **exported chat
+//! JSON-lines** shape and routes anything unrecognized to
+//! [`memscribe_core::EventKind::Unknown`] (losslessness).
+//!
+//! ## Exported record shape (one JSON object per line)
+//! - A leading control record `{kind:"session_start", cwd, git:{sha,branch},
+//!   toolVersion, model?, sessionId, ts}` binds the session/project.
+//! - `{kind:"session_end", sessionId, reason?, ts}` closes it.
+//! - Message records `{id, parentId, role:"user"|"assistant", ts, sessionId,
+//!   text, model?, usage:{input,output}?, toolCalls:[{id,name,args}]?,
+//!   toolResults:[{id,ok,output}]?, edits:[{path,oldText,newText,diff,added,
+//!   removed}]?}`.
+//!
+//! ## Mapping
+//! - `session_start` → [`EventKind::SessionStart`]; `session_end` →
+//!   [`EventKind::SessionEnd`].
+//! - `role:"user"` → [`EventKind::UserTurn`]; `role:"assistant"` →
+//!   [`EventKind::AssistantTurn`] (`text`, `model`, `usage`, `parts`).
+//! - `toolCalls[]` → [`EventKind::ToolCall`]; `toolResults[]` →
+//!   [`EventKind::ToolResult`] (`ok`); `edits[]` → [`EventKind::FileEdit`]
+//!   (`oldText`→`old`, `newText`→`new`, `diff`→`unified`, `added`/`removed`).
+//!
+//! `discover()` points at the real product paths (the binary store is not parsed
+//! in this model). The parser is deterministic, never panics, and dedups repeated
+//! records by their native id via [`ParseCtx::first_seen`].
+
+use crate::util;
+use memscribe_core::{
+    CaptureEvent, Diff, DiscoverCfg, EventKind, GitRef, ParseCtx, ParseError, ProjectRef,
+    RawRecord, SchemaVariant, SourceKind, TranscriptAdapter, TranscriptHandle, Usage,
+};
+use serde_json::Value;
+use std::path::PathBuf;
+
+/// Adapter for GitHub Copilot transcripts.
+#[derive(Debug, Default, Clone, Copy)]
+pub struct CopilotAdapter;
+
+impl TranscriptAdapter for CopilotAdapter {
+    fn source_kind(&self) -> SourceKind {
+        SourceKind::Copilot
+    }
+
+    fn discover(&self, cfg: &DiscoverCfg) -> Vec<TranscriptHandle> {
+        discover_handles(cfg)
+    }
+
+    fn parse(&self, raw: &RawRecord, ctx: &mut ParseCtx) -> Result<Vec<CaptureEvent>, ParseError> {
+        // Parse the line; blank lines yield nothing, invalid JSON is preserved as
+        // an Unknown (lossless) rather than failing the stream.
+        let Some(value) = util::parse_json_line(raw) else {
+            // Distinguish a blank line (skip) from invalid-but-present JSON.
+            if raw.as_str().map(str::trim).unwrap_or("").is_empty() {
+                return Ok(Vec::new());
+            }
+            return util::stub_parse(SourceKind::Copilot, raw, ctx);
+        };
+
+        Ok(parse_value(raw, ctx, value))
+    }
+
+    fn schema_fingerprint(&self, sample: &RawRecord) -> SchemaVariant {
+        fingerprint(sample)
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Discovery
+// ---------------------------------------------------------------------------
+
+/// The Copilot product paths we point discovery at. We do not parse the binary
+/// store in this model, but we surface its location so the runtime/UX can show
+/// where Copilot history lives.
+fn discover_handles(cfg: &DiscoverCfg) -> Vec<TranscriptHandle> {
+    let home = cfg.home_dir();
+    let mut handles = Vec::new();
+
+    // GitHub Copilot CLI config directory.
+    let cli_dir = home.join(".config").join("github-copilot");
+    // VS Code Copilot Chat workspace storage (handled in detail by the vscode
+    // adapter, but Copilot history physically lives here too).
+    let vscode_dir = home
+        .join(".config")
+        .join("Code")
+        .join("User")
+        .join("workspaceStorage");
+
+    for dir in [cli_dir, vscode_dir] {
+        let session_hint = dir.file_name().and_then(|s| s.to_str()).map(str::to_string);
+        handles.push(TranscriptHandle {
+            path: dir,
+            source: SourceKind::Copilot,
+            session_hint,
+            compressed: false,
+        });
+    }
+
+    handles
+}
+
+// ---------------------------------------------------------------------------
+// Fingerprinting
+// ---------------------------------------------------------------------------
+
+fn fingerprint(sample: &RawRecord) -> SchemaVariant {
+    let Some(value) = util::parse_json_line(sample) else {
+        return SchemaVariant::unknown(SourceKind::Copilot);
+    };
+
+    // A control record names the variant with high confidence.
+    if value.get("kind").and_then(Value::as_str) == Some("session_start") {
+        return SchemaVariant::certain(SourceKind::Copilot, "copilot/export-v1");
+    }
+
+    // A message record (role + id) is a reasonable but not definitive signal.
+    let has_role = value.get("role").and_then(Value::as_str).is_some();
+    let has_id = value.get("id").and_then(Value::as_str).is_some();
+    if has_role && has_id {
+        return SchemaVariant {
+            source: SourceKind::Copilot,
+            variant: "copilot/export-v1".to_string(),
+            confidence: 80,
+        };
+    }
+
+    SchemaVariant::unknown(SourceKind::Copilot)
+}
+
+// ---------------------------------------------------------------------------
+// Parsing
+// ---------------------------------------------------------------------------
+
+/// Parse one already-decoded JSON record into zero or more events.
+fn parse_value(raw: &RawRecord, ctx: &mut ParseCtx, value: Value) -> Vec<CaptureEvent> {
+    match value.get("kind").and_then(Value::as_str) {
+        Some("session_start") => parse_session_start(raw, ctx, &value),
+        Some("session_end") => parse_session_end(raw, ctx, &value),
+        // A control record we do not recognize → Unknown (lossless).
+        Some(_) => vec![util::unknown_event(SourceKind::Copilot, ctx, raw, value)],
+        None => match value.get("role").and_then(Value::as_str) {
+            Some("user") | Some("assistant") => parse_message(raw, ctx, value),
+            // Not a control record and not a known message → Unknown.
+            _ => vec![util::unknown_event(SourceKind::Copilot, ctx, raw, value)],
+        },
+    }
+}
+
+fn parse_session_start(raw: &RawRecord, ctx: &mut ParseCtx, value: &Value) -> Vec<CaptureEvent> {
+    // Bind the session id as soon as we learn it.
+    if ctx.session_id.is_none() {
+        if let Some(sid) = value.get("sessionId").and_then(Value::as_str) {
+            ctx.session_id = Some(sid.to_string());
+        }
+    }
+
+    let cwd = value
+        .get("cwd")
+        .and_then(Value::as_str)
+        .map(PathBuf::from)
+        .unwrap_or_else(|| PathBuf::from("."));
+    let git = parse_git(value.get("git"));
+    let model = str_field(value, "model");
+    let tool_version = str_field(value, "toolVersion");
+
+    // Populate the project binding from the session-start record.
+    if ctx.project.is_none() {
+        ctx.project = Some(ProjectRef {
+            cwd: cwd.clone(),
+            repo_root: None,
+            git: git.clone(),
+        });
+    }
+
+    let event_id = event_id_for(value, &raw.bytes);
+    if !ctx.first_seen(&event_id) {
+        return Vec::new();
+    }
+    let ts = ts_for(value);
+
+    vec![util::mk_event(
+        SourceKind::Copilot,
+        ctx,
+        raw,
+        event_id,
+        None,
+        ts,
+        EventKind::SessionStart {
+            cwd,
+            git,
+            model,
+            tool_version,
+        },
+    )]
+}
+
+fn parse_session_end(raw: &RawRecord, ctx: &mut ParseCtx, value: &Value) -> Vec<CaptureEvent> {
+    if ctx.session_id.is_none() {
+        if let Some(sid) = value.get("sessionId").and_then(Value::as_str) {
+            ctx.session_id = Some(sid.to_string());
+        }
+    }
+    let event_id = event_id_for(value, &raw.bytes);
+    if !ctx.first_seen(&event_id) {
+        return Vec::new();
+    }
+    let ts = ts_for(value);
+    let reason = str_field(value, "reason");
+
+    vec![util::mk_event(
+        SourceKind::Copilot,
+        ctx,
+        raw,
+        event_id,
+        None,
+        ts,
+        EventKind::SessionEnd { reason },
+    )]
+}
+
+/// Parse a message record into a turn event, plus any embedded tool calls, tool
+/// results, and file edits (each a distinct event with a derived id).
+fn parse_message(raw: &RawRecord, ctx: &mut ParseCtx, value: Value) -> Vec<CaptureEvent> {
+    if ctx.session_id.is_none() {
+        if let Some(sid) = value.get("sessionId").and_then(Value::as_str) {
+            ctx.session_id = Some(sid.to_string());
+        }
+    }
+
+    let msg_id = event_id_for(&value, &raw.bytes);
+    // Idempotency: a repeated message record (same id) is dropped wholesale,
+    // including all of its derived sub-events.
+    if !ctx.first_seen(&msg_id) {
+        return Vec::new();
+    }
+
+    let parent_id = str_field(&value, "parentId");
+    let ts = ts_for(&value);
+    let text = str_field(&value, "text").unwrap_or_default();
+    let role = value.get("role").and_then(Value::as_str).unwrap_or("");
+
+    let mut events = Vec::new();
+
+    // 1. The turn itself.
+    let kind = if role == "assistant" {
+        let model = str_field(&value, "model");
+        let usage = parse_usage(value.get("usage"));
+        EventKind::AssistantTurn {
+            text,
+            thinking: str_field(&value, "thinking"),
+            model,
+            usage,
+            parts: Vec::new(),
+        }
+    } else {
+        EventKind::UserTurn {
+            text,
+            parts: Vec::new(),
+        }
+    };
+    events.push(util::mk_event(
+        SourceKind::Copilot,
+        ctx,
+        raw,
+        msg_id.clone(),
+        parent_id.clone(),
+        ts,
+        kind,
+    ));
+
+    // 2. Tool calls embedded in the turn.
+    if let Some(calls) = value.get("toolCalls").and_then(Value::as_array) {
+        for (i, call) in calls.iter().enumerate() {
+            if let Some(ev) = tool_call_event(raw, ctx, &msg_id, ts, call, i) {
+                events.push(ev);
+            }
+        }
+    }
+
+    // 3. Tool results embedded in the turn. We record the success flag in the
+    //    context so that any sibling FileEdit can be paired with its outcome.
+    if let Some(results) = value.get("toolResults").and_then(Value::as_array) {
+        for (i, result) in results.iter().enumerate() {
+            if let Some(ev) = tool_result_event(raw, ctx, &msg_id, ts, result, i) {
+                events.push(ev);
+            }
+        }
+    }
+
+    // 4. File edits embedded in the turn.
+    if let Some(edits) = value.get("edits").and_then(Value::as_array) {
+        for (i, edit) in edits.iter().enumerate() {
+            if let Some(ev) = file_edit_event(raw, ctx, &msg_id, ts, edit, i) {
+                events.push(ev);
+            }
+        }
+    }
+
+    events
+}
+
+fn tool_call_event(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    msg_id: &str,
+    ts: memscribe_core::Timestamp,
+    call: &Value,
+    idx: usize,
+) -> Option<CaptureEvent> {
+    let call_id = str_field(call, "id").unwrap_or_else(|| format!("{msg_id}:call:{idx}"));
+    let name = str_field(call, "name").unwrap_or_default();
+    let args = call.get("args").cloned().unwrap_or(Value::Null);
+
+    // Remember the tool name for this call id (call/result pairing).
+    ctx.call_names.insert(call_id.clone(), name.clone());
+
+    let event_id = format!("{msg_id}:call:{call_id}");
+    if !ctx.first_seen(&event_id) {
+        return None;
+    }
+    Some(util::mk_event(
+        SourceKind::Copilot,
+        ctx,
+        raw,
+        event_id,
+        Some(msg_id.to_string()),
+        ts,
+        EventKind::ToolCall {
+            call_id,
+            name,
+            args,
+        },
+    ))
+}
+
+fn tool_result_event(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    msg_id: &str,
+    ts: memscribe_core::Timestamp,
+    result: &Value,
+    idx: usize,
+) -> Option<CaptureEvent> {
+    let call_id = str_field(result, "id").unwrap_or_else(|| format!("{msg_id}:result:{idx}"));
+    // `ok` defaults to true when absent (a present result with no flag is success).
+    let ok = result.get("ok").and_then(Value::as_bool).unwrap_or(true);
+    let output = result.get("output").cloned().unwrap_or(Value::Null);
+
+    // Record the outcome so a sibling FileEdit can be paired with it downstream.
+    ctx.call_ok.insert(call_id.clone(), ok);
+
+    let event_id = format!("{msg_id}:result:{call_id}");
+    if !ctx.first_seen(&event_id) {
+        return None;
+    }
+    Some(util::mk_event(
+        SourceKind::Copilot,
+        ctx,
+        raw,
+        event_id,
+        Some(msg_id.to_string()),
+        ts,
+        EventKind::ToolResult {
+            call_id,
+            ok,
+            output,
+        },
+    ))
+}
+
+fn file_edit_event(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    msg_id: &str,
+    ts: memscribe_core::Timestamp,
+    edit: &Value,
+    idx: usize,
+) -> Option<CaptureEvent> {
+    let path = str_field(edit, "path")?;
+    // An edit may name the originating tool call id so downstream can join its
+    // ToolResult.ok (a failed edit must not become an Episode).
+    let call_id = str_field(edit, "callId").or_else(|| str_field(edit, "call_id"));
+
+    let diff = Diff {
+        path: PathBuf::from(&path),
+        old: str_field(edit, "oldText"),
+        new: str_field(edit, "newText"),
+        unified: str_field(edit, "diff"),
+        added_lines: u32_field(edit, "added"),
+        removed_lines: u32_field(edit, "removed"),
+    };
+
+    let event_id = format!("{msg_id}:edit:{idx}:{path}");
+    if !ctx.first_seen(&event_id) {
+        return None;
+    }
+    Some(util::mk_event(
+        SourceKind::Copilot,
+        ctx,
+        raw,
+        event_id,
+        Some(msg_id.to_string()),
+        ts,
+        EventKind::FileEdit { call_id, diff },
+    ))
+}
+
+// ---------------------------------------------------------------------------
+// Field helpers (all total — never panic on missing/odd input)
+// ---------------------------------------------------------------------------
+
+/// The native event id, else a stable content hash of the raw bytes.
+fn event_id_for(value: &Value, bytes: &[u8]) -> String {
+    str_field(value, "id").unwrap_or_else(|| memscribe_core::content_id(bytes))
+}
+
+/// Pull a string field, treating empty/non-string as absent.
+fn str_field(value: &Value, key: &str) -> Option<String> {
+    value
+        .get(key)
+        .and_then(Value::as_str)
+        .map(str::to_string)
+        .filter(|s| !s.is_empty())
+}
+
+/// Pull a non-negative count as `u32`, clamping out-of-range/odd values to 0.
+fn u32_field(value: &Value, key: &str) -> u32 {
+    value
+        .get(key)
+        .and_then(Value::as_u64)
+        .and_then(|n| u32::try_from(n).ok())
+        .unwrap_or(0)
+}
+
+/// The record timestamp via the shared helper, tolerant of RFC3339 and epoch.
+fn ts_for(value: &Value) -> memscribe_core::Timestamp {
+    util::ts_from(value, &["ts", "timestamp", "time", "created_at"])
+}
+
+fn parse_git(value: Option<&Value>) -> Option<GitRef> {
+    let v = value?;
+    let sha = str_field(v, "sha")?;
+    Some(GitRef {
+        sha,
+        branch: str_field(v, "branch"),
+    })
+}
+
+fn parse_usage(value: Option<&Value>) -> Option<Usage> {
+    let v = value?;
+    let input_tokens = v.get("input").and_then(Value::as_u64);
+    let output_tokens = v.get("output").and_then(Value::as_u64);
+    if input_tokens.is_none() && output_tokens.is_none() {
+        return None;
+    }
+    Some(Usage {
+        input_tokens,
+        output_tokens,
+        cache_read_tokens: v.get("cacheRead").and_then(Value::as_u64),
+        cache_creation_tokens: v.get("cacheCreation").and_then(Value::as_u64),
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use memscribe_core::SourceLocation;
+
+    fn raw(s: &str) -> RawRecord {
+        RawRecord::from_line(s, SourceLocation::new("copilot.jsonl", 0, 1))
+    }
+
+    fn parse_all(lines: &[&str]) -> Vec<CaptureEvent> {
+        let adapter = CopilotAdapter;
+        let mut ctx = ParseCtx::new();
+        let mut out = Vec::new();
+        for line in lines {
+            out.extend(adapter.parse(&raw(line), &mut ctx).expect("never errs"));
+        }
+        out
+    }
+
+    const SESSION_START: &str = r#"{"kind":"session_start","cwd":"/Users/dev/projects/orbit","git":{"sha":"abc123","branch":"main"},"toolVersion":"copilot-cli 1.4.0","model":"gpt-4o","sessionId":"copilot-001","ts":"2026-06-22T10:00:00Z"}"#;
+
+    #[test]
+    fn session_start_binds_session_and_project() {
+        let adapter = CopilotAdapter;
+        let mut ctx = ParseCtx::new();
+        let evs = adapter.parse(&raw(SESSION_START), &mut ctx).unwrap();
+        assert_eq!(evs.len(), 1);
+        assert_eq!(evs[0].kind.tag(), "session_start");
+        assert_eq!(evs[0].session_id, "copilot-001");
+        assert_eq!(ctx.session_id.as_deref(), Some("copilot-001"));
+        let proj = ctx.project.as_ref().expect("project bound");
+        assert_eq!(proj.cwd, PathBuf::from("/Users/dev/projects/orbit"));
+        assert_eq!(proj.git.as_ref().map(|g| g.sha.as_str()), Some("abc123"));
+        match &evs[0].kind {
+            EventKind::SessionStart {
+                model,
+                tool_version,
+                ..
+            } => {
+                assert_eq!(model.as_deref(), Some("gpt-4o"));
+                assert_eq!(tool_version.as_deref(), Some("copilot-cli 1.4.0"));
+            }
+            other => panic!("expected session_start, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn normalized_event_sequence_for_decision_then_edit() {
+        let lines = [
+            SESSION_START,
+            r#"{"id":"m1","parentId":null,"role":"user","sessionId":"copilot-001","ts":"2026-06-22T10:00:05Z","text":"Let's use Postgres instead of MySQL."}"#,
+            r#"{"id":"m2","parentId":"m1","role":"assistant","sessionId":"copilot-001","ts":"2026-06-22T10:00:09Z","text":"Switching to Postgres.","model":"gpt-4o","usage":{"input":1000,"output":200},"edits":[{"path":"src/db.rs","oldText":"mysql","newText":"postgres","diff":"@@ -1 +1 @@\n-mysql\n+postgres","added":1,"removed":1}]}"#,
+            r#"{"kind":"session_end","sessionId":"copilot-001","reason":"done","ts":"2026-06-22T10:01:00Z"}"#,
+        ];
+        let evs = parse_all(&lines);
+        let tags: Vec<&str> = evs.iter().map(|e| e.kind.tag()).collect();
+        assert_eq!(
+            tags,
+            vec![
+                "session_start",
+                "user_turn",
+                "assistant_turn",
+                "file_edit",
+                "session_end"
+            ]
+        );
+    }
+
+    #[test]
+    fn decision_turn_then_file_edit() {
+        let lines = [
+            SESSION_START,
+            r#"{"id":"m1","role":"user","sessionId":"copilot-001","ts":"2026-06-22T10:00:05Z","text":"Use Redis for the cache."}"#,
+            r#"{"id":"m2","parentId":"m1","role":"assistant","sessionId":"copilot-001","ts":"2026-06-22T10:00:09Z","text":"Done.","edits":[{"path":"src/cache.rs","oldText":"a","newText":"b","diff":"d","added":3,"removed":2}]}"#,
+        ];
+        let evs = parse_all(&lines);
+        // user_turn precedes file_edit.
+        let user_idx = evs
+            .iter()
+            .position(|e| e.kind.tag() == "user_turn")
+            .unwrap();
+        let edit_idx = evs
+            .iter()
+            .position(|e| e.kind.tag() == "file_edit")
+            .unwrap();
+        assert!(user_idx < edit_idx);
+        match &evs[edit_idx].kind {
+            EventKind::FileEdit { diff, .. } => {
+                assert_eq!(diff.path, PathBuf::from("src/cache.rs"));
+                assert_eq!(diff.old.as_deref(), Some("a"));
+                assert_eq!(diff.new.as_deref(), Some("b"));
+                assert_eq!(diff.unified.as_deref(), Some("d"));
+                assert_eq!(diff.added_lines, 3);
+                assert_eq!(diff.removed_lines, 2);
+            }
+            other => panic!("expected file_edit, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn tool_call_and_result_with_ok_flag() {
+        let lines = [
+            SESSION_START,
+            r#"{"id":"m2","role":"assistant","sessionId":"copilot-001","ts":"2026-06-22T10:00:09Z","text":"Running.","toolCalls":[{"id":"c1","name":"apply_patch","args":{"path":"x.rs"}}],"toolResults":[{"id":"c1","ok":false,"output":"patch rejected"}]}"#,
+        ];
+        let evs = parse_all(&lines);
+        let tags: Vec<&str> = evs.iter().map(|e| e.kind.tag()).collect();
+        assert_eq!(
+            tags,
+            vec![
+                "session_start",
+                "assistant_turn",
+                "tool_call",
+                "tool_result"
+            ]
+        );
+        match &evs[3].kind {
+            EventKind::ToolResult { call_id, ok, .. } => {
+                assert_eq!(call_id, "c1");
+                assert!(!ok, "failed tool result must carry ok=false");
+            }
+            other => panic!("expected tool_result, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn assistant_usage_and_model_captured() {
+        let lines = [
+            SESSION_START,
+            r#"{"id":"m2","role":"assistant","sessionId":"copilot-001","ts":"2026-06-22T10:00:09Z","text":"Hi.","model":"gpt-4o-mini","usage":{"input":42,"output":7}}"#,
+        ];
+        let evs = parse_all(&lines);
+        match &evs[1].kind {
+            EventKind::AssistantTurn { model, usage, .. } => {
+                assert_eq!(model.as_deref(), Some("gpt-4o-mini"));
+                let u = usage.as_ref().expect("usage present");
+                assert_eq!(u.input_tokens, Some(42));
+                assert_eq!(u.output_tokens, Some(7));
+            }
+            other => panic!("expected assistant_turn, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn no_panic_on_garbage_input() {
+        let adapter = CopilotAdapter;
+        let mut ctx = ParseCtx::new();
+        for junk in [
+            "not json at all",
+            "{",
+            "[]",
+            "12345",
+            "null",
+            "true",
+            r#"{"role":"user"}"#,           // missing id/text
+            r#"{"kind":"weird_control"}"#,  // unknown control
+            r#"{"id":"e","role":"alien"}"#, // unknown role
+            r#"{"id":"x","role":"assistant","text":null,"edits":"not-an-array"}"#,
+            r#"{"kind":"session_start"}"#, // missing all fields
+        ] {
+            let res = adapter.parse(&raw(junk), &mut ctx);
+            assert!(res.is_ok(), "parse must never error on: {junk}");
+        }
+    }
+
+    #[test]
+    fn unrecognized_records_route_to_unknown() {
+        let evs = parse_all(&[
+            r#"{"kind":"telemetry_ping","seq":3}"#,
+            r#"{"id":"z","role":"system","text":"boot"}"#,
+        ]);
+        assert_eq!(evs.len(), 2);
+        assert!(evs.iter().all(|e| e.kind.tag() == "unknown"));
+    }
+
+    #[test]
+    fn blank_lines_skipped() {
+        let evs = parse_all(&["", "   ", "\t"]);
+        assert!(evs.is_empty());
+    }
+
+    #[test]
+    fn dedup_repeated_record_is_idempotent() {
+        let user = r#"{"id":"m1","role":"user","sessionId":"copilot-001","ts":"2026-06-22T10:00:05Z","text":"hello"}"#;
+        let adapter = CopilotAdapter;
+        let mut ctx = ParseCtx::new();
+        let first = adapter.parse(&raw(user), &mut ctx).unwrap();
+        assert_eq!(first.len(), 1);
+        // Re-ingesting the same record (same native id) yields nothing.
+        let second = adapter.parse(&raw(user), &mut ctx).unwrap();
+        assert!(second.is_empty(), "repeated record must dedup to empty");
+    }
+
+    #[test]
+    fn dedup_drops_derived_subevents_too() {
+        let msg = r#"{"id":"m2","role":"assistant","sessionId":"copilot-001","text":"x","edits":[{"path":"a.rs","oldText":"o","newText":"n","added":1,"removed":0}]}"#;
+        let adapter = CopilotAdapter;
+        let mut ctx = ParseCtx::new();
+        let first = adapter.parse(&raw(msg), &mut ctx).unwrap();
+        assert_eq!(first.len(), 2); // assistant_turn + file_edit
+        let second = adapter.parse(&raw(msg), &mut ctx).unwrap();
+        assert!(second.is_empty());
+    }
+
+    #[test]
+    fn seq_is_monotonic_across_subevents() {
+        let evs = parse_all(&[
+            SESSION_START,
+            r#"{"id":"m2","role":"assistant","sessionId":"copilot-001","text":"x","toolCalls":[{"id":"c1","name":"edit","args":{}}],"edits":[{"path":"a.rs","oldText":"o","newText":"n","added":1,"removed":0}]}"#,
+        ]);
+        let seqs: Vec<u64> = evs.iter().map(|e| e.seq).collect();
+        assert_eq!(seqs, vec![0, 1, 2, 3]);
+    }
+
+    #[test]
+    fn ban_turn_is_a_user_turn_with_verbatim_text() {
+        let evs = parse_all(&[
+            SESSION_START,
+            r#"{"id":"m1","role":"user","sessionId":"copilot-001","ts":"2026-06-22T10:00:05Z","text":"We will never add a dependency on left-pad."}"#,
+        ]);
+        match &evs[1].kind {
+            EventKind::UserTurn { text, .. } => {
+                assert_eq!(text, "We will never add a dependency on left-pad.");
+            }
+            other => panic!("expected user_turn, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn discover_points_at_product_paths() {
+        let cfg = DiscoverCfg {
+            home: Some(PathBuf::from("/home/dev")),
+            ..Default::default()
+        };
+        let handles = CopilotAdapter.discover(&cfg);
+        assert!(!handles.is_empty());
+        assert!(handles.iter().all(|h| h.source == SourceKind::Copilot));
+        assert!(handles.iter().any(|h| h.path.ends_with("github-copilot")));
+    }
+
+    #[test]
+    fn fingerprint_recognizes_session_start_and_messages() {
+        let fp = CopilotAdapter.schema_fingerprint(&raw(SESSION_START));
+        assert_eq!(fp.source, SourceKind::Copilot);
+        assert_eq!(fp.confidence, 100);
+        assert_eq!(fp.variant, "copilot/export-v1");
+
+        let msg = raw(r#"{"id":"m1","role":"user","text":"hi"}"#);
+        let fp2 = CopilotAdapter.schema_fingerprint(&msg);
+        assert_eq!(fp2.confidence, 80);
+
+        let junk = raw("not json");
+        let fp3 = CopilotAdapter.schema_fingerprint(&junk);
+        assert_eq!(fp3.confidence, 0);
+    }
+
+    #[test]
+    fn invariants_hold_on_happy_path() {
+        let evs = parse_all(&[
+            SESSION_START,
+            r#"{"id":"m1","role":"user","sessionId":"copilot-001","ts":"2026-06-22T10:00:05Z","text":"Use Postgres."}"#,
+            r#"{"id":"m2","parentId":"m1","role":"assistant","sessionId":"copilot-001","ts":"2026-06-22T10:00:09Z","text":"ok","edits":[{"path":"a.rs","oldText":"o","newText":"n","added":1,"removed":1}]}"#,
+        ]);
+        // Monotonic seq within the session.
+        let mut last = None;
+        for e in &evs {
+            if let Some(p) = last {
+                assert!(e.seq > p);
+            }
+            last = Some(e.seq);
+        }
+        // Unique event ids.
+        let mut seen = std::collections::HashSet::new();
+        for e in &evs {
+            assert!(seen.insert(e.event_id.clone()), "dup id {}", e.event_id);
+        }
+    }
+
+    // --- Fixture-parity guards (mirror fixtures/copilot/v1/*.jsonl verbatim) ---
+
+    #[test]
+    fn fixture_tool_failure_edit_has_failed_result() {
+        // The assistant record from fixtures/copilot/v1/tool_failure.jsonl: the
+        // edit's tool result failed, so the FileEdit must coexist with a
+        // ToolResult{ok:false} — that is the signal downstream uses to suppress
+        // a spurious Episode.
+        let line = r#"{"id":"msg-2","parentId":"msg-1","role":"assistant","ts":"2026-06-22T13:00:13Z","sessionId":"copilot-thread-004","text":"I'll apply the migration patch.","model":"gpt-4o","usage":{"input":720,"output":90},"toolCalls":[{"id":"call-z9","name":"apply_patch","args":{"path":"migrations/0007_email_not_null.sql"}}],"toolResults":[{"id":"call-z9","ok":false,"output":"error: patch did not apply cleanly: hunk #1 FAILED at line 3"}],"edits":[{"path":"migrations/0007_email_not_null.sql","callId":"call-z9","oldText":"email TEXT","newText":"email TEXT NOT NULL","diff":"@@ -3 +3 @@\n-email TEXT\n+email TEXT NOT NULL","added":1,"removed":1}]}"#;
+        let evs = parse_all(&[line]);
+        let tags: Vec<&str> = evs.iter().map(|e| e.kind.tag()).collect();
+        assert_eq!(
+            tags,
+            vec!["assistant_turn", "tool_call", "tool_result", "file_edit"]
+        );
+        // The failed result carries ok=false.
+        let failed = evs
+            .iter()
+            .find(|e| e.kind.tag() == "tool_result")
+            .expect("tool_result present");
+        match &failed.kind {
+            EventKind::ToolResult { ok, call_id, .. } => {
+                assert!(!ok, "tool result must be ok=false");
+                assert_eq!(call_id, "call-z9");
+            }
+            other => panic!("expected tool_result, got {other:?}"),
+        }
+        // The FileEdit links back to the failing call id for downstream pairing.
+        let edit = evs
+            .iter()
+            .find(|e| e.kind.tag() == "file_edit")
+            .expect("file_edit present");
+        match &edit.kind {
+            EventKind::FileEdit { call_id, .. } => {
+                assert_eq!(call_id.as_deref(), Some("call-z9"));
+            }
+            other => panic!("expected file_edit, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn fixture_rejected_alternative_edit_succeeds() {
+        // The assistant record from fixtures/copilot/v1/rejected_alternative.jsonl:
+        // the edit's tool result succeeded (ok=true).
+        let line = r#"{"id":"msg-2","parentId":"msg-1","role":"assistant","ts":"2026-06-22T11:00:14Z","sessionId":"copilot-thread-002","text":"Understood. I'll wire up the Stripe SDK and drop the PayPal client.","model":"gpt-4o","usage":{"input":980,"output":210},"toolCalls":[{"id":"call-a1","name":"apply_patch","args":{"path":"src/payments/provider.rs"}}],"toolResults":[{"id":"call-a1","ok":true,"output":"patch applied (1 file changed)"}],"edits":[{"path":"src/payments/provider.rs","callId":"call-a1","oldText":"use paypal_sdk::Client;","newText":"use stripe::Client;","diff":"@@ -1 +1 @@\n-use paypal_sdk::Client;\n+use stripe::Client;","added":1,"removed":1}]}"#;
+        let evs = parse_all(&[line]);
+        let ok = evs.iter().any(|e| {
+            matches!(&e.kind, EventKind::ToolResult { ok: true, call_id, .. } if call_id == "call-a1")
+        });
+        assert!(ok, "rejected_alternative edit result should be ok=true");
+    }
+}
diff --git a/crates/memscribe-adapters/src/cursor.rs b/crates/memscribe-adapters/src/cursor.rs
new file mode 100644
index 0000000..96df9c8
--- /dev/null
+++ b/crates/memscribe-adapters/src/cursor.rs
@@ -0,0 +1,820 @@
+//! Cursor adapter.
+//!
+//! Cursor is a VS Code fork; live chat state lives in the per-workspace
+//! `state.vscdb` SQLite store under
+//! `~/Library/Application Support/Cursor/User/workspaceStorage/<hash>/` (and,
+//! on newer builds, under `~/.cursor/`). That binary store is undocumented, so
+//! the first deterministic model targets an **exported JSON-lines** transcript
+//! with a stable `{role, text, ...}` shape; a SQLite reader can be layered into
+//! `memscribe-io` later. This parser pattern-matches the fields it needs and
+//! routes anything unrecognized to [`EventKind::Unknown`] so the stream stays
+//! lossless across Cursor-version churn.
+//!
+//! Record shape (one JSON object per line):
+//! - leading `{"kind":"session_start","cwd":..,"git":{"sha","branch"},
+//!   "toolVersion":..,"sessionId":..}` → [`EventKind::SessionStart`] and seeds
+//!   `ctx.project` / `ctx.session_id`.
+//! - message records
+//!   `{"id","parentId","role":"user"|"assistant","ts","sessionId","text",
+//!   "model","usage":{"input","output"},"toolCalls":[..],"toolResults":[..],
+//!   "edits":[..]}`. One record expands to multiple events, in a stable order:
+//!   the turn (`UserTurn` / `AssistantTurn`), then each `ToolCall`, each
+//!   `ToolResult`, then each `FileEdit`.
+//!
+//! `event_id` = the record's native `id`, else a `blake3(content)` fallback.
+
+use crate::util;
+use memscribe_core::{
+    content_id, CaptureEvent, Diff, DiscoverCfg, EventKind, GitRef, ParseCtx, ParseError, Part,
+    ProjectRef, RawRecord, SchemaVariant, SourceKind, TranscriptAdapter, TranscriptHandle, Usage,
+};
+use serde_json::Value;
+use std::path::PathBuf;
+
+const SRC: SourceKind = SourceKind::Cursor;
+
+/// Adapter for Cursor transcripts.
+#[derive(Debug, Default, Clone, Copy)]
+pub struct CursorAdapter;
+
+impl TranscriptAdapter for CursorAdapter {
+    fn source_kind(&self) -> SourceKind {
+        SRC
+    }
+
+    fn discover(&self, cfg: &DiscoverCfg) -> Vec<TranscriptHandle> {
+        let mut out = Vec::new();
+        let home = cfg.home_dir();
+        // Point at the real product locations. We do not parse the binary store
+        // in this model, but discovery should surface where it lives so the
+        // runtime can wire a SQLite reader without re-deriving these paths.
+        let roots = [
+            home.join("Library/Application Support/Cursor/User/workspaceStorage"),
+            home.join(".cursor"),
+        ];
+        for root in roots {
+            for entry in walkdir::WalkDir::new(&root)
+                .max_depth(3)
+                .into_iter()
+                .filter_map(std::result::Result::ok)
+            {
+                let path = entry.path();
+                let name = match path.file_name().and_then(|n| n.to_str()) {
+                    Some(n) => n,
+                    None => continue,
+                };
+                let is_store = name == "state.vscdb"
+                    || name.ends_with(".jsonl")
+                    || name.ends_with(".cursorchat");
+                if is_store && path.is_file() {
+                    let session_hint = path
+                        .parent()
+                        .and_then(|p| p.file_name())
+                        .and_then(|n| n.to_str())
+                        .map(str::to_string);
+                    out.push(TranscriptHandle {
+                        path: path.to_path_buf(),
+                        source: SRC,
+                        session_hint,
+                        compressed: false,
+                    });
+                }
+            }
+        }
+        // Deterministic order regardless of filesystem iteration order.
+        out.sort_by(|a, b| a.path.cmp(&b.path));
+        out
+    }
+
+    fn parse(&self, raw: &RawRecord, ctx: &mut ParseCtx) -> Result<Vec<CaptureEvent>, ParseError> {
+        // Blank lines yield nothing; invalid JSON is preserved verbatim as an
+        // Unknown so the stream is still lossless (never an error here).
+        let value = match util::parse_json_line(raw) {
+            Some(v) => v,
+            None => {
+                let s = raw.as_str().map(str::trim).unwrap_or("");
+                if s.is_empty() {
+                    return Ok(Vec::new());
+                }
+                return Ok(vec![util::unknown_event(
+                    SRC,
+                    ctx,
+                    raw,
+                    Value::String(s.to_string()),
+                )]);
+            }
+        };
+
+        // We only know how to parse JSON objects; anything else is Unknown.
+        let obj = match value.as_object() {
+            Some(o) => o,
+            None => return Ok(vec![util::unknown_event(SRC, ctx, raw, value)]),
+        };
+
+        // Seed session id from any record that carries one (records are parsed
+        // in file order, so the first one wins for the whole stream).
+        if ctx.session_id.is_none() {
+            if let Some(sid) = str_field(obj, "sessionId") {
+                ctx.session_id = Some(sid.to_string());
+            }
+        }
+
+        // Dispatch on the record discriminator. A `kind` of `session_start`
+        // (and a couple of tolerant aliases) means the session header; a `role`
+        // means a dialogue turn. Everything else is Unknown.
+        if let Some(kind) = str_field(obj, "kind") {
+            match kind {
+                "session_start" | "session-start" | "sessionStart" => {
+                    return Ok(parse_session_start(obj, ctx, raw));
+                }
+                "session_end" | "session-end" | "sessionEnd" => {
+                    return Ok(parse_session_end(obj, ctx, raw));
+                }
+                _ => {}
+            }
+        }
+
+        if str_field(obj, "role").is_some() {
+            return Ok(parse_message(obj, ctx, raw));
+        }
+
+        Ok(vec![util::unknown_event(SRC, ctx, raw, value)])
+    }
+
+    fn schema_fingerprint(&self, sample: &RawRecord) -> SchemaVariant {
+        match util::parse_json_line(sample)
+            .as_ref()
+            .and_then(Value::as_object)
+        {
+            Some(obj)
+                if obj.contains_key("role")
+                    || matches!(str_field(obj, "kind"), Some("session_start")) =>
+            {
+                SchemaVariant::certain(SRC, "cursor/export-v1")
+            }
+            _ => SchemaVariant::unknown(SRC),
+        }
+    }
+}
+
+/// Parse a `session_start` header: seeds `ctx.project` and emits `SessionStart`.
+fn parse_session_start(
+    obj: &serde_json::Map<String, Value>,
+    ctx: &mut ParseCtx,
+    raw: &RawRecord,
+) -> Vec<CaptureEvent> {
+    let cwd_str = str_field(obj, "cwd").unwrap_or(".");
+    let cwd = PathBuf::from(cwd_str);
+    let git = parse_git(obj.get("git"));
+    let model = str_field(obj, "model").map(str::to_string);
+    let tool_version = str_field(obj, "toolVersion")
+        .or_else(|| str_field(obj, "tool_version"))
+        .map(str::to_string);
+
+    // Bind the project for every subsequent event in this session.
+    ctx.project = Some(ProjectRef {
+        cwd: cwd.clone(),
+        repo_root: str_field(obj, "repoRoot")
+            .or_else(|| str_field(obj, "repo_root"))
+            .map(PathBuf::from),
+        git: git.clone(),
+    });
+
+    let event_id = event_id_for(obj, raw);
+    if !ctx.first_seen(&event_id) {
+        return Vec::new();
+    }
+    let ts = ts_for(obj);
+    vec![util::mk_event(
+        SRC,
+        ctx,
+        raw,
+        event_id,
+        None,
+        ts,
+        EventKind::SessionStart {
+            cwd,
+            git,
+            model,
+            tool_version,
+        },
+    )]
+}
+
+/// Parse a `session_end` header into a `SessionEnd` event.
+fn parse_session_end(
+    obj: &serde_json::Map<String, Value>,
+    ctx: &mut ParseCtx,
+    raw: &RawRecord,
+) -> Vec<CaptureEvent> {
+    let event_id = event_id_for(obj, raw);
+    if !ctx.first_seen(&event_id) {
+        return Vec::new();
+    }
+    let ts = ts_for(obj);
+    let reason = str_field(obj, "reason").map(str::to_string);
+    vec![util::mk_event(
+        SRC,
+        ctx,
+        raw,
+        event_id,
+        parent_field(obj),
+        ts,
+        EventKind::SessionEnd { reason },
+    )]
+}
+
+/// Parse a dialogue record into the turn event plus any embedded tool calls,
+/// tool results, and file edits — in a stable, deterministic order.
+fn parse_message(
+    obj: &serde_json::Map<String, Value>,
+    ctx: &mut ParseCtx,
+    raw: &RawRecord,
+) -> Vec<CaptureEvent> {
+    let base_id = event_id_for(obj, raw);
+    // Idempotency: a repeated record (same id) yields nothing.
+    if !ctx.first_seen(&base_id) {
+        return Vec::new();
+    }
+
+    let ts = ts_for(obj);
+    let parent = parent_field(obj);
+    let role = str_field(obj, "role").unwrap_or("");
+    let text = str_field(obj, "text").unwrap_or("").to_string();
+
+    let mut events = Vec::new();
+
+    // 1) The turn itself.
+    let turn_kind = match role {
+        "user" => EventKind::UserTurn {
+            text,
+            parts: text_parts(obj),
+        },
+        "assistant" => EventKind::AssistantTurn {
+            text,
+            thinking: str_field(obj, "thinking").map(str::to_string),
+            model: str_field(obj, "model").map(str::to_string),
+            usage: parse_usage(obj.get("usage")),
+            parts: text_parts(obj),
+        },
+        _ => {
+            // A role we don't recognize → Unknown, but still keep ordering.
+            EventKind::Unknown {
+                raw_type: role.to_string(),
+                raw: Value::Object(obj.clone()),
+            }
+        }
+    };
+    events.push(util::mk_event(
+        SRC,
+        ctx,
+        raw,
+        base_id.clone(),
+        parent.clone(),
+        ts,
+        turn_kind,
+    ));
+
+    // 2) Tool calls. Each gets a synthetic, deterministic id derived from the
+    //    turn id + the call id so it never collides with the turn or siblings.
+    if let Some(calls) = obj.get("toolCalls").and_then(Value::as_array) {
+        for (i, call) in calls.iter().enumerate() {
+            let call_obj = match call.as_object() {
+                Some(o) => o,
+                None => continue,
+            };
+            let call_id = str_field(call_obj, "id")
+                .map(str::to_string)
+                .unwrap_or_else(|| format!("{base_id}:call:{i}"));
+            let name = str_field(call_obj, "name").unwrap_or("").to_string();
+            let args = call_obj.get("args").cloned().unwrap_or(Value::Null);
+            // Remember the name so a later result can be paired by call_id.
+            ctx.call_names.insert(call_id.clone(), name.clone());
+            let ev_id = format!("{base_id}#toolcall:{call_id}");
+            if !ctx.first_seen(&ev_id) {
+                continue;
+            }
+            events.push(util::mk_event(
+                SRC,
+                ctx,
+                raw,
+                ev_id,
+                Some(base_id.clone()),
+                ts,
+                EventKind::ToolCall {
+                    call_id,
+                    name,
+                    args,
+                },
+            ));
+        }
+    }
+
+    // 3) Tool results.
+    if let Some(results) = obj.get("toolResults").and_then(Value::as_array) {
+        for (i, res) in results.iter().enumerate() {
+            let res_obj = match res.as_object() {
+                Some(o) => o,
+                None => continue,
+            };
+            let call_id = str_field(res_obj, "id")
+                .map(str::to_string)
+                .unwrap_or_else(|| format!("{base_id}:result:{i}"));
+            let ok = bool_field(res_obj, "ok").unwrap_or(true);
+            ctx.call_ok.insert(call_id.clone(), ok);
+            let output = res_obj.get("output").cloned().unwrap_or(Value::Null);
+            let ev_id = format!("{base_id}#toolresult:{call_id}");
+            if !ctx.first_seen(&ev_id) {
+                continue;
+            }
+            events.push(util::mk_event(
+                SRC,
+                ctx,
+                raw,
+                ev_id,
+                Some(base_id.clone()),
+                ts,
+                EventKind::ToolResult {
+                    call_id,
+                    ok,
+                    output,
+                },
+            ));
+        }
+    }
+
+    // 4) File edits.
+    if let Some(edits) = obj.get("edits").and_then(Value::as_array) {
+        for (i, edit) in edits.iter().enumerate() {
+            let edit_obj = match edit.as_object() {
+                Some(o) => o,
+                None => continue,
+            };
+            let path = str_field(edit_obj, "path").unwrap_or("").to_string();
+            let diff = Diff {
+                path: PathBuf::from(path),
+                old: str_field(edit_obj, "oldText").map(str::to_string),
+                new: str_field(edit_obj, "newText").map(str::to_string),
+                unified: str_field(edit_obj, "diff").map(str::to_string),
+                added_lines: u32_field(edit_obj, "added").unwrap_or(0),
+                removed_lines: u32_field(edit_obj, "removed").unwrap_or(0),
+            };
+            let call_id = str_field(edit_obj, "callId")
+                .or_else(|| str_field(edit_obj, "call_id"))
+                .map(str::to_string);
+            let ev_id = format!("{base_id}#edit:{i}");
+            if !ctx.first_seen(&ev_id) {
+                continue;
+            }
+            events.push(util::mk_event(
+                SRC,
+                ctx,
+                raw,
+                ev_id,
+                Some(base_id.clone()),
+                ts,
+                EventKind::FileEdit { call_id, diff },
+            ));
+        }
+    }
+
+    events
+}
+
+/// Build text/thinking [`Part`]s from a message (best-effort, never fails).
+fn text_parts(obj: &serde_json::Map<String, Value>) -> Vec<Part> {
+    let mut parts = Vec::new();
+    if let Some(t) = str_field(obj, "text") {
+        if !t.is_empty() {
+            parts.push(Part::Text {
+                text: t.to_string(),
+            });
+        }
+    }
+    if let Some(th) = str_field(obj, "thinking") {
+        if !th.is_empty() {
+            parts.push(Part::Thinking {
+                text: th.to_string(),
+            });
+        }
+    }
+    parts
+}
+
+/// Parse `usage:{input,output}` (also tolerant of token-suffixed keys).
+fn parse_usage(value: Option<&Value>) -> Option<Usage> {
+    let obj = value?.as_object()?;
+    let input = u64_field(obj, "input").or_else(|| u64_field(obj, "input_tokens"));
+    let output = u64_field(obj, "output").or_else(|| u64_field(obj, "output_tokens"));
+    if input.is_none() && output.is_none() {
+        return None;
+    }
+    Some(Usage {
+        input_tokens: input,
+        output_tokens: output,
+        cache_read_tokens: None,
+        cache_creation_tokens: None,
+    })
+}
+
+/// Parse a `{sha, branch}` git ref, if present.
+fn parse_git(value: Option<&Value>) -> Option<GitRef> {
+    let obj = value?.as_object()?;
+    let sha = str_field(obj, "sha")?.to_string();
+    Some(GitRef {
+        sha,
+        branch: str_field(obj, "branch").map(str::to_string),
+    })
+}
+
+// ---- small, total field accessors (no panics, no indexing) ----
+
+fn str_field<'a>(obj: &'a serde_json::Map<String, Value>, key: &str) -> Option<&'a str> {
+    obj.get(key).and_then(Value::as_str)
+}
+
+fn bool_field(obj: &serde_json::Map<String, Value>, key: &str) -> Option<bool> {
+    obj.get(key).and_then(Value::as_bool)
+}
+
+fn u64_field(obj: &serde_json::Map<String, Value>, key: &str) -> Option<u64> {
+    obj.get(key).and_then(Value::as_u64)
+}
+
+fn u32_field(obj: &serde_json::Map<String, Value>, key: &str) -> Option<u32> {
+    obj.get(key)
+        .and_then(Value::as_u64)
+        .map(|n| u32::try_from(n).unwrap_or(u32::MAX))
+}
+
+fn parent_field(obj: &serde_json::Map<String, Value>) -> Option<String> {
+    str_field(obj, "parentId")
+        .or_else(|| str_field(obj, "parent_id"))
+        .map(str::to_string)
+}
+
+/// The event id for a record: native `id`, else a stable content hash.
+fn event_id_for(obj: &serde_json::Map<String, Value>, raw: &RawRecord) -> String {
+    str_field(obj, "id")
+        .map(str::to_string)
+        .unwrap_or_else(|| content_id(&raw.bytes))
+}
+
+/// Timestamp from any of the common keys, falling back to the epoch.
+fn ts_for(obj: &serde_json::Map<String, Value>) -> memscribe_core::Timestamp {
+    util::ts_from(
+        &Value::Object(obj.clone()),
+        &["ts", "timestamp", "time", "created_at"],
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use memscribe_core::SourceLocation;
+
+    fn raw(s: &str, line: u64) -> RawRecord {
+        RawRecord::from_line(s, SourceLocation::new("cursor.jsonl", 0, line))
+    }
+
+    /// Parse a whole JSONL string through one shared context (file order),
+    /// returning the flattened event-kind tags.
+    fn parse_all(jsonl: &str) -> (Vec<CaptureEvent>, ParseCtx) {
+        let adapter = CursorAdapter;
+        let mut ctx = ParseCtx::new();
+        let mut events = Vec::new();
+        for (i, line) in jsonl.lines().enumerate() {
+            let r = raw(line, i as u64 + 1);
+            let evs = adapter.parse(&r, &mut ctx).expect("parse never errors");
+            events.extend(evs);
+        }
+        (events, ctx)
+    }
+
+    fn tags(events: &[CaptureEvent]) -> Vec<&'static str> {
+        events.iter().map(|e| e.kind.tag()).collect()
+    }
+
+    #[test]
+    fn session_start_then_decision_then_edit() {
+        let jsonl = r#"{"kind":"session_start","sessionId":"s1","cwd":"/work/app","git":{"sha":"abc123","branch":"main"},"toolVersion":"0.42.0"}
+{"id":"m1","role":"user","ts":"2026-06-22T10:00:00Z","sessionId":"s1","text":"Let's use Postgres instead of MySQL"}
+{"id":"m2","parentId":"m1","role":"assistant","ts":"2026-06-22T10:00:05Z","sessionId":"s1","text":"Switching to Postgres.","model":"cursor-fast","usage":{"input":12,"output":7},"edits":[{"path":"db/config.toml","oldText":"engine=mysql","newText":"engine=postgres","diff":"@@\n-engine=mysql\n+engine=postgres","added":1,"removed":1}]}"#;
+        let (events, ctx) = parse_all(jsonl);
+        assert_eq!(
+            tags(&events),
+            vec!["session_start", "user_turn", "assistant_turn", "file_edit"]
+        );
+        // Session + project were learned from the header.
+        assert_eq!(ctx.session_id.as_deref(), Some("s1"));
+        assert_eq!(events[1].session_id, "s1");
+        assert_eq!(events[1].project.cwd, PathBuf::from("/work/app"));
+        // seq is monotonic from file order.
+        assert_eq!(
+            events.iter().map(|e| e.seq).collect::<Vec<_>>(),
+            vec![0, 1, 2, 3]
+        );
+        // The decision turn is a UserTurn carrying the text verbatim.
+        match &events[1].kind {
+            EventKind::UserTurn { text, .. } => {
+                assert_eq!(text, "Let's use Postgres instead of MySQL");
+            }
+            other => panic!("expected user_turn, got {other:?}"),
+        }
+        // The FileEdit carries old/new/unified and line counts.
+        match &events[3].kind {
+            EventKind::FileEdit { diff, .. } => {
+                assert_eq!(diff.path, PathBuf::from("db/config.toml"));
+                assert_eq!(diff.old.as_deref(), Some("engine=mysql"));
+                assert_eq!(diff.new.as_deref(), Some("engine=postgres"));
+                assert_eq!(diff.added_lines, 1);
+                assert_eq!(diff.removed_lines, 1);
+            }
+            other => panic!("expected file_edit, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn assistant_usage_and_model_captured() {
+        let jsonl = r#"{"id":"a1","role":"assistant","sessionId":"s","text":"hi","model":"cursor-pro","usage":{"input":5,"output":9}}"#;
+        let (events, _) = parse_all(jsonl);
+        match &events[0].kind {
+            EventKind::AssistantTurn {
+                model, usage, text, ..
+            } => {
+                assert_eq!(text, "hi");
+                assert_eq!(model.as_deref(), Some("cursor-pro"));
+                let u = usage.as_ref().expect("usage");
+                assert_eq!(u.input_tokens, Some(5));
+                assert_eq!(u.output_tokens, Some(9));
+            }
+            other => panic!("expected assistant_turn, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn tool_call_then_result_pairing() {
+        let jsonl = r#"{"id":"t1","role":"assistant","sessionId":"s","text":"running","toolCalls":[{"id":"c1","name":"shell","args":{"cmd":"ls"}}],"toolResults":[{"id":"c1","ok":true,"output":"a\nb"}]}"#;
+        let (events, ctx) = parse_all(jsonl);
+        assert_eq!(
+            tags(&events),
+            vec!["assistant_turn", "tool_call", "tool_result"]
+        );
+        assert_eq!(ctx.call_names.get("c1").map(String::as_str), Some("shell"));
+        assert_eq!(ctx.call_ok.get("c1").copied(), Some(true));
+        match &events[2].kind {
+            EventKind::ToolResult { call_id, ok, .. } => {
+                assert_eq!(call_id, "c1");
+                assert!(*ok);
+            }
+            other => panic!("expected tool_result, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn failed_tool_result_is_marked_not_ok() {
+        // Mirrors the tool_failure fixture: edit's result failed.
+        let jsonl = r#"{"id":"f1","role":"assistant","sessionId":"s","text":"trying","toolCalls":[{"id":"e1","name":"edit","args":{"path":"x.rs"}}],"toolResults":[{"id":"e1","ok":false,"output":"permission denied"}],"edits":[{"path":"x.rs","oldText":"a","newText":"b","added":1,"removed":1}]}"#;
+        let (events, ctx) = parse_all(jsonl);
+        assert_eq!(
+            tags(&events),
+            vec!["assistant_turn", "tool_call", "tool_result", "file_edit"]
+        );
+        // The failed result is observable so downstream can suppress the Episode.
+        assert_eq!(ctx.call_ok.get("e1").copied(), Some(false));
+        match &events[2].kind {
+            EventKind::ToolResult { ok, .. } => assert!(!*ok),
+            other => panic!("expected tool_result, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn dedup_repeated_record_is_idempotent() {
+        let line = r#"{"id":"dup","role":"user","sessionId":"s","text":"hello"}"#;
+        let jsonl = format!("{line}\n{line}");
+        let (events, _) = parse_all(&jsonl);
+        // The second identical record yields nothing.
+        assert_eq!(tags(&events), vec!["user_turn"]);
+    }
+
+    #[test]
+    fn unrecognized_record_routes_to_unknown() {
+        let jsonl = r#"{"kind":"telemetry_ping","payload":{"x":1}}
+{"id":"w1","role":"wizard","sessionId":"s","text":"???"}"#;
+        let (events, _) = parse_all(jsonl);
+        // A record with neither a known kind nor a role → Unknown; a record with
+        // an unknown role also degrades to Unknown rather than panicking.
+        assert_eq!(tags(&events), vec!["unknown", "unknown"]);
+    }
+
+    #[test]
+    fn garbage_never_panics() {
+        let adapter = CursorAdapter;
+        let mut ctx = ParseCtx::new();
+        for bad in [
+            "",
+            "   ",
+            "not json at all",
+            "{",
+            "[1,2,3]",
+            "42",
+            "true",
+            "null",
+            r#"{"role":42}"#,
+            r#"{"kind":"session_start","git":"oops","cwd":12}"#,
+            r#"{"id":"x","role":"user","toolCalls":"not-an-array","edits":{"nope":1}}"#,
+        ] {
+            let r = raw(bad, 1);
+            let evs = adapter.parse(&r, &mut ctx).expect("never errors");
+            // Blank lines produce nothing; everything else is lossless (>=1).
+            if bad.trim().is_empty() {
+                assert!(evs.is_empty());
+            } else {
+                assert!(!evs.is_empty(), "lossless for {bad:?}");
+            }
+        }
+    }
+
+    #[test]
+    fn invalid_json_is_preserved_as_unknown() {
+        let (events, _) = parse_all("this is not json");
+        assert_eq!(tags(&events), vec!["unknown"]);
+        match &events[0].kind {
+            EventKind::Unknown { raw, .. } => {
+                assert_eq!(raw, &Value::String("this is not json".to_string()));
+            }
+            other => panic!("expected unknown, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn schema_fingerprint_detects_export() {
+        let adapter = CursorAdapter;
+        let hdr = raw(r#"{"kind":"session_start","cwd":"/x"}"#, 1);
+        let msg = raw(r#"{"id":"m","role":"user","text":"hi"}"#, 2);
+        let junk = raw(r#"{"kind":"telemetry"}"#, 3);
+        assert_eq!(adapter.schema_fingerprint(&hdr).confidence, 100);
+        assert_eq!(adapter.schema_fingerprint(&msg).confidence, 100);
+        assert_eq!(adapter.schema_fingerprint(&junk).confidence, 0);
+    }
+
+    #[test]
+    fn session_id_set_before_header_is_learned_from_record() {
+        // Even without a header, the first record carrying sessionId seeds ctx.
+        let jsonl = r#"{"id":"m1","role":"user","sessionId":"late","text":"hi"}"#;
+        let (events, ctx) = parse_all(jsonl);
+        assert_eq!(ctx.session_id.as_deref(), Some("late"));
+        assert_eq!(events[0].session_id, "late");
+    }
+
+    // ---- on-disk fixture conformance ----
+    //
+    // The fixtures under `fixtures/cursor/v1/` ARE this tool's real record
+    // shape and feed the Phase-2 conformance suite. These tests parse them
+    // through the live adapter to guarantee the two never drift apart.
+
+    fn fixture(name: &str) -> String {
+        let path = std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
+            .join("../../fixtures/cursor/v1")
+            .join(name);
+        std::fs::read_to_string(&path)
+            .unwrap_or_else(|e| panic!("read fixture {}: {e}", path.display()))
+    }
+
+    #[test]
+    fn fixture_happy_path_decision_then_edits() {
+        let (events, ctx) = parse_all(&fixture("happy_path_decision_then_edits.jsonl"));
+        assert_eq!(ctx.session_id.as_deref(), Some("cur-sess-001"));
+        // header, decision (user), assistant+call+result+edit, assistant+call+result+edit, end.
+        assert_eq!(
+            tags(&events),
+            vec![
+                "session_start",
+                "user_turn",
+                "assistant_turn",
+                "tool_call",
+                "tool_result",
+                "file_edit",
+                "assistant_turn",
+                "tool_call",
+                "tool_result",
+                "file_edit",
+                "session_end",
+            ]
+        );
+        // The decision turn is the user's, and a FileEdit follows.
+        match &events[1].kind {
+            EventKind::UserTurn { text, .. } => assert!(text.contains("Postgres")),
+            other => panic!("expected user_turn, got {other:?}"),
+        }
+        assert!(events.iter().any(|e| matches!(
+            &e.kind,
+            EventKind::FileEdit { diff, .. }
+                if diff.path == std::path::Path::new("config/database.toml")
+        )));
+        // Project binding came from the session_start header.
+        assert_eq!(
+            events[1].project.cwd,
+            PathBuf::from("/Users/dev/projects/orders-api")
+        );
+        // Every tool result in this fixture succeeded.
+        assert!(ctx.call_ok.values().all(|ok| *ok));
+    }
+
+    #[test]
+    fn fixture_rejected_alternative_parses() {
+        let (events, _) = parse_all(&fixture("rejected_alternative.jsonl"));
+        // Contains a user decision to reject Redux followed by an edit.
+        assert!(events.iter().any(
+            |e| matches!(&e.kind, EventKind::UserTurn { text, .. } if text.contains("reject"))
+        ));
+        assert!(events
+            .iter()
+            .any(|e| matches!(&e.kind, EventKind::FileEdit { .. })));
+    }
+
+    #[test]
+    fn fixture_ban_parses_decision_and_edit() {
+        let (events, _) = parse_all(&fixture("ban.jsonl"));
+        assert!(events.iter().any(|e| matches!(
+            &e.kind,
+            EventKind::UserTurn { text, .. } if text.contains("never")
+        )));
+        assert!(events
+            .iter()
+            .any(|e| matches!(&e.kind, EventKind::FileEdit { .. })));
+    }
+
+    #[test]
+    fn fixture_tool_failure_edit_has_failed_result() {
+        // The edit's tool result FAILED → downstream must NOT mint an Episode.
+        // At the event level: there IS a FileEdit, but it is LINKED by call_id to
+        // a ToolResult with ok=false, so the segmenter drops it (no spurious
+        // episode, §8.2).
+        let (events, ctx) = parse_all(&fixture("tool_failure.jsonl"));
+        // The edit event still exists (losslessness), keyed to call-edit-4.
+        let edit = events
+            .iter()
+            .find_map(|e| match &e.kind {
+                EventKind::FileEdit { call_id, diff } => Some((call_id.clone(), diff.clone())),
+                _ => None,
+            })
+            .expect("an edit event");
+        assert_eq!(edit.1.path, PathBuf::from("deploy.sh"));
+        // The edit is tied to the failing call so the segmenter can drop it.
+        assert_eq!(edit.0.as_deref(), Some("call-edit-4"));
+        // The failing result is observable by call_id → the gate for "no Episode".
+        assert_eq!(ctx.call_ok.get("call-edit-4").copied(), Some(false));
+        // And the ToolResult event itself is marked not-ok.
+        assert!(events.iter().any(|e| matches!(
+            &e.kind,
+            EventKind::ToolResult { call_id, ok: false, .. } if call_id == "call-edit-4"
+        )));
+    }
+
+    #[test]
+    fn fixture_tool_failure_yields_no_episode_via_segmenter() {
+        // End-to-end through the segmenter: the failed edit must NOT mint an
+        // Episode, and the happy path must still mint two.
+        use memscribe_core::gate::CommitmentGate;
+        use memscribe_core::segmenter::{DefaultSegmenter, Segmenter};
+
+        let gate = CommitmentGate::default();
+        let seg = DefaultSegmenter;
+
+        let (fail_events, _) = parse_all(&fixture("tool_failure.jsonl"));
+        let fail_seg = seg.segment(&fail_events, &gate);
+        assert_eq!(
+            fail_seg.episodes.len(),
+            0,
+            "a failed edit must produce no episode"
+        );
+
+        let (ok_events, _) = parse_all(&fixture("happy_path_decision_then_edits.jsonl"));
+        let ok_seg = seg.segment(&ok_events, &gate);
+        assert_eq!(
+            ok_seg.episodes.len(),
+            2,
+            "the happy path must still produce two episodes"
+        );
+    }
+
+    #[test]
+    fn all_fixtures_lossless_and_never_error() {
+        for name in [
+            "happy_path_decision_then_edits.jsonl",
+            "rejected_alternative.jsonl",
+            "ban.jsonl",
+            "tool_failure.jsonl",
+        ] {
+            let (events, _) = parse_all(&fixture(name));
+            // No record silently vanished: a non-empty fixture yields events,
+            // and none degraded to Unknown (the shapes are all recognized).
+            assert!(!events.is_empty(), "{name} produced no events");
+            assert!(
+                events.iter().all(|e| e.kind.tag() != "unknown"),
+                "{name} produced an Unknown event"
+            );
+        }
+    }
+}
diff --git a/crates/memscribe-adapters/src/gemini.rs b/crates/memscribe-adapters/src/gemini.rs
new file mode 100644
index 0000000..608ef7b
--- /dev/null
+++ b/crates/memscribe-adapters/src/gemini.rs
@@ -0,0 +1,823 @@
+//! Gemini CLI adapter.
+//!
+//! Transcripts: `~/.gemini/tmp/<hash>/chats/session-*.jsonl`, append-only JSONL;
+//! also a legacy single-blob `.json` history. Each non-control line is a message
+//! record: `{role: user|gemini|model, text|content|parts, timestamp, thoughts,
+//! tokens:{input,output}|tokenCount, toolCalls:[{name, args, resultDisplay}]}`.
+//! Control records: `{"$set":{...}}` (session/cwd metadata) and
+//! `{"$rewindTo": <id|index>}` (logical truncation).
+//!
+//! Mapping (whitepaper §5 + Appendix A):
+//! - `role:user` → [`EventKind::UserTurn`].
+//! - `role:gemini|model` → [`EventKind::AssistantTurn`] with `thinking` from
+//!   `thoughts`, `usage` from `tokens`, and structured `parts`.
+//! - nested `toolCalls[]` → a [`EventKind::ToolCall`], and when `resultDisplay`
+//!   is present a [`EventKind::ToolResult`]; a `FileDiff`-shaped `resultDisplay`
+//!   additionally yields a [`EventKind::FileEdit`].
+//! - `{"$rewindTo"}` → [`EventKind::Rewind`].
+//! - `{"$set"}` that carries a cwd/project → [`EventKind::SessionStart`], else
+//!   [`EventKind::Unknown`].
+//!
+//! Quirks: tolerate the legacy single-blob `.json` and the `$set`/`$rewindTo`
+//! control records; prefer `chats/*.jsonl` over `logs.json`. The parser never
+//! panics, is fully deterministic, and routes anything unrecognized to
+//! [`EventKind::Unknown`] so the stream stays lossless.
+
+use crate::util;
+use memscribe_core::{
+    content_id, CaptureEvent, Diff, DiscoverCfg, EventKind, GitRef, ParseCtx, ParseError, Part,
+    ProjectRef, RawRecord, SchemaVariant, SourceKind, TranscriptAdapter, TranscriptHandle, Usage,
+};
+use serde_json::Value;
+use std::path::{Path, PathBuf};
+use walkdir::WalkDir;
+
+const SOURCE: SourceKind = SourceKind::Gemini;
+
+/// Adapter for Google Gemini CLI transcripts.
+#[derive(Debug, Default, Clone, Copy)]
+pub struct GeminiAdapter;
+
+impl TranscriptAdapter for GeminiAdapter {
+    fn source_kind(&self) -> SourceKind {
+        SOURCE
+    }
+
+    fn discover(&self, cfg: &DiscoverCfg) -> Vec<TranscriptHandle> {
+        discover_transcripts(cfg)
+    }
+
+    fn parse(&self, raw: &RawRecord, ctx: &mut ParseCtx) -> Result<Vec<CaptureEvent>, ParseError> {
+        let Some(value) = util::parse_json_line(raw) else {
+            // Blank line or invalid JSON: nothing to emit (blank) or an Unknown
+            // for non-empty-but-unparseable bytes — keep losslessness.
+            let s = raw.as_str().map(str::trim).unwrap_or("");
+            if s.is_empty() {
+                return Ok(Vec::new());
+            }
+            return Ok(vec![util::unknown_event(
+                SOURCE,
+                ctx,
+                raw,
+                Value::String(s.to_string()),
+            )]);
+        };
+        Ok(parse_value(raw, ctx, value))
+    }
+
+    fn schema_fingerprint(&self, sample: &RawRecord) -> SchemaVariant {
+        match util::parse_json_line(sample) {
+            Some(v) if v.get("$set").is_some() || v.get("$rewindTo").is_some() => {
+                SchemaVariant::certain(SOURCE, "gemini/control")
+            }
+            Some(v) if v.get("role").is_some() => SchemaVariant::certain(SOURCE, "gemini/chat-v1"),
+            Some(_) => SchemaVariant::unknown(SOURCE),
+            None => SchemaVariant::unknown(SOURCE),
+        }
+    }
+}
+
+/// Discover Gemini transcripts under `<home>/.gemini/tmp/<hash>/`.
+///
+/// Prefers `chats/session-*.jsonl` (and any `chats/*.jsonl`) over the legacy
+/// `logs.json`; only when a project directory has no JSONL chat does it fall
+/// back to a `logs.json` / `*.json` blob. Output is sorted for determinism.
+fn discover_transcripts(cfg: &DiscoverCfg) -> Vec<TranscriptHandle> {
+    let root = cfg.home_dir().join(".gemini").join("tmp");
+    if !root.is_dir() {
+        return Vec::new();
+    }
+
+    let mut jsonl: Vec<PathBuf> = Vec::new();
+    let mut blob: Vec<PathBuf> = Vec::new();
+    for entry in WalkDir::new(&root)
+        .into_iter()
+        .filter_map(std::result::Result::ok)
+    {
+        let path = entry.path();
+        if !path.is_file() {
+            continue;
+        }
+        match path.extension().and_then(|e| e.to_str()) {
+            Some("jsonl") => jsonl.push(path.to_path_buf()),
+            Some("json") => blob.push(path.to_path_buf()),
+            _ => {}
+        }
+    }
+
+    // Project hash = the directory directly under `tmp/`. If any `.jsonl` chat
+    // exists for a project, drop that project's `.json` blobs (prefer chats).
+    let projects_with_jsonl: std::collections::HashSet<PathBuf> =
+        jsonl.iter().filter_map(|p| project_dir(&root, p)).collect();
+    blob.retain(|p| match project_dir(&root, p) {
+        Some(proj) => !projects_with_jsonl.contains(&proj),
+        None => true,
+    });
+
+    let mut handles: Vec<TranscriptHandle> = jsonl
+        .into_iter()
+        .chain(blob)
+        .map(|path| TranscriptHandle {
+            session_hint: session_hint_of(&path),
+            path,
+            source: SOURCE,
+            compressed: false,
+        })
+        .collect();
+    handles.sort_by(|a, b| a.path.cmp(&b.path));
+    handles
+}
+
+/// The project-hash directory directly beneath `tmp/` for a transcript path.
+fn project_dir(root: &Path, path: &Path) -> Option<PathBuf> {
+    let rel = path.strip_prefix(root).ok()?;
+    let first = rel.components().next()?;
+    Some(root.join(first.as_os_str()))
+}
+
+/// Derive a session hint from a `session-<id>.jsonl` filename.
+fn session_hint_of(path: &Path) -> Option<String> {
+    let stem = path.file_stem().and_then(|s| s.to_str())?;
+    Some(stem.strip_prefix("session-").unwrap_or(stem).to_string())
+}
+
+/// Parse one already-decoded JSON record into zero or more events.
+fn parse_value(raw: &RawRecord, ctx: &mut ParseCtx, value: Value) -> Vec<CaptureEvent> {
+    if value.get("$rewindTo").is_some() {
+        return parse_rewind(raw, ctx, value);
+    }
+    if value.get("$set").is_some() {
+        return parse_set(raw, ctx, value);
+    }
+    if value.get("role").is_some() {
+        return parse_message(raw, ctx, value);
+    }
+    vec![util::unknown_event(SOURCE, ctx, raw, value)]
+}
+
+/// `{"$rewindTo": <id|index>}` → [`EventKind::Rewind`].
+fn parse_rewind(raw: &RawRecord, ctx: &mut ParseCtx, value: Value) -> Vec<CaptureEvent> {
+    let target = value.get("$rewindTo");
+    let to_event = match target {
+        Some(Value::String(s)) => s.clone(),
+        Some(other) => other.to_string(),
+        None => String::new(),
+    };
+    let event_id = record_id(&value, &raw.bytes);
+    if !ctx.first_seen(&event_id) {
+        return Vec::new();
+    }
+    let ts = util::ts_from(&value, TS_KEYS);
+    vec![util::mk_event(
+        SOURCE,
+        ctx,
+        raw,
+        event_id,
+        None,
+        ts,
+        EventKind::Rewind { to_event },
+    )]
+}
+
+/// `{"$set": {...}}` → [`EventKind::SessionStart`] when it carries a cwd/project,
+/// otherwise [`EventKind::Unknown`] (lossless).
+fn parse_set(raw: &RawRecord, ctx: &mut ParseCtx, value: Value) -> Vec<CaptureEvent> {
+    let set = value.get("$set");
+    // Learn the session id if present anywhere in the $set payload.
+    if let Some(sid) = set
+        .and_then(|s| first_str(s, &["sessionId", "session_id", "id"]))
+        .or_else(|| first_str(&value, &["sessionId", "session_id"]))
+    {
+        if ctx.session_id.is_none() {
+            ctx.session_id = Some(sid);
+        }
+    }
+
+    let cwd = set.and_then(|s| first_str(s, &["cwd", "projectRoot", "project_root", "workspace"]));
+    let event_id = record_id(&value, &raw.bytes);
+    if !ctx.first_seen(&event_id) {
+        return Vec::new();
+    }
+    let ts = util::ts_from(&value, TS_KEYS);
+
+    let Some(cwd) = cwd else {
+        // A `$set` with no project binding is metadata we don't model yet.
+        return vec![util::unknown_event(SOURCE, ctx, raw, value)];
+    };
+
+    let git = git_ref_from(set.unwrap_or(&value));
+    let model = set.and_then(|s| first_str(s, &["model", "modelName", "model_name"]));
+    let tool_version = set.and_then(|s| first_str(s, &["version", "cliVersion", "cli_version"]));
+
+    // Stamp the project binding so every later event inherits it.
+    let repo_root = set
+        .and_then(|s| first_str(s, &["repoRoot", "repo_root"]))
+        .map(PathBuf::from);
+    ctx.project = Some(ProjectRef {
+        cwd: PathBuf::from(&cwd),
+        repo_root,
+        git: git.clone(),
+    });
+
+    vec![util::mk_event(
+        SOURCE,
+        ctx,
+        raw,
+        event_id,
+        None,
+        ts,
+        EventKind::SessionStart {
+            cwd: PathBuf::from(cwd),
+            git,
+            model,
+            tool_version,
+        },
+    )]
+}
+
+/// A message record (`role: user|gemini|model`) → one turn plus any nested
+/// `toolCalls[]` as ToolCall/ToolResult/FileEdit events.
+fn parse_message(raw: &RawRecord, ctx: &mut ParseCtx, value: Value) -> Vec<CaptureEvent> {
+    let role = value.get("role").and_then(Value::as_str).unwrap_or("");
+    let event_id = record_id(&value, &raw.bytes);
+    if !ctx.first_seen(&event_id) {
+        // Idempotency: a repeated record produces nothing on re-ingest.
+        return Vec::new();
+    }
+    let ts = util::ts_from(&value, TS_KEYS);
+    let text = flatten_text(&value);
+
+    let mut out = Vec::new();
+    let kind = match role {
+        "user" => EventKind::UserTurn {
+            text,
+            parts: message_parts(&value),
+        },
+        "gemini" | "model" | "assistant" => {
+            let thinking = first_str(&value, &["thoughts", "thinking", "reasoning"]);
+            EventKind::AssistantTurn {
+                text,
+                thinking,
+                model: first_str(&value, &["model", "modelName", "model_name"]),
+                usage: usage_from(&value),
+                parts: message_parts(&value),
+            }
+        }
+        _ => {
+            // An unrecognized role is still a valid record: keep it verbatim.
+            return vec![util::unknown_event(SOURCE, ctx, raw, value)];
+        }
+    };
+    out.push(util::mk_event(
+        SOURCE,
+        ctx,
+        raw,
+        event_id.clone(),
+        None,
+        ts,
+        kind,
+    ));
+
+    // Nested tool calls become their own events, parented to the turn.
+    if let Some(calls) = value.get("toolCalls").and_then(Value::as_array) {
+        for (i, call) in calls.iter().enumerate() {
+            out.extend(parse_tool_call(raw, ctx, &event_id, ts, i, call));
+        }
+    }
+    out
+}
+
+/// One nested `toolCalls[]` entry → a ToolCall and, when `resultDisplay` is
+/// present, a ToolResult (+ a FileEdit for a `FileDiff`-shaped result).
+fn parse_tool_call(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    turn_id: &str,
+    ts: memscribe_core::Timestamp,
+    index: usize,
+    call: &Value,
+) -> Vec<CaptureEvent> {
+    let name = first_str(call, &["name", "tool", "toolName"]).unwrap_or_default();
+    let args = call
+        .get("args")
+        .or_else(|| call.get("arguments"))
+        .or_else(|| call.get("input"))
+        .cloned()
+        .unwrap_or(Value::Null);
+    // A deterministic, stable call id: native id if present, else turn+index.
+    let call_id = first_str(call, &["callId", "call_id", "id"])
+        .unwrap_or_else(|| format!("{turn_id}:tool:{index}"));
+
+    let mut out = Vec::new();
+    out.push(util::mk_event(
+        SOURCE,
+        ctx,
+        raw,
+        format!("{call_id}:call"),
+        Some(turn_id.to_string()),
+        ts,
+        EventKind::ToolCall {
+            call_id: call_id.clone(),
+            name,
+            args,
+        },
+    ));
+
+    let Some(result) = call.get("resultDisplay").or_else(|| call.get("result")) else {
+        return out;
+    };
+
+    let ok = result_ok(call, result);
+    ctx.call_ok.insert(call_id.clone(), ok);
+    out.push(util::mk_event(
+        SOURCE,
+        ctx,
+        raw,
+        format!("{call_id}:result"),
+        Some(turn_id.to_string()),
+        ts,
+        EventKind::ToolResult {
+            call_id: call_id.clone(),
+            ok,
+            output: result.clone(),
+        },
+    ));
+
+    if let Some(diff) = file_diff_from(result) {
+        out.push(util::mk_event(
+            SOURCE,
+            ctx,
+            raw,
+            format!("{call_id}:edit"),
+            Some(turn_id.to_string()),
+            ts,
+            EventKind::FileEdit {
+                call_id: Some(call_id),
+                diff,
+            },
+        ));
+    }
+    out
+}
+
+/// Timestamp keys Gemini may use, in priority order.
+const TS_KEYS: &[&str] = &["timestamp", "time", "ts", "createdAt", "created_at"];
+
+/// A record's native id, falling back to a `blake3` content hash.
+fn record_id(value: &Value, bytes: &[u8]) -> String {
+    first_str(value, &["id", "messageId", "message_id", "uuid"])
+        .unwrap_or_else(|| content_id(bytes))
+}
+
+/// The first string-valued key from `keys` present on `value` (non-empty).
+fn first_str(value: &Value, keys: &[&str]) -> Option<String> {
+    for k in keys {
+        if let Some(s) = value.get(*k).and_then(Value::as_str) {
+            if !s.is_empty() {
+                return Some(s.to_string());
+            }
+        }
+    }
+    None
+}
+
+/// Flatten a message's textual content from `text`, `content`, or `parts[]`.
+fn flatten_text(value: &Value) -> String {
+    if let Some(s) = value.get("text").and_then(Value::as_str) {
+        return s.to_string();
+    }
+    if let Some(s) = value.get("content").and_then(Value::as_str) {
+        return s.to_string();
+    }
+    if let Some(parts) = value
+        .get("parts")
+        .or_else(|| value.get("content"))
+        .and_then(Value::as_array)
+    {
+        let mut buf = String::new();
+        for p in parts {
+            if let Some(s) = p.as_str() {
+                buf.push_str(s);
+            } else if let Some(s) = p.get("text").and_then(Value::as_str) {
+                buf.push_str(s);
+            }
+        }
+        return buf;
+    }
+    String::new()
+}
+
+/// Structured `parts[]`, preserving anything we don't recognize as [`Part::Other`].
+fn message_parts(value: &Value) -> Vec<Part> {
+    let Some(parts) = value.get("parts").and_then(Value::as_array) else {
+        return Vec::new();
+    };
+    parts
+        .iter()
+        .map(|p| {
+            if let Some(s) = p.as_str() {
+                Part::Text {
+                    text: s.to_string(),
+                }
+            } else if let Some(s) = p.get("text").and_then(Value::as_str) {
+                Part::Text {
+                    text: s.to_string(),
+                }
+            } else if let Some(s) = p.get("thought").and_then(Value::as_str) {
+                Part::Thinking {
+                    text: s.to_string(),
+                }
+            } else {
+                Part::Other { raw: p.clone() }
+            }
+        })
+        .collect()
+}
+
+/// Token usage from `tokens:{input,output,...}` or a flat `tokenCount`.
+fn usage_from(value: &Value) -> Option<Usage> {
+    if let Some(tokens) = value.get("tokens").filter(|v| v.is_object()) {
+        let usage = Usage {
+            input_tokens: u64_at(tokens, &["input", "inputTokens", "prompt", "promptTokens"]),
+            output_tokens: u64_at(
+                tokens,
+                &["output", "outputTokens", "completion", "completionTokens"],
+            ),
+            cache_read_tokens: u64_at(tokens, &["cacheRead", "cached", "cachedContentTokens"]),
+            cache_creation_tokens: u64_at(tokens, &["cacheCreation", "cacheWrite"]),
+        };
+        if usage != Usage::default() {
+            return Some(usage);
+        }
+    }
+    if let Some(total) = u64_at(value, &["tokenCount", "totalTokens"]) {
+        return Some(Usage {
+            output_tokens: Some(total),
+            ..Usage::default()
+        });
+    }
+    None
+}
+
+/// First unsigned-integer value among `keys` on `value`.
+fn u64_at(value: &Value, keys: &[&str]) -> Option<u64> {
+    for k in keys {
+        if let Some(n) = value.get(*k).and_then(Value::as_u64) {
+            return Some(n);
+        }
+    }
+    None
+}
+
+/// A git ref from `commit`/`sha` (+ optional `branch`) within a `$set` payload.
+fn git_ref_from(value: &Value) -> Option<GitRef> {
+    let sha = first_str(value, &["commit", "sha", "head", "gitCommit"])?;
+    Some(GitRef {
+        sha,
+        branch: first_str(value, &["branch", "gitBranch"]),
+    })
+}
+
+/// Whether a tool result is a success. A `FileDiff`-shaped result is a success
+/// by construction; otherwise an explicit `error`/`success`/`ok`/`status` field
+/// decides, defaulting to success when none is present.
+fn result_ok(call: &Value, result: &Value) -> bool {
+    for v in [call, result] {
+        if let Some(b) = v.get("success").and_then(Value::as_bool) {
+            return b;
+        }
+        if let Some(b) = v.get("ok").and_then(Value::as_bool) {
+            return b;
+        }
+        if let Some(b) = v.get("error").and_then(Value::as_bool) {
+            return !b;
+        }
+        if let Some(s) = v.get("status").and_then(Value::as_str) {
+            let s = s.to_ascii_lowercase();
+            if s == "error" || s == "failed" || s == "failure" || s == "rejected" {
+                return false;
+            }
+            if s == "success" || s == "ok" || s == "completed" {
+                return true;
+            }
+        }
+        // A non-empty `error` string/object means failure.
+        match v.get("error") {
+            Some(Value::String(s)) if !s.is_empty() => return false,
+            Some(Value::Object(o)) if !o.is_empty() => return false,
+            _ => {}
+        }
+    }
+    true
+}
+
+/// A normalized [`Diff`] from a `FileDiff`-shaped `resultDisplay`, if it looks
+/// like one (`fileName`/`filePath` plus diff content). Returns `None` for
+/// non-edit results.
+fn file_diff_from(result: &Value) -> Option<Diff> {
+    let obj = result.as_object()?;
+    let path = first_str(result, &["fileName", "filePath", "file", "path"])?;
+    let has_edit_shape = obj.contains_key("originalContent")
+        || obj.contains_key("newContent")
+        || obj.contains_key("fileDiff")
+        || obj.contains_key("diff")
+        || obj.contains_key("diffStat");
+    if !has_edit_shape {
+        return None;
+    }
+
+    let old = first_str(result, &["originalContent", "oldContent", "old"]);
+    let new = first_str(result, &["newContent", "new"]);
+    let unified = first_str(result, &["fileDiff", "diff", "unified"]);
+
+    let (added, removed) = diff_stat(result);
+    Some(Diff {
+        path: PathBuf::from(path),
+        old,
+        new,
+        unified,
+        added_lines: added,
+        removed_lines: removed,
+    })
+}
+
+/// Added/removed line counts from `diffStat:{added,removed}` or the
+/// `model_added_lines`/`model_removed_lines` shape.
+fn diff_stat(result: &Value) -> (u32, u32) {
+    if let Some(stat) = result.get("diffStat").filter(|v| v.is_object()) {
+        let added = u64_at(stat, &["added", "additions", "model_added_lines"]).unwrap_or(0);
+        let removed = u64_at(stat, &["removed", "deletions", "model_removed_lines"]).unwrap_or(0);
+        return (clamp_u32(added), clamp_u32(removed));
+    }
+    let added = u64_at(result, &["model_added_lines", "added"]).unwrap_or(0);
+    let removed = u64_at(result, &["model_removed_lines", "removed"]).unwrap_or(0);
+    (clamp_u32(added), clamp_u32(removed))
+}
+
+/// Saturate a `u64` line count into the model's `u32` field.
+fn clamp_u32(n: u64) -> u32 {
+    u32::try_from(n).unwrap_or(u32::MAX)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use memscribe_core::SourceLocation;
+
+    fn raw(line: &str) -> RawRecord {
+        RawRecord::from_line(line, SourceLocation::new("session-x.jsonl", 0, 1))
+    }
+
+    fn parse_all(lines: &[&str]) -> Vec<CaptureEvent> {
+        let adapter = GeminiAdapter;
+        let mut ctx = ParseCtx::new();
+        let mut out = Vec::new();
+        for (i, l) in lines.iter().enumerate() {
+            let r =
+                RawRecord::from_line(l, SourceLocation::new("session-x.jsonl", 0, i as u64 + 1));
+            out.extend(adapter.parse(&r, &mut ctx).expect("never errors"));
+        }
+        out
+    }
+
+    fn tags(events: &[CaptureEvent]) -> Vec<&'static str> {
+        events.iter().map(|e| e.kind.tag()).collect()
+    }
+
+    #[test]
+    fn set_with_cwd_is_session_start_and_binds_project() {
+        let line = r#"{"$set":{"sessionId":"sess-1","cwd":"/home/u/app","model":"gemini-2.5-pro","branch":"main","commit":"abc123"}}"#;
+        let evs = parse_all(&[line]);
+        assert_eq!(tags(&evs), ["session_start"]);
+        assert_eq!(evs[0].session_id, "sess-1");
+        assert_eq!(evs[0].project.cwd, PathBuf::from("/home/u/app"));
+        match &evs[0].kind {
+            EventKind::SessionStart {
+                cwd, git, model, ..
+            } => {
+                assert_eq!(cwd, &PathBuf::from("/home/u/app"));
+                assert_eq!(model.as_deref(), Some("gemini-2.5-pro"));
+                let git = git.as_ref().expect("git ref");
+                assert_eq!(git.sha, "abc123");
+                assert_eq!(git.branch.as_deref(), Some("main"));
+            }
+            other => panic!("expected SessionStart, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn set_without_project_is_unknown_not_session_start() {
+        let evs = parse_all(&[r#"{"$set":{"theme":"dark"}}"#]);
+        assert_eq!(tags(&evs), ["unknown"]);
+    }
+
+    #[test]
+    fn user_then_assistant_with_edit_yields_decision_then_file_edit() {
+        // A user decision turn, then an assistant turn whose tool call edits a
+        // file — the canonical happy path: UserTurn then (eventually) FileEdit.
+        let user = r#"{"id":"m1","role":"user","text":"Let's use Postgres instead of MySQL.","timestamp":"2026-06-22T10:00:00Z"}"#;
+        let asst = r#"{"id":"m2","role":"model","text":"Switching the driver.","thoughts":"swap the dep","tokens":{"input":12,"output":34},"timestamp":"2026-06-22T10:00:01Z","toolCalls":[{"name":"write_file","args":{"path":"db.rs"},"resultDisplay":{"fileName":"db.rs","originalContent":"mysql","newContent":"postgres","fileDiff":"@@ -1 +1 @@\n-mysql\n+postgres","diffStat":{"added":1,"removed":1}}}]}"#;
+        let evs = parse_all(&[user, asst]);
+        assert_eq!(
+            tags(&evs),
+            [
+                "user_turn",
+                "assistant_turn",
+                "tool_call",
+                "tool_result",
+                "file_edit",
+            ]
+        );
+
+        // UserTurn first.
+        match &evs[0].kind {
+            EventKind::UserTurn { text, .. } => {
+                assert!(text.contains("instead of MySQL"));
+            }
+            other => panic!("expected UserTurn, got {other:?}"),
+        }
+        // AssistantTurn carries thinking + usage.
+        match &evs[1].kind {
+            EventKind::AssistantTurn {
+                thinking, usage, ..
+            } => {
+                assert_eq!(thinking.as_deref(), Some("swap the dep"));
+                let u = usage.as_ref().expect("usage");
+                assert_eq!(u.input_tokens, Some(12));
+                assert_eq!(u.output_tokens, Some(34));
+            }
+            other => panic!("expected AssistantTurn, got {other:?}"),
+        }
+        // FileEdit normalized from the FileDiff.
+        match &evs[4].kind {
+            EventKind::FileEdit { call_id, diff } => {
+                assert!(call_id.is_some());
+                assert_eq!(diff.path, PathBuf::from("db.rs"));
+                assert_eq!(diff.old.as_deref(), Some("mysql"));
+                assert_eq!(diff.new.as_deref(), Some("postgres"));
+                assert_eq!(diff.added_lines, 1);
+                assert_eq!(diff.removed_lines, 1);
+                assert!(diff.unified.as_deref().unwrap().contains("+postgres"));
+            }
+            other => panic!("expected FileEdit, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn failed_tool_result_reports_ok_false() {
+        // An edit whose tool result failed must surface ToolResult.ok=false so
+        // the downstream segmenter suppresses the episode. The FileEdit still
+        // carries the same call_id the failed result keys on.
+        let line = r#"{"id":"m9","role":"model","text":"trying","toolCalls":[{"callId":"c7","name":"write_file","args":{"path":"x.rs"},"resultDisplay":{"fileName":"x.rs","newContent":"...","fileDiff":"@@","error":"permission denied"}}]}"#;
+        let evs = parse_all(&[line]);
+        assert_eq!(
+            tags(&evs),
+            ["assistant_turn", "tool_call", "tool_result", "file_edit"]
+        );
+        let result = evs
+            .iter()
+            .find(|e| matches!(e.kind, EventKind::ToolResult { .. }))
+            .unwrap();
+        let edit = evs
+            .iter()
+            .find(|e| matches!(e.kind, EventKind::FileEdit { .. }))
+            .unwrap();
+        match (&result.kind, &edit.kind) {
+            (
+                EventKind::ToolResult {
+                    call_id: rid, ok, ..
+                },
+                EventKind::FileEdit {
+                    call_id: Some(eid), ..
+                },
+            ) => {
+                assert!(!ok, "failed result must be ok=false");
+                assert_eq!(rid, eid, "edit and failed result must share call_id");
+            }
+            other => panic!("unexpected kinds: {other:?}"),
+        }
+    }
+
+    #[test]
+    fn rewind_control_record_maps_to_rewind() {
+        let evs = parse_all(&[r#"{"$rewindTo":"m1"}"#]);
+        assert_eq!(tags(&evs), ["rewind"]);
+        match &evs[0].kind {
+            EventKind::Rewind { to_event } => assert_eq!(to_event, "m1"),
+            other => panic!("expected Rewind, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn rewind_to_numeric_index_stringifies() {
+        let evs = parse_all(&[r#"{"$rewindTo":3}"#]);
+        match &evs[0].kind {
+            EventKind::Rewind { to_event } => assert_eq!(to_event, "3"),
+            other => panic!("expected Rewind, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn garbage_never_panics_and_is_lossless() {
+        // Invalid JSON, an empty object, a number, an unknown role, a blank
+        // line: none may panic, none may be silently dropped (except blanks).
+        let evs = parse_all(&[
+            "not json at all",
+            "{}",
+            "42",
+            r#"{"role":"system","text":"?"}"#,
+            "   ",
+            r#"{"foo":"bar"}"#,
+        ]);
+        // The blank line yields nothing; everything else is at least Unknown.
+        assert_eq!(evs.len(), 5);
+        assert!(evs.iter().all(|e| {
+            matches!(
+                e.kind.tag(),
+                "unknown" | "user_turn" | "assistant_turn" | "session_start"
+            )
+        }));
+        // The unknown-role record is preserved verbatim, not dropped.
+        assert!(evs.iter().any(|e| e.kind.tag() == "unknown"));
+    }
+
+    #[test]
+    fn repeated_record_is_deduped_for_idempotency() {
+        let line = r#"{"id":"dup-1","role":"user","text":"hello"}"#;
+        let once = parse_all(&[line]);
+        assert_eq!(tags(&once), ["user_turn"]);
+        // Re-ingesting the SAME record id within the session yields nothing.
+        let twice = parse_all(&[line, line]);
+        assert_eq!(tags(&twice), ["user_turn"]);
+    }
+
+    #[test]
+    fn parse_is_deterministic() {
+        let lines = [
+            r#"{"$set":{"sessionId":"s","cwd":"/w"}}"#,
+            r#"{"id":"a","role":"user","text":"go with Stripe instead of PayPal"}"#,
+            r#"{"id":"b","role":"model","text":"ok","toolCalls":[{"name":"edit","args":{},"resultDisplay":{"fileName":"a.rs","newContent":"x","fileDiff":"@@","diffStat":{"added":1,"removed":0}}}]}"#,
+        ];
+        let a = parse_all(&lines);
+        let b = parse_all(&lines);
+        assert_eq!(
+            serde_json::to_string(&a).unwrap(),
+            serde_json::to_string(&b).unwrap()
+        );
+    }
+
+    #[test]
+    fn seq_is_monotonic_across_a_message_with_tool_calls() {
+        let lines = [
+            r#"{"id":"a","role":"user","text":"hi"}"#,
+            r#"{"id":"b","role":"model","text":"editing","toolCalls":[{"name":"e","args":{},"resultDisplay":{"fileName":"f","newContent":"n","diff":"d"}}]}"#,
+        ];
+        let evs = parse_all(&lines);
+        for w in evs.windows(2) {
+            assert!(w[1].seq > w[0].seq, "seq must strictly increase");
+        }
+    }
+
+    #[test]
+    fn legacy_text_only_assistant_has_no_usage_when_absent() {
+        let evs = parse_all(&[r#"{"id":"z","role":"gemini","content":"plain reply"}"#]);
+        match &evs[0].kind {
+            EventKind::AssistantTurn { text, usage, .. } => {
+                assert_eq!(text, "plain reply");
+                assert!(usage.is_none());
+            }
+            other => panic!("expected AssistantTurn, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn token_count_flat_field_becomes_output_usage() {
+        let evs = parse_all(&[r#"{"id":"t","role":"model","text":"x","tokenCount":99}"#]);
+        match &evs[0].kind {
+            EventKind::AssistantTurn { usage, .. } => {
+                assert_eq!(usage.as_ref().unwrap().output_tokens, Some(99));
+            }
+            other => panic!("expected AssistantTurn, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn schema_fingerprint_classifies_records() {
+        let a = GeminiAdapter;
+        assert_eq!(
+            a.schema_fingerprint(&raw(r#"{"$set":{"cwd":"/w"}}"#))
+                .variant,
+            "gemini/control"
+        );
+        assert_eq!(
+            a.schema_fingerprint(&raw(r#"{"$rewindTo":1}"#)).variant,
+            "gemini/control"
+        );
+        assert_eq!(
+            a.schema_fingerprint(&raw(r#"{"role":"user","text":"hi"}"#))
+                .variant,
+            "gemini/chat-v1"
+        );
+        assert_eq!(a.schema_fingerprint(&raw("garbage")).confidence, 0);
+    }
+}
diff --git a/crates/memscribe-adapters/src/lib.rs b/crates/memscribe-adapters/src/lib.rs
new file mode 100644
index 0000000..9857030
--- /dev/null
+++ b/crates/memscribe-adapters/src/lib.rs
@@ -0,0 +1,37 @@
+//! # memscribe-adapters
+//!
+//! Per-tool transcript adapters. Each tool implements
+//! [`memscribe_core::TranscriptAdapter`]: where its logs live, and how to turn
+//! one raw record into normalized [`memscribe_core::CaptureEvent`]s. Parsers are
+//! **version-tolerant** — they route anything unrecognized to
+//! [`memscribe_core::EventKind::Unknown`] rather than failing — and **must never
+//! panic** (every parser has a fuzz target).
+//!
+//! Adapters are behind feature flags so a consumer can compile only the tools it
+//! needs. The [`registry`] assembles the set of enabled adapters.
+#![forbid(unsafe_code)]
+
+pub mod util;
+
+#[cfg(feature = "claude_code")]
+pub mod claude_code;
+#[cfg(feature = "codex")]
+pub mod codex;
+#[cfg(feature = "copilot")]
+pub mod copilot;
+#[cfg(feature = "cursor")]
+pub mod cursor;
+#[cfg(feature = "gemini")]
+pub mod gemini;
+#[cfg(feature = "otel")]
+pub mod otel;
+#[cfg(feature = "vscode")]
+pub mod vscode;
+#[cfg(feature = "windsurf")]
+pub mod windsurf;
+#[cfg(feature = "zed")]
+pub mod zed;
+
+pub mod registry;
+
+pub use registry::{adapter_for, all_adapters};
diff --git a/crates/memscribe-adapters/src/otel.rs b/crates/memscribe-adapters/src/otel.rs
new file mode 100644
index 0000000..3fe9ba2
--- /dev/null
+++ b/crates/memscribe-adapters/src/otel.rs
@@ -0,0 +1,806 @@
+//! OpenTelemetry GenAI adapter.
+//!
+//! Source: OTLP push (local collector / file), OTel GenAI semconv records as
+//! JSON / NDJSON — the universal fallback channel for any instrumented agent.
+//! Each line is a log record or span. Attributes come in two shapes and this
+//! adapter accepts both:
+//!
+//! - **flat**: `{ "gen_ai.input.messages": [...], "gen_ai.system": "...", ... }`
+//! - **OTLP nested**: `{ "attributes": [ { "key": "...",
+//!   "value": { "stringValue" | "intValue" | "arrayValue" } }, ... ] }`
+//!
+//! Mapping (whitepaper §5):
+//! - `gen_ai.input.messages` (role `user`) and `gen_ai.cli.user_prompt`
+//!   → [`EventKind::UserTurn`]
+//! - `gen_ai.output.messages` → [`EventKind::AssistantTurn`] (with model + usage)
+//! - `execute_tool` span → [`EventKind::ToolCall`] (+ [`EventKind::ToolResult`]
+//!   when result attributes are present)
+//! - `file_operation` span (and `execute_tool` edits) → [`EventKind::FileEdit`]
+//!   with `file.path`, `model_added_lines`/`code.added_lines`,
+//!   `model_removed_lines`/`code.removed_lines`
+//! - `gen_ai.conversation.id` → session id; record `time` → timestamp.
+//!
+//! Anything well-formed but unrecognized is routed to [`EventKind::Unknown`] so
+//! the stream stays lossless. The parser never panics.
+
+use crate::util;
+use memscribe_core::{
+    content_id, CaptureEvent, Diff, DiscoverCfg, EventKind, GitRef, ParseCtx, ParseError, Part,
+    ProjectRef, RawRecord, SchemaVariant, SourceKind, Timestamp, TranscriptAdapter,
+    TranscriptHandle, Usage,
+};
+use serde_json::Value;
+use std::path::PathBuf;
+
+/// Adapter for OpenTelemetry GenAI records.
+#[derive(Debug, Default, Clone, Copy)]
+pub struct OtelAdapter;
+
+impl TranscriptAdapter for OtelAdapter {
+    fn source_kind(&self) -> SourceKind {
+        SourceKind::Otel
+    }
+
+    fn discover(&self, _cfg: &DiscoverCfg) -> Vec<TranscriptHandle> {
+        // OTel records are pushed (collector / file tail) rather than discovered
+        // in a well-known per-tool directory, so there is nothing to glob.
+        Vec::new()
+    }
+
+    fn parse(&self, raw: &RawRecord, ctx: &mut ParseCtx) -> Result<Vec<CaptureEvent>, ParseError> {
+        let Some(value) = util::parse_json_line(raw) else {
+            // Blank line → nothing; non-JSON → lossless Unknown of the raw text.
+            let s = raw.as_str().map(str::trim).unwrap_or("");
+            if s.is_empty() {
+                return Ok(Vec::new());
+            }
+            return Ok(vec![util::unknown_event(
+                SourceKind::Otel,
+                ctx,
+                raw,
+                Value::String(s.to_string()),
+            )]);
+        };
+
+        // Normalize both shapes into a flat attribute view.
+        let attrs = Attrs::from_record(&value);
+
+        // Learn the session id and project binding as early as we can.
+        if ctx.session_id.is_none() {
+            if let Some(sid) = attrs.str("gen_ai.conversation.id") {
+                ctx.session_id = Some(sid.to_string());
+            }
+        }
+
+        let ts = attrs.timestamp();
+        let op = attrs.operation_name();
+
+        let events = match op.as_deref() {
+            Some("session.start") | Some("session_start") | Some("gen_ai.session.start") => {
+                vec![self.session_start(ctx, raw, &attrs, ts)]
+            }
+            Some("session.end") | Some("session_end") | Some("gen_ai.session.end") => {
+                vec![mk(
+                    ctx,
+                    raw,
+                    derive_id(raw, "session_end", 0),
+                    ts,
+                    EventKind::SessionEnd {
+                        reason: attrs.str("reason").map(str::to_string),
+                    },
+                )]
+            }
+            Some("execute_tool") | Some("gen_ai.execute_tool") => {
+                self.execute_tool(ctx, raw, &attrs, ts)
+            }
+            Some("file_operation") | Some("gen_ai.file_operation") => {
+                self.file_operation(ctx, raw, &attrs, ts)
+            }
+            // Chat / inference records carry the dialogue.
+            _ => self.dialogue(ctx, raw, &attrs, ts, &value),
+        };
+
+        // Dedup / idempotency: drop any event whose id we have already emitted.
+        Ok(self.dedup(ctx, events))
+    }
+
+    fn schema_fingerprint(&self, sample: &RawRecord) -> SchemaVariant {
+        let Some(value) = util::parse_json_line(sample) else {
+            return SchemaVariant::unknown(SourceKind::Otel);
+        };
+        // The OTLP nested shape carries an `attributes` array of {key,value}.
+        if value
+            .get("attributes")
+            .and_then(Value::as_array)
+            .is_some_and(|a| a.iter().any(|e| e.get("key").is_some()))
+        {
+            return SchemaVariant::certain(SourceKind::Otel, "otel/genai-otlp");
+        }
+        // The flat shape uses dotted `gen_ai.*` keys directly on the object.
+        if value
+            .as_object()
+            .is_some_and(|m| m.keys().any(|k| k.starts_with("gen_ai.")))
+        {
+            return SchemaVariant::certain(SourceKind::Otel, "otel/genai-flat");
+        }
+        SchemaVariant::unknown(SourceKind::Otel)
+    }
+}
+
+impl OtelAdapter {
+    fn session_start(
+        &self,
+        ctx: &mut ParseCtx,
+        raw: &RawRecord,
+        attrs: &Attrs,
+        ts: Timestamp,
+    ) -> CaptureEvent {
+        let cwd: PathBuf = attrs
+            .str("cwd")
+            .or_else(|| attrs.str("gen_ai.cli.cwd"))
+            .map(PathBuf::from)
+            .unwrap_or_else(|| PathBuf::from("."));
+        let git = attrs.str("git.commit").map(|sha| GitRef {
+            sha: sha.to_string(),
+            branch: attrs.str("git.branch").map(str::to_string),
+        });
+        // Stamp the project binding for the rest of the session.
+        if ctx.project.is_none() {
+            ctx.project = Some(ProjectRef {
+                cwd: cwd.clone(),
+                repo_root: attrs.str("repo_root").map(PathBuf::from),
+                git: git.clone(),
+            });
+        }
+        mk(
+            ctx,
+            raw,
+            derive_id(raw, "session_start", 0),
+            ts,
+            EventKind::SessionStart {
+                cwd,
+                git,
+                model: attrs.str("gen_ai.request.model").map(str::to_string),
+                tool_version: attrs
+                    .str("gen_ai.tool.version")
+                    .or_else(|| attrs.str("service.version"))
+                    .map(str::to_string),
+            },
+        )
+    }
+
+    /// `execute_tool` span → a ToolCall, plus a ToolResult when the span carries
+    /// result/error attributes, plus a FileEdit when it carries `file.path`.
+    fn execute_tool(
+        &self,
+        ctx: &mut ParseCtx,
+        raw: &RawRecord,
+        attrs: &Attrs,
+        ts: Timestamp,
+    ) -> Vec<CaptureEvent> {
+        let mut out = Vec::new();
+        let name = attrs.str("gen_ai.tool.name").unwrap_or("tool").to_string();
+        let call_id = attrs
+            .str("gen_ai.tool.call.id")
+            .map(str::to_string)
+            .unwrap_or_else(|| content_id(&raw.bytes));
+        let args = attrs
+            .value("gen_ai.tool.call.arguments")
+            .cloned()
+            .unwrap_or(Value::Null);
+
+        ctx.call_names.insert(call_id.clone(), name.clone());
+        out.push(mk(
+            ctx,
+            raw,
+            derive_id(raw, "tool_call", 0),
+            ts,
+            EventKind::ToolCall {
+                call_id: call_id.clone(),
+                name: name.clone(),
+                args,
+            },
+        ));
+
+        // A result is present iff the span reports a status/result/error.
+        let ok = attrs.tool_ok();
+        if let Some(ok) = ok {
+            ctx.call_ok.insert(call_id.clone(), ok);
+            let output = attrs
+                .value("gen_ai.tool.result")
+                .cloned()
+                .unwrap_or(Value::Null);
+            out.push(mk(
+                ctx,
+                raw,
+                derive_id(raw, "tool_result", 0),
+                ts,
+                EventKind::ToolResult {
+                    call_id: call_id.clone(),
+                    ok,
+                    output,
+                },
+            ));
+        }
+
+        // An edit-shaped tool span also yields a FileEdit (keyed to the call).
+        if let Some(diff) = attrs.file_diff() {
+            out.push(mk(
+                ctx,
+                raw,
+                derive_id(raw, "file_edit", 0),
+                ts,
+                EventKind::FileEdit {
+                    call_id: Some(call_id),
+                    diff,
+                },
+            ));
+        }
+        out
+    }
+
+    /// `file_operation` span → a FileEdit (no call id).
+    fn file_operation(
+        &self,
+        ctx: &mut ParseCtx,
+        raw: &RawRecord,
+        attrs: &Attrs,
+        ts: Timestamp,
+    ) -> Vec<CaptureEvent> {
+        match attrs.file_diff() {
+            Some(diff) => vec![mk(
+                ctx,
+                raw,
+                derive_id(raw, "file_edit", 0),
+                ts,
+                EventKind::FileEdit {
+                    call_id: None,
+                    diff,
+                },
+            )],
+            // A file_operation without a path is unrecognized → lossless Unknown.
+            None => vec![util::unknown_event(
+                SourceKind::Otel,
+                ctx,
+                raw,
+                attrs.raw().clone(),
+            )],
+        }
+    }
+
+    /// A chat / inference record: zero or more UserTurns from input messages and
+    /// the CLI prompt, then zero or more AssistantTurns from output messages.
+    fn dialogue(
+        &self,
+        ctx: &mut ParseCtx,
+        raw: &RawRecord,
+        attrs: &Attrs,
+        ts: Timestamp,
+        value: &Value,
+    ) -> Vec<CaptureEvent> {
+        let mut out = Vec::new();
+
+        // `gen_ai.cli.user_prompt` → one UserTurn.
+        if let Some(prompt) = attrs.str("gen_ai.cli.user_prompt") {
+            if !prompt.is_empty() {
+                out.push(mk(
+                    ctx,
+                    raw,
+                    derive_id(raw, "user_prompt", 0),
+                    ts,
+                    EventKind::UserTurn {
+                        text: prompt.to_string(),
+                        parts: vec![Part::Text {
+                            text: prompt.to_string(),
+                        }],
+                    },
+                ));
+            }
+        }
+
+        // `gen_ai.input.messages` → one UserTurn per user-role message.
+        if let Some(msgs) = attrs.array("gen_ai.input.messages") {
+            for (i, m) in msgs.iter().enumerate() {
+                if !is_user_role(m) {
+                    continue;
+                }
+                let text = message_text(m);
+                out.push(mk(
+                    ctx,
+                    raw,
+                    derive_id(raw, "input_msg", i),
+                    ts,
+                    EventKind::UserTurn {
+                        text: text.clone(),
+                        parts: vec![Part::Text { text }],
+                    },
+                ));
+            }
+        }
+
+        // `gen_ai.output.messages` → one AssistantTurn per message.
+        if let Some(msgs) = attrs.array("gen_ai.output.messages") {
+            let model = attrs.str("gen_ai.request.model").map(str::to_string);
+            let usage = attrs.usage();
+            for (i, m) in msgs.iter().enumerate() {
+                let text = message_text(m);
+                out.push(mk(
+                    ctx,
+                    raw,
+                    derive_id(raw, "output_msg", i),
+                    ts,
+                    EventKind::AssistantTurn {
+                        text: text.clone(),
+                        thinking: None,
+                        model: model.clone(),
+                        usage: usage.clone(),
+                        parts: vec![Part::Text { text }],
+                    },
+                ));
+            }
+        }
+
+        // A record with none of the recognized dialogue fields is preserved
+        // verbatim so the stream stays lossless.
+        if out.is_empty() {
+            out.push(util::unknown_event(
+                SourceKind::Otel,
+                ctx,
+                raw,
+                value.clone(),
+            ));
+        }
+        out
+    }
+
+    /// Drop events whose ids have already been emitted (dedup / idempotency).
+    fn dedup(&self, ctx: &mut ParseCtx, events: Vec<CaptureEvent>) -> Vec<CaptureEvent> {
+        events
+            .into_iter()
+            .filter(|e| ctx.first_seen(&e.event_id))
+            .collect()
+    }
+}
+
+/// Build a normalized event with the OTel source.
+fn mk(
+    ctx: &mut ParseCtx,
+    raw: &RawRecord,
+    event_id: String,
+    ts: Timestamp,
+    kind: EventKind,
+) -> CaptureEvent {
+    util::mk_event(SourceKind::Otel, ctx, raw, event_id, None, ts, kind)
+}
+
+/// A deterministic per-logical-event id: the record's content hash plus a stable
+/// `kind`/index suffix so multiple events from one record don't collide and a
+/// repeated record dedups to the same ids.
+fn derive_id(raw: &RawRecord, kind: &str, index: usize) -> String {
+    format!("{}:{kind}:{index}", content_id(&raw.bytes))
+}
+
+/// Is a `gen_ai.*.messages` entry a user-role message?
+fn is_user_role(m: &Value) -> bool {
+    m.get("role")
+        .and_then(Value::as_str)
+        .map(|r| r.eq_ignore_ascii_case("user"))
+        .unwrap_or(false)
+}
+
+/// Flatten a GenAI message's text from `content` (string or parts array) or
+/// `parts` (array of `{type,text}` / `{content}` / strings).
+fn message_text(m: &Value) -> String {
+    if let Some(s) = m.get("content").and_then(Value::as_str) {
+        return s.to_string();
+    }
+    let mut buf = String::new();
+    for key in ["parts", "content"] {
+        if let Some(arr) = m.get(key).and_then(Value::as_array) {
+            for p in arr {
+                if let Some(s) = p.as_str() {
+                    push_part(&mut buf, s);
+                } else if let Some(s) = p
+                    .get("text")
+                    .or_else(|| p.get("content"))
+                    .and_then(Value::as_str)
+                {
+                    push_part(&mut buf, s);
+                }
+            }
+        }
+    }
+    buf
+}
+
+fn push_part(buf: &mut String, s: &str) {
+    if !buf.is_empty() {
+        buf.push('\n');
+    }
+    buf.push_str(s);
+}
+
+/// A flat view over a GenAI record's attributes, hiding the flat-vs-OTLP shape.
+struct Attrs<'a> {
+    /// The flat record object, when the record is already flat.
+    flat: Option<&'a serde_json::Map<String, Value>>,
+    /// Materialized {key → value} from the OTLP `attributes` array, when nested.
+    nested: Option<std::collections::HashMap<String, Value>>,
+    /// The original record (for lossless Unknown fallbacks).
+    raw: &'a Value,
+}
+
+impl<'a> Attrs<'a> {
+    fn from_record(value: &'a Value) -> Self {
+        // OTLP nested: an `attributes: [{key, value:{...}}]` array.
+        if let Some(arr) = value.get("attributes").and_then(Value::as_array) {
+            let mut map = std::collections::HashMap::new();
+            for entry in arr {
+                if let Some(key) = entry.get("key").and_then(Value::as_str) {
+                    if let Some(v) = entry.get("value").map(otlp_value) {
+                        map.insert(key.to_string(), v);
+                    }
+                }
+            }
+            return Attrs {
+                flat: None,
+                nested: Some(map),
+                raw: value,
+            };
+        }
+        Attrs {
+            flat: value.as_object(),
+            nested: None,
+            raw: value,
+        }
+    }
+
+    fn raw(&self) -> &'a Value {
+        self.raw
+    }
+
+    fn value(&self, key: &str) -> Option<&Value> {
+        if let Some(m) = self.flat {
+            return m.get(key);
+        }
+        self.nested.as_ref().and_then(|m| m.get(key))
+    }
+
+    fn str(&self, key: &str) -> Option<&str> {
+        self.value(key).and_then(Value::as_str)
+    }
+
+    fn array(&self, key: &str) -> Option<&Vec<Value>> {
+        self.value(key).and_then(Value::as_array)
+    }
+
+    fn u64(&self, key: &str) -> Option<u64> {
+        let v = self.value(key)?;
+        if let Some(n) = v.as_u64() {
+            return Some(n);
+        }
+        // OTLP intValue is often a stringified integer.
+        v.as_str().and_then(|s| s.trim().parse::<u64>().ok())
+    }
+
+    fn u32(&self, key: &str) -> Option<u32> {
+        self.u64(key).and_then(|n| u32::try_from(n).ok())
+    }
+
+    fn operation_name(&self) -> Option<String> {
+        self.str("gen_ai.operation.name")
+            .or_else(|| self.str("operation.name"))
+            .or_else(|| self.str("name"))
+            .map(str::to_string)
+    }
+
+    fn timestamp(&self) -> Timestamp {
+        // Prefer record-level time fields (which may live outside `attributes`).
+        util::ts_from(
+            self.raw,
+            &[
+                "time",
+                "timestamp",
+                "timeUnixNano",
+                "observedTimeUnixNano",
+                "ts",
+            ],
+        )
+    }
+
+    fn usage(&self) -> Option<Usage> {
+        let input = self.u64("gen_ai.usage.input_tokens");
+        let output = self.u64("gen_ai.usage.output_tokens");
+        if input.is_none() && output.is_none() {
+            return None;
+        }
+        Some(Usage {
+            input_tokens: input,
+            output_tokens: output,
+            cache_read_tokens: self.u64("gen_ai.usage.cache_read_tokens"),
+            cache_creation_tokens: self.u64("gen_ai.usage.cache_creation_tokens"),
+        })
+    }
+
+    /// The success flag of a tool span, if any result/error attribute is present.
+    /// `None` means "no result observed on this span".
+    fn tool_ok(&self) -> Option<bool> {
+        if let Some(status) = self
+            .str("gen_ai.tool.result.status")
+            .or_else(|| self.str("otel.status_code"))
+            .or_else(|| self.str("status"))
+        {
+            let s = status.trim().to_ascii_lowercase();
+            return Some(!matches!(s.as_str(), "error" | "failed" | "failure" | "ko"));
+        }
+        if self.value("error.type").is_some() || self.value("exception.type").is_some() {
+            return Some(false);
+        }
+        if self.value("gen_ai.tool.result").is_some() {
+            return Some(true);
+        }
+        None
+    }
+
+    /// A normalized diff from a file-edit span, if it carries a `file.path`.
+    fn file_diff(&self) -> Option<Diff> {
+        let path = self
+            .str("file.path")
+            .or_else(|| self.str("code.filepath"))?;
+        let added = self
+            .u32("model_added_lines")
+            .or_else(|| self.u32("code.added_lines"))
+            .unwrap_or(0);
+        let removed = self
+            .u32("model_removed_lines")
+            .or_else(|| self.u32("code.removed_lines"))
+            .unwrap_or(0);
+        Some(Diff {
+            path: PathBuf::from(path),
+            old: None,
+            new: None,
+            unified: None,
+            added_lines: added,
+            removed_lines: removed,
+        })
+    }
+}
+
+/// Collapse an OTLP `value` object (`stringValue` / `intValue` / `boolValue` /
+/// `doubleValue` / `arrayValue` / `kvlistValue`) to a plain JSON value.
+fn otlp_value(v: &Value) -> Value {
+    if let Some(s) = v.get("stringValue") {
+        return s.clone();
+    }
+    if let Some(i) = v.get("intValue") {
+        return i.clone();
+    }
+    if let Some(b) = v.get("boolValue") {
+        return b.clone();
+    }
+    if let Some(d) = v.get("doubleValue") {
+        return d.clone();
+    }
+    if let Some(arr) = v.get("arrayValue").and_then(|a| a.get("values")) {
+        if let Some(items) = arr.as_array() {
+            return Value::Array(items.iter().map(otlp_value).collect());
+        }
+    }
+    if let Some(kv) = v.get("kvlistValue").and_then(|k| k.get("values")) {
+        if let Some(items) = kv.as_array() {
+            let mut map = serde_json::Map::new();
+            for item in items {
+                if let Some(key) = item.get("key").and_then(Value::as_str) {
+                    if let Some(val) = item.get("value").map(otlp_value) {
+                        map.insert(key.to_string(), val);
+                    }
+                }
+            }
+            return Value::Object(map);
+        }
+    }
+    // Already a plain scalar/array, or an unrecognized shape: pass through.
+    v.clone()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use memscribe_core::SourceLocation;
+
+    fn raw(s: &str) -> RawRecord {
+        RawRecord::from_line(s, SourceLocation::new("otel.jsonl", 0, 1))
+    }
+
+    fn parse_all(lines: &[&str]) -> (Vec<CaptureEvent>, ParseCtx) {
+        let adapter = OtelAdapter;
+        let mut ctx = ParseCtx::new();
+        let mut out = Vec::new();
+        for l in lines {
+            let evs = adapter.parse(&raw(l), &mut ctx).expect("never errors");
+            out.extend(evs);
+        }
+        (out, ctx)
+    }
+
+    fn tags(evs: &[CaptureEvent]) -> Vec<&'static str> {
+        evs.iter().map(|e| e.kind.tag()).collect()
+    }
+
+    const SESSION_START: &str = r#"{"time":"2026-06-22T10:00:00Z","gen_ai.operation.name":"session.start","gen_ai.conversation.id":"sess-1","gen_ai.request.model":"claude-opus-4-8","cwd":"/home/dev/svc","repo_root":"/home/dev/svc","git.commit":"abc1234","git.branch":"main"}"#;
+    const USER_PROMPT: &str = r#"{"time":"2026-06-22T10:00:05Z","gen_ai.operation.name":"chat","gen_ai.conversation.id":"sess-1","gen_ai.cli.user_prompt":"Let's use Postgres instead of MySQL."}"#;
+    const FILE_EDIT: &str = r#"{"time":"2026-06-22T10:00:15Z","gen_ai.operation.name":"file_operation","gen_ai.conversation.id":"sess-1","file.path":"db/config.rs","code.added_lines":12,"code.removed_lines":4}"#;
+
+    #[test]
+    fn normalized_sequence_for_a_small_session() {
+        let assistant = r#"{"gen_ai.operation.name":"chat","gen_ai.conversation.id":"sess-1","gen_ai.request.model":"claude-opus-4-8","gen_ai.usage.input_tokens":42,"gen_ai.usage.output_tokens":8,"gen_ai.output.messages":[{"role":"assistant","content":"Switching to Postgres."}]}"#;
+        let session_end = r#"{"gen_ai.operation.name":"session.end","gen_ai.conversation.id":"sess-1","reason":"done"}"#;
+        let (evs, ctx) = parse_all(&[
+            SESSION_START,
+            USER_PROMPT,
+            assistant,
+            FILE_EDIT,
+            session_end,
+        ]);
+        assert_eq!(
+            tags(&evs),
+            vec![
+                "session_start",
+                "user_turn",
+                "assistant_turn",
+                "file_edit",
+                "session_end"
+            ]
+        );
+        // Session id is learned from gen_ai.conversation.id and stamped.
+        assert_eq!(ctx.session_id.as_deref(), Some("sess-1"));
+        assert!(evs.iter().all(|e| e.session_id == "sess-1"));
+        // Project binding came from the session-start record.
+        let proj = ctx.project.expect("project set at session start");
+        assert_eq!(proj.cwd, PathBuf::from("/home/dev/svc"));
+        assert_eq!(proj.git.as_ref().map(|g| g.sha.as_str()), Some("abc1234"));
+    }
+
+    #[test]
+    fn decision_then_edit_produces_user_turn_then_file_edit() {
+        let (evs, _) = parse_all(&[USER_PROMPT, FILE_EDIT]);
+        assert_eq!(tags(&evs), vec!["user_turn", "file_edit"]);
+        match &evs[0].kind {
+            EventKind::UserTurn { text, .. } => {
+                assert!(text.contains("Postgres"));
+            }
+            other => panic!("expected user_turn, got {other:?}"),
+        }
+        match &evs[1].kind {
+            EventKind::FileEdit { diff, call_id } => {
+                assert_eq!(diff.path, PathBuf::from("db/config.rs"));
+                assert_eq!(diff.added_lines, 12);
+                assert_eq!(diff.removed_lines, 4);
+                assert!(call_id.is_none());
+            }
+            other => panic!("expected file_edit, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn input_messages_become_user_turns() {
+        let line = r#"{"gen_ai.operation.name":"chat","gen_ai.conversation.id":"s","gen_ai.input.messages":[{"role":"user","content":"hello"},{"role":"system","content":"ignore"},{"role":"user","parts":[{"type":"text","text":"world"}]}]}"#;
+        let (evs, _) = parse_all(&[line]);
+        // Only the two user-role messages map to UserTurns.
+        assert_eq!(tags(&evs), vec!["user_turn", "user_turn"]);
+        match &evs[1].kind {
+            EventKind::UserTurn { text, .. } => assert_eq!(text, "world"),
+            other => panic!("expected user_turn, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn execute_tool_emits_call_result_and_edit() {
+        let line = r#"{"gen_ai.operation.name":"execute_tool","gen_ai.conversation.id":"s","gen_ai.tool.name":"edit_file","gen_ai.tool.call.id":"c1","file.path":"a.rs","model_added_lines":3,"model_removed_lines":1,"gen_ai.tool.result":"ok"}"#;
+        let (evs, _) = parse_all(&[line]);
+        assert_eq!(tags(&evs), vec!["tool_call", "tool_result", "file_edit"]);
+        match &evs[1].kind {
+            EventKind::ToolResult { ok, call_id, .. } => {
+                assert!(*ok);
+                assert_eq!(call_id, "c1");
+            }
+            other => panic!("expected tool_result, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn failed_tool_result_is_marked_not_ok() {
+        // tool_failure scenario: the ToolResult must be ok=false so downstream
+        // produces no spurious Episode.
+        let line = r#"{"gen_ai.operation.name":"execute_tool","gen_ai.conversation.id":"s","gen_ai.tool.name":"edit_file","gen_ai.tool.call.id":"cf","file.path":"a.rs","model_added_lines":3,"model_removed_lines":1,"gen_ai.tool.result.status":"error","error.type":"PatchConflict","gen_ai.tool.result":"hunk failed"}"#;
+        let (evs, _) = parse_all(&[line]);
+        assert_eq!(tags(&evs), vec!["tool_call", "tool_result", "file_edit"]);
+        match &evs[1].kind {
+            EventKind::ToolResult { ok, .. } => assert!(!*ok, "failed tool must be ok=false"),
+            other => panic!("expected tool_result, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn otlp_nested_shape_is_supported() {
+        // The same edit, expressed in the OTLP attributes-array shape.
+        let line = r#"{"timeUnixNano":"1750586400000000000","attributes":[{"key":"gen_ai.operation.name","value":{"stringValue":"file_operation"}},{"key":"gen_ai.conversation.id","value":{"stringValue":"nested-1"}},{"key":"file.path","value":{"stringValue":"src/main.rs"}},{"key":"code.added_lines","value":{"intValue":"7"}},{"key":"code.removed_lines","value":{"intValue":"2"}}]}"#;
+        let (evs, ctx) = parse_all(&[line]);
+        assert_eq!(tags(&evs), vec!["file_edit"]);
+        assert_eq!(ctx.session_id.as_deref(), Some("nested-1"));
+        match &evs[0].kind {
+            EventKind::FileEdit { diff, .. } => {
+                assert_eq!(diff.path, PathBuf::from("src/main.rs"));
+                assert_eq!(diff.added_lines, 7);
+                assert_eq!(diff.removed_lines, 2);
+            }
+            other => panic!("expected file_edit, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn unrecognized_record_routes_to_unknown() {
+        let line = r#"{"gen_ai.operation.name":"telemetry.heartbeat","gen_ai.conversation.id":"s","foo":"bar"}"#;
+        let (evs, _) = parse_all(&[line]);
+        assert_eq!(tags(&evs), vec!["unknown"]);
+    }
+
+    #[test]
+    fn garbage_input_never_panics_and_stays_lossless() {
+        let adapter = OtelAdapter;
+        let mut ctx = ParseCtx::new();
+        // Non-JSON garbage → one lossless Unknown.
+        let evs = adapter
+            .parse(&raw("}{ not json at all <<<"), &mut ctx)
+            .expect("never errors");
+        assert_eq!(tags(&evs), vec!["unknown"]);
+        // Blank line → nothing.
+        let evs = adapter.parse(&raw("   "), &mut ctx).expect("never errors");
+        assert!(evs.is_empty());
+        // Truncated / weird JSON values must not panic.
+        for g in [
+            "null",
+            "[]",
+            "123",
+            "\"a string\"",
+            r#"{"attributes":"not-an-array"}"#,
+            r#"{"gen_ai.input.messages":42}"#,
+            r#"{"gen_ai.output.messages":[{}]}"#,
+        ] {
+            let _ = adapter.parse(&raw(g), &mut ctx).expect("never errors");
+        }
+    }
+
+    #[test]
+    fn repeated_record_dedups_to_empty() {
+        let adapter = OtelAdapter;
+        let mut ctx = ParseCtx::new();
+        let first = adapter.parse(&raw(FILE_EDIT), &mut ctx).expect("ok");
+        assert_eq!(tags(&first), vec!["file_edit"]);
+        // The very same record again → idempotent, emits nothing.
+        let second = adapter.parse(&raw(FILE_EDIT), &mut ctx).expect("ok");
+        assert!(second.is_empty(), "repeat must dedup to empty");
+    }
+
+    #[test]
+    fn ban_prompt_is_carried_as_user_turn() {
+        let line = r#"{"gen_ai.operation.name":"chat","gen_ai.conversation.id":"s","gen_ai.cli.user_prompt":"We will never add a dependency on left-pad."}"#;
+        let (evs, _) = parse_all(&[line]);
+        assert_eq!(tags(&evs), vec!["user_turn"]);
+        match &evs[0].kind {
+            EventKind::UserTurn { text, .. } => assert!(text.contains("never add a dependency")),
+            other => panic!("expected user_turn, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn fingerprint_distinguishes_flat_and_otlp() {
+        let adapter = OtelAdapter;
+        let flat = adapter.schema_fingerprint(&raw(USER_PROMPT));
+        assert_eq!(flat.variant, "otel/genai-flat");
+        assert_eq!(flat.confidence, 100);
+        let nested = raw(r#"{"attributes":[{"key":"gen_ai.system","value":{"stringValue":"x"}}]}"#);
+        let otlp = adapter.schema_fingerprint(&nested);
+        assert_eq!(otlp.variant, "otel/genai-otlp");
+    }
+}
diff --git a/crates/memscribe-adapters/src/registry.rs b/crates/memscribe-adapters/src/registry.rs
new file mode 100644
index 0000000..d6fb08b
--- /dev/null
+++ b/crates/memscribe-adapters/src/registry.rs
@@ -0,0 +1,56 @@
+//! The adapter registry: assemble the set of enabled adapters and resolve one
+//! by [`SourceKind`].
+
+use memscribe_core::{SourceKind, TranscriptAdapter};
+
+/// Every enabled adapter, in a stable order.
+#[must_use]
+#[allow(clippy::vec_init_then_push)] // pushes are cfg-gated; a vec! literal won't work
+pub fn all_adapters() -> Vec<Box<dyn TranscriptAdapter>> {
+    let mut v: Vec<Box<dyn TranscriptAdapter>> = Vec::new();
+    #[cfg(feature = "claude_code")]
+    v.push(Box::new(crate::claude_code::ClaudeCodeAdapter));
+    #[cfg(feature = "codex")]
+    v.push(Box::new(crate::codex::CodexAdapter));
+    #[cfg(feature = "gemini")]
+    v.push(Box::new(crate::gemini::GeminiAdapter));
+    #[cfg(feature = "otel")]
+    v.push(Box::new(crate::otel::OtelAdapter));
+    #[cfg(feature = "cursor")]
+    v.push(Box::new(crate::cursor::CursorAdapter));
+    #[cfg(feature = "windsurf")]
+    v.push(Box::new(crate::windsurf::WindsurfAdapter));
+    #[cfg(feature = "zed")]
+    v.push(Box::new(crate::zed::ZedAdapter));
+    #[cfg(feature = "vscode")]
+    v.push(Box::new(crate::vscode::VsCodeAdapter));
+    #[cfg(feature = "copilot")]
+    v.push(Box::new(crate::copilot::CopilotAdapter));
+    v
+}
+
+/// Resolve the adapter for a given source, if its feature is enabled.
+#[must_use]
+pub fn adapter_for(kind: SourceKind) -> Option<Box<dyn TranscriptAdapter>> {
+    match kind {
+        #[cfg(feature = "claude_code")]
+        SourceKind::ClaudeCode => Some(Box::new(crate::claude_code::ClaudeCodeAdapter)),
+        #[cfg(feature = "codex")]
+        SourceKind::Codex => Some(Box::new(crate::codex::CodexAdapter)),
+        #[cfg(feature = "gemini")]
+        SourceKind::Gemini => Some(Box::new(crate::gemini::GeminiAdapter)),
+        #[cfg(feature = "otel")]
+        SourceKind::Otel => Some(Box::new(crate::otel::OtelAdapter)),
+        #[cfg(feature = "cursor")]
+        SourceKind::Cursor => Some(Box::new(crate::cursor::CursorAdapter)),
+        #[cfg(feature = "windsurf")]
+        SourceKind::Windsurf => Some(Box::new(crate::windsurf::WindsurfAdapter)),
+        #[cfg(feature = "zed")]
+        SourceKind::Zed => Some(Box::new(crate::zed::ZedAdapter)),
+        #[cfg(feature = "vscode")]
+        SourceKind::VsCode => Some(Box::new(crate::vscode::VsCodeAdapter)),
+        #[cfg(feature = "copilot")]
+        SourceKind::Copilot => Some(Box::new(crate::copilot::CopilotAdapter)),
+        _ => None,
+    }
+}
diff --git a/crates/memscribe-adapters/src/util.rs b/crates/memscribe-adapters/src/util.rs
new file mode 100644
index 0000000..322fd3c
--- /dev/null
+++ b/crates/memscribe-adapters/src/util.rs
@@ -0,0 +1,186 @@
+//! Shared, deterministic helpers for adapters. **Do not add tool-specific logic
+//! here** — keep that in each tool's module so adapters can be maintained
+//! independently. These helpers build normalized events with correct
+//! provenance, sequencing, and ids.
+
+use memscribe_core::{
+    content_id, CaptureEvent, EventKind, ParseCtx, ParseError, RawRecord, SourceKind, Timestamp,
+    SCHEMA_VERSION,
+};
+use time::format_description::well_known::Rfc3339;
+use time::OffsetDateTime;
+
+/// Parse a JSONL line into a JSON value. Returns `None` for blank lines or
+/// invalid JSON (the caller decides whether that is an `Unknown` or a skip).
+#[must_use]
+pub fn parse_json_line(raw: &RawRecord) -> Option<serde_json::Value> {
+    let s = raw.as_str()?.trim();
+    if s.is_empty() {
+        return None;
+    }
+    serde_json::from_str(s).ok()
+}
+
+/// Parse a timestamp from RFC3339, or from epoch seconds/milliseconds. Returns
+/// `None` if neither parses.
+#[must_use]
+pub fn parse_ts(s: &str) -> Option<Timestamp> {
+    if let Ok(t) = OffsetDateTime::parse(s.trim(), &Rfc3339) {
+        return Some(t);
+    }
+    let n: i64 = s.trim().parse().ok()?;
+    // Heuristic: values above ~year 2286 in seconds are really milliseconds.
+    let (secs, nanos) = if n.abs() > 10_000_000_000 {
+        (n / 1000, (n % 1000) * 1_000_000)
+    } else {
+        (n, 0)
+    };
+    OffsetDateTime::from_unix_timestamp(secs)
+        .ok()
+        .map(|t| t + time::Duration::nanoseconds(nanos))
+}
+
+/// Extract an RFC3339/epoch timestamp from a JSON object under any of the given
+/// keys, falling back to the Unix epoch (so output stays deterministic even when
+/// a record carries no timestamp).
+#[must_use]
+pub fn ts_from(value: &serde_json::Value, keys: &[&str]) -> Timestamp {
+    for k in keys {
+        if let Some(v) = value.get(*k) {
+            if let Some(s) = v.as_str() {
+                if let Some(t) = parse_ts(s) {
+                    return t;
+                }
+            } else if let Some(n) = v.as_i64() {
+                if let Some(t) = parse_ts(&n.to_string()) {
+                    return t;
+                }
+            }
+        }
+    }
+    OffsetDateTime::UNIX_EPOCH
+}
+
+/// The Unix epoch — a stable default timestamp.
+#[must_use]
+pub fn epoch() -> Timestamp {
+    OffsetDateTime::UNIX_EPOCH
+}
+
+/// Build a normalized [`CaptureEvent`], allocating the monotonic `seq` from the
+/// context and stamping the session/project binding.
+#[allow(clippy::too_many_arguments)]
+#[must_use]
+pub fn mk_event(
+    source: SourceKind,
+    ctx: &mut ParseCtx,
+    raw: &RawRecord,
+    event_id: String,
+    parent_id: Option<String>,
+    timestamp: Timestamp,
+    kind: EventKind,
+) -> CaptureEvent {
+    let seq = ctx.alloc_seq();
+    let session_id = ctx
+        .session_id
+        .clone()
+        .unwrap_or_else(|| "unknown".to_string());
+    CaptureEvent {
+        schema_version: SCHEMA_VERSION,
+        source,
+        session_id,
+        seq,
+        event_id,
+        parent_id,
+        timestamp,
+        project: ctx.project_or_default(),
+        kind,
+        provenance: raw.location.clone(),
+    }
+}
+
+/// Build an [`EventKind::Unknown`] event from a raw JSON value — the lossless,
+/// version-tolerant fallback every adapter uses for records it does not yet
+/// understand.
+#[must_use]
+pub fn unknown_event(
+    source: SourceKind,
+    ctx: &mut ParseCtx,
+    raw: &RawRecord,
+    value: serde_json::Value,
+) -> CaptureEvent {
+    let raw_type = value
+        .get("type")
+        .and_then(|v| v.as_str())
+        .map(str::to_string)
+        .unwrap_or_else(|| "unknown".to_string());
+    let timestamp = ts_from(&value, &["timestamp", "time", "ts", "created_at"]);
+    let id = content_id(&raw.bytes);
+    mk_event(
+        source,
+        ctx,
+        raw,
+        id,
+        None,
+        timestamp,
+        EventKind::Unknown {
+            raw_type,
+            raw: value,
+        },
+    )
+}
+
+/// The default skeleton parse: emit exactly one `Unknown` event per non-blank
+/// record (so the stream is lossless even before a real parser exists). Tool
+/// modules replace this with real parsing but should preserve the losslessness
+/// guarantee for records they do not recognize.
+///
+/// # Errors
+/// Never returns an error — present for signature parity with `parse`.
+pub fn stub_parse(
+    source: SourceKind,
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+) -> Result<Vec<CaptureEvent>, ParseError> {
+    let s = raw.as_str().map(str::trim).unwrap_or("");
+    if s.is_empty() {
+        return Ok(Vec::new());
+    }
+    let value =
+        serde_json::from_str(s).unwrap_or_else(|_| serde_json::Value::String(s.to_string()));
+    Ok(vec![unknown_event(source, ctx, raw, value)])
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use memscribe_core::SourceLocation;
+
+    fn raw(s: &str) -> RawRecord {
+        RawRecord::from_line(s, SourceLocation::new("t.jsonl", 0, 1))
+    }
+
+    #[test]
+    fn parse_ts_rfc3339_and_epoch() {
+        assert!(parse_ts("2026-06-22T10:00:00Z").is_some());
+        assert!(parse_ts("1750000000").is_some());
+        assert!(parse_ts("1750000000000").is_some());
+        assert!(parse_ts("not a time").is_none());
+    }
+
+    #[test]
+    fn stub_parse_is_lossless_for_nonblank() {
+        let mut ctx = ParseCtx::new();
+        let evs = stub_parse(SourceKind::Unknown, &raw("{\"type\":\"x\"}"), &mut ctx).unwrap();
+        assert_eq!(evs.len(), 1);
+        assert_eq!(evs[0].kind.tag(), "unknown");
+    }
+
+    #[test]
+    fn stub_parse_skips_blank() {
+        let mut ctx = ParseCtx::new();
+        assert!(stub_parse(SourceKind::Unknown, &raw("   "), &mut ctx)
+            .unwrap()
+            .is_empty());
+    }
+}
diff --git a/crates/memscribe-adapters/src/vscode.rs b/crates/memscribe-adapters/src/vscode.rs
new file mode 100644
index 0000000..e200612
--- /dev/null
+++ b/crates/memscribe-adapters/src/vscode.rs
@@ -0,0 +1,796 @@
+//! VS Code adapter (Copilot Chat / chat sessions).
+//!
+//! VS Code stores chat sessions under
+//! `<user>/workspaceStorage/<hash>/chatSessions/*.json` (and
+//! `chatEditingSessions` for edits). That on-disk store is an undocumented,
+//! version-churning JSON blob, so this adapter parses two shapes:
+//!
+//! 1. A stable, **exported** chat JSON-lines shape (one record per line) that a
+//!    companion exporter writes — a leading `{kind:session_start, cwd, git,
+//!    toolVersion}` followed by message records `{id, parentId, role, ts,
+//!    sessionId, text, model, usage, toolCalls, toolResults, edits}`.
+//! 2. The **native** `chatSessions` JSON shape, where a single object carries
+//!    `{version, requesterUsername, responderUsername, requests:[{message,
+//!    response}]}`; each request maps to a `UserTurn` and its response to an
+//!    `AssistantTurn`.
+//!
+//! Anything unrecognized-but-valid routes to [`memscribe_core::EventKind::Unknown`]
+//! via [`util::unknown_event`], so the stream stays lossless across VS Code
+//! version churn. The parser is fully deterministic and never panics.
+
+use crate::util;
+use memscribe_core::{
+    content_id, CaptureEvent, Diff, DiscoverCfg, EventKind, GitRef, ParseCtx, ParseError,
+    ProjectRef, RawRecord, SchemaVariant, SourceKind, TranscriptAdapter, TranscriptHandle, Usage,
+};
+use serde_json::Value;
+use std::path::PathBuf;
+
+const SOURCE: SourceKind = SourceKind::VsCode;
+
+/// Adapter for VS Code chat-session transcripts.
+#[derive(Debug, Default, Clone, Copy)]
+pub struct VsCodeAdapter;
+
+impl TranscriptAdapter for VsCodeAdapter {
+    fn source_kind(&self) -> SourceKind {
+        SOURCE
+    }
+
+    fn discover(&self, cfg: &DiscoverCfg) -> Vec<TranscriptHandle> {
+        // Point at the real product path; we don't parse the binary store here.
+        // `Application Support/Code/User/workspaceStorage/<hash>/chatSessions/*.json`
+        let home = cfg.home_dir();
+        let base = home
+            .join("Library")
+            .join("Application Support")
+            .join("Code")
+            .join("User")
+            .join("workspaceStorage");
+        let mut handles = Vec::new();
+        // Walk workspaceStorage/<hash>/chatSessions/*.json deterministically.
+        let mut hashes: Vec<PathBuf> = Vec::new();
+        if let Ok(entries) = std::fs::read_dir(&base) {
+            for entry in entries.flatten() {
+                let p = entry.path();
+                if p.is_dir() {
+                    hashes.push(p);
+                }
+            }
+        }
+        hashes.sort();
+        for ws in hashes {
+            let sessions_dir = ws.join("chatSessions");
+            let mut files: Vec<PathBuf> = Vec::new();
+            if let Ok(entries) = std::fs::read_dir(&sessions_dir) {
+                for entry in entries.flatten() {
+                    let p = entry.path();
+                    if p.extension().and_then(|e| e.to_str()) == Some("json") {
+                        files.push(p);
+                    }
+                }
+            }
+            files.sort();
+            for f in files {
+                let session_hint = f.file_stem().and_then(|s| s.to_str()).map(str::to_string);
+                handles.push(TranscriptHandle {
+                    path: f,
+                    source: SOURCE,
+                    session_hint,
+                    compressed: false,
+                });
+            }
+        }
+        handles
+    }
+
+    fn parse(&self, raw: &RawRecord, ctx: &mut ParseCtx) -> Result<Vec<CaptureEvent>, ParseError> {
+        let Some(value) = util::parse_json_line(raw) else {
+            // Blank line → nothing; non-JSON but non-blank → lossless Unknown.
+            let s = raw.as_str().map(str::trim).unwrap_or("");
+            if s.is_empty() {
+                return Ok(Vec::new());
+            }
+            return Ok(vec![util::unknown_event(
+                SOURCE,
+                ctx,
+                raw,
+                Value::String(s.to_string()),
+            )]);
+        };
+
+        // The native chatSessions shape: a single object with a `requests` array.
+        if value.get("requests").and_then(Value::as_array).is_some() {
+            return Ok(parse_native_session(raw, ctx, &value));
+        }
+
+        // Otherwise treat it as one exported JSON-lines record.
+        Ok(parse_exported_record(raw, ctx, value))
+    }
+
+    fn schema_fingerprint(&self, sample: &RawRecord) -> SchemaVariant {
+        let Some(value) = util::parse_json_line(sample) else {
+            return SchemaVariant::unknown(SOURCE);
+        };
+        if value.get("requests").and_then(Value::as_array).is_some() {
+            return SchemaVariant::certain(SOURCE, "vscode/chat-sessions-native");
+        }
+        if value.get("kind").and_then(Value::as_str) == Some("session_start")
+            || value.get("role").and_then(Value::as_str).is_some()
+        {
+            return SchemaVariant::certain(SOURCE, "vscode/exported-jsonl-v1");
+        }
+        SchemaVariant::unknown(SOURCE)
+    }
+}
+
+/// Parse one record of the exported JSON-lines shape into zero or more events.
+fn parse_exported_record(raw: &RawRecord, ctx: &mut ParseCtx, value: Value) -> Vec<CaptureEvent> {
+    let kind = value.get("kind").and_then(Value::as_str);
+
+    // Record-level dedup / idempotency: a repeated record yields nothing.
+    let record_id = record_event_id(raw, &value);
+    if !ctx.first_seen(&record_id) {
+        return Vec::new();
+    }
+
+    match kind {
+        Some("session_start") => {
+            apply_session_start(ctx, &value);
+            let ts = util::ts_from(&value, &["ts", "timestamp", "time"]);
+            let git = parse_git(value.get("git"));
+            let cwd = string_field(&value, "cwd").unwrap_or_else(|| ".".to_string());
+            let model = string_field(&value, "model");
+            let tool_version = string_field(&value, "toolVersion");
+            vec![util::mk_event(
+                SOURCE,
+                ctx,
+                raw,
+                record_id,
+                None,
+                ts,
+                EventKind::SessionStart {
+                    cwd: PathBuf::from(cwd),
+                    git,
+                    model,
+                    tool_version,
+                },
+            )]
+        }
+        Some("session_end") => {
+            adopt_session(ctx, &value);
+            let ts = util::ts_from(&value, &["ts", "timestamp", "time"]);
+            let reason = string_field(&value, "reason");
+            vec![util::mk_event(
+                SOURCE,
+                ctx,
+                raw,
+                record_id,
+                None,
+                ts,
+                EventKind::SessionEnd { reason },
+            )]
+        }
+        _ => parse_message_record(raw, ctx, &value, record_id),
+    }
+}
+
+/// Parse one message record (`role: user|assistant`, with optional toolCalls,
+/// toolResults, and edits) into an ordered list of events.
+fn parse_message_record(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    value: &Value,
+    record_id: String,
+) -> Vec<CaptureEvent> {
+    adopt_session(ctx, value);
+    let ts = util::ts_from(value, &["ts", "timestamp", "time"]);
+    let parent_id = string_field(value, "parentId");
+    let role = value.get("role").and_then(Value::as_str);
+    let text = string_field(value, "text").unwrap_or_default();
+
+    let mut out = Vec::new();
+
+    match role {
+        Some("user") => out.push(util::mk_event(
+            SOURCE,
+            ctx,
+            raw,
+            record_id.clone(),
+            parent_id.clone(),
+            ts,
+            EventKind::UserTurn {
+                text,
+                parts: Vec::new(),
+            },
+        )),
+        Some("assistant") => out.push(util::mk_event(
+            SOURCE,
+            ctx,
+            raw,
+            record_id.clone(),
+            parent_id.clone(),
+            ts,
+            EventKind::AssistantTurn {
+                text,
+                thinking: None,
+                model: string_field(value, "model"),
+                usage: parse_usage(value.get("usage")),
+                parts: Vec::new(),
+            },
+        )),
+        // A record with edits/tool data but no recognized role is still
+        // valuable; if it carries no actionable role and nothing else, fall
+        // through to the sub-records below and, if none, emit Unknown.
+        _ => {
+            let has_children = value
+                .get("toolCalls")
+                .and_then(Value::as_array)
+                .is_some_and(|a| !a.is_empty())
+                || value
+                    .get("toolResults")
+                    .and_then(Value::as_array)
+                    .is_some_and(|a| !a.is_empty())
+                || value
+                    .get("edits")
+                    .and_then(Value::as_array)
+                    .is_some_and(|a| !a.is_empty());
+            if !has_children {
+                return vec![util::unknown_event(SOURCE, ctx, raw, value.clone())];
+            }
+        }
+    }
+
+    // Tool calls.
+    if let Some(calls) = value.get("toolCalls").and_then(Value::as_array) {
+        for (i, call) in calls.iter().enumerate() {
+            let call_id =
+                string_field(call, "id").unwrap_or_else(|| format!("{record_id}:call:{i}"));
+            let name = string_field(call, "name").unwrap_or_default();
+            let args = call.get("args").cloned().unwrap_or(Value::Null);
+            ctx.call_names.insert(call_id.clone(), name.clone());
+            out.push(util::mk_event(
+                SOURCE,
+                ctx,
+                raw,
+                derive_id(&record_id, "toolcall", i, &call_id),
+                Some(record_id.clone()),
+                ts,
+                EventKind::ToolCall {
+                    call_id,
+                    name,
+                    args,
+                },
+            ));
+        }
+    }
+
+    // Tool results.
+    if let Some(results) = value.get("toolResults").and_then(Value::as_array) {
+        for (i, res) in results.iter().enumerate() {
+            let call_id =
+                string_field(res, "id").unwrap_or_else(|| format!("{record_id}:result:{i}"));
+            let ok = res.get("ok").and_then(Value::as_bool).unwrap_or(true);
+            let output = res.get("output").cloned().unwrap_or(Value::Null);
+            ctx.call_ok.insert(call_id.clone(), ok);
+            out.push(util::mk_event(
+                SOURCE,
+                ctx,
+                raw,
+                derive_id(&record_id, "toolresult", i, &call_id),
+                Some(record_id.clone()),
+                ts,
+                EventKind::ToolResult {
+                    call_id,
+                    ok,
+                    output,
+                },
+            ));
+        }
+    }
+
+    // File edits.
+    if let Some(edits) = value.get("edits").and_then(Value::as_array) {
+        for (i, edit) in edits.iter().enumerate() {
+            let diff = parse_edit(edit);
+            let edit_id = string_field(edit, "id")
+                .or_else(|| diff.path.to_str().map(str::to_string))
+                .unwrap_or_default();
+            // Link the edit to its originating tool call when the export carries
+            // one (`callId`/`call_id`). This is what lets the segmenter drop an
+            // edit whose paired ToolResult failed (ok=false) — "a tool failure →
+            // no spurious episode" (§8.2). Absent the field, the edit stands on
+            // its own (the happy path), matching the prior behavior.
+            let call_id = string_field(edit, "callId").or_else(|| string_field(edit, "call_id"));
+            out.push(util::mk_event(
+                SOURCE,
+                ctx,
+                raw,
+                derive_id(&record_id, "edit", i, &edit_id),
+                Some(record_id.clone()),
+                ts,
+                EventKind::FileEdit { call_id, diff },
+            ));
+        }
+    }
+
+    out
+}
+
+/// Parse the native `chatSessions` shape: `requests[].message` → `UserTurn`,
+/// `requests[].response[]` → `AssistantTurn`.
+fn parse_native_session(raw: &RawRecord, ctx: &mut ParseCtx, value: &Value) -> Vec<CaptureEvent> {
+    // Record-level idempotency on the whole session object.
+    let record_id = record_event_id(raw, value);
+    if !ctx.first_seen(&record_id) {
+        return Vec::new();
+    }
+    adopt_session(ctx, value);
+    if ctx.session_id.is_none() {
+        // Native files have no `sessionId`; derive a stable one from content.
+        ctx.session_id = Some(format!("vscode-{}", &record_id[..record_id.len().min(16)]));
+    }
+
+    let ts = util::ts_from(value, &["ts", "timestamp", "time"]);
+    let responder = string_field(value, "responderUsername");
+    let mut out = Vec::new();
+
+    let Some(requests) = value.get("requests").and_then(Value::as_array) else {
+        return vec![util::unknown_event(SOURCE, ctx, raw, value.clone())];
+    };
+
+    for (i, req) in requests.iter().enumerate() {
+        // User turn from `message`.
+        let user_text = req
+            .get("message")
+            .map(flatten_native_text)
+            .unwrap_or_default();
+        out.push(util::mk_event(
+            SOURCE,
+            ctx,
+            raw,
+            format!("{record_id}:req:{i}:user"),
+            None,
+            ts,
+            EventKind::UserTurn {
+                text: user_text,
+                parts: Vec::new(),
+            },
+        ));
+
+        // Assistant turn from `response` (an array of parts).
+        let resp_text = flatten_native_response(req.get("response"));
+        out.push(util::mk_event(
+            SOURCE,
+            ctx,
+            raw,
+            format!("{record_id}:req:{i}:asst"),
+            Some(format!("{record_id}:req:{i}:user")),
+            ts,
+            EventKind::AssistantTurn {
+                text: resp_text,
+                thinking: None,
+                model: responder.clone(),
+                usage: None,
+                parts: Vec::new(),
+            },
+        ));
+    }
+
+    out
+}
+
+// ---------------------------------------------------------------------------
+// helpers
+// ---------------------------------------------------------------------------
+
+/// The deterministic record-level event id: the native `id` field, else a
+/// `blake3` of the raw bytes (per the format spec: `event_id = id else content_id`).
+fn record_event_id(raw: &RawRecord, value: &Value) -> String {
+    string_field(value, "id").unwrap_or_else(|| content_id(&raw.bytes))
+}
+
+/// Derive a stable, collision-free child event id under a parent record id.
+fn derive_id(record_id: &str, kind: &str, index: usize, native: &str) -> String {
+    format!("{record_id}:{kind}:{index}:{native}")
+}
+
+/// Read a string field, ignoring empty/non-string values.
+fn string_field(value: &Value, key: &str) -> Option<String> {
+    value
+        .get(key)
+        .and_then(Value::as_str)
+        .map(str::to_string)
+        .filter(|s| !s.is_empty())
+}
+
+/// Set `ctx.session_id` from `sessionId` if not already set.
+fn adopt_session(ctx: &mut ParseCtx, value: &Value) {
+    if ctx.session_id.is_none() {
+        if let Some(sid) = string_field(value, "sessionId") {
+            ctx.session_id = Some(sid);
+        }
+    }
+}
+
+/// Apply a session-start record to the context: session id + project binding.
+fn apply_session_start(ctx: &mut ParseCtx, value: &Value) {
+    adopt_session(ctx, value);
+    let cwd = string_field(value, "cwd").unwrap_or_else(|| ".".to_string());
+    let git = parse_git(value.get("git"));
+    ctx.project = Some(ProjectRef {
+        cwd: PathBuf::from(cwd),
+        repo_root: None,
+        git,
+    });
+}
+
+/// Parse a `{sha, branch}` git object.
+fn parse_git(value: Option<&Value>) -> Option<GitRef> {
+    let obj = value?;
+    let sha = string_field(obj, "sha")?;
+    let branch = string_field(obj, "branch");
+    Some(GitRef { sha, branch })
+}
+
+/// Parse a `{input, output}` usage object into [`Usage`].
+fn parse_usage(value: Option<&Value>) -> Option<Usage> {
+    let obj = value?;
+    let input_tokens = obj.get("input").and_then(Value::as_u64);
+    let output_tokens = obj.get("output").and_then(Value::as_u64);
+    if input_tokens.is_none() && output_tokens.is_none() {
+        return None;
+    }
+    Some(Usage {
+        input_tokens,
+        output_tokens,
+        cache_read_tokens: None,
+        cache_creation_tokens: None,
+    })
+}
+
+/// Parse an `{path, oldText, newText, diff, added, removed}` edit into a [`Diff`].
+fn parse_edit(edit: &Value) -> Diff {
+    let path = string_field(edit, "path").unwrap_or_default();
+    Diff {
+        path: PathBuf::from(path),
+        old: string_field(edit, "oldText"),
+        new: string_field(edit, "newText"),
+        unified: string_field(edit, "diff"),
+        added_lines: edit
+            .get("added")
+            .and_then(Value::as_u64)
+            .unwrap_or(0)
+            .min(u64::from(u32::MAX)) as u32,
+        removed_lines: edit
+            .get("removed")
+            .and_then(Value::as_u64)
+            .unwrap_or(0)
+            .min(u64::from(u32::MAX)) as u32,
+    }
+}
+
+/// Flatten a native `message` object (`{text, parts:[{kind:text,text}]}`).
+fn flatten_native_text(message: &Value) -> String {
+    if let Some(t) = string_field(message, "text") {
+        return t;
+    }
+    flatten_text_parts(message.get("parts"))
+}
+
+/// Flatten a native `response` (an array of `{kind:text,text}` parts).
+fn flatten_native_response(response: Option<&Value>) -> String {
+    match response {
+        Some(Value::Array(_)) => flatten_text_parts(response),
+        Some(Value::String(s)) => s.clone(),
+        Some(obj @ Value::Object(_)) => flatten_native_text(obj),
+        _ => String::new(),
+    }
+}
+
+/// Concatenate the `text` of every `{kind:"text", text}` part in an array.
+fn flatten_text_parts(parts: Option<&Value>) -> String {
+    let Some(arr) = parts.and_then(Value::as_array) else {
+        return String::new();
+    };
+    let mut chunks: Vec<String> = Vec::new();
+    for part in arr {
+        if part.get("kind").and_then(Value::as_str) == Some("text") {
+            if let Some(t) = string_field(part, "text") {
+                chunks.push(t);
+            }
+        }
+    }
+    chunks.join("")
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use memscribe_core::SourceLocation;
+
+    fn raw(line: &str) -> RawRecord {
+        RawRecord::from_line(line, SourceLocation::new("vscode.jsonl", 0, 1))
+    }
+
+    /// Run a slice of JSONL lines through the adapter, threading one context.
+    fn run(lines: &[&str]) -> Vec<CaptureEvent> {
+        let adapter = VsCodeAdapter;
+        let mut ctx = ParseCtx::new();
+        let mut out = Vec::new();
+        for line in lines {
+            let evs = adapter.parse(&raw(line), &mut ctx).expect("never errors");
+            out.extend(evs);
+        }
+        out
+    }
+
+    fn tags(events: &[CaptureEvent]) -> Vec<&'static str> {
+        events.iter().map(|e| e.kind.tag()).collect()
+    }
+
+    const SESSION_START: &str = r#"{"kind":"session_start","sessionId":"s1","cwd":"/work","git":{"sha":"abc","branch":"main"},"toolVersion":"1.92.0","model":"gpt-4o"}"#;
+
+    #[test]
+    fn session_start_sets_session_and_project() {
+        let evs = run(&[SESSION_START]);
+        assert_eq!(tags(&evs), vec!["session_start"]);
+        assert_eq!(evs[0].session_id, "s1");
+        assert_eq!(evs[0].project.cwd, PathBuf::from("/work"));
+        let git = evs[0].project.git.as_ref().expect("git set from start");
+        assert_eq!(git.sha, "abc");
+        assert_eq!(git.branch.as_deref(), Some("main"));
+        match &evs[0].kind {
+            EventKind::SessionStart {
+                model,
+                tool_version,
+                ..
+            } => {
+                assert_eq!(model.as_deref(), Some("gpt-4o"));
+                assert_eq!(tool_version.as_deref(), Some("1.92.0"));
+            }
+            other => panic!("expected session_start, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn normalized_sequence_decision_then_edits() {
+        let user = r#"{"id":"m1","role":"user","ts":"2026-06-22T10:00:00Z","sessionId":"s1","text":"Let's use Postgres instead of MySQL."}"#;
+        let asst = r#"{"id":"m2","parentId":"m1","role":"assistant","ts":"2026-06-22T10:00:05Z","sessionId":"s1","text":"Switching now.","model":"gpt-4o","usage":{"input":10,"output":3},"edits":[{"path":"src/db.ts","oldText":"mysql","newText":"postgres","added":1,"removed":1}]}"#;
+        let evs = run(&[SESSION_START, user, asst]);
+        // session_start, user_turn, assistant_turn, file_edit
+        assert_eq!(
+            tags(&evs),
+            vec!["session_start", "user_turn", "assistant_turn", "file_edit"]
+        );
+        // The decision turn is a UserTurn carrying the decision text.
+        match &evs[1].kind {
+            EventKind::UserTurn { text, .. } => {
+                assert!(text.contains("Postgres"));
+            }
+            other => panic!("expected user_turn, got {other:?}"),
+        }
+        // The edit is a FileEdit with the diff fields mapped.
+        match &evs[3].kind {
+            EventKind::FileEdit { diff, .. } => {
+                assert_eq!(diff.path, PathBuf::from("src/db.ts"));
+                assert_eq!(diff.old.as_deref(), Some("mysql"));
+                assert_eq!(diff.new.as_deref(), Some("postgres"));
+                assert_eq!(diff.added_lines, 1);
+                assert_eq!(diff.removed_lines, 1);
+            }
+            other => panic!("expected file_edit, got {other:?}"),
+        }
+        // The edit's parent links back to the assistant turn record id.
+        assert_eq!(evs[3].parent_id.as_deref(), Some("m2"));
+    }
+
+    #[test]
+    fn tool_call_then_result_failure() {
+        let asst = r#"{"id":"t2","role":"assistant","ts":"2026-06-22T13:00:07Z","sessionId":"s4","text":"applying","toolCalls":[{"id":"c1","name":"applyEdit","args":{"path":"x"}}],"edits":[{"path":"x","callId":"c1","oldText":"a","newText":"b","added":1,"removed":1}]}"#;
+        let res = r#"{"id":"t3","role":"assistant","ts":"2026-06-22T13:00:09Z","sessionId":"s4","text":"","toolResults":[{"id":"c1","ok":false,"output":"FAILED"}]}"#;
+        let evs = run(&[asst, res]);
+        // assistant_turn, tool_call, file_edit, assistant_turn, tool_result
+        assert_eq!(
+            tags(&evs),
+            vec![
+                "assistant_turn",
+                "tool_call",
+                "file_edit",
+                "assistant_turn",
+                "tool_result"
+            ]
+        );
+        // The edit is linked to the failing call by call_id, so the segmenter
+        // drops it (no spurious episode for a failed edit).
+        match &evs[2].kind {
+            EventKind::FileEdit { call_id, .. } => assert_eq!(call_id.as_deref(), Some("c1")),
+            other => panic!("expected file_edit, got {other:?}"),
+        }
+        // The failed result must carry ok=false (so no Episode is produced
+        // downstream for the failed edit).
+        match &evs[4].kind {
+            EventKind::ToolResult { ok, call_id, .. } => {
+                assert!(!ok);
+                assert_eq!(call_id, "c1");
+            }
+            other => panic!("expected tool_result, got {other:?}"),
+        }
+    }
+
+    /// On-disk fixture conformance: the `tool_failure` fixture must, end-to-end
+    /// through the segmenter, mint NO episode; the happy path must still mint two.
+    fn vscode_fixture(name: &str) -> String {
+        let path = std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
+            .join("../../fixtures/vscode/v1")
+            .join(name);
+        std::fs::read_to_string(&path)
+            .unwrap_or_else(|e| panic!("read fixture {}: {e}", path.display()))
+    }
+
+    fn run_file(jsonl: &str) -> Vec<CaptureEvent> {
+        let adapter = VsCodeAdapter;
+        let mut ctx = ParseCtx::new();
+        let mut out = Vec::new();
+        for line in jsonl.lines() {
+            let evs = adapter.parse(&raw(line), &mut ctx).expect("never errors");
+            out.extend(evs);
+        }
+        out
+    }
+
+    #[test]
+    fn fixture_tool_failure_yields_no_episode_via_segmenter() {
+        use memscribe_core::gate::CommitmentGate;
+        use memscribe_core::segmenter::{DefaultSegmenter, Segmenter};
+
+        let gate = CommitmentGate::default();
+        let seg = DefaultSegmenter;
+
+        let fail_events = run_file(&vscode_fixture("tool_failure.jsonl"));
+        // The failed edit is linked to a failing ToolResult by call_id.
+        assert!(fail_events.iter().any(|e| matches!(
+            &e.kind,
+            EventKind::FileEdit { call_id, .. } if call_id.as_deref() == Some("call-edit-1")
+        )));
+        let fail_seg = seg.segment(&fail_events, &gate);
+        assert_eq!(
+            fail_seg.episodes.len(),
+            0,
+            "a failed edit must produce no episode"
+        );
+
+        let ok_events = run_file(&vscode_fixture("happy_path_decision_then_edits.jsonl"));
+        let ok_seg = seg.segment(&ok_events, &gate);
+        assert_eq!(
+            ok_seg.episodes.len(),
+            2,
+            "the happy path must still produce two episodes"
+        );
+    }
+
+    #[test]
+    fn never_panics_on_garbage() {
+        // Invalid JSON, empty, and structurally-weird-but-valid inputs.
+        let garbage = run(&[
+            "not json at all {{{",
+            "",
+            "   ",
+            "42",
+            "true",
+            r#"{"role":12345}"#,
+            r#"{"kind":"session_start"}"#,
+            r#"{"requests":"not-an-array"}"#,
+            r#"{"id":"x","edits":[{}]}"#,
+        ]);
+        // Nothing panicked; every non-blank record produced at least an event.
+        // Blank lines produce nothing, so the count is < the input count but > 0.
+        assert!(!garbage.is_empty());
+        // A non-JSON line is preserved as Unknown (lossless).
+        assert!(garbage.iter().any(|e| e.kind.tag() == "unknown"));
+    }
+
+    #[test]
+    fn unrecognized_valid_record_routes_to_unknown() {
+        let evs = run(&[r#"{"id":"weird","kind":"telemetry","payload":{"a":1}}"#]);
+        assert_eq!(tags(&evs), vec!["unknown"]);
+        match &evs[0].kind {
+            EventKind::Unknown { raw_type, raw } => {
+                assert_eq!(raw_type, "unknown");
+                assert!(raw.get("payload").is_some());
+            }
+            other => panic!("expected unknown, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn dedup_repeated_record_is_idempotent() {
+        let user =
+            r#"{"id":"m1","role":"user","ts":"2026-06-22T10:00:00Z","sessionId":"s1","text":"hi"}"#;
+        // Same record twice → only one event.
+        let evs = run(&[SESSION_START, user, user]);
+        assert_eq!(tags(&evs), vec!["session_start", "user_turn"]);
+        // Sequence numbers are still monotonic and gap-free for what was kept.
+        assert_eq!(evs[0].seq, 0);
+        assert_eq!(evs[1].seq, 1);
+    }
+
+    #[test]
+    fn idempotent_record_with_children_dedups_whole_record() {
+        let asst = r#"{"id":"m2","role":"assistant","ts":"2026-06-22T10:00:05Z","sessionId":"s1","text":"x","edits":[{"path":"a.ts","oldText":"1","newText":"2","added":1,"removed":0}]}"#;
+        let evs = run(&[asst, asst]);
+        // First time: assistant_turn + file_edit. Second time: nothing.
+        assert_eq!(tags(&evs), vec!["assistant_turn", "file_edit"]);
+    }
+
+    #[test]
+    fn child_event_ids_do_not_collide_with_turn() {
+        let asst = r#"{"id":"m9","role":"assistant","ts":"2026-06-22T10:00:05Z","sessionId":"s1","text":"x","toolCalls":[{"id":"m9","name":"t","args":{}}]}"#;
+        // The tool-call's native id collides with the record id; derivation must
+        // keep the events distinct so both survive.
+        let evs = run(&[asst]);
+        assert_eq!(tags(&evs), vec!["assistant_turn", "tool_call"]);
+        assert_ne!(evs[0].event_id, evs[1].event_id);
+    }
+
+    #[test]
+    fn epoch_ms_timestamp_parses() {
+        let user = r#"{"id":"m1","role":"user","ts":1750000000000,"sessionId":"s1","text":"hi"}"#;
+        let evs = run(&[user]);
+        assert_eq!(tags(&evs), vec!["user_turn"]);
+        // 1_750_000_000_000 ms = 2025-06-15ish — well after the epoch.
+        assert!(evs[0].timestamp.unix_timestamp() > 1_700_000_000);
+    }
+
+    #[test]
+    fn native_chatsession_shape_maps_requests() {
+        let native = r#"{"version":3,"requesterUsername":"dev","responderUsername":"Copilot","requests":[{"message":{"text":"Add a health check","parts":[{"kind":"text","text":"Add a health check"}]},"response":[{"kind":"text","text":"Adding GET /healthz."},{"kind":"text","text":" Done."}]}]}"#;
+        let evs = run(&[native]);
+        assert_eq!(tags(&evs), vec!["user_turn", "assistant_turn"]);
+        match &evs[0].kind {
+            EventKind::UserTurn { text, .. } => assert_eq!(text, "Add a health check"),
+            other => panic!("expected user_turn, got {other:?}"),
+        }
+        match &evs[1].kind {
+            EventKind::AssistantTurn { text, model, .. } => {
+                assert_eq!(text, "Adding GET /healthz. Done.");
+                assert_eq!(model.as_deref(), Some("Copilot"));
+            }
+            other => panic!("expected assistant_turn, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn schema_fingerprint_distinguishes_shapes() {
+        let adapter = VsCodeAdapter;
+        let exported = adapter.schema_fingerprint(&raw(SESSION_START));
+        assert_eq!(exported.variant, "vscode/exported-jsonl-v1");
+        assert_eq!(exported.confidence, 100);
+
+        let native = adapter.schema_fingerprint(&raw(r#"{"requests":[]}"#));
+        assert_eq!(native.variant, "vscode/chat-sessions-native");
+
+        let unknown = adapter.schema_fingerprint(&raw("garbage"));
+        assert_eq!(unknown.confidence, 0);
+    }
+
+    #[test]
+    fn session_id_falls_back_to_unknown_without_start() {
+        // A bare message with no sessionId and no prior session_start.
+        let user = r#"{"id":"m1","role":"user","ts":"2026-06-22T10:00:00Z","text":"hi"}"#;
+        let evs = run(&[user]);
+        assert_eq!(evs[0].session_id, "unknown");
+    }
+
+    #[test]
+    fn determinism_same_input_same_output() {
+        let lines = [
+            SESSION_START,
+            r#"{"id":"m1","role":"user","ts":"2026-06-22T10:00:00Z","sessionId":"s1","text":"a"}"#,
+            r#"{"id":"m2","parentId":"m1","role":"assistant","ts":"2026-06-22T10:00:05Z","sessionId":"s1","text":"b","edits":[{"path":"p","oldText":"x","newText":"y","added":2,"removed":1}]}"#,
+        ];
+        let a = run(&lines);
+        let b = run(&lines);
+        assert_eq!(a, b);
+    }
+}
diff --git a/crates/memscribe-adapters/src/windsurf.rs b/crates/memscribe-adapters/src/windsurf.rs
new file mode 100644
index 0000000..fc641ad
--- /dev/null
+++ b/crates/memscribe-adapters/src/windsurf.rs
@@ -0,0 +1,527 @@
+//! Windsurf (Codeium) adapter.
+//!
+//! Windsurf is a VS Code-based editor whose Cascade agent stores chat in an
+//! undocumented binary/SQLite store under `~/.codeium/windsurf/` and
+//! `~/Library/Application Support/Windsurf/User/`. We do not parse that binary
+//! store in this model. Instead this adapter targets a deterministic **exported
+//! Cascade chat JSON-Lines** shape (one JSON object per line) and routes any
+//! unrecognized-but-valid record to [`memscribe_core::EventKind::Unknown`] so the
+//! stream stays lossless.
+//!
+//! Record shape (see `fixtures/windsurf/v1/`):
+//! - a leading session header: `{"kind":"session_start","cwd":..,"git":{"sha","branch"},"toolVersion":..,"sessionId":..,"model":..}`
+//! - message records: `{"id","parentId","role":"user"|"assistant","ts","sessionId","text","model","usage":{"input","output"},"toolCalls":[{"id","name","args"}],"toolResults":[{"id","ok","output"}],"edits":[{"path","oldText","newText","diff","added","removed"}]}`
+//!
+//! Mapping: `session_start` → `SessionStart`; `role:user` → `UserTurn`;
+//! `role:assistant` → `AssistantTurn`; each `toolCalls[]` → `ToolCall`; each
+//! `toolResults[]` → `ToolResult{ok}`; each `edits[]` → `FileEdit`.
+//!
+//! Hard rules honored: never panics (no unwrap/expect/indexing on parsed input);
+//! deterministic (no clock/random/global state); `ctx.session_id` is set from the
+//! first record carrying it; `ctx.project` is populated from the session-start
+//! record; repeated records dedup via `ctx.first_seen(event_id)`.
+
+use crate::util;
+use memscribe_core::{
+    content_id, CaptureEvent, Diff, DiscoverCfg, EventKind, GitRef, ParseCtx, ParseError,
+    ProjectRef, RawRecord, SchemaVariant, SourceKind, TranscriptHandle, Usage,
+};
+use std::path::PathBuf;
+
+/// Adapter for Windsurf transcripts.
+#[derive(Debug, Default, Clone, Copy)]
+pub struct WindsurfAdapter;
+
+const SOURCE: SourceKind = SourceKind::Windsurf;
+
+impl memscribe_core::TranscriptAdapter for WindsurfAdapter {
+    fn source_kind(&self) -> SourceKind {
+        SOURCE
+    }
+
+    fn discover(&self, cfg: &DiscoverCfg) -> Vec<TranscriptHandle> {
+        // The real product stores chat in a binary/SQLite store; we do not parse
+        // it here, but discovery still points at the on-disk locations so a
+        // future exporter / probe has the canonical paths. Order is stable.
+        let home = cfg.home_dir();
+        let mut out = Vec::new();
+        let candidates = [
+            home.join(".codeium").join("windsurf"),
+            home.join("Library")
+                .join("Application Support")
+                .join("Windsurf")
+                .join("User"),
+        ];
+        for path in candidates {
+            out.push(TranscriptHandle {
+                path,
+                source: SOURCE,
+                session_hint: None,
+                compressed: false,
+            });
+        }
+        out
+    }
+
+    fn parse(&self, raw: &RawRecord, ctx: &mut ParseCtx) -> Result<Vec<CaptureEvent>, ParseError> {
+        // Blank lines / invalid JSON: skip (blank) or fall through to a string
+        // Unknown so nothing is lost.
+        let Some(value) = util::parse_json_line(raw) else {
+            let s = raw.as_str().map(str::trim).unwrap_or("");
+            if s.is_empty() {
+                return Ok(Vec::new());
+            }
+            let v = serde_json::Value::String(s.to_string());
+            return Ok(vec![util::unknown_event(SOURCE, ctx, raw, v)]);
+        };
+
+        let kind = str_field(&value, "kind");
+        let role = str_field(&value, "role");
+
+        if kind.as_deref() == Some("session_start") {
+            return Ok(parse_session_start(raw, ctx, value));
+        }
+        match role.as_deref() {
+            Some("user") => Ok(parse_message(raw, ctx, value, false)),
+            Some("assistant") => Ok(parse_message(raw, ctx, value, true)),
+            // A valid JSON record we don't recognize: lossless Unknown.
+            _ => Ok(vec![util::unknown_event(SOURCE, ctx, raw, value)]),
+        }
+    }
+
+    fn schema_fingerprint(&self, sample: &RawRecord) -> SchemaVariant {
+        match util::parse_json_line(sample) {
+            Some(v)
+                if str_field(&v, "kind").as_deref() == Some("session_start")
+                    || str_field(&v, "role").is_some() =>
+            {
+                SchemaVariant::certain(SOURCE, "windsurf/cascade-export-v1")
+            }
+            _ => SchemaVariant::unknown(SOURCE),
+        }
+    }
+}
+
+/// Parse the session-start header: stamp `ctx.session_id` and `ctx.project`,
+/// then emit a `SessionStart` event (deduped).
+fn parse_session_start(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    value: serde_json::Value,
+) -> Vec<CaptureEvent> {
+    if let Some(sid) = str_field(&value, "sessionId") {
+        if ctx.session_id.is_none() {
+            ctx.session_id = Some(sid);
+        }
+    }
+
+    let cwd = str_field(&value, "cwd")
+        .map(PathBuf::from)
+        .unwrap_or_else(|| PathBuf::from("."));
+    let git = parse_git(value.get("git"));
+    let model = str_field(&value, "model");
+    let tool_version = str_field(&value, "toolVersion");
+
+    // Bind the project for every event in this session.
+    ctx.project = Some(ProjectRef {
+        cwd: cwd.clone(),
+        repo_root: None,
+        git: git.clone(),
+    });
+
+    let event_id = event_id_for(&value, raw);
+    if !ctx.first_seen(&event_id) {
+        return Vec::new();
+    }
+    let ts = ts_of(&value);
+    vec![util::mk_event(
+        SOURCE,
+        ctx,
+        raw,
+        event_id,
+        None,
+        ts,
+        EventKind::SessionStart {
+            cwd,
+            git,
+            model,
+            tool_version,
+        },
+    )]
+}
+
+/// Parse a user/assistant message record into its turn event plus any nested
+/// tool calls, tool results, and file edits (in a deterministic order).
+fn parse_message(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    value: serde_json::Value,
+    assistant: bool,
+) -> Vec<CaptureEvent> {
+    // Pick up a session id if the header was missing.
+    if ctx.session_id.is_none() {
+        if let Some(sid) = str_field(&value, "sessionId") {
+            ctx.session_id = Some(sid);
+        }
+    }
+
+    let turn_id = event_id_for(&value, raw);
+    // Idempotency: a repeated record (same id) emits nothing.
+    if !ctx.first_seen(&turn_id) {
+        return Vec::new();
+    }
+
+    let parent_id = str_field(&value, "parentId");
+    let ts = ts_of(&value);
+    let text = str_field(&value, "text").unwrap_or_default();
+
+    let mut out = Vec::new();
+
+    let turn_kind = if assistant {
+        let model = str_field(&value, "model");
+        let usage = parse_usage(value.get("usage"));
+        EventKind::AssistantTurn {
+            text,
+            thinking: None,
+            model,
+            usage,
+            parts: Vec::new(),
+        }
+    } else {
+        EventKind::UserTurn {
+            text,
+            parts: Vec::new(),
+        }
+    };
+    out.push(util::mk_event(
+        SOURCE,
+        ctx,
+        raw,
+        turn_id.clone(),
+        parent_id,
+        ts,
+        turn_kind,
+    ));
+
+    // Tool calls — record call name so a later result/edit can pair by call_id.
+    if let Some(calls) = value.get("toolCalls").and_then(|v| v.as_array()) {
+        for (i, call) in calls.iter().enumerate() {
+            let call_id = str_field(call, "id").unwrap_or_else(|| format!("{turn_id}/call/{i}"));
+            let name = str_field(call, "name").unwrap_or_default();
+            let args = call.get("args").cloned().unwrap_or(serde_json::Value::Null);
+            ctx.call_names.insert(call_id.clone(), name.clone());
+            out.push(util::mk_event(
+                SOURCE,
+                ctx,
+                raw,
+                format!("{turn_id}#toolcall:{call_id}"),
+                Some(turn_id.clone()),
+                ts,
+                EventKind::ToolCall {
+                    call_id,
+                    name,
+                    args,
+                },
+            ));
+        }
+    }
+
+    // Tool results — `ok` flag is recorded so edits can detect tool failures.
+    if let Some(results) = value.get("toolResults").and_then(|v| v.as_array()) {
+        for (i, res) in results.iter().enumerate() {
+            let call_id = str_field(res, "id").unwrap_or_else(|| format!("{turn_id}/result/{i}"));
+            let ok = res
+                .get("ok")
+                .and_then(serde_json::Value::as_bool)
+                .unwrap_or(true);
+            let output = res
+                .get("output")
+                .cloned()
+                .unwrap_or(serde_json::Value::Null);
+            ctx.call_ok.insert(call_id.clone(), ok);
+            out.push(util::mk_event(
+                SOURCE,
+                ctx,
+                raw,
+                format!("{turn_id}#toolresult:{call_id}"),
+                Some(turn_id.clone()),
+                ts,
+                EventKind::ToolResult {
+                    call_id,
+                    ok,
+                    output,
+                },
+            ));
+        }
+    }
+
+    // File edits — normalized to FileEdit{diff}.
+    if let Some(edits) = value.get("edits").and_then(|v| v.as_array()) {
+        for (i, edit) in edits.iter().enumerate() {
+            let path = str_field(edit, "path").unwrap_or_default();
+            let diff = Diff {
+                path: PathBuf::from(path),
+                old: str_field(edit, "oldText"),
+                new: str_field(edit, "newText"),
+                unified: str_field(edit, "diff"),
+                added_lines: u32_field(edit, "added"),
+                removed_lines: u32_field(edit, "removed"),
+            };
+            let call_id = str_field(edit, "id");
+            out.push(util::mk_event(
+                SOURCE,
+                ctx,
+                raw,
+                format!("{turn_id}#edit:{i}"),
+                Some(turn_id.clone()),
+                ts,
+                EventKind::FileEdit { call_id, diff },
+            ));
+        }
+    }
+
+    out
+}
+
+/// The event id: tool-native `id` when present, else a content hash of the bytes.
+fn event_id_for(value: &serde_json::Value, raw: &RawRecord) -> String {
+    str_field(value, "id").unwrap_or_else(|| content_id(&raw.bytes))
+}
+
+/// The record timestamp, via the shared `parse_ts` over the common keys.
+fn ts_of(value: &serde_json::Value) -> memscribe_core::Timestamp {
+    util::ts_from(value, &["ts", "timestamp", "time", "created_at"])
+}
+
+/// Read a string field, returning `None` when absent or not a string.
+fn str_field(value: &serde_json::Value, key: &str) -> Option<String> {
+    value.get(key).and_then(|v| v.as_str()).map(str::to_string)
+}
+
+/// Read a non-negative integer field as `u32`, clamped, defaulting to 0.
+fn u32_field(value: &serde_json::Value, key: &str) -> u32 {
+    value
+        .get(key)
+        .and_then(serde_json::Value::as_u64)
+        .map(|n| u32::try_from(n).unwrap_or(u32::MAX))
+        .unwrap_or(0)
+}
+
+/// Parse the optional `git` object into a `GitRef`.
+fn parse_git(value: Option<&serde_json::Value>) -> Option<GitRef> {
+    let g = value?;
+    let sha = str_field(g, "sha")?;
+    let branch = str_field(g, "branch");
+    Some(GitRef { sha, branch })
+}
+
+/// Parse the optional `usage` object into a `Usage`.
+fn parse_usage(value: Option<&serde_json::Value>) -> Option<Usage> {
+    let u = value?;
+    let input_tokens = u.get("input").and_then(serde_json::Value::as_u64);
+    let output_tokens = u.get("output").and_then(serde_json::Value::as_u64);
+    Some(Usage {
+        input_tokens,
+        output_tokens,
+        cache_read_tokens: None,
+        cache_creation_tokens: None,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use memscribe_core::{SourceLocation, TranscriptAdapter};
+    use std::path::Path;
+
+    fn raw(s: &str) -> RawRecord {
+        RawRecord::from_line(s, SourceLocation::new("cascade.jsonl", 0, 1))
+    }
+
+    fn parse_all(lines: &[&str]) -> Vec<CaptureEvent> {
+        let a = WindsurfAdapter;
+        let mut ctx = ParseCtx::new();
+        let mut out = Vec::new();
+        for l in lines {
+            out.extend(a.parse(&raw(l), &mut ctx).expect("parse never errors"));
+        }
+        out
+    }
+
+    const SESSION_START: &str = r#"{"kind":"session_start","sessionId":"ws-1","cwd":"/home/dev/proj","git":{"sha":"abc123","branch":"main"},"toolVersion":"1.2.3","model":"cascade-base"}"#;
+
+    #[test]
+    fn session_start_sets_session_and_project() {
+        let evs = parse_all(&[SESSION_START]);
+        assert_eq!(evs.len(), 1);
+        let e = &evs[0];
+        assert_eq!(e.kind.tag(), "session_start");
+        assert_eq!(e.session_id, "ws-1");
+        assert_eq!(e.project.cwd, PathBuf::from("/home/dev/proj"));
+        let git = e.project.git.as_ref().expect("git bound");
+        assert_eq!(git.sha, "abc123");
+        assert_eq!(git.branch.as_deref(), Some("main"));
+        match &e.kind {
+            EventKind::SessionStart {
+                model,
+                tool_version,
+                ..
+            } => {
+                assert_eq!(model.as_deref(), Some("cascade-base"));
+                assert_eq!(tool_version.as_deref(), Some("1.2.3"));
+            }
+            other => panic!("expected session_start, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn normalized_sequence_user_then_assistant_with_tools_and_edit() {
+        let user = r#"{"id":"u1","role":"user","ts":"2026-06-22T10:00:00Z","sessionId":"ws-1","text":"Let's use Postgres instead of MySQL"}"#;
+        let asst = r#"{"id":"a1","parentId":"u1","role":"assistant","ts":"2026-06-22T10:00:05Z","sessionId":"ws-1","text":"On it.","model":"cascade-base","usage":{"input":10,"output":4},"toolCalls":[{"id":"c1","name":"edit_file","args":{"path":"db.rs"}}],"toolResults":[{"id":"c1","ok":true,"output":"done"}],"edits":[{"id":"c1","path":"db.rs","oldText":"mysql","newText":"postgres","diff":"@@ -1 +1 @@","added":1,"removed":1}]}"#;
+        let tags: Vec<&str> = parse_all(&[SESSION_START, user, asst])
+            .iter()
+            .map(|e| e.kind.tag())
+            .collect();
+        assert_eq!(
+            tags,
+            vec![
+                "session_start",
+                "user_turn",
+                "assistant_turn",
+                "tool_call",
+                "tool_result",
+                "file_edit",
+            ]
+        );
+    }
+
+    #[test]
+    fn decision_then_edit_produces_user_turn_then_file_edit() {
+        let user = r#"{"id":"u1","role":"user","ts":"2026-06-22T10:00:00Z","sessionId":"ws-1","text":"Let's use Postgres instead of MySQL","edits":[{"path":"schema.sql","oldText":"a","newText":"b","added":1,"removed":1}]}"#;
+        let evs = parse_all(&[user]);
+        assert_eq!(evs.len(), 2);
+        assert_eq!(evs[0].kind.tag(), "user_turn");
+        assert_eq!(evs[1].kind.tag(), "file_edit");
+        match &evs[0].kind {
+            EventKind::UserTurn { text, .. } => {
+                assert_eq!(text, "Let's use Postgres instead of MySQL");
+            }
+            other => panic!("expected user_turn, got {other:?}"),
+        }
+        match &evs[1].kind {
+            EventKind::FileEdit { diff, .. } => {
+                assert_eq!(diff.path, PathBuf::from("schema.sql"));
+                assert_eq!(diff.added_lines, 1);
+                assert_eq!(diff.removed_lines, 1);
+            }
+            other => panic!("expected file_edit, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn assistant_usage_and_model_are_copied() {
+        let asst = r#"{"id":"a1","role":"assistant","ts":"2026-06-22T10:00:05Z","sessionId":"ws-1","text":"hi","model":"cascade-pro","usage":{"input":100,"output":42}}"#;
+        let evs = parse_all(&[asst]);
+        match &evs[0].kind {
+            EventKind::AssistantTurn { model, usage, .. } => {
+                assert_eq!(model.as_deref(), Some("cascade-pro"));
+                let u = usage.as_ref().expect("usage present");
+                assert_eq!(u.input_tokens, Some(100));
+                assert_eq!(u.output_tokens, Some(42));
+            }
+            other => panic!("expected assistant_turn, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn tool_failure_result_marks_not_ok() {
+        let asst = r#"{"id":"a1","role":"assistant","ts":"2026-06-22T10:00:05Z","sessionId":"ws-1","text":"trying","toolResults":[{"id":"c1","ok":false,"output":"permission denied"}],"edits":[{"id":"c1","path":"locked.rs","oldText":"x","newText":"y","added":1,"removed":1}]}"#;
+        let evs = parse_all(&[asst]);
+        let tr = evs
+            .iter()
+            .find(|e| e.kind.tag() == "tool_result")
+            .expect("tool_result present");
+        match &tr.kind {
+            EventKind::ToolResult { ok, .. } => assert!(!ok),
+            other => panic!("expected tool_result, got {other:?}"),
+        }
+        // The edit is still captured (losslessness); episode-building downstream
+        // decides not to mint an Episode for a failed edit.
+        assert!(evs.iter().any(|e| e.kind.tag() == "file_edit"));
+    }
+
+    #[test]
+    fn unrecognized_valid_record_becomes_unknown() {
+        let weird = r#"{"kind":"telemetry","payload":{"latency_ms":12}}"#;
+        let evs = parse_all(&[weird]);
+        assert_eq!(evs.len(), 1);
+        assert_eq!(evs[0].kind.tag(), "unknown");
+    }
+
+    #[test]
+    fn garbage_never_panics_and_is_lossless() {
+        // Invalid JSON, a bare scalar, and blank input.
+        let a = WindsurfAdapter;
+        let mut ctx = ParseCtx::new();
+        let garbage = a.parse(&raw("{not json at all"), &mut ctx).unwrap();
+        assert_eq!(garbage.len(), 1);
+        assert_eq!(garbage[0].kind.tag(), "unknown");
+        let scalar = a.parse(&raw("42"), &mut ctx).unwrap();
+        assert_eq!(scalar.len(), 1);
+        assert_eq!(scalar[0].kind.tag(), "unknown");
+        let blank = a.parse(&raw("   "), &mut ctx).unwrap();
+        assert!(blank.is_empty());
+    }
+
+    #[test]
+    fn repeated_record_is_deduped() {
+        let user = r#"{"id":"u1","role":"user","ts":"2026-06-22T10:00:00Z","sessionId":"ws-1","text":"hello"}"#;
+        let evs = parse_all(&[user, user]);
+        assert_eq!(evs.len(), 1, "second identical record dedups to empty");
+        assert_eq!(evs[0].kind.tag(), "user_turn");
+    }
+
+    #[test]
+    fn seq_is_monotonic_and_deterministic() {
+        let user = r#"{"id":"u1","role":"user","ts":"2026-06-22T10:00:00Z","sessionId":"ws-1","text":"a"}"#;
+        let asst = r#"{"id":"a1","role":"assistant","ts":"2026-06-22T10:00:01Z","sessionId":"ws-1","text":"b"}"#;
+        let evs = parse_all(&[SESSION_START, user, asst]);
+        let seqs: Vec<u64> = evs.iter().map(|e| e.seq).collect();
+        assert_eq!(seqs, vec![0, 1, 2]);
+    }
+
+    #[test]
+    fn no_id_falls_back_to_content_hash() {
+        let rec =
+            r#"{"role":"user","ts":"2026-06-22T10:00:00Z","sessionId":"ws-1","text":"no id here"}"#;
+        let evs = parse_all(&[rec]);
+        assert_eq!(evs.len(), 1);
+        assert!(!evs[0].event_id.is_empty());
+        // Deterministic: same bytes → same id.
+        let again = parse_all(&[rec]);
+        assert_eq!(evs[0].event_id, again[0].event_id);
+    }
+
+    #[test]
+    fn discover_points_at_real_product_paths() {
+        let cfg = DiscoverCfg {
+            home: Some(PathBuf::from("/home/dev")),
+            ..DiscoverCfg::default()
+        };
+        let handles = WindsurfAdapter.discover(&cfg);
+        assert_eq!(handles.len(), 2);
+        assert!(handles.iter().all(|h| h.source == SourceKind::Windsurf));
+        assert!(handles
+            .iter()
+            .any(|h| h.path == Path::new("/home/dev/.codeium/windsurf")));
+    }
+
+    #[test]
+    fn schema_fingerprint_recognizes_cascade_export() {
+        let fp = WindsurfAdapter.schema_fingerprint(&raw(SESSION_START));
+        assert_eq!(fp.variant, "windsurf/cascade-export-v1");
+        assert_eq!(fp.confidence, 100);
+        let unknown = WindsurfAdapter.schema_fingerprint(&raw("{not json"));
+        assert_eq!(unknown.confidence, 0);
+    }
+}
diff --git a/crates/memscribe-adapters/src/zed.rs b/crates/memscribe-adapters/src/zed.rs
new file mode 100644
index 0000000..0b1ac5d
--- /dev/null
+++ b/crates/memscribe-adapters/src/zed.rs
@@ -0,0 +1,776 @@
+//! Zed adapter.
+//!
+//! Zed's assistant/agent stores threads under its application support directory
+//! (`~/Library/Application Support/Zed/threads/` on macOS, `~/.local/share/zed/`
+//! on Linux) in an undocumented SQLite/JSON store. We do not parse that binary
+//! store in this model; instead this adapter targets an **exported JSON-lines**
+//! thread shape and routes anything unrecognized to [`EventKind::Unknown`] so the
+//! stream stays lossless across Zed's frequent format churn.
+//!
+//! ## Exported record shape (one JSON object per line)
+//!
+//! A leading session header:
+//! ```json
+//! {"kind":"session_start","cwd":"…","git":{"sha":"…","branch":"…"},
+//!  "toolVersion":"zed 0.182.0","sessionId":"…","ts":"2026-06-22T10:00:00Z"}
+//! ```
+//! followed by message records:
+//! ```json
+//! {"id":"…","parentId":"…","role":"user|assistant","ts":"…","sessionId":"…",
+//!  "text":"…","model":"…","usage":{"input":N,"output":N},
+//!  "toolCalls":[{"id":"…","name":"…","args":{…}}],
+//!  "toolResults":[{"id":"…","ok":true,"output":…}],
+//!  "edits":[{"path":"…","oldText":"…","newText":"…","diff":"…",
+//!            "added":N,"removed":N}]}
+//! ```
+//! and an optional `{"kind":"session_end","reason":"…"}` trailer.
+//!
+//! ## Mapping
+//! - `kind:session_start` → [`EventKind::SessionStart`] (also binds
+//!   `ctx.session_id` and `ctx.project`).
+//! - `role:user` → [`EventKind::UserTurn`].
+//! - `role:assistant` → [`EventKind::AssistantTurn`] (`text`, `model`, `usage`).
+//! - `toolCalls[]` → [`EventKind::ToolCall`].
+//! - `toolResults[]` → [`EventKind::ToolResult`] (`ok`), and the `ok` flag is
+//!   remembered so downstream can suppress episodes for failed edits.
+//! - `edits[]` → [`EventKind::FileEdit`] with a normalized [`Diff`].
+//!
+//! ## Invariants
+//! Never panics (no `unwrap`/`expect`/indexing on parsed input); fully
+//! deterministic (no clock/random/global state); deduplicates by record id via
+//! [`ParseCtx::first_seen`]; any valid-but-unrecognized record becomes
+//! [`EventKind::Unknown`].
+
+use crate::util;
+use memscribe_core::{
+    CaptureEvent, Diff, DiscoverCfg, EventKind, GitRef, ParseCtx, ParseError, ProjectRef,
+    RawRecord, SchemaVariant, SourceKind, TranscriptAdapter, TranscriptHandle, Usage,
+};
+use std::path::PathBuf;
+
+/// Adapter for Zed transcripts.
+#[derive(Debug, Default, Clone, Copy)]
+pub struct ZedAdapter;
+
+impl TranscriptAdapter for ZedAdapter {
+    fn source_kind(&self) -> SourceKind {
+        SourceKind::Zed
+    }
+
+    fn discover(&self, cfg: &DiscoverCfg) -> Vec<TranscriptHandle> {
+        let home = cfg.home_dir();
+        // Zed's real on-disk thread stores. We point at them so the runtime can
+        // surface where Zed history lives even though this model parses exported
+        // JSONL rather than the binary store.
+        let roots = [
+            home.join("Library/Application Support/Zed/threads"),
+            home.join(".local/share/zed/threads"),
+            home.join(".local/share/zed"),
+        ];
+        let mut handles = Vec::new();
+        for root in roots {
+            if !root.is_dir() {
+                continue;
+            }
+            for entry in walkdir::WalkDir::new(&root)
+                .max_depth(4)
+                .into_iter()
+                .filter_map(Result::ok)
+            {
+                let path = entry.path();
+                if !path.is_file() {
+                    continue;
+                }
+                let ext_ok = path
+                    .extension()
+                    .and_then(|e| e.to_str())
+                    .map(|e| matches!(e, "jsonl" | "json" | "ndjson"))
+                    .unwrap_or(false);
+                if !ext_ok {
+                    continue;
+                }
+                let session_hint = path
+                    .file_stem()
+                    .and_then(|s| s.to_str())
+                    .map(str::to_string);
+                handles.push(TranscriptHandle {
+                    path: path.to_path_buf(),
+                    source: SourceKind::Zed,
+                    session_hint,
+                    compressed: false,
+                });
+            }
+        }
+        // Deterministic ordering across platforms / filesystem iteration order.
+        handles.sort_by(|a, b| a.path.cmp(&b.path));
+        handles
+    }
+
+    fn parse(&self, raw: &RawRecord, ctx: &mut ParseCtx) -> Result<Vec<CaptureEvent>, ParseError> {
+        let Some(value) = util::parse_json_line(raw) else {
+            // Blank line → nothing; non-JSON garbage → lossless Unknown.
+            let s = raw.as_str().map(str::trim).unwrap_or("");
+            if s.is_empty() {
+                return Ok(Vec::new());
+            }
+            let v = serde_json::Value::String(s.to_string());
+            return Ok(vec![util::unknown_event(SourceKind::Zed, ctx, raw, v)]);
+        };
+
+        // `kind`-tagged control records (session lifecycle).
+        if let Some(kind) = value.get("kind").and_then(|v| v.as_str()) {
+            match kind {
+                "session_start" => return Ok(parse_session_start(raw, ctx, &value)),
+                "session_end" => return Ok(parse_session_end(raw, ctx, &value)),
+                _ => return Ok(vec![util::unknown_event(SourceKind::Zed, ctx, raw, value)]),
+            }
+        }
+
+        // Otherwise it should be a `role`-tagged message record.
+        if value.get("role").and_then(|v| v.as_str()).is_some() {
+            return Ok(parse_message(raw, ctx, &value));
+        }
+
+        // Valid JSON we don't recognize → Unknown (losslessness).
+        Ok(vec![util::unknown_event(SourceKind::Zed, ctx, raw, value)])
+    }
+
+    fn schema_fingerprint(&self, sample: &RawRecord) -> SchemaVariant {
+        let Some(value) = util::parse_json_line(sample) else {
+            return SchemaVariant::unknown(SourceKind::Zed);
+        };
+        let looks_like_zed = value.get("kind").and_then(|v| v.as_str()) == Some("session_start")
+            || (value.get("role").is_some()
+                && (value.get("toolCalls").is_some()
+                    || value.get("toolResults").is_some()
+                    || value.get("edits").is_some()
+                    || value.get("sessionId").is_some()));
+        if looks_like_zed {
+            SchemaVariant::certain(SourceKind::Zed, "zed/export-v1")
+        } else {
+            SchemaVariant::unknown(SourceKind::Zed)
+        }
+    }
+}
+
+/// Parse a `kind:session_start` header, binding session + project on `ctx`.
+fn parse_session_start(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    value: &serde_json::Value,
+) -> Vec<CaptureEvent> {
+    if let Some(sid) = value.get("sessionId").and_then(|v| v.as_str()) {
+        ctx.session_id = Some(sid.to_string());
+    }
+    let cwd = value
+        .get("cwd")
+        .and_then(|v| v.as_str())
+        .map(PathBuf::from)
+        .unwrap_or_else(|| PathBuf::from("."));
+    let git = parse_git(value.get("git"));
+    let model = value
+        .get("model")
+        .and_then(|v| v.as_str())
+        .map(str::to_string);
+    let tool_version = value
+        .get("toolVersion")
+        .and_then(|v| v.as_str())
+        .map(str::to_string);
+
+    // Bind the project for every subsequent event in this session.
+    ctx.project = Some(ProjectRef {
+        cwd: cwd.clone(),
+        repo_root: None,
+        git: git.clone(),
+    });
+
+    let event_id = event_id_for(value, raw);
+    if !ctx.first_seen(&event_id) {
+        return Vec::new();
+    }
+    let ts = util::ts_from(value, &["ts", "timestamp", "time"]);
+    vec![util::mk_event(
+        SourceKind::Zed,
+        ctx,
+        raw,
+        event_id,
+        None,
+        ts,
+        EventKind::SessionStart {
+            cwd,
+            git,
+            model,
+            tool_version,
+        },
+    )]
+}
+
+/// Parse a `kind:session_end` trailer.
+fn parse_session_end(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    value: &serde_json::Value,
+) -> Vec<CaptureEvent> {
+    if let Some(sid) = value.get("sessionId").and_then(|v| v.as_str()) {
+        if ctx.session_id.is_none() {
+            ctx.session_id = Some(sid.to_string());
+        }
+    }
+    let event_id = event_id_for(value, raw);
+    if !ctx.first_seen(&event_id) {
+        return Vec::new();
+    }
+    let ts = util::ts_from(value, &["ts", "timestamp", "time"]);
+    let reason = value
+        .get("reason")
+        .and_then(|v| v.as_str())
+        .map(str::to_string);
+    vec![util::mk_event(
+        SourceKind::Zed,
+        ctx,
+        raw,
+        event_id,
+        None,
+        ts,
+        EventKind::SessionEnd { reason },
+    )]
+}
+
+/// Parse a `role`-tagged message record into its turn plus any embedded
+/// tool calls, tool results, and file edits (one record fans out to many events).
+fn parse_message(
+    raw: &RawRecord,
+    ctx: &mut ParseCtx,
+    value: &serde_json::Value,
+) -> Vec<CaptureEvent> {
+    if let Some(sid) = value.get("sessionId").and_then(|v| v.as_str()) {
+        if ctx.session_id.is_none() {
+            ctx.session_id = Some(sid.to_string());
+        }
+    }
+
+    let record_id = event_id_for(value, raw);
+    // Idempotency: a repeated record (same id) yields nothing.
+    if !ctx.first_seen(&record_id) {
+        return Vec::new();
+    }
+
+    let ts = util::ts_from(value, &["ts", "timestamp", "time", "created_at"]);
+    let parent_id = value
+        .get("parentId")
+        .and_then(|v| v.as_str())
+        .filter(|s| !s.is_empty())
+        .map(str::to_string);
+    let role = value.get("role").and_then(|v| v.as_str()).unwrap_or("");
+    let text = value
+        .get("text")
+        .and_then(|v| v.as_str())
+        .unwrap_or("")
+        .to_string();
+
+    let mut events = Vec::new();
+
+    // 1) The turn itself.
+    let turn_kind = match role {
+        "user" => EventKind::UserTurn {
+            text,
+            parts: Vec::new(),
+        },
+        "assistant" => {
+            let model = value
+                .get("model")
+                .and_then(|v| v.as_str())
+                .map(str::to_string);
+            let thinking = value
+                .get("thinking")
+                .and_then(|v| v.as_str())
+                .map(str::to_string);
+            let usage = parse_usage(value.get("usage"));
+            EventKind::AssistantTurn {
+                text,
+                thinking,
+                model,
+                usage,
+                parts: Vec::new(),
+            }
+        }
+        _ => {
+            // Unknown role → lossless Unknown for the whole record.
+            return vec![util::unknown_event(
+                SourceKind::Zed,
+                ctx,
+                raw,
+                value.clone(),
+            )];
+        }
+    };
+    events.push(util::mk_event(
+        SourceKind::Zed,
+        ctx,
+        raw,
+        record_id.clone(),
+        parent_id,
+        ts,
+        turn_kind,
+    ));
+
+    // 2) Tool calls embedded in the turn.
+    if let Some(calls) = value.get("toolCalls").and_then(|v| v.as_array()) {
+        for (i, call) in calls.iter().enumerate() {
+            let call_id = call
+                .get("id")
+                .and_then(|v| v.as_str())
+                .map(str::to_string)
+                .unwrap_or_else(|| format!("{record_id}#call{i}"));
+            let name = call
+                .get("name")
+                .and_then(|v| v.as_str())
+                .unwrap_or("")
+                .to_string();
+            let args = call.get("args").cloned().unwrap_or(serde_json::Value::Null);
+            // Remember the call name for pairing with results/edits.
+            ctx.call_names.insert(call_id.clone(), name.clone());
+            let child_id = format!("{record_id}:call:{call_id}");
+            if !ctx.first_seen(&child_id) {
+                continue;
+            }
+            events.push(util::mk_event(
+                SourceKind::Zed,
+                ctx,
+                raw,
+                child_id,
+                Some(record_id.clone()),
+                ts,
+                EventKind::ToolCall {
+                    call_id,
+                    name,
+                    args,
+                },
+            ));
+        }
+    }
+
+    // 3) Tool results embedded in the turn.
+    if let Some(results) = value.get("toolResults").and_then(|v| v.as_array()) {
+        for (i, result) in results.iter().enumerate() {
+            let call_id = result
+                .get("id")
+                .and_then(|v| v.as_str())
+                .map(str::to_string)
+                .unwrap_or_else(|| format!("{record_id}#res{i}"));
+            // `ok` defaults to true when omitted; an explicit `false` marks failure.
+            let ok = result.get("ok").and_then(|v| v.as_bool()).unwrap_or(true);
+            let output = result
+                .get("output")
+                .cloned()
+                .unwrap_or(serde_json::Value::Null);
+            // Remember success/failure so downstream can suppress failed-edit episodes.
+            ctx.call_ok.insert(call_id.clone(), ok);
+            let child_id = format!("{record_id}:result:{call_id}");
+            if !ctx.first_seen(&child_id) {
+                continue;
+            }
+            events.push(util::mk_event(
+                SourceKind::Zed,
+                ctx,
+                raw,
+                child_id,
+                Some(record_id.clone()),
+                ts,
+                EventKind::ToolResult {
+                    call_id,
+                    ok,
+                    output,
+                },
+            ));
+        }
+    }
+
+    // 4) File edits embedded in the turn.
+    if let Some(edits) = value.get("edits").and_then(|v| v.as_array()) {
+        for (i, edit) in edits.iter().enumerate() {
+            let path = edit
+                .get("path")
+                .and_then(|v| v.as_str())
+                .unwrap_or("")
+                .to_string();
+            let old = edit
+                .get("oldText")
+                .and_then(|v| v.as_str())
+                .map(str::to_string);
+            let new = edit
+                .get("newText")
+                .and_then(|v| v.as_str())
+                .map(str::to_string);
+            let unified = edit
+                .get("diff")
+                .and_then(|v| v.as_str())
+                .map(str::to_string);
+            let added_lines = edit.get("added").and_then(|v| v.as_u64()).unwrap_or(0) as u32;
+            let removed_lines = edit.get("removed").and_then(|v| v.as_u64()).unwrap_or(0) as u32;
+            // Correlate the edit to a tool call in the same record, if exactly one
+            // exists (so downstream can join the edit to its result's `ok` flag).
+            let call_id = edit
+                .get("callId")
+                .and_then(|v| v.as_str())
+                .map(str::to_string)
+                .or_else(|| sole_tool_call_id(value));
+            let child_id = format!("{record_id}:edit:{i}");
+            if !ctx.first_seen(&child_id) {
+                continue;
+            }
+            events.push(util::mk_event(
+                SourceKind::Zed,
+                ctx,
+                raw,
+                child_id,
+                Some(record_id.clone()),
+                ts,
+                EventKind::FileEdit {
+                    call_id,
+                    diff: Diff {
+                        path: PathBuf::from(path),
+                        old,
+                        new,
+                        unified,
+                        added_lines,
+                        removed_lines,
+                    },
+                },
+            ));
+        }
+    }
+
+    events
+}
+
+/// Resolve the `event_id`: tool-native `id` when present, else a content hash.
+fn event_id_for(value: &serde_json::Value, raw: &RawRecord) -> String {
+    value
+        .get("id")
+        .and_then(|v| v.as_str())
+        .filter(|s| !s.is_empty())
+        .map(str::to_string)
+        .unwrap_or_else(|| memscribe_core::content_id(&raw.bytes))
+}
+
+/// Parse an optional `git` object into a [`GitRef`]. A missing/blank sha yields
+/// `None` rather than an empty ref.
+fn parse_git(value: Option<&serde_json::Value>) -> Option<GitRef> {
+    let g = value?;
+    let sha = g.get("sha").and_then(|v| v.as_str())?;
+    if sha.is_empty() {
+        return None;
+    }
+    let branch = g
+        .get("branch")
+        .and_then(|v| v.as_str())
+        .filter(|s| !s.is_empty())
+        .map(str::to_string);
+    Some(GitRef {
+        sha: sha.to_string(),
+        branch,
+    })
+}
+
+/// Parse an optional `usage` object. Returns `None` when no fields are present.
+fn parse_usage(value: Option<&serde_json::Value>) -> Option<Usage> {
+    let u = value?;
+    let input_tokens = u
+        .get("input")
+        .or_else(|| u.get("input_tokens"))
+        .and_then(|v| v.as_u64());
+    let output_tokens = u
+        .get("output")
+        .or_else(|| u.get("output_tokens"))
+        .and_then(|v| v.as_u64());
+    if input_tokens.is_none() && output_tokens.is_none() {
+        return None;
+    }
+    Some(Usage {
+        input_tokens,
+        output_tokens,
+        cache_read_tokens: None,
+        cache_creation_tokens: None,
+    })
+}
+
+/// If a record carries exactly one tool call, return its id — used to correlate
+/// a sibling file edit to that call. Returns `None` for zero or many calls.
+fn sole_tool_call_id(value: &serde_json::Value) -> Option<String> {
+    let calls = value.get("toolCalls").and_then(|v| v.as_array())?;
+    if calls.len() != 1 {
+        return None;
+    }
+    calls
+        .first()
+        .and_then(|c| c.get("id"))
+        .and_then(|v| v.as_str())
+        .map(str::to_string)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use memscribe_core::SourceLocation;
+
+    fn raw(s: &str) -> RawRecord {
+        RawRecord::from_line(s, SourceLocation::new("zed.jsonl", 0, 1))
+    }
+
+    /// Parse a whole JSONL string through one shared ctx, mirroring runtime use.
+    fn parse_all(lines: &str) -> (Vec<CaptureEvent>, ParseCtx) {
+        let adapter = ZedAdapter;
+        let mut ctx = ParseCtx::new();
+        let mut out = Vec::new();
+        for line in lines.lines() {
+            let evs = adapter.parse(&raw(line), &mut ctx).expect("never errors");
+            out.extend(evs);
+        }
+        (out, ctx)
+    }
+
+    fn tags(evs: &[CaptureEvent]) -> Vec<&'static str> {
+        evs.iter().map(|e| e.kind.tag()).collect()
+    }
+
+    #[test]
+    fn session_start_binds_session_and_project() {
+        let line = r#"{"kind":"session_start","cwd":"/w/orbit","git":{"sha":"abc","branch":"main"},"toolVersion":"zed 0.1","sessionId":"s1","ts":"2026-06-22T10:00:00Z"}"#;
+        let (evs, ctx) = parse_all(line);
+        assert_eq!(tags(&evs), vec!["session_start"]);
+        assert_eq!(ctx.session_id.as_deref(), Some("s1"));
+        assert_eq!(evs[0].session_id, "s1");
+        match &evs[0].kind {
+            EventKind::SessionStart {
+                cwd,
+                git,
+                tool_version,
+                ..
+            } => {
+                assert_eq!(cwd.as_path(), std::path::Path::new("/w/orbit"));
+                assert_eq!(git.as_ref().map(|g| g.sha.as_str()), Some("abc"));
+                assert_eq!(tool_version.as_deref(), Some("zed 0.1"));
+            }
+            other => panic!("expected SessionStart, got {other:?}"),
+        }
+        // Project propagated from session start.
+        assert_eq!(evs[0].project.cwd, std::path::Path::new("/w/orbit"));
+    }
+
+    #[test]
+    fn decision_turn_then_edit_sequence() {
+        let lines = concat!(
+            r#"{"kind":"session_start","cwd":"/w","git":{"sha":"a"},"sessionId":"s","ts":"2026-06-22T10:00:00Z"}"#,
+            "\n",
+            r#"{"id":"u1","role":"user","ts":"2026-06-22T10:00:05Z","sessionId":"s","text":"Let's use Postgres instead of MySQL"}"#,
+            "\n",
+            r#"{"id":"a1","parentId":"u1","role":"assistant","ts":"2026-06-22T10:00:09Z","sessionId":"s","text":"ok","model":"m","usage":{"input":10,"output":3},"edits":[{"path":"src/db.rs","oldText":"mysql","newText":"postgres","diff":"d","added":1,"removed":1}]}"#,
+        );
+        let (evs, _) = parse_all(lines);
+        // A decision (UserTurn) followed by a FileEdit must appear in order.
+        assert_eq!(
+            tags(&evs),
+            vec!["session_start", "user_turn", "assistant_turn", "file_edit"]
+        );
+        // The user decision text is preserved verbatim.
+        match &evs[1].kind {
+            EventKind::UserTurn { text, .. } => assert!(text.contains("Postgres")),
+            other => panic!("expected UserTurn, got {other:?}"),
+        }
+        // The edit normalizes old/new/unified + line counts.
+        match &evs[3].kind {
+            EventKind::FileEdit { diff, .. } => {
+                assert_eq!(diff.path, PathBuf::from("src/db.rs"));
+                assert_eq!(diff.old.as_deref(), Some("mysql"));
+                assert_eq!(diff.new.as_deref(), Some("postgres"));
+                assert_eq!(diff.unified.as_deref(), Some("d"));
+                assert_eq!(diff.added_lines, 1);
+                assert_eq!(diff.removed_lines, 1);
+            }
+            other => panic!("expected FileEdit, got {other:?}"),
+        }
+        // Seq is monotonic across the fanned-out events.
+        let seqs: Vec<u64> = evs.iter().map(|e| e.seq).collect();
+        assert_eq!(seqs, vec![0, 1, 2, 3]);
+    }
+
+    #[test]
+    fn tool_call_and_result_ok_recorded() {
+        let lines = concat!(
+            r#"{"id":"a","role":"assistant","sessionId":"s","text":"calling","toolCalls":[{"id":"c1","name":"read_file","args":{"path":"x"}}]}"#,
+            "\n",
+            r#"{"id":"b","role":"assistant","sessionId":"s","text":"got it","toolResults":[{"id":"c1","ok":true,"output":"data"}]}"#,
+        );
+        let (evs, ctx) = parse_all(lines);
+        assert_eq!(
+            tags(&evs),
+            vec![
+                "assistant_turn",
+                "tool_call",
+                "assistant_turn",
+                "tool_result"
+            ]
+        );
+        assert_eq!(
+            ctx.call_names.get("c1").map(String::as_str),
+            Some("read_file")
+        );
+        assert_eq!(ctx.call_ok.get("c1"), Some(&true));
+    }
+
+    #[test]
+    fn failed_tool_result_marks_call_not_ok() {
+        // An edit whose tool result failed: the edit is captured but the result's
+        // ok:false is recorded so downstream can suppress the episode.
+        let lines = concat!(
+            r#"{"id":"a","role":"assistant","sessionId":"s","text":"editing","toolCalls":[{"id":"c9","name":"edit_file","args":{}}],"edits":[{"path":"src/c.rs","oldText":"x","newText":"y","added":1,"removed":1}]}"#,
+            "\n",
+            r#"{"id":"b","role":"assistant","sessionId":"s","text":"failed","toolResults":[{"id":"c9","ok":false,"output":"locked"}]}"#,
+        );
+        let (evs, ctx) = parse_all(lines);
+        assert_eq!(
+            tags(&evs),
+            vec![
+                "assistant_turn",
+                "tool_call",
+                "file_edit",
+                "assistant_turn",
+                "tool_result"
+            ]
+        );
+        // The edit was correlated to the sole tool call in its record.
+        match &evs[2].kind {
+            EventKind::FileEdit { call_id, .. } => {
+                assert_eq!(call_id.as_deref(), Some("c9"));
+            }
+            other => panic!("expected FileEdit, got {other:?}"),
+        }
+        // The failure is recorded against the call id → downstream drops the episode.
+        assert_eq!(ctx.call_ok.get("c9"), Some(&false));
+        match &evs[4].kind {
+            EventKind::ToolResult { ok, .. } => assert!(!ok),
+            other => panic!("expected ToolResult, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn dedup_repeated_record_is_idempotent() {
+        let line = r#"{"id":"u1","role":"user","sessionId":"s","text":"hi"}"#;
+        let adapter = ZedAdapter;
+        let mut ctx = ParseCtx::new();
+        let first = adapter.parse(&raw(line), &mut ctx).unwrap();
+        let second = adapter.parse(&raw(line), &mut ctx).unwrap();
+        assert_eq!(tags(&first), vec!["user_turn"]);
+        assert!(second.is_empty(), "repeated record must yield nothing");
+    }
+
+    #[test]
+    fn unknown_record_kind_is_lossless() {
+        let line = r#"{"kind":"telemetry_ping","payload":42}"#;
+        let (evs, _) = parse_all(line);
+        assert_eq!(tags(&evs), vec!["unknown"]);
+        match &evs[0].kind {
+            EventKind::Unknown { raw, .. } => {
+                assert_eq!(raw.get("payload").and_then(|v| v.as_i64()), Some(42));
+            }
+            other => panic!("expected Unknown, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn unknown_role_is_lossless() {
+        let line = r#"{"id":"x","role":"system","sessionId":"s","text":"boot"}"#;
+        let (evs, _) = parse_all(line);
+        assert_eq!(tags(&evs), vec!["unknown"]);
+    }
+
+    #[test]
+    fn garbage_input_never_panics_and_is_lossless() {
+        let adapter = ZedAdapter;
+        let mut ctx = ParseCtx::new();
+        // Non-JSON line → Unknown, no panic.
+        let g = adapter.parse(&raw("}{ not json at all"), &mut ctx).unwrap();
+        assert_eq!(tags(&g), vec!["unknown"]);
+        // Blank line → nothing.
+        let blank = adapter.parse(&raw("   "), &mut ctx).unwrap();
+        assert!(blank.is_empty());
+        // Truncated / weird JSON shapes must not panic.
+        for s in [
+            "{",
+            "[1,2,3]",
+            "null",
+            "12345",
+            r#"{"role":"assistant"}"#,
+            r#"{"role":"assistant","edits":[{}]}"#,
+            r#"{"kind":"session_start"}"#,
+            r#"{"role":"assistant","toolCalls":"not-an-array"}"#,
+            r#"{"role":"assistant","usage":"oops"}"#,
+        ] {
+            let _ = adapter.parse(&raw(s), &mut ctx).unwrap();
+        }
+    }
+
+    #[test]
+    fn session_end_emits_session_end() {
+        let line = r#"{"kind":"session_end","sessionId":"s","reason":"user_closed","ts":"2026-06-22T10:01:30Z"}"#;
+        let (evs, _) = parse_all(line);
+        assert_eq!(tags(&evs), vec!["session_end"]);
+        match &evs[0].kind {
+            EventKind::SessionEnd { reason } => assert_eq!(reason.as_deref(), Some("user_closed")),
+            other => panic!("expected SessionEnd, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn no_id_falls_back_to_content_hash() {
+        let line = r#"{"role":"user","sessionId":"s","text":"anon"}"#;
+        let (evs, _) = parse_all(line);
+        assert_eq!(tags(&evs), vec!["user_turn"]);
+        // 64-hex blake3 content id (no native id present).
+        assert_eq!(evs[0].event_id.len(), 64);
+        assert!(evs[0].event_id.chars().all(|c| c.is_ascii_hexdigit()));
+    }
+
+    #[test]
+    fn fingerprint_recognizes_zed_export() {
+        let adapter = ZedAdapter;
+        let start = raw(r#"{"kind":"session_start","sessionId":"s","cwd":"/w"}"#);
+        assert_eq!(adapter.schema_fingerprint(&start).confidence, 100);
+        let msg = raw(r#"{"id":"a","role":"assistant","sessionId":"s","edits":[]}"#);
+        assert_eq!(adapter.schema_fingerprint(&msg).confidence, 100);
+        let foreign = raw(r#"{"type":"summary","text":"x"}"#);
+        assert_eq!(adapter.schema_fingerprint(&foreign).confidence, 0);
+    }
+
+    #[test]
+    fn full_happy_path_fixture_shape_parses() {
+        // Mirrors fixtures/zed/v1/happy_path_decision_then_edits.jsonl in shape.
+        let lines = concat!(
+            r#"{"kind":"session_start","cwd":"/w","git":{"sha":"a","branch":"main"},"toolVersion":"zed 0.1","sessionId":"t1","ts":"2026-06-22T10:00:00Z"}"#,
+            "\n",
+            r#"{"id":"m1","parentId":null,"role":"user","ts":"2026-06-22T10:00:05Z","sessionId":"t1","text":"Let's use Postgres instead of MySQL."}"#,
+            "\n",
+            r#"{"id":"m2","parentId":"m1","role":"assistant","ts":"2026-06-22T10:00:09Z","sessionId":"t1","text":"ok","model":"m","usage":{"input":1,"output":1},"edits":[{"path":"a.rs","oldText":"x","newText":"y","diff":"d","added":1,"removed":1},{"path":"b.rs","oldText":"p","newText":"q","diff":"d2","added":2,"removed":1}]}"#,
+            "\n",
+            r#"{"kind":"session_end","sessionId":"t1","ts":"2026-06-22T10:01:30Z","reason":"user_closed"}"#,
+        );
+        let (evs, _) = parse_all(lines);
+        assert_eq!(
+            tags(&evs),
+            vec![
+                "session_start",
+                "user_turn",
+                "assistant_turn",
+                "file_edit",
+                "file_edit",
+                "session_end"
+            ]
+        );
+        // Every event carries the bound session + project.
+        assert!(evs.iter().all(|e| e.session_id == "t1"));
+        assert!(evs
+            .iter()
+            .all(|e| e.project.cwd == std::path::Path::new("/w")));
+    }
+}
diff --git a/crates/memscribe-cli/Cargo.toml b/crates/memscribe-cli/Cargo.toml
new file mode 100644
index 0000000..7e37530
--- /dev/null
+++ b/crates/memscribe-cli/Cargo.toml
@@ -0,0 +1,36 @@
+[package]
+name = "memscribe-cli"
+version.workspace = true
+edition.workspace = true
+rust-version.workspace = true
+license.workspace = true
+repository.workspace = true
+description = "The `memscribe` binary: watch / hook / parse / replay / verify / redact."
+
+[lints]
+workspace = true
+
+[[bin]]
+name = "memscribe"
+path = "src/main.rs"
+
+[dependencies]
+memscribe-core = { workspace = true }
+memscribe-adapters = { workspace = true }
+memscribe-sink = { workspace = true }
+memscribe-io = { workspace = true, features = ["watch", "cursor-store"] }
+memscribe-testkit = { workspace = true }
+clap = { workspace = true }
+ctrlc = "3"
+anyhow = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+toml = "0.8"
+zstd = { workspace = true }
+tracing = { workspace = true }
+tracing-subscriber = { workspace = true }
+
+[dev-dependencies]
+assert_cmd = { workspace = true }
+predicates = { workspace = true }
+tempfile = { workspace = true }
diff --git a/crates/memscribe-cli/README.md b/crates/memscribe-cli/README.md
new file mode 100644
index 0000000..8f4b6e7
--- /dev/null
+++ b/crates/memscribe-cli/README.md
@@ -0,0 +1,131 @@
+# memscribe (CLI)
+
+The `memscribe` binary — the daemon and the toolbox for Memscribe's
+deterministic, zero-LLM transcript capture. It wraps the workspace crates
+(`memscribe-core`, `-adapters`, `-io`, `-sink`) behind six subcommands.
+
+```console
+cargo run -p memscribe-cli -- <command> [args]
+# or, once installed:
+memscribe <command> [args]
+```
+
+The pipeline is deterministic and never calls a model. By default the redaction
+pass is **on**, so secrets are stripped before anything is written. See the
+workspace [ARCHITECTURE.md](../../ARCHITECTURE.md) for the pipeline and
+[memscribe.example.toml](../../memscribe.example.toml) for the config surface.
+
+---
+
+## Commands
+
+### `watch` — the steady-state capture daemon
+
+Tail discovered transcripts (and serve the hook endpoint), preparing nodes to a
+sink as they arrive.
+
+```console
+memscribe watch [--tools claude,codex,gemini] [--sink ndjson|sqlite|memdb] \
+                [--out FILE|-] [--root DIR ...] [--once] [--config memscribe.toml]
+```
+
+| Flag | Default | Meaning |
+|------|---------|---------|
+| `--tools` | every adapter | Comma-separated tool slugs to watch (`SourceKind::parse` values). |
+| `--sink` | `ndjson` | Sink target: `ndjson`, `sqlite`, or `memdb` (`memdb` needs the `memdb` feature). |
+| `--out` | `-` (stdout) | Where prepared nodes go (a file for `ndjson`/`sqlite`); `-` is stdout. |
+| `--root` | `$HOME` | Directory root(s) to scan for transcripts; repeatable. |
+| `--once` | off | Drain what already exists and exit, instead of tailing live. |
+| `--config` | — | Path to a `memscribe.toml` (see `memscribe.example.toml`). |
+
+### `hook` — the hook handler
+
+Reads a hook payload from stdin, records it, and exits `0` immediately. It never
+blocks the agent and never invokes a model. Agents wire this as their hook
+command.
+
+```console
+memscribe hook < payload.json
+```
+
+### `parse` — one-shot parse a transcript to NDJSON
+
+The workhorse for tests and debugging: run one transcript file through the
+adapter and the full pipeline, emitting prepared nodes as NDJSON on stdout.
+
+```console
+memscribe parse <file> [--as TOOL] [--no-redact]
+```
+
+- `--as TOOL` forces a specific adapter (`claude_code`, `codex`, `gemini`,
+  `otel`, `cursor`, `windsurf`, `zed`, `vscode`, `copilot`). Omit it to infer
+  the tool from the path; if inference fails the command tells you to pass
+  `--as`.
+- `--no-redact` emits verbatim content (used by golden tests that assert on
+  exact text). Redaction is on otherwise.
+
+```console
+memscribe parse ~/.claude/projects/foo/session.jsonl --as claude_code
+```
+
+### `replay` — re-run preparation over a historical session
+
+Re-prepares a transcript file with the current pipeline (redaction on). Useful
+after an adapter or pipeline change to see the new node stream for an old
+session.
+
+```console
+memscribe replay <file> [--as TOOL]
+```
+
+### `verify` — the conformance smoke suite
+
+Parses every fixture under `fixtures/` and prints a per-tool `CASES / OK / NODES`
+table, exiting non-zero on any failure. This is the fast, shellable summary the
+daemon ships with; full cross-tool conformance and the §8.3 invariants live in
+the testkit (`cargo test -p memscribe-testkit`).
+
+```console
+memscribe verify
+memscribe verify --capture     # (planned) snapshot a live session into a new fixture
+```
+
+### `redact` — preview the redaction pass
+
+Reads a file and prints it with secrets replaced by `[REDACTED:<label>]`,
+warning on stderr if anything was stripped. `--no-content` elides all text and
+keeps only structure.
+
+```console
+memscribe redact session.jsonl
+memscribe redact session.jsonl --no-content
+```
+
+---
+
+## Logging
+
+Logs go to **stderr** (stdout is reserved for node output), filtered by the
+standard `RUST_LOG` env var. The default level is `warn`.
+
+```console
+RUST_LOG=debug memscribe parse session.jsonl --as codex
+```
+
+---
+
+## Build features
+
+The CLI builds every adapter by default. The MemDB sink is feature-gated in
+`memscribe-sink` and **off by default**; `--sink memdb` only does anything once
+that feature is compiled in.
+
+```console
+cargo build -p memscribe-cli     # ndjson sink (default), all adapters
+```
+
+The CLI does not yet expose a passthrough feature for `memscribe-sink/memdb`, so
+enabling the MemDB sink is a build-config follow-up: add a
+`memdb = ["memscribe-sink/memdb"]` feature to `crates/memscribe-cli/Cargo.toml`
+and build with `--features memdb`. Until then, `--sink ndjson` (the default) and
+`--sink sqlite` are the available targets from the binary.
diff --git a/crates/memscribe-cli/src/config.rs b/crates/memscribe-cli/src/config.rs
new file mode 100644
index 0000000..52393f2
--- /dev/null
+++ b/crates/memscribe-cli/src/config.rs
@@ -0,0 +1,458 @@
+//! The `memscribe.toml` config loader (whitepaper §10).
+//!
+//! Parses a [`Config`] from the on-disk TOML and resolves it into the concrete
+//! runtime knobs the pipeline already exposes — there is no bespoke runtime
+//! state here, only a deterministic mapping onto existing types:
+//!
+//! | TOML section          | runtime type                                  |
+//! |-----------------------|-----------------------------------------------|
+//! | `[capture]`           | tool set + [`DiscoverCfg`] (`home`, `project_filter`) |
+//! | `[tools.*.overrides]` | [`DiscoverCfg::overrides`]                    |
+//! | `[[gate.rules]]`      | [`CommitmentGate::from_triples`]              |
+//! | `[redact]` / patterns | [`Redactor::from_patterns`] / `no_content`    |
+//! | `[ingest]`            | parsed-and-stored cadence (may be unused yet) |
+//! | `[sink]`              | sink target + path/endpoint                   |
+//!
+//! Every section is optional: a value left out falls back to the compiled
+//! default (`CommitmentGate::default_table`, `Redactor::default`, the NDJSON
+//! sink), so a minimal or empty config is valid and changes nothing.
+//!
+//! The schema mirrors the committed `memscribe.example.toml` one-to-one; that
+//! file is the conformance fixture for [`Config::load`] (see the tests).
+
+use anyhow::{Context, Result};
+use memscribe_core::{CommitmentGate, DiscoverCfg, MarkerCategory, Redactor};
+use serde::Deserialize;
+use std::collections::HashMap;
+use std::path::{Path, PathBuf};
+use std::time::Duration;
+
+/// The parsed `memscribe.toml`. All sections are optional.
+#[derive(Debug, Default, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct Config {
+    /// Which tools to tail and where their transcripts live.
+    #[serde(default)]
+    pub capture: CaptureCfg,
+    /// Per-tool path overrides, keyed by tool slug then env-var name.
+    #[serde(default)]
+    pub tools: HashMap<String, ToolCfg>,
+    /// The commitment-marker gate rule table (replaces the default when present).
+    #[serde(default)]
+    pub gate: Option<GateCfg>,
+    /// The redaction pass configuration.
+    #[serde(default)]
+    pub redact: Option<RedactCfg>,
+    /// Retention / ingest cadence (parsed and stored; may be unused).
+    #[serde(default)]
+    pub ingest: IngestCfg,
+    /// Where prepared nodes go.
+    #[serde(default)]
+    pub sink: SinkCfg,
+}
+
+/// The `[capture]` section.
+#[derive(Debug, Default, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct CaptureCfg {
+    /// The adapter slugs to enable (e.g. `["claude_code", "codex"]`). Empty =
+    /// "use the CLI `--tools` value / every adapter".
+    #[serde(default)]
+    pub tools: Vec<String>,
+    /// Restrict discovery to a single project root (`DiscoverCfg.project_filter`).
+    #[serde(default)]
+    pub project_filter: Option<PathBuf>,
+    /// Override `$HOME` for discovery (`DiscoverCfg.home`).
+    #[serde(default)]
+    pub home: Option<PathBuf>,
+}
+
+/// A `[tools.<slug>]` table: currently just per-tool path overrides.
+#[derive(Debug, Default, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct ToolCfg {
+    /// Native-env-var → path overrides (e.g. `CODEX_HOME`, `CLAUDE_CONFIG_DIR`).
+    #[serde(default)]
+    pub overrides: HashMap<String, PathBuf>,
+}
+
+/// The `[gate]` section: a replacement commitment-marker rule table.
+#[derive(Debug, Default, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct GateCfg {
+    /// The `[[gate.rules]]` array.
+    #[serde(default)]
+    pub rules: Vec<GateRuleCfg>,
+}
+
+/// One `[[gate.rules]]` entry: the `(id, category, pattern)` triple
+/// [`CommitmentGate::from_triples`] consumes.
+#[derive(Debug, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct GateRuleCfg {
+    /// The rule id (e.g. `decision_verb.use`).
+    pub id: String,
+    /// The marker category (snake_case, matching [`MarkerCategory`]).
+    pub category: MarkerCategory,
+    /// The case-insensitive regex pattern.
+    pub pattern: String,
+}
+
+/// The `[redact]` section.
+#[derive(Debug, Default, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct RedactCfg {
+    /// Structure-only mode: elide ALL verbatim text (the `--no-content` flag).
+    #[serde(default)]
+    pub no_content: bool,
+    /// The `[[redact.patterns]]` array (replaces the default set when present).
+    #[serde(default)]
+    pub patterns: Vec<RedactPatternCfg>,
+}
+
+/// One `[[redact.patterns]]` entry: a `(label, pattern)` pair.
+#[derive(Debug, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct RedactPatternCfg {
+    /// The label, surfaced as `[REDACTED:<label>]`.
+    pub label: String,
+    /// The regex pattern to strip.
+    pub pattern: String,
+}
+
+/// The `[ingest]` section: retention / cadence. Parsed and stored even though
+/// the current daemon does not yet read every field (the loader is the contract;
+/// wiring each knob is incremental).
+#[derive(Debug, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct IngestCfg {
+    /// How often the tailer re-scans for new bytes (seconds).
+    #[serde(default = "default_poll_interval_secs")]
+    pub poll_interval_secs: u64,
+    /// Resume each file from its last byte offset instead of re-reading.
+    #[serde(default = "default_true")]
+    pub resume_from_offset: bool,
+    /// Cold-start lookback window in days; `0`/omitted = full history.
+    #[serde(default)]
+    pub backfill_days: u64,
+}
+
+impl Default for IngestCfg {
+    fn default() -> Self {
+        IngestCfg {
+            poll_interval_secs: default_poll_interval_secs(),
+            resume_from_offset: true,
+            backfill_days: 0,
+        }
+    }
+}
+
+fn default_poll_interval_secs() -> u64 {
+    5
+}
+fn default_true() -> bool {
+    true
+}
+
+/// The `[sink]` section.
+#[derive(Debug, Default, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct SinkCfg {
+    /// `ndjson` (default) | `sqlite` | `memdb`.
+    #[serde(default)]
+    pub target: Option<String>,
+    /// `[sink.ndjson]`.
+    #[serde(default)]
+    pub ndjson: Option<NdjsonSinkCfg>,
+    /// `[sink.sqlite]`.
+    #[serde(default)]
+    pub sqlite: Option<SqliteSinkCfg>,
+    /// `[sink.memdb]`.
+    #[serde(default)]
+    pub memdb: Option<MemdbSinkCfg>,
+}
+
+/// `[sink.ndjson]`.
+#[derive(Debug, Default, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct NdjsonSinkCfg {
+    /// Output path; omit for stdout.
+    #[serde(default)]
+    pub path: Option<PathBuf>,
+}
+
+/// `[sink.sqlite]`.
+#[derive(Debug, Default, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct SqliteSinkCfg {
+    /// The SQLite database path.
+    #[serde(default)]
+    pub path: Option<PathBuf>,
+}
+
+/// `[sink.memdb]`.
+#[derive(Debug, Default, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct MemdbSinkCfg {
+    /// The MemDB endpoint (`host:port`).
+    #[serde(default)]
+    pub endpoint: Option<String>,
+}
+
+impl Config {
+    /// Parse a [`Config`] from a TOML file on disk.
+    ///
+    /// # Errors
+    /// Returns an error if the file cannot be read or is not valid TOML for this
+    /// schema.
+    pub fn load(path: &Path) -> Result<Self> {
+        let text = std::fs::read_to_string(path)
+            .with_context(|| format!("reading config {}", path.display()))?;
+        Self::parse_str(&text).with_context(|| format!("parsing config {}", path.display()))
+    }
+
+    /// Parse a [`Config`] from a TOML string (the unit-testable core of
+    /// [`Self::load`]).
+    ///
+    /// # Errors
+    /// Returns the underlying TOML error on a schema mismatch.
+    pub fn parse_str(text: &str) -> Result<Self> {
+        let cfg: Config = toml::from_str(text).context("invalid memscribe.toml")?;
+        Ok(cfg)
+    }
+
+    /// Build the [`DiscoverCfg`] this config implies: `home`, `project_filter`,
+    /// and the union of every tool's `overrides`.
+    #[must_use]
+    pub fn discover_cfg(&self) -> DiscoverCfg {
+        let mut overrides: HashMap<String, PathBuf> = HashMap::new();
+        for tool in self.tools.values() {
+            for (k, v) in &tool.overrides {
+                overrides.insert(k.clone(), v.clone());
+            }
+        }
+        DiscoverCfg {
+            home: self.capture.home.clone(),
+            overrides,
+            project_filter: self.capture.project_filter.clone(),
+        }
+    }
+
+    /// Build the commitment gate: the config's `[[gate.rules]]` if any are given,
+    /// otherwise the compiled default table.
+    ///
+    /// # Errors
+    /// Returns an error if a configured pattern fails to compile.
+    pub fn build_gate(&self) -> Result<CommitmentGate> {
+        match &self.gate {
+            Some(g) if !g.rules.is_empty() => {
+                let triples = g
+                    .rules
+                    .iter()
+                    .map(|r| (r.id.clone(), r.category, r.pattern.clone()));
+                CommitmentGate::from_triples(triples)
+                    .context("compiling a configured [[gate.rules]] pattern")
+            }
+            _ => Ok(CommitmentGate::default_table()),
+        }
+    }
+
+    /// Build the redactor implied by `[redact]`. Custom `[[redact.patterns]]`
+    /// replace the default set; with none given the default patterns are used.
+    /// `no_content` is honored either way. Returns `None` only if a future config
+    /// disables redaction (currently redaction is always on).
+    ///
+    /// # Errors
+    /// Returns an error if a configured pattern fails to compile.
+    pub fn build_redactor(&self) -> Result<Redactor> {
+        match &self.redact {
+            Some(r) if !r.patterns.is_empty() => {
+                let pairs = r
+                    .patterns
+                    .iter()
+                    .map(|p| (p.label.clone(), p.pattern.clone()));
+                Redactor::from_patterns(pairs, r.no_content)
+                    .context("compiling a configured [[redact.patterns]] pattern")
+            }
+            Some(r) => Ok(Redactor::with_default_patterns(r.no_content)),
+            None => Ok(Redactor::default()),
+        }
+    }
+
+    /// The configured sink target, defaulting to `ndjson`.
+    #[must_use]
+    pub fn sink_target(&self) -> &str {
+        self.sink.target.as_deref().unwrap_or("ndjson")
+    }
+
+    /// The configured output path for the selected sink, if the config names one
+    /// (`[sink.ndjson].path` or `[sink.sqlite].path`). `None` means "use the CLI
+    /// `--out` value".
+    #[must_use]
+    pub fn sink_out_path(&self) -> Option<PathBuf> {
+        match self.sink_target() {
+            "sqlite" => self.sink.sqlite.as_ref().and_then(|s| s.path.clone()),
+            _ => self.sink.ndjson.as_ref().and_then(|s| s.path.clone()),
+        }
+    }
+
+    /// The tool slugs the config selects, if `[capture].tools` is non-empty.
+    #[must_use]
+    pub fn capture_tools(&self) -> &[String] {
+        &self.capture.tools
+    }
+
+    /// The tailer poll interval implied by `[ingest].poll_interval_secs`.
+    #[must_use]
+    pub fn poll_interval(&self) -> Duration {
+        Duration::from_secs(self.ingest.poll_interval_secs)
+    }
+
+    /// Whether `[ingest].resume_from_offset` is set (resume vs. re-read).
+    #[must_use]
+    pub fn resume_from_offset(&self) -> bool {
+        self.ingest.resume_from_offset
+    }
+
+    /// The cold-start backfill window in days (`[ingest].backfill_days`); `0`
+    /// means "ingest the full available history".
+    #[must_use]
+    pub fn backfill_days(&self) -> u64 {
+        self.ingest.backfill_days
+    }
+
+    /// The configured MemDB endpoint, when the sink target is `memdb`
+    /// (`[sink.memdb].endpoint`). Surfaced so a future `memdb` sink can consume
+    /// it; `None` otherwise.
+    #[must_use]
+    pub fn memdb_endpoint(&self) -> Option<&str> {
+        if self.sink_target() == "memdb" {
+            self.sink.memdb.as_ref().and_then(|m| m.endpoint.as_deref())
+        } else {
+            None
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The committed example config must parse with this schema — it is the
+    /// loader's conformance fixture.
+    #[test]
+    fn example_toml_parses() {
+        let root = Path::new(env!("CARGO_MANIFEST_DIR"))
+            .ancestors()
+            .nth(2)
+            .unwrap()
+            .to_path_buf();
+        let example = root.join("memscribe.example.toml");
+        let cfg = Config::load(&example).expect("memscribe.example.toml must parse");
+
+        // Spot-check the round-trip against the committed values.
+        assert_eq!(cfg.capture_tools(), ["claude_code", "codex", "gemini"]);
+        assert_eq!(cfg.sink_target(), "ndjson");
+
+        // The example ships the 8 default gate rules and the default redact set.
+        let gate = cfg.build_gate().expect("example gate rules compile");
+        assert!(gate.rule_count() >= 8);
+        let redactor = cfg
+            .build_redactor()
+            .expect("example redact patterns compile");
+        assert!(redactor.contains_secret("sk-ant-AAAAAAAAAAAAAAAAAAAA"));
+
+        // Per-tool overrides land in the DiscoverCfg.
+        let disc = cfg.discover_cfg();
+        assert!(disc.overrides.contains_key("CODEX_HOME"));
+        assert!(disc.overrides.contains_key("CLAUDE_CONFIG_DIR"));
+
+        // Ingest cadence parses.
+        assert_eq!(cfg.ingest.poll_interval_secs, 5);
+        assert!(cfg.ingest.resume_from_offset);
+    }
+
+    #[test]
+    fn empty_config_is_all_defaults() {
+        let cfg = Config::parse_str("").expect("empty config parses");
+        assert!(cfg.capture_tools().is_empty());
+        assert_eq!(cfg.sink_target(), "ndjson");
+        assert!(cfg.sink_out_path().is_none());
+        // Default gate + default redactor.
+        assert_eq!(cfg.build_gate().unwrap().rule_count(), 8);
+        assert!(cfg
+            .build_redactor()
+            .unwrap()
+            .contains_secret("sk-ant-AAAAAAAAAAAAAAAAAAAA"));
+    }
+
+    #[test]
+    fn custom_gate_replaces_the_default_table() {
+        let text = r#"
+[[gate.rules]]
+id = "custom.banana"
+category = "decision_verb"
+pattern = "banana"
+"#;
+        let cfg = Config::parse_str(text).unwrap();
+        let gate = cfg.build_gate().unwrap();
+        assert_eq!(gate.rule_count(), 1);
+        assert!(gate.admits("we will use banana"));
+        assert!(!gate.admits("let's go with redis"));
+    }
+
+    #[test]
+    fn custom_redaction_pattern_is_applied() {
+        let text = r#"
+[redact]
+no_content = false
+
+[[redact.patterns]]
+label = "banana_token"
+pattern = "BANANA-[0-9]{4}"
+"#;
+        let cfg = Config::parse_str(text).unwrap();
+        let r = cfg.build_redactor().unwrap();
+        let out = r.redact_text("here is BANANA-1234 in the log");
+        assert!(out.contains("[REDACTED:banana_token]"));
+        // The default patterns are REPLACED, so a default secret survives.
+        assert!(!r.contains_secret("AKIAABCDEFGHIJKLMNOP"));
+    }
+
+    #[test]
+    fn invalid_pattern_surfaces_as_error() {
+        let text = r#"
+[[gate.rules]]
+id = "bad"
+category = "imperative"
+pattern = "("
+"#;
+        let cfg = Config::parse_str(text).unwrap();
+        assert!(cfg.build_gate().is_err());
+    }
+
+    #[test]
+    fn sink_path_selection_follows_target() {
+        let text = r#"
+[sink]
+target = "sqlite"
+
+[sink.sqlite]
+path = "/tmp/x.db"
+"#;
+        let cfg = Config::parse_str(text).unwrap();
+        assert_eq!(cfg.sink_target(), "sqlite");
+        assert_eq!(cfg.sink_out_path().unwrap(), PathBuf::from("/tmp/x.db"));
+    }
+
+    #[test]
+    fn unknown_category_is_rejected() {
+        let text = r#"
+[[gate.rules]]
+id = "x"
+category = "not_a_category"
+pattern = "x"
+"#;
+        assert!(Config::parse_str(text).is_err());
+    }
+}
diff --git a/crates/memscribe-cli/src/main.rs b/crates/memscribe-cli/src/main.rs
new file mode 100644
index 0000000..e526244
--- /dev/null
+++ b/crates/memscribe-cli/src/main.rs
@@ -0,0 +1,821 @@
+//! The `memscribe` binary — the daemon and the toolbox (whitepaper §10).
+#![forbid(unsafe_code)]
+
+mod config;
+
+use anyhow::{anyhow, bail, Context, Result};
+use clap::{Parser, Subcommand};
+use config::Config;
+use memscribe_core::{DefaultPipeline, PreparedNode, Redactor, Sink, SourceKind};
+use memscribe_io::cursor_store::persistent::SqliteOffsetStore;
+use memscribe_io::discover::find_transcripts;
+use memscribe_io::tailer::LiveTailer;
+use memscribe_sink::NdjsonSink;
+use std::io::Read;
+use std::path::{Path, PathBuf};
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::Arc;
+use std::time::Duration;
+
+/// Deterministic, zero-LLM transcript capture for AI coding agents.
+#[derive(Parser, Debug)]
+#[command(name = "memscribe", version, about, long_about = None)]
+struct Cli {
+    #[command(subcommand)]
+    command: Command,
+}
+
+#[derive(Subcommand, Debug)]
+enum Command {
+    /// Run the tailers + hook server; the steady-state capture daemon.
+    Watch {
+        /// Tools to watch (e.g. `claude,codex,gemini`). Empty = every adapter.
+        #[arg(long, value_delimiter = ',')]
+        tools: Vec<String>,
+        /// Sink target: `ndjson`, `sqlite`, or `memdb`.
+        #[arg(long, default_value = "ndjson")]
+        sink: String,
+        /// Where prepared nodes go (a file for `ndjson`/`sqlite`); `-` is stdout.
+        #[arg(long, default_value = "-")]
+        out: PathBuf,
+        /// Directory roots to scan for transcripts (default: `$HOME`).
+        #[arg(long = "root", value_name = "DIR")]
+        roots: Vec<PathBuf>,
+        /// Drain what already exists and exit, rather than tailing live.
+        #[arg(long)]
+        once: bool,
+        /// Path to a `memscribe.toml` config.
+        #[arg(long)]
+        config: Option<PathBuf>,
+    },
+    /// The hook handler agents invoke (reads stdin, records, returns 0).
+    Hook,
+    /// One-shot parse a transcript to NDJSON (the workhorse for tests/debugging).
+    Parse {
+        /// The transcript file to parse.
+        file: PathBuf,
+        /// Force a specific tool adapter (e.g. `claude_code`).
+        #[arg(long = "as", value_name = "TOOL")]
+        source: Option<String>,
+        /// Do not run the redaction pass (emit verbatim).
+        #[arg(long)]
+        no_redact: bool,
+    },
+    /// Re-run preparation over a historical session (file path).
+    Replay {
+        /// A transcript file to replay.
+        target: PathBuf,
+        /// Force a specific tool adapter.
+        #[arg(long = "as", value_name = "TOOL")]
+        source: Option<String>,
+    },
+    /// Run the conformance suite; with `--capture`, snapshot a live session.
+    Verify {
+        /// Snapshot this transcript into a new fixture instead of asserting.
+        /// Pass the live/sample session file to capture.
+        #[arg(long, value_name = "SESSION_FILE")]
+        capture: Option<PathBuf>,
+        /// Force a specific tool adapter for the captured session (e.g.
+        /// `claude_code`). Inferred from the path when omitted.
+        #[arg(long = "as", value_name = "TOOL")]
+        source: Option<String>,
+        /// The fixture base name to write (defaults to the session file stem).
+        #[arg(long, value_name = "NAME")]
+        name: Option<String>,
+        /// Also write the prepared nodes alongside the raw transcript
+        /// (`<name>.nodes.ndjson`) so the captured corpus carries expected output.
+        #[arg(long)]
+        with_nodes: bool,
+    },
+    /// Show what the redaction pass would strip.
+    Redact {
+        /// The file to inspect.
+        file: PathBuf,
+        /// Elide all content (structure-only `--no-content` mode).
+        #[arg(long)]
+        no_content: bool,
+    },
+}
+
+fn main() -> Result<()> {
+    tracing_subscriber::fmt()
+        .with_env_filter(
+            tracing_subscriber::EnvFilter::try_from_default_env()
+                .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new("warn")),
+        )
+        .with_writer(std::io::stderr)
+        .init();
+
+    let cli = Cli::parse();
+    match cli.command {
+        Command::Parse {
+            file,
+            source,
+            no_redact,
+        } => cmd_parse(&file, source.as_deref(), no_redact),
+        Command::Replay { target, source } => cmd_parse(&target, source.as_deref(), false),
+        Command::Redact { file, no_content } => cmd_redact(&file, no_content),
+        Command::Hook => cmd_hook(),
+        Command::Verify {
+            capture,
+            source,
+            name,
+            with_nodes,
+        } => match capture {
+            Some(session) => {
+                cmd_verify_capture(&session, source.as_deref(), name.as_deref(), with_nodes)
+            }
+            None => cmd_verify(),
+        },
+        Command::Watch {
+            tools,
+            sink,
+            out,
+            roots,
+            once,
+            config,
+        } => cmd_watch(&tools, &sink, &out, &roots, once, config.as_deref()),
+    }
+}
+
+/// Resolve the tool adapter from an explicit `--as` flag or by inferring from
+/// the file path.
+fn resolve_source(source: Option<&str>, file: &Path) -> Result<SourceKind> {
+    if let Some(s) = source {
+        return SourceKind::parse(s).ok_or_else(|| anyhow!("unknown tool `{s}`"));
+    }
+    infer_source(file).ok_or_else(|| {
+        anyhow!(
+            "could not infer the tool from `{}`; pass --as <tool>",
+            file.display()
+        )
+    })
+}
+
+/// Best-effort tool inference from a transcript path. Returns `None` when no
+/// marker is recognizable (the caller decides whether that is fatal).
+fn infer_source(file: &Path) -> Option<SourceKind> {
+    let p = file.to_string_lossy().to_ascii_lowercase();
+    let inferred = if p.contains(".codex") || p.contains("codex") {
+        SourceKind::Codex
+    } else if p.contains(".claude") || p.contains("claude") {
+        SourceKind::ClaudeCode
+    } else if p.contains(".gemini") || p.contains("gemini") {
+        SourceKind::Gemini
+    } else if p.contains("cursor") {
+        SourceKind::Cursor
+    } else if p.contains("windsurf") {
+        SourceKind::Windsurf
+    } else if p.contains("zed") {
+        SourceKind::Zed
+    } else if p.contains("copilot") {
+        SourceKind::Copilot
+    } else if p.contains("vscode") || p.contains("code") {
+        SourceKind::VsCode
+    } else if p.contains("otel") || p.ends_with(".ndjson") {
+        SourceKind::Otel
+    } else {
+        return None;
+    };
+    Some(inferred)
+}
+
+fn cmd_parse(file: &Path, source: Option<&str>, no_redact: bool) -> Result<()> {
+    let kind = resolve_source(source, file)?;
+    let adapter = memscribe_adapters::adapter_for(kind)
+        .ok_or_else(|| anyhow!("the `{kind}` adapter is not compiled into this build"))?;
+    let records =
+        memscribe_io::read_records(file).with_context(|| format!("reading {}", file.display()))?;
+
+    let pipeline = if no_redact {
+        DefaultPipeline::without_redaction()
+    } else {
+        DefaultPipeline::new()
+    };
+    let nodes = pipeline.run_records(adapter.as_ref(), &records);
+
+    let mut sink = NdjsonSink::stdout();
+    for n in &nodes {
+        sink.emit(n)?;
+    }
+    sink.flush()?;
+    Ok(())
+}
+
+fn cmd_redact(file: &Path, no_content: bool) -> Result<()> {
+    let redactor = Redactor::with_default_patterns(no_content);
+    let content =
+        std::fs::read_to_string(file).with_context(|| format!("reading {}", file.display()))?;
+    let had_secret = redactor.contains_secret(&content);
+    print!("{}", redactor.redact_text(&content));
+    if had_secret {
+        eprintln!("memscribe: redacted one or more secrets");
+    }
+    Ok(())
+}
+
+fn cmd_hook() -> Result<()> {
+    // Read the hook payload, record it, and exit 0 immediately — never block the
+    // agent, never invoke a model.
+    let mut buf = Vec::new();
+    std::io::stdin().read_to_end(&mut buf)?;
+    if let Some(p) = memscribe_io::hook::record_hook(&buf) {
+        tracing::debug!(
+            event = ?p.payload.hook_event_name,
+            transcript = ?p.transcript_path,
+            "memscribe hook received"
+        );
+    }
+    Ok(())
+}
+
+// ---------------------------------------------------------------------------
+// verify
+// ---------------------------------------------------------------------------
+
+/// Parse every fixture under `fixtures/` and assert each yields a non-error node
+/// stream, printing a per-tool PASS/FAIL table. Exits non-zero on any failure.
+///
+/// Full cross-tool conformance (golden equality, the §8.3 invariants) lives in
+/// the testkit (`cargo test -p memscribe-testkit`); this is the fast, shellable
+/// smoke summary the daemon ships with.
+fn cmd_verify() -> Result<()> {
+    let fixtures = memscribe_testkit::golden::fixtures_dir();
+    if !fixtures.is_dir() {
+        bail!("no fixtures directory at {}", fixtures.display());
+    }
+
+    // Aggregate per tool: (cases, ok-cases, total-nodes, sample failure).
+    use std::collections::BTreeMap;
+    #[derive(Default)]
+    struct Tally {
+        cases: usize,
+        ok: usize,
+        nodes: usize,
+        first_error: Option<String>,
+    }
+    let mut by_tool: BTreeMap<String, Tally> = BTreeMap::new();
+
+    // Walk fixtures/<tool>/<version>/<case>.jsonl deterministically.
+    let mut tool_dirs: Vec<PathBuf> = std::fs::read_dir(&fixtures)
+        .with_context(|| format!("reading {}", fixtures.display()))?
+        .filter_map(Result::ok)
+        .map(|e| e.path())
+        .filter(|p| p.is_dir())
+        .collect();
+    tool_dirs.sort();
+
+    for tool_dir in tool_dirs {
+        let tool_slug = tool_dir
+            .file_name()
+            .and_then(|n| n.to_str())
+            .unwrap_or("?")
+            .to_string();
+        let kind = SourceKind::parse(&tool_slug);
+
+        let files = find_transcripts(&tool_dir, &["jsonl", "json", "zst"]);
+        for disc in files {
+            let tally = by_tool.entry(tool_slug.clone()).or_default();
+            tally.cases += 1;
+            match verify_one(kind, &disc.path) {
+                Ok(n) => {
+                    tally.ok += 1;
+                    tally.nodes += n;
+                }
+                Err(e) => {
+                    if tally.first_error.is_none() {
+                        tally.first_error = Some(format!("{}: {e}", disc.path.display()));
+                    }
+                }
+            }
+        }
+    }
+
+    // Render the table.
+    println!(
+        "{:<14} {:>6} {:>6} {:>8}  STATUS",
+        "TOOL", "CASES", "OK", "NODES"
+    );
+    let mut any_fail = false;
+    for (tool, t) in &by_tool {
+        let pass = t.ok == t.cases && t.cases > 0;
+        any_fail |= !pass;
+        println!(
+            "{:<14} {:>6} {:>6} {:>8}  {}",
+            tool,
+            t.cases,
+            t.ok,
+            t.nodes,
+            if pass { "PASS" } else { "FAIL" }
+        );
+        if let Some(err) = &t.first_error {
+            println!("    └─ first failure: {err}");
+        }
+    }
+
+    if by_tool.is_empty() {
+        bail!("no fixtures found under {}", fixtures.display());
+    }
+    if any_fail {
+        bail!("verify: one or more tools FAILED conformance smoke");
+    }
+    println!("verify: all {} tool(s) PASS", by_tool.len());
+    Ok(())
+}
+
+/// `verify --capture <session-file> --as <tool>` (whitepaper M6).
+///
+/// Snapshot a live/sample transcript into a NEW fixture so the corpus can grow
+/// from real sessions: resolve the tool, parse the transcript to confirm it
+/// yields a clean event stream (we never capture a broken sample), then copy the
+/// raw transcript verbatim into `fixtures/<tool>/captured/<name>.jsonl`. With
+/// `--with-nodes`, the prepared nodes are written alongside as
+/// `<name>.nodes.ndjson` so the captured case carries its expected output.
+fn cmd_verify_capture(
+    session: &Path,
+    source: Option<&str>,
+    name: Option<&str>,
+    with_nodes: bool,
+) -> Result<()> {
+    // Resolve the adapter the same way `parse` does (explicit `--as`, else infer).
+    let kind = resolve_source(source, session)?;
+    if memscribe_adapters::adapter_for(kind).is_none() {
+        bail!("the `{kind}` adapter is not compiled into this build");
+    }
+
+    // Read the raw transcript bytes (verbatim — the fixture is the source of
+    // truth) and the decompressed bytes for the conformance check.
+    let raw =
+        std::fs::read(session).with_context(|| format!("reading session {}", session.display()))?;
+    let bytes = read_decompressed(session)?;
+
+    // Confirm the sample is well-formed BEFORE we admit it to the corpus: the
+    // event stream must satisfy the §8.3 invariants and every prepared node must
+    // serialize. A broken sample is rejected rather than captured.
+    verify_one(Some(kind), session).with_context(|| {
+        format!(
+            "refusing to capture {}: it does not pass conformance",
+            session.display()
+        )
+    })?;
+
+    // Choose the fixture name: explicit `--name`, else the session file stem.
+    let base = match name {
+        Some(n) => n.to_string(),
+        None => session
+            .file_stem()
+            .and_then(|s| s.to_str())
+            .map(str::to_string)
+            .ok_or_else(|| anyhow!("could not derive a fixture name from {}", session.display()))?,
+    };
+
+    // fixtures/<tool>/captured/<name>.jsonl — a stable, discoverable home for
+    // sessions captured from the wild, kept apart from the curated version dirs.
+    // `MEMSCRIBE_FIXTURES_DIR` overrides the corpus root (used by tests to stay
+    // hermetic — capture into a tempdir instead of the repo's `fixtures/`).
+    let fixtures_root = std::env::var_os("MEMSCRIBE_FIXTURES_DIR")
+        .map(PathBuf::from)
+        .unwrap_or_else(memscribe_testkit::golden::fixtures_dir);
+    let dest_dir = fixtures_root.join(kind.as_str()).join("captured");
+    std::fs::create_dir_all(&dest_dir)
+        .with_context(|| format!("creating fixture dir {}", dest_dir.display()))?;
+
+    // Preserve the source extension (a `.zst` rollout stays compressed) so the
+    // reader treats the captured fixture exactly like the original.
+    let ext = session
+        .extension()
+        .and_then(|e| e.to_str())
+        .unwrap_or("jsonl");
+    let dest = dest_dir.join(format!("{base}.{ext}"));
+    std::fs::write(&dest, &raw).with_context(|| format!("writing fixture {}", dest.display()))?;
+
+    if with_nodes {
+        // Prepare the nodes from the decompressed bytes and write them as NDJSON
+        // next to the raw transcript.
+        let nodes = memscribe_testkit::prepare_nodes(kind, &bytes, session);
+        let node_count = nodes.len();
+        let nodes_path = dest_dir.join(format!("{base}.nodes.ndjson"));
+        let mut out = String::new();
+        for n in &nodes {
+            out.push_str(&serde_json::to_string(n).context("node failed to serialize")?);
+            out.push('\n');
+        }
+        std::fs::write(&nodes_path, out)
+            .with_context(|| format!("writing nodes {}", nodes_path.display()))?;
+        println!(
+            "captured {} → {} ({} node(s) → {})",
+            session.display(),
+            dest.display(),
+            node_count,
+            nodes_path.display()
+        );
+    } else {
+        println!("captured {} → {}", session.display(), dest.display());
+    }
+    Ok(())
+}
+
+/// Parse one fixture and assert the whitepaper §8.3 conformance invariants on
+/// the **event** stream (the real losslessness floor — note that some scenarios,
+/// e.g. `tool_failure` and `no_commitment_marker`, correctly elevate *zero*
+/// prepared nodes, so a non-empty node count is deliberately NOT required), plus
+/// clean node serialization. Returns the prepared-node count on success.
+fn verify_one(kind: Option<SourceKind>, path: &Path) -> Result<usize> {
+    use memscribe_testkit::invariants::{
+        check_determinism, check_lossless, check_monotonic_seq, check_unique_event_ids,
+    };
+    use memscribe_testkit::{count_nonblank_lines, parse_events, prepare_nodes};
+
+    // Resolve the adapter: explicit tool slug from the directory, else infer.
+    let kind = match kind {
+        Some(k) => k,
+        None => infer_source(path)
+            .ok_or_else(|| anyhow!("could not resolve a tool for {}", path.display()))?,
+    };
+    if memscribe_adapters::adapter_for(kind).is_none() {
+        bail!("the `{kind}` adapter is not compiled in");
+    }
+
+    // Read decompressed bytes so the testkit (which works on raw bytes) sees the
+    // same content the one-shot reader would, including `.zst` rollouts.
+    let bytes = read_decompressed(path)?;
+
+    // Events: the lossless, deterministic, monotonic, deduped stream.
+    let events = parse_events(kind, &bytes, path);
+    check_lossless(count_nonblank_lines(&bytes), &events).map_err(|e| anyhow!(e))?;
+    check_monotonic_seq(&events).map_err(|e| anyhow!(e))?;
+    check_unique_event_ids(&events).map_err(|e| anyhow!(e))?;
+    // Determinism: a second parse must be byte-identical.
+    let events2 = parse_events(kind, &bytes, path);
+    check_determinism(&events, &events2).map_err(|e| anyhow!(e))?;
+
+    // Nodes: prepare and confirm every node serializes cleanly (round-trip floor).
+    let nodes = prepare_nodes(kind, &bytes, path);
+    for n in &nodes {
+        serde_json::to_string(n).context("node failed to serialize")?;
+    }
+    Ok(nodes.len())
+}
+
+/// Read a transcript's bytes, transparently decompressing a `.zst` rollout.
+fn read_decompressed(path: &Path) -> Result<Vec<u8>> {
+    let raw = std::fs::read(path).with_context(|| format!("reading {}", path.display()))?;
+    if path.extension().and_then(|e| e.to_str()) == Some("zst") {
+        zstd::decode_all(&raw[..]).with_context(|| format!("decompressing {}", path.display()))
+    } else {
+        Ok(raw)
+    }
+}
+
+// ---------------------------------------------------------------------------
+// watch
+// ---------------------------------------------------------------------------
+
+/// The set of file extensions transcripts use across tools. Discovery is by
+/// extension because adapter `discover()` is a stub in the initial model.
+const TRANSCRIPT_EXTS: &[&str] = &["jsonl", "json", "ndjson", "zst"];
+
+/// Run the live capture daemon: discover transcripts for the requested `tools`
+/// under `roots`, tail them with the crash-safe [`LiveTailer`], prepare each
+/// appended batch through the pipeline, and write the prepared nodes to `sink`.
+///
+/// Responds to Ctrl-C cleanly (the tail loop breaks on the next tick and the
+/// sink is flushed) and never panics — discovery/IO problems are logged and the
+/// affected file is skipped.
+fn cmd_watch(
+    tools: &[String],
+    sink_kind: &str,
+    out: &Path,
+    roots: &[PathBuf],
+    once: bool,
+    config: Option<&Path>,
+) -> Result<()> {
+    // Load the config (if any). It feeds: (a) the tool set, (b) per-tool path
+    // overrides + home/project_filter into a DiscoverCfg, (c) the gate rule
+    // table, (d) the redaction patterns, and (e) the sink selection. A missing
+    // section falls back to the compiled default, so a partial config is fine.
+    let cfg = match config {
+        Some(p) => {
+            Some(Config::load(p).with_context(|| format!("loading config {}", p.display()))?)
+        }
+        None => None,
+    };
+
+    // Surface the parsed ingest cadence (retention/cadence is parsed-and-stored;
+    // the tailer cadence wiring lands incrementally, so log it for now rather
+    // than silently dropping it).
+    if let Some(c) = &cfg {
+        tracing::info!(
+            poll_interval_secs = c.poll_interval().as_secs(),
+            resume_from_offset = c.resume_from_offset(),
+            backfill_days = c.backfill_days(),
+            memdb_endpoint = c.memdb_endpoint().unwrap_or("-"),
+            "memscribe.toml loaded"
+        );
+    }
+
+    // Build the pipeline from config: a config-driven gate + redactor when given,
+    // else the compiled defaults (redaction stays on either way).
+    let pipeline = match &cfg {
+        Some(c) => DefaultPipeline::new()
+            .with_gate(c.build_gate()?)
+            .with_redactor(Some(c.build_redactor()?)),
+        None => DefaultPipeline::new(),
+    };
+
+    // Resolve the set of tools we will accept. Precedence: CLI `--tools`, then
+    // the config's `[capture].tools`, then every compiled adapter.
+    let tool_slugs: Vec<String> = if !tools.iter().any(|t| !t.trim().is_empty()) {
+        cfg.as_ref()
+            .map(|c| c.capture_tools().to_vec())
+            .unwrap_or_default()
+    } else {
+        tools.to_vec()
+    };
+    let wanted = resolve_wanted_tools(&tool_slugs)?;
+
+    // The discovery config the adapters consult (per-tool overrides + home +
+    // project_filter). When no config is present this is the default cfg.
+    let disc_cfg = cfg.as_ref().map(Config::discover_cfg).unwrap_or_default();
+
+    // Resolve scan roots: explicit `--root`s, else the config/`$HOME` home dir.
+    let roots: Vec<PathBuf> = if roots.is_empty() {
+        vec![disc_cfg.home_dir()]
+    } else {
+        roots.to_vec()
+    };
+
+    // Discover candidate transcripts and bucket each to a tool. Two sources are
+    // merged: (1) the extension walk over `roots` (the default discovery), and
+    // (2) each wanted adapter's own `discover(&disc_cfg)` — which is what
+    // actually consumes the config's per-tool path overrides.
+    let mut targets: Vec<(SourceKind, PathBuf)> = Vec::new();
+    for root in &roots {
+        for disc in find_transcripts(root, TRANSCRIPT_EXTS) {
+            if let Some(kind) = infer_source(&disc.path) {
+                if wanted.contains(&kind) {
+                    targets.push((kind, disc.path));
+                }
+            }
+        }
+    }
+    for kind in &wanted {
+        if let Some(adapter) = memscribe_adapters::adapter_for(*kind) {
+            for handle in adapter.discover(&disc_cfg) {
+                targets.push((handle.source, handle.path));
+            }
+        }
+    }
+    // Deterministic order, deduped by path (the same file is never tailed twice
+    // even if two roots overlap). `SourceKind` is not `Ord`, so key on the path.
+    targets.sort_by(|a, b| a.1.cmp(&b.1));
+    targets.dedup_by(|a, b| a.1 == b.1);
+
+    let tool_list: Vec<&str> = wanted.iter().map(SourceKind::as_str).collect();
+    eprintln!(
+        "memscribe watch: {} tool(s) [{}], {} root(s), {} transcript(s) discovered",
+        wanted.len(),
+        tool_list.join(","),
+        roots.len(),
+        targets.len()
+    );
+
+    // Sink selection: the config's `[sink]` target/path, else the CLI flags.
+    let (sink_kind, sink_out): (String, PathBuf) = match &cfg {
+        Some(c) => (
+            c.sink_target().to_string(),
+            c.sink_out_path().unwrap_or_else(|| out.to_path_buf()),
+        ),
+        None => (sink_kind.to_string(), out.to_path_buf()),
+    };
+    let mut sink = build_sink(&sink_kind, &sink_out)?;
+
+    if once {
+        // One-shot: drain everything that already exists once, then exit. This
+        // is the documented simplification for the initial model.
+        return watch_once(&targets, &pipeline, sink.as_mut());
+    }
+
+    watch_live(&targets, &pipeline, sink.as_mut())
+}
+
+/// The wanted-tool set from `--tools` (empty = every adapter that is compiled).
+fn resolve_wanted_tools(tools: &[String]) -> Result<Vec<SourceKind>> {
+    if tools.is_empty() {
+        return Ok(memscribe_adapters::all_adapters()
+            .iter()
+            .map(|a| a.source_kind())
+            .collect());
+    }
+    let mut out = Vec::new();
+    for t in tools {
+        let t = t.trim();
+        if t.is_empty() {
+            continue;
+        }
+        let kind = SourceKind::parse(t).ok_or_else(|| anyhow!("unknown tool `{t}`"))?;
+        if memscribe_adapters::adapter_for(kind).is_none() {
+            bail!("the `{kind}` adapter is not compiled into this build");
+        }
+        if !out.contains(&kind) {
+            out.push(kind);
+        }
+    }
+    if out.is_empty() {
+        bail!("no valid tools requested");
+    }
+    Ok(out)
+}
+
+/// Build the chosen sink. `ndjson` (default) writes one JSON node per line to a
+/// file or stdout; `sqlite`/`memdb` are recognized but gated on their features.
+fn build_sink(kind: &str, out: &Path) -> Result<Box<dyn Sink>> {
+    match kind {
+        "ndjson" => {
+            if out == Path::new("-") {
+                Ok(Box::new(NdjsonSink::stdout()))
+            } else {
+                Ok(Box::new(NdjsonSink::file(out).with_context(|| {
+                    format!("opening ndjson sink at {}", out.display())
+                })?))
+            }
+        }
+        "sqlite" => {
+            if out == Path::new("-") {
+                bail!("the `sqlite` sink needs a file path; pass `--out <file.sqlite>`");
+            }
+            Ok(Box::new(memscribe_sink::SqliteSink::open(out).map_err(
+                |e| anyhow!("opening sqlite sink at {}: {e}", out.display()),
+            )?))
+        }
+        "memdb" => bail!(
+            "the `memdb` sink is not compiled into this build (Memtrace enables it via the \
+             `memdb-sink` feature)"
+        ),
+        other => bail!("unknown sink `{other}`; expected `ndjson`, `sqlite`, or `memdb`"),
+    }
+}
+
+/// One-shot: read every existing transcript fully and emit its prepared nodes
+/// through the (possibly config-driven) `pipeline`.
+fn watch_once(
+    targets: &[(SourceKind, PathBuf)],
+    pipeline: &DefaultPipeline,
+    sink: &mut dyn Sink,
+) -> Result<()> {
+    let mut total = 0usize;
+    for (kind, path) in targets {
+        match prepare_file(*kind, path, pipeline) {
+            Ok(nodes) => {
+                for n in &nodes {
+                    sink.emit(n)?;
+                }
+                total += nodes.len();
+                tracing::info!(tool = %kind, path = %path.display(), nodes = nodes.len(), "drained");
+            }
+            Err(e) => {
+                tracing::warn!(tool = %kind, path = %path.display(), error = %e, "skipping transcript");
+            }
+        }
+    }
+    sink.flush()?;
+    eprintln!(
+        "memscribe watch --once: emitted {total} node(s) from {} transcript(s)",
+        targets.len()
+    );
+    Ok(())
+}
+
+/// Live tailing: register every discovered transcript with one [`LiveTailer`]
+/// (backed by a persistent SQLite cursor so restarts resume), drain what already
+/// exists, then loop emitting prepared nodes for each appended batch until
+/// Ctrl-C. Each batch is routed to its file's adapter by path.
+fn watch_live(
+    targets: &[(SourceKind, PathBuf)],
+    pipeline: &DefaultPipeline,
+    sink: &mut dyn Sink,
+) -> Result<()> {
+    // A path → tool map so a tailer batch (which carries provenance paths) is
+    // routed to the right adapter.
+    use std::collections::HashMap;
+    let by_path: HashMap<PathBuf, SourceKind> =
+        targets.iter().map(|(k, p)| (p.clone(), *k)).collect();
+
+    // Persistent offset cursor under the user's state dir, so a restart resumes
+    // exactly where it left off (zero loss, zero dup — whitepaper §8.5).
+    let store = open_cursor_store()?;
+    let mut tailer = LiveTailer::new(store, Duration::from_millis(200))
+        .map_err(|e| anyhow!("creating the live tailer: {e}"))?;
+    tailer
+        .watch_paths(by_path.keys())
+        .map_err(|e| anyhow!("registering transcripts to watch: {e}"))?;
+
+    // Clean Ctrl-C: flip a flag the tail loop checks on each tick, then break.
+    let stop = Arc::new(AtomicBool::new(false));
+    {
+        let stop = Arc::clone(&stop);
+        // Best-effort: if a handler is already installed (e.g. in tests), don't
+        // make that fatal.
+        let _ = ctrlc::set_handler(move || {
+            stop.store(true, Ordering::SeqCst);
+        });
+    }
+
+    // Drain pre-existing content once before going live.
+    let pre = tailer.poll_existing();
+    emit_batch(&pre, &by_path, pipeline, sink)?;
+    sink.flush()?;
+    eprintln!(
+        "memscribe watch: tailing {} transcript(s); press Ctrl-C to stop",
+        by_path.len()
+    );
+
+    // The blocking tail loop. We drive `poll` directly (rather than `run`) so the
+    // `stop` flag is checked on EVERY tick — including idle timeouts, which `run`
+    // skips. That is what makes Ctrl-C responsive on an otherwise-quiet tree:
+    // without it, an idle watcher would never observe the stop signal until the
+    // next append. A short tick keeps shutdown latency at one tick.
+    let tick = Duration::from_millis(200);
+    loop {
+        if stop.load(Ordering::SeqCst) {
+            break;
+        }
+        match tailer.poll(tick) {
+            // `Some(empty)` is a timeout — loop back and re-check `stop`.
+            Some(batch) if batch.is_empty() => continue,
+            Some(batch) => {
+                emit_batch(&batch, &by_path, pipeline, sink)?;
+                sink.flush()?;
+            }
+            // The watcher shut down (sender dropped) — end of stream.
+            None => break,
+        }
+    }
+
+    sink.flush()?;
+    eprintln!("memscribe watch: stopped cleanly");
+    Ok(())
+}
+
+/// Open the persistent cursor store under the OS state dir, creating parents.
+fn open_cursor_store() -> Result<SqliteOffsetStore> {
+    let base = std::env::var_os("HOME")
+        .map(PathBuf::from)
+        .unwrap_or_else(|| PathBuf::from("."));
+    let dir = base.join(".local/state/memscribe");
+    std::fs::create_dir_all(&dir)
+        .with_context(|| format!("creating cursor dir {}", dir.display()))?;
+    let db = dir.join("cursors.sqlite");
+    SqliteOffsetStore::open(&db).map_err(|e| anyhow!("opening cursor store {}: {e}", db.display()))
+}
+
+/// Group a tailer batch by source file, route each group to its adapter, and
+/// emit the prepared nodes. A record whose file we don't have a tool for is
+/// skipped (logged), never fatal.
+fn emit_batch(
+    batch: &[memscribe_core::RawRecord],
+    by_path: &std::collections::HashMap<PathBuf, SourceKind>,
+    pipeline: &DefaultPipeline,
+    sink: &mut dyn Sink,
+) -> Result<()> {
+    use std::collections::BTreeMap;
+    // Preserve per-file order; BTreeMap keeps deterministic file ordering.
+    let mut grouped: BTreeMap<PathBuf, Vec<memscribe_core::RawRecord>> = BTreeMap::new();
+    for rec in batch {
+        grouped
+            .entry(rec.location.file.clone())
+            .or_default()
+            .push(rec.clone());
+    }
+    for (file, recs) in grouped {
+        let kind = match by_path.get(&file).copied().or_else(|| infer_source(&file)) {
+            Some(k) => k,
+            None => {
+                tracing::warn!(path = %file.display(), "no adapter for tailed file; skipping batch");
+                continue;
+            }
+        };
+        let Some(adapter) = memscribe_adapters::adapter_for(kind) else {
+            continue;
+        };
+        let nodes = pipeline.run_records(adapter.as_ref(), &recs);
+        for n in &nodes {
+            sink.emit(n)?;
+        }
+    }
+    Ok(())
+}
+
+/// Read a whole transcript file and prepare its nodes through `pipeline`
+/// (redaction **on** — the safe default for anything that lands in a sink).
+fn prepare_file(
+    kind: SourceKind,
+    path: &Path,
+    pipeline: &DefaultPipeline,
+) -> Result<Vec<PreparedNode>> {
+    let adapter = memscribe_adapters::adapter_for(kind)
+        .ok_or_else(|| anyhow!("the `{kind}` adapter is not compiled in"))?;
+    let records =
+        memscribe_io::read_records(path).with_context(|| format!("reading {}", path.display()))?;
+    Ok(pipeline.run_records(adapter.as_ref(), &records))
+}
diff --git a/crates/memscribe-cli/tests/cli.rs b/crates/memscribe-cli/tests/cli.rs
new file mode 100644
index 0000000..38a968f
--- /dev/null
+++ b/crates/memscribe-cli/tests/cli.rs
@@ -0,0 +1,192 @@
+//! Black-box integration tests for the `memscribe` binary (assert_cmd).
+//!
+//! These drive the real compiled CLI end-to-end against the repo's fixtures and
+//! temp files, asserting the user-visible contracts: `parse` emits valid NDJSON,
+//! `--as` forces an adapter, `redact` strips a seeded secret, `hook` consumes
+//! stdin and exits 0, unknown tools fail cleanly, and `--help` works.
+
+use assert_cmd::Command;
+use predicates::prelude::*;
+use std::path::{Path, PathBuf};
+
+/// The workspace root (three levels up from this test crate's manifest dir:
+/// `crates/memscribe-cli` → `crates` → workspace).
+fn workspace_root() -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR"))
+        .ancestors()
+        .nth(2)
+        .map(Path::to_path_buf)
+        .expect("workspace root resolvable from CARGO_MANIFEST_DIR")
+}
+
+/// A known-good Claude Code fixture (its path contains `claude`, so the tool is
+/// inferable without `--as`).
+fn claude_fixture() -> PathBuf {
+    workspace_root().join("fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl")
+}
+
+fn memscribe() -> Command {
+    Command::cargo_bin("memscribe").expect("the `memscribe` binary builds")
+}
+
+/// Every non-blank stdout line must be a parseable JSON object.
+fn assert_all_lines_are_json_objects(stdout: &[u8]) {
+    let text = String::from_utf8(stdout.to_vec()).expect("stdout is valid UTF-8");
+    let mut lines = 0usize;
+    for line in text.lines() {
+        if line.trim().is_empty() {
+            continue;
+        }
+        lines += 1;
+        let value: serde_json::Value = serde_json::from_str(line)
+            .unwrap_or_else(|e| panic!("line is not JSON: {line:?}: {e}"));
+        assert!(
+            value.is_object(),
+            "each NDJSON line must be a JSON object, got: {line:?}"
+        );
+    }
+    assert!(lines > 0, "parse must emit at least one NDJSON node");
+}
+
+#[test]
+fn parse_fixture_emits_valid_ndjson() {
+    let assert = memscribe()
+        .arg("parse")
+        .arg(claude_fixture())
+        .assert()
+        .success();
+    assert_all_lines_are_json_objects(&assert.get_output().stdout);
+}
+
+#[test]
+fn parse_with_explicit_as_tool() {
+    // Force the adapter rather than inferring; output must still be valid NDJSON.
+    let assert = memscribe()
+        .arg("parse")
+        .arg("--as")
+        .arg("claude_code")
+        .arg(claude_fixture())
+        .assert()
+        .success();
+    assert_all_lines_are_json_objects(&assert.get_output().stdout);
+}
+
+#[test]
+fn parse_unknown_tool_errors_cleanly() {
+    // An unrecognized `--as` value must fail with a clear message, not a panic.
+    memscribe()
+        .arg("parse")
+        .arg("--as")
+        .arg("not-a-real-tool")
+        .arg(claude_fixture())
+        .assert()
+        .failure()
+        .stderr(predicate::str::contains("unknown tool"));
+}
+
+#[test]
+fn redact_strips_a_seeded_secret() {
+    // Seed a realistically-shaped Anthropic key (the default pattern requires
+    // 16+ trailing chars) and assert the redactor removes it from stdout.
+    let secret = "sk-ant-api03-AAAAAAAAAAAAAAAAAAAAAAAA";
+    let body = format!("here is a key: {secret}\nplain second line\n");
+
+    let dir = tempfile::tempdir().unwrap();
+    let file = dir.path().join("leaky.txt");
+    std::fs::write(&file, &body).unwrap();
+
+    let assert = memscribe().arg("redact").arg(&file).assert().success();
+    let stdout = String::from_utf8(assert.get_output().stdout.clone()).unwrap();
+
+    assert!(
+        !stdout.contains(secret),
+        "the seeded secret must be gone from stdout, got:\n{stdout}"
+    );
+    // Non-secret content survives.
+    assert!(
+        stdout.contains("plain second line"),
+        "non-secret content must be preserved"
+    );
+}
+
+#[test]
+fn hook_reads_stdin_and_exits_zero() {
+    // The hook handler must consume stdin and exit 0 immediately — never block.
+    memscribe()
+        .arg("hook")
+        .write_stdin(
+            r#"{"session_id":"s1","transcript_path":"/tmp/x.jsonl","hook_event_name":"PostToolUse","tool_name":"Edit"}"#,
+        )
+        .assert()
+        .success();
+}
+
+#[test]
+fn hook_with_invalid_json_still_exits_zero() {
+    // Invalid JSON on stdin must not crash the hook — it still exits 0.
+    memscribe()
+        .arg("hook")
+        .write_stdin("this is not json")
+        .assert()
+        .success();
+}
+
+#[test]
+fn help_works() {
+    memscribe()
+        .arg("--help")
+        .assert()
+        .success()
+        .stdout(predicate::str::contains("memscribe"))
+        .stdout(predicate::str::contains("watch"))
+        .stdout(predicate::str::contains("parse"));
+}
+
+#[test]
+fn verify_passes_over_fixtures() {
+    // `verify` parses every fixture and prints a per-tool PASS/FAIL table; with
+    // the shipped (green) fixtures it must succeed and report PASS.
+    memscribe()
+        .arg("verify")
+        .assert()
+        .success()
+        .stdout(predicate::str::contains("PASS"))
+        .stdout(predicate::str::contains("claude_code"));
+}
+
+#[test]
+fn watch_once_drains_a_discovered_transcript() {
+    // `watch --once --root <dir>` discovers the seeded transcript, parses it, and
+    // emits prepared nodes to stdout (the ndjson sink). It must exit cleanly and
+    // not hang (no live tail loop in `--once`).
+    let dir = tempfile::tempdir().unwrap();
+    // The filename carries `claude` so the tool is inferable.
+    let transcript = dir.path().join("claude-session.jsonl");
+    std::fs::copy(claude_fixture(), &transcript).unwrap();
+
+    let assert = memscribe()
+        .arg("watch")
+        .arg("--once")
+        .arg("--tools")
+        .arg("claude_code")
+        .arg("--root")
+        .arg(dir.path())
+        .assert()
+        .success();
+
+    // Whatever it emitted must be valid NDJSON (a non-empty transcript yields
+    // at least one node).
+    assert_all_lines_are_json_objects(&assert.get_output().stdout);
+}
+
+#[test]
+fn watch_unknown_tool_errors_cleanly() {
+    memscribe()
+        .arg("watch")
+        .arg("--once")
+        .arg("--tools")
+        .arg("not-a-real-tool")
+        .assert()
+        .failure()
+        .stderr(predicate::str::contains("unknown tool"));
+}
diff --git a/crates/memscribe-cli/tests/cli_config.rs b/crates/memscribe-cli/tests/cli_config.rs
new file mode 100644
index 0000000..0eae596
--- /dev/null
+++ b/crates/memscribe-cli/tests/cli_config.rs
@@ -0,0 +1,262 @@
+//! Integration tests for the `memscribe.toml` config loader and
+//! `verify --capture` (assert_cmd). Hermetic: every test writes its inputs into
+//! a `tempdir` and asserts the user-visible behavior of the real compiled CLI.
+
+use assert_cmd::Command;
+use predicates::prelude::*;
+use std::path::{Path, PathBuf};
+
+fn memscribe() -> Command {
+    Command::cargo_bin("memscribe").expect("the `memscribe` binary builds")
+}
+
+/// The workspace root (three levels up from this test crate's manifest dir).
+fn workspace_root() -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR"))
+        .ancestors()
+        .nth(2)
+        .map(Path::to_path_buf)
+        .expect("workspace root resolvable from CARGO_MANIFEST_DIR")
+}
+
+/// A known-good Claude Code fixture (path contains `claude`, so it is inferable).
+fn claude_fixture() -> PathBuf {
+    workspace_root().join("fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl")
+}
+
+/// One Claude-Code-shaped user turn whose text carries `body` verbatim. The path
+/// inference and adapter both recognize this as `claude_code`.
+fn claude_user_turn(body: &str) -> String {
+    // A single, minimal JSONL line — a user message with plain-string content.
+    format!(
+        r#"{{"type":"user","uuid":"11111111-1111-4111-8111-111111111111","parentUuid":null,"timestamp":"2026-06-22T10:00:00.000Z","sessionId":"sess-cfg-001","cwd":"/tmp/proj","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{{"role":"user","content":{}}}}}"#,
+        serde_json::to_string(body).unwrap()
+    )
+}
+
+/// Collect the emitted NDJSON nodes from a successful run's stdout.
+fn nodes_from_stdout(stdout: &[u8]) -> Vec<serde_json::Value> {
+    let text = String::from_utf8(stdout.to_vec()).expect("stdout is valid UTF-8");
+    text.lines()
+        .filter(|l| !l.trim().is_empty())
+        .map(|l| serde_json::from_str(l).expect("each NDJSON line parses"))
+        .collect()
+}
+
+/// `watch --once --config <toml>` must consume the config: a CUSTOM commitment
+/// rule (replacing the default table) gates a turn the defaults would ignore, and
+/// a CUSTOM redaction pattern (replacing the default set) strips a secret the
+/// defaults would miss. This is the end-to-end round-trip of TASK A.
+#[test]
+fn config_custom_gate_and_redaction_round_trip() {
+    let dir = tempfile::tempdir().unwrap();
+
+    // A user turn whose only commitment marker is the custom word `banana`
+    // (no default verb fires on it), carrying a secret only the custom pattern
+    // matches.
+    let transcript = dir.path().join("claude-session.jsonl");
+    std::fs::write(
+        &transcript,
+        format!(
+            "{}\n",
+            claude_user_turn("we will banana the orders service; key BANANA-1234")
+        ),
+    )
+    .unwrap();
+
+    // A config that REPLACES the default gate + redaction tables.
+    let config = dir.path().join("memscribe.toml");
+    std::fs::write(
+        &config,
+        r#"
+[capture]
+tools = ["claude_code"]
+
+[[gate.rules]]
+id = "custom.banana"
+category = "decision_verb"
+pattern = "banana"
+
+[redact]
+no_content = false
+
+[[redact.patterns]]
+label = "banana_token"
+pattern = "BANANA-[0-9]{4}"
+"#,
+    )
+    .unwrap();
+
+    let assert = memscribe()
+        .arg("watch")
+        .arg("--once")
+        .arg("--config")
+        .arg(&config)
+        .arg("--root")
+        .arg(dir.path())
+        .assert()
+        .success();
+
+    let nodes = nodes_from_stdout(&assert.get_output().stdout);
+    let blob = serde_json::to_string(&nodes).unwrap();
+
+    // The custom gate admitted the `banana` turn → at least one node was emitted.
+    assert!(
+        !nodes.is_empty(),
+        "the custom gate rule must admit the `banana` turn; got no nodes"
+    );
+    // The custom redaction pattern replaced the secret with its label.
+    assert!(
+        blob.contains("[REDACTED:banana_token]"),
+        "the custom redaction pattern must fire; nodes:\n{blob}"
+    );
+    // And the raw secret must be gone.
+    assert!(
+        !blob.contains("BANANA-1234"),
+        "the raw custom secret must not survive; nodes:\n{blob}"
+    );
+}
+
+/// The config's `[capture].tools` selects the tool set when `--tools` is omitted.
+/// With a `claude_code`-only config and a discoverable claude transcript, the run
+/// drains it and emits valid NDJSON.
+#[test]
+fn config_capture_tools_selects_adapter() {
+    let dir = tempfile::tempdir().unwrap();
+    let transcript = dir.path().join("claude-session.jsonl");
+    std::fs::copy(claude_fixture(), &transcript).unwrap();
+
+    let config = dir.path().join("memscribe.toml");
+    std::fs::write(
+        &config,
+        r#"
+[capture]
+tools = ["claude_code"]
+"#,
+    )
+    .unwrap();
+
+    let assert = memscribe()
+        .arg("watch")
+        .arg("--once")
+        .arg("--config")
+        .arg(&config)
+        .arg("--root")
+        .arg(dir.path())
+        .assert()
+        .success();
+
+    let nodes = nodes_from_stdout(&assert.get_output().stdout);
+    assert!(
+        !nodes.is_empty(),
+        "the discovered transcript must yield nodes"
+    );
+}
+
+/// A malformed config must fail with a clear message, not a panic.
+#[test]
+fn config_invalid_pattern_fails_cleanly() {
+    let dir = tempfile::tempdir().unwrap();
+    let config = dir.path().join("memscribe.toml");
+    std::fs::write(
+        &config,
+        r#"
+[[gate.rules]]
+id = "bad"
+category = "imperative"
+pattern = "("
+"#,
+    )
+    .unwrap();
+
+    memscribe()
+        .arg("watch")
+        .arg("--once")
+        .arg("--config")
+        .arg(&config)
+        .arg("--root")
+        .arg(dir.path())
+        .assert()
+        .failure();
+}
+
+/// `verify --capture <session> --as <tool>` must snapshot the transcript into a
+/// new fixture under `<corpus>/<tool>/captured/<name>.<ext>`. The corpus root is
+/// redirected to a tempdir via `MEMSCRIBE_FIXTURES_DIR`, so the test is fully
+/// hermetic and never touches the repo's real `fixtures/`.
+#[test]
+fn verify_capture_writes_a_fixture() {
+    let corpus = tempfile::tempdir().unwrap();
+    let unique = "sample_session";
+
+    let captured_dir = corpus.path().join("claude_code").join("captured");
+    let dest = captured_dir.join(format!("{unique}.jsonl"));
+    let nodes_dest = captured_dir.join(format!("{unique}.nodes.ndjson"));
+
+    memscribe()
+        .env("MEMSCRIBE_FIXTURES_DIR", corpus.path())
+        .arg("verify")
+        .arg("--capture")
+        .arg(claude_fixture())
+        .arg("--as")
+        .arg("claude_code")
+        .arg("--name")
+        .arg(unique)
+        .arg("--with-nodes")
+        .assert()
+        .success()
+        .stdout(predicate::str::contains("captured"));
+
+    // The raw transcript fixture was written, byte-for-byte equal to the source.
+    assert!(
+        dest.is_file(),
+        "captured fixture must exist at {}",
+        dest.display()
+    );
+    let written = std::fs::read(&dest).unwrap();
+    let original = std::fs::read(claude_fixture()).unwrap();
+    assert_eq!(
+        written, original,
+        "captured fixture must be a verbatim copy"
+    );
+
+    // `--with-nodes` wrote the prepared nodes alongside, as valid NDJSON.
+    assert!(
+        nodes_dest.is_file(),
+        "prepared nodes must be written at {}",
+        nodes_dest.display()
+    );
+    let nodes_text = std::fs::read_to_string(&nodes_dest).unwrap();
+    let mut lines = 0usize;
+    for line in nodes_text.lines() {
+        if line.trim().is_empty() {
+            continue;
+        }
+        lines += 1;
+        let v: serde_json::Value = serde_json::from_str(line).expect("nodes line parses");
+        assert!(v.is_object(), "each captured node is a JSON object");
+    }
+    assert!(
+        lines > 0,
+        "the captured sample must yield at least one node"
+    );
+}
+
+/// `verify --capture` on a session it cannot resolve must fail cleanly (it never
+/// admits an unresolvable sample to the corpus).
+#[test]
+fn verify_capture_unknown_tool_fails() {
+    let dir = tempfile::tempdir().unwrap();
+    let session = dir.path().join("mystery.jsonl");
+    std::fs::write(&session, "{}\n").unwrap();
+
+    memscribe()
+        .arg("verify")
+        .arg("--capture")
+        .arg(&session)
+        .arg("--as")
+        .arg("not-a-real-tool")
+        .assert()
+        .failure()
+        .stderr(predicate::str::contains("unknown tool"));
+}
diff --git a/crates/memscribe-core/Cargo.toml b/crates/memscribe-core/Cargo.toml
new file mode 100644
index 0000000..be70cd7
--- /dev/null
+++ b/crates/memscribe-core/Cargo.toml
@@ -0,0 +1,22 @@
+[package]
+name = "memscribe-core"
+version.workspace = true
+edition.workspace = true
+rust-version.workspace = true
+license.workspace = true
+repository.workspace = true
+description = "The deterministic, zero-LLM contract and pipeline for Memscribe: model, traits, gate, segmenter, binder, node-prep."
+
+[lints]
+workspace = true
+
+[dependencies]
+serde = { workspace = true }
+serde_json = { workspace = true }
+time = { workspace = true }
+thiserror = { workspace = true }
+blake3 = { workspace = true }
+regex = { workspace = true }
+
+[dev-dependencies]
+proptest = { workspace = true }
diff --git a/crates/memscribe-core/src/adapter.rs b/crates/memscribe-core/src/adapter.rs
new file mode 100644
index 0000000..27697df
--- /dev/null
+++ b/crates/memscribe-core/src/adapter.rs
@@ -0,0 +1,179 @@
+//! The adapter layer contract (whitepaper §5).
+//!
+//! Each tool implements [`TranscriptAdapter`]: where its logs live, and how to
+//! turn one raw record into normalized events. Parsers are **version-tolerant**:
+//! they pattern-match on the fields they need and route anything unrecognized to
+//! [`crate::model::EventKind::Unknown`] rather than failing the stream. A parser
+//! **must never panic**.
+
+use crate::error::ParseError;
+use crate::model::{CaptureEvent, ProjectRef, SourceKind, SourceLocation};
+use std::collections::{HashMap, HashSet};
+use std::path::PathBuf;
+
+/// A raw, unparsed record as produced by a Source: a JSONL line, a hook stdin
+/// blob, or an OTLP record — carrying the provenance needed to replay it.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct RawRecord {
+    /// The raw bytes of the record.
+    pub bytes: Vec<u8>,
+    /// Where the record came from.
+    pub location: SourceLocation,
+}
+
+impl RawRecord {
+    /// Construct a raw record.
+    pub fn new(bytes: impl Into<Vec<u8>>, location: SourceLocation) -> Self {
+        RawRecord {
+            bytes: bytes.into(),
+            location,
+        }
+    }
+
+    /// The record as UTF-8 text, if valid.
+    #[must_use]
+    pub fn as_str(&self) -> Option<&str> {
+        std::str::from_utf8(&self.bytes).ok()
+    }
+
+    /// Construct from a string and a location (convenience for tests).
+    pub fn from_line(line: &str, location: SourceLocation) -> Self {
+        RawRecord::new(line.as_bytes().to_vec(), location)
+    }
+}
+
+/// Mutable per-session context threaded through parsing. It assigns the
+/// monotonic `seq`, dedups by `event_id`, resolves tool-call/result pairing by
+/// `call_id`, and carries the session-start project binding.
+#[derive(Clone, Debug, Default)]
+pub struct ParseCtx {
+    /// The session id, set once known.
+    pub session_id: Option<String>,
+    /// The next sequence number to assign.
+    pub next_seq: u64,
+    /// Event ids already emitted (for dedup / idempotency).
+    pub seen_event_ids: HashSet<String>,
+    /// The project binding captured at session start.
+    pub project: Option<ProjectRef>,
+    /// Map of `call_id` → tool name, for pairing calls with results/edits.
+    pub call_names: HashMap<String, String>,
+    /// Map of `call_id` → success flag, from observed tool results.
+    pub call_ok: HashMap<String, bool>,
+}
+
+impl ParseCtx {
+    /// A fresh context.
+    #[must_use]
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Allocate the next monotonic sequence number for this session.
+    pub fn alloc_seq(&mut self) -> u64 {
+        let s = self.next_seq;
+        self.next_seq += 1;
+        s
+    }
+
+    /// Record an `event_id` as seen; returns `true` if it was new (not a dup).
+    pub fn first_seen(&mut self, event_id: &str) -> bool {
+        self.seen_event_ids.insert(event_id.to_string())
+    }
+
+    /// The project ref to stamp on an event, defaulting to the cwd.
+    #[must_use]
+    pub fn project_or_default(&self) -> ProjectRef {
+        self.project
+            .clone()
+            .unwrap_or_else(|| ProjectRef::from_cwd("."))
+    }
+}
+
+/// Where a tool's transcripts live and how to discover them.
+#[derive(Clone, Debug, Default)]
+pub struct DiscoverCfg {
+    /// Override for `$HOME` (used by tests and sandboxes).
+    pub home: Option<PathBuf>,
+    /// Per-tool path overrides (e.g. `CODEX_HOME`, `CLAUDE_CONFIG_DIR`).
+    pub overrides: HashMap<String, PathBuf>,
+    /// Restrict discovery to a single project root, if set.
+    pub project_filter: Option<PathBuf>,
+}
+
+impl DiscoverCfg {
+    /// The effective home directory (override, then `$HOME`, then `.`).
+    #[must_use]
+    pub fn home_dir(&self) -> PathBuf {
+        if let Some(h) = &self.home {
+            return h.clone();
+        }
+        std::env::var_os("HOME")
+            .map(PathBuf::from)
+            .unwrap_or_else(|| PathBuf::from("."))
+    }
+}
+
+/// A discovered transcript file.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct TranscriptHandle {
+    /// The transcript file path.
+    pub path: PathBuf,
+    /// The tool that produced it.
+    pub source: SourceKind,
+    /// A session-id hint derived from the path, if any.
+    pub session_hint: Option<String>,
+    /// Whether the file is zstd-compressed (e.g. Codex cold rollouts).
+    pub compressed: bool,
+}
+
+/// The result of fingerprinting a raw record, so the corpus and runtime can
+/// version-gate the parser.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct SchemaVariant {
+    /// The tool the variant belongs to.
+    pub source: SourceKind,
+    /// A variant identifier (e.g. `claude_code/2.1`, `codex/rollout-v2`).
+    pub variant: String,
+    /// Confidence 0..=100 that the fingerprint is correct.
+    pub confidence: u8,
+}
+
+impl SchemaVariant {
+    /// A variant with full confidence.
+    #[must_use]
+    pub fn certain(source: SourceKind, variant: impl Into<String>) -> Self {
+        SchemaVariant {
+            source,
+            variant: variant.into(),
+            confidence: 100,
+        }
+    }
+
+    /// An unknown variant (zero confidence).
+    #[must_use]
+    pub fn unknown(source: SourceKind) -> Self {
+        SchemaVariant {
+            source,
+            variant: "unknown".to_string(),
+            confidence: 0,
+        }
+    }
+}
+
+/// Each tool implements this trait. See the module docs for the contract.
+pub trait TranscriptAdapter: Send + Sync {
+    /// The tool this adapter handles.
+    fn source_kind(&self) -> SourceKind;
+
+    /// Locate live & historical transcripts (handles globbing, project hashing,
+    /// rotation, `.zst`).
+    fn discover(&self, cfg: &DiscoverCfg) -> Vec<TranscriptHandle>;
+
+    /// Parse ONE raw record into zero or more normalized events. Must never
+    /// panic; unknowns route to [`crate::model::EventKind::Unknown`].
+    fn parse(&self, raw: &RawRecord, ctx: &mut ParseCtx) -> Result<Vec<CaptureEvent>, ParseError>;
+
+    /// Fingerprint a sample record so the corpus and runtime can version-gate
+    /// the parser.
+    fn schema_fingerprint(&self, sample: &RawRecord) -> SchemaVariant;
+}
diff --git a/crates/memscribe-core/src/binder.rs b/crates/memscribe-core/src/binder.rs
new file mode 100644
index 0000000..2c26673
--- /dev/null
+++ b/crates/memscribe-core/src/binder.rs
@@ -0,0 +1,465 @@
+//! The binder: decision ↔ edit, with PROV (whitepaper §3, §8.8).
+//!
+//! For each episode, the binder finds the most recent preceding decision in the
+//! same session whose timestamp is `≤` the episode timestamp, and writes a
+//! [`BindingEdge`] with a [`ProvRecord`] that satisfies the temporal invariant
+//! `t_use ≤ t_gen`. A binding recorded from the deterministic stream is
+//! [`FactStatus::DeterministicallyDerived`]. An episode with no preceding
+//! decision produces no spurious binding.
+
+use crate::node::{BindingEdge, CorrelationTuple, FactStatus, NodeId, ProvRecord, Relation};
+use crate::segmenter::Segmentation;
+use std::collections::HashMap;
+use std::path::PathBuf;
+
+/// Minimal-support threshold for emitting a [`CorrelationTuple`] (§6/§8.8,
+/// "the correct correlation tuple ... when computable"). A decision must bind to
+/// **at least this many episodes in the batch** before a contingency table over
+/// it carries enough mass to be meaningful; below it, `correlation` is left
+/// `None`. A single degenerate arc (one decision → one episode) is therefore
+/// never assigned a correlation.
+const MIN_CORRELATION_SUPPORT: usize = 2;
+
+/// Decimal places every correlation field is rounded to, so the tuple is a
+/// byte-stable function of the integer contingency counts (no float-formatting
+/// drift across runs).
+const CORRELATION_DECIMALS: i32 = 6;
+
+/// The binder stage.
+pub trait Binder {
+    /// Produce binding edges from a segmentation.
+    fn bind(&self, seg: &Segmentation) -> Vec<BindingEdge>;
+}
+
+/// The default deterministic binder.
+#[derive(Debug, Default)]
+pub struct DefaultBinder;
+
+impl Binder for DefaultBinder {
+    fn bind(&self, seg: &Segmentation) -> Vec<BindingEdge> {
+        let mut edges = Vec::new();
+        // Parallel to `edges`: the bound decision's id and the episode path, so a
+        // second pass can compute the co-occurrence contingency table.
+        let mut arcs: Vec<(NodeId, PathBuf)> = Vec::new();
+
+        for ep in &seg.episodes {
+            // Find the latest decision in the same session that precedes the
+            // episode in time (t_use ≤ t_gen). A decision that was rewound away
+            // or compacted out (`superseded_by` set) no longer governs current
+            // edits, so it is skipped — the binding falls through to the most
+            // recent *non-superseded* preceding decision instead (§8.2).
+            let mut best: Option<&crate::segmenter::DecisionCandidate> = None;
+            for dec in &seg.decisions {
+                if dec.session_id != ep.session_id {
+                    continue;
+                }
+                if dec.record.superseded_by.is_some() {
+                    continue;
+                }
+                if dec.timestamp > ep.timestamp {
+                    continue;
+                }
+                // Most-recent-preceding wins; on a timestamp tie the higher
+                // turn_seq (later in the stream) wins, so the result is a
+                // deterministic function of the input regardless of iteration
+                // order.
+                let better = match best {
+                    None => true,
+                    Some(b) => (dec.timestamp, dec.turn_seq) > (b.timestamp, b.turn_seq),
+                };
+                if better {
+                    best = Some(dec);
+                }
+            }
+
+            let Some(dec) = best else { continue };
+
+            arcs.push((dec.node_id.clone(), ep.episode.path.clone()));
+            edges.push(BindingEdge {
+                from: dec.node_id.clone(),
+                to: ep.node_id.clone(),
+                relation: Relation::Produced,
+                prov: ProvRecord {
+                    used_session: dec.session_id.clone(),
+                    used_decision: Some(dec.node_id.clone()),
+                    was_generated_by_session: ep.session_id.clone(),
+                    t_use: dec.timestamp,
+                    t_gen: ep.timestamp,
+                },
+                fact_status: FactStatus::DeterministicallyDerived,
+                correlation: None,
+            });
+        }
+
+        // Second pass: compute each edge's correlation tuple from the batch-wide
+        // co-occurrence contingency table (§6/§8.8). `correlation` stays `None`
+        // below the minimal-support threshold ("when computable").
+        attach_correlations(&mut edges, &arcs);
+
+        // Deterministic ordering, independent of episode/decision discovery order.
+        edges.sort_by(|a, b| a.from.cmp(&b.from).then_with(|| a.to.cmp(&b.to)));
+        edges
+    }
+}
+
+/// Attach a [`CorrelationTuple`] to every edge whose decision meets the
+/// minimal-support threshold, computed from the batch-wide co-occurrence
+/// contingency table.
+///
+/// The "population" is the set of bound arcs (one per binding edge). For a given
+/// edge `decision D → episode with path P`, the 2×2 table over that population
+/// is, for each arc:
+/// - row: is the arc's decision == D?
+/// - col: is the arc's path == P?
+///
+/// yielding counts `n11` (D & P), `n10` (D & ¬P), `n01` (¬D & P),
+/// `n00` (¬D & ¬P), with `N = n11+n10+n01+n00`. From these:
+/// - `support    = n11 / N`
+/// - `confidence = n11 / (n11 + n10)` — `P(P | D)`
+/// - `lift       = confidence / P(P)` where `P(P) = (n11 + n01) / N`
+/// - `phi        = (n11·n00 − n10·n01) / sqrt(row·row·col·col marginals)`
+/// - `p          = erfc(sqrt(chi² / 2))`, `chi² = N · phi²` (1 dof)
+///
+/// All fields are rounded to [`CORRELATION_DECIMALS`] so the tuple is a stable
+/// function of the integer counts. Degenerate tables (a zero marginal makes phi
+/// undefined; `lift` undefined when `P(P)=0`) collapse to neutral, finite values
+/// (`phi = 0`, `p = 1`, `lift = 0`) rather than NaN — still deterministic.
+fn attach_correlations(edges: &mut [BindingEdge], arcs: &[(NodeId, PathBuf)]) {
+    // How many arcs each decision binds, so we can apply the support threshold.
+    let mut bound_by_decision: HashMap<&NodeId, usize> = HashMap::new();
+    for (dec, _) in arcs {
+        *bound_by_decision.entry(dec).or_insert(0) += 1;
+    }
+
+    let total = arcs.len();
+    if total == 0 {
+        return;
+    }
+
+    for (edge, (dec, path)) in edges.iter_mut().zip(arcs.iter()) {
+        // "when computable": require enough co-occurring episodes for the
+        // decision before the table carries meaning.
+        if bound_by_decision.get(dec).copied().unwrap_or(0) < MIN_CORRELATION_SUPPORT {
+            continue;
+        }
+
+        let mut n11 = 0u64; // D & P
+        let mut n10 = 0u64; // D & ¬P
+        let mut n01 = 0u64; // ¬D & P
+        let mut n00 = 0u64; // ¬D & ¬P
+        for (d, p) in arcs {
+            let is_d = d == dec;
+            let is_p = p == path;
+            match (is_d, is_p) {
+                (true, true) => n11 += 1,
+                (true, false) => n10 += 1,
+                (false, true) => n01 += 1,
+                (false, false) => n00 += 1,
+            }
+        }
+
+        edge.correlation = Some(contingency_to_correlation(n11, n10, n01, n00, total as u64));
+    }
+}
+
+/// Turn an integer 2×2 contingency table into a rounded [`CorrelationTuple`].
+/// Pure and deterministic.
+fn contingency_to_correlation(n11: u64, n10: u64, n01: u64, n00: u64, n: u64) -> CorrelationTuple {
+    let f = |x: u64| x as f64;
+    let n_f = f(n).max(1.0);
+
+    let support = f(n11) / n_f;
+
+    let antecedent = n11 + n10; // arcs with this decision
+    let confidence = if antecedent == 0 {
+        0.0
+    } else {
+        f(n11) / f(antecedent)
+    };
+
+    let p_consequent = f(n11 + n01) / n_f; // P(path = P)
+    let lift = if p_consequent == 0.0 {
+        0.0
+    } else {
+        confidence / p_consequent
+    };
+
+    // phi = (n11 n00 − n10 n01) / sqrt(product of the four marginals).
+    let row1 = f(n11 + n10);
+    let row0 = f(n01 + n00);
+    let col1 = f(n11 + n01);
+    let col0 = f(n10 + n00);
+    let denom = (row1 * row0 * col1 * col0).sqrt();
+    let phi = if denom == 0.0 {
+        0.0
+    } else {
+        (f(n11) * f(n00) - f(n10) * f(n01)) / denom
+    };
+
+    // chi-square (1 dof) = N · phi²; two-sided p-value = erfc(sqrt(chi²/2)).
+    let chi2 = n_f * phi * phi;
+    let p = erfc((chi2 / 2.0).sqrt());
+
+    CorrelationTuple {
+        support: round(support),
+        confidence: round(confidence),
+        lift: round(lift),
+        phi: round(phi),
+        p: round(p),
+    }
+}
+
+/// Round to [`CORRELATION_DECIMALS`] places — stable across runs.
+fn round(x: f64) -> f64 {
+    if !x.is_finite() {
+        return 0.0;
+    }
+    let scale = 10f64.powi(CORRELATION_DECIMALS);
+    (x * scale).round() / scale
+}
+
+/// Complementary error function — the Numerical-Recipes `erfcc` rational
+/// approximation (fractional error < 1.2e-7 everywhere). Uses only `exp` and
+/// arithmetic, so it is a deterministic function of its input on a given target.
+fn erfc(x: f64) -> f64 {
+    let z = x.abs();
+    let t = 1.0 / (1.0 + 0.5 * z);
+    let ans = t
+        * (-z * z - 1.265_512_23
+            + t * (1.000_023_68
+                + t * (0.374_091_96
+                    + t * (0.096_784_18
+                        + t * (-0.186_288_06
+                            + t * (0.278_868_07
+                                + t * (-1.135_203_98
+                                    + t * (1.488_515_87
+                                        + t * (-0.822_152_23 + t * 0.170_872_77)))))))))
+            .exp();
+    if x >= 0.0 {
+        ans
+    } else {
+        2.0 - ans
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::model::Diff;
+    use crate::node::{CodeEpisode, DecisionRecord, FactStatus as FS};
+    use crate::segmenter::{DecisionCandidate, EpisodeRecord, Segmentation};
+    use time::OffsetDateTime;
+
+    fn ts(secs: i64) -> OffsetDateTime {
+        OffsetDateTime::from_unix_timestamp(1_700_000_000 + secs).unwrap()
+    }
+
+    fn decision(id: &str, session: &str, turn_seq: u64, t: i64) -> DecisionCandidate {
+        DecisionCandidate {
+            record: DecisionRecord {
+                epitome: format!("decision {id}"),
+                considered_options: vec![],
+                is_ban: false,
+                superseded_by: None,
+                confirmation: None,
+                source_span: turn_seq..turn_seq + 1,
+                fact_status: FS::Observed,
+            },
+            node_id: NodeId::new(id),
+            turn_seq,
+            timestamp: ts(t),
+            session_id: session.to_string(),
+        }
+    }
+
+    fn superseded(mut d: DecisionCandidate, marker: &str) -> DecisionCandidate {
+        d.record.superseded_by = Some(NodeId::new(marker));
+        d
+    }
+
+    fn episode(id: &str, session: &str, seq: u64, t: i64, path: &str) -> EpisodeRecord {
+        EpisodeRecord {
+            episode: CodeEpisode {
+                path: PathBuf::from(path),
+                diff: Diff::for_path(path),
+                git: None,
+                episode_id: id.to_string(),
+            },
+            node_id: NodeId::new(format!("episode:{id}")),
+            seq,
+            timestamp: ts(t),
+            session_id: session.to_string(),
+        }
+    }
+
+    // ---- Task C: structural scenarios that must still hold ----
+
+    #[test]
+    fn interleaved_arcs_bind_to_most_recent_preceding_decision() {
+        // Two decisions in one session; an edit after each binds to the nearest
+        // preceding decision, not the older one.
+        let seg = Segmentation {
+            conversations: vec![],
+            decisions: vec![
+                decision("dec:A", "s1", 1, 10),
+                decision("dec:B", "s1", 3, 30),
+            ],
+            episodes: vec![
+                episode("e1", "s1", 2, 20, "a.rs"), // after A, before B → A
+                episode("e2", "s1", 4, 40, "b.rs"), // after B → B
+            ],
+        };
+        let edges = DefaultBinder.bind(&seg);
+        let a = edges.iter().find(|e| e.to.0 == "episode:e1").unwrap();
+        let b = edges.iter().find(|e| e.to.0 == "episode:e2").unwrap();
+        assert_eq!(a.from.0, "dec:A");
+        assert_eq!(b.from.0, "dec:B");
+    }
+
+    #[test]
+    fn multi_edit_one_decision_n_episodes_n_bindings() {
+        let seg = Segmentation {
+            conversations: vec![],
+            decisions: vec![decision("dec:A", "s1", 1, 10)],
+            episodes: vec![
+                episode("e1", "s1", 2, 20, "a.rs"),
+                episode("e2", "s1", 3, 30, "b.rs"),
+                episode("e3", "s1", 4, 40, "c.rs"),
+            ],
+        };
+        let edges = DefaultBinder.bind(&seg);
+        assert_eq!(edges.len(), 3);
+        assert!(edges.iter().all(|e| e.from.0 == "dec:A"));
+    }
+
+    #[test]
+    fn no_decision_no_spurious_binding() {
+        let seg = Segmentation {
+            conversations: vec![],
+            decisions: vec![],
+            episodes: vec![episode("e1", "s1", 2, 20, "a.rs")],
+        };
+        assert!(DefaultBinder.bind(&seg).is_empty());
+    }
+
+    // ---- Task A (binder side): superseded decisions don't bind ----
+
+    #[test]
+    fn superseded_decision_is_skipped_and_falls_through() {
+        // Decision A is rewound away; decision B governs. The edit after B binds
+        // to B, never to the superseded A.
+        let seg = Segmentation {
+            conversations: vec![],
+            decisions: vec![
+                superseded(decision("dec:A", "s1", 1, 10), "rewind:s1:2"),
+                decision("dec:B", "s1", 5, 50),
+            ],
+            episodes: vec![episode("e1", "s1", 6, 60, "a.rs")],
+        };
+        let edges = DefaultBinder.bind(&seg);
+        assert_eq!(edges.len(), 1);
+        assert_eq!(edges[0].from.0, "dec:B");
+    }
+
+    #[test]
+    fn only_superseded_decision_yields_no_binding() {
+        let seg = Segmentation {
+            conversations: vec![],
+            decisions: vec![superseded(decision("dec:A", "s1", 1, 10), "rewind:s1:2")],
+            episodes: vec![episode("e1", "s1", 6, 60, "a.rs")],
+        };
+        assert!(DefaultBinder.bind(&seg).is_empty());
+    }
+
+    // ---- Task B: correlation tuple ----
+
+    #[test]
+    fn repeated_cooccurrence_yields_some_correlation_with_sane_values() {
+        // One decision binds three episodes (>= MIN_CORRELATION_SUPPORT), two of
+        // which touch the same path. Correlation must be Some and in-range.
+        let seg = Segmentation {
+            conversations: vec![],
+            decisions: vec![decision("dec:A", "s1", 1, 10)],
+            episodes: vec![
+                episode("e1", "s1", 2, 20, "auth.rs"),
+                episode("e2", "s1", 3, 30, "auth.rs"),
+                episode("e3", "s1", 4, 40, "db.rs"),
+            ],
+        };
+        let edges = DefaultBinder.bind(&seg);
+        assert_eq!(edges.len(), 3);
+        for e in &edges {
+            let c = e.correlation.as_ref().expect("correlation present");
+            assert!((0.0..=1.0).contains(&c.support), "support {}", c.support);
+            assert!(
+                (0.0..=1.0).contains(&c.confidence),
+                "confidence {}",
+                c.confidence
+            );
+            assert!(c.lift >= 0.0, "lift {}", c.lift);
+            assert!((-1.0..=1.0).contains(&c.phi), "phi {}", c.phi);
+            assert!((0.0..=1.0).contains(&c.p), "p {}", c.p);
+        }
+        // fact_status stays DeterministicallyDerived even with correlation set.
+        assert!(edges
+            .iter()
+            .all(|e| e.fact_status == FactStatus::DeterministicallyDerived));
+    }
+
+    #[test]
+    fn single_degenerate_arc_yields_none() {
+        // A lone decision→episode arc has support 1 < threshold → no correlation.
+        let seg = Segmentation {
+            conversations: vec![],
+            decisions: vec![decision("dec:A", "s1", 1, 10)],
+            episodes: vec![episode("e1", "s1", 2, 20, "a.rs")],
+        };
+        let edges = DefaultBinder.bind(&seg);
+        assert_eq!(edges.len(), 1);
+        assert!(edges[0].correlation.is_none());
+    }
+
+    #[test]
+    fn correlation_is_deterministic_across_runs() {
+        let seg = Segmentation {
+            conversations: vec![],
+            decisions: vec![
+                decision("dec:A", "s1", 1, 10),
+                decision("dec:B", "s1", 5, 50),
+            ],
+            episodes: vec![
+                episode("e1", "s1", 2, 20, "auth.rs"),
+                episode("e2", "s1", 3, 30, "auth.rs"),
+                episode("e3", "s1", 4, 40, "db.rs"),
+                episode("e4", "s1", 6, 60, "db.rs"),
+                episode("e5", "s1", 7, 70, "auth.rs"),
+            ],
+        };
+        let a = DefaultBinder.bind(&seg);
+        let b = DefaultBinder.bind(&seg);
+        let ja = serde_json::to_string(&a).unwrap();
+        let jb = serde_json::to_string(&b).unwrap();
+        assert_eq!(ja, jb);
+    }
+
+    #[test]
+    fn contingency_table_matches_hand_computed_values() {
+        // n11=2, n10=1, n01=1, n00=2, N=6. Standard market-basket arithmetic.
+        let c = contingency_to_correlation(2, 1, 1, 2, 6);
+        assert!((c.support - 0.333_333).abs() < 1e-6);
+        // confidence = 2/3
+        assert!((c.confidence - 0.666_667).abs() < 1e-6);
+        // P(P) = 3/6 = 0.5 → lift = (2/3)/0.5 = 1.333333
+        assert!((c.lift - 1.333_333).abs() < 1e-6);
+        // phi = (2*2 - 1*1)/sqrt(3*3*3*3) = 3/9 = 0.333333
+        assert!((c.phi - 0.333_333).abs() < 1e-6);
+        assert!((0.0..=1.0).contains(&c.p));
+    }
+
+    #[test]
+    fn erfc_is_sane_at_known_points() {
+        assert!((erfc(0.0) - 1.0).abs() < 1e-6);
+        assert!(erfc(5.0) < 1e-6);
+        assert!((erfc(-5.0) - 2.0).abs() < 1e-6);
+    }
+}
diff --git a/crates/memscribe-core/src/error.rs b/crates/memscribe-core/src/error.rs
new file mode 100644
index 0000000..7c54719
--- /dev/null
+++ b/crates/memscribe-core/src/error.rs
@@ -0,0 +1,49 @@
+//! Error types for the deterministic pipeline. Adapters return [`ParseError`]
+//! only for genuinely malformed input; unrecognized-but-well-formed records are
+//! routed to [`crate::model::EventKind::Unknown`], never an error.
+
+use thiserror::Error;
+
+/// A parse failure. Adapters must reserve this for malformed bytes — anything
+/// merely *unrecognized* becomes [`crate::model::EventKind::Unknown`] so the
+/// stream stays lossless and version-tolerant.
+#[derive(Debug, Error, Clone, PartialEq, Eq)]
+pub enum ParseError {
+    /// The record could not be parsed at all (e.g. invalid UTF-8 / broken JSON).
+    #[error("malformed record at {location}: {message}")]
+    Malformed { location: String, message: String },
+    /// JSON deserialization failed.
+    #[error("json error: {0}")]
+    Json(String),
+    /// The schema variant is recognized but not supported by this adapter.
+    #[error("unsupported schema variant: {0}")]
+    UnsupportedSchema(String),
+    /// An I/O problem occurred while reading the record.
+    #[error("io error: {0}")]
+    Io(String),
+}
+
+/// A failure while writing to a [`crate::sink::Sink`].
+#[derive(Debug, Error, Clone, PartialEq, Eq)]
+pub enum SinkError {
+    /// The underlying write failed.
+    #[error("sink write failed: {0}")]
+    Write(String),
+    /// Flushing buffered nodes failed.
+    #[error("sink flush failed: {0}")]
+    Flush(String),
+    /// A node could not be serialized.
+    #[error("serialization failed: {0}")]
+    Serialize(String),
+}
+
+/// A top-level pipeline failure, wrapping a parse or sink error.
+#[derive(Debug, Error)]
+pub enum PipelineError {
+    /// A record failed to parse.
+    #[error(transparent)]
+    Parse(#[from] ParseError),
+    /// A node failed to reach the sink.
+    #[error(transparent)]
+    Sink(#[from] SinkError),
+}
diff --git a/crates/memscribe-core/src/gate.rs b/crates/memscribe-core/src/gate.rs
new file mode 100644
index 0000000..eee2e58
--- /dev/null
+++ b/crates/memscribe-core/src/gate.rs
@@ -0,0 +1,214 @@
+//! The deterministic commitment-marker gate (whitepaper Appendix B).
+//!
+//! A small, inspectable rule table over user turns. Each rule is a category plus
+//! a regular expression. Evaluating a turn is a **pure function of the turn
+//! text** — no global state — which is the property the gate-purity test
+//! asserts. A match elevates the turn-span to a Conversation node and seeds a
+//! candidate Decision; a non-match retains the verbatim turn at low salience but
+//! creates no node.
+
+use crate::node::{CommitmentMarker, MarkerCategory};
+use regex::Regex;
+
+/// One rule in the commitment-marker table.
+pub struct GateRule {
+    /// The rule id (e.g. `decision_verb.use`).
+    pub id: String,
+    /// The category the rule expresses.
+    pub category: MarkerCategory,
+    /// The case-insensitive pattern.
+    pub pattern: Regex,
+}
+
+impl std::fmt::Debug for GateRule {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("GateRule")
+            .field("id", &self.id)
+            .field("category", &self.category)
+            .field("pattern", &self.pattern.as_str())
+            .finish()
+    }
+}
+
+/// The default rule table, as `(id, category, pattern)` triples. Patterns are
+/// compiled case-insensitively. Kept here so the table is inspectable and
+/// unit-tested per rule.
+#[must_use]
+pub fn default_rules() -> Vec<(&'static str, MarkerCategory, &'static str)> {
+    use MarkerCategory::*;
+    vec![
+        // Explicit decision verbs.
+        (
+            "decision_verb.use",
+            DecisionVerb,
+            r"\b(?:use|using|adopt|adopts|go with|let'?s go with|switch to|migrate to)\b",
+        ),
+        (
+            "decision_verb.decide",
+            DecisionVerb,
+            r"\b(?:decide(?:d)?|we(?:'ll| will) choose|choose|chose|settle on|going to use|pick(?:ed)?)\b",
+        ),
+        // Rejected alternatives.
+        ("rejection.instead_of", Rejection, r"\binstead of\b"),
+        (
+            "rejection.rather_than",
+            Rejection,
+            r"\b(?:rather than|as opposed to|in favor of)\b",
+        ),
+        // Bans (Kruchten anticrisis).
+        (
+            "ban.negated_use",
+            Ban,
+            r"\b(?:never|do not|don'?t|won'?t|will not|must not|should not|shouldn'?t|no longer)\s+(?:use|add|adopt|depend|introduce|rely)\b",
+        ),
+        (
+            "ban.no_dependency",
+            Ban,
+            r"\bno (?:new )?dependenc(?:y|ies)\b",
+        ),
+        // Imperatives.
+        (
+            "imperative.must_always_never",
+            Imperative,
+            r"\b(?:must|always|never|shall|required to|need to)\b",
+        ),
+        // Memory directives.
+        (
+            "memory.remember",
+            Memory,
+            r"\b(?:remember that|keep in mind|note that|for future reference|don'?t forget)\b",
+        ),
+    ]
+}
+
+/// The commitment gate: a compiled, ordered rule table.
+#[derive(Debug)]
+pub struct CommitmentGate {
+    rules: Vec<GateRule>,
+}
+
+impl Default for CommitmentGate {
+    fn default() -> Self {
+        Self::default_table()
+    }
+}
+
+impl CommitmentGate {
+    /// Build the gate from the default rule table.
+    ///
+    /// # Panics
+    /// Never in practice — the default patterns are compile-time constants and
+    /// are exercised by tests; a malformed default is a build-breaking bug.
+    #[must_use]
+    pub fn default_table() -> Self {
+        Self::from_triples(default_rules()).expect("default gate rules must compile")
+    }
+
+    /// Build a gate from `(id, category, pattern)` triples (config-driven).
+    ///
+    /// # Errors
+    /// Returns the underlying regex error if any pattern fails to compile.
+    pub fn from_triples<S: AsRef<str>>(
+        triples: impl IntoIterator<Item = (S, MarkerCategory, S)>,
+    ) -> Result<Self, regex::Error> {
+        let mut rules = Vec::new();
+        for (id, category, pattern) in triples {
+            let pattern = Regex::new(&format!("(?i){}", pattern.as_ref()))?;
+            rules.push(GateRule {
+                id: id.as_ref().to_string(),
+                category,
+                pattern,
+            });
+        }
+        Ok(CommitmentGate { rules })
+    }
+
+    /// The number of rules in the table.
+    #[must_use]
+    pub fn rule_count(&self) -> usize {
+        self.rules.len()
+    }
+
+    /// Evaluate a turn's text against the rule table. **Pure**: depends only on
+    /// `text`. Returns the markers that fired, in rule order (deterministic).
+    #[must_use]
+    pub fn evaluate(&self, text: &str) -> Vec<CommitmentMarker> {
+        let mut out = Vec::new();
+        for rule in &self.rules {
+            if let Some(m) = rule.pattern.find(text) {
+                out.push(CommitmentMarker {
+                    rule_id: rule.id.clone(),
+                    category: rule.category,
+                    matched_text: m.as_str().to_string(),
+                    offset: m.start(),
+                });
+            }
+        }
+        out
+    }
+
+    /// Whether the turn is admitted (any marker fired).
+    #[must_use]
+    pub fn admits(&self, text: &str) -> bool {
+        self.rules.iter().any(|r| r.pattern.is_match(text))
+    }
+
+    /// Whether any fired marker is a ban.
+    #[must_use]
+    pub fn is_ban(&self, markers: &[CommitmentMarker]) -> bool {
+        markers.iter().any(|m| m.category == MarkerCategory::Ban)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn default_table_compiles_and_has_rules() {
+        let gate = CommitmentGate::default_table();
+        assert!(gate.rule_count() >= 8);
+    }
+
+    #[test]
+    fn decision_verb_fires() {
+        let gate = CommitmentGate::default_table();
+        let m = gate.evaluate("Let's go with Postgres for storage.");
+        assert!(!m.is_empty());
+        assert!(m.iter().any(|x| x.category == MarkerCategory::DecisionVerb));
+    }
+
+    #[test]
+    fn rejection_fires() {
+        let gate = CommitmentGate::default_table();
+        let m = gate.evaluate("Use Stripe instead of PayPal.");
+        assert!(m.iter().any(|x| x.category == MarkerCategory::Rejection));
+        assert!(m.iter().any(|x| x.category == MarkerCategory::DecisionVerb));
+    }
+
+    #[test]
+    fn ban_fires_and_is_detected() {
+        let gate = CommitmentGate::default_table();
+        let m = gate.evaluate("We will never add a dependency on left-pad.");
+        assert!(gate.is_ban(&m), "expected a ban marker: {m:?}");
+    }
+
+    #[test]
+    fn plain_chatter_does_not_admit() {
+        let gate = CommitmentGate::default_table();
+        assert!(!gate.admits("Thanks, that looks good to me."));
+    }
+
+    #[test]
+    fn evaluate_is_pure_and_repeatable() {
+        let gate = CommitmentGate::default_table();
+        let t = "We must always use prepared statements instead of string concatenation.";
+        assert_eq!(gate.evaluate(t), gate.evaluate(t));
+    }
+
+    #[test]
+    fn case_insensitive() {
+        let gate = CommitmentGate::default_table();
+        assert!(gate.admits("LET'S GO WITH redis"));
+    }
+}
diff --git a/crates/memscribe-core/src/lib.rs b/crates/memscribe-core/src/lib.rs
new file mode 100644
index 0000000..2fde053
--- /dev/null
+++ b/crates/memscribe-core/src/lib.rs
@@ -0,0 +1,50 @@
+//! # memscribe-core
+//!
+//! The deterministic, zero-LLM **contract** at the heart of Memscribe.
+//!
+//! This crate defines the thin waist that decouples per-tool adapters from
+//! everything downstream:
+//!
+//! - The normalized event model ([`CaptureEvent`] / [`EventKind`]) — the system
+//!   of record produced by adapters.
+//! - The output contract ([`node::PreparedNode`]) — the typed nodes the
+//!   consumer layer (MemCortex) ingests, each carrying a [`node::FactStatus`].
+//! - The [`TranscriptAdapter`] and [`Sink`] traits — the two plug points.
+//! - The deterministic pipeline: [`gate`] → [`segmenter`] → [`binder`] →
+//!   [`nodeprep`], plus a [`redact`] pass.
+//!
+//! Everything here is a pure function of its input. No model is ever called.
+//! That is what makes Memscribe golden-file, property, and fuzz testable.
+#![forbid(unsafe_code)]
+
+pub mod adapter;
+pub mod binder;
+pub mod error;
+pub mod gate;
+pub mod model;
+pub mod node;
+pub mod nodeprep;
+pub mod pipeline;
+pub mod redact;
+pub mod segmenter;
+pub mod sink;
+
+pub use adapter::{
+    DiscoverCfg, ParseCtx, RawRecord, SchemaVariant, TranscriptAdapter, TranscriptHandle,
+};
+pub use binder::{Binder, DefaultBinder};
+pub use error::{ParseError, PipelineError, SinkError};
+pub use gate::{CommitmentGate, GateRule};
+pub use model::{
+    content_id, CaptureEvent, Diff, EventKind, GitRef, Part, ProjectRef, SourceKind,
+    SourceLocation, Timestamp, Usage, SCHEMA_VERSION,
+};
+pub use node::{
+    BindingEdge, CheckRef, CodeEpisode, CommitmentMarker, ConversationSpan, CorrelationTuple,
+    DecisionRecord, FactStatus, MarkerCategory, NodeId, Opt, PreparedNode, ProvRecord, Relation,
+};
+pub use nodeprep::{DefaultNodePrep, NodePrep};
+pub use pipeline::DefaultPipeline;
+pub use redact::Redactor;
+pub use segmenter::{DecisionCandidate, DefaultSegmenter, EpisodeRecord, Segmentation, Segmenter};
+pub use sink::{Sink, VecSink};
diff --git a/crates/memscribe-core/src/model.rs b/crates/memscribe-core/src/model.rs
new file mode 100644
index 0000000..ee5ab99
--- /dev/null
+++ b/crates/memscribe-core/src/model.rs
@@ -0,0 +1,394 @@
+//! The normalized event model — the system of record produced by adapters.
+//!
+//! Every field is copied verbatim from the source; none is generated by
+//! Memscribe. This is the thin waist that decouples adapters from everything
+//! downstream (whitepaper §4).
+
+use serde::{Deserialize, Serialize};
+use std::ops::Range;
+use std::path::PathBuf;
+use time::OffsetDateTime;
+
+/// The current event schema version. Bumped on any breaking change to
+/// [`CaptureEvent`] / [`EventKind`]. Consumers gate on this independently of the
+/// crate's semver.
+pub const SCHEMA_VERSION: u16 = 1;
+
+/// An RFC3339 timestamp copied verbatim from the source.
+pub type Timestamp = OffsetDateTime;
+
+/// A stable blake3 content hash, hex-encoded — used as an `event_id` fallback
+/// when a tool provides no native id, and to derive deterministic episode ids.
+#[must_use]
+pub fn content_id(bytes: &[u8]) -> String {
+    blake3::hash(bytes).to_hex().to_string()
+}
+
+/// Which tool produced an event.
+#[derive(Clone, Copy, Debug, Serialize, Deserialize, PartialEq, Eq, Hash)]
+#[serde(rename_all = "snake_case")]
+pub enum SourceKind {
+    /// Anthropic Claude Code.
+    ClaudeCode,
+    /// OpenAI Codex CLI.
+    Codex,
+    /// Google Gemini CLI.
+    Gemini,
+    /// Any OpenTelemetry GenAI-instrumented agent.
+    Otel,
+    /// Cursor (VS Code fork).
+    Cursor,
+    /// Windsurf (Codeium).
+    Windsurf,
+    /// Zed editor agent.
+    Zed,
+    /// VS Code (Copilot Chat / chat sessions).
+    VsCode,
+    /// GitHub Copilot.
+    Copilot,
+    /// Source could not be determined.
+    Unknown,
+}
+
+impl SourceKind {
+    /// The stable snake_case identifier (matches serde and CLI `--as` values).
+    #[must_use]
+    pub fn as_str(&self) -> &'static str {
+        match self {
+            SourceKind::ClaudeCode => "claude_code",
+            SourceKind::Codex => "codex",
+            SourceKind::Gemini => "gemini",
+            SourceKind::Otel => "otel",
+            SourceKind::Cursor => "cursor",
+            SourceKind::Windsurf => "windsurf",
+            SourceKind::Zed => "zed",
+            SourceKind::VsCode => "vscode",
+            SourceKind::Copilot => "copilot",
+            SourceKind::Unknown => "unknown",
+        }
+    }
+
+    /// Parse a `SourceKind` from a CLI/`--as` identifier (tolerant of aliases).
+    #[must_use]
+    pub fn parse(s: &str) -> Option<SourceKind> {
+        match s.trim().to_ascii_lowercase().as_str() {
+            "claude_code" | "claude" | "claude-code" => Some(SourceKind::ClaudeCode),
+            "codex" => Some(SourceKind::Codex),
+            "gemini" => Some(SourceKind::Gemini),
+            "otel" | "opentelemetry" => Some(SourceKind::Otel),
+            "cursor" => Some(SourceKind::Cursor),
+            "windsurf" => Some(SourceKind::Windsurf),
+            "zed" => Some(SourceKind::Zed),
+            "vscode" | "vs_code" | "vs-code" | "code" => Some(SourceKind::VsCode),
+            "copilot" | "github_copilot" => Some(SourceKind::Copilot),
+            "unknown" => Some(SourceKind::Unknown),
+            _ => None,
+        }
+    }
+}
+
+impl std::fmt::Display for SourceKind {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(self.as_str())
+    }
+}
+
+/// A git reference captured at session start: commit sha and (optional) branch.
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
+pub struct GitRef {
+    /// The commit sha at the time of capture.
+    pub sha: String,
+    /// The branch name, if known.
+    pub branch: Option<String>,
+}
+
+/// The project a session is bound to (captured at session start).
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
+pub struct ProjectRef {
+    /// The working directory the agent ran in.
+    pub cwd: PathBuf,
+    /// The repository root, if resolved.
+    pub repo_root: Option<PathBuf>,
+    /// The git ref at session start, if any.
+    pub git: Option<GitRef>,
+}
+
+impl ProjectRef {
+    /// A minimal project ref for a bare working directory.
+    #[must_use]
+    pub fn from_cwd(cwd: impl Into<PathBuf>) -> Self {
+        ProjectRef {
+            cwd: cwd.into(),
+            repo_root: None,
+            git: None,
+        }
+    }
+}
+
+/// A pointer back into the source bytes, for replay and audit.
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
+pub struct SourceLocation {
+    /// The transcript file the record came from.
+    pub file: PathBuf,
+    /// Byte offset of the record's start within the file.
+    pub byte_offset: u64,
+    /// 1-based line number of the record within the file.
+    pub line_no: u64,
+}
+
+impl SourceLocation {
+    /// Construct a source location.
+    #[must_use]
+    pub fn new(file: impl Into<PathBuf>, byte_offset: u64, line_no: u64) -> Self {
+        SourceLocation {
+            file: file.into(),
+            byte_offset,
+            line_no,
+        }
+    }
+}
+
+impl std::fmt::Display for SourceLocation {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(
+            f,
+            "{}:{}@{}",
+            self.file.display(),
+            self.line_no,
+            self.byte_offset
+        )
+    }
+}
+
+/// A content part within a turn. Multimodal- and version-tolerant: anything
+/// unrecognized is preserved verbatim as [`Part::Other`].
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
+#[serde(tag = "part", rename_all = "snake_case")]
+pub enum Part {
+    /// Plain text content.
+    Text {
+        /// The verbatim text.
+        text: String,
+    },
+    /// Model-private reasoning, where the tool exposes it.
+    Thinking {
+        /// The verbatim thinking text.
+        text: String,
+    },
+    /// A tool invocation embedded in the turn.
+    ToolUse {
+        /// The tool-native call id.
+        call_id: String,
+        /// The tool name.
+        name: String,
+        /// The raw arguments.
+        args: serde_json::Value,
+    },
+    /// A tool result embedded in the turn.
+    ToolResult {
+        /// The tool-native call id this result answers.
+        call_id: String,
+        /// The raw output.
+        output: serde_json::Value,
+    },
+    /// An image/attachment reference (content not captured).
+    Image {
+        /// The media type, if known.
+        media_type: Option<String>,
+    },
+    /// Any unrecognized part, preserved verbatim.
+    Other {
+        /// The raw JSON of the part.
+        raw: serde_json::Value,
+    },
+}
+
+/// Token usage for an assistant turn, where the tool reports it.
+#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq, Eq)]
+pub struct Usage {
+    /// Input/prompt tokens.
+    pub input_tokens: Option<u64>,
+    /// Output/completion tokens.
+    pub output_tokens: Option<u64>,
+    /// Cache-read tokens.
+    pub cache_read_tokens: Option<u64>,
+    /// Cache-creation tokens.
+    pub cache_creation_tokens: Option<u64>,
+}
+
+/// A normalized file edit. The diff is copied from the source when available,
+/// or computed deterministically (e.g. via `similar`) when only old/new are
+/// present.
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
+pub struct Diff {
+    /// The path that was edited.
+    pub path: PathBuf,
+    /// The prior contents, if the source provides them.
+    pub old: Option<String>,
+    /// The new contents, if the source provides them.
+    pub new: Option<String>,
+    /// Unified-diff text — from the source, or computed deterministically.
+    pub unified: Option<String>,
+    /// Count of added lines.
+    pub added_lines: u32,
+    /// Count of removed lines.
+    pub removed_lines: u32,
+}
+
+impl Diff {
+    /// A minimal diff for a path with no captured content.
+    #[must_use]
+    pub fn for_path(path: impl Into<PathBuf>) -> Self {
+        Diff {
+            path: path.into(),
+            old: None,
+            new: None,
+            unified: None,
+            added_lines: 0,
+            removed_lines: 0,
+        }
+    }
+}
+
+/// The kind of a normalized event. [`EventKind::Unknown`] is load-bearing: an
+/// unrecognized record type or new field is preserved verbatim and flagged,
+/// never discarded — this is how the stream stays lossless across tool-version
+/// churn (whitepaper §4).
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
+#[serde(tag = "event", rename_all = "snake_case")]
+pub enum EventKind {
+    /// The session opened: cwd, git, model, tool version.
+    SessionStart {
+        /// Working directory.
+        cwd: PathBuf,
+        /// Git ref at session start.
+        git: Option<GitRef>,
+        /// The model in use, if reported.
+        model: Option<String>,
+        /// The tool version, if reported.
+        tool_version: Option<String>,
+    },
+    /// A user message.
+    UserTurn {
+        /// The flattened text of the turn.
+        text: String,
+        /// The structured parts, if any.
+        #[serde(default)]
+        parts: Vec<Part>,
+    },
+    /// An assistant message.
+    AssistantTurn {
+        /// The flattened text of the turn.
+        text: String,
+        /// Model-private reasoning, if exposed.
+        thinking: Option<String>,
+        /// The model that produced the turn.
+        model: Option<String>,
+        /// Token usage, if reported.
+        usage: Option<Usage>,
+        /// The structured parts, if any.
+        #[serde(default)]
+        parts: Vec<Part>,
+    },
+    /// A tool call.
+    ToolCall {
+        /// The tool-native call id.
+        call_id: String,
+        /// The tool name.
+        name: String,
+        /// The raw arguments.
+        args: serde_json::Value,
+    },
+    /// A tool result.
+    ToolResult {
+        /// The call id this result answers.
+        call_id: String,
+        /// Whether the tool succeeded.
+        ok: bool,
+        /// The raw output.
+        output: serde_json::Value,
+    },
+    /// A file edit, normalized from Edit/Write/apply_patch/replace.
+    FileEdit {
+        /// The originating tool call id, if any.
+        call_id: Option<String>,
+        /// The normalized diff.
+        diff: Diff,
+    },
+    /// Model-side history compaction — flagged, never stored as truth.
+    Compaction {
+        /// The (inclusive-start, exclusive-end) seq range that was replaced.
+        replaced: Range<u64>,
+    },
+    /// A user rewind — logical truncation back to an earlier event.
+    Rewind {
+        /// The event id rewound to.
+        to_event: String,
+    },
+    /// The session ended.
+    SessionEnd {
+        /// The reason, if reported.
+        reason: Option<String>,
+    },
+    /// Forward-compat: an unrecognized record, preserved verbatim and flagged.
+    Unknown {
+        /// The raw record type string, if any.
+        raw_type: String,
+        /// The full raw record.
+        raw: serde_json::Value,
+    },
+}
+
+impl EventKind {
+    /// A stable tag string for the variant — used in dedup keys and tests.
+    #[must_use]
+    pub fn tag(&self) -> &'static str {
+        match self {
+            EventKind::SessionStart { .. } => "session_start",
+            EventKind::UserTurn { .. } => "user_turn",
+            EventKind::AssistantTurn { .. } => "assistant_turn",
+            EventKind::ToolCall { .. } => "tool_call",
+            EventKind::ToolResult { .. } => "tool_result",
+            EventKind::FileEdit { .. } => "file_edit",
+            EventKind::Compaction { .. } => "compaction",
+            EventKind::Rewind { .. } => "rewind",
+            EventKind::SessionEnd { .. } => "session_end",
+            EventKind::Unknown { .. } => "unknown",
+        }
+    }
+}
+
+/// A normalized capture event. Stable, versioned (`schema_version`) public
+/// contract (whitepaper §4).
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
+pub struct CaptureEvent {
+    /// The event schema version (see [`SCHEMA_VERSION`]).
+    pub schema_version: u16,
+    /// Which tool produced the event.
+    pub source: SourceKind,
+    /// The tool-native session/thread id.
+    pub session_id: String,
+    /// Monotonic per-session sequence, assigned from file order.
+    pub seq: u64,
+    /// Tool-native id (e.g. a Claude Code uuid), or a `blake3(content)` fallback.
+    pub event_id: String,
+    /// DAG link where the tool provides one (e.g. Claude Code `parentUuid`).
+    pub parent_id: Option<String>,
+    /// RFC3339 timestamp from the source.
+    #[serde(with = "time::serde::rfc3339")]
+    pub timestamp: OffsetDateTime,
+    /// The project binding captured at session start.
+    pub project: ProjectRef,
+    /// The event payload.
+    pub kind: EventKind,
+    /// Pointer back into the source bytes, for replay and audit.
+    pub provenance: SourceLocation,
+}
+
+impl CaptureEvent {
+    /// A dedup key: tool-native `event_id` is authoritative within a session.
+    #[must_use]
+    pub fn dedup_key(&self) -> (&str, &str) {
+        (self.session_id.as_str(), self.event_id.as_str())
+    }
+}
diff --git a/crates/memscribe-core/src/node.rs b/crates/memscribe-core/src/node.rs
new file mode 100644
index 0000000..f687a1b
--- /dev/null
+++ b/crates/memscribe-core/src/node.rs
@@ -0,0 +1,266 @@
+//! The output contract: prepared nodes (whitepaper §6).
+//!
+//! Memscribe only ever produces nodes with `Observed` or
+//! `DeterministicallyDerived` fact-status. It does the deterministic
+//! preparation and *flags* everything that would require inference
+//! (fine-grained decision typing, concept naming) for the consumer to handle
+//! later. That is what keeps the module zero-LLM and its output golden-testable.
+
+use crate::model::{Diff, GitRef, SourceLocation};
+use serde::{Deserialize, Serialize};
+use std::ops::Range;
+use std::path::PathBuf;
+use time::OffsetDateTime;
+
+/// A stable id for a prepared node. Derived deterministically from the source
+/// (session id + span), so the same input always yields the same id.
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq, Hash, PartialOrd, Ord)]
+pub struct NodeId(pub String);
+
+impl NodeId {
+    /// Construct a node id.
+    pub fn new(s: impl Into<String>) -> Self {
+        NodeId(s.into())
+    }
+    /// The id as a string slice.
+    #[must_use]
+    pub fn as_str(&self) -> &str {
+        &self.0
+    }
+}
+
+impl std::fmt::Display for NodeId {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(&self.0)
+    }
+}
+
+/// The epistemic status of a node or edge. Memscribe emits only the first two;
+/// the latter two are *flags* for a downstream inference layer, never values
+/// Memscribe itself computes by guessing.
+#[derive(Clone, Copy, Debug, Serialize, Deserialize, PartialEq, Eq)]
+#[serde(rename_all = "snake_case")]
+pub enum FactStatus {
+    /// Verbatim from the source.
+    Observed,
+    /// Computed by a deterministic function of observed data.
+    DeterministicallyDerived,
+    /// Ranked by a statistical measure (downstream).
+    StatisticallyRanked,
+    /// An LLM hypothesis (downstream); Memscribe only ever *flags* this.
+    LlmHypothesis,
+}
+
+/// The category of a deterministic commitment marker.
+#[derive(Clone, Copy, Debug, Serialize, Deserialize, PartialEq, Eq)]
+#[serde(rename_all = "snake_case")]
+pub enum MarkerCategory {
+    /// Explicit decision verb ("use", "let's go with", "decide").
+    DecisionVerb,
+    /// A rejected alternative ("instead of X", "rather than").
+    Rejection,
+    /// A ban ("we will NOT / never use X") — Kruchten anticrisis.
+    Ban,
+    /// An imperative ("must", "always", "never", "shall").
+    Imperative,
+    /// A memory directive ("remember that", "keep in mind").
+    Memory,
+    /// Assistant-proposal-then-user-confirmation.
+    Confirmation,
+}
+
+/// Which deterministic commitment marker fired on a turn, and where.
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
+pub struct CommitmentMarker {
+    /// The rule id that matched (e.g. `decision_verb.use`).
+    pub rule_id: String,
+    /// The marker category.
+    pub category: MarkerCategory,
+    /// The verbatim text span that matched.
+    pub matched_text: String,
+    /// Byte offset of the match within the turn text.
+    pub offset: usize,
+}
+
+/// A gated, verbatim dialogue span (always `Observed`).
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
+pub struct ConversationSpan {
+    /// The session the span belongs to.
+    pub session_id: String,
+    /// The (inclusive-start, exclusive-end) turn-seq range.
+    pub turn_range: Range<u64>,
+    /// The verbatim dialogue text.
+    pub text: String,
+    /// Which deterministic markers fired.
+    pub markers: Vec<CommitmentMarker>,
+    /// Always [`FactStatus::Observed`].
+    pub fact_status: FactStatus,
+    /// Provenance pointers for replay & audit.
+    pub provenance: Vec<SourceLocation>,
+}
+
+/// A considered option within a decision.
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
+pub struct Opt {
+    /// The option text (verbatim span).
+    pub text: String,
+    /// Whether this option was the one chosen.
+    pub chosen: bool,
+}
+
+/// A pointer to a confirmation check (an ArchUnit rule, test, or schema check)
+/// named in a decision.
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
+pub struct CheckRef {
+    /// The kind of check (`archunit` | `test` | `schema`).
+    pub kind: String,
+    /// The named target.
+    pub target: String,
+}
+
+/// A decision parsed deterministically from a gated turn. The schema follows
+/// IBIS / QOC / MADR / Kruchten. Prose typing that requires inference is left to
+/// the consumer; only verbatim spans and structural flags are populated here.
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
+pub struct DecisionRecord {
+    /// The decision sentence (a verbatim span).
+    pub epitome: String,
+    /// Options parsed from "instead of X", "vs", or explicit lists.
+    pub considered_options: Vec<Opt>,
+    /// True for a ban ("we will NOT / never use X").
+    pub is_ban: bool,
+    /// A pointer to a node that supersedes this decision, if known.
+    pub superseded_by: Option<NodeId>,
+    /// A named confirmation check, if the decision references one.
+    pub confirmation: Option<CheckRef>,
+    /// The exact turn span (no accreted context).
+    pub source_span: Range<u64>,
+    /// `Observed` for the verbatim text. Element-typing uncertainty is flagged
+    /// downstream as [`FactStatus::LlmHypothesis`], never guessed here.
+    pub fact_status: FactStatus,
+}
+
+/// A code edit episode: the path, the diff, and the git sha
+/// (`DeterministicallyDerived`).
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
+pub struct CodeEpisode {
+    /// The edited path.
+    pub path: PathBuf,
+    /// The normalized diff.
+    pub diff: Diff,
+    /// The git ref at edit time, if known.
+    pub git: Option<GitRef>,
+    /// A deterministic id for the episode.
+    pub episode_id: String,
+}
+
+/// The relation a binding edge expresses.
+#[derive(Clone, Copy, Debug, Serialize, Deserialize, PartialEq, Eq)]
+#[serde(rename_all = "snake_case")]
+pub enum Relation {
+    /// A decision/conversation produced an episode.
+    Produced,
+    /// A decision governs an episode.
+    Governs,
+    /// An episode is derived from a decision/conversation.
+    DerivedFrom,
+    /// Two nodes are statistically correlated.
+    CorrelatedWith,
+}
+
+/// A PROV record: `used(session, decision)` + `wasGeneratedBy(diff, session)`
+/// with the temporal invariant `t_use ≤ t_gen`.
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
+pub struct ProvRecord {
+    /// The session that used the decision.
+    pub used_session: String,
+    /// The decision node that was used, if any.
+    pub used_decision: Option<NodeId>,
+    /// The session that generated the edit.
+    pub was_generated_by_session: String,
+    /// When the decision was used.
+    #[serde(with = "time::serde::rfc3339")]
+    pub t_use: OffsetDateTime,
+    /// When the edit was generated. Invariant: `t_use ≤ t_gen`.
+    #[serde(with = "time::serde::rfc3339")]
+    pub t_gen: OffsetDateTime,
+}
+
+impl ProvRecord {
+    /// Whether the temporal invariant `t_use ≤ t_gen` holds.
+    #[must_use]
+    pub fn is_temporally_valid(&self) -> bool {
+        self.t_use <= self.t_gen
+    }
+}
+
+/// A correlation measure between two nodes, when computable.
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
+pub struct CorrelationTuple {
+    /// Support.
+    pub support: f64,
+    /// Confidence.
+    pub confidence: f64,
+    /// Lift.
+    pub lift: f64,
+    /// Phi coefficient.
+    pub phi: f64,
+    /// p-value.
+    pub p: f64,
+}
+
+/// A binding edge: decision/conversation → episode, with PROV, fact-status, and
+/// (optional) correlation.
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
+pub struct BindingEdge {
+    /// The source node.
+    pub from: NodeId,
+    /// The target node.
+    pub to: NodeId,
+    /// The relation.
+    pub relation: Relation,
+    /// The PROV record.
+    pub prov: ProvRecord,
+    /// `DeterministicallyDerived` when recorded live; else downgraded.
+    pub fact_status: FactStatus,
+    /// A correlation tuple, when computable.
+    pub correlation: Option<CorrelationTuple>,
+}
+
+/// The typed data the consumer layer (MemCortex) ingests.
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
+#[serde(tag = "node", rename_all = "snake_case")]
+pub enum PreparedNode {
+    /// A gated, verbatim dialogue span.
+    Conversation(ConversationSpan),
+    /// A deterministically-parsed decision.
+    Decision(DecisionRecord),
+    /// A code edit episode.
+    Episode(CodeEpisode),
+    /// A decision/conversation → episode binding.
+    Binding(BindingEdge),
+}
+
+impl PreparedNode {
+    /// A stable tag for the node variant — used in tests and ordering.
+    #[must_use]
+    pub fn tag(&self) -> &'static str {
+        match self {
+            PreparedNode::Conversation(_) => "conversation",
+            PreparedNode::Decision(_) => "decision",
+            PreparedNode::Episode(_) => "episode",
+            PreparedNode::Binding(_) => "binding",
+        }
+    }
+
+    /// The node's fact-status.
+    #[must_use]
+    pub fn fact_status(&self) -> FactStatus {
+        match self {
+            PreparedNode::Conversation(c) => c.fact_status,
+            PreparedNode::Decision(d) => d.fact_status,
+            PreparedNode::Episode(_) => FactStatus::DeterministicallyDerived,
+            PreparedNode::Binding(b) => b.fact_status,
+        }
+    }
+}
diff --git a/crates/memscribe-core/src/nodeprep.rs b/crates/memscribe-core/src/nodeprep.rs
new file mode 100644
index 0000000..89d3d56
--- /dev/null
+++ b/crates/memscribe-core/src/nodeprep.rs
@@ -0,0 +1,78 @@
+//! Node preparation: assemble the final [`PreparedNode`] stream (whitepaper §3,
+//! §8.8).
+//!
+//! The node-prep stage takes a [`Segmentation`] and the binder's edges and emits
+//! a deterministically-ordered stream of prepared nodes. Order is chronological
+//! by originating seq, with a stable per-kind tiebreak, so the output is
+//! byte-stable for golden tests.
+
+use crate::node::{BindingEdge, PreparedNode};
+use crate::segmenter::Segmentation;
+
+/// The node-prep stage.
+pub trait NodePrep {
+    /// Assemble the final prepared-node stream.
+    fn prepare(&self, seg: &Segmentation, bindings: Vec<BindingEdge>) -> Vec<PreparedNode>;
+}
+
+/// The default deterministic node-prep.
+#[derive(Debug, Default)]
+pub struct DefaultNodePrep;
+
+/// A node carrying its deterministic sort key: `(primary_seq, kind_rank)`.
+struct Keyed {
+    seq: u64,
+    rank: u8,
+    secondary: String,
+    node: PreparedNode,
+}
+
+impl NodePrep for DefaultNodePrep {
+    fn prepare(&self, seg: &Segmentation, bindings: Vec<BindingEdge>) -> Vec<PreparedNode> {
+        let mut keyed: Vec<Keyed> = Vec::new();
+
+        for c in &seg.conversations {
+            keyed.push(Keyed {
+                seq: c.turn_range.start,
+                rank: 0,
+                secondary: c.session_id.clone(),
+                node: PreparedNode::Conversation(c.clone()),
+            });
+        }
+        for d in &seg.decisions {
+            keyed.push(Keyed {
+                seq: d.turn_seq,
+                rank: 1,
+                secondary: d.node_id.0.clone(),
+                node: PreparedNode::Decision(d.record.clone()),
+            });
+        }
+        for e in &seg.episodes {
+            keyed.push(Keyed {
+                seq: e.seq,
+                rank: 2,
+                secondary: e.node_id.0.clone(),
+                node: PreparedNode::Episode(e.episode.clone()),
+            });
+        }
+        for b in bindings {
+            // Bindings sort just after the episode they generate.
+            let secondary = format!("{}->{}", b.from, b.to);
+            keyed.push(Keyed {
+                seq: u64::MAX,
+                rank: 3,
+                secondary,
+                node: PreparedNode::Binding(b),
+            });
+        }
+
+        keyed.sort_by(|a, b| {
+            a.seq
+                .cmp(&b.seq)
+                .then(a.rank.cmp(&b.rank))
+                .then_with(|| a.secondary.cmp(&b.secondary))
+        });
+
+        keyed.into_iter().map(|k| k.node).collect()
+    }
+}
diff --git a/crates/memscribe-core/src/pipeline.rs b/crates/memscribe-core/src/pipeline.rs
new file mode 100644
index 0000000..4e1a41c
--- /dev/null
+++ b/crates/memscribe-core/src/pipeline.rs
@@ -0,0 +1,149 @@
+//! The linear, deterministic pipeline (whitepaper §3).
+//!
+//! `Source → Adapter` produces a normalized [`CaptureEvent`] stream; this module
+//! turns that stream into [`PreparedNode`]s via Gate → Segmenter → Binder →
+//! NodePrep, applies the optional redaction pass, and writes to a [`Sink`].
+//! Everything here is pure and synchronous given the event stream.
+
+use crate::adapter::{ParseCtx, RawRecord, TranscriptAdapter};
+use crate::binder::{Binder, DefaultBinder};
+use crate::error::{PipelineError, SinkError};
+use crate::gate::CommitmentGate;
+use crate::model::CaptureEvent;
+use crate::node::PreparedNode;
+use crate::nodeprep::{DefaultNodePrep, NodePrep};
+use crate::redact::Redactor;
+use crate::segmenter::{DefaultSegmenter, Segmenter};
+use crate::sink::Sink;
+
+/// The default pipeline with the standard stages. Construct with [`Self::new`]
+/// (redaction on by default) or [`Self::without_redaction`].
+#[derive(Debug)]
+pub struct DefaultPipeline {
+    /// The commitment-marker gate.
+    pub gate: CommitmentGate,
+    /// The segmenter stage.
+    pub segmenter: DefaultSegmenter,
+    /// The binder stage.
+    pub binder: DefaultBinder,
+    /// The node-prep stage.
+    pub nodeprep: DefaultNodePrep,
+    /// The redaction pass, if enabled.
+    pub redactor: Option<Redactor>,
+}
+
+impl Default for DefaultPipeline {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl DefaultPipeline {
+    /// A pipeline with default stages and redaction **on** (the safe default).
+    #[must_use]
+    pub fn new() -> Self {
+        DefaultPipeline {
+            gate: CommitmentGate::default_table(),
+            segmenter: DefaultSegmenter,
+            binder: DefaultBinder,
+            nodeprep: DefaultNodePrep,
+            redactor: Some(Redactor::default()),
+        }
+    }
+
+    /// A pipeline with redaction disabled (e.g. for golden tests that assert on
+    /// verbatim content).
+    #[must_use]
+    pub fn without_redaction() -> Self {
+        DefaultPipeline {
+            redactor: None,
+            ..Self::new()
+        }
+    }
+
+    /// Replace the gate (e.g. with a config-driven rule table).
+    #[must_use]
+    pub fn with_gate(mut self, gate: CommitmentGate) -> Self {
+        self.gate = gate;
+        self
+    }
+
+    /// Replace the redactor (e.g. `--no-content` mode), or pass `None` to
+    /// disable redaction.
+    #[must_use]
+    pub fn with_redactor(mut self, redactor: Option<Redactor>) -> Self {
+        self.redactor = redactor;
+        self
+    }
+
+    /// Transform a normalized event stream into prepared nodes. **Pure**: the
+    /// output is an exact function of `events`.
+    #[must_use]
+    pub fn prepare_events(&self, events: &[CaptureEvent]) -> Vec<PreparedNode> {
+        let seg = self.segmenter.segment(events, &self.gate);
+        let bindings = self.binder.bind(&seg);
+        let mut nodes = self.nodeprep.prepare(&seg, bindings);
+        if let Some(r) = &self.redactor {
+            for n in &mut nodes {
+                r.redact_node(n);
+            }
+        }
+        nodes
+    }
+
+    /// Run an adapter over raw records (skipping malformed records — adapters
+    /// route unrecognized-but-valid records to `Unknown`, so a real `Err` here
+    /// is a genuinely broken line that is skipped-and-flagged), then prepare the
+    /// resulting nodes.
+    #[must_use]
+    pub fn run_records(
+        &self,
+        adapter: &dyn TranscriptAdapter,
+        records: &[RawRecord],
+    ) -> Vec<PreparedNode> {
+        let (events, _ctx) = parse_records(adapter, records);
+        self.prepare_events(&events)
+    }
+
+    /// Run the full pipeline to a sink. Returns the number of nodes emitted.
+    ///
+    /// # Errors
+    /// Returns a [`PipelineError`] if the sink fails to emit or flush.
+    pub fn run_to_sink(
+        &self,
+        adapter: &dyn TranscriptAdapter,
+        records: &[RawRecord],
+        sink: &mut dyn Sink,
+    ) -> Result<usize, PipelineError> {
+        let nodes = self.run_records(adapter, records);
+        emit_all(sink, &nodes)?;
+        Ok(nodes.len())
+    }
+}
+
+/// Parse a batch of raw records with a fresh context, collecting the normalized
+/// events. Malformed records (a real `Err`) are skipped; the stream stays
+/// lossless for well-formed input because adapters emit `Unknown` rather than
+/// erroring on unrecognized records.
+#[must_use]
+pub fn parse_records(
+    adapter: &dyn TranscriptAdapter,
+    records: &[RawRecord],
+) -> (Vec<CaptureEvent>, ParseCtx) {
+    let mut ctx = ParseCtx::new();
+    let mut events = Vec::new();
+    for r in records {
+        if let Ok(evs) = adapter.parse(r, &mut ctx) {
+            events.extend(evs);
+        }
+    }
+    (events, ctx)
+}
+
+/// Emit every node to a sink, then flush.
+fn emit_all(sink: &mut dyn Sink, nodes: &[PreparedNode]) -> Result<(), SinkError> {
+    for node in nodes {
+        sink.emit(node)?;
+    }
+    sink.flush()
+}
diff --git a/crates/memscribe-core/src/redact.rs b/crates/memscribe-core/src/redact.rs
new file mode 100644
index 0000000..7bd05dd
--- /dev/null
+++ b/crates/memscribe-core/src/redact.rs
@@ -0,0 +1,167 @@
+//! The redaction pass (whitepaper §8.6, §11).
+//!
+//! A deterministic pass that strips known secret patterns (API keys, tokens,
+//! private-key blocks, `.env`-style assignments) from node text *before* the
+//! sink. A `--no-content` mode elides all verbatim text and keeps only
+//! structure. Redaction is on by default for known secret patterns.
+
+use crate::node::PreparedNode;
+use regex::Regex;
+
+/// Default secret-detecting patterns, as `(label, pattern)` pairs.
+#[must_use]
+fn default_patterns() -> Vec<(&'static str, &'static str)> {
+    vec![
+        ("anthropic_key", r"sk-ant-[A-Za-z0-9_-]{16,}"),
+        ("openai_key", r"sk-[A-Za-z0-9]{20,}"),
+        ("aws_access_key", r"AKIA[0-9A-Z]{16}"),
+        ("github_token", r"gh[pousr]_[A-Za-z0-9]{20,}"),
+        ("slack_token", r"xox[baprs]-[A-Za-z0-9-]{10,}"),
+        ("google_api_key", r"AIza[0-9A-Za-z_-]{35}"),
+        ("bearer_token", r"(?i)bearer\s+[A-Za-z0-9._~+/-]{16,}=*"),
+        (
+            "assignment_secret",
+            r#"(?i)\b(?:api[_-]?key|secret|token|password|passwd|access[_-]?key)\b\s*[=:]\s*[^\s'"]{6,}"#,
+        ),
+        (
+            "private_key_block",
+            r"-----BEGIN (?:RSA |EC |OPENSSH |DSA |PGP )?PRIVATE KEY-----[\s\S]*?-----END (?:RSA |EC |OPENSSH |DSA |PGP )?PRIVATE KEY-----",
+        ),
+    ]
+}
+
+/// The redactor.
+#[derive(Debug)]
+pub struct Redactor {
+    patterns: Vec<(String, Regex)>,
+    no_content: bool,
+}
+
+impl Default for Redactor {
+    fn default() -> Self {
+        Self::with_default_patterns(false)
+    }
+}
+
+impl Redactor {
+    /// Build a redactor from the default patterns. `no_content` elides all text.
+    ///
+    /// # Panics
+    /// Never in practice — the default patterns are exercised by tests.
+    #[must_use]
+    pub fn with_default_patterns(no_content: bool) -> Self {
+        let patterns = default_patterns()
+            .into_iter()
+            .map(|(label, pat)| {
+                (
+                    label.to_string(),
+                    Regex::new(pat).expect("default redaction patterns must compile"),
+                )
+            })
+            .collect();
+        Redactor {
+            patterns,
+            no_content,
+        }
+    }
+
+    /// Build a redactor from custom `(label, pattern)` pairs.
+    ///
+    /// # Errors
+    /// Returns the regex error if any pattern fails to compile.
+    pub fn from_patterns<S: AsRef<str>>(
+        pairs: impl IntoIterator<Item = (S, S)>,
+        no_content: bool,
+    ) -> Result<Self, regex::Error> {
+        let mut patterns = Vec::new();
+        for (label, pat) in pairs {
+            patterns.push((label.as_ref().to_string(), Regex::new(pat.as_ref())?));
+        }
+        Ok(Redactor {
+            patterns,
+            no_content,
+        })
+    }
+
+    /// Whether `--no-content` mode is on.
+    #[must_use]
+    pub fn is_no_content(&self) -> bool {
+        self.no_content
+    }
+
+    /// Redact a string. In `no_content` mode, returns a structural placeholder.
+    /// Otherwise replaces each secret match with `[REDACTED:<label>]`.
+    /// Deterministic: patterns are applied in a fixed order.
+    #[must_use]
+    pub fn redact_text(&self, s: &str) -> String {
+        if self.no_content {
+            return "[content elided]".to_string();
+        }
+        let mut out = s.to_string();
+        for (label, re) in &self.patterns {
+            out = re
+                .replace_all(&out, format!("[REDACTED:{label}]").as_str())
+                .into_owned();
+        }
+        out
+    }
+
+    /// Whether the text contains any secret (before redaction).
+    #[must_use]
+    pub fn contains_secret(&self, s: &str) -> bool {
+        self.patterns.iter().any(|(_, re)| re.is_match(s))
+    }
+
+    /// Redact a prepared node in place (text, epitome, and diff contents).
+    pub fn redact_node(&self, node: &mut PreparedNode) {
+        match node {
+            PreparedNode::Conversation(c) => {
+                c.text = self.redact_text(&c.text);
+            }
+            PreparedNode::Decision(d) => {
+                d.epitome = self.redact_text(&d.epitome);
+                for opt in &mut d.considered_options {
+                    opt.text = self.redact_text(&opt.text);
+                }
+            }
+            PreparedNode::Episode(e) => {
+                if let Some(old) = e.diff.old.take() {
+                    e.diff.old = Some(self.redact_text(&old));
+                }
+                if let Some(new) = e.diff.new.take() {
+                    e.diff.new = Some(self.redact_text(&new));
+                }
+                if let Some(u) = e.diff.unified.take() {
+                    e.diff.unified = Some(self.redact_text(&u));
+                }
+            }
+            PreparedNode::Binding(_) => {}
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn strips_known_keys() {
+        let r = Redactor::default();
+        let out = r.redact_text("export OPENAI_API_KEY=sk-abcdefghijklmnopqrstuvwx1234");
+        assert!(!out.contains("sk-abcdefghijklmnopqrst"));
+        assert!(out.contains("[REDACTED:"));
+    }
+
+    #[test]
+    fn no_content_elides_everything() {
+        let r = Redactor::with_default_patterns(true);
+        assert_eq!(r.redact_text("anything at all"), "[content elided]");
+    }
+
+    #[test]
+    fn redaction_is_deterministic() {
+        let r = Redactor::default();
+        let s = "token: ghp_abcdefghijklmnopqrstuvwxyz0123456789";
+        assert_eq!(r.redact_text(s), r.redact_text(s));
+    }
+}
diff --git a/crates/memscribe-core/src/segmenter.rs b/crates/memscribe-core/src/segmenter.rs
new file mode 100644
index 0000000..b5520a5
--- /dev/null
+++ b/crates/memscribe-core/src/segmenter.rs
@@ -0,0 +1,511 @@
+//! The segmenter: arc / turn-span bounds (whitepaper §3).
+//!
+//! Given the normalized event stream and the gate, the segmenter bounds
+//! turn-spans, elevates gated turns to [`ConversationSpan`]s, seeds candidate
+//! [`DecisionRecord`]s by parsing the turn text deterministically, and collects
+//! the file edits as [`crate::node::CodeEpisode`]s. It performs no inference —
+//! every field is a verbatim span or a deterministic function of one.
+
+use crate::gate::CommitmentGate;
+use crate::model::{content_id, CaptureEvent, EventKind};
+use crate::node::{CodeEpisode, ConversationSpan, DecisionRecord, FactStatus, NodeId, Opt};
+use std::collections::HashMap;
+use time::OffsetDateTime;
+
+/// A candidate decision seeded from a gated turn, with the metadata the binder
+/// needs to wire PROV edges.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct DecisionCandidate {
+    /// The parsed decision record.
+    pub record: DecisionRecord,
+    /// The deterministic node id.
+    pub node_id: NodeId,
+    /// The turn seq this decision was parsed from.
+    pub turn_seq: u64,
+    /// The decision's timestamp (used as `t_use`).
+    pub timestamp: OffsetDateTime,
+    /// The session the decision belongs to.
+    pub session_id: String,
+}
+
+/// A code edit episode with the metadata the binder needs.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct EpisodeRecord {
+    /// The prepared episode.
+    pub episode: CodeEpisode,
+    /// The deterministic node id.
+    pub node_id: NodeId,
+    /// The seq of the originating edit event.
+    pub seq: u64,
+    /// The episode timestamp (used as `t_gen`).
+    pub timestamp: OffsetDateTime,
+    /// The session the episode belongs to.
+    pub session_id: String,
+}
+
+/// The result of segmentation: conversations, decision candidates, episodes.
+#[derive(Clone, Debug, Default, PartialEq, Eq)]
+pub struct Segmentation {
+    /// Gated, verbatim dialogue spans.
+    pub conversations: Vec<ConversationSpan>,
+    /// Candidate decisions seeded from gated turns.
+    pub decisions: Vec<DecisionCandidate>,
+    /// Code edit episodes.
+    pub episodes: Vec<EpisodeRecord>,
+}
+
+/// The segmenter stage.
+pub trait Segmenter {
+    /// Segment a per-session event stream (sorted by `seq`) into spans.
+    fn segment(&self, events: &[CaptureEvent], gate: &CommitmentGate) -> Segmentation;
+}
+
+/// The default deterministic segmenter.
+#[derive(Debug, Default)]
+pub struct DefaultSegmenter;
+
+impl Segmenter for DefaultSegmenter {
+    fn segment(&self, events: &[CaptureEvent], gate: &CommitmentGate) -> Segmentation {
+        let mut seg = Segmentation::default();
+
+        // A FileEdit whose tool call failed (ToolResult.ok == false) must not
+        // become an episode — "a tool failure → no spurious episode" (§8.2).
+        let mut call_ok: HashMap<String, bool> = HashMap::new();
+        for ev in events {
+            if let EventKind::ToolResult { call_id, ok, .. } = &ev.kind {
+                call_ok.insert(call_id.clone(), *ok);
+            }
+        }
+
+        // Rewind / Compaction supersede markers (§8.2): the verbatim history of
+        // the affected turns is still emitted, but in the *current view* any
+        // decision whose source turn falls in a rewound-away region or inside a
+        // Compaction.replaced range no longer governs current edits. We resolve
+        // those regions here, keyed by `(session_id, turn_seq)`, so each parsed
+        // decision can be stamped with `superseded_by` deterministically.
+        let supersedes = resolve_supersede_markers(events);
+
+        for ev in events {
+            match &ev.kind {
+                EventKind::UserTurn { text, .. } => {
+                    let markers = gate.evaluate(text);
+                    if markers.is_empty() {
+                        continue; // retained verbatim at the event layer; no node
+                    }
+                    let turn_range = ev.seq..ev.seq + 1;
+                    seg.conversations.push(ConversationSpan {
+                        session_id: ev.session_id.clone(),
+                        turn_range: turn_range.clone(),
+                        text: text.clone(),
+                        markers: markers.clone(),
+                        fact_status: FactStatus::Observed,
+                        provenance: vec![ev.provenance.clone()],
+                    });
+
+                    let is_ban = gate.is_ban(&markers);
+                    // A decision whose source turn was rewound away or compacted
+                    // out is superseded in the current view.
+                    let superseded_by = supersedes.get(&(ev.session_id.clone(), ev.seq)).cloned();
+                    let record = DecisionRecord {
+                        epitome: epitome_of(text, markers.first().map(|m| m.offset).unwrap_or(0)),
+                        considered_options: parse_options(text),
+                        is_ban,
+                        superseded_by,
+                        confirmation: None,
+                        source_span: turn_range,
+                        // Observed for the verbatim text; uncertain element
+                        // typing is flagged downstream, never guessed here.
+                        fact_status: FactStatus::Observed,
+                    };
+                    seg.decisions.push(DecisionCandidate {
+                        node_id: NodeId::new(format!("decision:{}:{}", ev.session_id, ev.seq)),
+                        record,
+                        turn_seq: ev.seq,
+                        timestamp: ev.timestamp,
+                        session_id: ev.session_id.clone(),
+                    });
+                }
+                EventKind::FileEdit { call_id, diff } => {
+                    // Drop edits from a failed tool call.
+                    if let Some(cid) = call_id {
+                        if call_ok.get(cid) == Some(&false) {
+                            continue;
+                        }
+                    }
+                    let episode_id = content_id(
+                        format!("{}:{}:{}", ev.session_id, ev.seq, diff.path.display()).as_bytes(),
+                    );
+                    let git = ev.project.git.clone();
+                    seg.episodes.push(EpisodeRecord {
+                        episode: CodeEpisode {
+                            path: diff.path.clone(),
+                            diff: diff.clone(),
+                            git,
+                            episode_id: episode_id.clone(),
+                        },
+                        node_id: NodeId::new(format!("episode:{}", episode_id)),
+                        seq: ev.seq,
+                        timestamp: ev.timestamp,
+                        session_id: ev.session_id.clone(),
+                    });
+                }
+                _ => {}
+            }
+        }
+
+        seg
+    }
+}
+
+/// Resolve every `(session_id, turn_seq)` that is rewound away or compacted out
+/// to a deterministic supersede-marker [`NodeId`] (§8.2).
+///
+/// Semantics:
+/// - **Rewind**: a `Rewind { to_event }` at seq `R` logically truncates the
+///   session back to the event whose `event_id == to_event`, at seq `T`. Every
+///   turn strictly after the target and up to (and including) the rewind point —
+///   the half-open interval `(T, R]` in seq terms — is rewound away. The target
+///   turn itself survives (we rewound *to* it). The marker id is
+///   `rewind:<session>:<R>`. If `to_event` cannot be resolved to a seq in the
+///   same session, the rewind is a flagged no-op (no turns superseded) — never
+///   a panic.
+/// - **Compaction**: a `Compaction { replaced }` carries a `[start, end)` seq
+///   range directly; every turn seq in that range is compacted out. The marker
+///   id is `compaction:<session>:<start>-<end>`.
+///
+/// When a turn is covered by multiple markers, the *last* event in stream order
+/// wins (the most recent truncation governs the current view), keeping the
+/// result a deterministic function of event order.
+fn resolve_supersede_markers(events: &[CaptureEvent]) -> HashMap<(String, u64), NodeId> {
+    // event_id → seq, per session, so a Rewind target resolves to a turn seq.
+    let mut seq_of: HashMap<(&str, &str), u64> = HashMap::new();
+    for ev in events {
+        seq_of
+            .entry((ev.session_id.as_str(), ev.event_id.as_str()))
+            .or_insert(ev.seq);
+    }
+
+    let mut out: HashMap<(String, u64), NodeId> = HashMap::new();
+    for ev in events {
+        match &ev.kind {
+            EventKind::Rewind { to_event } => {
+                let rewind_seq = ev.seq;
+                let Some(&target_seq) = seq_of.get(&(ev.session_id.as_str(), to_event.as_str()))
+                else {
+                    // Unknown target → flagged no-op, panic-free.
+                    continue;
+                };
+                if target_seq >= rewind_seq {
+                    // Target at or after the rewind point: nothing to truncate.
+                    continue;
+                }
+                let marker = NodeId::new(format!("rewind:{}:{}", ev.session_id, rewind_seq));
+                for other in events {
+                    if other.session_id == ev.session_id
+                        && other.seq > target_seq
+                        && other.seq <= rewind_seq
+                    {
+                        out.insert((other.session_id.clone(), other.seq), marker.clone());
+                    }
+                }
+            }
+            EventKind::Compaction { replaced } => {
+                let marker = NodeId::new(format!(
+                    "compaction:{}:{}-{}",
+                    ev.session_id, replaced.start, replaced.end
+                ));
+                for other in events {
+                    if other.session_id == ev.session_id && replaced.contains(&other.seq) {
+                        out.insert((other.session_id.clone(), other.seq), marker.clone());
+                    }
+                }
+            }
+            _ => {}
+        }
+    }
+    out
+}
+
+/// Extract the decision sentence containing `offset` — a verbatim span, bounded
+/// by sentence terminators. Deterministic.
+fn epitome_of(text: &str, offset: usize) -> String {
+    let bytes = text.as_bytes();
+    let offset = offset.min(text.len());
+    // Walk back to the start of the sentence.
+    let mut start = 0usize;
+    for i in (0..offset).rev() {
+        if matches!(bytes[i], b'.' | b'!' | b'?' | b'\n') {
+            start = i + 1;
+            break;
+        }
+    }
+    // Walk forward to the end of the sentence.
+    let mut end = text.len();
+    for (i, b) in bytes.iter().enumerate().skip(offset) {
+        if matches!(b, b'.' | b'!' | b'?' | b'\n') {
+            end = i + 1;
+            break;
+        }
+    }
+    // Snap to char boundaries to stay panic-free on multibyte input.
+    while start < text.len() && !text.is_char_boundary(start) {
+        start += 1;
+    }
+    while end < text.len() && !text.is_char_boundary(end) {
+        end += 1;
+    }
+    text[start..end].trim().to_string()
+}
+
+/// Deterministically parse considered options from decision prose: the chosen
+/// option from a use/go-with verb, and the rejected option from "instead of X".
+fn parse_options(text: &str) -> Vec<Opt> {
+    let mut opts = Vec::new();
+
+    // Rejected: "instead of X" / "rather than X".
+    for marker in ["instead of", "rather than", "as opposed to"] {
+        if let Some(opt) = capture_after(text, marker) {
+            opts.push(Opt {
+                text: opt,
+                chosen: false,
+            });
+        }
+    }
+
+    // Chosen: "use X" / "go with X" / "switch to X" / "adopt X".
+    for marker in ["go with", "switch to", "migrate to", "adopt", "use"] {
+        if let Some(opt) = capture_after(text, marker) {
+            // Avoid double-listing the rejected phrase.
+            if !opts.iter().any(|o| o.text.eq_ignore_ascii_case(&opt)) {
+                opts.push(Opt {
+                    text: opt,
+                    chosen: true,
+                });
+                break;
+            }
+        }
+    }
+
+    opts
+}
+
+/// Capture the noun phrase immediately after a marker phrase, up to a clause or
+/// sentence boundary. Case-insensitive match, verbatim capture. Deterministic.
+fn capture_after(text: &str, marker: &str) -> Option<String> {
+    let lower = text.to_ascii_lowercase();
+    let pos = lower.find(marker)?;
+    let after = pos + marker.len();
+    // Walk to a char boundary.
+    let mut start = after;
+    while start < text.len() && !text.is_char_boundary(start) {
+        start += 1;
+    }
+    let rest = &text[start..];
+    let trimmed_start = rest.len() - rest.trim_start().len();
+    let phrase: String = rest[trimmed_start..]
+        .chars()
+        .take_while(|c| !matches!(c, '.' | ',' | '!' | '?' | ';' | ':' | '\n'))
+        .collect();
+    let phrase = phrase.trim();
+    // Capture just the option token (typically a single library/service name,
+    // possibly hyphenated/dotted like `left-pad`). Keeping it to one token stays
+    // deterministic and avoids swallowing the rest of the sentence.
+    phrase.split_whitespace().next().map(str::to_string)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn options_parse_chosen_and_rejected() {
+        let opts = parse_options("Let's use Stripe instead of PayPal for billing.");
+        assert!(opts.iter().any(|o| o.text == "Stripe" && o.chosen));
+        assert!(opts.iter().any(|o| o.text == "PayPal" && !o.chosen));
+    }
+
+    #[test]
+    fn epitome_is_the_containing_sentence() {
+        let t = "Some preamble. We must use prepared statements. Thanks.";
+        let e = epitome_of(t, t.find("must").unwrap());
+        assert_eq!(e, "We must use prepared statements.");
+    }
+
+    #[test]
+    fn epitome_is_panic_free_on_multibyte() {
+        let t = "café — we will use Postgres ☕";
+        let _ = epitome_of(t, 3);
+    }
+
+    // ---- Task A: rewind / compaction supersede-marking ----
+
+    use crate::binder::{Binder, DefaultBinder};
+    use crate::model::{
+        CaptureEvent, Diff, EventKind, ProjectRef, SourceKind, SourceLocation, SCHEMA_VERSION,
+    };
+    use time::OffsetDateTime;
+
+    fn ts(secs: i64) -> OffsetDateTime {
+        OffsetDateTime::from_unix_timestamp(1_700_000_000 + secs).unwrap()
+    }
+
+    fn ev(seq: u64, event_id: &str, kind: EventKind) -> CaptureEvent {
+        CaptureEvent {
+            schema_version: SCHEMA_VERSION,
+            source: SourceKind::ClaudeCode,
+            session_id: "s1".to_string(),
+            seq,
+            event_id: event_id.to_string(),
+            parent_id: None,
+            timestamp: ts(seq as i64),
+            project: ProjectRef::from_cwd("/repo"),
+            kind,
+            provenance: SourceLocation::new("t.jsonl", 0, seq + 1),
+        }
+    }
+
+    fn user(seq: u64, event_id: &str, text: &str) -> CaptureEvent {
+        ev(
+            seq,
+            event_id,
+            EventKind::UserTurn {
+                text: text.to_string(),
+                parts: vec![],
+            },
+        )
+    }
+
+    fn edit(seq: u64, path: &str) -> CaptureEvent {
+        ev(
+            seq,
+            &format!("edit-{seq}"),
+            EventKind::FileEdit {
+                call_id: None,
+                diff: Diff::for_path(path),
+            },
+        )
+    }
+
+    #[test]
+    fn rewind_supersedes_decisions_in_rewound_region_and_binds_only_to_survivor() {
+        // Decision A at turn 1; rewind back to turn 1 happens at turn 3 (so turn 2
+        // is rewound away); decision B at turn 5; edit at turn 6.
+        let events = vec![
+            user(1, "m1", "Let's use Postgres for storage."),
+            user(2, "m2", "Actually we must use MySQL instead."),
+            ev(
+                3,
+                "r1",
+                EventKind::Rewind {
+                    to_event: "m1".into(),
+                },
+            ),
+            user(5, "m5", "We will use Redis for the cache."),
+            edit(6, "cache.rs"),
+        ];
+        let gate = CommitmentGate::default_table();
+        let seg = DefaultSegmenter.segment(&events, &gate);
+
+        // Turn 2's decision is in the rewound region (2 in (1, 3]) → superseded.
+        let d2 = seg
+            .decisions
+            .iter()
+            .find(|d| d.turn_seq == 2)
+            .expect("turn-2 decision exists");
+        assert_eq!(
+            d2.record.superseded_by,
+            Some(NodeId::new("rewind:s1:3")),
+            "turn-2 decision must be superseded by the rewind marker"
+        );
+
+        // The rewound-to target (turn 1) survives.
+        let d1 = seg.decisions.iter().find(|d| d.turn_seq == 1).unwrap();
+        assert!(d1.record.superseded_by.is_none());
+
+        // Decision B (turn 5) survives.
+        let d5 = seg.decisions.iter().find(|d| d.turn_seq == 5).unwrap();
+        assert!(d5.record.superseded_by.is_none());
+
+        // Verbatim conversation span for the superseded turn 2 is still present.
+        assert!(
+            seg.conversations.iter().any(|c| c.turn_range.start == 2),
+            "verbatim span for the rewound turn must be preserved"
+        );
+
+        // The edit binds to the most-recent NON-superseded decision (B), never A2.
+        let edges = DefaultBinder.bind(&seg);
+        assert_eq!(edges.len(), 1);
+        assert_eq!(edges[0].from, NodeId::new("decision:s1:5"));
+    }
+
+    #[test]
+    fn compaction_range_supersedes_contained_decisions() {
+        // Compaction replaces seqs [1, 3): turns 1 and 2 are compacted out.
+        let events = vec![
+            user(1, "m1", "Let's use Postgres for storage."),
+            user(2, "m2", "We must always use prepared statements."),
+            ev(3, "c1", EventKind::Compaction { replaced: 1..3 }),
+            user(4, "m4", "We will use Redis for the cache."),
+        ];
+        let gate = CommitmentGate::default_table();
+        let seg = DefaultSegmenter.segment(&events, &gate);
+
+        let marker = NodeId::new("compaction:s1:1-3");
+        for seq in [1u64, 2] {
+            let d = seg.decisions.iter().find(|d| d.turn_seq == seq).unwrap();
+            assert_eq!(
+                d.record.superseded_by,
+                Some(marker.clone()),
+                "turn {seq} should be compacted out"
+            );
+        }
+        // Turn 4 is outside the range → survives.
+        let d4 = seg.decisions.iter().find(|d| d.turn_seq == 4).unwrap();
+        assert!(d4.record.superseded_by.is_none());
+
+        // Verbatim spans for the compacted turns are still present (lossless).
+        assert!(seg.conversations.iter().any(|c| c.turn_range.start == 1));
+        assert!(seg.conversations.iter().any(|c| c.turn_range.start == 2));
+    }
+
+    #[test]
+    fn rewind_to_unknown_target_is_a_flagged_no_op() {
+        // The rewind target does not exist → no decision is superseded; panic-free.
+        let events = vec![
+            user(1, "m1", "Let's use Postgres for storage."),
+            ev(
+                2,
+                "r1",
+                EventKind::Rewind {
+                    to_event: "does-not-exist".into(),
+                },
+            ),
+        ];
+        let gate = CommitmentGate::default_table();
+        let seg = DefaultSegmenter.segment(&events, &gate);
+        assert!(seg
+            .decisions
+            .iter()
+            .all(|d| d.record.superseded_by.is_none()));
+    }
+
+    #[test]
+    fn supersede_marking_is_deterministic_across_runs() {
+        let events = vec![
+            user(1, "m1", "Let's use Postgres."),
+            user(2, "m2", "We must use MySQL instead."),
+            ev(
+                3,
+                "r1",
+                EventKind::Rewind {
+                    to_event: "m1".into(),
+                },
+            ),
+            ev(4, "c1", EventKind::Compaction { replaced: 2..3 }),
+        ];
+        let gate = CommitmentGate::default_table();
+        let a = DefaultSegmenter.segment(&events, &gate);
+        let b = DefaultSegmenter.segment(&events, &gate);
+        assert_eq!(a, b);
+    }
+}
diff --git a/crates/memscribe-core/src/sink.rs b/crates/memscribe-core/src/sink.rs
new file mode 100644
index 0000000..b15a56a
--- /dev/null
+++ b/crates/memscribe-core/src/sink.rs
@@ -0,0 +1,59 @@
+//! The [`Sink`] trait — the single seam that decouples Memscribe from MemDB
+//! (whitepaper §6).
+//!
+//! Nothing in the pipeline knows what a sink does with a node. Concrete sinks
+//! (NDJSON, SQLite, and a feature-gated MemDB sink) live in `memscribe-sink`.
+//! Because the canonical default is NDJSON, the entire module is observable and
+//! testable without MemDB present.
+
+use crate::error::SinkError;
+use crate::node::PreparedNode;
+
+/// A consumer of prepared nodes.
+pub trait Sink: Send {
+    /// Emit one prepared node.
+    fn emit(&mut self, node: &PreparedNode) -> Result<(), SinkError>;
+    /// Flush any buffered nodes.
+    fn flush(&mut self) -> Result<(), SinkError>;
+
+    /// Emit every node in a slice, then flush. Convenience for batch use. Kept
+    /// object-safe (concrete slice, no generics) so `&mut dyn Sink` works.
+    fn emit_all(&mut self, nodes: &[PreparedNode]) -> Result<(), SinkError> {
+        for node in nodes {
+            self.emit(node)?;
+        }
+        self.flush()
+    }
+}
+
+/// An in-memory sink that collects nodes. Useful for tests, the conformance
+/// harness, and `replay`.
+#[derive(Debug, Default, Clone)]
+pub struct VecSink {
+    /// The collected nodes, in emission order.
+    pub nodes: Vec<PreparedNode>,
+}
+
+impl VecSink {
+    /// A fresh, empty collecting sink.
+    #[must_use]
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Consume the sink and return the collected nodes.
+    #[must_use]
+    pub fn into_nodes(self) -> Vec<PreparedNode> {
+        self.nodes
+    }
+}
+
+impl Sink for VecSink {
+    fn emit(&mut self, node: &PreparedNode) -> Result<(), SinkError> {
+        self.nodes.push(node.clone());
+        Ok(())
+    }
+    fn flush(&mut self) -> Result<(), SinkError> {
+        Ok(())
+    }
+}
diff --git a/crates/memscribe-core/tests/pipeline_properties.rs b/crates/memscribe-core/tests/pipeline_properties.rs
new file mode 100644
index 0000000..7eef0bd
--- /dev/null
+++ b/crates/memscribe-core/tests/pipeline_properties.rs
@@ -0,0 +1,223 @@
+//! Whitepaper §8.3 pipeline invariants as `proptest` properties.
+//!
+//! These drive `DefaultPipeline::without_redaction().prepare_events` over
+//! randomly-constructed `Vec<CaptureEvent>` (built directly, not parsed from
+//! bytes) and assert the pipeline's core guarantees:
+//!
+//! - **Determinism** — `prepare_events(x)` is byte-identical across runs.
+//! - **Idempotency** — preparing the same event stream twice yields identical
+//!   node sets (the pipeline is a pure function with no accumulating state).
+//! - **Temporal validity** — every emitted `Binding` carries a PROV record with
+//!   `t_use <= t_gen` (`prov.is_temporally_valid()`), the §8.8 invariant the
+//!   binder is responsible for.
+//!
+//! Events are generated with strictly increasing `seq` and non-decreasing
+//! timestamps within a session, mirroring how a real adapter assigns them from
+//! file order — the precondition the binder's `t_use <= t_gen` search relies on.
+
+use memscribe_core::model::{Diff, EventKind};
+use memscribe_core::{
+    CaptureEvent, DefaultPipeline, PreparedNode, ProjectRef, SourceKind, SourceLocation,
+    SCHEMA_VERSION,
+};
+use proptest::prelude::*;
+use std::path::PathBuf;
+use time::OffsetDateTime;
+
+/// A base instant; per-event timestamps are this plus the event's `seq` seconds,
+/// so timestamps are non-decreasing in lockstep with `seq` within a session.
+fn base_time() -> OffsetDateTime {
+    // 2026-06-22T12:00:00Z, expressed via Unix time to avoid macro feature reqs.
+    OffsetDateTime::from_unix_timestamp(1_781_697_600).expect("valid unix timestamp")
+}
+
+/// The kind of synthetic event to generate. The two that drive the
+/// segmenter/binder are a decision-bearing user turn and a file edit; the rest
+/// are filler that must pass through without producing spurious bindings.
+#[derive(Clone, Debug)]
+enum GenEvent {
+    /// A user turn whose text fires (or does not fire) the commitment gate.
+    UserTurn(String),
+    /// A file edit on a path — becomes a `CodeEpisode`.
+    FileEdit(String),
+    /// An assistant turn (filler; no node).
+    Assistant(String),
+}
+
+/// Decision-bearing and plain user-turn texts, so the gate is exercised both
+/// ways (admitted → Decision/Conversation; plain → no node).
+fn user_text() -> impl Strategy<Value = String> {
+    prop_oneof![
+        Just("let's go with Postgres for storage".to_string()),
+        Just("use Stripe instead of PayPal".to_string()),
+        Just("we will never add a dependency on left-pad".to_string()),
+        Just("we must always use prepared statements".to_string()),
+        Just("thanks, that looks good to me".to_string()),
+        "[a-zA-Z ]{0,40}".prop_map(|s| s),
+    ]
+}
+
+fn edit_path() -> impl Strategy<Value = String> {
+    prop_oneof![
+        Just("src/main.rs".to_string()),
+        Just("src/lib.rs".to_string()),
+        Just("Cargo.toml".to_string()),
+        "[a-z]{1,8}\\.rs".prop_map(|s| format!("src/{s}")),
+    ]
+}
+
+fn gen_event() -> impl Strategy<Value = GenEvent> {
+    prop_oneof![
+        user_text().prop_map(GenEvent::UserTurn),
+        edit_path().prop_map(GenEvent::FileEdit),
+        "[a-z ]{0,20}".prop_map(GenEvent::Assistant),
+    ]
+}
+
+/// Build a real `Vec<CaptureEvent>` from generated descriptors, assigning a
+/// strictly-increasing `seq` and a non-decreasing timestamp (base + seq seconds)
+/// within a single session — the shape an adapter produces from file order.
+fn build_events(session_id: &str, gen: &[GenEvent]) -> Vec<CaptureEvent> {
+    let base = base_time();
+    let project = ProjectRef::from_cwd(".");
+    gen.iter()
+        .enumerate()
+        .map(|(i, g)| {
+            let seq = i as u64;
+            let timestamp = base + time::Duration::seconds(seq as i64);
+            let (event_id, kind) = match g {
+                GenEvent::UserTurn(text) => (
+                    format!("u-{seq}"),
+                    EventKind::UserTurn {
+                        text: text.clone(),
+                        parts: Vec::new(),
+                    },
+                ),
+                GenEvent::FileEdit(path) => (
+                    format!("e-{seq}"),
+                    EventKind::FileEdit {
+                        call_id: None,
+                        diff: Diff::for_path(path.as_str()),
+                    },
+                ),
+                GenEvent::Assistant(text) => (
+                    format!("a-{seq}"),
+                    EventKind::AssistantTurn {
+                        text: text.clone(),
+                        thinking: None,
+                        model: None,
+                        usage: None,
+                        parts: Vec::new(),
+                    },
+                ),
+            };
+            CaptureEvent {
+                schema_version: SCHEMA_VERSION,
+                source: SourceKind::ClaudeCode,
+                session_id: session_id.to_string(),
+                seq,
+                event_id,
+                parent_id: None,
+                timestamp,
+                project: project.clone(),
+                kind,
+                provenance: SourceLocation::new(PathBuf::from("gen.jsonl"), seq, seq + 1),
+            }
+        })
+        .collect()
+}
+
+proptest! {
+    #![proptest_config(ProptestConfig::with_cases(256))]
+
+    /// Determinism: `prepare_events` is a pure function — two runs over the same
+    /// event stream serialize byte-identically.
+    #[test]
+    fn prepare_events_is_deterministic(gen in proptest::collection::vec(gen_event(), 0..16)) {
+        let events = build_events("s1", &gen);
+        let pipeline = DefaultPipeline::without_redaction();
+
+        let a = pipeline.prepare_events(&events);
+        let b = pipeline.prepare_events(&events);
+
+        let ja = serde_json::to_string(&a).unwrap();
+        let jb = serde_json::to_string(&b).unwrap();
+        prop_assert_eq!(ja, jb);
+    }
+
+    /// Idempotency: a freshly-constructed pipeline over the same events yields an
+    /// identical node set — preparing twice does not accumulate state or drift.
+    #[test]
+    fn prepare_events_is_idempotent(gen in proptest::collection::vec(gen_event(), 0..16)) {
+        let events = build_events("s1", &gen);
+
+        let first = DefaultPipeline::without_redaction().prepare_events(&events);
+        let second = DefaultPipeline::without_redaction().prepare_events(&events);
+
+        // Same pipeline instance, applied again, must also match.
+        let pipeline = DefaultPipeline::without_redaction();
+        let third = pipeline.prepare_events(&events);
+        let fourth = pipeline.prepare_events(&events);
+
+        prop_assert_eq!(
+            serde_json::to_string(&first).unwrap(),
+            serde_json::to_string(&second).unwrap()
+        );
+        prop_assert_eq!(
+            serde_json::to_string(&third).unwrap(),
+            serde_json::to_string(&fourth).unwrap()
+        );
+        prop_assert_eq!(
+            serde_json::to_string(&first).unwrap(),
+            serde_json::to_string(&third).unwrap()
+        );
+    }
+
+    /// Temporal validity (§8.8): every emitted `Binding` satisfies the PROV
+    /// invariant `t_use <= t_gen` via `prov.is_temporally_valid()`.
+    #[test]
+    fn every_binding_is_temporally_valid(
+        gen in proptest::collection::vec(gen_event(), 0..24),
+    ) {
+        let events = build_events("s1", &gen);
+        let nodes = DefaultPipeline::without_redaction().prepare_events(&events);
+
+        for node in &nodes {
+            if let PreparedNode::Binding(edge) = node {
+                prop_assert!(
+                    edge.prov.is_temporally_valid(),
+                    "binding {:?} -> {:?} violates t_use <= t_gen: t_use={}, t_gen={}",
+                    edge.from,
+                    edge.to,
+                    edge.prov.t_use,
+                    edge.prov.t_gen
+                );
+            }
+        }
+    }
+
+    /// Cross-session temporal validity: even with two interleaved sessions, every
+    /// binding's PROV invariant holds (the binder must never bind across
+    /// sessions, so `t_use <= t_gen` cannot be violated by session interleaving).
+    #[test]
+    fn bindings_valid_across_two_sessions(
+        a in proptest::collection::vec(gen_event(), 0..12),
+        b in proptest::collection::vec(gen_event(), 0..12),
+    ) {
+        let mut events = build_events("sA", &a);
+        events.extend(build_events("sB", &b));
+
+        let nodes = DefaultPipeline::without_redaction().prepare_events(&events);
+
+        for node in &nodes {
+            if let PreparedNode::Binding(edge) = node {
+                prop_assert!(edge.prov.is_temporally_valid());
+                // The binder binds within a session only.
+                prop_assert_eq!(
+                    &edge.prov.used_session,
+                    &edge.prov.was_generated_by_session
+                );
+            }
+        }
+    }
+}
diff --git a/crates/memscribe-io/Cargo.toml b/crates/memscribe-io/Cargo.toml
new file mode 100644
index 0000000..2379c21
--- /dev/null
+++ b/crates/memscribe-io/Cargo.toml
@@ -0,0 +1,38 @@
+[package]
+name = "memscribe-io"
+version.workspace = true
+edition.workspace = true
+rust-version.workspace = true
+license.workspace = true
+repository.workspace = true
+description = "Source layer for Memscribe: file reader/tailer (offset resume), hook server, and an optional OTLP receiver."
+
+[lints]
+workspace = true
+
+[features]
+default = []
+# The live notify-based tailer + persisted offset cursor store.
+watch = ["dep:notify", "dep:notify-debouncer-full", "dep:tokio"]
+# The persisted offset store backed by SQLite.
+cursor-store = ["dep:rusqlite"]
+# A loopback-only HTTP OTLP receiver (std TcpListener, no async/tonic deps).
+# Off by default: the default build opens no network ports (whitepaper §11).
+# Reuses the already-present serde_json dependency; no new crates pulled in.
+otlp = []
+
+[dependencies]
+memscribe-core = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+time = { workspace = true }
+tracing = { workspace = true }
+zstd = { workspace = true }
+walkdir = { workspace = true }
+tokio = { workspace = true, optional = true }
+notify = { workspace = true, optional = true }
+notify-debouncer-full = { workspace = true, optional = true }
+rusqlite = { workspace = true, optional = true }
+
+[dev-dependencies]
+tempfile = { workspace = true }
diff --git a/crates/memscribe-io/src/cursor_store.rs b/crates/memscribe-io/src/cursor_store.rs
new file mode 100644
index 0000000..7f00851
--- /dev/null
+++ b/crates/memscribe-io/src/cursor_store.rs
@@ -0,0 +1,96 @@
+//! The persisted byte-offset cursor store. The tailer keeps one offset per
+//! transcript file so a restart resumes exactly where it left off — no
+//! duplicates, no loss (whitepaper §7, §8.5).
+
+use std::collections::HashMap;
+
+/// A keyed store of byte offsets (key = a transcript file identity).
+pub trait OffsetStore {
+    /// The last persisted offset for `key`, if any.
+    fn get(&self, key: &str) -> Option<u64>;
+    /// Persist `offset` for `key`.
+    fn set(&mut self, key: &str, offset: u64);
+}
+
+/// An in-memory offset store (tests, ephemeral runs).
+#[derive(Debug, Default, Clone)]
+pub struct MemoryOffsetStore {
+    map: HashMap<String, u64>,
+}
+
+impl MemoryOffsetStore {
+    /// A fresh, empty store.
+    #[must_use]
+    pub fn new() -> Self {
+        Self::default()
+    }
+}
+
+impl OffsetStore for MemoryOffsetStore {
+    fn get(&self, key: &str) -> Option<u64> {
+        self.map.get(key).copied()
+    }
+    fn set(&mut self, key: &str, offset: u64) {
+        self.map.insert(key.to_string(), offset);
+    }
+}
+
+/// A SQLite-backed persistent offset store (feature `cursor-store`).
+#[cfg(feature = "cursor-store")]
+pub mod persistent {
+    use super::OffsetStore;
+    use rusqlite::Connection;
+    use std::path::Path;
+
+    /// A durable offset store backed by SQLite.
+    pub struct SqliteOffsetStore {
+        conn: Connection,
+    }
+
+    impl SqliteOffsetStore {
+        /// Open (or create) the offset store at `path`.
+        ///
+        /// # Errors
+        /// Returns a rusqlite error if the database cannot be opened.
+        pub fn open(path: impl AsRef<Path>) -> rusqlite::Result<Self> {
+            let conn = Connection::open(path)?;
+            conn.execute_batch(
+                "CREATE TABLE IF NOT EXISTS offsets (key TEXT PRIMARY KEY, offset INTEGER NOT NULL);",
+            )?;
+            Ok(SqliteOffsetStore { conn })
+        }
+    }
+
+    impl OffsetStore for SqliteOffsetStore {
+        fn get(&self, key: &str) -> Option<u64> {
+            self.conn
+                .query_row("SELECT offset FROM offsets WHERE key = ?1", [key], |r| {
+                    r.get::<_, i64>(0)
+                })
+                .ok()
+                .map(|v| v as u64)
+        }
+        fn set(&mut self, key: &str, offset: u64) {
+            let _ = self.conn.execute(
+                "INSERT INTO offsets(key, offset) VALUES(?1, ?2)
+                 ON CONFLICT(key) DO UPDATE SET offset = ?2",
+                rusqlite::params![key, offset as i64],
+            );
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn memory_store_roundtrips() {
+        let mut s = MemoryOffsetStore::new();
+        assert_eq!(s.get("a"), None);
+        s.set("a", 42);
+        assert_eq!(s.get("a"), Some(42));
+        s.set("a", 100);
+        assert_eq!(s.get("a"), Some(100));
+    }
+}
diff --git a/crates/memscribe-io/src/discover.rs b/crates/memscribe-io/src/discover.rs
new file mode 100644
index 0000000..4aebfb2
--- /dev/null
+++ b/crates/memscribe-io/src/discover.rs
@@ -0,0 +1,139 @@
+//! Transcript discovery.
+//!
+//! A small, dependency-light helper that walks a directory tree and collects the
+//! transcript files a [`crate::tailer::LiveTailer`] should watch (or the one-shot
+//! reader should replay). It is generic over the set of extensions a tool uses
+//! (`jsonl`, `json`, `zst`, ...) and reports, per file, whether it is a `.zst`
+//! cold rollout so the caller can route it to the decompressing reader.
+//!
+//! Discovery is deterministic (results are sorted by path), panic-free, and
+//! tolerant of unreadable subtrees: directories it cannot descend are skipped
+//! rather than aborting the whole walk.
+
+use std::path::{Path, PathBuf};
+use walkdir::WalkDir;
+
+/// One discovered transcript file.
+#[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord)]
+pub struct Discovered {
+    /// Absolute (or root-relative) path to the transcript file.
+    pub path: PathBuf,
+    /// Whether the file is a `.zst` (Codex cold rollout) needing decompression.
+    pub is_zst: bool,
+}
+
+/// Whether a path's extension matches one of `exts` (case-insensitive, no dot).
+fn ext_matches(path: &Path, exts: &[&str]) -> bool {
+    match path.extension().and_then(|e| e.to_str()) {
+        Some(ext) => exts.iter().any(|want| want.eq_ignore_ascii_case(ext)),
+        None => false,
+    }
+}
+
+/// Recursively find transcript files under `root` whose extension is in `exts`.
+///
+/// `exts` are bare extensions without the leading dot (e.g. `["jsonl", "zst"]`)
+/// and are matched case-insensitively. Symlinks are not followed (avoids cycles
+/// and surprise escapes from the watched tree). The result is sorted for
+/// determinism. A `.zst` file is flagged via [`Discovered::is_zst`] regardless of
+/// whether `"zst"` itself was requested — e.g. `transcript.jsonl.zst` matched by
+/// the `"zst"` ext is still a zst.
+///
+/// Unreadable entries are silently skipped; this never panics and never returns
+/// an error — a missing or non-directory `root` simply yields an empty list.
+#[must_use]
+pub fn find_transcripts(root: impl AsRef<Path>, exts: &[&str]) -> Vec<Discovered> {
+    let mut out: Vec<Discovered> = WalkDir::new(root.as_ref())
+        .follow_links(false)
+        .into_iter()
+        .filter_map(Result::ok)
+        .filter(|e| e.file_type().is_file())
+        .map(|e| e.into_path())
+        .filter(|p| ext_matches(p, exts))
+        .map(|path| {
+            let is_zst = path.extension().and_then(|e| e.to_str()) == Some("zst");
+            Discovered { path, is_zst }
+        })
+        .collect();
+    out.sort();
+    out.dedup();
+    out
+}
+
+/// The just the paths of [`find_transcripts`], convenient for seeding a watcher.
+#[must_use]
+pub fn find_transcript_paths(root: impl AsRef<Path>, exts: &[&str]) -> Vec<PathBuf> {
+    find_transcripts(root, exts)
+        .into_iter()
+        .map(|d| d.path)
+        .collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::fs;
+
+    fn touch(path: &Path) {
+        if let Some(parent) = path.parent() {
+            fs::create_dir_all(parent).unwrap();
+        }
+        fs::write(path, b"x").unwrap();
+    }
+
+    #[test]
+    fn finds_nested_by_extension_and_flags_zst() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path();
+        touch(&root.join("a.jsonl"));
+        touch(&root.join("nested/b.jsonl"));
+        touch(&root.join("nested/deep/c.jsonl.zst"));
+        touch(&root.join("ignore.txt"));
+        touch(&root.join("nested/ignore.md"));
+
+        let found = find_transcripts(root, &["jsonl", "zst"]);
+        assert_eq!(found.len(), 3, "three transcripts, txt/md ignored");
+
+        // Deterministic order (sorted by path).
+        let mut sorted = found.clone();
+        sorted.sort();
+        assert_eq!(found, sorted);
+
+        let zst: Vec<_> = found.iter().filter(|d| d.is_zst).collect();
+        assert_eq!(zst.len(), 1);
+        assert!(zst[0].path.ends_with("c.jsonl.zst"));
+    }
+
+    #[test]
+    fn extension_match_is_case_insensitive() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path();
+        touch(&root.join("upper.JSONL"));
+        let found = find_transcripts(root, &["jsonl"]);
+        assert_eq!(found.len(), 1);
+    }
+
+    #[test]
+    fn missing_root_yields_empty_not_panic() {
+        let dir = tempfile::tempdir().unwrap();
+        let missing = dir.path().join("does-not-exist");
+        assert!(find_transcripts(&missing, &["jsonl"]).is_empty());
+    }
+
+    #[test]
+    fn paths_helper_returns_only_paths() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path();
+        touch(&root.join("a.jsonl"));
+        let paths = find_transcript_paths(root, &["jsonl"]);
+        assert_eq!(paths.len(), 1);
+        assert!(paths[0].ends_with("a.jsonl"));
+    }
+
+    #[test]
+    fn empty_exts_matches_nothing() {
+        let dir = tempfile::tempdir().unwrap();
+        touch(&dir.path().join("a.jsonl"));
+        assert!(find_transcripts(dir.path(), &[]).is_empty());
+    }
+}
diff --git a/crates/memscribe-io/src/hook.rs b/crates/memscribe-io/src/hook.rs
new file mode 100644
index 0000000..40d8357
--- /dev/null
+++ b/crates/memscribe-io/src/hook.rs
@@ -0,0 +1,135 @@
+//! The hook handler (whitepaper §7).
+//!
+//! A `memscribe hook` subcommand registered as the tools' hook handler reads
+//! event JSON on stdin, records it, and — critically — captures the
+//! `transcript_path` and the live edit event so the Binder can write a PROV
+//! `wasGeneratedBy` record at the moment of the edit. It exits 0 immediately,
+//! never blocks the agent, and never invokes a model.
+
+use memscribe_core::{RawRecord, SourceLocation};
+use serde::Deserialize;
+
+/// The common shape of a tool hook payload delivered on stdin. Each tool's
+/// schema differs, so all fields are optional and unknown fields are retained.
+#[derive(Debug, Clone, Default, Deserialize)]
+pub struct HookPayload {
+    /// The tool-native session id, if present.
+    pub session_id: Option<String>,
+    /// The path to the live transcript the hook fired for.
+    pub transcript_path: Option<String>,
+    /// The hook event name (e.g. `PostToolUse`, `UserPromptSubmit`).
+    pub hook_event_name: Option<String>,
+    /// The working directory, if present.
+    pub cwd: Option<String>,
+    /// Any remaining fields, preserved verbatim.
+    #[serde(flatten)]
+    pub rest: serde_json::Map<String, serde_json::Value>,
+}
+
+/// Parse a hook payload from stdin bytes. Returns `None` if the bytes are not
+/// valid JSON (the handler still exits 0 — it never blocks the agent).
+#[must_use]
+pub fn parse_hook_payload(bytes: &[u8]) -> Option<HookPayload> {
+    serde_json::from_slice(bytes).ok()
+}
+
+/// Wrap raw hook bytes as a [`RawRecord`] with a synthetic hook provenance, so
+/// the same parsing path can consume it.
+#[must_use]
+pub fn hook_record(bytes: &[u8]) -> RawRecord {
+    RawRecord::new(bytes.to_vec(), SourceLocation::new("<hook stdin>", 0, 1))
+}
+
+/// The outcome of [`record_hook`]: the parsed payload, the transcript path it
+/// captured (if the payload carried one), and a synthetic [`RawRecord`] a caller
+/// can hand straight to the Binder to record a live edit.
+#[derive(Debug, Clone)]
+pub struct RecordedHook {
+    /// The parsed hook payload (common fields + preserved `rest`).
+    pub payload: HookPayload,
+    /// The live transcript path the hook fired for, if the payload carried one.
+    pub transcript_path: Option<String>,
+    /// A synthetic record wrapping the hook bytes, located at the transcript
+    /// path when known (so PROV provenance points at the real file) and at
+    /// `<hook stdin>` otherwise.
+    pub record: RawRecord,
+}
+
+/// Parse hook stdin bytes, capture the `transcript_path`, and build a synthetic
+/// [`RawRecord`] so a caller can feed the Binder a live edit at the moment the
+/// hook fired.
+///
+/// Returns `None` only when the bytes are not valid JSON — mirroring
+/// [`parse_hook_payload`], the handler still exits 0 and never blocks the agent.
+/// When a `transcript_path` is present the synthetic record's provenance points
+/// at that file (offset/line `0`/`1`, since the exact in-file location is not
+/// yet known at hook time); otherwise it falls back to `<hook stdin>`.
+#[must_use]
+pub fn record_hook(bytes: &[u8]) -> Option<RecordedHook> {
+    let payload = parse_hook_payload(bytes)?;
+    let transcript_path = payload.transcript_path.clone();
+    let location = match transcript_path.as_deref() {
+        Some(path) => SourceLocation::new(path, 0, 1),
+        None => SourceLocation::new("<hook stdin>", 0, 1),
+    };
+    let record = RawRecord::new(bytes.to_vec(), location);
+    Some(RecordedHook {
+        payload,
+        transcript_path,
+        record,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn parses_common_fields_and_keeps_rest() {
+        let payload = br#"{"session_id":"abc","transcript_path":"/t.jsonl","hook_event_name":"PostToolUse","tool_name":"Edit"}"#;
+        let p = parse_hook_payload(payload).unwrap();
+        assert_eq!(p.session_id.as_deref(), Some("abc"));
+        assert_eq!(p.hook_event_name.as_deref(), Some("PostToolUse"));
+        assert!(p.rest.contains_key("tool_name"));
+    }
+
+    #[test]
+    fn invalid_json_is_none_not_panic() {
+        assert!(parse_hook_payload(b"not json").is_none());
+    }
+
+    #[test]
+    fn record_hook_captures_transcript_path_and_locates_record_there() {
+        let payload = br#"{"session_id":"s1","transcript_path":"/home/u/.codex/sessions/x.jsonl","hook_event_name":"PostToolUse","tool_name":"Edit"}"#;
+        let rec = record_hook(payload).expect("valid json");
+        assert_eq!(
+            rec.transcript_path.as_deref(),
+            Some("/home/u/.codex/sessions/x.jsonl")
+        );
+        // The synthetic record carries the raw bytes verbatim...
+        assert_eq!(rec.record.bytes, payload.to_vec());
+        // ...and its provenance points at the captured transcript, not stdin.
+        assert_eq!(
+            rec.record.location.file,
+            std::path::PathBuf::from("/home/u/.codex/sessions/x.jsonl")
+        );
+        assert_eq!(rec.record.location.line_no, 1);
+        assert_eq!(rec.payload.session_id.as_deref(), Some("s1"));
+    }
+
+    #[test]
+    fn record_hook_without_transcript_falls_back_to_stdin() {
+        let payload = br#"{"hook_event_name":"UserPromptSubmit"}"#;
+        let rec = record_hook(payload).expect("valid json");
+        assert!(rec.transcript_path.is_none());
+        assert_eq!(
+            rec.record.location.file,
+            std::path::PathBuf::from("<hook stdin>")
+        );
+    }
+
+    #[test]
+    fn record_hook_invalid_json_is_none_not_panic() {
+        assert!(record_hook(b"<<<not json>>>").is_none());
+    }
+}
diff --git a/crates/memscribe-io/src/lib.rs b/crates/memscribe-io/src/lib.rs
new file mode 100644
index 0000000..caba577
--- /dev/null
+++ b/crates/memscribe-io/src/lib.rs
@@ -0,0 +1,31 @@
+//! # memscribe-io
+//!
+//! The source layer. A `Source` is conceptually just a stream of
+//! [`memscribe_core::RawRecord`]s; the rest of the pipeline does not know or
+//! care which source produced the bytes.
+//!
+//! - [`records`] — read a transcript file (or bytes) into `RawRecord`s with
+//!   exact byte/line provenance, transparently decompressing `.zst`. This is the
+//!   one-shot reader used by `memscribe parse` and the test harness.
+//! - [`tailer`] — the live, notify-based file tailer with a persisted byte-offset
+//!   cursor so restarts resume exactly where they left off (feature `watch`).
+//! - [`hook`] — the hook handler agents invoke (reads event JSON on stdin,
+//!   records it, returns immediately).
+//! - [`discover`] — walk a directory tree for transcript files by extension,
+//!   reporting which are `.zst` cold rollouts.
+//! - [`cursor_store`] — the persisted offset store (feature `cursor-store`).
+//! - [`otlp`] — a loopback-only HTTP OTLP receiver that ingests pushed
+//!   OpenTelemetry GenAI records into `RawRecord`s (feature `otlp`). Off by
+//!   default so the default build opens no network ports.
+#![forbid(unsafe_code)]
+
+pub mod cursor_store;
+pub mod discover;
+pub mod hook;
+#[cfg(feature = "otlp")]
+pub mod otlp;
+pub mod records;
+pub mod tailer;
+
+pub use records::{read_records, read_records_from_bytes};
+pub use tailer::poll_new_records;
diff --git a/crates/memscribe-io/src/otlp.rs b/crates/memscribe-io/src/otlp.rs
new file mode 100644
index 0000000..97768c4
--- /dev/null
+++ b/crates/memscribe-io/src/otlp.rs
@@ -0,0 +1,663 @@
+//! A loopback-only OTLP/HTTP receiver (whitepaper §7, §11).
+//!
+//! Behind the `otlp` cargo feature, this exposes a tiny HTTP/1.1 endpoint that
+//! ingests OpenTelemetry GenAI records pushed over the network and turns each
+//! one into a [`memscribe_core::RawRecord`] — exactly the shape the existing
+//! OTel adapter ([`memscribe_adapters::otel`]) already knows how to parse. The
+//! pipeline downstream of the source layer is unchanged: bytes in, normalized
+//! events out.
+//!
+//! ## Design (deliberately minimal)
+//!
+//! - **No async runtime, no tonic, no `hyper`.** A receiver that only needs to
+//!   accept short JSON POSTs on loopback does not justify pulling a full async
+//!   stack into Memscribe. We use [`std::net::TcpListener`] with a hand-rolled
+//!   HTTP/1.1 request reader that understands exactly the two things a pusher
+//!   sends: a request line, headers, and a `Content-Length`-delimited body.
+//! - **Loopback only (§11).** [`OtlpReceiver::bind`] rejects any address whose
+//!   IP is not a loopback address, so the default build never opens a port that
+//!   is reachable off-host. There is no flag to widen this.
+//! - **Lossless + version-tolerant, like every adapter.** The body may be
+//!   OTLP/HTTP-JSON (one JSON object, or a JSON array of records) or NDJSON
+//!   (one JSON record per line). Each record becomes one `RawRecord` carrying a
+//!   synthetic [`SourceLocation`] (`otlp://<peer>` with a per-record line
+//!   number) so the adapter and any audit replay see the same provenance shape
+//!   they get from a file.
+//! - **Never panics on malformed input.** A body that is not valid JSON / not
+//!   valid UTF-8 yields `400 Bad Request` and the server keeps serving. The
+//!   accept loop swallows per-connection I/O errors rather than tearing down.
+//!
+//! ## Usage
+//!
+//! ```no_run
+//! use memscribe_io::otlp::OtlpReceiver;
+//! use std::sync::mpsc;
+//!
+//! let recv = OtlpReceiver::bind("127.0.0.1:0").unwrap();
+//! let (tx, rx) = mpsc::channel();
+//! let handle = recv.serve_into(tx);
+//! // ... another part of the program drains `rx` for batches of RawRecords ...
+//! handle.shutdown();
+//! ```
+
+use memscribe_core::{RawRecord, SourceLocation};
+use std::io::{self, BufRead, BufReader, Read, Write};
+use std::net::{Shutdown, SocketAddr, TcpListener, TcpStream, ToSocketAddrs};
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::mpsc::Sender;
+use std::sync::Arc;
+use std::thread::{self, JoinHandle};
+
+/// Maximum request body we will buffer, in bytes. OTel GenAI records are small;
+/// this is a guard against a misbehaving or hostile local pusher, not a tuning
+/// knob. Bodies larger than this are rejected with `413 Payload Too Large`.
+const MAX_BODY_BYTES: usize = 8 * 1024 * 1024;
+
+/// Maximum number of header lines we will read before giving up on a request.
+const MAX_HEADER_LINES: usize = 256;
+
+/// A loopback-only HTTP receiver for OpenTelemetry GenAI records.
+///
+/// Construct with [`OtlpReceiver::bind`], then drive it with one of the serve
+/// methods. Each accepted POST body is decoded into zero or more
+/// [`RawRecord`]s that the OTel adapter can parse.
+#[derive(Debug)]
+pub struct OtlpReceiver {
+    listener: TcpListener,
+    local_addr: SocketAddr,
+}
+
+impl OtlpReceiver {
+    /// Bind the receiver to `addr`, which **must** resolve to a loopback
+    /// address (`127.0.0.0/8` or `::1`). Pass port `0` for an OS-assigned
+    /// ephemeral port and read it back with [`OtlpReceiver::local_addr`].
+    ///
+    /// # Errors
+    /// Returns an [`io::Error`] if the address does not resolve, resolves to a
+    /// non-loopback address ([`io::ErrorKind::InvalidInput`]), or the socket
+    /// cannot be bound.
+    pub fn bind(addr: impl ToSocketAddrs) -> io::Result<Self> {
+        let mut last_err = io::Error::new(
+            io::ErrorKind::InvalidInput,
+            "no socket address resolved from input",
+        );
+        for candidate in addr.to_socket_addrs()? {
+            if !candidate.ip().is_loopback() {
+                last_err = io::Error::new(
+                    io::ErrorKind::InvalidInput,
+                    format!(
+                        "OTLP receiver refuses to bind non-loopback address {candidate} \
+                         (whitepaper §11: loopback only)"
+                    ),
+                );
+                continue;
+            }
+            match TcpListener::bind(candidate) {
+                Ok(listener) => {
+                    let local_addr = listener.local_addr()?;
+                    return Ok(OtlpReceiver {
+                        listener,
+                        local_addr,
+                    });
+                }
+                Err(e) => last_err = e,
+            }
+        }
+        Err(last_err)
+    }
+
+    /// The address the receiver is actually listening on (resolves port `0`).
+    #[must_use]
+    pub fn local_addr(&self) -> SocketAddr {
+        self.local_addr
+    }
+
+    /// Run the blocking accept loop on the current thread, handing every batch
+    /// of decoded [`RawRecord`]s to `on_batch`. Returns when `should_stop`
+    /// observes `true` (set it from another thread, then poke the listener —
+    /// see [`OtlpReceiver::serve_into`] / [`ServeHandle`] for the wired-up
+    /// version).
+    ///
+    /// A connection that yields no records (empty or malformed body) still gets
+    /// a response; `on_batch` is simply not called for it.
+    pub fn serve_blocking<F>(&self, mut on_batch: F, should_stop: &AtomicBool)
+    where
+        F: FnMut(Vec<RawRecord>),
+    {
+        for stream in self.listener.incoming() {
+            if should_stop.load(Ordering::SeqCst) {
+                break;
+            }
+            match stream {
+                Ok(stream) => {
+                    // A single malformed/hostile connection must never take the
+                    // server down: handle it, log nothing fancy, keep serving.
+                    let records = handle_connection(stream, self.local_addr);
+                    if let Ok(records) = records {
+                        if !records.is_empty() {
+                            on_batch(records);
+                        }
+                    }
+                }
+                // Accept errors (e.g. the wake-up poke on shutdown) are not
+                // fatal; re-check the stop flag and continue.
+                Err(_) => {
+                    if should_stop.load(Ordering::SeqCst) {
+                        break;
+                    }
+                }
+            }
+        }
+    }
+
+    /// Spawn the accept loop on a background thread and stream every decoded
+    /// batch of [`RawRecord`]s into `sink`. Returns a [`ServeHandle`] for
+    /// graceful shutdown.
+    ///
+    /// The channel send is the back-pressure boundary: if the receiver of
+    /// `sink` is dropped, the loop stops on the next batch.
+    #[must_use]
+    pub fn serve_into(self, sink: Sender<Vec<RawRecord>>) -> ServeHandle {
+        let stop = Arc::new(AtomicBool::new(false));
+        let local_addr = self.local_addr;
+        let thread_stop = Arc::clone(&stop);
+        let join = thread::spawn(move || {
+            self.serve_blocking(
+                |batch| {
+                    // A dropped consumer means "stop": flip the flag so the
+                    // loop exits after this connection.
+                    if sink.send(batch).is_err() {
+                        thread_stop.store(true, Ordering::SeqCst);
+                    }
+                },
+                &thread_stop,
+            );
+        });
+        ServeHandle {
+            stop,
+            local_addr,
+            join: Some(join),
+        }
+    }
+}
+
+/// Handle to a running [`OtlpReceiver`] accept loop, for graceful shutdown.
+///
+/// Dropping the handle requests shutdown and joins the loop thread.
+#[derive(Debug)]
+pub struct ServeHandle {
+    stop: Arc<AtomicBool>,
+    local_addr: SocketAddr,
+    join: Option<JoinHandle<()>>,
+}
+
+impl ServeHandle {
+    /// The address the underlying receiver is listening on.
+    #[must_use]
+    pub fn local_addr(&self) -> SocketAddr {
+        self.local_addr
+    }
+
+    /// Request shutdown and block until the accept loop has stopped.
+    ///
+    /// Sets the stop flag, then opens (and immediately drops) one loopback
+    /// connection to wake the blocking `accept()` so the loop can observe the
+    /// flag and return promptly.
+    pub fn shutdown(mut self) {
+        self.signal_and_join();
+    }
+
+    fn signal_and_join(&mut self) {
+        self.stop.store(true, Ordering::SeqCst);
+        // Poke the listener so a thread parked in `accept()` wakes up.
+        if let Ok(stream) = TcpStream::connect(self.local_addr) {
+            let _ = stream.shutdown(Shutdown::Both);
+        }
+        if let Some(join) = self.join.take() {
+            let _ = join.join();
+        }
+    }
+}
+
+impl Drop for ServeHandle {
+    fn drop(&mut self) {
+        if self.join.is_some() {
+            self.signal_and_join();
+        }
+    }
+}
+
+/// Read and answer a single HTTP/1.1 request, returning the decoded records.
+///
+/// Always writes a response (so the client never hangs) and never propagates a
+/// decode failure as an error — a bad body is a `400`, not a server fault. The
+/// only `Err` returned is a genuine transport failure on this one connection,
+/// which the caller treats as "skip and keep serving".
+fn handle_connection(stream: TcpStream, local_addr: SocketAddr) -> io::Result<Vec<RawRecord>> {
+    let peer = stream
+        .peer_addr()
+        .map_or_else(|_| local_addr.to_string(), |p| p.to_string());
+    let mut reader = BufReader::new(stream);
+
+    match read_request(&mut reader) {
+        Ok(Request::Post { body }) => match decode_body(&body, &peer) {
+            Ok(records) => {
+                let n = records.len();
+                write_response(
+                    reader.get_mut(),
+                    200,
+                    "OK",
+                    &format!("{{\"accepted\":{n}}}"),
+                )?;
+                Ok(records)
+            }
+            Err(BodyError::TooLarge) => {
+                write_response(
+                    reader.get_mut(),
+                    413,
+                    "Payload Too Large",
+                    "{\"error\":\"body too large\"}",
+                )?;
+                Ok(Vec::new())
+            }
+            Err(BodyError::Malformed) => {
+                write_response(
+                    reader.get_mut(),
+                    400,
+                    "Bad Request",
+                    "{\"error\":\"malformed body\"}",
+                )?;
+                Ok(Vec::new())
+            }
+        },
+        Ok(Request::Other) => {
+            // Any non-POST verb (GET health check, etc.) is answered but
+            // produces no records.
+            write_response(
+                reader.get_mut(),
+                405,
+                "Method Not Allowed",
+                "{\"error\":\"POST only\"}",
+            )?;
+            Ok(Vec::new())
+        }
+        Err(RequestError::TooLarge) => {
+            write_response(
+                reader.get_mut(),
+                413,
+                "Payload Too Large",
+                "{\"error\":\"body too large\"}",
+            )?;
+            Ok(Vec::new())
+        }
+        Err(RequestError::Malformed) => {
+            write_response(
+                reader.get_mut(),
+                400,
+                "Bad Request",
+                "{\"error\":\"malformed request\"}",
+            )?;
+            Ok(Vec::new())
+        }
+        // A genuine transport error (peer hung up mid-headers, etc.): best
+        // effort, swallow it and move on without crashing.
+        Err(RequestError::Io) => Ok(Vec::new()),
+    }
+}
+
+/// The parsed request, reduced to what the receiver cares about.
+enum Request {
+    Post { body: Vec<u8> },
+    Other,
+}
+
+enum RequestError {
+    Malformed,
+    TooLarge,
+    /// A genuine transport failure on this connection. We only branch on the
+    /// *kind* — the underlying error is not surfaced, since the policy is the
+    /// same for every transport fault: drop this connection, keep serving.
+    Io,
+}
+
+impl From<io::Error> for RequestError {
+    fn from(_: io::Error) -> Self {
+        RequestError::Io
+    }
+}
+
+/// Parse the request line + headers, then read a `Content-Length`-delimited
+/// body. Deliberately strict and tiny: we only support exactly what a local
+/// OTLP pusher needs (no chunked transfer-encoding, no keep-alive).
+fn read_request<R: Read>(reader: &mut BufReader<R>) -> Result<Request, RequestError> {
+    let mut request_line = String::new();
+    let n = reader.read_line(&mut request_line)?;
+    if n == 0 {
+        // Connection opened and closed with no data (e.g. the shutdown poke).
+        return Err(RequestError::Malformed);
+    }
+    let method = request_line
+        .split_whitespace()
+        .next()
+        .ok_or(RequestError::Malformed)?;
+    let is_post = method.eq_ignore_ascii_case("POST");
+
+    let mut content_length: Option<usize> = None;
+    for _ in 0..MAX_HEADER_LINES {
+        let mut line = String::new();
+        let read = reader.read_line(&mut line)?;
+        if read == 0 {
+            break;
+        }
+        let trimmed = line.trim_end_matches(['\r', '\n']);
+        if trimmed.is_empty() {
+            // Blank line terminates the header block.
+            break;
+        }
+        if let Some((name, value)) = trimmed.split_once(':') {
+            if name.trim().eq_ignore_ascii_case("content-length") {
+                content_length = Some(
+                    value
+                        .trim()
+                        .parse::<usize>()
+                        .map_err(|_| RequestError::Malformed)?,
+                );
+            }
+        }
+    }
+
+    if !is_post {
+        return Ok(Request::Other);
+    }
+
+    let len = content_length.unwrap_or(0);
+    if len > MAX_BODY_BYTES {
+        return Err(RequestError::TooLarge);
+    }
+    let mut body = vec![0u8; len];
+    reader.read_exact(&mut body)?;
+    Ok(Request::Post { body })
+}
+
+/// Why a body could not be turned into records.
+enum BodyError {
+    Malformed,
+    TooLarge,
+}
+
+/// Turn a POSTed body into zero or more [`RawRecord`]s.
+///
+/// Accepts three shapes, in priority order:
+/// 1. A single JSON **array** of records → one `RawRecord` per element.
+/// 2. A single JSON **object** (OTLP/HTTP-JSON, possibly pretty-printed) → one
+///    `RawRecord` holding the object serialized to a single compact line.
+/// 3. **NDJSON** — one JSON value per line → one `RawRecord` per non-blank line.
+///
+/// Every emitted record's bytes are a single compact JSON line, which is
+/// exactly what the OTel adapter parses (it reads one JSON value per record).
+fn decode_body(body: &[u8], peer: &str) -> Result<Vec<RawRecord>, BodyError> {
+    if body.len() > MAX_BODY_BYTES {
+        return Err(BodyError::TooLarge);
+    }
+    let text = std::str::from_utf8(body).map_err(|_| BodyError::Malformed)?;
+    let trimmed = text.trim();
+    if trimmed.is_empty() {
+        return Ok(Vec::new());
+    }
+
+    // Shape 1 & 2: the whole body is one JSON value.
+    if let Ok(value) = serde_json::from_str::<serde_json::Value>(trimmed) {
+        return Ok(records_from_value(value, peer));
+    }
+
+    // Shape 3: NDJSON. Each non-blank line must itself be valid JSON, otherwise
+    // the body is malformed and we reject the whole thing (the pusher sent us
+    // something we cannot losslessly route).
+    let mut out = Vec::new();
+    let mut line_no = 1u64;
+    for line in trimmed.lines() {
+        let line = line.trim();
+        if line.is_empty() {
+            continue;
+        }
+        // Validate each line is JSON so a half-written stream becomes a 400
+        // rather than silently producing Unknown events downstream.
+        if serde_json::from_str::<serde_json::Value>(line).is_err() {
+            return Err(BodyError::Malformed);
+        }
+        out.push(mk_record(line.as_bytes().to_vec(), peer, line_no));
+        line_no += 1;
+    }
+    Ok(out)
+}
+
+/// Expand a single parsed JSON value into records: arrays fan out element-wise,
+/// everything else (object / scalar) becomes one record.
+fn records_from_value(value: serde_json::Value, peer: &str) -> Vec<RawRecord> {
+    match value {
+        serde_json::Value::Array(items) => items
+            .into_iter()
+            .enumerate()
+            .map(|(i, item)| {
+                let line = serde_json::to_string(&item).unwrap_or_else(|_| item.to_string());
+                mk_record(line.into_bytes(), peer, (i as u64) + 1)
+            })
+            .collect(),
+        other => {
+            let line = serde_json::to_string(&other).unwrap_or_else(|_| other.to_string());
+            vec![mk_record(line.into_bytes(), peer, 1)]
+        }
+    }
+}
+
+/// Build one `RawRecord` with a synthetic `otlp://<peer>` source location.
+fn mk_record(bytes: Vec<u8>, peer: &str, line_no: u64) -> RawRecord {
+    RawRecord::new(
+        bytes,
+        SourceLocation::new(format!("otlp://{peer}"), 0, line_no),
+    )
+}
+
+/// Write a minimal HTTP/1.1 response with a JSON body and close the connection.
+fn write_response<W: Write>(
+    writer: &mut W,
+    status: u16,
+    reason: &str,
+    body: &str,
+) -> io::Result<()> {
+    let response = format!(
+        "HTTP/1.1 {status} {reason}\r\n\
+         Content-Type: application/json\r\n\
+         Content-Length: {len}\r\n\
+         Connection: close\r\n\
+         \r\n\
+         {body}",
+        len = body.len(),
+    );
+    writer.write_all(response.as_bytes())?;
+    writer.flush()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::net::TcpStream;
+    use std::sync::mpsc;
+    use std::time::Duration;
+
+    /// POST `body` to `addr` and return the raw HTTP response text.
+    fn post(addr: SocketAddr, content_type: &str, body: &str) -> String {
+        let mut stream = TcpStream::connect(addr).expect("connect to receiver");
+        stream
+            .set_read_timeout(Some(Duration::from_secs(5)))
+            .expect("set read timeout");
+        let request = format!(
+            "POST /v1/logs HTTP/1.1\r\n\
+             Host: 127.0.0.1\r\n\
+             Content-Type: {content_type}\r\n\
+             Content-Length: {len}\r\n\
+             Connection: close\r\n\
+             \r\n\
+             {body}",
+            len = body.len(),
+        );
+        stream.write_all(request.as_bytes()).expect("write request");
+        stream.flush().expect("flush request");
+        let mut response = String::new();
+        stream.read_to_string(&mut response).expect("read response");
+        response
+    }
+
+    const GENAI_RECORD: &str = r#"{"time":"2026-06-22T10:00:05Z","gen_ai.operation.name":"chat","gen_ai.conversation.id":"sess-1","gen_ai.cli.user_prompt":"Use Postgres."}"#;
+
+    #[test]
+    fn bind_rejects_non_loopback() {
+        // 0.0.0.0 binds every interface — must be refused (whitepaper §11).
+        let err = OtlpReceiver::bind("0.0.0.0:0").expect_err("must refuse non-loopback");
+        assert_eq!(err.kind(), io::ErrorKind::InvalidInput);
+    }
+
+    #[test]
+    fn bind_accepts_loopback_ephemeral_port() {
+        let recv = OtlpReceiver::bind("127.0.0.1:0").expect("bind loopback");
+        assert!(recv.local_addr().ip().is_loopback());
+        assert_ne!(recv.local_addr().port(), 0, "ephemeral port resolved");
+    }
+
+    #[test]
+    fn post_single_json_object_yields_one_roundtripping_record() {
+        let recv = OtlpReceiver::bind("127.0.0.1:0").expect("bind");
+        let addr = recv.local_addr();
+        let (tx, rx) = mpsc::channel();
+        let handle = recv.serve_into(tx);
+
+        let response = post(addr, "application/json", GENAI_RECORD);
+        assert!(response.starts_with("HTTP/1.1 200"), "got: {response}");
+
+        let batch = rx.recv_timeout(Duration::from_secs(5)).expect("a batch");
+        assert_eq!(batch.len(), 1, "one record for one object");
+        let rec = &batch[0];
+
+        // The record's bytes must be valid JSON that round-trips to the same
+        // logical value the pusher sent — that is what the adapter parses.
+        let sent: serde_json::Value = serde_json::from_str(GENAI_RECORD).unwrap();
+        let got: serde_json::Value =
+            serde_json::from_slice(&rec.bytes).expect("record bytes are JSON");
+        assert_eq!(got, sent, "record round-trips the posted GenAI record");
+
+        // Provenance points at the loopback peer, not a file.
+        assert!(rec.location.file.to_string_lossy().starts_with("otlp://"));
+
+        handle.shutdown();
+    }
+
+    #[test]
+    fn post_json_array_fans_out_to_one_record_each() {
+        let recv = OtlpReceiver::bind("127.0.0.1:0").expect("bind");
+        let addr = recv.local_addr();
+        let (tx, rx) = mpsc::channel();
+        let handle = recv.serve_into(tx);
+
+        let body = format!("[{GENAI_RECORD},{GENAI_RECORD}]");
+        let response = post(addr, "application/json", &body);
+        assert!(response.starts_with("HTTP/1.1 200"), "got: {response}");
+
+        let batch = rx.recv_timeout(Duration::from_secs(5)).expect("a batch");
+        assert_eq!(batch.len(), 2, "two array elements → two records");
+        for rec in &batch {
+            serde_json::from_slice::<serde_json::Value>(&rec.bytes)
+                .expect("each record is valid JSON");
+        }
+
+        handle.shutdown();
+    }
+
+    #[test]
+    fn post_ndjson_yields_one_record_per_line() {
+        let recv = OtlpReceiver::bind("127.0.0.1:0").expect("bind");
+        let addr = recv.local_addr();
+        let (tx, rx) = mpsc::channel();
+        let handle = recv.serve_into(tx);
+
+        let body = format!("{GENAI_RECORD}\n{GENAI_RECORD}\n{GENAI_RECORD}\n");
+        let response = post(addr, "application/x-ndjson", &body);
+        assert!(response.starts_with("HTTP/1.1 200"), "got: {response}");
+
+        let batch = rx.recv_timeout(Duration::from_secs(5)).expect("a batch");
+        assert_eq!(batch.len(), 3, "three lines → three records");
+        // Line numbers in provenance are 1-based and monotonic.
+        assert_eq!(batch[0].location.line_no, 1);
+        assert_eq!(batch[2].location.line_no, 3);
+
+        handle.shutdown();
+    }
+
+    #[test]
+    fn malformed_body_returns_400_and_server_keeps_serving() {
+        let recv = OtlpReceiver::bind("127.0.0.1:0").expect("bind");
+        let addr = recv.local_addr();
+        let (tx, rx) = mpsc::channel();
+        let handle = recv.serve_into(tx);
+
+        // Garbage body → 400, no record, no panic.
+        let bad = post(addr, "application/json", "{not valid json at all");
+        assert!(bad.starts_with("HTTP/1.1 400"), "got: {bad}");
+        assert!(
+            rx.recv_timeout(Duration::from_millis(300)).is_err(),
+            "malformed body emits no batch"
+        );
+
+        // The server is still alive: a good POST afterwards still works.
+        let good = post(addr, "application/json", GENAI_RECORD);
+        assert!(good.starts_with("HTTP/1.1 200"), "got: {good}");
+        let batch = rx
+            .recv_timeout(Duration::from_secs(5))
+            .expect("server still serving after a bad request");
+        assert_eq!(batch.len(), 1);
+
+        handle.shutdown();
+    }
+
+    #[test]
+    fn empty_body_is_accepted_with_no_records() {
+        let recv = OtlpReceiver::bind("127.0.0.1:0").expect("bind");
+        let addr = recv.local_addr();
+        let (tx, rx) = mpsc::channel();
+        let handle = recv.serve_into(tx);
+
+        let response = post(addr, "application/json", "");
+        assert!(response.starts_with("HTTP/1.1 200"), "got: {response}");
+        assert!(
+            rx.recv_timeout(Duration::from_millis(300)).is_err(),
+            "empty body emits no batch"
+        );
+
+        handle.shutdown();
+    }
+
+    #[test]
+    fn non_post_method_is_rejected_without_records() {
+        let recv = OtlpReceiver::bind("127.0.0.1:0").expect("bind");
+        let addr = recv.local_addr();
+        let (tx, rx) = mpsc::channel();
+        let handle = recv.serve_into(tx);
+
+        let mut stream = TcpStream::connect(addr).expect("connect");
+        stream
+            .write_all(b"GET /health HTTP/1.1\r\nHost: 127.0.0.1\r\nConnection: close\r\n\r\n")
+            .expect("write");
+        stream.flush().expect("flush");
+        let mut response = String::new();
+        stream.read_to_string(&mut response).expect("read");
+        assert!(
+            response.starts_with("HTTP/1.1 405"),
+            "GET should be 405, got: {response}"
+        );
+        assert!(
+            rx.recv_timeout(Duration::from_millis(300)).is_err(),
+            "GET emits no batch"
+        );
+
+        handle.shutdown();
+    }
+}
diff --git a/crates/memscribe-io/src/records.rs b/crates/memscribe-io/src/records.rs
new file mode 100644
index 0000000..71618fd
--- /dev/null
+++ b/crates/memscribe-io/src/records.rs
@@ -0,0 +1,110 @@
+//! One-shot record reading with exact provenance.
+//!
+//! Splits a transcript into newline-delimited [`RawRecord`]s, each carrying its
+//! byte offset and 1-based line number, transparently decompressing `.zst`
+//! (Codex cold rollouts). This is the deterministic reader behind `memscribe
+//! parse` and the golden-file harness.
+
+use memscribe_core::{RawRecord, SourceLocation};
+use std::io;
+use std::path::Path;
+
+/// Split raw bytes into newline-delimited records with byte/line provenance.
+/// A trailing newline yields no extra record; blank lines are preserved as
+/// empty records (adapters skip them) so byte offsets stay exact.
+#[must_use]
+pub fn read_records_from_bytes(bytes: &[u8], path: &Path) -> Vec<RawRecord> {
+    let mut out = Vec::new();
+    let mut offset: u64 = 0;
+    let mut line_no: u64 = 1;
+    let mut start = 0usize;
+    for i in 0..bytes.len() {
+        if bytes[i] == b'\n' {
+            let line = &bytes[start..i];
+            out.push(RawRecord::new(
+                line.to_vec(),
+                SourceLocation::new(path, offset, line_no),
+            ));
+            offset = (i + 1) as u64;
+            line_no += 1;
+            start = i + 1;
+        }
+    }
+    // Trailing bytes with no final newline form one last record.
+    if start < bytes.len() {
+        out.push(RawRecord::new(
+            bytes[start..].to_vec(),
+            SourceLocation::new(path, offset, line_no),
+        ));
+    }
+    out
+}
+
+/// Read a transcript file into records, decompressing `.zst` transparently.
+///
+/// # Errors
+/// Returns an [`io::Error`] if the file cannot be read or decompressed.
+pub fn read_records(path: impl AsRef<Path>) -> io::Result<Vec<RawRecord>> {
+    let path = path.as_ref();
+    let bytes = std::fs::read(path)?;
+    let bytes = if path.extension().and_then(|e| e.to_str()) == Some("zst") {
+        zstd::decode_all(&bytes[..])?
+    } else {
+        bytes
+    };
+    Ok(read_records_from_bytes(&bytes, path))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::path::Path;
+
+    #[test]
+    fn splits_lines_with_offsets() {
+        let recs = read_records_from_bytes(b"a\nbb\nccc", Path::new("t.jsonl"));
+        assert_eq!(recs.len(), 3);
+        assert_eq!(recs[0].location.byte_offset, 0);
+        assert_eq!(recs[0].location.line_no, 1);
+        assert_eq!(recs[1].location.byte_offset, 2);
+        assert_eq!(recs[1].location.line_no, 2);
+        assert_eq!(recs[2].location.byte_offset, 5);
+        assert_eq!(recs[2].as_str(), Some("ccc"));
+    }
+
+    #[test]
+    fn trailing_newline_yields_no_empty_record() {
+        let recs = read_records_from_bytes(b"a\nb\n", Path::new("t.jsonl"));
+        assert_eq!(recs.len(), 2);
+    }
+
+    /// Offset-resumption property (whitepaper §8.3): splitting at any record
+    /// boundary and concatenating yields the same records as reading the whole.
+    #[test]
+    fn split_then_concat_equals_whole() {
+        let data = b"one\ntwo\nthree\nfour";
+        let whole: Vec<Vec<u8>> = read_records_from_bytes(data, Path::new("t"))
+            .into_iter()
+            .map(|r| r.bytes)
+            .collect();
+        for split in 0..=data.len() {
+            // Only split on record boundaries (start of file, end of file, or
+            // immediately after a newline) — that is where a real tailer resumes.
+            let on_boundary = split == 0 || split == data.len() || data[split - 1] == b'\n';
+            if !on_boundary {
+                continue;
+            }
+            let mut combined: Vec<Vec<u8>> =
+                read_records_from_bytes(&data[..split], Path::new("t"))
+                    .into_iter()
+                    .map(|r| r.bytes)
+                    .collect();
+            combined.extend(
+                read_records_from_bytes(&data[split..], Path::new("t"))
+                    .into_iter()
+                    .map(|r| r.bytes),
+            );
+            assert_eq!(combined, whole, "split at {split}");
+        }
+    }
+}
diff --git a/crates/memscribe-io/src/tailer.rs b/crates/memscribe-io/src/tailer.rs
new file mode 100644
index 0000000..927fbeb
--- /dev/null
+++ b/crates/memscribe-io/src/tailer.rs
@@ -0,0 +1,655 @@
+//! The file tailer.
+//!
+//! [`poll_new_records`] is the crash-safe core: it reads the bytes appended to a
+//! transcript since the last persisted offset, stops at the last complete line
+//! (so a half-written final line is never emitted — it resumes there next poll),
+//! advances the offset, and returns the new records with file-relative
+//! provenance. The notify-based live watcher (feature `watch`) is layered on top
+//! of this and is the fleet's task to wire.
+
+use crate::cursor_store::OffsetStore;
+use crate::records::read_records_from_bytes;
+use memscribe_core::RawRecord;
+use std::io::{self, Read, Seek, SeekFrom};
+use std::path::Path;
+
+/// Read records appended since the last persisted offset for `key`, advancing
+/// the offset to the last complete line. Returns the new records (empty if none
+/// or only a partial trailing line is available).
+///
+/// # Errors
+/// Returns an [`io::Error`] if the file cannot be opened or read.
+pub fn poll_new_records(
+    path: &Path,
+    store: &mut dyn OffsetStore,
+    key: &str,
+) -> io::Result<Vec<RawRecord>> {
+    let mut file = std::fs::File::open(path)?;
+    let len = file.metadata()?.len();
+    let start = store.get(key).unwrap_or(0).min(len);
+    if start >= len {
+        return Ok(Vec::new());
+    }
+    file.seek(SeekFrom::Start(start))?;
+    let mut buf = Vec::new();
+    file.read_to_end(&mut buf)?;
+
+    // Only consume up to the last newline; a partial final line is left for the
+    // next poll. This is what makes restart mid-line lossless and duplicate-free.
+    let consumable = buf.iter().rposition(|&b| b == b'\n').map_or(0, |i| i + 1);
+    if consumable == 0 {
+        return Ok(Vec::new());
+    }
+
+    let mut recs = read_records_from_bytes(&buf[..consumable], path);
+    for r in &mut recs {
+        // Provenance offsets become file-absolute.
+        r.location.byte_offset += start;
+    }
+    store.set(key, start + consumable as u64);
+    Ok(recs)
+}
+
+/// The live, notify-based tailer (feature `watch`).
+///
+/// [`LiveTailer`] watches a set of transcript files and, on each debounced
+/// create/modify event, delegates to the crash-safe [`poll_new_records`] core to
+/// emit only the records appended since the last persisted offset. It owns one
+/// [`OffsetStore`] for all watched files (keyed by path), so a restart resumes
+/// exactly where it left off.
+///
+/// Watching is done at the *directory* level (the parent of each watched file),
+/// which is what makes rotation and late-created files work: when a tool
+/// truncates-and-rewrites or only creates the transcript after the tailer has
+/// started, the directory watch still delivers the event and the file is picked
+/// up on the next poll. A file whose offset is ahead of a shrunken file (log
+/// rotation / truncation) is handled by [`poll_new_records`], which clamps the
+/// start offset to the current length.
+///
+/// The API is blocking and deterministic:
+/// - [`LiveTailer::poll`] waits up to a timeout for the next debounced batch and
+///   returns the emitted records (empty on timeout).
+/// - [`LiveTailer::run`] loops, invoking a callback for each batch, until the
+///   watcher is dropped or the callback returns [`ControlFlow::Break`].
+///
+/// It never panics: watcher/IO errors are surfaced as `tracing` warnings and the
+/// affected batch is skipped rather than unwinding.
+#[cfg(feature = "watch")]
+pub mod live {
+    use super::poll_new_records;
+    use crate::cursor_store::OffsetStore;
+    use memscribe_core::RawRecord;
+    use notify::{EventKind, RecursiveMode, Watcher};
+    use notify_debouncer_full::{new_debouncer, DebounceEventResult, Debouncer, FileIdMap};
+    use std::collections::BTreeSet;
+    use std::ops::ControlFlow;
+    use std::path::{Path, PathBuf};
+    use std::sync::mpsc::{Receiver, RecvTimeoutError};
+    use std::time::Duration;
+
+    /// What a [`LiveTailer`] should do after a batch is handled by `run`.
+    pub type Batch = Vec<RawRecord>;
+
+    /// A canonical key for a watched path. We use the lossy string form of the
+    /// path the caller registered, so the offset store key is stable across
+    /// polls (we deliberately do not canonicalize on disk — a rotated/recreated
+    /// file keeps the same logical key).
+    fn key_for(path: &Path) -> String {
+        path.to_string_lossy().into_owned()
+    }
+
+    /// The live notify-based tailer. Generic over the [`OffsetStore`] so it can
+    /// drive either the in-memory or the SQLite-backed cursor.
+    pub struct LiveTailer<S: OffsetStore> {
+        store: S,
+        // Files we emit records for, by their registered (logical) path.
+        interest: BTreeSet<PathBuf>,
+        // Parent directories actually handed to the OS watcher.
+        watched_dirs: BTreeSet<PathBuf>,
+        debouncer: Debouncer<notify::RecommendedWatcher, FileIdMap>,
+        rx: Receiver<DebounceEventResult>,
+    }
+
+    impl<S: OffsetStore> LiveTailer<S> {
+        /// Create a tailer with the given offset store and debounce timeout.
+        ///
+        /// `timeout` is the notify-debouncer-full coalescing window; a value
+        /// around 100–300ms keeps latency low while collapsing the burst of
+        /// events a single append produces.
+        ///
+        /// # Errors
+        /// Returns a [`notify::Error`] if the OS watcher cannot be created.
+        pub fn new(store: S, timeout: Duration) -> notify::Result<Self> {
+            let (tx, rx) = std::sync::mpsc::channel();
+            // `std::sync::mpsc::Sender` implements `DebounceEventHandler`.
+            let debouncer = new_debouncer(timeout, None, tx)?;
+            Ok(Self {
+                store,
+                interest: BTreeSet::new(),
+                watched_dirs: BTreeSet::new(),
+                debouncer,
+                rx,
+            })
+        }
+
+        /// Register a single transcript file to tail. The file need not exist yet
+        /// — the nearest existing ancestor directory is watched recursively, so a
+        /// later-created file (even one whose parent directory does not exist yet)
+        /// is still picked up. Re-registering a path is idempotent.
+        ///
+        /// # Errors
+        /// Returns a [`notify::Error`] only if a *new* directory cannot be watched;
+        /// an already-watched directory is a no-op. A path with no existing
+        /// ancestor at all is recorded in the interest set and retried lazily on
+        /// the next [`watch_path`] / [`poll_existing`] — it never errors.
+        ///
+        /// [`poll_existing`]: LiveTailer::poll_existing
+        pub fn watch_path(&mut self, path: impl AsRef<Path>) -> notify::Result<()> {
+            let path = path.as_ref().to_path_buf();
+            self.interest.insert(path.clone());
+            self.ensure_ancestor_watch(&path)
+        }
+
+        /// Watch the nearest *existing* ancestor directory of `path`. If the
+        /// immediate parent exists we watch it non-recursively (cheap, exact). If
+        /// it does not yet exist (e.g. a tool that creates `~/.tool/sessions/`
+        /// lazily), we climb to the nearest existing ancestor and watch *that*
+        /// recursively, so the eventual creation of the file is observed. Already
+        /// watched directories are a no-op.
+        fn ensure_ancestor_watch(&mut self, path: &Path) -> notify::Result<()> {
+            // The directory the file lives in (own parent, or "." for a bare name).
+            let parent = path
+                .parent()
+                .filter(|p| !p.as_os_str().is_empty())
+                .map_or_else(|| PathBuf::from("."), Path::to_path_buf);
+
+            // Walk up to the first ancestor that exists on disk right now.
+            let mut candidate = parent.as_path();
+            let mut recursive = RecursiveMode::NonRecursive;
+            loop {
+                if candidate.is_dir() {
+                    break;
+                }
+                match candidate.parent() {
+                    Some(up) if !up.as_os_str().is_empty() => {
+                        candidate = up;
+                        // We dropped below the exact parent → must watch
+                        // recursively to catch the not-yet-created subtree.
+                        recursive = RecursiveMode::Recursive;
+                    }
+                    _ => {
+                        // No existing ancestor at all (very unusual). Keep it in
+                        // the interest set; it is retried on the next register.
+                        return Ok(());
+                    }
+                }
+            }
+
+            let dir = candidate.to_path_buf();
+            if self.watched_dirs.insert(dir.clone()) {
+                self.debouncer.watcher().watch(&dir, recursive)?;
+            }
+            Ok(())
+        }
+
+        /// Register many transcript files at once (see [`watch_path`]).
+        ///
+        /// [`watch_path`]: LiveTailer::watch_path
+        ///
+        /// # Errors
+        /// Returns the first [`notify::Error`] encountered while registering.
+        pub fn watch_paths<P: AsRef<Path>>(
+            &mut self,
+            paths: impl IntoIterator<Item = P>,
+        ) -> notify::Result<()> {
+            for p in paths {
+                self.watch_path(p)?;
+            }
+            Ok(())
+        }
+
+        /// Immediately poll every watched file once for records appended since the
+        /// last offset, without waiting for an event. Useful to drain content that
+        /// already existed (or was written before the watcher started) on startup.
+        ///
+        /// Also lazily (re)establishes the directory watch for any interested path
+        /// whose ancestor directory has since come into existence — closing the
+        /// window where a file registered before its directory existed would
+        /// otherwise never be watched.
+        ///
+        /// Records are returned in deterministic (sorted-by-path) order.
+        pub fn poll_existing(&mut self) -> Batch {
+            let mut out = Vec::new();
+            // `interest` is a BTreeSet → iteration is sorted → deterministic.
+            let paths: Vec<PathBuf> = self.interest.iter().cloned().collect();
+            for path in &paths {
+                // Best-effort: retry the ancestor watch in case the directory was
+                // created after registration. A failure here is non-fatal.
+                let _ = self.ensure_ancestor_watch(path);
+            }
+            for path in paths {
+                self.drain_path(&path, &mut out);
+            }
+            out
+        }
+
+        /// Block up to `timeout` for the next debounced batch and return the
+        /// records it produced. Returns an empty batch on timeout (so the caller
+        /// can interleave other work) and an empty batch — never an error — if the
+        /// watcher reported a recoverable error (it is logged via `tracing`).
+        ///
+        /// Returns `None` only when the watcher has shut down and no further events
+        /// will ever arrive (the sender was dropped), signalling end-of-stream.
+        pub fn poll(&mut self, timeout: Duration) -> Option<Batch> {
+            match self.rx.recv_timeout(timeout) {
+                Ok(result) => Some(self.handle_result(result)),
+                Err(RecvTimeoutError::Timeout) => Some(Vec::new()),
+                Err(RecvTimeoutError::Disconnected) => None,
+            }
+        }
+
+        /// Run the blocking tail loop, invoking `on_batch` for each non-empty
+        /// debounced batch. The loop ends when the watcher disconnects or
+        /// `on_batch` returns [`ControlFlow::Break`]. Empty batches (timeouts) are
+        /// not delivered to the callback.
+        ///
+        /// `tick` bounds how long a single `recv` blocks, so the loop stays
+        /// responsive to a `Break` even when the tree is idle.
+        pub fn run<F>(&mut self, tick: Duration, mut on_batch: F)
+        where
+            F: FnMut(Batch) -> ControlFlow<()>,
+        {
+            loop {
+                match self.poll(tick) {
+                    Some(batch) => {
+                        if batch.is_empty() {
+                            continue;
+                        }
+                        if on_batch(batch).is_break() {
+                            return;
+                        }
+                    }
+                    None => return,
+                }
+            }
+        }
+
+        /// Test seam: feed a synthetic debounced *modify* event for `paths`
+        /// through the exact same translation path the real watcher uses, so the
+        /// event→records logic can be tested deterministically without depending
+        /// on the platform's (notoriously timing-dependent) filesystem
+        /// notifications. Not part of the public API.
+        #[cfg(test)]
+        pub(crate) fn handle_synthetic_modify<P: AsRef<Path>>(
+            &mut self,
+            paths: impl IntoIterator<Item = P>,
+        ) -> Batch {
+            use notify::event::{Event, EventKind as Ek, ModifyKind};
+            use notify_debouncer_full::DebouncedEvent;
+            let mut event = Event::new(Ek::Modify(ModifyKind::Any));
+            for p in paths {
+                event = event.add_path(p.as_ref().to_path_buf());
+            }
+            let debounced = DebouncedEvent::new(event, std::time::Instant::now());
+            self.handle_result(Ok(vec![debounced]))
+        }
+
+        /// Turn one debounced result into the records to emit. Errors are logged
+        /// and yield an empty batch — the tailer is panic-free by contract.
+        fn handle_result(&mut self, result: DebounceEventResult) -> Batch {
+            let mut out = Vec::new();
+            let events = match result {
+                Ok(events) => events,
+                Err(errors) => {
+                    for e in errors {
+                        tracing::warn!(error = %e, "live tailer watch error; skipping batch");
+                    }
+                    return out;
+                }
+            };
+
+            // Collect the distinct interested paths touched by create/modify
+            // events, then drain each once in deterministic (sorted) order.
+            let mut touched: BTreeSet<PathBuf> = BTreeSet::new();
+            for ev in events {
+                if !matches!(
+                    ev.kind,
+                    EventKind::Create(_) | EventKind::Modify(_) | EventKind::Any
+                ) {
+                    continue;
+                }
+                for p in &ev.paths {
+                    if self.interest.contains(p) {
+                        touched.insert(p.clone());
+                    }
+                }
+            }
+            for path in touched {
+                self.drain_path(&path, &mut out);
+            }
+            out
+        }
+
+        /// Poll one path, appending its new records to `out`. A missing file
+        /// (late create / rotation gap) is treated as "no records yet", not an
+        /// error.
+        ///
+        /// Handles **truncation/rotation**: if the file is now shorter than the
+        /// persisted offset, the transcript was rewritten in place (the tool
+        /// truncated and started over). The stored offset is stale, so we reset it
+        /// to `0` before delegating to the crash-safe core — which then reads the
+        /// fresh content from the start. This never re-emits old content (the old
+        /// bytes are gone) and never loses the new content.
+        fn drain_path(&mut self, path: &Path, out: &mut Batch) {
+            let key = key_for(path);
+            if let Ok(meta) = std::fs::metadata(path) {
+                let len = meta.len();
+                if self.store.get(&key).is_some_and(|off| off > len) {
+                    tracing::debug!(
+                        path = %path.display(),
+                        "transcript shrank below cursor; treating as rotation, resetting offset"
+                    );
+                    self.store.set(&key, 0);
+                }
+            }
+            match poll_new_records(path, &mut self.store, &key) {
+                Ok(mut recs) => out.append(&mut recs),
+                Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
+                    // File not present yet (or mid-rotation). Nothing to emit.
+                }
+                Err(e) => {
+                    tracing::warn!(path = %path.display(), error = %e, "live tailer poll error");
+                }
+            }
+        }
+
+        /// Borrow the underlying offset store (e.g. to flush/inspect).
+        pub fn store(&self) -> &S {
+            &self.store
+        }
+
+        /// Mutably borrow the underlying offset store.
+        pub fn store_mut(&mut self) -> &mut S {
+            &mut self.store
+        }
+    }
+}
+
+#[cfg(feature = "watch")]
+pub use live::LiveTailer;
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::cursor_store::MemoryOffsetStore;
+    use std::io::Write;
+
+    #[test]
+    fn resumes_from_offset_with_no_loss_or_dup() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("s.jsonl");
+        let mut store = MemoryOffsetStore::new();
+
+        std::fs::write(&path, b"a\nb\n").unwrap();
+        let first = poll_new_records(&path, &mut store, "s").unwrap();
+        assert_eq!(first.len(), 2);
+
+        // Append more; only the new records come back.
+        let mut f = std::fs::OpenOptions::new()
+            .append(true)
+            .open(&path)
+            .unwrap();
+        f.write_all(b"c\n").unwrap();
+        let second = poll_new_records(&path, &mut store, "s").unwrap();
+        assert_eq!(second.len(), 1);
+        assert_eq!(second[0].as_str(), Some("c"));
+    }
+
+    #[test]
+    fn partial_final_line_is_held_back() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("s.jsonl");
+        let mut store = MemoryOffsetStore::new();
+
+        std::fs::write(&path, b"a\nb\npartial").unwrap();
+        let recs = poll_new_records(&path, &mut store, "s").unwrap();
+        assert_eq!(recs.len(), 2, "partial trailing line must be held back");
+
+        // Once the line is completed, it is delivered exactly once.
+        let mut f = std::fs::OpenOptions::new()
+            .append(true)
+            .open(&path)
+            .unwrap();
+        f.write_all(b"_done\n").unwrap();
+        let recs = poll_new_records(&path, &mut store, "s").unwrap();
+        assert_eq!(recs.len(), 1);
+        assert_eq!(recs[0].as_str(), Some("partial_done"));
+    }
+
+    /// Crash/replay invariant (whitepaper §8.5): a process that dies after
+    /// consuming some complete lines but while a final line is still being
+    /// written must, on resume with the *persisted* offset, emit every record
+    /// exactly once — zero loss, zero duplication — regardless of where the
+    /// crash truncated the stream.
+    ///
+    /// We simulate this by persisting the offset to a fresh store on each
+    /// "process restart" (the offset is the only state that survives a crash),
+    /// growing the file one fragment at a time, and crashing mid-line.
+    #[test]
+    fn crash_mid_line_resumes_with_zero_loss_zero_dup() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("s.jsonl");
+
+        // The full intended transcript: three complete records.
+        let full = b"first\nsecond\nthird\n";
+
+        // The "crash schedule": a sequence of byte prefixes of `full` that the
+        // file is observed at across restarts. Several land *inside* a line
+        // (truncated final line) — the dangerous case.
+        let crash_prefixes = [
+            6usize, // "first\n"            (boundary)
+            9,      // "first\nsec"         (mid second line — TRUNCATED)
+            13,     // "first\nsecond"      (mid second line — TRUNCATED, no \n)
+            14,     // "first\nsecond\n"    (boundary)
+            17,     // "first\nsecond\nthi" (mid third line — TRUNCATED)
+            full.len(),
+        ];
+
+        // The single piece of state that survives a crash.
+        let mut persisted_offset: u64 = 0;
+        let mut emitted: Vec<String> = Vec::new();
+
+        for &prefix in &crash_prefixes {
+            // The file as it exists at this crash point.
+            std::fs::write(&path, &full[..prefix]).unwrap();
+
+            // "Restart": a brand-new store seeded ONLY with the persisted offset.
+            let mut store = MemoryOffsetStore::new();
+            if persisted_offset > 0 {
+                store.set("s", persisted_offset);
+            }
+
+            let recs = poll_new_records(&path, &mut store, "s").unwrap();
+            for r in &recs {
+                emitted.push(r.as_str().unwrap().to_string());
+            }
+
+            // Persist whatever offset the tailer advanced to (it only advances
+            // past complete lines), exactly as a real crash-safe cursor would.
+            persisted_offset = store.get("s").unwrap_or(persisted_offset);
+        }
+
+        // Zero loss + zero dup: every record once, in order.
+        assert_eq!(
+            emitted,
+            vec![
+                "first".to_string(),
+                "second".to_string(),
+                "third".to_string()
+            ],
+            "each record must be emitted exactly once across crashes"
+        );
+        // The offset finished exactly at EOF.
+        assert_eq!(persisted_offset, full.len() as u64);
+    }
+
+    /// A re-poll at the *same* persisted offset (e.g. the watcher fires twice for
+    /// one append, or the process restarts without any new bytes) must be a
+    /// no-op — this is the duplicate-suppression half of the invariant.
+    #[test]
+    fn repoll_without_new_bytes_emits_nothing() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("s.jsonl");
+        let mut store = MemoryOffsetStore::new();
+
+        std::fs::write(&path, b"x\ny\n").unwrap();
+        assert_eq!(poll_new_records(&path, &mut store, "s").unwrap().len(), 2);
+        // No new bytes; repeated polls (incl. a spurious watcher wakeup).
+        assert!(poll_new_records(&path, &mut store, "s").unwrap().is_empty());
+        assert!(poll_new_records(&path, &mut store, "s").unwrap().is_empty());
+    }
+}
+
+#[cfg(all(test, feature = "watch"))]
+mod live_tests {
+    //! These tests exercise the `LiveTailer` translation logic deterministically
+    //! via the `handle_synthetic_modify` seam. They deliberately do **not** rely
+    //! on the operating system actually delivering filesystem notifications:
+    //! FSEvents/inotify delivery for short-lived temp files is timing-dependent
+    //! and would make the suite flaky (and, on a watcher that never fires, hang).
+    //! The real OS path (`watch_path` → debouncer → `poll`) is wired and
+    //! compiled; here we validate that a debounced modify event produces exactly
+    //! the newly-appended records, with the same crash-safe offset semantics the
+    //! core guarantees.
+
+    use super::live::LiveTailer;
+    use crate::cursor_store::MemoryOffsetStore;
+    use std::io::Write;
+    use std::time::Duration;
+
+    fn append(path: &std::path::Path, bytes: &[u8]) {
+        let mut f = std::fs::OpenOptions::new()
+            .create(true)
+            .append(true)
+            .open(path)
+            .unwrap();
+        f.write_all(bytes).unwrap();
+        f.flush().unwrap();
+    }
+
+    fn texts(batch: &[memscribe_core::RawRecord]) -> Vec<String> {
+        batch
+            .iter()
+            .map(|r| r.as_str().unwrap().to_string())
+            .collect()
+    }
+
+    #[test]
+    fn new_and_watch_path_succeed_without_error() {
+        // Constructing the real OS watcher and registering paths must work even
+        // for not-yet-existing files (the parent dir is watched).
+        let dir = tempfile::tempdir().unwrap();
+        let mut tailer =
+            LiveTailer::new(MemoryOffsetStore::new(), Duration::from_millis(50)).unwrap();
+        tailer
+            .watch_paths([dir.path().join("a.jsonl"), dir.path().join("sub/b.jsonl")])
+            .unwrap();
+        // Re-registering a path / a sibling in an already-watched dir is a no-op.
+        tailer.watch_path(dir.path().join("a.jsonl")).unwrap();
+    }
+
+    #[test]
+    fn poll_existing_drains_pre_written_content_once() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("pre.jsonl");
+        std::fs::write(&path, b"a\nb\n").unwrap();
+
+        let mut tailer =
+            LiveTailer::new(MemoryOffsetStore::new(), Duration::from_millis(50)).unwrap();
+        tailer.watch_path(&path).unwrap();
+
+        assert_eq!(texts(&tailer.poll_existing()), vec!["a", "b"]);
+        // A second drain with no new bytes yields nothing (offset persisted).
+        assert!(tailer.poll_existing().is_empty());
+    }
+
+    #[test]
+    fn modify_event_emits_only_new_records() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("live.jsonl");
+        std::fs::write(&path, b"seed\n").unwrap();
+
+        let mut tailer =
+            LiveTailer::new(MemoryOffsetStore::new(), Duration::from_millis(50)).unwrap();
+        tailer.watch_path(&path).unwrap();
+        // Consume the seed so we only observe the live append.
+        let _ = tailer.poll_existing();
+
+        append(&path, b"live-one\nlive-two\n");
+        let got = tailer.handle_synthetic_modify([&path]);
+        assert_eq!(texts(&got), vec!["live-one", "live-two"]);
+        // A duplicate event for the same (unchanged) file emits nothing — the
+        // offset already advanced. This is the dedup guarantee under spurious
+        // debouncer wakeups.
+        assert!(tailer.handle_synthetic_modify([&path]).is_empty());
+    }
+
+    #[test]
+    fn late_created_file_is_picked_up() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("late.jsonl"); // does not exist yet
+
+        let mut tailer =
+            LiveTailer::new(MemoryOffsetStore::new(), Duration::from_millis(50)).unwrap();
+        tailer.watch_path(&path).unwrap();
+        // A modify event for a not-yet-existing file is harmless (NotFound is
+        // swallowed) and emits nothing.
+        assert!(tailer.handle_synthetic_modify([&path]).is_empty());
+
+        // Now it appears; the create/modify event tails it from byte 0.
+        append(&path, b"born-late\n");
+        let got = tailer.handle_synthetic_modify([&path]);
+        assert_eq!(texts(&got), vec!["born-late"]);
+    }
+
+    #[test]
+    fn rotation_truncation_does_not_lose_or_duplicate() {
+        // A tool that truncates-and-rewrites its transcript (log rotation) must
+        // not crash the tailer or replay stale content. poll_new_records clamps
+        // the start offset to the (now smaller) length.
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("rot.jsonl");
+        std::fs::write(&path, b"old-a\nold-b\n").unwrap();
+
+        let mut tailer =
+            LiveTailer::new(MemoryOffsetStore::new(), Duration::from_millis(50)).unwrap();
+        tailer.watch_path(&path).unwrap();
+        assert_eq!(texts(&tailer.poll_existing()), vec!["old-a", "old-b"]);
+
+        // Rotation: the file shrinks below the persisted offset, then new lines.
+        std::fs::write(&path, b"new-a\n").unwrap();
+        let got = tailer.handle_synthetic_modify([&path]);
+        assert_eq!(
+            texts(&got),
+            vec!["new-a"],
+            "post-rotation content tails cleanly"
+        );
+    }
+
+    #[test]
+    fn only_interested_paths_in_event_are_drained() {
+        let dir = tempfile::tempdir().unwrap();
+        let watched = dir.path().join("watched.jsonl");
+        let sibling = dir.path().join("sibling.jsonl");
+        std::fs::write(&watched, b"mine\n").unwrap();
+        std::fs::write(&sibling, b"not-mine\n").unwrap();
+
+        let mut tailer =
+            LiveTailer::new(MemoryOffsetStore::new(), Duration::from_millis(50)).unwrap();
+        tailer.watch_path(&watched).unwrap();
+
+        // An event mentioning both paths only drains the registered one.
+        let got = tailer.handle_synthetic_modify([watched.clone(), sibling.clone()]);
+        assert_eq!(texts(&got), vec!["mine"]);
+    }
+}
diff --git a/crates/memscribe-sink/Cargo.toml b/crates/memscribe-sink/Cargo.toml
new file mode 100644
index 0000000..6e6413b
--- /dev/null
+++ b/crates/memscribe-sink/Cargo.toml
@@ -0,0 +1,30 @@
+[package]
+name = "memscribe-sink"
+version.workspace = true
+edition.workspace = true
+rust-version.workspace = true
+license.workspace = true
+repository.workspace = true
+description = "Sink implementations for Memscribe: NDJSON (default), SQLite, and a feature-gated MemDB sink."
+
+[lints]
+workspace = true
+
+[features]
+default = ["sqlite"]
+sqlite = ["dep:rusqlite"]
+# Enable the MemDB sink. Memtrace turns this on; it is off by default so the
+# crate is fully usable without MemDB present.
+memdb = []
+
+[dependencies]
+memscribe-core = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+time = { workspace = true }
+rusqlite = { workspace = true, optional = true }
+
+[dev-dependencies]
+memscribe-core = { workspace = true }
+tempfile = { workspace = true }
+serde_json = { workspace = true }
diff --git a/crates/memscribe-sink/MEMDB.md b/crates/memscribe-sink/MEMDB.md
new file mode 100644
index 0000000..6ebe052
--- /dev/null
+++ b/crates/memscribe-sink/MEMDB.md
@@ -0,0 +1,265 @@
+# Writing Memscribe into MemDB
+
+How `memscribe-sink` ingests prepared nodes into **MemDB**, and — just as
+important — how Memscribe stays **fully usable without MemDB**.
+
+This document is grounded in the real MemDB SDK as it exists in the indexed
+`memdb` repo. Every type and method named below is cited with its source
+location so the wiring can be implemented against the actual API rather than a
+guess. All paths are relative to
+`memdb/memcore-rs/crates/`.
+
+---
+
+## 0. The headline: MemDB is optional
+
+Memscribe's pipeline writes through one trait — `memscribe_core::Sink`
+(`memscribe-core/src/sink.rs`) — with exactly two methods, `emit(&PreparedNode)`
+and `flush()`. Nothing upstream of the sink knows or cares what the sink does.
+
+The crate ships **three** sinks, and the default has no external dependency at
+all:
+
+| Sink | Cargo feature | Default? | External service |
+|------|---------------|----------|------------------|
+| `NdjsonSink` | *(always built)* | **yes** (canonical) | none — one JSON line per node |
+| `SqliteSink` | `sqlite` | on | none — local file/`:memory:` |
+| `MemDbSink`  | `memdb`  | **off** | MemDB (gRPC) |
+
+In `Cargo.toml`:
+
+```toml
+[features]
+default = ["sqlite"]
+sqlite  = ["dep:rusqlite"]
+memdb   = []   # OFF by default
+```
+
+Consequences:
+
+- `cargo build` / `cargo test -p memscribe-sink` never compiles a line of
+  MemDB code. `memdb.rs` is behind `#[cfg(feature = "memdb")]` in
+  `src/lib.rs`, so the MemDB types do not even exist in the default build.
+- The whole module is observable and golden-testable with `NdjsonSink`: one
+  prepared node ⇒ one line of canonical JSON. CI, fixtures, and `replay` use
+  it with zero infrastructure.
+- `SqliteSink` gives a queryable local store (indexed by variant tag and
+  `FactStatus`, deduplicated by a deterministic primary key) for anyone who
+  wants SQL without standing up MemDB.
+- **Memtrace** — and only Memtrace — flips on `--features memdb` to route the
+  same `PreparedNode` stream into MemDB. No other consumer pays for the
+  dependency.
+
+So the answer to "what if I don't have MemDB?" is: nothing changes. NDJSON is
+the system of record's transport; MemDB is an opt-in destination.
+
+---
+
+## 1. The MemDB API we write into
+
+The SDK is `memcore_client` (`memcore-client/src/lib.rs`). The relevant public
+surface:
+
+### Connecting
+
+```rust
+// memcore-client/src/lib.rs:97
+MemcoreClient::connect(endpoint: impl Into<String>) -> Result<Self, ClientError>
+// accepts "http://host:port" / "https://host:port"; 5s connect timeout;
+// 64 MiB max message size (MAX_GRPC_MESSAGE_BYTES, lib.rs:92).
+```
+
+### The bi-temporal record header
+
+Every record on disk in MemDB is fronted by a fixed 32-byte header
+(`memcore-core/src/lib.rs:237`, layout documented there):
+
+```rust
+// memcore-core/src/lib.rs
+pub struct RecordHeader {
+    pub rid:        Rid,        // u64  — server-assigned (pass Rid(0) on create)
+    pub valid_at:   Micros,     // i64 µs since Unix epoch  → VALID time
+    pub invalid_at: Micros,     // STILL_VALID = "not superseded"
+    pub episode_id: EpisodeId,  // u32  — the arc/episode
+    pub schema_ver: SchemaVer,  // u16
+    pub kind:       RecordKind, // Node=1, Edge=2, EdgeSegment=3, Episode=4, VectorBlob=5
+    pub flags:      u8,
+}
+```
+
+Supporting newtypes (all `memcore-core/src/lib.rs`):
+
+- `Micros(pub i64)` — microseconds since the Unix epoch (`:158`).
+- `STILL_VALID: Micros = Micros::MAX` — the "never superseded" sentinel
+  (`:172`); `invalid_at = STILL_VALID` means the row is currently live.
+- `EpisodeId(pub u32)` (`:289`).
+- `SchemaVer(pub u16)` (`:149`).
+- `RecordKind` — closed enum, no extension point (`:201`).
+- `AsOf::now()` = `AsOf(STILL_VALID)`, `AsOf::at(Micros)` (`:179`) — the read
+  side's as-of clock. `RecordHeader::visible_at(t)` is `valid_at <= t <
+  invalid_at` with the `STILL_VALID` branch (`:260`).
+
+### Creating records (with typed properties)
+
+```rust
+// memcore-client/src/lib.rs:492
+MemcoreClient::create_record_with_properties(
+    header:     RecordHeader,
+    body:       Vec<u8>,
+    properties: Vec<Property>,
+) -> Result<Rid, ClientError>
+```
+
+This is the ergonomic create: it builds the prost `CreateRequest`
+(`header`, `body`, `properties`, leaving `fencing_token`/`durability`/
+`session_id` at server defaults) and returns the server-assigned `Rid`. The
+typed properties feed MemDB's property index, which is what backs Memtrace's
+`find_symbol` / `find_code`. Properties are built with `PropertyBuilder`
+(`memcore-client` re-exports it at `lib.rs:43`); from the SDK's own doc example
+(`lib.rs:465-487`):
+
+```rust
+let properties = PropertyBuilder::new()
+    .string("name", "validateToken")
+    .string("file_path", "src/auth.ts")
+    .int("start_line", 42)
+    .build();
+let rid = client
+    .create_record_with_properties(header, body, properties)
+    .await?;
+```
+
+There is also a lower-level `create_record(CreateRequest)` (`lib.rs:237`) if we
+ever need to set fencing/durability/session explicitly.
+
+### Episodes and counting
+
+- `record_episode(RecordEpisodeRequest) -> RecordEpisodeResponse`
+  (`lib.rs:279`) — registers an episode (the arc) with its coordination
+  payload. This is where an `Episode` node's `episode_id` is minted/anchored
+  before the record rows that reference it land.
+- `count_records(kind: RecordKind, as_of: AsOf) -> CountRecordsAck`
+  (`lib.rs:408`) — bi-temporal `count(*)` over a kind shard; `AsOf::now()` +
+  no filter takes an O(1) counter path, any historical `as_of` takes the
+  scan path. This is the natural post-ingest assertion (see
+  `memcore-client/tests/count.rs`, which builds a `node_header(valid_at)` and
+  round-trips 50 inserts → count 50).
+
+---
+
+## 2. Mapping a `PreparedNode` onto a `RecordHeader`
+
+Memscribe emits four `PreparedNode` variants (`memscribe-core/src/node.rs:233`):
+`Conversation`, `Decision`, `Episode`, `Binding`. The sink's job is to derive a
+correct **bi-temporal header** for each. The two axes are distinct and must not
+be conflated:
+
+- **`valid_at` (valid time)** — when the fact was *true in the world*: the
+  turn/episode time. This comes **from the node**, never from the clock.
+- **`transaction_at` (transaction time)** — when MemDB *learned* the fact: our
+  **ingest** instant. One `MemDbSink` stamps a single `transaction_at` for the
+  whole batch (constructor arg, `src/memdb.rs`) so a replayed transcript lands
+  at one coherent transaction instant. On the wire this is simply the
+  wall-clock at which `create_record` is issued; MemDB does not take it as an
+  explicit `RecordHeader` field, so the sink keeps it in its own `BiTemporal`
+  header for audit and defers to the RPC instant on send.
+- **`episode_id` (the arc)** — `RecordHeader::episode_id`.
+
+The current derivation (`src/memdb.rs`, `header_for`) — intentionally
+conservative, deriving anchors only where they are *intrinsic* to the prepared
+node and never inferred:
+
+| Variant | `valid_at` | `episode_id` | Maps to MemDB |
+|---------|-----------|--------------|---------------|
+| `Episode(CodeEpisode)` | *(none on prepared node)* | `Some(e.episode_id)` | `RecordKind::Episode` (4); register via `record_episode`, then a `Node`/`Episode` record with `episode_id` set |
+| `Binding(BindingEdge)` | `Some(b.prov.t_gen)` — the `wasGeneratedBy` instant | `None` | `RecordKind::Edge` (2) `from → to`, typed by `relation` |
+| `Decision(DecisionRecord)` | *(none — only turn-seq spans)* | `None` | `RecordKind::Node` (1) |
+| `Conversation(ConversationSpan)` | *(none — only turn-seq spans)* | `None` | `RecordKind::Node` (1) |
+
+Why these choices:
+
+- **Episode → `episode_id`.** A `CodeEpisode` *is* an arc; its `episode_id`
+  (`node.rs:154`) is the deterministic id Memtrace keys co-change/provenance
+  off, so it maps straight to `RecordHeader::episode_id`. The prepared struct
+  carries no in-band `OffsetDateTime`, so `valid_at` is left for the consumer
+  (it falls back to `transaction_at`); the git sha / path travel in the body
+  and properties.
+- **Binding → `valid_at = prov.t_gen`.** A `BindingEdge` is a PROV edge with a
+  `ProvRecord` (`node.rs:173`) whose invariant is `t_use <= t_gen`
+  (`ProvRecord::is_temporally_valid`). The edit was *generated* at `t_gen`, so
+  that is the binding's valid time. A binding is an edge, not an arc, so
+  `episode_id` is `None`.
+- **Decision / Conversation.** These carry only `Range<u64>` turn-seq spans
+  (`node.rs:91`, `:137`), not wall-clock timestamps, so there is no intrinsic
+  `valid_at` to set without inference — and Memscribe is a zero-LLM,
+  no-guessing module (`node.rs:1-7`). Their valid time defaults to
+  `transaction_at` downstream. (If a future `PreparedNode` carries a real
+  timestamp, set `valid_at` from it here.)
+
+Converting `valid_at` to the wire type: `Micros((odt.unix_timestamp_nanos() /
+1_000) as i64)`, with `invalid_at = STILL_VALID` for a live row. `episode_id`
+(currently a deterministic `String`) is resolved to the `EpisodeId(u32)`
+returned/anchored by `record_episode`.
+
+---
+
+## 3. End-to-end ingest sketch (when `record_episode` and the client are wired)
+
+```rust
+// feature = "memdb"
+let client = MemcoreClient::connect("http://127.0.0.1:7878").await?;
+
+for record in sink.records() {          // each is a MemDbRecord { header, node_json }
+    let valid_at = record.header.valid_at
+        .map(|t| Micros((t.unix_timestamp_nanos() / 1_000) as i64))
+        .unwrap_or(Micros(/* ingest fallback */));
+
+    // Resolve the arc id (anchored once per episode via record_episode).
+    let episode_id = resolve_episode_id(&record.header.episode_id, &client).await?;
+
+    let header = RecordHeader {
+        rid: Rid(0),                    // server assigns
+        valid_at,
+        invalid_at: STILL_VALID,
+        episode_id,
+        schema_ver: SchemaVer(1),
+        kind: kind_for(&record),        // Node / Edge / Episode
+        flags: 0,
+    };
+    let props = property_rows(&record); // name/file_path/start_line/… via PropertyBuilder
+    client
+        .create_record_with_properties(header, record.node_json.into_bytes(), props)
+        .await?;
+}
+// Post-ingest assertion mirrors memcore-client/tests/count.rs:
+let ack = client.count_records(RecordKind::Node, AsOf::now()).await?;
+```
+
+The current `MemDbSink` stops one step short of this: it prepares every node
+into an in-memory `MemDbRecord` carrying the **correct bi-temporal shape**
+(`BiTemporal { valid_at, transaction_at, episode_id }`), which Memtrace's own
+integration test asserts against. Swapping the `Vec<MemDbRecord>` push for the
+`create_record_with_properties` call above is the only remaining wiring, and it
+is isolated entirely inside this one feature-gated module — no other crate
+changes.
+
+---
+
+## 4. Test posture
+
+- **Default build** (`cargo test -p memscribe-sink`): exercises `NdjsonSink`
+  and `SqliteSink` only. No MemDB, no network. `SqliteSink` is tested by
+  inserting each of the four `PreparedNode` variants and reading them back,
+  asserting the stored `fact_status` and JSON round-trip per variant, plus
+  primary-key upsert behaviour.
+- **MemDB build** (`cargo test -p memscribe-sink --features memdb`): adds the
+  `MemDbSink` bi-temporal tests — an `Episode` node yields a header with
+  `episode_id` set, and a `Binding` node yields `valid_at = prov.t_gen` — with
+  `transaction_at` shared across a batch. These run **without** a live MemDB
+  because the sink prepares records in memory; the real gRPC round-trip is
+  covered on Memtrace's side against a `MockEngine` (the pattern in
+  `memcore-client/tests/count.rs`).
+
+The invariant this preserves: turning the `memdb` feature on adds a
+destination; it never changes what a prepared node *is*, and turning it off
+leaves a complete, testable system on NDJSON.
diff --git a/crates/memscribe-sink/src/lib.rs b/crates/memscribe-sink/src/lib.rs
new file mode 100644
index 0000000..84d3bac
--- /dev/null
+++ b/crates/memscribe-sink/src/lib.rs
@@ -0,0 +1,21 @@
+//! # memscribe-sink
+//!
+//! Concrete [`memscribe_core::Sink`] implementations. The canonical, default
+//! sink is [`NdjsonSink`] — one JSON node per line — which makes the whole
+//! module observable and testable without any external service. [`SqliteSink`]
+//! (feature `sqlite`) gives a queryable local store, and `MemDbSink` (feature
+//! `memdb`, off by default) writes into MemDB for Memtrace.
+#![forbid(unsafe_code)]
+
+pub mod ndjson;
+pub use ndjson::NdjsonSink;
+
+#[cfg(feature = "sqlite")]
+pub mod sqlite;
+#[cfg(feature = "sqlite")]
+pub use sqlite::SqliteSink;
+
+#[cfg(feature = "memdb")]
+pub mod memdb;
+#[cfg(feature = "memdb")]
+pub use memdb::MemDbSink;
diff --git a/crates/memscribe-sink/src/memdb.rs b/crates/memscribe-sink/src/memdb.rs
new file mode 100644
index 0000000..71ea8aa
--- /dev/null
+++ b/crates/memscribe-sink/src/memdb.rs
@@ -0,0 +1,233 @@
+//! The MemDB sink (feature `memdb`, **off by default**).
+//!
+//! This is the single seam Memtrace turns on to ingest [`PreparedNode`]s into
+//! MemDB with their **bi-temporal** headers. The two time axes (see `MEMDB.md`
+//! and MemDB's `memcore_core::RecordHeader`) are:
+//!
+//! * **`valid_at`** — *valid time*: when the fact was true in the world. For
+//!   Memscribe that is the **turn / episode time** — the moment the agent edit
+//!   happened or the dialogue turn occurred — **not** when we ingested it.
+//! * **`transaction_at`** — *transaction time*: when MemDB learned the fact,
+//!   i.e. our **ingest time**. One sink instance stamps a single
+//!   `transaction_at` for the whole batch so a replayed transcript lands at one
+//!   coherent transaction instant.
+//! * **`episode_id`** — the **arc / episode** the node belongs to. MemDB keys
+//!   its episodic, co-change and provenance machinery off this id
+//!   (`RecordHeader::episode_id`).
+//!
+//! Memscribe is fully usable **without** MemDB: NDJSON is the default sink and
+//! this module only compiles when the `memdb` feature is enabled. Until the real
+//! gRPC client (`memcore_client::MemcoreClient`) is wired in (see `MEMDB.md` for
+//! the field-by-field mapping onto `RecordHeader` / `create_record_with_properties`),
+//! this sink prepares each node into an in-memory [`MemDbRecord`] carrying the
+//! correct bi-temporal shape, which Memtrace's own integration test asserts
+//! against.
+
+use memscribe_core::{PreparedNode, Sink, SinkError};
+use time::OffsetDateTime;
+
+/// Bi-temporal coordinates written alongside each node in MemDB.
+///
+/// These map directly onto MemDB's `memcore_core::RecordHeader`: `valid_at` →
+/// `RecordHeader::valid_at`, `transaction_at` is the wall-clock at which the
+/// `create_record` RPC is issued, and `episode_id` → `RecordHeader::episode_id`.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct BiTemporal {
+    /// The turn/episode time (**valid time**), when derivable from the node.
+    /// `None` when the node carries no intrinsic valid-time anchor (the consumer
+    /// then falls back to `transaction_at`).
+    pub valid_at: Option<OffsetDateTime>,
+    /// The ingest time (**transaction time**).
+    pub transaction_at: OffsetDateTime,
+    /// The arc/episode id, when the node carries one.
+    pub episode_id: Option<String>,
+}
+
+/// A node prepared for MemDB ingestion.
+#[derive(Debug, Clone)]
+pub struct MemDbRecord {
+    /// The bi-temporal header.
+    pub header: BiTemporal,
+    /// The node's canonical JSON.
+    pub node_json: String,
+}
+
+/// A sink that prepares nodes for MemDB ingestion.
+pub struct MemDbSink {
+    records: Vec<MemDbRecord>,
+    transaction_at: OffsetDateTime,
+}
+
+impl MemDbSink {
+    /// A sink stamping `transaction_at` as the ingest time. (Pass a fixed value
+    /// for deterministic tests; pass `OffsetDateTime::now_utc()` in production.)
+    #[must_use]
+    pub fn new(transaction_at: OffsetDateTime) -> Self {
+        MemDbSink {
+            records: Vec::new(),
+            transaction_at,
+        }
+    }
+
+    /// The records prepared so far.
+    #[must_use]
+    pub fn records(&self) -> &[MemDbRecord] {
+        &self.records
+    }
+
+    /// Derive the bi-temporal header for a node.
+    ///
+    /// The valid-time / episode anchors are intrinsic to the node and never
+    /// guessed:
+    ///
+    /// * [`Episode`](PreparedNode::Episode) carries an `episode_id` (the arc) but
+    ///   no in-band timestamp on the prepared struct, so its valid time is left
+    ///   to the consumer; `episode_id` is set.
+    /// * [`Binding`](PreparedNode::Binding) is a PROV edge: its valid time is the
+    ///   moment the bound edit was generated, `prov.t_gen` (the `wasGeneratedBy`
+    ///   instant). It is not itself an episode, so `episode_id` is `None`.
+    /// * [`Decision`](PreparedNode::Decision) and
+    ///   [`Conversation`](PreparedNode::Conversation) have no intrinsic
+    ///   `OffsetDateTime` on the prepared node (only turn-seq spans), so valid
+    ///   time falls back to `transaction_at` downstream; both leave `episode_id`
+    ///   unset here.
+    fn header_for(&self, node: &PreparedNode) -> BiTemporal {
+        let (valid_at, episode_id) = match node {
+            PreparedNode::Episode(e) => (None, Some(e.episode_id.clone())),
+            PreparedNode::Binding(b) => (Some(b.prov.t_gen), None),
+            PreparedNode::Decision(_) | PreparedNode::Conversation(_) => (None, None),
+        };
+        BiTemporal {
+            valid_at,
+            transaction_at: self.transaction_at,
+            episode_id,
+        }
+    }
+}
+
+impl Sink for MemDbSink {
+    fn emit(&mut self, node: &PreparedNode) -> Result<(), SinkError> {
+        let node_json =
+            serde_json::to_string(node).map_err(|e| SinkError::Serialize(e.to_string()))?;
+        self.records.push(MemDbRecord {
+            header: self.header_for(node),
+            node_json,
+        });
+        Ok(())
+    }
+
+    fn flush(&mut self) -> Result<(), SinkError> {
+        Ok(())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use memscribe_core::{
+        BindingEdge, CodeEpisode, ConversationSpan, DecisionRecord, Diff, FactStatus, NodeId,
+        ProvRecord, Relation,
+    };
+    use time::macros::datetime;
+
+    const INGEST: OffsetDateTime = datetime!(2026-06-22 12:00:00 UTC);
+    const GEN: OffsetDateTime = datetime!(2026-06-22 09:30:00 UTC);
+
+    fn episode() -> PreparedNode {
+        PreparedNode::Episode(CodeEpisode {
+            path: "a.rs".into(),
+            diff: Diff::for_path("a.rs"),
+            git: None,
+            episode_id: "ep-42".into(),
+        })
+    }
+
+    fn binding() -> PreparedNode {
+        PreparedNode::Binding(BindingEdge {
+            from: NodeId::new("decision:1"),
+            to: NodeId::new("episode:ep-42"),
+            relation: Relation::Produced,
+            prov: ProvRecord {
+                used_session: "sess-1".into(),
+                used_decision: Some(NodeId::new("decision:1")),
+                was_generated_by_session: "sess-1".into(),
+                t_use: datetime!(2026-06-22 09:00:00 UTC),
+                t_gen: GEN,
+            },
+            fact_status: FactStatus::DeterministicallyDerived,
+            correlation: None,
+        })
+    }
+
+    fn decision() -> PreparedNode {
+        PreparedNode::Decision(DecisionRecord {
+            epitome: "use postgres".into(),
+            considered_options: Vec::new(),
+            is_ban: false,
+            superseded_by: None,
+            confirmation: None,
+            source_span: 1..2,
+            fact_status: FactStatus::Observed,
+        })
+    }
+
+    fn conversation() -> PreparedNode {
+        PreparedNode::Conversation(ConversationSpan {
+            session_id: "sess-1".into(),
+            turn_range: 0..3,
+            text: "let's use postgres".into(),
+            markers: Vec::new(),
+            fact_status: FactStatus::Observed,
+            provenance: Vec::new(),
+        })
+    }
+
+    #[test]
+    fn episode_node_sets_episode_id() {
+        let mut sink = MemDbSink::new(INGEST);
+        sink.emit(&episode()).unwrap();
+        sink.flush().unwrap();
+        let header = &sink.records()[0].header;
+        assert_eq!(header.episode_id.as_deref(), Some("ep-42"));
+        assert_eq!(header.transaction_at, INGEST);
+        // An episode has no intrinsic valid timestamp on the prepared node.
+        assert_eq!(header.valid_at, None);
+    }
+
+    #[test]
+    fn binding_node_valid_at_is_prov_t_gen() {
+        let mut sink = MemDbSink::new(INGEST);
+        sink.emit(&binding()).unwrap();
+        sink.flush().unwrap();
+        let header = &sink.records()[0].header;
+        // valid time = when the bound edit was generated.
+        assert_eq!(header.valid_at, Some(GEN));
+        // transaction time = ingest time, independent of valid time.
+        assert_eq!(header.transaction_at, INGEST);
+        // A binding is an edge, not an episode.
+        assert_eq!(header.episode_id, None);
+    }
+
+    #[test]
+    fn decision_and_conversation_have_no_intrinsic_anchors() {
+        let mut sink = MemDbSink::new(INGEST);
+        sink.emit(&decision()).unwrap();
+        sink.emit(&conversation()).unwrap();
+        sink.flush().unwrap();
+        for rec in sink.records() {
+            assert_eq!(rec.header.valid_at, None);
+            assert_eq!(rec.header.episode_id, None);
+            assert_eq!(rec.header.transaction_at, INGEST);
+        }
+    }
+
+    #[test]
+    fn transaction_at_is_shared_across_a_batch() {
+        let mut sink = MemDbSink::new(INGEST);
+        sink.emit_all(&[episode(), binding(), decision()]).unwrap();
+        assert_eq!(sink.records().len(), 3);
+        for rec in sink.records() {
+            assert_eq!(rec.header.transaction_at, INGEST);
+        }
+    }
+}
diff --git a/crates/memscribe-sink/src/ndjson.rs b/crates/memscribe-sink/src/ndjson.rs
new file mode 100644
index 0000000..d7ded0b
--- /dev/null
+++ b/crates/memscribe-sink/src/ndjson.rs
@@ -0,0 +1,112 @@
+//! The NDJSON sink — the canonical, audit-friendly default. Writes one JSON
+//! `PreparedNode` per line to any [`std::io::Write`] (stdout, a file, a buffer).
+
+use memscribe_core::{PreparedNode, Sink, SinkError};
+use std::fs::File;
+use std::io::{self, BufWriter, Write};
+use std::path::Path;
+
+/// A sink that serializes each node as one line of JSON.
+pub struct NdjsonSink<W: Write> {
+    writer: W,
+    count: usize,
+}
+
+impl<W: Write> NdjsonSink<W> {
+    /// Wrap an arbitrary writer.
+    pub fn new(writer: W) -> Self {
+        NdjsonSink { writer, count: 0 }
+    }
+
+    /// The number of nodes emitted so far.
+    #[must_use]
+    pub fn count(&self) -> usize {
+        self.count
+    }
+
+    /// Consume the sink and return the inner writer.
+    pub fn into_inner(self) -> W {
+        self.writer
+    }
+}
+
+impl NdjsonSink<BufWriter<io::Stdout>> {
+    /// An NDJSON sink writing to stdout.
+    #[must_use]
+    pub fn stdout() -> Self {
+        NdjsonSink::new(BufWriter::new(io::stdout()))
+    }
+}
+
+impl NdjsonSink<BufWriter<File>> {
+    /// An NDJSON sink writing to a file at `path` (truncating any existing file).
+    ///
+    /// # Errors
+    /// Returns a [`SinkError`] if the file cannot be created.
+    pub fn file(path: impl AsRef<Path>) -> Result<Self, SinkError> {
+        let f = File::create(path).map_err(|e| SinkError::Write(e.to_string()))?;
+        Ok(NdjsonSink::new(BufWriter::new(f)))
+    }
+}
+
+impl<W: Write + Send> Sink for NdjsonSink<W> {
+    fn emit(&mut self, node: &PreparedNode) -> Result<(), SinkError> {
+        let line = serde_json::to_string(node).map_err(|e| SinkError::Serialize(e.to_string()))?;
+        self.writer
+            .write_all(line.as_bytes())
+            .map_err(|e| SinkError::Write(e.to_string()))?;
+        self.writer
+            .write_all(b"\n")
+            .map_err(|e| SinkError::Write(e.to_string()))?;
+        self.count += 1;
+        Ok(())
+    }
+
+    fn flush(&mut self) -> Result<(), SinkError> {
+        self.writer
+            .flush()
+            .map_err(|e| SinkError::Flush(e.to_string()))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use memscribe_core::{CodeEpisode, Diff, PreparedNode};
+
+    #[test]
+    fn writes_one_line_per_node() {
+        let node = PreparedNode::Episode(CodeEpisode {
+            path: "src/lib.rs".into(),
+            diff: Diff::for_path("src/lib.rs"),
+            git: None,
+            episode_id: "abc".into(),
+        });
+        let mut sink = NdjsonSink::new(Vec::new());
+        sink.emit(&node).unwrap();
+        sink.emit(&node).unwrap();
+        sink.flush().unwrap();
+        let out = String::from_utf8(sink.into_inner()).unwrap();
+        assert_eq!(out.lines().count(), 2);
+        // Each line is valid JSON.
+        for line in out.lines() {
+            let _: serde_json::Value = serde_json::from_str(line).unwrap();
+        }
+    }
+
+    #[test]
+    fn roundtrips_through_json() {
+        let node = PreparedNode::Episode(CodeEpisode {
+            path: "a.rs".into(),
+            diff: Diff::for_path("a.rs"),
+            git: None,
+            episode_id: "id".into(),
+        });
+        let mut sink = NdjsonSink::new(Vec::new());
+        sink.emit(&node).unwrap();
+        sink.flush().unwrap();
+        let out = String::from_utf8(sink.into_inner()).unwrap();
+        let back: PreparedNode = serde_json::from_str(out.trim()).unwrap();
+        assert_eq!(back, node);
+    }
+}
diff --git a/crates/memscribe-sink/src/sqlite.rs b/crates/memscribe-sink/src/sqlite.rs
new file mode 100644
index 0000000..9a95ea9
--- /dev/null
+++ b/crates/memscribe-sink/src/sqlite.rs
@@ -0,0 +1,249 @@
+//! The SQLite sink (feature `sqlite`) — a queryable local store with no external
+//! service. Each node is stored as JSON alongside its variant tag, its
+//! [`FactStatus`](memscribe_core::FactStatus), and a deterministic primary key
+//! derived from the node's own identity (so re-emitting the same node is an
+//! upsert, not a duplicate).
+
+use memscribe_core::{PreparedNode, Sink, SinkError};
+use rusqlite::Connection;
+use std::path::Path;
+
+/// A sink that writes nodes into a local SQLite database.
+pub struct SqliteSink {
+    conn: Connection,
+    count: usize,
+}
+
+impl SqliteSink {
+    /// Open (or create) a SQLite database at `path`.
+    ///
+    /// # Errors
+    /// Returns a [`SinkError`] if the database cannot be opened or initialized.
+    pub fn open(path: impl AsRef<Path>) -> Result<Self, SinkError> {
+        let conn = Connection::open(path).map_err(|e| SinkError::Write(e.to_string()))?;
+        Self::init(conn)
+    }
+
+    /// An in-memory SQLite database (for tests).
+    ///
+    /// # Errors
+    /// Returns a [`SinkError`] if the database cannot be initialized.
+    pub fn in_memory() -> Result<Self, SinkError> {
+        let conn = Connection::open_in_memory().map_err(|e| SinkError::Write(e.to_string()))?;
+        Self::init(conn)
+    }
+
+    fn init(conn: Connection) -> Result<Self, SinkError> {
+        // `pk` is the node's own stable identity (see `primary_key`), so an
+        // `INSERT OR REPLACE` re-emit of the same node updates in place rather
+        // than appending a duplicate row. `fact_status` is indexed alongside
+        // `node_type` so consumers can filter on epistemic status without
+        // re-parsing the JSON blob.
+        conn.execute_batch(
+            "CREATE TABLE IF NOT EXISTS nodes (
+                pk          TEXT PRIMARY KEY,
+                node_type   TEXT NOT NULL,
+                fact_status TEXT NOT NULL,
+                json        TEXT NOT NULL
+            );
+            CREATE INDEX IF NOT EXISTS idx_nodes_type ON nodes(node_type);
+            CREATE INDEX IF NOT EXISTS idx_nodes_fact_status ON nodes(fact_status);",
+        )
+        .map_err(|e| SinkError::Write(e.to_string()))?;
+        Ok(SqliteSink { conn, count: 0 })
+    }
+
+    /// The number of `emit` calls accepted so far. Note this counts emissions,
+    /// not distinct rows: re-emitting a node with the same [`primary_key`] is an
+    /// upsert, so the table may hold fewer rows than this value.
+    ///
+    /// [`primary_key`]: SqliteSink::primary_key
+    #[must_use]
+    pub fn count(&self) -> usize {
+        self.count
+    }
+
+    /// Count the rows currently stored for a given variant tag (e.g. `"episode"`,
+    /// `"binding"`, `"decision"`, `"conversation"` — see [`PreparedNode::tag`]).
+    ///
+    /// # Errors
+    /// Returns a [`SinkError`] if the query fails.
+    pub fn query_count_by_type(&self, node_type: &str) -> Result<u64, SinkError> {
+        let n: i64 = self
+            .conn
+            .query_row(
+                "SELECT COUNT(*) FROM nodes WHERE node_type = ?1",
+                rusqlite::params![node_type],
+                |r| r.get(0),
+            )
+            .map_err(|e| SinkError::Write(e.to_string()))?;
+        // COUNT(*) is non-negative; the cast is lossless for any real table size.
+        Ok(n.max(0) as u64)
+    }
+
+    /// A deterministic primary key for a node, derived from the node's own
+    /// identity so the same logical node always maps to the same row. The key is
+    /// namespaced by the variant tag so two different variants can never collide.
+    ///
+    /// The identity per variant:
+    /// * `Episode`   → its `episode_id`.
+    /// * `Binding`   → `from→to` plus the relation (a directed, typed edge).
+    /// * `Decision`  → the session-agnostic turn span it was parsed from.
+    /// * `Conversation` → session id plus its turn range.
+    #[must_use]
+    pub fn primary_key(node: &PreparedNode) -> String {
+        let tag = node.tag();
+        match node {
+            PreparedNode::Episode(e) => format!("{tag}:{}", e.episode_id),
+            PreparedNode::Binding(b) => {
+                format!("{tag}:{}->{}:{:?}", b.from, b.to, b.relation)
+            }
+            PreparedNode::Decision(d) => {
+                format!("{tag}:{}..{}", d.source_span.start, d.source_span.end)
+            }
+            PreparedNode::Conversation(c) => {
+                format!(
+                    "{tag}:{}:{}..{}",
+                    c.session_id, c.turn_range.start, c.turn_range.end
+                )
+            }
+        }
+    }
+}
+
+impl Sink for SqliteSink {
+    fn emit(&mut self, node: &PreparedNode) -> Result<(), SinkError> {
+        let json = serde_json::to_string(node).map_err(|e| SinkError::Serialize(e.to_string()))?;
+        let fact = format!("{:?}", node.fact_status());
+        let pk = Self::primary_key(node);
+        self.conn
+            .execute(
+                "INSERT OR REPLACE INTO nodes (pk, node_type, fact_status, json)
+                 VALUES (?1, ?2, ?3, ?4)",
+                rusqlite::params![pk, node.tag(), fact, json],
+            )
+            .map_err(|e| SinkError::Write(e.to_string()))?;
+        self.count += 1;
+        Ok(())
+    }
+
+    fn flush(&mut self) -> Result<(), SinkError> {
+        Ok(())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use memscribe_core::{
+        BindingEdge, CodeEpisode, ConversationSpan, DecisionRecord, Diff, FactStatus, NodeId,
+        PreparedNode, ProvRecord, Relation,
+    };
+    use time::OffsetDateTime;
+
+    fn episode() -> PreparedNode {
+        PreparedNode::Episode(CodeEpisode {
+            path: "a.rs".into(),
+            diff: Diff::for_path("a.rs"),
+            git: None,
+            episode_id: "ep-1".into(),
+        })
+    }
+
+    fn conversation() -> PreparedNode {
+        PreparedNode::Conversation(ConversationSpan {
+            session_id: "sess-1".into(),
+            turn_range: 0..3,
+            text: "let's use postgres".into(),
+            markers: Vec::new(),
+            fact_status: FactStatus::Observed,
+            provenance: Vec::new(),
+        })
+    }
+
+    fn decision() -> PreparedNode {
+        PreparedNode::Decision(DecisionRecord {
+            epitome: "use postgres".into(),
+            considered_options: Vec::new(),
+            is_ban: false,
+            superseded_by: None,
+            confirmation: None,
+            source_span: 1..2,
+            fact_status: FactStatus::Observed,
+        })
+    }
+
+    fn binding() -> PreparedNode {
+        PreparedNode::Binding(BindingEdge {
+            from: NodeId::new("decision:1"),
+            to: NodeId::new("episode:ep-1"),
+            relation: Relation::Produced,
+            prov: ProvRecord {
+                used_session: "sess-1".into(),
+                used_decision: Some(NodeId::new("decision:1")),
+                was_generated_by_session: "sess-1".into(),
+                t_use: OffsetDateTime::UNIX_EPOCH,
+                t_gen: OffsetDateTime::UNIX_EPOCH,
+            },
+            fact_status: FactStatus::DeterministicallyDerived,
+            correlation: None,
+        })
+    }
+
+    #[test]
+    fn inserts_nodes() {
+        let mut sink = SqliteSink::in_memory().unwrap();
+        sink.emit(&episode()).unwrap();
+        sink.flush().unwrap();
+        assert_eq!(sink.count(), 1);
+        let n: i64 = sink
+            .conn
+            .query_row("SELECT COUNT(*) FROM nodes", [], |r| r.get(0))
+            .unwrap();
+        assert_eq!(n, 1);
+    }
+
+    #[test]
+    fn stores_all_four_variants_and_reads_them_back() {
+        let mut sink = SqliteSink::in_memory().unwrap();
+        let nodes = [conversation(), decision(), episode(), binding()];
+        for n in &nodes {
+            sink.emit(n).unwrap();
+        }
+        sink.flush().unwrap();
+
+        // One row per variant tag.
+        for tag in ["conversation", "decision", "episode", "binding"] {
+            assert_eq!(sink.query_count_by_type(tag).unwrap(), 1, "tag {tag}");
+        }
+        assert_eq!(sink.query_count_by_type("nonexistent").unwrap(), 0);
+
+        // The stored fact_status matches each node's own status, and the JSON
+        // round-trips back to the exact node.
+        for original in &nodes {
+            let pk = SqliteSink::primary_key(original);
+            let (fact, json): (String, String) = sink
+                .conn
+                .query_row(
+                    "SELECT fact_status, json FROM nodes WHERE pk = ?1",
+                    rusqlite::params![pk],
+                    |r| Ok((r.get(0)?, r.get(1)?)),
+                )
+                .unwrap();
+            assert_eq!(fact, format!("{:?}", original.fact_status()));
+            let back: PreparedNode = serde_json::from_str(&json).unwrap();
+            assert_eq!(&back, original);
+        }
+    }
+
+    #[test]
+    fn primary_key_is_stable_so_re_emit_upserts() {
+        let mut sink = SqliteSink::in_memory().unwrap();
+        sink.emit(&episode()).unwrap();
+        sink.emit(&episode()).unwrap();
+        sink.flush().unwrap();
+        // Two emits, but the same identity → a single row.
+        assert_eq!(sink.count(), 2);
+        assert_eq!(sink.query_count_by_type("episode").unwrap(), 1);
+    }
+}
diff --git a/crates/memscribe-testkit/Cargo.toml b/crates/memscribe-testkit/Cargo.toml
new file mode 100644
index 0000000..8eda58b
--- /dev/null
+++ b/crates/memscribe-testkit/Cargo.toml
@@ -0,0 +1,23 @@
+[package]
+name = "memscribe-testkit"
+version.workspace = true
+edition.workspace = true
+rust-version.workspace = true
+license.workspace = true
+repository.workspace = true
+description = "Test harness for Memscribe: golden-file loaders, invariant checks, conformance scenarios, synthetic generators."
+
+[lints]
+workspace = true
+
+[dependencies]
+memscribe-core = { workspace = true }
+memscribe-adapters = { workspace = true }
+memscribe-sink = { workspace = true }
+memscribe-io = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+
+[dev-dependencies]
+insta = { workspace = true }
+proptest = { workspace = true }
diff --git a/crates/memscribe-testkit/src/golden.rs b/crates/memscribe-testkit/src/golden.rs
new file mode 100644
index 0000000..711be92
--- /dev/null
+++ b/crates/memscribe-testkit/src/golden.rs
@@ -0,0 +1,138 @@
+//! Golden-file fixture resolution and comparison (whitepaper §8.1).
+//!
+//! Layout:
+//! ```text
+//! fixtures/<tool>/<version>/<case>.jsonl            # input
+//! fixtures-expected/<tool>/<version>/<case>.events.json   # normalized events
+//! fixtures-expected/<tool>/<version>/<case>.nodes.json    # prepared nodes
+//! ```
+
+use std::path::{Path, PathBuf};
+
+/// The workspace root (two levels up from this crate's manifest dir).
+#[must_use]
+pub fn workspace_root() -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR"))
+        .ancestors()
+        .nth(2)
+        .map(Path::to_path_buf)
+        .unwrap_or_else(|| PathBuf::from("."))
+}
+
+/// The `fixtures/` directory.
+#[must_use]
+pub fn fixtures_dir() -> PathBuf {
+    workspace_root().join("fixtures")
+}
+
+/// The `fixtures-expected/` directory.
+#[must_use]
+pub fn fixtures_expected_dir() -> PathBuf {
+    workspace_root().join("fixtures-expected")
+}
+
+/// A single golden case: the input transcript and its expected outputs.
+#[derive(Debug, Clone)]
+pub struct GoldenCase {
+    /// The tool slug (e.g. `claude_code`).
+    pub tool: String,
+    /// The tool version slug (e.g. `2.1`).
+    pub version: String,
+    /// The case name.
+    pub case: String,
+}
+
+impl GoldenCase {
+    /// Construct a golden case descriptor.
+    pub fn new(
+        tool: impl Into<String>,
+        version: impl Into<String>,
+        case: impl Into<String>,
+    ) -> Self {
+        GoldenCase {
+            tool: tool.into(),
+            version: version.into(),
+            case: case.into(),
+        }
+    }
+
+    /// The input transcript path.
+    #[must_use]
+    pub fn input_path(&self) -> PathBuf {
+        fixtures_dir()
+            .join(&self.tool)
+            .join(&self.version)
+            .join(format!("{}.jsonl", self.case))
+    }
+
+    /// The expected normalized-events path.
+    #[must_use]
+    pub fn expected_events_path(&self) -> PathBuf {
+        fixtures_expected_dir()
+            .join(&self.tool)
+            .join(&self.version)
+            .join(format!("{}.events.json", self.case))
+    }
+
+    /// The expected prepared-nodes path.
+    #[must_use]
+    pub fn expected_nodes_path(&self) -> PathBuf {
+        fixtures_expected_dir()
+            .join(&self.tool)
+            .join(&self.version)
+            .join(format!("{}.nodes.json", self.case))
+    }
+
+    /// Read the input transcript bytes.
+    ///
+    /// # Errors
+    /// Returns an [`std::io::Error`] if the fixture is missing.
+    pub fn read_input(&self) -> std::io::Result<Vec<u8>> {
+        std::fs::read(self.input_path())
+    }
+}
+
+/// Discover every `*.jsonl` fixture under `fixtures/`, returning golden-case
+/// descriptors. Useful for a data-driven test that iterates all cases.
+#[must_use]
+pub fn discover_cases() -> Vec<GoldenCase> {
+    let root = fixtures_dir();
+    let mut cases = Vec::new();
+    let Ok(tools) = std::fs::read_dir(&root) else {
+        return cases;
+    };
+    for tool in tools.flatten() {
+        if !tool.path().is_dir() {
+            continue;
+        }
+        let tool_name = tool.file_name().to_string_lossy().to_string();
+        let Ok(versions) = std::fs::read_dir(tool.path()) else {
+            continue;
+        };
+        for version in versions.flatten() {
+            if !version.path().is_dir() {
+                continue;
+            }
+            let version_name = version.file_name().to_string_lossy().to_string();
+            let Ok(files) = std::fs::read_dir(version.path()) else {
+                continue;
+            };
+            for file in files.flatten() {
+                let path = file.path();
+                if path.extension().and_then(|e| e.to_str()) == Some("jsonl") {
+                    if let Some(stem) = path.file_stem().and_then(|s| s.to_str()) {
+                        cases.push(GoldenCase::new(&tool_name, &version_name, stem));
+                    }
+                }
+            }
+        }
+    }
+    cases.sort_by(|a, b| {
+        (a.tool.clone(), a.version.clone(), a.case.clone()).cmp(&(
+            b.tool.clone(),
+            b.version.clone(),
+            b.case.clone(),
+        ))
+    });
+    cases
+}
diff --git a/crates/memscribe-testkit/src/invariants.rs b/crates/memscribe-testkit/src/invariants.rs
new file mode 100644
index 0000000..49df143
--- /dev/null
+++ b/crates/memscribe-testkit/src/invariants.rs
@@ -0,0 +1,90 @@
+//! Reusable invariant checks (whitepaper §8.3). These are written as `Result`
+//! returns so they can be used both in `#[test]`s and in `proptest` bodies.
+
+use memscribe_core::CaptureEvent;
+use std::collections::HashMap;
+
+/// `seq` is strictly increasing within a session and matches file order.
+///
+/// # Errors
+/// Returns a message describing the first violation.
+pub fn check_monotonic_seq(events: &[CaptureEvent]) -> Result<(), String> {
+    let mut last: HashMap<&str, u64> = HashMap::new();
+    for ev in events {
+        if let Some(prev) = last.get(ev.session_id.as_str()) {
+            if ev.seq <= *prev {
+                return Err(format!(
+                    "seq not strictly increasing in session {}: {} after {}",
+                    ev.session_id, ev.seq, prev
+                ));
+            }
+        }
+        last.insert(ev.session_id.as_str(), ev.seq);
+    }
+    Ok(())
+}
+
+/// Losslessness: every non-blank source record maps to at least one event.
+///
+/// # Errors
+/// Returns a message if fewer events than records were produced.
+pub fn check_lossless(nonblank_record_count: usize, events: &[CaptureEvent]) -> Result<(), String> {
+    if events.len() < nonblank_record_count {
+        return Err(format!(
+            "lossy: {} records produced only {} events",
+            nonblank_record_count,
+            events.len()
+        ));
+    }
+    Ok(())
+}
+
+/// Idempotency by `event_id`: re-ingesting the same input yields the same set of
+/// `(session_id, event_id)` keys with no duplicates introduced.
+///
+/// # Errors
+/// Returns a message if duplicate dedup keys are present.
+pub fn check_unique_event_ids(events: &[CaptureEvent]) -> Result<(), String> {
+    let mut seen = std::collections::HashSet::new();
+    for ev in events {
+        let key = (ev.session_id.clone(), ev.event_id.clone());
+        if !seen.insert(key) {
+            return Err(format!(
+                "duplicate event_id {} in session {}",
+                ev.event_id, ev.session_id
+            ));
+        }
+    }
+    Ok(())
+}
+
+/// Determinism: two parses of the same input are byte-identical (serialized).
+///
+/// # Errors
+/// Returns a message if the two event vectors differ.
+pub fn check_determinism(a: &[CaptureEvent], b: &[CaptureEvent]) -> Result<(), String> {
+    let ja = serde_json::to_string(a).map_err(|e| e.to_string())?;
+    let jb = serde_json::to_string(b).map_err(|e| e.to_string())?;
+    if ja != jb {
+        return Err("parse is not deterministic across runs".to_string());
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::parse_events;
+    use memscribe_core::SourceKind;
+    use std::path::Path;
+
+    #[test]
+    fn stub_stream_satisfies_invariants() {
+        let jsonl = b"{\"type\":\"a\"}\n{\"type\":\"b\"}\n";
+        let events = parse_events(SourceKind::ClaudeCode, jsonl, Path::new("t.jsonl"));
+        check_monotonic_seq(&events).unwrap();
+        check_lossless(2, &events).unwrap();
+        let again = parse_events(SourceKind::ClaudeCode, jsonl, Path::new("t.jsonl"));
+        check_determinism(&events, &again).unwrap();
+    }
+}
diff --git a/crates/memscribe-testkit/src/lib.rs b/crates/memscribe-testkit/src/lib.rs
new file mode 100644
index 0000000..6921d0e
--- /dev/null
+++ b/crates/memscribe-testkit/src/lib.rs
@@ -0,0 +1,56 @@
+//! # memscribe-testkit
+//!
+//! The test harness that makes Memscribe's determinism a property you can run.
+//! It provides:
+//!
+//! - [`parse_events`] / [`prepare_nodes`] — drive a tool's adapter and the
+//!   pipeline over raw bytes, the way every golden and property test does.
+//! - [`invariants`] — reusable checks for the whitepaper §8.3 invariants
+//!   (determinism, monotonic seq, losslessness, idempotency).
+//! - [`golden`] — fixture path resolution and load/compare helpers.
+//! - [`scenarios`] — the cross-tool conformance scenario catalog (§8.2).
+#![forbid(unsafe_code)]
+
+pub mod golden;
+pub mod invariants;
+pub mod scenarios;
+
+use memscribe_adapters::adapter_for;
+use memscribe_core::{
+    pipeline::parse_records, CaptureEvent, DefaultPipeline, PreparedNode, SourceKind,
+};
+use memscribe_io::read_records_from_bytes;
+use std::path::Path;
+
+/// Parse a tool's transcript bytes into the normalized event stream.
+///
+/// # Panics
+/// Panics if the adapter for `tool` is not compiled into the build.
+#[must_use]
+pub fn parse_events(tool: SourceKind, jsonl: &[u8], path: &Path) -> Vec<CaptureEvent> {
+    let recs = read_records_from_bytes(jsonl, path);
+    let adapter = adapter_for(tool).expect("adapter feature must be enabled for this tool");
+    let (events, _ctx) = parse_records(adapter.as_ref(), &recs);
+    events
+}
+
+/// Parse and prepare a tool's transcript bytes into the prepared-node stream
+/// (redaction off, so tests can assert on verbatim content).
+///
+/// # Panics
+/// Panics if the adapter for `tool` is not compiled into the build.
+#[must_use]
+pub fn prepare_nodes(tool: SourceKind, jsonl: &[u8], path: &Path) -> Vec<PreparedNode> {
+    let recs = read_records_from_bytes(jsonl, path);
+    let adapter = adapter_for(tool).expect("adapter feature must be enabled for this tool");
+    DefaultPipeline::without_redaction().run_records(adapter.as_ref(), &recs)
+}
+
+/// Count non-blank lines — the lower bound on events for the losslessness check.
+#[must_use]
+pub fn count_nonblank_lines(jsonl: &[u8]) -> usize {
+    String::from_utf8_lossy(jsonl)
+        .lines()
+        .filter(|l| !l.trim().is_empty())
+        .count()
+}
diff --git a/crates/memscribe-testkit/src/scenarios.rs b/crates/memscribe-testkit/src/scenarios.rs
new file mode 100644
index 0000000..31e046e
--- /dev/null
+++ b/crates/memscribe-testkit/src/scenarios.rs
@@ -0,0 +1,74 @@
+//! The cross-tool conformance scenario catalog (whitepaper §8.2).
+//!
+//! A canonical set of scenarios — authored once, captured from every tool — that
+//! must normalize to the same shape regardless of which tool produced them. Each
+//! scenario names the invariant the conformance suite asserts.
+
+/// One canonical conformance scenario.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct Scenario {
+    /// A stable slug used as the fixture `<case>` name.
+    pub slug: &'static str,
+    /// What the scenario exercises and the shape it must normalize to.
+    pub expectation: &'static str,
+}
+
+/// The canonical scenarios every adapter's fixtures must cover.
+pub const SCENARIOS: &[Scenario] = &[
+    Scenario {
+        slug: "happy_path_decision_then_edits",
+        expectation: "a decision turn followed by edits to N files → 1 Decision, N Episodes, N Bindings",
+    },
+    Scenario {
+        slug: "rejected_alternative",
+        expectation: "\"use Stripe instead of PayPal\" → considered_options populated, the unchosen one marked",
+    },
+    Scenario {
+        slug: "ban",
+        expectation: "\"we will NOT add a dependency on X\" → is_ban = true",
+    },
+    Scenario {
+        slug: "interleaved_arcs",
+        expectation: "two decisions, edits to overlapping files → correct per-decision binding",
+    },
+    Scenario {
+        slug: "multi_edit_single_commit",
+        expectation: "a single commit touching several files → several Episodes",
+    },
+    Scenario {
+        slug: "tool_failure",
+        expectation: "edit rejected (ToolResult.ok = false) → no spurious Episode",
+    },
+    Scenario {
+        slug: "rewind_compaction",
+        expectation: "rewind/compaction flagged, verbatim history preserved, current view honors it",
+    },
+    Scenario {
+        slug: "subagent_thread",
+        expectation: "a subagent thread → attributed, not merged",
+    },
+    Scenario {
+        slug: "no_commitment_marker",
+        expectation: "a turn with no marker → no Conversation node elevated, verbatim turn still retained",
+    },
+];
+
+/// The scenario slugs, for iterating fixtures.
+#[must_use]
+pub fn slugs() -> Vec<&'static str> {
+    SCENARIOS.iter().map(|s| s.slug).collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn catalog_is_complete_and_unique() {
+        let mut seen = std::collections::HashSet::new();
+        for s in SCENARIOS {
+            assert!(seen.insert(s.slug), "duplicate scenario slug {}", s.slug);
+        }
+        assert_eq!(SCENARIOS.len(), 9);
+    }
+}
diff --git a/crates/memscribe-testkit/tests/conformance.rs b/crates/memscribe-testkit/tests/conformance.rs
new file mode 100644
index 0000000..dcfbf55
--- /dev/null
+++ b/crates/memscribe-testkit/tests/conformance.rs
@@ -0,0 +1,827 @@
+//! Cross-tool conformance suite (whitepaper §8.2).
+//!
+//! The thesis Memscribe sells is *adapter interchangeability*: a decision turn,
+//! a rejection, a ban, or a failed edit should normalize to the **same
+//! structural shape** no matter which of the nine tools produced the transcript.
+//! These tests prove that against the canonical fixtures by driving every tool's
+//! adapter through the real pipeline (`testkit::prepare_nodes`) and comparing the
+//! resulting node shapes.
+//!
+//! Where a scenario's fixtures were authored *identically* across tools
+//! (`happy_path_decision_then_edits`), we assert the full shape is byte-identical
+//! across all nine. Where fixture content legitimately differs per tool (the
+//! `rejected_alternative` corpus uses different examples), we assert the weaker
+//! cross-tool invariant the scenario actually guarantees. For `tool_failure`, all
+//! nine tools now uphold the same zero-episode invariant: a failed edit
+//! (`ToolResult.ok = false`, linked to its `FileEdit` by `call_id`) produces no
+//! spurious `CodeEpisode` and therefore no binding. See
+//! `tool_failure_yields_no_spurious_episode`.
+
+use memscribe_core::node::{BindingEdge, DecisionRecord, PreparedNode};
+use memscribe_core::SourceKind;
+use memscribe_testkit::golden::fixtures_dir;
+use memscribe_testkit::{parse_events, prepare_nodes};
+use std::collections::BTreeSet;
+use std::path::{Path, PathBuf};
+
+/// Every tool and the version slug its fixtures live under. Driven dynamically
+/// against the fixtures on disk so a new tool/version is picked up automatically.
+const TOOLS: &[(SourceKind, &str)] = &[
+    (SourceKind::ClaudeCode, "2.0"),
+    (SourceKind::Codex, "v2"),
+    (SourceKind::Gemini, "v1"),
+    (SourceKind::Otel, "genai"),
+    (SourceKind::Cursor, "v1"),
+    (SourceKind::Windsurf, "v1"),
+    (SourceKind::Zed, "v1"),
+    (SourceKind::VsCode, "v1"),
+    (SourceKind::Copilot, "v1"),
+];
+
+/// The three first-class CLIs that carry the five additional §8.2 scenarios
+/// (`interleaved_arcs`, `multi_edit_single_commit`, `rewind_compaction`,
+/// `subagent_thread`, `no_commitment_marker`). The OTel/IDE adapters do not
+/// author these fixtures, so the extended-scenario tests are scoped to this set.
+const THREE_CLIS: &[(SourceKind, &str)] = &[
+    (SourceKind::ClaudeCode, "2.0"),
+    (SourceKind::Codex, "v2"),
+    (SourceKind::Gemini, "v1"),
+];
+
+/// A structural fingerprint of a prepared-node stream: counts per variant plus
+/// the decision-level flags the conformance contract pins. Two tools that
+/// normalize the same scenario to the same `Shape` are interchangeable behind the
+/// contract for that scenario.
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct Shape {
+    conversations: usize,
+    decisions: usize,
+    episodes: usize,
+    bindings: usize,
+    is_ban: bool,
+    chosen: BTreeSet<String>,
+    rejected: BTreeSet<String>,
+}
+
+impl Shape {
+    fn of(nodes: &[PreparedNode]) -> Self {
+        let mut shape = Shape {
+            conversations: 0,
+            decisions: 0,
+            episodes: 0,
+            bindings: 0,
+            is_ban: false,
+            chosen: BTreeSet::new(),
+            rejected: BTreeSet::new(),
+        };
+        for n in nodes {
+            match n {
+                PreparedNode::Conversation(_) => shape.conversations += 1,
+                PreparedNode::Decision(d) => {
+                    shape.decisions += 1;
+                    shape.fold_decision(d);
+                }
+                PreparedNode::Episode(_) => shape.episodes += 1,
+                PreparedNode::Binding(_) => shape.bindings += 1,
+            }
+        }
+        shape
+    }
+
+    fn fold_decision(&mut self, d: &DecisionRecord) {
+        if d.is_ban {
+            self.is_ban = true;
+        }
+        for opt in &d.considered_options {
+            if opt.chosen {
+                self.chosen.insert(opt.text.clone());
+            } else {
+                self.rejected.insert(opt.text.clone());
+            }
+        }
+    }
+}
+
+/// Resolve a fixture's input path and the stable *relative* path we feed the
+/// pipeline (so provenance and any path-derived ids are machine-independent).
+fn fixture_paths(tool: SourceKind, version: &str, case: &str) -> (PathBuf, PathBuf) {
+    let file = format!("{case}.jsonl");
+    let abs = fixtures_dir().join(tool.as_str()).join(version).join(&file);
+    let rel = Path::new("fixtures")
+        .join(tool.as_str())
+        .join(version)
+        .join(&file);
+    (abs, rel)
+}
+
+/// Drive a tool's adapter + pipeline over a fixture and return the prepared
+/// nodes, or `None` when the fixture is absent for that tool.
+fn nodes_for(tool: SourceKind, version: &str, case: &str) -> Option<Vec<PreparedNode>> {
+    let (abs, rel) = fixture_paths(tool, version, case);
+    let bytes = std::fs::read(abs).ok()?;
+    Some(prepare_nodes(tool, &bytes, &rel))
+}
+
+/// The shape a tool normalizes a scenario to, or `None` if the fixture is absent.
+fn shape_for(tool: SourceKind, version: &str, case: &str) -> Option<Shape> {
+    nodes_for(tool, version, case).map(|n| Shape::of(&n))
+}
+
+/// Drive a tool's adapter over a fixture and return the normalized event stream,
+/// or `None` when the fixture is absent. Used to assert event-layer invariants
+/// (verbatim retention) that hold *below* the gate.
+fn events_for(
+    tool: SourceKind,
+    version: &str,
+    case: &str,
+) -> Option<Vec<memscribe_core::CaptureEvent>> {
+    let (abs, rel) = fixture_paths(tool, version, case);
+    let bytes = std::fs::read(abs).ok()?;
+    Some(parse_events(tool, &bytes, &rel))
+}
+
+/// Collect every binding edge in a prepared-node stream, in stream order.
+fn bindings_of(nodes: &[PreparedNode]) -> Vec<&BindingEdge> {
+    nodes
+        .iter()
+        .filter_map(|n| match n {
+            PreparedNode::Binding(b) => Some(b),
+            _ => None,
+        })
+        .collect()
+}
+
+/// Collect every decision record in a prepared-node stream, in stream order.
+fn decisions_of(nodes: &[PreparedNode]) -> Vec<&DecisionRecord> {
+    nodes
+        .iter()
+        .filter_map(|n| match n {
+            PreparedNode::Decision(d) => Some(d),
+            _ => None,
+        })
+        .collect()
+}
+
+/// Whether the node stream marks at least one decision as superseded
+/// (`superseded_by` is `Some`) — the structural signature of a rewind/compaction
+/// the adapter surfaced to the segmenter as a typed `Rewind`/`Compaction` event.
+fn has_superseded_decision(nodes: &[PreparedNode]) -> bool {
+    decisions_of(nodes)
+        .iter()
+        .any(|d| d.superseded_by.is_some())
+}
+
+// ---------------------------------------------------------------------------
+// happy_path_decision_then_edits — authored identically across all nine tools.
+// This is the strongest interchangeability claim: byte-identical shape.
+// ---------------------------------------------------------------------------
+
+#[test]
+fn happy_path_normalizes_to_identical_shape_across_every_tool() {
+    let case = "happy_path_decision_then_edits";
+
+    // The canonical shape: one gated conversation, one decision (Postgres chosen
+    // over MySQL, not a ban), two file-edit episodes, two bindings.
+    let mut chosen = BTreeSet::new();
+    chosen.insert("Postgres".to_string());
+    let mut rejected = BTreeSet::new();
+    rejected.insert("MySQL".to_string());
+    let canonical = Shape {
+        conversations: 1,
+        decisions: 1,
+        episodes: 2,
+        bindings: 2,
+        is_ban: false,
+        chosen,
+        rejected,
+    };
+
+    let mut seen = 0;
+    for &(tool, version) in TOOLS {
+        let Some(shape) = shape_for(tool, version, case) else {
+            panic!("missing {case} fixture for {tool}");
+        };
+        assert_eq!(
+            shape, canonical,
+            "{tool} normalized {case} to a different shape than the contract; \
+             adapters must be interchangeable for this scenario"
+        );
+        seen += 1;
+    }
+    assert_eq!(
+        seen,
+        TOOLS.len(),
+        "every tool must carry the {case} fixture"
+    );
+}
+
+// ---------------------------------------------------------------------------
+// ban — fixture content differs per tool, but the *ban flag* is the invariant
+// the scenario guarantees, and it must hold for all nine. We also assert the
+// gate elevated exactly one conversation + one decision per tool.
+// ---------------------------------------------------------------------------
+
+#[test]
+fn ban_sets_is_ban_true_for_every_tool() {
+    let case = "ban";
+    let mut seen = 0;
+    for &(tool, version) in TOOLS {
+        let Some(shape) = shape_for(tool, version, case) else {
+            panic!("missing {case} fixture for {tool}");
+        };
+        assert!(
+            shape.is_ban,
+            "{tool} failed to flag the {case} scenario as a ban (is_ban must be true)"
+        );
+        assert_eq!(
+            shape.conversations, 1,
+            "{tool} {case}: expected exactly one gated conversation"
+        );
+        assert_eq!(
+            shape.decisions, 1,
+            "{tool} {case}: expected exactly one decision carrying the ban"
+        );
+        seen += 1;
+    }
+    assert_eq!(
+        seen,
+        TOOLS.len(),
+        "every tool must carry the {case} fixture"
+    );
+}
+
+// ---------------------------------------------------------------------------
+// rejected_alternative — the scenario invariant is that *when* the gate
+// elevates a decision from a rejection-marked turn, the considered options carry
+// both a chosen and a rejected alternative. The canonical "Stripe instead of
+// PayPal" corpus is shared by the four tools that authored it identically; the
+// rest use different examples (and codex's phrasing intentionally does not trip
+// the gate). We assert the shared sub-corpus is identical, and that no tool
+// fabricates an option set out of thin air.
+// ---------------------------------------------------------------------------
+
+#[test]
+fn rejected_alternative_shared_corpus_is_identical() {
+    let case = "rejected_alternative";
+    // Tools whose fixtures use the canonical "Stripe instead of PayPal" text.
+    let canonical_corpus = [
+        SourceKind::ClaudeCode,
+        SourceKind::Gemini,
+        SourceKind::Otel,
+        SourceKind::Copilot,
+    ];
+
+    let mut chosen = BTreeSet::new();
+    chosen.insert("Stripe".to_string());
+    let mut rejected = BTreeSet::new();
+    rejected.insert("PayPal".to_string());
+
+    let mut reference: Option<Shape> = None;
+    for tool in canonical_corpus {
+        let version = version_of(tool);
+        let Some(shape) = shape_for(tool, version, case) else {
+            panic!("missing {case} fixture for {tool}");
+        };
+        assert_eq!(
+            shape.chosen, chosen,
+            "{tool} {case}: chosen option set diverged from the shared corpus"
+        );
+        assert_eq!(
+            shape.rejected, rejected,
+            "{tool} {case}: rejected option set diverged from the shared corpus"
+        );
+        assert!(!shape.is_ban, "{tool} {case}: a rejection is not a ban");
+        match &reference {
+            None => reference = Some(shape),
+            Some(r) => assert_eq!(
+                &shape, r,
+                "{tool} {case}: full shape diverged from the shared corpus"
+            ),
+        }
+    }
+}
+
+/// Across *all* tools, a `rejected_alternative` fixture must never invent an
+/// option set without an originating decision: any chosen/rejected option
+/// implies at least one decision node. This is the contract that keeps the
+/// option lists `Observed`, never guessed.
+#[test]
+fn rejected_alternative_options_imply_a_decision() {
+    let case = "rejected_alternative";
+    for &(tool, version) in TOOLS {
+        let Some(nodes) = nodes_for(tool, version, case) else {
+            panic!("missing {case} fixture for {tool}");
+        };
+        let shape = Shape::of(&nodes);
+        if !shape.chosen.is_empty() || !shape.rejected.is_empty() {
+            assert!(
+                shape.decisions >= 1,
+                "{tool} {case}: produced options with no decision node — options must \
+                 derive from an observed decision, never be fabricated"
+            );
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// tool_failure — the scenario invariant: a rejected edit (ToolResult.ok = false)
+// must NOT produce a spurious Episode, and therefore no binding either. This now
+// holds *uniformly* for all nine tools: every adapter links a failed edit's
+// `FileEdit` to its failing `ToolResult` by `call_id`, so the segmenter drops it.
+// ---------------------------------------------------------------------------
+
+#[test]
+fn tool_failure_yields_no_spurious_episode() {
+    let case = "tool_failure";
+
+    let mut seen = 0;
+    for &(tool, version) in TOOLS {
+        let Some(shape) = shape_for(tool, version, case) else {
+            panic!("missing {case} fixture for {tool}");
+        };
+        assert_eq!(
+            shape.episodes, 0,
+            "{tool} {case}: a failed edit minted a spurious Episode — the adapter \
+             must link the FileEdit to its failing ToolResult (ok=false) by call_id \
+             so the segmenter drops it"
+        );
+        assert_eq!(
+            shape.bindings, 0,
+            "{tool} {case}: no episode but a binding survived — bindings must not \
+             outlive their episode"
+        );
+        seen += 1;
+    }
+    assert_eq!(
+        seen,
+        TOOLS.len(),
+        "every tool must carry the {case} fixture"
+    );
+}
+
+/// Look up a tool's fixture version slug from the driving table.
+fn version_of(tool: SourceKind) -> &'static str {
+    TOOLS
+        .iter()
+        .find(|(t, _)| *t == tool)
+        .map(|(_, v)| *v)
+        .unwrap_or_else(|| panic!("no version registered for {tool}"))
+}
+
+// ===========================================================================
+// The five additional §8.2 scenarios, asserted for the three first-class CLIs
+// (claude_code, codex, gemini). Where a tool's transcript format only carries an
+// *analog* of the scenario (e.g. a context-compaction notice with no
+// machine-resolvable replaced-range, or a subagent modeled as a nested thread on
+// one session id), we assert the invariant that *genuinely* holds for that tool
+// and PIN the divergence in an explicit assertion + comment — never silently
+// weakened to "anything goes".
+// ===========================================================================
+
+// ---------------------------------------------------------------------------
+// interleaved_arcs — two decisions, edits to overlapping files. The invariant:
+// *each edit binds to its own decision*, i.e. the most-recent decision that
+// precedes it in time, so the two arcs do not collapse into one. Concretely:
+//   - at least two distinct decisions each govern at least one edit;
+//   - the very first edit binds to the FIRST decision (the earlier arc), and the
+//     final edit binds to the SECOND decision (the later arc);
+//   - no binding points at a decision that does not exist in the stream.
+// ---------------------------------------------------------------------------
+
+#[test]
+fn interleaved_arcs_each_edit_binds_to_its_own_decision() {
+    let case = "interleaved_arcs";
+    let mut seen = 0;
+    for &(tool, version) in THREE_CLIS {
+        let Some(nodes) = nodes_for(tool, version, case) else {
+            panic!("missing {case} fixture for {tool}");
+        };
+        let decisions = decisions_of(&nodes);
+        let bindings = bindings_of(&nodes);
+
+        assert!(
+            decisions.len() >= 2,
+            "{tool} {case}: expected at least two decisions (two arcs), got {}",
+            decisions.len()
+        );
+        assert!(
+            bindings.len() >= 2,
+            "{tool} {case}: expected at least two bindings across the two arcs, got {}",
+            bindings.len()
+        );
+
+        // Every binding's source must be an observed decision, never fabricated:
+        // a binding's PROV `used_decision` must equal its `from`.
+        for b in &bindings {
+            assert_eq!(
+                b.prov.used_decision.as_ref(),
+                Some(&b.from),
+                "{tool} {case}: a binding's PROV used_decision must equal its `from` \
+                 (the governing decision), never a fabricated source"
+            );
+        }
+
+        // The set of distinct governing decisions must be > 1: the arcs are
+        // genuinely interleaved, not all folded onto a single decision.
+        let governing: BTreeSet<&str> = bindings.iter().map(|b| b.from.as_str()).collect();
+        assert!(
+            governing.len() >= 2,
+            "{tool} {case}: every edit collapsed onto a single decision — the two \
+             arcs were not bound independently (governing = {governing:?})"
+        );
+
+        // The first edit binds to the earliest arc and the last edit to the
+        // latest arc: ordering is by `t_gen` (the edit time), then by `from`.
+        let mut by_time = bindings.clone();
+        by_time.sort_by(|a, b| {
+            a.prov
+                .t_gen
+                .cmp(&b.prov.t_gen)
+                .then_with(|| a.from.as_str().cmp(b.from.as_str()))
+        });
+        let first_from = by_time.first().unwrap().from.as_str();
+        let last_from = by_time.last().unwrap().from.as_str();
+        assert_ne!(
+            first_from, last_from,
+            "{tool} {case}: the earliest and latest edits must bind to *different* \
+             decisions (the interleaving invariant)"
+        );
+        // And `t_use <= t_gen` must hold for every arc.
+        for b in &bindings {
+            assert!(
+                b.prov.is_temporally_valid(),
+                "{tool} {case}: a binding violated t_use <= t_gen"
+            );
+        }
+        seen += 1;
+    }
+    assert_eq!(seen, THREE_CLIS.len());
+}
+
+// ---------------------------------------------------------------------------
+// multi_edit_single_commit — one decision, N file edits. The invariant:
+// 1 Decision / N Episodes / N Bindings, with every binding sourced from that one
+// decision. This is the canonical "one commitment fans out to many files" shape.
+// ---------------------------------------------------------------------------
+
+#[test]
+fn multi_edit_single_commit_one_decision_n_episodes_n_bindings() {
+    let case = "multi_edit_single_commit";
+    let mut seen = 0;
+    for &(tool, version) in THREE_CLIS {
+        let Some(nodes) = nodes_for(tool, version, case) else {
+            panic!("missing {case} fixture for {tool}");
+        };
+        let shape = Shape::of(&nodes);
+
+        assert_eq!(
+            shape.decisions, 1,
+            "{tool} {case}: a single-commit fan-out must elevate exactly one decision"
+        );
+        assert!(
+            shape.episodes >= 2,
+            "{tool} {case}: expected several episodes (one per edited file), got {}",
+            shape.episodes
+        );
+        assert_eq!(
+            shape.bindings, shape.episodes,
+            "{tool} {case}: N episodes must produce N bindings (1 decision → N edits)"
+        );
+        assert!(
+            !shape.is_ban,
+            "{tool} {case}: a fan-out commit is not a ban"
+        );
+
+        // All bindings share the one decision as their source.
+        let bindings = bindings_of(&nodes);
+        let governing: BTreeSet<&str> = bindings.iter().map(|b| b.from.as_str()).collect();
+        assert_eq!(
+            governing.len(),
+            1,
+            "{tool} {case}: every edit must bind to the *same* single decision, \
+             found {} distinct sources",
+            governing.len()
+        );
+        seen += 1;
+    }
+    assert_eq!(seen, THREE_CLIS.len());
+}
+
+// ---------------------------------------------------------------------------
+// rewind_compaction — a decision is made, then rewound/compacted away, then a
+// replacement decision is made and the edit lands.
+//
+// The full structural invariant (pre-rewind decision is *superseded* —
+// `superseded_by = Some` — and does NOT bind; verbatim history preserved; the
+// final edit binds to the post-rewind decision) requires the adapter to surface
+// the rewind/compaction as a typed `Rewind { to_event }` / `Compaction
+// { replaced }` event that the segmenter can resolve to a turn-seq region.
+//
+//   * gemini's `$rewindTo` carries a resolvable target event id → the segmenter
+//     supersedes the rewound decision. We assert the FULL invariant here.
+//
+//   * claude_code's `summary` line and codex's `compacted`
+//     (`replaced_response_ids` = opaque string ids) are context-compaction
+//     *notices* with no machine-resolvable replaced-range, so both adapters route
+//     them to `Unknown` and NOTHING is superseded. This is a real, pinned
+//     divergence. The invariant that still genuinely holds for these two: the
+//     verbatim conversation of every pivot turn is preserved, and the FINAL edit
+//     binds to the LATEST (post-pivot) decision, never to the stale earlier one —
+//     so the current view honors the pivot even without a structural supersede
+//     marker. We assert exactly that and PIN the absence of the marker.
+// ---------------------------------------------------------------------------
+
+#[test]
+fn rewind_compaction_gemini_supersedes_pre_rewind_decision() {
+    // Gemini is the one CLI whose format carries a resolvable rewind target, so
+    // it is the tool that exercises the full supersede-and-skip invariant.
+    let nodes = nodes_for(SourceKind::Gemini, "v1", "rewind_compaction")
+        .expect("gemini rewind_compaction fixture present");
+    let decisions = decisions_of(&nodes);
+    let bindings = bindings_of(&nodes);
+
+    // Exactly one decision is superseded (the pre-rewind MongoDB choice), and at
+    // least one survives (the post-rewind Postgres choice).
+    let superseded: Vec<&&DecisionRecord> = decisions
+        .iter()
+        .filter(|d| d.superseded_by.is_some())
+        .collect();
+    assert_eq!(
+        superseded.len(),
+        1,
+        "gemini rewind_compaction: exactly one (pre-rewind) decision must be superseded"
+    );
+    let marker = superseded[0].superseded_by.as_ref().unwrap().as_str();
+    assert!(
+        marker.starts_with("rewind:"),
+        "gemini rewind_compaction: the supersede marker must be a rewind marker, got {marker:?}"
+    );
+
+    // The superseded decision must NOT govern any edit: the binder skips
+    // superseded decisions and falls through to the survivor. We derive the
+    // superseded turn-seq (the `decision:<session>:<seq>` suffix) and assert no
+    // binding sources from it.
+    let superseded_seqs: BTreeSet<u64> = decisions
+        .iter()
+        .filter(|d| d.superseded_by.is_some())
+        .map(|d| d.source_span.start)
+        .collect();
+    assert!(
+        !superseded_seqs.is_empty(),
+        "gemini rewind_compaction: expected at least one superseded turn-seq"
+    );
+    for b in &bindings {
+        let from_is_superseded = superseded_seqs
+            .iter()
+            .any(|seq| b.from.as_str().ends_with(&format!(":{seq}")));
+        assert!(
+            !from_is_superseded,
+            "gemini rewind_compaction: binding {} sources a superseded decision",
+            b.from
+        );
+    }
+
+    // Verbatim history is preserved: a conversation span exists for BOTH the
+    // rewound turn and the surviving turn (losslessness across the rewind).
+    let convo_count = nodes
+        .iter()
+        .filter(|n| matches!(n, PreparedNode::Conversation(_)))
+        .count();
+    assert!(
+        convo_count >= 2,
+        "gemini rewind_compaction: verbatim spans for both the rewound and the \
+         surviving turn must be preserved, found {convo_count}"
+    );
+}
+
+#[test]
+fn rewind_compaction_claude_and_codex_preserve_history_and_honor_the_pivot() {
+    // PINNED DIVERGENCE: neither claude_code's `summary` nor codex's `compacted`
+    // notice resolves to a typed Rewind/Compaction, so NO decision is superseded.
+    // We assert the invariants that genuinely hold and explicitly pin the absence
+    // of the structural supersede marker.
+    for tool in [SourceKind::ClaudeCode, SourceKind::Codex] {
+        let version = version_of(tool);
+        let nodes = nodes_for(tool, version, "rewind_compaction")
+            .unwrap_or_else(|| panic!("{tool} rewind_compaction fixture present"));
+
+        // Pinned: the compaction notice did not mint a supersede marker.
+        assert!(
+            !has_superseded_decision(&nodes),
+            "{tool} rewind_compaction: this adapter is expected to route its \
+             context-compaction notice to Unknown (no resolvable replaced-range), \
+             so NO decision should be superseded. If this fires, the adapter began \
+             emitting a typed Rewind/Compaction and this pin must be revisited."
+        );
+
+        // Verbatim history preserved: every decision turn kept its conversation
+        // span (at least the pre- and post-pivot decisions are both present).
+        let convos = nodes
+            .iter()
+            .filter(|n| matches!(n, PreparedNode::Conversation(_)))
+            .count();
+        let decisions = decisions_of(&nodes);
+        assert!(
+            convos >= decisions.len() && decisions.len() >= 2,
+            "{tool} rewind_compaction: both pivot turns must be retained verbatim \
+             (convos={convos}, decisions={})",
+            decisions.len()
+        );
+
+        // The current view honors the pivot: the FINAL edit binds to the LATEST
+        // decision (largest source-turn seq), never to the stale earlier one.
+        let bindings = bindings_of(&nodes);
+        assert!(
+            !bindings.is_empty(),
+            "{tool} rewind_compaction: the post-pivot edit must still bind"
+        );
+        let latest_decision_seq = decisions
+            .iter()
+            .map(|d| d.source_span.start)
+            .max()
+            .expect("at least one decision");
+        let final_binding = bindings
+            .iter()
+            .max_by_key(|b| b.prov.t_gen)
+            .expect("at least one binding");
+        assert!(
+            final_binding
+                .from
+                .as_str()
+                .ends_with(&format!(":{latest_decision_seq}")),
+            "{tool} rewind_compaction: the final edit must bind to the latest \
+             (post-pivot) decision :{latest_decision_seq}, got {}",
+            final_binding.from
+        );
+    }
+}
+
+// ---------------------------------------------------------------------------
+// subagent_thread — work is delegated to a subagent / nested thread.
+//
+// PINNED DIVERGENCE: the documented goal is "subagent nodes carry the *distinct*
+// session id, not merged". In practice none of the three CLIs surface a separate
+// `CaptureEvent.session_id` for the delegated work:
+//   * claude_code MERGES the `isSidechain:true` session (`sess-subagent-008b`)
+//     into the parent `sess-main-008`;
+//   * codex carries the `thread_id` only inside an Unknown `turn_context` record,
+//     not as a session;
+//   * gemini drops the nested `threadId` entirely.
+// So at the node layer the subagent work is co-attributed to ONE session id.
+//
+// The invariant that genuinely holds (and that we assert): the subagent's edit
+// is attributed to the same single session as its *own* governing decision and
+// binds to THAT decision (the in-thread commitment), not to the dispatch turn or
+// to a foreign session. The delegated work is captured, never dropped or
+// cross-attributed. We pin the single-session-merge explicitly.
+// ---------------------------------------------------------------------------
+
+#[test]
+fn subagent_thread_is_captured_and_bound_within_one_session() {
+    let case = "subagent_thread";
+    let mut seen = 0;
+    for &(tool, version) in THREE_CLIS {
+        let Some(nodes) = nodes_for(tool, version, case) else {
+            panic!("missing {case} fixture for {tool}");
+        };
+
+        // The subagent's edit is captured (not dropped) and binds.
+        let episodes = nodes
+            .iter()
+            .filter(|n| matches!(n, PreparedNode::Episode(_)))
+            .count();
+        assert!(
+            episodes >= 1,
+            "{tool} {case}: the subagent's edit must be captured as an episode"
+        );
+        let bindings = bindings_of(&nodes);
+        assert!(
+            !bindings.is_empty(),
+            "{tool} {case}: the subagent edit must bind to its governing decision"
+        );
+
+        // PINNED: every node carries a single, shared session id (subagent work is
+        // merged into one session, not split). Drawn from event-layer session ids.
+        let events = events_for(tool, version, case).expect("events parse");
+        let sessions: BTreeSet<&str> = events.iter().map(|e| e.session_id.as_str()).collect();
+        assert_eq!(
+            sessions.len(),
+            1,
+            "{tool} {case}: expected a single normalized session id (subagent merged \
+             into the parent). If this fires, an adapter began emitting a distinct \
+             subagent session and this pin should become the stronger 'distinct \
+             session' assertion (sessions = {sessions:?})"
+        );
+        let the_session = *sessions.iter().next().unwrap();
+
+        // The binding stays within that one session, and its source is the
+        // subagent's own decision (PROV used_decision == from), never fabricated.
+        for b in &bindings {
+            assert_eq!(
+                b.prov.used_session, the_session,
+                "{tool} {case}: binding used_session must be the single shared session"
+            );
+            assert_eq!(
+                b.prov.was_generated_by_session, the_session,
+                "{tool} {case}: episode session must be the single shared session"
+            );
+            assert_eq!(
+                b.prov.used_decision.as_ref(),
+                Some(&b.from),
+                "{tool} {case}: the subagent edit must bind to an observed decision"
+            );
+            assert!(
+                b.prov.is_temporally_valid(),
+                "{tool} {case}: subagent binding violated t_use <= t_gen"
+            );
+        }
+        seen += 1;
+    }
+    assert_eq!(seen, THREE_CLIS.len());
+}
+
+// ---------------------------------------------------------------------------
+// no_commitment_marker — a turn with no commitment marker, followed by an edit.
+// The invariant: 0 Conversation nodes and 0 Decision nodes are elevated (the gate
+// did not fire), but the edit episode is STILL present, and the verbatim user
+// turn is still retained at the event layer (losslessness below the gate).
+// ---------------------------------------------------------------------------
+
+#[test]
+fn no_commitment_marker_elevates_nothing_but_keeps_the_edit() {
+    let case = "no_commitment_marker";
+    let mut seen = 0;
+    for &(tool, version) in THREE_CLIS {
+        let Some(nodes) = nodes_for(tool, version, case) else {
+            panic!("missing {case} fixture for {tool}");
+        };
+        let shape = Shape::of(&nodes);
+
+        assert_eq!(
+            shape.conversations, 0,
+            "{tool} {case}: an unmarked turn must not elevate a Conversation node"
+        );
+        assert_eq!(
+            shape.decisions, 0,
+            "{tool} {case}: an unmarked turn must not elevate a Decision node"
+        );
+        assert!(
+            shape.episodes >= 1,
+            "{tool} {case}: the edit episode must still be present despite no marker"
+        );
+        // With no decision to govern it, the lone edit produces no binding.
+        assert_eq!(
+            shape.bindings, 0,
+            "{tool} {case}: an unbound edit must not fabricate a binding"
+        );
+
+        // Losslessness below the gate: the verbatim user turn is still retained as
+        // a normalized event even though it produced no node.
+        let events = events_for(tool, version, case).expect("events parse");
+        let user_turns = events
+            .iter()
+            .filter(|e| matches!(e.kind, memscribe_core::EventKind::UserTurn { .. }))
+            .count();
+        assert!(
+            user_turns >= 1,
+            "{tool} {case}: the unmarked user turn must still be retained verbatim \
+             at the event layer (lossless capture below the gate)"
+        );
+        seen += 1;
+    }
+    assert_eq!(seen, THREE_CLIS.len());
+}
+
+// ---------------------------------------------------------------------------
+// Determinism guard: prepared-node output is a pure function of the input bytes.
+// Re-running every fixture twice must yield byte-identical node streams. This is
+// the property the golden snapshots silently depend on.
+// ---------------------------------------------------------------------------
+
+#[test]
+fn prepared_nodes_are_deterministic_across_runs() {
+    for &(tool, version) in TOOLS {
+        for case in [
+            "happy_path_decision_then_edits",
+            "rejected_alternative",
+            "ban",
+            "tool_failure",
+            // The five additional §8.2 scenarios (present only for the three CLIs;
+            // absent fixtures are skipped via the `continue` below).
+            "interleaved_arcs",
+            "multi_edit_single_commit",
+            "rewind_compaction",
+            "subagent_thread",
+            "no_commitment_marker",
+        ] {
+            let Some(first) = nodes_for(tool, version, case) else {
+                continue;
+            };
+            let second = nodes_for(tool, version, case).expect("fixture read twice");
+            assert_eq!(
+                first, second,
+                "{tool} {case}: prepared-node output is not deterministic across runs"
+            );
+        }
+    }
+}
diff --git a/crates/memscribe-testkit/tests/golden.rs b/crates/memscribe-testkit/tests/golden.rs
new file mode 100644
index 0000000..e874fd1
--- /dev/null
+++ b/crates/memscribe-testkit/tests/golden.rs
@@ -0,0 +1,102 @@
+//! Golden snapshots for every fixture (whitepaper §8.1).
+//!
+//! For each `fixtures/<tool>/<version>/<case>.jsonl` we snapshot **two** things
+//! with `insta`:
+//!
+//! 1. the normalized [`CaptureEvent`] stream (`testkit::parse_events`), and
+//! 2. the prepared [`PreparedNode`] stream (`testkit::prepare_nodes`).
+//!
+//! The committed `.snap` files under `tests/snapshots/` are the golden record. A
+//! future diff is then unambiguous: either a real regression in an adapter / the
+//! pipeline, or an *intended* format change that the author re-accepts with
+//! `cargo insta accept` (or `INSTA_UPDATE=always`). Because the pipeline is a
+//! pure function of the input bytes and the *relative* fixture path we feed it,
+//! these snapshots are byte-stable and machine-independent.
+//!
+//! First run (writes the snapshots):
+//! ```text
+//! INSTA_UPDATE=always cargo test -p memscribe-testkit --test golden
+//! ```
+//! Then re-run without the env var to prove they are stable.
+
+use insta::assert_json_snapshot;
+use memscribe_core::SourceKind;
+use memscribe_testkit::golden::{discover_cases, fixtures_dir, GoldenCase};
+use memscribe_testkit::{parse_events, prepare_nodes};
+use std::path::{Path, PathBuf};
+
+/// The stable, machine-independent path we feed the pipeline so provenance and
+/// any path-derived ids are identical on every machine.
+fn relative_input_path(case: &GoldenCase) -> PathBuf {
+    Path::new("fixtures")
+        .join(&case.tool)
+        .join(&case.version)
+        .join(format!("{}.jsonl", case.case))
+}
+
+/// Resolve the tool slug to a [`SourceKind`], skipping any fixture directory
+/// that is not a known tool (so a stray directory never fails the suite).
+fn source_kind(case: &GoldenCase) -> Option<SourceKind> {
+    SourceKind::parse(&case.tool).filter(|k| *k != SourceKind::Unknown)
+}
+
+/// A stable per-fixture snapshot name. Slashes/dots in the slugs are normalized
+/// so insta writes one `.snap` file per fixture under `tests/snapshots/`.
+fn snap_name(prefix: &str, case: &GoldenCase) -> String {
+    let sanitize = |s: &str| s.replace(['.', '/', '\\', ' '], "_");
+    format!(
+        "{prefix}__{}__{}__{}",
+        sanitize(&case.tool),
+        sanitize(&case.version),
+        sanitize(&case.case),
+    )
+}
+
+#[test]
+fn golden_snapshots_for_every_fixture() {
+    // Pin the snapshot directory and strip insta's auto-prepended module path so
+    // the file names are exactly `snapshots/<prefix>__<tool>__<version>__<case>.snap`.
+    let mut settings = insta::Settings::clone_current();
+    settings.set_snapshot_path("snapshots");
+    settings.set_prepend_module_to_snapshot(false);
+
+    let cases = discover_cases();
+    assert!(
+        !cases.is_empty(),
+        "no fixtures discovered under {}",
+        fixtures_dir().display()
+    );
+
+    let mut snapped = 0;
+    settings.bind(|| {
+        for case in &cases {
+            let Some(kind) = source_kind(case) else {
+                panic!(
+                    "fixture tool slug {:?} does not resolve to a known SourceKind",
+                    case.tool
+                );
+            };
+
+            let bytes = case
+                .read_input()
+                .unwrap_or_else(|e| panic!("read fixture {:?}: {e}", case.input_path()));
+            let rel = relative_input_path(case);
+
+            // 1. Normalized events.
+            let events = parse_events(kind, &bytes, &rel);
+            assert_json_snapshot!(snap_name("events", case), events);
+
+            // 2. Prepared nodes (redaction off, so content is verbatim).
+            let nodes = prepare_nodes(kind, &bytes, &rel);
+            assert_json_snapshot!(snap_name("nodes", case), nodes);
+
+            snapped += 1;
+        }
+    });
+
+    assert_eq!(
+        snapped,
+        cases.len(),
+        "every discovered fixture must be snapshotted"
+    );
+}
diff --git a/crates/memscribe-testkit/tests/properties.proptest-regressions b/crates/memscribe-testkit/tests/properties.proptest-regressions
new file mode 100644
index 0000000..9f07df5
--- /dev/null
+++ b/crates/memscribe-testkit/tests/properties.proptest-regressions
@@ -0,0 +1,8 @@
+# Seeds for failure cases proptest has generated in the past. It is
+# automatically read and these particular cases re-run before any
+# novel cases are generated.
+#
+# It is recommended to check this file in to source control so that
+# everyone who runs the test benefits from these saved cases.
+cc 14175008d79450aed54071dcb2eb9366f3f52f7f959521ce014c59812311195a # shrinks to tool = Windsurf, bytes = [123, 34, 116, 121, 112, 101, 34, 58, 34, 117, 115, 101, 114, 34, 44, 34, 114, 111, 108, 101, 34, 58, 34, 115, 121, 115, 116, 101, 109, 34, 44, 34, 116, 101, 120, 116, 34, 58, 34, 108, 101, 116, 39, 115, 32, 103, 111, 32, 119, 105, 116, 104, 32, 80, 111, 115, 116, 103, 114, 101, 115, 32, 102, 111, 114, 32, 115, 116, 111, 114, 97, 103, 101, 34, 44, 34, 109, 101, 115, 115, 97, 103, 101, 34, 58, 123, 34, 114, 111, 108, 101, 34, 58, 34, 117, 115, 101, 114, 34, 44, 34, 99, 111, 110, 116, 101, 110, 116, 34, 58, 34, 108, 101, 116, 39, 115, 32, 103, 111, 32, 119, 105, 116, 104, 32, 80, 111, 115, 116, 103, 114, 101, 115, 32, 102, 111, 114, 32, 115, 116, 111, 114, 97, 103, 101, 34, 125, 44, 34, 116, 105, 109, 101, 115, 116, 97, 109, 112, 34, 58, 34, 50, 48, 50, 54, 45, 48, 54, 45, 50, 50, 84, 49, 50, 58, 48, 48, 58, 48, 48, 90, 34, 125]
+cc be690137d2d29796150116664535b12a0f68636bb99c9bd079f42f65af08b58f # shrinks to tool = Gemini, bytes = [123, 34, 116, 121, 112, 101, 34, 58, 34, 116, 111, 111, 108, 95, 117, 115, 101, 34, 44, 34, 114, 111, 108, 101, 34, 58, 34, 115, 121, 115, 116, 101, 109, 34, 44, 34, 116, 101, 120, 116, 34, 58, 34, 119, 101, 32, 119, 105, 108, 108, 32, 110, 101, 118, 101, 114, 32, 97, 100, 100, 32, 97, 32, 100, 101, 112, 101, 110, 100, 101, 110, 99, 121, 32, 111, 110, 32, 108, 101, 102, 116, 45, 112, 97, 100, 34, 44, 34, 109, 101, 115, 115, 97, 103, 101, 34, 58, 123, 34, 114, 111, 108, 101, 34, 58, 34, 117, 115, 101, 114, 34, 44, 34, 99, 111, 110, 116, 101, 110, 116, 34, 58, 34, 119, 101, 32, 119, 105, 108, 108, 32, 110, 101, 118, 101, 114, 32, 97, 100, 100, 32, 97, 32, 100, 101, 112, 101, 110, 100, 101, 110, 99, 121, 32, 111, 110, 32, 108, 101, 102, 116, 45, 112, 97, 100, 34, 125, 44, 34, 116, 105, 109, 101, 115, 116, 97, 109, 112, 34, 58, 34, 50, 48, 50, 54, 45, 48, 54, 45, 50, 50, 84, 49, 50, 58, 48, 48, 58, 48, 48, 90, 34, 125, 10, 123, 34, 116, 121, 112, 101, 34, 58, 34, 116, 111, 111, 108, 95, 117, 115, 101, 34, 44, 34, 114, 111, 108, 101, 34, 58, 34, 115, 121, 115, 116, 101, 109, 34, 44, 34, 116, 101, 120, 116, 34, 58, 34, 119, 101, 32, 119, 105, 108, 108, 32, 110, 101, 118, 101, 114, 32, 97, 100, 100, 32, 97, 32, 100, 101, 112, 101, 110, 100, 101, 110, 99, 121, 32, 111, 110, 32, 108, 101, 102, 116, 45, 112, 97, 100, 34, 44, 34, 109, 101, 115, 115, 97, 103, 101, 34, 58, 123, 34, 114, 111, 108, 101, 34, 58, 34, 117, 115, 101, 114, 34, 44, 34, 99, 111, 110, 116, 101, 110, 116, 34, 58, 34, 119, 101, 32, 119, 105, 108, 108, 32, 110, 101, 118, 101, 114, 32, 97, 100, 100, 32, 97, 32, 100, 101, 112, 101, 110, 100, 101, 110, 99, 121, 32, 111, 110, 32, 108, 101, 102, 116, 45, 112, 97, 100, 34, 125, 44, 34, 116, 105, 109, 101, 115, 116, 97, 109, 112, 34, 58, 34, 50, 48, 50, 54, 45, 48, 54, 45, 50, 50, 84, 49, 50, 58, 48, 48, 58, 48, 48, 90, 34, 125]
diff --git a/crates/memscribe-testkit/tests/properties.rs b/crates/memscribe-testkit/tests/properties.rs
new file mode 100644
index 0000000..69b30bf
--- /dev/null
+++ b/crates/memscribe-testkit/tests/properties.rs
@@ -0,0 +1,369 @@
+//! Whitepaper §8.3 invariants, expressed as `proptest` properties.
+//!
+//! These complement the per-stage unit tests and the golden-file suite: instead
+//! of fixed inputs they assert that the §8.3 invariants hold for *arbitrary*
+//! generated input — both plausible JSONL records and raw arbitrary bytes.
+//!
+//! The invariants checked here:
+//!
+//! 1. **Determinism** — `parse` / `prepare` are byte-identical across runs and
+//!    depend only on their input (pure functions, thread-independent).
+//! 2. **Idempotency by `event_id`** — concatenating a *uniquely-identified*
+//!    record stream with itself through one parse pass dedups back to the
+//!    single-pass event set (adapters dedup recognized records via
+//!    `ctx.first_seen`).
+//! 3. **Monotonic `seq`** — strictly increasing per session, for any stream.
+//! 4. **Losslessness** — every distinct non-blank record yields `>= 1` event.
+//! 5. **Gate purity** — `CommitmentGate::evaluate` depends only on the text, is
+//!    repeatable, and is independent of evaluation order/context.
+//! 6. **Offset resumption** — splitting a buffer at a newline boundary and
+//!    concatenating the two reads equals reading the whole, at the byte level.
+//!
+//! A note on the dedup/losslessness contract (ground truth in the adapters):
+//! recognized records are deduplicated once on their native id (via
+//! `ctx.first_seen`), while *unrecognized* records are routed to
+//! `EventKind::Unknown` and preserved verbatim **without** dedup — losslessness
+//! for unknown data outranks idempotency. Two records with identical bytes
+//! therefore collapse to one event when recognized (same content-hash fallback
+//! id) but are both retained when unknown. To test idempotency and losslessness
+//! against a well-defined record identity, the generators below stamp a unique
+//! native id on every record, so distinct records never collide.
+
+use memscribe_adapters::adapter_for;
+use memscribe_core::{pipeline::parse_records, CaptureEvent, CommitmentGate, SourceKind};
+use memscribe_io::read_records_from_bytes;
+use memscribe_testkit::golden::{discover_cases, GoldenCase};
+use memscribe_testkit::invariants::{
+    check_determinism, check_lossless, check_monotonic_seq, check_unique_event_ids,
+};
+use memscribe_testkit::{count_nonblank_lines, parse_events};
+use proptest::prelude::*;
+use std::path::Path;
+
+// ---------------------------------------------------------------------------
+// Strategies: plausible JSONL records and raw arbitrary bytes.
+// ---------------------------------------------------------------------------
+
+/// Plausible turn text, including the commitment-marker vocabulary so the gate
+/// and segmenter get exercised, plus arbitrary free text.
+fn turn_text() -> impl Strategy<Value = String> {
+    prop_oneof![
+        Just("let's go with Postgres for storage".to_string()),
+        Just("use Stripe instead of PayPal".to_string()),
+        Just("we will never add a dependency on left-pad".to_string()),
+        Just("we must always use prepared statements".to_string()),
+        Just("remember that the cache TTL is 60s".to_string()),
+        Just("thanks, that looks good".to_string()),
+        // Arbitrary printable-ish text, to stress the gate/segmenter spans.
+        "[a-zA-Z0-9 ,.!?'\\-]{0,80}".prop_map(|s| s),
+    ]
+}
+
+/// A Claude Code record shape (`type`/`uuid`/`message`). Claude Code is the
+/// reference adapter for these properties because it recognizes the shape (so
+/// records are not all Unknown) and deduplicates once on `uuid` — giving every
+/// record a well-defined, stable identity. The `idx` makes the `uuid` unique so
+/// distinct records never collide on the content-hash fallback.
+fn claude_record(idx: usize) -> impl Strategy<Value = serde_json::Value> {
+    (prop_oneof![Just("user"), Just("assistant")], turn_text()).prop_map(move |(role, text)| {
+        serde_json::json!({
+            "type": role,
+            "uuid": format!("evt-{idx}"),
+            "timestamp": "2026-06-22T12:00:00.000Z",
+            "message": { "role": role, "content": text },
+        })
+    })
+}
+
+/// A JSONL document of uniquely-identified Claude Code records, newline-joined.
+/// May or may not end with a trailing newline (both are valid reader inputs).
+fn claude_document() -> impl Strategy<Value = Vec<u8>> {
+    (0usize..12)
+        .prop_flat_map(|n| {
+            let recs: Vec<_> = (0..n).map(claude_record).collect();
+            (recs, any::<bool>())
+        })
+        .prop_map(|(records, trailing_nl)| {
+            let mut doc = records
+                .iter()
+                .map(serde_json::Value::to_string)
+                .collect::<Vec<_>>()
+                .join("\n");
+            if trailing_nl && !doc.is_empty() {
+                doc.push('\n');
+            }
+            doc.into_bytes()
+        })
+}
+
+/// Raw arbitrary bytes — including invalid UTF-8 and random newlines — so the
+/// reader and adapters are stressed on input that is *not* well-formed JSONL.
+fn arbitrary_bytes() -> impl Strategy<Value = Vec<u8>> {
+    proptest::collection::vec(any::<u8>(), 0..256)
+}
+
+/// The tools whose adapters are compiled in. Driving generated bytes through
+/// each one widens coverage of the normalization contract.
+fn any_tool() -> impl Strategy<Value = SourceKind> {
+    prop_oneof![
+        Just(SourceKind::ClaudeCode),
+        Just(SourceKind::Codex),
+        Just(SourceKind::Gemini),
+        Just(SourceKind::Otel),
+        Just(SourceKind::Cursor),
+        Just(SourceKind::Windsurf),
+        Just(SourceKind::Zed),
+        Just(SourceKind::VsCode),
+        Just(SourceKind::Copilot),
+    ]
+}
+
+// ---------------------------------------------------------------------------
+// Properties.
+// ---------------------------------------------------------------------------
+
+proptest! {
+    #![proptest_config(ProptestConfig::with_cases(256))]
+
+    /// (1) Determinism: parsing the same generated JSONL twice is byte-identical
+    /// (serialized), and the result depends only on the input — re-parsing in a
+    /// freshly spawned thread yields the identical bytes (pure, thread-free).
+    #[test]
+    fn parse_is_deterministic_and_thread_independent(bytes in claude_document()) {
+        let path = Path::new("gen.jsonl");
+        let a = parse_events(SourceKind::ClaudeCode, &bytes, path);
+        let b = parse_events(SourceKind::ClaudeCode, &bytes, path);
+        check_determinism(&a, &b).map_err(TestCaseError::fail)?;
+
+        // Thread-independence: the same pure function, run on another thread,
+        // must produce the identical serialized output (no ambient/thread state).
+        let bytes_for_thread = bytes.clone();
+        let serialized_main = serde_json::to_string(&a).unwrap();
+        let serialized_thread = std::thread::spawn(move || {
+            let evts = parse_events(SourceKind::ClaudeCode, &bytes_for_thread, Path::new("gen.jsonl"));
+            serde_json::to_string(&evts).unwrap()
+        })
+        .join()
+        .expect("parse thread must not panic");
+        prop_assert_eq!(serialized_main, serialized_thread);
+    }
+
+    /// (1) Determinism over arbitrary (possibly non-UTF8) bytes, through every
+    /// adapter: the reader + adapter must still be a pure function and never
+    /// panic on malformed input.
+    #[test]
+    fn parse_is_deterministic_over_arbitrary_bytes(
+        tool in any_tool(),
+        bytes in arbitrary_bytes(),
+    ) {
+        let path = Path::new("gen.bin");
+        let a = parse_events(tool, &bytes, path);
+        let b = parse_events(tool, &bytes, path);
+        check_determinism(&a, &b).map_err(TestCaseError::fail)?;
+    }
+
+    /// (2) Idempotency by `event_id`: feeding a uniquely-identified record stream
+    /// concatenated with itself through a single parse pass (one shared
+    /// `ParseCtx`) dedups back to the single-pass event set. Adapters dedup
+    /// recognized records via `ctx.first_seen`, so the doubled stream must not
+    /// introduce duplicate `(session_id, event_id)` keys and must yield exactly
+    /// the single-pass dedup-key set.
+    #[test]
+    fn parse_is_idempotent_by_event_id(bytes in claude_document()) {
+        let path = Path::new("gen.jsonl");
+
+        let single_recs = read_records_from_bytes(&bytes, path);
+        let adapter = adapter_for(SourceKind::ClaudeCode).expect("adapter must be compiled");
+
+        let (single, _ctx) = parse_records(adapter.as_ref(), &single_recs);
+
+        // Build a doubled stream: the same records again, in one parse pass.
+        let mut doubled_recs = single_recs.clone();
+        doubled_recs.extend(single_recs.iter().cloned());
+        let (doubled, _ctx2) = parse_records(adapter.as_ref(), &doubled_recs);
+
+        // No duplicate dedup keys were introduced by the second copy.
+        check_unique_event_ids(&doubled).map_err(TestCaseError::fail)?;
+
+        // The dedup-key SET is identical between the single and doubled passes.
+        let key_set = |evs: &[CaptureEvent]| -> std::collections::BTreeSet<(String, String)> {
+            evs.iter()
+                .map(|e| (e.session_id.clone(), e.event_id.clone()))
+                .collect()
+        };
+        prop_assert_eq!(key_set(&single), key_set(&doubled));
+    }
+
+    /// (3) Monotonic `seq`: holds for any generated record stream, per session.
+    /// Checked over every adapter so the property is not Claude-specific.
+    #[test]
+    fn seq_is_monotonic(tool in any_tool(), bytes in claude_document()) {
+        let events = parse_events(tool, &bytes, Path::new("gen.jsonl"));
+        check_monotonic_seq(&events).map_err(TestCaseError::fail)?;
+    }
+
+    /// (3) Monotonic `seq` over arbitrary bytes too: even a malformed stream that
+    /// produces a pile of Unknown events keeps `seq` strictly increasing.
+    #[test]
+    fn seq_is_monotonic_over_arbitrary_bytes(tool in any_tool(), bytes in arbitrary_bytes()) {
+        let events = parse_events(tool, &bytes, Path::new("gen.bin"));
+        check_monotonic_seq(&events).map_err(TestCaseError::fail)?;
+    }
+
+    /// (4) Losslessness: every distinct non-blank source record yields at least
+    /// one event. The generator stamps a unique id on each record, so no two
+    /// distinct records collapse via the content-hash dedup fallback; the event
+    /// count is therefore `>=` the non-blank record count.
+    #[test]
+    fn parse_is_lossless(bytes in claude_document()) {
+        let events = parse_events(SourceKind::ClaudeCode, &bytes, Path::new("gen.jsonl"));
+        let nonblank = count_nonblank_lines(&bytes);
+        check_lossless(nonblank, &events).map_err(TestCaseError::fail)?;
+    }
+
+    /// (5) Gate purity: `evaluate(s)` depends only on `s`. Two calls are equal,
+    /// and evaluating the same text on an independently-constructed gate (a
+    /// different "context") yields the identical markers — no order/context
+    /// dependence.
+    #[test]
+    fn gate_is_pure(text in turn_text()) {
+        let gate = CommitmentGate::default_table();
+        let a = gate.evaluate(&text);
+        let b = gate.evaluate(&text);
+        prop_assert_eq!(&a, &b);
+
+        // A second, independently constructed gate must agree exactly.
+        let other = CommitmentGate::default_table();
+        prop_assert_eq!(&a, &other.evaluate(&text));
+    }
+
+    /// (5) Gate purity under interleaving: evaluating two texts in either order
+    /// on the same gate yields per-text results identical to evaluating each in
+    /// isolation. This is the "depends only on s, not on call history" property.
+    #[test]
+    fn gate_is_order_independent(t1 in turn_text(), t2 in turn_text()) {
+        let gate = CommitmentGate::default_table();
+
+        let iso1 = gate.evaluate(&t1);
+        let iso2 = gate.evaluate(&t2);
+
+        // Forward order, then reverse order, on the same instance.
+        let f1 = gate.evaluate(&t1);
+        let f2 = gate.evaluate(&t2);
+        let r2 = gate.evaluate(&t2);
+        let r1 = gate.evaluate(&t1);
+
+        prop_assert_eq!(&iso1, &f1);
+        prop_assert_eq!(&iso2, &f2);
+        prop_assert_eq!(&iso1, &r1);
+        prop_assert_eq!(&iso2, &r2);
+    }
+
+    /// (5) Gate purity over arbitrary text: evaluate must be repeatable and
+    /// never panic on any string, including unusual unicode.
+    #[test]
+    fn gate_is_pure_over_arbitrary_text(text in ".{0,200}") {
+        let gate = CommitmentGate::default_table();
+        prop_assert_eq!(gate.evaluate(&text), gate.evaluate(&text));
+    }
+
+    /// (6) Offset resumption: for any split of the buffer at a newline boundary,
+    /// `read_records_from_bytes(prefix) ++ read_records_from_bytes(rest)` equals
+    /// `read_records_from_bytes(whole)` at the byte level. This is what lets a
+    /// live tailer resume from a persisted offset without losing or duplicating
+    /// records.
+    #[test]
+    fn offset_resumption_holds(bytes in arbitrary_bytes()) {
+        let path = Path::new("gen.bin");
+        let whole: Vec<Vec<u8>> = read_records_from_bytes(&bytes, path)
+            .into_iter()
+            .map(|r| r.bytes)
+            .collect();
+
+        // Newline boundaries are the only points a real tailer resumes at.
+        let mut boundaries = vec![0usize, bytes.len()];
+        for (i, b) in bytes.iter().enumerate() {
+            if *b == b'\n' {
+                boundaries.push(i + 1);
+            }
+        }
+        boundaries.sort_unstable();
+        boundaries.dedup();
+
+        for split in boundaries {
+            let mut combined: Vec<Vec<u8>> = read_records_from_bytes(&bytes[..split], path)
+                .into_iter()
+                .map(|r| r.bytes)
+                .collect();
+            combined.extend(
+                read_records_from_bytes(&bytes[split..], path)
+                    .into_iter()
+                    .map(|r| r.bytes),
+            );
+            prop_assert_eq!(&combined, &whole);
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Fixture-driven properties: the same invariants, but exercised on the real
+// committed fixtures (every tool / version / case). These anchor the proptest
+// strategies on inputs adapters fully recognize.
+// ---------------------------------------------------------------------------
+
+/// Resolve the `SourceKind` for a fixture's tool slug, skipping any case whose
+/// slug is not a known source (defensive — `discover_cases` walks the dir).
+fn tool_for_case(c: &GoldenCase) -> Option<SourceKind> {
+    SourceKind::parse(&c.tool)
+}
+
+/// Dedup-key set over a normalized event stream.
+fn key_set(evs: &[CaptureEvent]) -> std::collections::BTreeSet<(String, String)> {
+    evs.iter()
+        .map(|e| (e.session_id.clone(), e.event_id.clone()))
+        .collect()
+}
+
+#[test]
+fn fixtures_satisfy_section_8_3_invariants() {
+    let cases = discover_cases();
+    assert!(
+        !cases.is_empty(),
+        "expected committed fixtures under fixtures/"
+    );
+
+    for case in &cases {
+        let Some(tool) = tool_for_case(case) else {
+            continue;
+        };
+        let bytes = case.read_input().expect("fixture readable");
+        let path = case.input_path();
+        let label = format!("{}/{}/{}", case.tool, case.version, case.case);
+
+        // Determinism.
+        let a = parse_events(tool, &bytes, &path);
+        let b = parse_events(tool, &bytes, &path);
+        check_determinism(&a, &b).unwrap_or_else(|e| panic!("{label}: {e}"));
+
+        // Monotonic seq.
+        check_monotonic_seq(&a).unwrap_or_else(|e| panic!("{label}: {e}"));
+
+        // Losslessness.
+        let nonblank = count_nonblank_lines(&bytes);
+        check_lossless(nonblank, &a).unwrap_or_else(|e| panic!("{label}: {e}"));
+
+        // A single normal parse pass introduces no duplicate dedup keys: every
+        // recognized record is deduped on its native id, and the per-record
+        // content-hash fallback is unique within a real transcript.
+        check_unique_event_ids(&a).unwrap_or_else(|e| panic!("{label}: {e}"));
+
+        // Idempotency: re-parsing the identical bytes yields the identical
+        // dedup-key set (true idempotency of the parse function).
+        let reparse = parse_events(tool, &bytes, &path);
+        assert_eq!(
+            key_set(&a),
+            key_set(&reparse),
+            "{label}: re-parsing changed the event_id key set",
+        );
+    }
+}
diff --git a/crates/memscribe-testkit/tests/redaction.rs b/crates/memscribe-testkit/tests/redaction.rs
new file mode 100644
index 0000000..7e1dbd6
--- /dev/null
+++ b/crates/memscribe-testkit/tests/redaction.rs
@@ -0,0 +1,381 @@
+//! Redaction & privacy tests (whitepaper §8.6) and the cross-version corpus
+//! (whitepaper §8.7).
+//!
+//! ## §8.6 — redaction
+//! Drive the **full** [`DefaultPipeline::new`] (redaction **on** by default) over
+//! transcripts that carry real secret shapes — an Anthropic key, an AWS access
+//! key, a GitHub token, a bearer token, a PEM private-key block, an `API_KEY=`
+//! assignment, and `.env`-style lines — inside both *user-turn text* and *diffs*.
+//! The guarantee under test: **no secret substring survives in any emitted node's
+//! serialized JSON**. A second test proves that a no-content [`Redactor`] elides
+//! all verbatim text while leaving the node *structure* (kinds and counts)
+//! unchanged.
+//!
+//! ## §8.7 — cross-version tolerance
+//! For the three primary tools we add a *second* fixture version with a slightly
+//! different record shape (Claude Code 2.1 with string `content`, a legacy Codex
+//! `v1` pre-rollout shape, a Gemini `legacy_json` `$set`/`$rewindTo` corpus) and
+//! assert the adapter still parses it *losslessly* — every non-blank record maps
+//! to at least one event, nothing panics, and unrecognized records/fields route
+//! to [`EventKind::Unknown`] rather than failing.
+
+use memscribe_core::{
+    CaptureEvent, DefaultPipeline, EventKind, PreparedNode, Redactor, SourceKind,
+};
+use memscribe_testkit::golden::fixtures_dir;
+use memscribe_testkit::{count_nonblank_lines, parse_events};
+use std::path::{Path, PathBuf};
+
+// ---------------------------------------------------------------------------
+// Shared secret catalog
+// ---------------------------------------------------------------------------
+
+/// The verbatim secret substrings that must never survive the pipeline. Each is
+/// matched by a default [`Redactor`] pattern; the tests assert none of these
+/// appears in any emitted node's JSON after redaction.
+const SECRETS: &[&str] = &[
+    // Anthropic API key (`sk-ant-…`).
+    "sk-ant-api03-AAAAAAAAAAAAAAAAAAAAAAAA",
+    // AWS access key id (`AKIA` + 16).
+    "AKIAIOSFODNN7EXAMPLE",
+    // GitHub personal access token (`ghp_…`).
+    "ghp_1234567890abcdefghijklmnopqrstuvwx",
+    // A bearer token value.
+    "abcdefghijklmnop1234567890",
+    // The PEM private-key body line (must be elided with the block).
+    "MIIEowIBAAKCAQEAsupersecretkeymaterial",
+    // `API_KEY=` assignment value (in a diff).
+    "supersecretapikeyvalue123",
+    // `.env`-style assignment values.
+    "topsecretenvvalue",
+    "hunter2envpassword",
+];
+
+/// A PEM private-key block embedded in a diff. The whole block (BEGIN…END) must
+/// be elided, including the body line tracked in [`SECRETS`].
+const PRIVATE_KEY_BLOCK: &str = "-----BEGIN RSA PRIVATE KEY-----\\nMIIEowIBAAKCAQEAsupersecretkeymaterial\\nB2hY9kRdeadbeefcafef00d\\n-----END RSA PRIVATE KEY-----";
+
+/// A Claude Code transcript whose **user-turn text** and **diffs** both carry
+/// secrets. The first user turn is a decision ("use … instead of …") so it is
+/// elevated; the second is a memory directive ("remember that …"). Two assistant
+/// edits carry a PEM block and an `API_KEY=` assignment in their diffs.
+fn secret_bearing_claude_transcript() -> Vec<u8> {
+    // Decision turn — secrets in user text (Anthropic key + bearer token).
+    let decision = r#"{"type":"user","uuid":"r1","parentUuid":null,"timestamp":"2026-06-22T10:00:00.000Z","sessionId":"sess-redact-001","cwd":"/repo","gitBranch":"main","version":"2.0.5","message":{"role":"user","content":"Let's use the Anthropic API instead of OpenAI. My key is sk-ant-api03-AAAAAAAAAAAAAAAAAAAAAAAA and the call uses Authorization: Bearer abcdefghijklmnop1234567890."}}"#;
+    // Memory directive turn — AWS key + GitHub token + .env assignments.
+    let memory = r#"{"type":"user","uuid":"r2","parentUuid":"r1","timestamp":"2026-06-22T10:00:01.000Z","sessionId":"sess-redact-001","version":"2.0.5","message":{"role":"user","content":"Remember that the deploy creds are AKIAIOSFODNN7EXAMPLE and ghp_1234567890abcdefghijklmnopqrstuvwx. The .env has SECRET=topsecretenvvalue and password: hunter2envpassword."}}"#;
+    // Edit #1 — a PEM private-key block lands in the diff (old/new/unified). The
+    // `{pk}` interpolation puts the BEGIN…END block into `newString`.
+    let edit_pk = format!(
+        r#"{{"type":"assistant","uuid":"r3","parentUuid":"r2","timestamp":"2026-06-22T10:00:02.000Z","sessionId":"sess-redact-001","version":"2.0.5","message":{{"role":"assistant","model":"claude-opus-4-8","content":[{{"type":"tool_use","id":"call_pk","name":"Write","input":{{"file_path":"/repo/key.pem"}}}}]}},"toolUseResult":{{"filePath":"/repo/key.pem","oldString":"","newString":"{PRIVATE_KEY_BLOCK}","structuredPatch":[{{"oldStart":1,"oldLines":0,"newStart":1,"newLines":4,"lines":["+-----BEGIN RSA PRIVATE KEY-----","+MIIEowIBAAKCAQEAsupersecretkeymaterial","+B2hY9kRdeadbeefcafef00d","+-----END RSA PRIVATE KEY-----"]}}]}}}}"#
+    );
+    // Edit #2 — an API_KEY= assignment lands in the diff.
+    let edit_cfg = r#"{"type":"assistant","uuid":"r4","parentUuid":"r3","timestamp":"2026-06-22T10:00:03.000Z","sessionId":"sess-redact-001","version":"2.0.5","message":{"role":"assistant","model":"claude-opus-4-8","content":[{"type":"tool_use","id":"call_cfg","name":"Edit","input":{"file_path":"/repo/config.rs"}}]},"toolUseResult":{"filePath":"/repo/config.rs","oldString":"const API_KEY=PLACEHOLDER","newString":"const API_KEY=supersecretapikeyvalue123","structuredPatch":[{"oldStart":1,"oldLines":1,"newStart":1,"newLines":1,"lines":["-const API_KEY=PLACEHOLDER","+const API_KEY=supersecretapikeyvalue123"]}]}}"#;
+
+    [decision, memory, &edit_pk, edit_cfg]
+        .join("\n")
+        .into_bytes()
+}
+
+/// Serialize a node stream to one JSON blob for substring scanning.
+fn nodes_to_json(nodes: &[PreparedNode]) -> String {
+    serde_json::to_string(nodes).expect("nodes serialize")
+}
+
+// ---------------------------------------------------------------------------
+// §8.6 — redaction strips every secret from emitted node JSON
+// ---------------------------------------------------------------------------
+
+#[test]
+fn full_pipeline_redacts_every_secret_from_node_json() {
+    let bytes = secret_bearing_claude_transcript();
+    let path = Path::new("sess-redact-001.jsonl");
+    let events = parse_events(SourceKind::ClaudeCode, &bytes, path);
+
+    // Grounding: without redaction, the secrets DO reach emitted nodes. If a
+    // secret fails to appear here, the fixture is not exercising the pass and
+    // the redaction assertion below would be vacuous — so we fail loudly first.
+    let plain = DefaultPipeline::without_redaction().prepare_events(&events);
+    assert!(!plain.is_empty(), "fixture must produce nodes");
+    let plain_json = nodes_to_json(&plain);
+    for secret in SECRETS {
+        assert!(
+            plain_json.contains(secret),
+            "precondition: secret {secret:?} must reach a node BEFORE redaction \
+             (otherwise the redaction assertion is vacuous)"
+        );
+    }
+
+    // The real guarantee: the default pipeline has redaction ON, and no secret
+    // substring may survive in ANY emitted node's serialized JSON.
+    let redacted = DefaultPipeline::new().prepare_events(&events);
+    assert_eq!(
+        redacted.len(),
+        plain.len(),
+        "redaction must not add or drop nodes"
+    );
+    let redacted_json = nodes_to_json(&redacted);
+    for secret in SECRETS {
+        assert!(
+            !redacted_json.contains(secret),
+            "secret {secret:?} survived redaction in: {redacted_json}"
+        );
+    }
+    // The redaction marker proves the pass actually fired (not that the text was
+    // merely absent for some unrelated reason).
+    assert!(
+        redacted_json.contains("[REDACTED:"),
+        "redaction markers must be present"
+    );
+}
+
+#[test]
+fn redactor_strips_each_pattern_in_isolation() {
+    // A direct, per-pattern check against the default redactor, independent of
+    // segmenter/gate behavior — locks each individual pattern.
+    let r = Redactor::default();
+    let cases = [
+        "sk-ant-api03-AAAAAAAAAAAAAAAAAAAAAAAA",
+        "AKIAIOSFODNN7EXAMPLE",
+        "ghp_1234567890abcdefghijklmnopqrstuvwx",
+        "Authorization: Bearer abcdefghijklmnop1234567890",
+        "const API_KEY=supersecretapikeyvalue123",
+        "SECRET=topsecretenvvalue",
+        "password: hunter2envpassword",
+        "-----BEGIN RSA PRIVATE KEY-----\nMIIEowIBAAKCAQEAsupersecretkeymaterial\n-----END RSA PRIVATE KEY-----",
+    ];
+    for case in cases {
+        assert!(
+            r.contains_secret(case),
+            "redactor must detect secret in {case:?}"
+        );
+        let out = r.redact_text(case);
+        for secret in SECRETS {
+            assert!(
+                !out.contains(secret),
+                "pattern leaked {secret:?} from {case:?} -> {out:?}"
+            );
+        }
+        assert!(
+            out.contains("[REDACTED:"),
+            "redaction marker missing for {case:?}"
+        );
+    }
+}
+
+// ---------------------------------------------------------------------------
+// §8.6 — no-content mode elides verbatim text but preserves structure
+// ---------------------------------------------------------------------------
+
+#[test]
+fn no_content_elides_text_but_keeps_node_kinds_and_counts() {
+    let bytes = secret_bearing_claude_transcript();
+    let path = Path::new("sess-redact-001.jsonl");
+    let events = parse_events(SourceKind::ClaudeCode, &bytes, path);
+
+    // Baseline structure (redaction off): node count and the ordered kind list.
+    let plain = DefaultPipeline::without_redaction().prepare_events(&events);
+    let plain_kinds: Vec<&'static str> = plain.iter().map(PreparedNode::tag).collect();
+
+    // No-content redactor: elides all verbatim text, keeps structure.
+    let no_content =
+        DefaultPipeline::new().with_redactor(Some(Redactor::with_default_patterns(true)));
+    let elided = no_content.prepare_events(&events);
+    let elided_kinds: Vec<&'static str> = elided.iter().map(PreparedNode::tag).collect();
+
+    // Structure unchanged: same number of nodes, same kinds in the same order.
+    assert_eq!(
+        elided.len(),
+        plain.len(),
+        "no-content must not change node count"
+    );
+    assert_eq!(
+        elided_kinds, plain_kinds,
+        "no-content must not change node kinds/order"
+    );
+
+    // All verbatim text is elided to the structural placeholder, and NO secret
+    // (and indeed no original prose) survives anywhere.
+    let elided_json = nodes_to_json(&elided);
+    assert!(
+        elided_json.contains("[content elided]"),
+        "no-content placeholder must be present: {elided_json}"
+    );
+    for secret in SECRETS {
+        assert!(
+            !elided_json.contains(secret),
+            "no-content leaked secret {secret:?}: {elided_json}"
+        );
+    }
+    // A representative non-secret prose fragment is also gone, proving full
+    // elision rather than mere secret-stripping.
+    assert!(
+        !elided_json.contains("Anthropic API instead of OpenAI"),
+        "no-content must elide ALL verbatim text, not just secrets"
+    );
+}
+
+// ---------------------------------------------------------------------------
+// §8.7 — cross-version corpus: version tolerance for the three primary tools
+// ---------------------------------------------------------------------------
+
+/// The path to a cross-version fixture under `fixtures/<tool>/<version>/`.
+fn version_fixture(tool: &str, version: &str, case: &str) -> PathBuf {
+    fixtures_dir()
+        .join(tool)
+        .join(version)
+        .join(format!("{case}.jsonl"))
+}
+
+/// Parse a fixture and assert the version-tolerance guarantee:
+/// - the file is read and produces events (no panic),
+/// - losslessness: at least as many events as non-blank records,
+/// - any unrecognized record/field is preserved as [`EventKind::Unknown`],
+///   never dropped or errored.
+fn assert_version_tolerant(
+    tool: SourceKind,
+    fixture: &Path,
+    expect_unknown: bool,
+) -> Vec<CaptureEvent> {
+    let bytes = std::fs::read(fixture)
+        .unwrap_or_else(|e| panic!("read fixture {}: {e}", fixture.display()));
+    let events = parse_events(tool, &bytes, fixture);
+
+    let nonblank = count_nonblank_lines(&bytes);
+    assert!(
+        events.len() >= nonblank,
+        "{}: lossy — {} non-blank records produced only {} events",
+        fixture.display(),
+        nonblank,
+        events.len()
+    );
+
+    if expect_unknown {
+        assert!(
+            events
+                .iter()
+                .any(|e| matches!(e.kind, EventKind::Unknown { .. })),
+            "{}: an unrecognized record must route to Unknown, not be dropped",
+            fixture.display()
+        );
+    }
+
+    // The full pipeline must run over the (possibly novel) shape without panic.
+    let nodes = DefaultPipeline::new().prepare_events(&events);
+    let _ = nodes_to_json(&nodes);
+
+    events
+}
+
+#[test]
+fn claude_code_2_1_string_content_parses_losslessly() {
+    // 2.1: `content` is a plain string (not blocks), `gitHead` replaces `gitSha`,
+    // and a `telemetry` record type is unrecognized → Unknown.
+    let fixture = version_fixture("claude_code", "2.1", "version_tolerance");
+    let events = assert_version_tolerant(SourceKind::ClaudeCode, &fixture, true);
+
+    // The string-`content` user turn is still a UserTurn with verbatim text.
+    let decision = events.iter().find_map(|e| match &e.kind {
+        EventKind::UserTurn { text, .. } if text.contains("Postgres instead of MySQL") => {
+            Some(text.clone())
+        }
+        _ => None,
+    });
+    assert!(
+        decision.is_some(),
+        "string-content user turn must parse to a UserTurn"
+    );
+
+    // The string-`content` ASSISTANT turn (a renamed shape) is still an
+    // AssistantTurn carrying its verbatim text.
+    let asst = events.iter().find_map(|e| match &e.kind {
+        EventKind::AssistantTurn { text, .. } if text.contains("Switching the orders service") => {
+            Some(text.clone())
+        }
+        _ => None,
+    });
+    assert!(
+        asst.is_some(),
+        "string-content assistant turn must parse to an AssistantTurn"
+    );
+
+    // The unknown `telemetry` record is preserved verbatim, with its raw type.
+    let unknown = events.iter().find_map(|e| match &e.kind {
+        EventKind::Unknown { raw_type, .. } => Some(raw_type.clone()),
+        _ => None,
+    });
+    assert_eq!(
+        unknown.as_deref(),
+        Some("telemetry"),
+        "the telemetry record must survive as Unknown with its raw type"
+    );
+}
+
+#[test]
+fn codex_v1_pre_rollout_shape_parses_losslessly() {
+    // v1: only `session_meta` is recognized; the pre-rollout `record_type` lines
+    // and a `kind`-tagged state line have no `type`+`payload` shape → Unknown.
+    let fixture = version_fixture("codex", "v1", "version_tolerance");
+    let events = assert_version_tolerant(SourceKind::Codex, &fixture, true);
+
+    // The recognized session_meta still opens the session.
+    assert!(
+        events
+            .iter()
+            .any(|e| matches!(e.kind, EventKind::SessionStart { .. })),
+        "session_meta must still produce a SessionStart in the legacy corpus"
+    );
+    // The legacy `record_type` lines are preserved as Unknown (not dropped).
+    let unknown_count = events
+        .iter()
+        .filter(|e| matches!(e.kind, EventKind::Unknown { .. }))
+        .count();
+    assert!(
+        unknown_count >= 3,
+        "all three non-session_meta legacy records must survive as Unknown, got {unknown_count}"
+    );
+}
+
+#[test]
+fn gemini_legacy_json_control_lines_parse_losslessly() {
+    // legacy_json: `$set` opens the session, `content` (not `text`) carries the
+    // turn body, `role:system` is an unknown role → Unknown, and `$rewindTo` is
+    // a control line → Rewind.
+    let fixture = version_fixture("gemini", "legacy_json", "version_tolerance");
+    let events = assert_version_tolerant(SourceKind::Gemini, &fixture, true);
+
+    // `$set` with a cwd opens the session.
+    assert!(
+        events
+            .iter()
+            .any(|e| matches!(e.kind, EventKind::SessionStart { .. })),
+        "$set with cwd must open the session"
+    );
+    // The legacy `content` field is read as the user-turn text.
+    let user = events.iter().find_map(|e| match &e.kind {
+        EventKind::UserTurn { text, .. } if text.contains("Postgres instead of MySQL") => {
+            Some(text.clone())
+        }
+        _ => None,
+    });
+    assert!(
+        user.is_some(),
+        "legacy `content` field must populate the UserTurn text"
+    );
+
+    // The `$rewindTo` control line maps to a Rewind, not a dropped record.
+    assert!(
+        events
+            .iter()
+            .any(|e| matches!(e.kind, EventKind::Rewind { .. })),
+        "$rewindTo must map to a Rewind event"
+    );
+
+    // The unknown `role:system` record is preserved, not dropped.
+    assert!(
+        events
+            .iter()
+            .any(|e| matches!(e.kind, EventKind::Unknown { .. })),
+        "an unknown role must survive as Unknown"
+    );
+}
diff --git a/crates/memscribe-testkit/tests/robustness.rs b/crates/memscribe-testkit/tests/robustness.rs
new file mode 100644
index 0000000..ec5e6ba
--- /dev/null
+++ b/crates/memscribe-testkit/tests/robustness.rs
@@ -0,0 +1,304 @@
+//! Non-nightly robustness suite for every adapter parser (whitepaper §8.4).
+//!
+//! The `fuzz/` crate gives us coverage-guided fuzzing, but it needs a nightly
+//! toolchain and `cargo-fuzz`. This suite is its workspace-resident, stable-Rust
+//! counterpart so CI gets robustness value on every run: it feeds arbitrary and
+//! deliberately *mutated* bytes — random noise, truncated JSON, deeply nested
+//! JSON, gigantic numbers, invalid UTF-8 — to **every** adapter's `parse()` and
+//! asserts the §8.4 parser contract:
+//!
+//! 1. **No panic.** A parser must never panic on any input (we wrap each call in
+//!    [`std::panic::catch_unwind`]).
+//! 2. **Bounded time.** A parser must terminate; we run it on a worker thread and
+//!    fail if it does not finish within a generous wall-clock budget.
+//! 3. **Stream survival.** A single malformed line must be *skipped* (an `Err`
+//!    that [`parse_records`] drops) or routed to `Unknown` — never abort the
+//!    surrounding stream. We sandwich a malformed line between two well-formed
+//!    ones and assert the good events still come through.
+//!
+//! Every check runs against all nine adapters via [`adapter_for`].
+
+use memscribe_adapters::adapter_for;
+use memscribe_core::model::SourceLocation;
+use memscribe_core::pipeline::parse_records;
+use memscribe_core::{ParseCtx, RawRecord, SourceKind, TranscriptAdapter};
+use proptest::prelude::*;
+use std::panic::{catch_unwind, AssertUnwindSafe};
+use std::path::Path;
+use std::sync::mpsc;
+use std::thread;
+use std::time::Duration;
+
+/// The nine adapters this suite must cover, in registry order.
+const ALL_KINDS: &[SourceKind] = &[
+    SourceKind::ClaudeCode,
+    SourceKind::Codex,
+    SourceKind::Gemini,
+    SourceKind::Otel,
+    SourceKind::Cursor,
+    SourceKind::Windsurf,
+    SourceKind::Zed,
+    SourceKind::VsCode,
+    SourceKind::Copilot,
+];
+
+/// Wall-clock budget for a single `parse()` call on one record. Parsing one
+/// line of even pathological JSON is microseconds of work; a multi-second
+/// budget only ever trips on a genuine hang or runaway recursion.
+const PARSE_BUDGET: Duration = Duration::from_secs(5);
+
+/// Resolve the adapter for `kind`, asserting its feature is compiled in (the
+/// default workspace build enables all nine).
+fn adapter(kind: SourceKind) -> Box<dyn TranscriptAdapter> {
+    adapter_for(kind).unwrap_or_else(|| panic!("adapter feature for {kind} must be enabled"))
+}
+
+/// Drive one record through one adapter, enforcing **no panic** and **bounded
+/// time**. Returns nothing — a violation is a test failure, not a value.
+///
+/// The parse runs on a dedicated worker thread so a non-terminating parser can
+/// be detected via a receive timeout instead of hanging the whole test binary.
+/// The worker re-creates the adapter from `kind` (adapters are zero-sized unit
+/// structs, so this is free) to keep everything `Send`.
+fn assert_parse_is_safe(kind: SourceKind, bytes: Vec<u8>) {
+    let (tx, rx) = mpsc::channel::<()>();
+    let worker = thread::spawn(move || {
+        let adapter = adapter(kind);
+        let loc = SourceLocation::new("robustness://input", 0, 1);
+        let raw = RawRecord::new(bytes, loc);
+        let mut ctx = ParseCtx::new();
+        // A panic inside `parse` is caught here so the worker thread always
+        // sends its completion signal; we convert it into an explicit failure.
+        let outcome = catch_unwind(AssertUnwindSafe(|| {
+            let _ = adapter.parse(&raw, &mut ctx);
+            let _ = adapter.schema_fingerprint(&raw);
+        }));
+        // Ignore send errors: if the receiver already timed out and went away,
+        // the main thread has already failed the test.
+        let _ = tx.send(());
+        outcome
+    });
+
+    match rx.recv_timeout(PARSE_BUDGET) {
+        Ok(()) => {
+            // The worker finished in time; surface any panic it caught.
+            match worker.join() {
+                Ok(Ok(())) => {}
+                Ok(Err(_)) => panic!("{kind} adapter panicked while parsing a mutated record"),
+                Err(_) => panic!("{kind} adapter worker thread itself panicked"),
+            }
+        }
+        Err(mpsc::RecvTimeoutError::Timeout) => {
+            panic!("{kind} adapter did not terminate within {PARSE_BUDGET:?} on a mutated record");
+        }
+        Err(mpsc::RecvTimeoutError::Disconnected) => {
+            panic!("{kind} adapter worker thread vanished without completing");
+        }
+    }
+}
+
+// --- Adversarial corpus -----------------------------------------------------
+
+/// A minimal, well-formed JSON object. Not a valid record for any specific
+/// adapter, but well-formed enough that a tolerant parser routes it to
+/// `Unknown` (or `Ok([])`) rather than erroring — our "good anchor" line for the
+/// stream-survival check.
+const GOOD_ANCHOR: &[u8] = br#"{"type":"unknown_but_well_formed","v":1}"#;
+
+/// Deterministic adversarial inputs that every adapter must survive. These are
+/// the named mutation classes from the task, materialized as concrete bytes so
+/// the corpus is reproducible and reviewable.
+fn adversarial_corpus() -> Vec<(&'static str, Vec<u8>)> {
+    // Valid JSON prefix followed by an invalid UTF-8 tail.
+    let json_with_invalid_utf8 = {
+        let mut v = br#"{"text":""#.to_vec();
+        v.extend_from_slice(&[0xff, 0xff]);
+        v.extend_from_slice(br#""}"#);
+        v
+    };
+
+    vec![
+        // Empty and whitespace-only.
+        ("empty", Vec::new()),
+        ("whitespace", b"   \t  ".to_vec()),
+        // Random / non-JSON bytes.
+        ("random_ascii", b"not json at all, just words".to_vec()),
+        ("control_bytes", vec![0x00, 0x01, 0x02, 0x07, 0x1b, 0x7f]),
+        // Invalid UTF-8 (only expressible via raw bytes — `as_str()` must
+        // return None and the parser must still not panic).
+        ("invalid_utf8", vec![0xff, 0xfe, 0xfd, 0xc0, 0x80]),
+        ("lone_surrogate_bytes", vec![0xed, 0xa0, 0x80]),
+        ("json_with_invalid_utf8", json_with_invalid_utf8),
+        // Truncated JSON in several shapes.
+        (
+            "truncated_object",
+            br#"{"type":"user","message":{"role":"#.to_vec(),
+        ),
+        ("truncated_string", br#"{"text":"unterminated"#.to_vec()),
+        ("truncated_array", b"[1,2,3".to_vec()),
+        ("dangling_comma", br#"{"a":1,}"#.to_vec()),
+        ("just_open_brace", b"{".to_vec()),
+        // Deeply nested JSON — exercises any recursive descent for stack safety.
+        ("deep_array", deep_nested_array(2_000)),
+        ("deep_object", deep_nested_object(2_000)),
+        // Huge numbers — beyond i64/u64/f64 range and absurd precision.
+        (
+            "huge_integer",
+            format!(r#"{{"n":{}}}"#, "9".repeat(400)).into_bytes(),
+        ),
+        ("huge_exponent", br#"{"n":1e400,"m":-1e-400}"#.to_vec()),
+        (
+            "huge_precision",
+            format!(r#"{{"n":0.{}}}"#, "1".repeat(500)).into_bytes(),
+        ),
+        (
+            "huge_timestamp",
+            br#"{"timestamp":99999999999999999999,"ts":"+999999-01-01T00:00:00Z"}"#.to_vec(),
+        ),
+        // A plausible-but-wrong record: right-shaped keys, garbage values.
+        (
+            "type_confusion",
+            br#"{"type":12345,"message":[],"timestamp":true,"usage":"nope"}"#.to_vec(),
+        ),
+        // A long single line, to make sure nothing is quadratic enough to time
+        // out.
+        (
+            "long_text",
+            format!(r#"{{"text":"{}"}}"#, "a".repeat(50_000)).into_bytes(),
+        ),
+    ]
+}
+
+/// `[[[...]]]` nested `depth` deep — a stack-depth stressor for recursive
+/// JSON parsers. `serde_json` has its own recursion limit; the point is the
+/// adapter must not panic regardless of how `serde_json` reports it.
+fn deep_nested_array(depth: usize) -> Vec<u8> {
+    let mut s = String::with_capacity(depth * 2 + 2);
+    for _ in 0..depth {
+        s.push('[');
+    }
+    for _ in 0..depth {
+        s.push(']');
+    }
+    s.into_bytes()
+}
+
+/// `{"a":{"a":{...}}}` nested `depth` deep.
+fn deep_nested_object(depth: usize) -> Vec<u8> {
+    let mut s = String::new();
+    for _ in 0..depth {
+        s.push_str(r#"{"a":"#);
+    }
+    s.push('1');
+    for _ in 0..depth {
+        s.push('}');
+    }
+    s.into_bytes()
+}
+
+// --- Tests ------------------------------------------------------------------
+
+/// Every adapter survives every deterministic adversarial input: no panic,
+/// bounded time. This is the reproducible core of the suite.
+#[test]
+fn every_adapter_survives_adversarial_corpus() {
+    for &kind in ALL_KINDS {
+        for (name, bytes) in adversarial_corpus() {
+            // A fresh clone per call; `assert_parse_is_safe` takes ownership and
+            // moves the bytes onto the worker thread.
+            let label = format!("{kind}/{name}");
+            assert_parse_is_safe(kind, bytes);
+            // Touch `label` so a failing case is identifiable in a backtrace via
+            // the panic message above; this also keeps the binding live.
+            let _ = label;
+        }
+    }
+}
+
+/// A malformed line sandwiched between two well-formed ones must not abort the
+/// stream: [`parse_records`] skips the bad record and keeps going. We assert the
+/// good anchors still produce at least as many events as a clean two-line run,
+/// proving the malformed middle line neither erased earlier output nor stopped
+/// later parsing.
+#[test]
+fn malformed_line_is_skipped_not_fatal_for_every_adapter() {
+    let path = Path::new("robustness://stream");
+    for &kind in ALL_KINDS {
+        let a = adapter(kind);
+
+        // Baseline: two good anchor lines, no malformed middle.
+        let clean = records_from_lines(&[GOOD_ANCHOR, GOOD_ANCHOR], path);
+        let (clean_events, _) = parse_records(a.as_ref(), &clean);
+
+        for (name, bad) in adversarial_corpus() {
+            // good, MALFORMED, good — the malformed line must not take the
+            // surrounding records down with it.
+            let mixed = records_from_lines(&[GOOD_ANCHOR, &bad, GOOD_ANCHOR], path);
+
+            // The whole stream must parse without panicking.
+            let result = catch_unwind(AssertUnwindSafe(|| parse_records(a.as_ref(), &mixed)));
+            let (mixed_events, _) = result.unwrap_or_else(|_| {
+                panic!("{kind} aborted the stream on a malformed `{name}` line")
+            });
+
+            // Stream survival: the two good anchors still produced their events;
+            // the malformed middle did not erase or block them.
+            assert!(
+                mixed_events.len() >= clean_events.len(),
+                "{kind}: malformed `{name}` line suppressed good events \
+                 (clean={}, mixed={})",
+                clean_events.len(),
+                mixed_events.len(),
+            );
+        }
+    }
+}
+
+/// Build a single multi-line transcript from the given line byte-slices and
+/// split it back into records exactly the way the real reader does, so the
+/// per-record provenance (offsets, line numbers) matches production.
+fn records_from_lines(lines: &[&[u8]], path: &Path) -> Vec<RawRecord> {
+    let mut buf: Vec<u8> = Vec::new();
+    for (i, line) in lines.iter().enumerate() {
+        if i > 0 {
+            buf.push(b'\n');
+        }
+        buf.extend_from_slice(line);
+    }
+    memscribe_io::read_records_from_bytes(&buf, path)
+}
+
+proptest! {
+    // Keep the case count modest: each case spawns a worker thread per adapter,
+    // so 64 cases × 9 adapters is a few hundred guarded parses — plenty of
+    // coverage without making the suite slow.
+    #![proptest_config(ProptestConfig::with_cases(64))]
+
+    /// Arbitrary byte vectors (including invalid UTF-8) never panic or hang any
+    /// adapter. This is the property-based analogue of the cargo-fuzz targets.
+    #[test]
+    fn arbitrary_bytes_never_panic_any_adapter(bytes in proptest::collection::vec(any::<u8>(), 0..512)) {
+        for &kind in ALL_KINDS {
+            assert_parse_is_safe(kind, bytes.clone());
+        }
+    }
+
+    /// Mutated-JSON inputs: start from a well-formed-ish object and let proptest
+    /// splice random bytes in, exercising the "almost valid" region of the input
+    /// space that pure-random bytes rarely reach.
+    #[test]
+    fn mutated_json_never_panics_any_adapter(
+        seed in proptest::collection::vec(any::<u8>(), 0..64),
+        cut in 0usize..40,
+    ) {
+        let mut bytes = br#"{"type":"assistant","message":{"role":"assistant","content":[{"type":"text","text":"hi"}]},"timestamp":"2026-06-22T10:00:00Z"}"#.to_vec();
+        // Truncate at an arbitrary point, then append random noise: a cheap way
+        // to manufacture truncated-then-garbled JSON.
+        let at = cut.min(bytes.len());
+        bytes.truncate(at);
+        bytes.extend_from_slice(&seed);
+        for &kind in ALL_KINDS {
+            assert_parse_is_safe(kind, bytes.clone());
+        }
+    }
+}
diff --git a/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__ban.snap b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__ban.snap
new file mode 100644
index 0000000..05a35d5
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__ban.snap
@@ -0,0 +1,110 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-ban-003",
+    "seq": 0,
+    "event_id": "29c0476fba3dbcbcbbe1770f2bd1a1f6f6c96a1422d857bd6ac0188784e5f3e8",
+    "parent_id": null,
+    "timestamp": "2026-06-22T12:15:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/core",
+      "repo_root": null,
+      "git": {
+        "sha": "deadbeefdeadbeefdeadbeefdeadbeefdeadbeef",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme/core",
+      "git": {
+        "sha": "deadbeefdeadbeefdeadbeefdeadbeefdeadbeef",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "2.0.5"
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/ban.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-ban-003",
+    "seq": 1,
+    "event_id": "bbbbbbb1-0000-4000-8000-000000000001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T12:15:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/core",
+      "repo_root": null,
+      "git": {
+        "sha": "deadbeefdeadbeefdeadbeefdeadbeefdeadbeef",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "We will never add a dependency on left-pad in this repo. Use the std string padding instead.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "We will never add a dependency on left-pad in this repo. Use the std string padding instead."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/ban.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-ban-003",
+    "seq": 2,
+    "event_id": "bbbbbbb2-0000-4000-8000-000000000002",
+    "parent_id": "bbbbbbb1-0000-4000-8000-000000000001",
+    "timestamp": "2026-06-22T12:15:03Z",
+    "project": {
+      "cwd": "/Users/dev/acme/core",
+      "repo_root": null,
+      "git": {
+        "sha": "deadbeefdeadbeefdeadbeefdeadbeefdeadbeef",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Noted. I'll keep padding on the standard library and avoid third-party micro-dependencies.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 210,
+        "output_tokens": 48,
+        "cache_read_tokens": 0,
+        "cache_creation_tokens": 0
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Noted. I'll keep padding on the standard library and avoid third-party micro-dependencies."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/ban.jsonl",
+      "byte_offset": 415,
+      "line_no": 2
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..5b7607b
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__happy_path_decision_then_edits.snap
@@ -0,0 +1,424 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-happy-001",
+    "seq": 0,
+    "event_id": "8099926f555f38a5e3198f983b067f9fc295de21160ebb6f4d3799966845b3b8",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme/api",
+      "git": {
+        "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "2.0.5"
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-happy-001",
+    "seq": 1,
+    "event_id": "11111111-1111-4111-8111-111111111111",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use Postgres instead of MySQL for the orders service.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's use Postgres instead of MySQL for the orders service."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-happy-001",
+    "seq": 2,
+    "event_id": "22222222-2222-4222-8222-222222222222",
+    "parent_id": "11111111-1111-4111-8111-111111111111",
+    "timestamp": "2026-06-22T10:00:03.12Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Switching the orders service to Postgres. I'll update the connection config and the schema module.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 482,
+        "output_tokens": 96,
+        "cache_read_tokens": 120,
+        "cache_creation_tokens": 40
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Switching the orders service to Postgres. I'll update the connection config and the schema module."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_01edit_config",
+          "name": "Edit",
+          "args": {
+            "file_path": "/Users/dev/acme/api/src/db/config.rs",
+            "old_string": "mysql://localhost:3306/orders",
+            "new_string": "postgres://localhost:5432/orders"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 383,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-happy-001",
+    "seq": 3,
+    "event_id": "da3654fc1eb3fc44085b23c7c8bd3f245765145b80abff5674f69e12341133f2",
+    "parent_id": "22222222-2222-4222-8222-222222222222",
+    "timestamp": "2026-06-22T10:00:03.12Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_01edit_config",
+      "name": "Edit",
+      "args": {
+        "file_path": "/Users/dev/acme/api/src/db/config.rs",
+        "old_string": "mysql://localhost:3306/orders",
+        "new_string": "postgres://localhost:5432/orders"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 383,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-happy-001",
+    "seq": 4,
+    "event_id": "33333333-3333-4333-8333-333333333333",
+    "parent_id": "22222222-2222-4222-8222-222222222222",
+    "timestamp": "2026-06-22T10:00:03.48Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "",
+      "parts": [
+        {
+          "part": "tool_result",
+          "call_id": "toolu_01edit_config",
+          "output": "The file /Users/dev/acme/api/src/db/config.rs has been updated."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1174,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-happy-001",
+    "seq": 5,
+    "event_id": "f81ef847b5a424e5bc8745c9288de5aba856ddb249328d1bbe1395e3e7780465",
+    "parent_id": "33333333-3333-4333-8333-333333333333",
+    "timestamp": "2026-06-22T10:00:03.48Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "toolu_01edit_config",
+      "ok": true,
+      "output": "The file /Users/dev/acme/api/src/db/config.rs has been updated."
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1174,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-happy-001",
+    "seq": 6,
+    "event_id": "9a994fec70e293568cfd331491b45b7217a5801585b2b005da33781648b2ddb6",
+    "parent_id": "33333333-3333-4333-8333-333333333333",
+    "timestamp": "2026-06-22T10:00:03.48Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "toolu_01edit_config",
+      "diff": {
+        "path": "/Users/dev/acme/api/src/db/config.rs",
+        "old": "mysql://localhost:3306/orders",
+        "new": "postgres://localhost:5432/orders",
+        "unified": "-    const DSN: &str = \"mysql://localhost:3306/orders\";\n+    const DSN: &str = \"postgres://localhost:5432/orders\";",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1174,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-happy-001",
+    "seq": 7,
+    "event_id": "44444444-4444-4444-8444-444444444444",
+    "parent_id": "33333333-3333-4333-8333-333333333333",
+    "timestamp": "2026-06-22T10:00:06.9Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Now updating the schema driver import.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 640,
+        "output_tokens": 140,
+        "cache_read_tokens": 480,
+        "cache_creation_tokens": 0
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Now updating the schema driver import."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_02edit_schema",
+          "name": "Edit",
+          "args": {
+            "file_path": "/Users/dev/acme/api/src/db/schema.rs",
+            "old_string": "use mysql::Pool;",
+            "new_string": "use postgres::Client;"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 2039,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-happy-001",
+    "seq": 8,
+    "event_id": "2603db709a643db6056c328854240f71d391ce6afbe595b209921fd97d790278",
+    "parent_id": "44444444-4444-4444-8444-444444444444",
+    "timestamp": "2026-06-22T10:00:06.9Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_02edit_schema",
+      "name": "Edit",
+      "args": {
+        "file_path": "/Users/dev/acme/api/src/db/schema.rs",
+        "old_string": "use mysql::Pool;",
+        "new_string": "use postgres::Client;"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 2039,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-happy-001",
+    "seq": 9,
+    "event_id": "55555555-5555-4555-8555-555555555555",
+    "parent_id": "44444444-4444-4444-8444-444444444444",
+    "timestamp": "2026-06-22T10:00:07.21Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "",
+      "parts": [
+        {
+          "part": "tool_result",
+          "call_id": "toolu_02edit_schema",
+          "output": "The file /Users/dev/acme/api/src/db/schema.rs has been updated."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 2746,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-happy-001",
+    "seq": 10,
+    "event_id": "13300cdcd771b3946c86e3b533a4dad91c5002e024f56e986e69b52881335e4f",
+    "parent_id": "55555555-5555-4555-8555-555555555555",
+    "timestamp": "2026-06-22T10:00:07.21Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "toolu_02edit_schema",
+      "ok": true,
+      "output": "The file /Users/dev/acme/api/src/db/schema.rs has been updated."
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 2746,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-happy-001",
+    "seq": 11,
+    "event_id": "8e21f86302f70c7c2075fac8eefc50dd74e602d3027ef6a204bf61562acad7ba",
+    "parent_id": "55555555-5555-4555-8555-555555555555",
+    "timestamp": "2026-06-22T10:00:07.21Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "toolu_02edit_schema",
+      "diff": {
+        "path": "/Users/dev/acme/api/src/db/schema.rs",
+        "old": "use mysql::Pool;",
+        "new": "use postgres::Client;",
+        "unified": "-use mysql::Pool;\n+use postgres::Client;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 2746,
+      "line_no": 5
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__interleaved_arcs.snap b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__interleaved_arcs.snap
new file mode 100644
index 0000000..3c492ef
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__interleaved_arcs.snap
@@ -0,0 +1,633 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 0,
+    "event_id": "3a1423ea740421651cd930118216a75453f1f7661de6d315a86406e48b9d2d88",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme/api",
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "2.0.5"
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 1,
+    "event_id": "d1aaaaa1-0000-4000-8000-000000000001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use argon2 for password hashing in the auth module.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's use argon2 for password hashing in the auth module."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 2,
+    "event_id": "d1aaaaa2-0000-4000-8000-000000000002",
+    "parent_id": "d1aaaaa1-0000-4000-8000-000000000001",
+    "timestamp": "2026-06-22T14:00:03Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Switching the hasher to argon2 in the auth module.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 300,
+        "output_tokens": 70,
+        "cache_read_tokens": 0,
+        "cache_creation_tokens": 120
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Switching the hasher to argon2 in the auth module."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_arc_a_auth",
+          "name": "Edit",
+          "args": {
+            "file_path": "/Users/dev/acme/api/src/auth.rs",
+            "old_string": "use bcrypt::hash;",
+            "new_string": "use argon2::hash;"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 387,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 3,
+    "event_id": "d360d7fe354ddb00e4a56bc26d64486e1e853bb0d18fc62f3e7684ff0e8809d5",
+    "parent_id": "d1aaaaa2-0000-4000-8000-000000000002",
+    "timestamp": "2026-06-22T14:00:03Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_arc_a_auth",
+      "name": "Edit",
+      "args": {
+        "file_path": "/Users/dev/acme/api/src/auth.rs",
+        "old_string": "use bcrypt::hash;",
+        "new_string": "use argon2::hash;"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 387,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 4,
+    "event_id": "d1aaaaa3-0000-4000-8000-000000000003",
+    "parent_id": "d1aaaaa2-0000-4000-8000-000000000002",
+    "timestamp": "2026-06-22T14:00:03.5Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "",
+      "parts": [
+        {
+          "part": "tool_result",
+          "call_id": "toolu_arc_a_auth",
+          "output": "The file /Users/dev/acme/api/src/auth.rs has been updated."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 1100,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 5,
+    "event_id": "f23451a2b3e5837d27f6183c6417bbf74dbdd915d40a4eb24b9228fcda7ff3ef",
+    "parent_id": "d1aaaaa3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T14:00:03.5Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "toolu_arc_a_auth",
+      "ok": true,
+      "output": "The file /Users/dev/acme/api/src/auth.rs has been updated."
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 1100,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 6,
+    "event_id": "a78a0e286df422ff98681b4b8fd006cdefa69203c5da74e28b5e1b4ac9ea05ab",
+    "parent_id": "d1aaaaa3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T14:00:03.5Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "toolu_arc_a_auth",
+      "diff": {
+        "path": "/Users/dev/acme/api/src/auth.rs",
+        "old": "use bcrypt::hash;",
+        "new": "use argon2::hash;",
+        "unified": "-use bcrypt::hash;\n+use argon2::hash;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 1100,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 7,
+    "event_id": "d1aaaaa4-0000-4000-8000-000000000004",
+    "parent_id": "d1aaaaa3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T14:01:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Actually, let's switch to scrypt for the auth module and tune the cost in config.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Actually, let's switch to scrypt for the auth module and tune the cost in config."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 1853,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 8,
+    "event_id": "d1aaaaa5-0000-4000-8000-000000000005",
+    "parent_id": "d1aaaaa4-0000-4000-8000-000000000004",
+    "timestamp": "2026-06-22T14:01:04Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Re-pointing the auth hasher to scrypt.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 360,
+        "output_tokens": 90,
+        "cache_read_tokens": 300,
+        "cache_creation_tokens": 0
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Re-pointing the auth hasher to scrypt."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_arc_b_auth",
+          "name": "Edit",
+          "args": {
+            "file_path": "/Users/dev/acme/api/src/auth.rs",
+            "old_string": "use argon2::hash;",
+            "new_string": "use scrypt::hash;"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 2246,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 9,
+    "event_id": "a719d5a45c6af5850df2e6257d5e7996517ff24e4de7a44ed12416619a4e288a",
+    "parent_id": "d1aaaaa5-0000-4000-8000-000000000005",
+    "timestamp": "2026-06-22T14:01:04Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_arc_b_auth",
+      "name": "Edit",
+      "args": {
+        "file_path": "/Users/dev/acme/api/src/auth.rs",
+        "old_string": "use argon2::hash;",
+        "new_string": "use scrypt::hash;"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 2246,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 10,
+    "event_id": "d1aaaaa6-0000-4000-8000-000000000006",
+    "parent_id": "d1aaaaa5-0000-4000-8000-000000000005",
+    "timestamp": "2026-06-22T14:01:04.5Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "",
+      "parts": [
+        {
+          "part": "tool_result",
+          "call_id": "toolu_arc_b_auth",
+          "output": "The file /Users/dev/acme/api/src/auth.rs has been updated."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 2947,
+      "line_no": 6
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 11,
+    "event_id": "a56f837bcc848d86066ce95c684beb3901b7e56c918c1293b40f80a7eca63c73",
+    "parent_id": "d1aaaaa6-0000-4000-8000-000000000006",
+    "timestamp": "2026-06-22T14:01:04.5Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "toolu_arc_b_auth",
+      "ok": true,
+      "output": "The file /Users/dev/acme/api/src/auth.rs has been updated."
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 2947,
+      "line_no": 6
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 12,
+    "event_id": "80107084a3da3c00da57c882d34f636980a1d0b8ccbb03732a5dbf92c2c032ae",
+    "parent_id": "d1aaaaa6-0000-4000-8000-000000000006",
+    "timestamp": "2026-06-22T14:01:04.5Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "toolu_arc_b_auth",
+      "diff": {
+        "path": "/Users/dev/acme/api/src/auth.rs",
+        "old": "use argon2::hash;",
+        "new": "use scrypt::hash;",
+        "unified": "-use argon2::hash;\n+use scrypt::hash;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 2947,
+      "line_no": 6
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 13,
+    "event_id": "d1aaaaa7-0000-4000-8000-000000000007",
+    "parent_id": "d1aaaaa6-0000-4000-8000-000000000006",
+    "timestamp": "2026-06-22T14:01:08Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Tuning the scrypt cost parameter in config.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 410,
+        "output_tokens": 85,
+        "cache_read_tokens": 360,
+        "cache_creation_tokens": 0
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Tuning the scrypt cost parameter in config."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_arc_b_config",
+          "name": "Edit",
+          "args": {
+            "file_path": "/Users/dev/acme/api/src/config.rs",
+            "old_string": "const HASH_COST: u32 = 10;",
+            "new_string": "const HASH_COST: u32 = 15;"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 3700,
+      "line_no": 7
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 14,
+    "event_id": "0ef15c7fdec7df16bc3b5eb3b0dd370b67dab8b1c45f1eea3950c21d4a61917b",
+    "parent_id": "d1aaaaa7-0000-4000-8000-000000000007",
+    "timestamp": "2026-06-22T14:01:08Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_arc_b_config",
+      "name": "Edit",
+      "args": {
+        "file_path": "/Users/dev/acme/api/src/config.rs",
+        "old_string": "const HASH_COST: u32 = 10;",
+        "new_string": "const HASH_COST: u32 = 15;"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 3700,
+      "line_no": 7
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 15,
+    "event_id": "d1aaaaa8-0000-4000-8000-000000000008",
+    "parent_id": "d1aaaaa7-0000-4000-8000-000000000007",
+    "timestamp": "2026-06-22T14:01:08.5Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "",
+      "parts": [
+        {
+          "part": "tool_result",
+          "call_id": "toolu_arc_b_config",
+          "output": "The file /Users/dev/acme/api/src/config.rs has been updated."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 4428,
+      "line_no": 8
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 16,
+    "event_id": "560a8c8a172b4fc34b59e0a118d1579a7b384e5d3e4feb08c9c5c020116a15f6",
+    "parent_id": "d1aaaaa8-0000-4000-8000-000000000008",
+    "timestamp": "2026-06-22T14:01:08.5Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "toolu_arc_b_config",
+      "ok": true,
+      "output": "The file /Users/dev/acme/api/src/config.rs has been updated."
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 4428,
+      "line_no": 8
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-interleaved-005",
+    "seq": 17,
+    "event_id": "ccaf7c474b51d55025baba21ea22f808b5197b85cea212ce045f9dc9a3a2a83e",
+    "parent_id": "d1aaaaa8-0000-4000-8000-000000000008",
+    "timestamp": "2026-06-22T14:01:08.5Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "1111111122222222333333334444444455555555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "toolu_arc_b_config",
+      "diff": {
+        "path": "/Users/dev/acme/api/src/config.rs",
+        "old": "const HASH_COST: u32 = 10;",
+        "new": "const HASH_COST: u32 = 15;",
+        "unified": "-const HASH_COST: u32 = 10;\n+const HASH_COST: u32 = 15;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+      "byte_offset": 4428,
+      "line_no": 8
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__multi_edit_single_commit.snap b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__multi_edit_single_commit.snap
new file mode 100644
index 0000000..2cf1070
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__multi_edit_single_commit.snap
@@ -0,0 +1,601 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 0,
+    "event_id": "a7734c5a3cdff1577e96f755e63ddb3fafc1a063520a4ce402c323b567db59a7",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme/api",
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "2.0.5"
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 1,
+    "event_id": "d2bbbbb1-0000-4000-8000-000000000001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's migrate to anyhow for error handling across the request, handler, and router modules.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's migrate to anyhow for error handling across the request, handler, and router modules."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 2,
+    "event_id": "d2bbbbb2-0000-4000-8000-000000000002",
+    "parent_id": "d2bbbbb1-0000-4000-8000-000000000001",
+    "timestamp": "2026-06-22T15:00:03Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "I'll convert each module to anyhow::Result. Starting with request.rs.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 420,
+        "output_tokens": 110,
+        "cache_read_tokens": 0,
+        "cache_creation_tokens": 180
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "I'll convert each module to anyhow::Result. Starting with request.rs."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_multi_request",
+          "name": "Edit",
+          "args": {
+            "file_path": "/Users/dev/acme/api/src/request.rs",
+            "old_string": "use crate::error::AppError;",
+            "new_string": "use anyhow::Result;"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 419,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 3,
+    "event_id": "c95f8a23152120475186b7183f79e6f935eebb0fd2767aacbae284600fc8a84c",
+    "parent_id": "d2bbbbb2-0000-4000-8000-000000000002",
+    "timestamp": "2026-06-22T15:00:03Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_multi_request",
+      "name": "Edit",
+      "args": {
+        "file_path": "/Users/dev/acme/api/src/request.rs",
+        "old_string": "use crate::error::AppError;",
+        "new_string": "use anyhow::Result;"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 419,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 4,
+    "event_id": "d2bbbbb3-0000-4000-8000-000000000003",
+    "parent_id": "d2bbbbb2-0000-4000-8000-000000000002",
+    "timestamp": "2026-06-22T15:00:03.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "",
+      "parts": [
+        {
+          "part": "tool_result",
+          "call_id": "toolu_multi_request",
+          "output": "The file /Users/dev/acme/api/src/request.rs has been updated."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 1168,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 5,
+    "event_id": "20efc68a28fc836e1300f425cd2e1bdc0a78eae06ef78591f59ab943fb8630df",
+    "parent_id": "d2bbbbb3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T15:00:03.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "toolu_multi_request",
+      "ok": true,
+      "output": "The file /Users/dev/acme/api/src/request.rs has been updated."
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 1168,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 6,
+    "event_id": "a0703cd2f3f9be8afe7fffdba3bfc07d6261b77af17f6611a8f8c237cb86bed6",
+    "parent_id": "d2bbbbb3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T15:00:03.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "toolu_multi_request",
+      "diff": {
+        "path": "/Users/dev/acme/api/src/request.rs",
+        "old": "use crate::error::AppError;",
+        "new": "use anyhow::Result;",
+        "unified": "-use crate::error::AppError;\n+use anyhow::Result;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 1168,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 7,
+    "event_id": "d2bbbbb4-0000-4000-8000-000000000004",
+    "parent_id": "d2bbbbb3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T15:00:06Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Now handler.rs.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 480,
+        "output_tokens": 95,
+        "cache_read_tokens": 420,
+        "cache_creation_tokens": 0
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Now handler.rs."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_multi_handler",
+          "name": "Edit",
+          "args": {
+            "file_path": "/Users/dev/acme/api/src/handler.rs",
+            "old_string": "use crate::error::AppError;",
+            "new_string": "use anyhow::Result;"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 1974,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 8,
+    "event_id": "5c52d560ed043d41d9c2ba3a849e2285f7c0b633d4fcd65fbe66b218367eb6a2",
+    "parent_id": "d2bbbbb4-0000-4000-8000-000000000004",
+    "timestamp": "2026-06-22T15:00:06Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_multi_handler",
+      "name": "Edit",
+      "args": {
+        "file_path": "/Users/dev/acme/api/src/handler.rs",
+        "old_string": "use crate::error::AppError;",
+        "new_string": "use anyhow::Result;"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 1974,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 9,
+    "event_id": "d2bbbbb5-0000-4000-8000-000000000005",
+    "parent_id": "d2bbbbb4-0000-4000-8000-000000000004",
+    "timestamp": "2026-06-22T15:00:06.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "",
+      "parts": [
+        {
+          "part": "tool_result",
+          "call_id": "toolu_multi_handler",
+          "output": "The file /Users/dev/acme/api/src/handler.rs has been updated."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 2668,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 10,
+    "event_id": "2acc0f8513f070532d9222d84c1a56cf557546a2cbcbe21154879778ec6420a7",
+    "parent_id": "d2bbbbb5-0000-4000-8000-000000000005",
+    "timestamp": "2026-06-22T15:00:06.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "toolu_multi_handler",
+      "ok": true,
+      "output": "The file /Users/dev/acme/api/src/handler.rs has been updated."
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 2668,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 11,
+    "event_id": "a7e8736565115cd1477920b81020f7fc260d736559d61efd378d60d077fef12a",
+    "parent_id": "d2bbbbb5-0000-4000-8000-000000000005",
+    "timestamp": "2026-06-22T15:00:06.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "toolu_multi_handler",
+      "diff": {
+        "path": "/Users/dev/acme/api/src/handler.rs",
+        "old": "use crate::error::AppError;",
+        "new": "use anyhow::Result;",
+        "unified": "-use crate::error::AppError;\n+use anyhow::Result;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 2668,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 12,
+    "event_id": "d2bbbbb6-0000-4000-8000-000000000006",
+    "parent_id": "d2bbbbb5-0000-4000-8000-000000000005",
+    "timestamp": "2026-06-22T15:00:09Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "And finally router.rs.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 540,
+        "output_tokens": 95,
+        "cache_read_tokens": 480,
+        "cache_creation_tokens": 0
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "And finally router.rs."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_multi_router",
+          "name": "Edit",
+          "args": {
+            "file_path": "/Users/dev/acme/api/src/router.rs",
+            "old_string": "use crate::error::AppError;",
+            "new_string": "use anyhow::Result;"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 3474,
+      "line_no": 6
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 13,
+    "event_id": "22d3bd186b096ab920ba071c27e311e5c953d814f6a67f9f78aa831d852fe15b",
+    "parent_id": "d2bbbbb6-0000-4000-8000-000000000006",
+    "timestamp": "2026-06-22T15:00:09Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_multi_router",
+      "name": "Edit",
+      "args": {
+        "file_path": "/Users/dev/acme/api/src/router.rs",
+        "old_string": "use crate::error::AppError;",
+        "new_string": "use anyhow::Result;"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 3474,
+      "line_no": 6
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 14,
+    "event_id": "d2bbbbb7-0000-4000-8000-000000000007",
+    "parent_id": "d2bbbbb6-0000-4000-8000-000000000006",
+    "timestamp": "2026-06-22T15:00:09.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "",
+      "parts": [
+        {
+          "part": "tool_result",
+          "call_id": "toolu_multi_router",
+          "output": "The file /Users/dev/acme/api/src/router.rs has been updated."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 4173,
+      "line_no": 7
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 15,
+    "event_id": "9191e66261f113d4f52c3dc83949f94cf8c36f4c856266a8d6483ab7e0815e43",
+    "parent_id": "d2bbbbb7-0000-4000-8000-000000000007",
+    "timestamp": "2026-06-22T15:00:09.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "toolu_multi_router",
+      "ok": true,
+      "output": "The file /Users/dev/acme/api/src/router.rs has been updated."
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 4173,
+      "line_no": 7
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-multiedit-006",
+    "seq": 16,
+    "event_id": "8d6814cc10b737c815002e5f3bb8992d8ba7656ea98618710ea80992b841246a",
+    "parent_id": "d2bbbbb7-0000-4000-8000-000000000007",
+    "timestamp": "2026-06-22T15:00:09.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "toolu_multi_router",
+      "diff": {
+        "path": "/Users/dev/acme/api/src/router.rs",
+        "old": "use crate::error::AppError;",
+        "new": "use anyhow::Result;",
+        "unified": "-use crate::error::AppError;\n+use anyhow::Result;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+      "byte_offset": 4173,
+      "line_no": 7
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__no_commitment_marker.snap b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__no_commitment_marker.snap
new file mode 100644
index 0000000..5f69677
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__no_commitment_marker.snap
@@ -0,0 +1,247 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-nomarker-009",
+    "seq": 0,
+    "event_id": "1ea3dae75813cc75cab54b295ece07cf4f1d3a6891415dd522607b9f5dfed176",
+    "parent_id": null,
+    "timestamp": "2026-06-22T18:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0a0b0c0d0e0f10111213141516171819202122ab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme/api",
+      "git": {
+        "sha": "0a0b0c0d0e0f10111213141516171819202122ab",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "2.0.5"
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/no_commitment_marker.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-nomarker-009",
+    "seq": 1,
+    "event_id": "d5fffff1-0000-4000-8000-000000000001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T18:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0a0b0c0d0e0f10111213141516171819202122ab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Thanks, that looks good to me. Please tidy the trailing whitespace in the handler.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Thanks, that looks good to me. Please tidy the trailing whitespace in the handler."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/no_commitment_marker.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-nomarker-009",
+    "seq": 2,
+    "event_id": "d5fffff2-0000-4000-8000-000000000002",
+    "parent_id": "d5fffff1-0000-4000-8000-000000000001",
+    "timestamp": "2026-06-22T18:00:03Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0a0b0c0d0e0f10111213141516171819202122ab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Tidying the trailing whitespace in the handler now.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 210,
+        "output_tokens": 45,
+        "cache_read_tokens": 0,
+        "cache_creation_tokens": 70
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Tidying the trailing whitespace in the handler now."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_nomarker_fmt",
+          "name": "Edit",
+          "args": {
+            "file_path": "/Users/dev/acme/api/src/handler.rs",
+            "old_string": "    let body = read(); ",
+            "new_string": "    let body = read();"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/no_commitment_marker.jsonl",
+      "byte_offset": 409,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-nomarker-009",
+    "seq": 3,
+    "event_id": "fbedd67c6680789bb13865302012615a92673fbaab3b144766caa097a6665b75",
+    "parent_id": "d5fffff2-0000-4000-8000-000000000002",
+    "timestamp": "2026-06-22T18:00:03Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0a0b0c0d0e0f10111213141516171819202122ab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_nomarker_fmt",
+      "name": "Edit",
+      "args": {
+        "file_path": "/Users/dev/acme/api/src/handler.rs",
+        "old_string": "    let body = read(); ",
+        "new_string": "    let body = read();"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/no_commitment_marker.jsonl",
+      "byte_offset": 409,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-nomarker-009",
+    "seq": 4,
+    "event_id": "d5fffff3-0000-4000-8000-000000000003",
+    "parent_id": "d5fffff2-0000-4000-8000-000000000002",
+    "timestamp": "2026-06-22T18:00:03.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0a0b0c0d0e0f10111213141516171819202122ab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "",
+      "parts": [
+        {
+          "part": "tool_result",
+          "call_id": "toolu_nomarker_fmt",
+          "output": "The file /Users/dev/acme/api/src/handler.rs has been updated."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/no_commitment_marker.jsonl",
+      "byte_offset": 1135,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-nomarker-009",
+    "seq": 5,
+    "event_id": "2b68e60c216b40d4a5278c97e5cd6710a7799f2af7ca7dddbd33cd28ca416d32",
+    "parent_id": "d5fffff3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T18:00:03.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0a0b0c0d0e0f10111213141516171819202122ab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "toolu_nomarker_fmt",
+      "ok": true,
+      "output": "The file /Users/dev/acme/api/src/handler.rs has been updated."
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/no_commitment_marker.jsonl",
+      "byte_offset": 1135,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-nomarker-009",
+    "seq": 6,
+    "event_id": "18b9ef56fb1ddf29da302fe4ff9ae87e39291712a4d634a22c5aa50e8fef4ff1",
+    "parent_id": "d5fffff3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T18:00:03.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0a0b0c0d0e0f10111213141516171819202122ab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "toolu_nomarker_fmt",
+      "diff": {
+        "path": "/Users/dev/acme/api/src/handler.rs",
+        "old": "    let body = read(); ",
+        "new": "    let body = read();",
+        "unified": "-    let body = read(); \n+    let body = read();",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/no_commitment_marker.jsonl",
+      "byte_offset": 1135,
+      "line_no": 3
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__rejected_alternative.snap
new file mode 100644
index 0000000..944bba3
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__rejected_alternative.snap
@@ -0,0 +1,245 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-reject-002",
+    "seq": 0,
+    "event_id": "d807c123098e115d7a263e94e914e84cb23020ec52a0011156965f43b1457275",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:30:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/web",
+      "repo_root": null,
+      "git": {
+        "sha": "0f1e2d3c4b5a69788796a5b4c3d2e1f00f1e2d3c",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme/web",
+      "git": {
+        "sha": "0f1e2d3c4b5a69788796a5b4c3d2e1f00f1e2d3c",
+        "branch": "feat/payments"
+      },
+      "model": null,
+      "tool_version": "2.0.5"
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rejected_alternative.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-reject-002",
+    "seq": 1,
+    "event_id": "aaaaaaa1-0000-4000-8000-000000000001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:30:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/web",
+      "repo_root": null,
+      "git": {
+        "sha": "0f1e2d3c4b5a69788796a5b4c3d2e1f00f1e2d3c",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "We'll go with Stripe instead of PayPal for the checkout flow.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "We'll go with Stripe instead of PayPal for the checkout flow."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rejected_alternative.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-reject-002",
+    "seq": 2,
+    "event_id": "aaaaaaa2-0000-4000-8000-000000000002",
+    "parent_id": "aaaaaaa1-0000-4000-8000-000000000001",
+    "timestamp": "2026-06-22T11:30:04.5Z",
+    "project": {
+      "cwd": "/Users/dev/acme/web",
+      "repo_root": null,
+      "git": {
+        "sha": "0f1e2d3c4b5a69788796a5b4c3d2e1f00f1e2d3c",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Understood. I'll wire the Stripe client into the checkout handler.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 310,
+        "output_tokens": 72,
+        "cache_read_tokens": 0,
+        "cache_creation_tokens": 210
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Understood. I'll wire the Stripe client into the checkout handler."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_stripe_edit",
+          "name": "Write",
+          "args": {
+            "file_path": "/Users/dev/acme/web/src/checkout/provider.ts",
+            "content": "export { stripe as paymentProvider } from './stripe';\n"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rejected_alternative.jsonl",
+      "byte_offset": 395,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-reject-002",
+    "seq": 3,
+    "event_id": "896e82e3555ad7e5d102d107a9ef7f3255af79db45cc2805a6b9caea3d191fde",
+    "parent_id": "aaaaaaa2-0000-4000-8000-000000000002",
+    "timestamp": "2026-06-22T11:30:04.5Z",
+    "project": {
+      "cwd": "/Users/dev/acme/web",
+      "repo_root": null,
+      "git": {
+        "sha": "0f1e2d3c4b5a69788796a5b4c3d2e1f00f1e2d3c",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_stripe_edit",
+      "name": "Write",
+      "args": {
+        "file_path": "/Users/dev/acme/web/src/checkout/provider.ts",
+        "content": "export { stripe as paymentProvider } from './stripe';\n"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rejected_alternative.jsonl",
+      "byte_offset": 395,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-reject-002",
+    "seq": 4,
+    "event_id": "aaaaaaa3-0000-4000-8000-000000000003",
+    "parent_id": "aaaaaaa2-0000-4000-8000-000000000002",
+    "timestamp": "2026-06-22T11:30:04.9Z",
+    "project": {
+      "cwd": "/Users/dev/acme/web",
+      "repo_root": null,
+      "git": {
+        "sha": "0f1e2d3c4b5a69788796a5b4c3d2e1f00f1e2d3c",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "",
+      "parts": [
+        {
+          "part": "tool_result",
+          "call_id": "toolu_stripe_edit",
+          "output": "File created successfully at: /Users/dev/acme/web/src/checkout/provider.ts"
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rejected_alternative.jsonl",
+      "byte_offset": 1145,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-reject-002",
+    "seq": 5,
+    "event_id": "530c11d4ddb3c1751e197bd422144309e21894df01521fa0f01c7e39d1e450d4",
+    "parent_id": "aaaaaaa3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T11:30:04.9Z",
+    "project": {
+      "cwd": "/Users/dev/acme/web",
+      "repo_root": null,
+      "git": {
+        "sha": "0f1e2d3c4b5a69788796a5b4c3d2e1f00f1e2d3c",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "toolu_stripe_edit",
+      "ok": true,
+      "output": "File created successfully at: /Users/dev/acme/web/src/checkout/provider.ts"
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rejected_alternative.jsonl",
+      "byte_offset": 1145,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-reject-002",
+    "seq": 6,
+    "event_id": "86682f1c1a7b186f9608e4887b5ae045dddc0b220c33c525588b9077b2d784ed",
+    "parent_id": "aaaaaaa3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T11:30:04.9Z",
+    "project": {
+      "cwd": "/Users/dev/acme/web",
+      "repo_root": null,
+      "git": {
+        "sha": "0f1e2d3c4b5a69788796a5b4c3d2e1f00f1e2d3c",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "toolu_stripe_edit",
+      "diff": {
+        "path": "/Users/dev/acme/web/src/checkout/provider.ts",
+        "old": "",
+        "new": "export { stripe as paymentProvider } from './stripe';\n",
+        "unified": "+export { stripe as paymentProvider } from './stripe';",
+        "added_lines": 1,
+        "removed_lines": 0
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rejected_alternative.jsonl",
+      "byte_offset": 1145,
+      "line_no": 3
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__rewind_compaction.snap b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__rewind_compaction.snap
new file mode 100644
index 0000000..598212b
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__rewind_compaction.snap
@@ -0,0 +1,487 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-rewind-007",
+    "seq": 0,
+    "event_id": "b391cb2cf51db7d594ac1fcf022fde9cc85fb47886fa0a9e723e7c88615e6af9",
+    "parent_id": null,
+    "timestamp": "2026-06-22T16:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme/api",
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "2.0.5"
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-rewind-007",
+    "seq": 1,
+    "event_id": "d3ccccc1-0000-4000-8000-000000000001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T16:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use rustls for the TLS layer in the gateway.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's use rustls for the TLS layer in the gateway."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-rewind-007",
+    "seq": 2,
+    "event_id": "d3ccccc2-0000-4000-8000-000000000002",
+    "parent_id": "d3ccccc1-0000-4000-8000-000000000001",
+    "timestamp": "2026-06-22T16:00:03Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Pointing the gateway TLS layer at rustls.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 300,
+        "output_tokens": 60,
+        "cache_read_tokens": 0,
+        "cache_creation_tokens": 110
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Pointing the gateway TLS layer at rustls."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_rewind_rustls",
+          "name": "Edit",
+          "args": {
+            "file_path": "/Users/dev/acme/api/src/gateway/tls.rs",
+            "old_string": "use native_tls::TlsConnector;",
+            "new_string": "use rustls::ClientConfig;"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+      "byte_offset": 375,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-rewind-007",
+    "seq": 3,
+    "event_id": "8a5c3dbd85949394fbcbabbc3ecc79d3ab9439257e5b020fb1dc6b3c31fea761",
+    "parent_id": "d3ccccc2-0000-4000-8000-000000000002",
+    "timestamp": "2026-06-22T16:00:03Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_rewind_rustls",
+      "name": "Edit",
+      "args": {
+        "file_path": "/Users/dev/acme/api/src/gateway/tls.rs",
+        "old_string": "use native_tls::TlsConnector;",
+        "new_string": "use rustls::ClientConfig;"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+      "byte_offset": 375,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-rewind-007",
+    "seq": 4,
+    "event_id": "d3ccccc3-0000-4000-8000-000000000003",
+    "parent_id": "d3ccccc2-0000-4000-8000-000000000002",
+    "timestamp": "2026-06-22T16:00:03.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "",
+      "parts": [
+        {
+          "part": "tool_result",
+          "call_id": "toolu_rewind_rustls",
+          "output": "The file /Users/dev/acme/api/src/gateway/tls.rs has been updated."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+      "byte_offset": 1104,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-rewind-007",
+    "seq": 5,
+    "event_id": "a73e4c896a7b430909c81178667ecfa4866848062ea6420c8e6636ce335183db",
+    "parent_id": "d3ccccc3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T16:00:03.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "toolu_rewind_rustls",
+      "ok": true,
+      "output": "The file /Users/dev/acme/api/src/gateway/tls.rs has been updated."
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+      "byte_offset": 1104,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-rewind-007",
+    "seq": 6,
+    "event_id": "7d74a8b92b65a8c5b05afbaa6bfe24eb5d6275316222f8bcc6b336ac68e42ad1",
+    "parent_id": "d3ccccc3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T16:00:03.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "toolu_rewind_rustls",
+      "diff": {
+        "path": "/Users/dev/acme/api/src/gateway/tls.rs",
+        "old": "use native_tls::TlsConnector;",
+        "new": "use rustls::ClientConfig;",
+        "unified": "-use native_tls::TlsConnector;\n+use rustls::ClientConfig;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+      "byte_offset": 1104,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-rewind-007",
+    "seq": 7,
+    "event_id": "8d385489e1af193ff19ad3748689a39c404bc7f3e5987c9c09a905668014fb9e",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "unknown",
+      "raw_type": "summary",
+      "raw": {
+        "type": "summary",
+        "summary": "The user chose rustls for the gateway TLS layer and the connector import was updated. Conversation compacted to free context.",
+        "leafUuid": "d3ccccc3-0000-4000-8000-000000000003"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+      "byte_offset": 1921,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-rewind-007",
+    "seq": 8,
+    "event_id": "d3ccccc5-0000-4000-8000-000000000005",
+    "parent_id": "d3ccccc3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T16:05:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "On reflection, let's switch to native-tls after all for the gateway to match the rest of the fleet.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "On reflection, let's switch to native-tls after all for the gateway to match the rest of the fleet."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+      "byte_offset": 2128,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-rewind-007",
+    "seq": 9,
+    "event_id": "d3ccccc6-0000-4000-8000-000000000006",
+    "parent_id": "d3ccccc5-0000-4000-8000-000000000005",
+    "timestamp": "2026-06-22T16:05:04Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Reverting the gateway TLS layer back to native-tls.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 350,
+        "output_tokens": 70,
+        "cache_read_tokens": 0,
+        "cache_creation_tokens": 140
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Reverting the gateway TLS layer back to native-tls."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_rewind_native",
+          "name": "Edit",
+          "args": {
+            "file_path": "/Users/dev/acme/api/src/gateway/tls.rs",
+            "old_string": "use rustls::ClientConfig;",
+            "new_string": "use native_tls::TlsConnector;"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+      "byte_offset": 2534,
+      "line_no": 6
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-rewind-007",
+    "seq": 10,
+    "event_id": "78a838103a7e67f73cfeef3ab21c3a0a657827f94b1555081ac0e166d0b6a081",
+    "parent_id": "d3ccccc6-0000-4000-8000-000000000006",
+    "timestamp": "2026-06-22T16:05:04Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_rewind_native",
+      "name": "Edit",
+      "args": {
+        "file_path": "/Users/dev/acme/api/src/gateway/tls.rs",
+        "old_string": "use rustls::ClientConfig;",
+        "new_string": "use native_tls::TlsConnector;"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+      "byte_offset": 2534,
+      "line_no": 6
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-rewind-007",
+    "seq": 11,
+    "event_id": "d3ccccc7-0000-4000-8000-000000000007",
+    "parent_id": "d3ccccc6-0000-4000-8000-000000000006",
+    "timestamp": "2026-06-22T16:05:04.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "",
+      "parts": [
+        {
+          "part": "tool_result",
+          "call_id": "toolu_rewind_native",
+          "output": "The file /Users/dev/acme/api/src/gateway/tls.rs has been updated."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+      "byte_offset": 3273,
+      "line_no": 7
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-rewind-007",
+    "seq": 12,
+    "event_id": "bb9b8b960f8fcb07bed41d97b399a9d5056ae027b6d0f4b249e509b0448d0b9c",
+    "parent_id": "d3ccccc7-0000-4000-8000-000000000007",
+    "timestamp": "2026-06-22T16:05:04.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "toolu_rewind_native",
+      "ok": true,
+      "output": "The file /Users/dev/acme/api/src/gateway/tls.rs has been updated."
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+      "byte_offset": 3273,
+      "line_no": 7
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-rewind-007",
+    "seq": 13,
+    "event_id": "528920ad3e5d73535a1183741c180665f17901d9dc903114de2fec6cb7da79d4",
+    "parent_id": "d3ccccc7-0000-4000-8000-000000000007",
+    "timestamp": "2026-06-22T16:05:04.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "9999888877776666555544443333222211110000",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "toolu_rewind_native",
+      "diff": {
+        "path": "/Users/dev/acme/api/src/gateway/tls.rs",
+        "old": "use rustls::ClientConfig;",
+        "new": "use native_tls::TlsConnector;",
+        "unified": "-use rustls::ClientConfig;\n+use native_tls::TlsConnector;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+      "byte_offset": 3273,
+      "line_no": 7
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__subagent_thread.snap b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__subagent_thread.snap
new file mode 100644
index 0000000..de299fe
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__subagent_thread.snap
@@ -0,0 +1,399 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-main-008",
+    "seq": 0,
+    "event_id": "f7756c9f370ddf7742449e3a7c9d38836fab85c735a603e29769a0447241316e",
+    "parent_id": null,
+    "timestamp": "2026-06-22T17:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "abcdefab12345678abcdefab12345678abcdefab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme/api",
+      "git": {
+        "sha": "abcdefab12345678abcdefab12345678abcdefab",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "2.0.5"
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/subagent_thread.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-main-008",
+    "seq": 1,
+    "event_id": "d4ddddd1-0000-4000-8000-000000000001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T17:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "abcdefab12345678abcdefab12345678abcdefab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use a dedicated search subagent to migrate the indexer to tantivy.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's use a dedicated search subagent to migrate the indexer to tantivy."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/subagent_thread.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-main-008",
+    "seq": 2,
+    "event_id": "d4ddddd2-0000-4000-8000-000000000002",
+    "parent_id": "d4ddddd1-0000-4000-8000-000000000001",
+    "timestamp": "2026-06-22T17:00:03Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "abcdefab12345678abcdefab12345678abcdefab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Dispatching a search subagent to handle the indexer migration.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 260,
+        "output_tokens": 55,
+        "cache_read_tokens": 0,
+        "cache_creation_tokens": 90
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Dispatching a search subagent to handle the indexer migration."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_dispatch_subagent",
+          "name": "Task",
+          "args": {
+            "description": "Migrate indexer to tantivy",
+            "subagent_type": "search"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/subagent_thread.jsonl",
+      "byte_offset": 395,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-main-008",
+    "seq": 3,
+    "event_id": "ffb4da23ea65520882fa0ea995d97b4a08b6cd3cfdddf667579aa803b3de5d30",
+    "parent_id": "d4ddddd2-0000-4000-8000-000000000002",
+    "timestamp": "2026-06-22T17:00:03Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "abcdefab12345678abcdefab12345678abcdefab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_dispatch_subagent",
+      "name": "Task",
+      "args": {
+        "description": "Migrate indexer to tantivy",
+        "subagent_type": "search"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/subagent_thread.jsonl",
+      "byte_offset": 395,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-main-008",
+    "seq": 4,
+    "event_id": "d4eeeee1-0000-4000-8000-000000000011",
+    "parent_id": null,
+    "timestamp": "2026-06-22T17:00:05Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "abcdefab12345678abcdefab12345678abcdefab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's adopt tantivy as the search index backend in the indexer module.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's adopt tantivy as the search index backend in the indexer module."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/subagent_thread.jsonl",
+      "byte_offset": 1075,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-main-008",
+    "seq": 5,
+    "event_id": "d4eeeee2-0000-4000-8000-000000000012",
+    "parent_id": "d4eeeee1-0000-4000-8000-000000000011",
+    "timestamp": "2026-06-22T17:00:08Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "abcdefab12345678abcdefab12345678abcdefab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Swapping the indexer backend to tantivy.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 280,
+        "output_tokens": 65,
+        "cache_read_tokens": 0,
+        "cache_creation_tokens": 100
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Swapping the indexer backend to tantivy."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_sub_indexer",
+          "name": "Edit",
+          "args": {
+            "file_path": "/Users/dev/acme/api/src/indexer.rs",
+            "old_string": "use tantivy_lite::Index;",
+            "new_string": "use tantivy::Index;"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/subagent_thread.jsonl",
+      "byte_offset": 1472,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-main-008",
+    "seq": 6,
+    "event_id": "37a2b025bf9334d617bfcd15ef9c4f4d71c6f3681ada947401666150380b5311",
+    "parent_id": "d4eeeee2-0000-4000-8000-000000000012",
+    "timestamp": "2026-06-22T17:00:08Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "abcdefab12345678abcdefab12345678abcdefab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_sub_indexer",
+      "name": "Edit",
+      "args": {
+        "file_path": "/Users/dev/acme/api/src/indexer.rs",
+        "old_string": "use tantivy_lite::Index;",
+        "new_string": "use tantivy::Index;"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/subagent_thread.jsonl",
+      "byte_offset": 1472,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-main-008",
+    "seq": 7,
+    "event_id": "d4eeeee3-0000-4000-8000-000000000013",
+    "parent_id": "d4eeeee2-0000-4000-8000-000000000012",
+    "timestamp": "2026-06-22T17:00:08.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "abcdefab12345678abcdefab12345678abcdefab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "",
+      "parts": [
+        {
+          "part": "tool_result",
+          "call_id": "toolu_sub_indexer",
+          "output": "The file /Users/dev/acme/api/src/indexer.rs has been updated."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/subagent_thread.jsonl",
+      "byte_offset": 2185,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-main-008",
+    "seq": 8,
+    "event_id": "122aeba483f743e488f76e18554afe7dd8c9780e15b11528d6a61d7b08d05aeb",
+    "parent_id": "d4eeeee3-0000-4000-8000-000000000013",
+    "timestamp": "2026-06-22T17:00:08.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "abcdefab12345678abcdefab12345678abcdefab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "toolu_sub_indexer",
+      "ok": true,
+      "output": "The file /Users/dev/acme/api/src/indexer.rs has been updated."
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/subagent_thread.jsonl",
+      "byte_offset": 2185,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-main-008",
+    "seq": 9,
+    "event_id": "c7efff91f683749911e13fa88c774818eeab2ba921d0fbc1b4f65632ee6e22fc",
+    "parent_id": "d4eeeee3-0000-4000-8000-000000000013",
+    "timestamp": "2026-06-22T17:00:08.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "abcdefab12345678abcdefab12345678abcdefab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "toolu_sub_indexer",
+      "diff": {
+        "path": "/Users/dev/acme/api/src/indexer.rs",
+        "old": "use tantivy_lite::Index;",
+        "new": "use tantivy::Index;",
+        "unified": "-use tantivy_lite::Index;\n+use tantivy::Index;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/subagent_thread.jsonl",
+      "byte_offset": 2185,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-main-008",
+    "seq": 10,
+    "event_id": "d4ddddd3-0000-4000-8000-000000000003",
+    "parent_id": "d4ddddd2-0000-4000-8000-000000000002",
+    "timestamp": "2026-06-22T17:01:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "abcdefab12345678abcdefab12345678abcdefab",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "The search subagent finished the tantivy migration in the indexer.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 320,
+        "output_tokens": 40,
+        "cache_read_tokens": 260,
+        "cache_creation_tokens": 0
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "The search subagent finished the tantivy migration in the indexer."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/subagent_thread.jsonl",
+      "byte_offset": 2967,
+      "line_no": 6
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__tool_failure.snap
new file mode 100644
index 0000000..a91d802
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_0__tool_failure.snap
@@ -0,0 +1,247 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-fail-004",
+    "seq": 0,
+    "event_id": "28085cd8047ba49caf4c43e715e2c58af169629bf9bdc661165ddb573f24d0c1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:45:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "feedface00000000feedface00000000feedface",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme/api",
+      "git": {
+        "sha": "feedface00000000feedface00000000feedface",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "2.0.5"
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/tool_failure.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-fail-004",
+    "seq": 1,
+    "event_id": "ccccccc1-0000-4000-8000-000000000001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:45:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "feedface00000000feedface00000000feedface",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use a connection pool in the orders repository.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's use a connection pool in the orders repository."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/tool_failure.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-fail-004",
+    "seq": 2,
+    "event_id": "ccccccc2-0000-4000-8000-000000000002",
+    "parent_id": "ccccccc1-0000-4000-8000-000000000001",
+    "timestamp": "2026-06-22T13:45:03Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "feedface00000000feedface00000000feedface",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "I'll add the pool to the repository module.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 270,
+        "output_tokens": 80,
+        "cache_read_tokens": 0,
+        "cache_creation_tokens": 0
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "I'll add the pool to the repository module."
+        },
+        {
+          "part": "tool_use",
+          "call_id": "toolu_fail_edit",
+          "name": "Edit",
+          "args": {
+            "file_path": "/Users/dev/acme/api/src/orders/repo.rs",
+            "old_string": "let conn = Connection::new();",
+            "new_string": "let conn = pool.get()?;"
+          }
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/tool_failure.jsonl",
+      "byte_offset": 376,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-fail-004",
+    "seq": 3,
+    "event_id": "0406757d8b9aa666a1818db67009daa652893c5d2946f62154114f86dd1cfd58",
+    "parent_id": "ccccccc2-0000-4000-8000-000000000002",
+    "timestamp": "2026-06-22T13:45:03Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "feedface00000000feedface00000000feedface",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "toolu_fail_edit",
+      "name": "Edit",
+      "args": {
+        "file_path": "/Users/dev/acme/api/src/orders/repo.rs",
+        "old_string": "let conn = Connection::new();",
+        "new_string": "let conn = pool.get()?;"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/tool_failure.jsonl",
+      "byte_offset": 376,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-fail-004",
+    "seq": 4,
+    "event_id": "ccccccc3-0000-4000-8000-000000000003",
+    "parent_id": "ccccccc2-0000-4000-8000-000000000002",
+    "timestamp": "2026-06-22T13:45:03.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "feedface00000000feedface00000000feedface",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "",
+      "parts": [
+        {
+          "part": "tool_result",
+          "call_id": "toolu_fail_edit",
+          "output": "Error: String to replace not found in file. The string 'let conn = Connection::new();' does not appear in /Users/dev/acme/api/src/orders/repo.rs."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/tool_failure.jsonl",
+      "byte_offset": 1097,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-fail-004",
+    "seq": 5,
+    "event_id": "347d0ed2c43c443afcbbeaa0ce20d58a734305ccaa64065232e1a96394d34200",
+    "parent_id": "ccccccc3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T13:45:03.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "feedface00000000feedface00000000feedface",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "toolu_fail_edit",
+      "ok": false,
+      "output": "Error: String to replace not found in file. The string 'let conn = Connection::new();' does not appear in /Users/dev/acme/api/src/orders/repo.rs."
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/tool_failure.jsonl",
+      "byte_offset": 1097,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-fail-004",
+    "seq": 6,
+    "event_id": "2574a1c617ead0d1ef60771f17d099e5dafde107aafad122fff2bf6657c320e9",
+    "parent_id": "ccccccc3-0000-4000-8000-000000000003",
+    "timestamp": "2026-06-22T13:45:03.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "feedface00000000feedface00000000feedface",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "toolu_fail_edit",
+      "diff": {
+        "path": "/Users/dev/acme/api/src/orders/repo.rs",
+        "old": "let conn = Connection::new();",
+        "new": "let conn = pool.get()?;",
+        "unified": "-    let conn = Connection::new();\n+    let conn = pool.get()?;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.0/tool_failure.jsonl",
+      "byte_offset": 1097,
+      "line_no": 3
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_1__version_tolerance.snap b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_1__version_tolerance.snap
new file mode 100644
index 0000000..3c36d13
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__claude_code__2_1__version_tolerance.snap
@@ -0,0 +1,180 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-v21-001",
+    "seq": 0,
+    "event_id": "8708f20fad4b5b81aa235ab722e5bba2e0438c431146ee860a8c390b8acbedc5",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme/api",
+      "git": {
+        "sha": "",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "2.1.0"
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.1/version_tolerance.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-v21-001",
+    "seq": 1,
+    "event_id": "21111111-1111-4111-8111-111111111111",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use Postgres instead of MySQL for the orders service.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's use Postgres instead of MySQL for the orders service."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.1/version_tolerance.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-v21-001",
+    "seq": 2,
+    "event_id": "22222221-2222-4222-8222-222222222222",
+    "parent_id": "21111111-1111-4111-8111-111111111111",
+    "timestamp": "2026-06-22T11:00:03.12Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Switching the orders service to Postgres now.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 482,
+        "output_tokens": 96,
+        "cache_read_tokens": 120,
+        "cache_creation_tokens": 40
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Switching the orders service to Postgres now."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.1/version_tolerance.jsonl",
+      "byte_offset": 444,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-v21-001",
+    "seq": 3,
+    "event_id": "6d2a8432dfe95c9015787e8aae177e5c79f876d7361641235f3d67a9975e8a2a",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:00:03.48Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "unknown",
+      "raw_type": "telemetry",
+      "raw": {
+        "type": "telemetry",
+        "uuid": "23333331-3333-4333-8333-333333333333",
+        "parentUuid": "22222221-2222-4222-8222-222222222222",
+        "timestamp": "2026-06-22T11:00:03.480Z",
+        "sessionId": "sess-v21-001",
+        "version": "2.1.0",
+        "payload": {
+          "kind": "latency",
+          "ms": 1200
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.1/version_tolerance.jsonl",
+      "byte_offset": 988,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "claude_code",
+    "session_id": "sess-v21-001",
+    "seq": 4,
+    "event_id": "24444441-4444-4444-8444-444444444444",
+    "parent_id": "22222221-2222-4222-8222-222222222222",
+    "timestamp": "2026-06-22T11:00:07.21Z",
+    "project": {
+      "cwd": "/Users/dev/acme/api",
+      "repo_root": null,
+      "git": {
+        "sha": "",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Looks good, ship it.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Looks good, ship it."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/claude_code/2.1/version_tolerance.jsonl",
+      "byte_offset": 1230,
+      "line_no": 4
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__codex__v1__version_tolerance.snap b/crates/memscribe-testkit/tests/snapshots/events__codex__v1__version_tolerance.snap
new file mode 100644
index 0000000..f9a276e
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__codex__v1__version_tolerance.snap
@@ -0,0 +1,135 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0001legacy",
+    "seq": 0,
+    "event_id": "session_meta:01JZ8codexsess0001legacy",
+    "parent_id": null,
+    "timestamp": "2026-06-22T09:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme-api",
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "0.1.0"
+    },
+    "provenance": {
+      "file": "fixtures/codex/v1/version_tolerance.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0001legacy",
+    "seq": 1,
+    "event_id": "8a9e7f9763d69f931e5992d22ffa877274ab85c5eab6e6ffd32a91e008c9533f",
+    "parent_id": null,
+    "timestamp": "2026-06-22T09:00:05.12Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "unknown",
+      "raw_type": "unknown",
+      "raw": {
+        "timestamp": "2026-06-22T09:00:05.120Z",
+        "record_type": "input",
+        "role": "user",
+        "text": "Let's use Postgres instead of MySQL for the persistence layer."
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v1/version_tolerance.jsonl",
+      "byte_offset": 258,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0001legacy",
+    "seq": 2,
+    "event_id": "f774cfc5829730294915b0f7e7f2427aa89b140c17eaa6c9038783f2fb865d66",
+    "parent_id": null,
+    "timestamp": "2026-06-22T09:00:09.88Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "unknown",
+      "raw_type": "unknown",
+      "raw": {
+        "timestamp": "2026-06-22T09:00:09.880Z",
+        "record_type": "output",
+        "role": "assistant",
+        "text": "Switching to Postgres."
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v1/version_tolerance.jsonl",
+      "byte_offset": 407,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0001legacy",
+    "seq": 3,
+    "event_id": "faebbccbcb7656e1b0d9d6f228ccb7cc4d617198913e24977dcaef237e8d56f3",
+    "parent_id": null,
+    "timestamp": "2026-06-22T09:00:12.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "unknown",
+      "raw_type": "unknown",
+      "raw": {
+        "timestamp": "2026-06-22T09:00:12.400Z",
+        "kind": "state",
+        "payload": {
+          "phase": "awaiting_input"
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v1/version_tolerance.jsonl",
+      "byte_offset": 522,
+      "line_no": 4
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__codex__v2__ban.snap b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__ban.snap
new file mode 100644
index 0000000..3550f08
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__ban.snap
@@ -0,0 +1,208 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0003ban",
+    "seq": 0,
+    "event_id": "session_meta:01JZ8codexsess0003ban",
+    "parent_id": null,
+    "timestamp": "2026-06-22T16:20:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "c3d4e5f60718293a4b5c6d7e8f90010203040506",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme-api",
+      "git": {
+        "sha": "c3d4e5f60718293a4b5c6d7e8f90010203040506",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "0.6.2"
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/ban.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0003ban",
+    "seq": 1,
+    "event_id": "60cfe394ca6c0aaa7a76a91648594ae96d383367677989cea6b5ad6b9c5cb2fd",
+    "parent_id": null,
+    "timestamp": "2026-06-22T16:20:07.33Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "c3d4e5f60718293a4b5c6d7e8f90010203040506",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Policy: we will never add a dependency on left-pad. Do not introduce it under any circumstances, even transitively if you can avoid it.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Policy: we will never add a dependency on left-pad. Do not introduce it under any circumstances, even transitively if you can avoid it."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/ban.jsonl",
+      "byte_offset": 255,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0003ban",
+    "seq": 2,
+    "event_id": "fc8db1a0f1d355dc38c71a6fd9437e63157739c6b474a6887bd92a4ee57e2b8e",
+    "parent_id": null,
+    "timestamp": "2026-06-22T16:20:12.77Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "c3d4e5f60718293a4b5c6d7e8f90010203040506",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Acknowledged. I will never add left-pad as a dependency.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Acknowledged. I will never add left-pad as a dependency."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/ban.jsonl",
+      "byte_offset": 541,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0003ban",
+    "seq": 3,
+    "event_id": "call_check_001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T16:20:40.9Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "c3d4e5f60718293a4b5c6d7e8f90010203040506",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_check_001",
+      "name": "shell",
+      "args": {
+        "command": [
+          "rg",
+          "left-pad",
+          "package.json",
+          "Cargo.toml"
+        ]
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/ban.jsonl",
+      "byte_offset": 754,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0003ban",
+    "seq": 4,
+    "event_id": "call_check_001:output",
+    "parent_id": "call_check_001",
+    "timestamp": "2026-06-22T16:20:41.22Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "c3d4e5f60718293a4b5c6d7e8f90010203040506",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_check_001",
+      "ok": false,
+      "output": {
+        "output": "",
+        "metadata": {
+          "exit_code": 1
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/ban.jsonl",
+      "byte_offset": 977,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0003ban",
+    "seq": 5,
+    "event_id": "4df8c878f6537a8f6ebfe2847cb6a9e7604f73c374f15b94cf5072adb64353f1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T16:20:44Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "c3d4e5f60718293a4b5c6d7e8f90010203040506",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Confirmed: left-pad is not present in the manifests.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Confirmed: left-pad is not present in the manifests."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/ban.jsonl",
+      "byte_offset": 1160,
+      "line_no": 6
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__codex__v2__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..ce90561
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__happy_path_decision_then_edits.snap
@@ -0,0 +1,343 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0001happypath",
+    "seq": 0,
+    "event_id": "session_meta:01JZ8codexsess0001happypath",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "feat/storage-layer"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme-api",
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "feat/storage-layer"
+      },
+      "model": null,
+      "tool_version": "0.6.2"
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0001happypath",
+    "seq": 1,
+    "event_id": "44f62eea08bdb9079bb6d8ce8efdc2d14d9005fb1d04c8cc86b4bb5c517affef",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:00:05.12Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "feat/storage-layer"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use Postgres instead of MySQL for the persistence layer. Swap the driver and the connection URL.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's use Postgres instead of MySQL for the persistence layer. Swap the driver and the connection URL."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 275,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0001happypath",
+    "seq": 2,
+    "event_id": "1df61243852bfd57e7e6b29dafade4c4b7372fc86b7cc150932a09ccc095c34f",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:00:09.88Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "feat/storage-layer"
+      }
+    },
+    "kind": {
+      "event": "unknown",
+      "raw_type": "response_item",
+      "raw": {
+        "timestamp": "2026-06-22T14:00:09.880Z",
+        "type": "response_item",
+        "payload": {
+          "type": "reasoning",
+          "summary": [
+            "The user decided on Postgres. I'll update the Cargo deps and the connection string."
+          ]
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 528,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0001happypath",
+    "seq": 3,
+    "event_id": "eabfd13ca8aff0186a54e7ca801698a6f392fc558e8440607b4e981d871df742",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:00:12.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "feat/storage-layer"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Switching to Postgres. I'll update Cargo.toml and the database module.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Switching to Postgres. I'll update Cargo.toml and the database module."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 721,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0001happypath",
+    "seq": 4,
+    "event_id": "call_apply_patch_001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:00:18.64Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "feat/storage-layer"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_apply_patch_001",
+      "name": "apply_patch",
+      "args": {
+        "input": "*** Begin Patch\n*** Update File: Cargo.toml\n@@\n-mysql = \"24\"\n+sqlx = { version = \"0.7\", features = [\"postgres\", \"runtime-tokio\"] }\n*** Update File: src/db.rs\n@@\n-let url = \"mysql://localhost:3306/acme\";\n-let pool = mysql::Pool::new(url)?;\n+let url = \"postgres://localhost:5432/acme\";\n+let pool = sqlx::PgPool::connect(url).await?;\n*** End Patch\n"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 948,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0001happypath",
+    "seq": 5,
+    "event_id": "b95d795c08871aafd5a6dc9ee0a39dcd9742868bb515858c02bb4a8b23e7cd2c",
+    "parent_id": "call_apply_patch_001",
+    "timestamp": "2026-06-22T14:00:18.64Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "feat/storage-layer"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_apply_patch_001",
+      "diff": {
+        "path": "Cargo.toml",
+        "old": null,
+        "new": null,
+        "unified": "@@\n-mysql = \"24\"\n+sqlx = { version = \"0.7\", features = [\"postgres\", \"runtime-tokio\"] }",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 948,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0001happypath",
+    "seq": 6,
+    "event_id": "912a555fdfd6ef8ca9d60f8b60251d6270fe7dea12f749bf8bda1122b68adb07",
+    "parent_id": "call_apply_patch_001",
+    "timestamp": "2026-06-22T14:00:18.64Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "feat/storage-layer"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_apply_patch_001",
+      "diff": {
+        "path": "src/db.rs",
+        "old": null,
+        "new": null,
+        "unified": "@@\n-let url = \"mysql://localhost:3306/acme\";\n-let pool = mysql::Pool::new(url)?;\n+let url = \"postgres://localhost:5432/acme\";\n+let pool = sqlx::PgPool::connect(url).await?;",
+        "added_lines": 2,
+        "removed_lines": 2
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 948,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0001happypath",
+    "seq": 7,
+    "event_id": "call_apply_patch_001:output",
+    "parent_id": "call_apply_patch_001",
+    "timestamp": "2026-06-22T14:00:19.01Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "feat/storage-layer"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_apply_patch_001",
+      "ok": true,
+      "output": {
+        "output": "Applied patch to 2 files: Cargo.toml, src/db.rs",
+        "metadata": {
+          "exit_code": 0,
+          "duration_ms": 42
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1537,
+      "line_no": 6
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0001happypath",
+    "seq": 8,
+    "event_id": "7a127d992906e8773bdde097d47b4a320acc727f19ff1a9f7641e987fdbd535c",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:00:25.3Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "feat/storage-layer"
+      }
+    },
+    "kind": {
+      "event": "unknown",
+      "raw_type": "event_msg",
+      "raw": {
+        "timestamp": "2026-06-22T14:00:25.300Z",
+        "type": "event_msg",
+        "payload": {
+          "type": "token_count",
+          "input_tokens": 1840,
+          "output_tokens": 260
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1790,
+      "line_no": 7
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0001happypath",
+    "seq": 9,
+    "event_id": "cd85e04b441bb669401e9e985b8d47f3a9d8d45e63a0fc0642b52a14f24dd923",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:00:26Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+        "branch": "feat/storage-layer"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Done. The persistence layer now uses Postgres via sqlx.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Done. The persistence layer now uses Postgres via sqlx."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1923,
+      "line_no": 8
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__codex__v2__interleaved_arcs.snap b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__interleaved_arcs.snap
new file mode 100644
index 0000000..7d49531
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__interleaved_arcs.snap
@@ -0,0 +1,506 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 0,
+    "event_id": "session_meta:01JZ8codexsess0005interleaved",
+    "parent_id": null,
+    "timestamp": "2026-06-22T18:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme-api",
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      },
+      "model": null,
+      "tool_version": "0.6.2"
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 1,
+    "event_id": "3a9543e8ab68ed9d503da81b7b3df749f66082bbef5063caab3e35a400e02c57",
+    "parent_id": null,
+    "timestamp": "2026-06-22T18:00:06.1Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use argon2 for password hashing. Update the auth module and the shared crypto helper accordingly.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's use argon2 for password hashing. Update the auth module and the shared crypto helper accordingly."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 277,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 2,
+    "event_id": "0cf556d8c3ac63ee2fbd39c361907134f97138cbf77f8a7ef4016c616b1335df",
+    "parent_id": null,
+    "timestamp": "2026-06-22T18:00:10.3Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Switching the hasher to argon2 in the auth module and the shared crypto helper.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Switching the hasher to argon2 in the auth module and the shared crypto helper."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 531,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 3,
+    "event_id": "call_apply_patch_arc_a",
+    "parent_id": null,
+    "timestamp": "2026-06-22T18:00:16.45Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_apply_patch_arc_a",
+      "name": "apply_patch",
+      "args": {
+        "input": "*** Begin Patch\n*** Update File: src/auth.rs\n@@\n-use bcrypt::hash;\n+use argon2::Argon2;\n*** Update File: src/shared/crypto.rs\n@@\n-pub const HASH_COST: u32 = 12;\n+pub const ARGON2_MEM_KIB: u32 = 19456;\n*** End Patch\n"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 767,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 4,
+    "event_id": "a0e26c4acdd223389282b50baa8c7931d5889707cddb86bc0bd6974aaa82289c",
+    "parent_id": "call_apply_patch_arc_a",
+    "timestamp": "2026-06-22T18:00:16.45Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_apply_patch_arc_a",
+      "diff": {
+        "path": "src/auth.rs",
+        "old": null,
+        "new": null,
+        "unified": "@@\n-use bcrypt::hash;\n+use argon2::Argon2;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 767,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 5,
+    "event_id": "803f07edb1bb1408291fed4e5b7f792b1c466b76d34919831479378b62e54e45",
+    "parent_id": "call_apply_patch_arc_a",
+    "timestamp": "2026-06-22T18:00:16.45Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_apply_patch_arc_a",
+      "diff": {
+        "path": "src/shared/crypto.rs",
+        "old": null,
+        "new": null,
+        "unified": "@@\n-pub const HASH_COST: u32 = 12;\n+pub const ARGON2_MEM_KIB: u32 = 19456;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 767,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 6,
+    "event_id": "call_apply_patch_arc_a:output",
+    "parent_id": "call_apply_patch_arc_a",
+    "timestamp": "2026-06-22T18:00:16.7Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_apply_patch_arc_a",
+      "ok": true,
+      "output": {
+        "output": "Applied patch to 2 files: src/auth.rs, src/shared/crypto.rs",
+        "metadata": {
+          "exit_code": 0,
+          "duration_ms": 38
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 1188,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 7,
+    "event_id": "26e81aff3f0f096fefea314b9f51a29757da810801b900a5139dc0afdccc56e0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T18:00:24Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "argon2 is now wired into auth and the shared crypto helper.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "argon2 is now wired into auth and the shared crypto helper."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 1455,
+      "line_no": 6
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 8,
+    "event_id": "baf0eaeeb14764e809ad5d4e6a29ce866a63a459e54e201bf7968415a4175bfa",
+    "parent_id": null,
+    "timestamp": "2026-06-22T18:00:40.2Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Now let's switch to a 30-minute session TTL. Update the session manager and bump the shared default in the crypto helper's config block.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Now let's switch to a 30-minute session TTL. Update the session manager and bump the shared default in the crypto helper's config block."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 1671,
+      "line_no": 7
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 9,
+    "event_id": "c1e4f1509eb39a2d703f2c2b2ea780ec4ad23171bb2219e2eab82d603498c0c6",
+    "parent_id": null,
+    "timestamp": "2026-06-22T18:00:45.6Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Setting the session TTL to 30 minutes in the session manager and the shared config default.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Setting the session TTL to 30 minutes in the session manager and the shared config default."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 1958,
+      "line_no": 8
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 10,
+    "event_id": "call_apply_patch_arc_b",
+    "parent_id": null,
+    "timestamp": "2026-06-22T18:00:52.8Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_apply_patch_arc_b",
+      "name": "apply_patch",
+      "args": {
+        "input": "*** Begin Patch\n*** Update File: src/session.rs\n@@\n-const SESSION_TTL_SECS: u64 = 3600;\n+const SESSION_TTL_SECS: u64 = 1800;\n*** Update File: src/shared/crypto.rs\n@@\n-pub const DEFAULT_TTL_SECS: u64 = 3600;\n+pub const DEFAULT_TTL_SECS: u64 = 1800;\n*** End Patch\n"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 2206,
+      "line_no": 9
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 11,
+    "event_id": "d0e90e7be5b17e3c9ca5e5bb7b97bcb72a28cf9693d1301d67738d6654081d07",
+    "parent_id": "call_apply_patch_arc_b",
+    "timestamp": "2026-06-22T18:00:52.8Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_apply_patch_arc_b",
+      "diff": {
+        "path": "src/session.rs",
+        "old": null,
+        "new": null,
+        "unified": "@@\n-const SESSION_TTL_SECS: u64 = 3600;\n+const SESSION_TTL_SECS: u64 = 1800;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 2206,
+      "line_no": 9
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 12,
+    "event_id": "5d99ef3371bc0b1b0eddb09ebb51affbc9e6c879e567c66bbf8e406bc1700a2a",
+    "parent_id": "call_apply_patch_arc_b",
+    "timestamp": "2026-06-22T18:00:52.8Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_apply_patch_arc_b",
+      "diff": {
+        "path": "src/shared/crypto.rs",
+        "old": null,
+        "new": null,
+        "unified": "@@\n-pub const DEFAULT_TTL_SECS: u64 = 3600;\n+pub const DEFAULT_TTL_SECS: u64 = 1800;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 2206,
+      "line_no": 9
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 13,
+    "event_id": "call_apply_patch_arc_b:output",
+    "parent_id": "call_apply_patch_arc_b",
+    "timestamp": "2026-06-22T18:00:53.05Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_apply_patch_arc_b",
+      "ok": true,
+      "output": {
+        "output": "Applied patch to 2 files: src/session.rs, src/shared/crypto.rs",
+        "metadata": {
+          "exit_code": 0,
+          "duration_ms": 41
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 2674,
+      "line_no": 10
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "seq": 14,
+    "event_id": "ed27b3a06441eed92458353bd4aa0cd6af913d19b0a75bd5e1d88d1226a546f4",
+    "parent_id": null,
+    "timestamp": "2026-06-22T18:00:58Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+        "branch": "feat/auth-overhaul"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Done. Session TTL is now 30 minutes in both the manager and the shared default.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Done. Session TTL is now 30 minutes in both the manager and the shared default."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+      "byte_offset": 2944,
+      "line_no": 11
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__codex__v2__multi_edit_single_commit.snap b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__multi_edit_single_commit.snap
new file mode 100644
index 0000000..3309a55
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__multi_edit_single_commit.snap
@@ -0,0 +1,342 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0006multiedit",
+    "seq": 0,
+    "event_id": "session_meta:01JZ8codexsess0006multiedit",
+    "parent_id": null,
+    "timestamp": "2026-06-22T19:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f60718293a4b5c6d7e8f90010203040506070809",
+        "branch": "feat/rename-client"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme-api",
+      "git": {
+        "sha": "f60718293a4b5c6d7e8f90010203040506070809",
+        "branch": "feat/rename-client"
+      },
+      "model": null,
+      "tool_version": "0.6.2"
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/multi_edit_single_commit.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0006multiedit",
+    "seq": 1,
+    "event_id": "4c4422382382a3aac72495c54c35c3310b1657c2faa2ba6e98c1c5b70f0c9503",
+    "parent_id": null,
+    "timestamp": "2026-06-22T19:00:07.25Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f60718293a4b5c6d7e8f90010203040506070809",
+        "branch": "feat/rename-client"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's rename the HttpClient type to ApiClient everywhere. Use ApiClient as the canonical name across the codebase.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's rename the HttpClient type to ApiClient everywhere. Use ApiClient as the canonical name across the codebase."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/multi_edit_single_commit.jsonl",
+      "byte_offset": 275,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0006multiedit",
+    "seq": 2,
+    "event_id": "2f0e6a9facde7c4b76c97b67edb4b8e9015669c2bb284c7cbd468b1b84d38313",
+    "parent_id": null,
+    "timestamp": "2026-06-22T19:00:12.9Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f60718293a4b5c6d7e8f90010203040506070809",
+        "branch": "feat/rename-client"
+      }
+    },
+    "kind": {
+      "event": "unknown",
+      "raw_type": "response_item",
+      "raw": {
+        "timestamp": "2026-06-22T19:00:12.900Z",
+        "type": "response_item",
+        "payload": {
+          "type": "reasoning",
+          "summary": [
+            "This is a single logical rename commit touching several files. I'll do it in one apply_patch."
+          ]
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/multi_edit_single_commit.jsonl",
+      "byte_offset": 540,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0006multiedit",
+    "seq": 3,
+    "event_id": "c3a571ed3694d33e37409ce3688362694d7a07d365f761f9d734b4d679cfc4cb",
+    "parent_id": null,
+    "timestamp": "2026-06-22T19:00:15.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f60718293a4b5c6d7e8f90010203040506070809",
+        "branch": "feat/rename-client"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Renaming HttpClient to ApiClient across client.rs, handlers.rs, and main.rs in one patch.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Renaming HttpClient to ApiClient across client.rs, handlers.rs, and main.rs in one patch."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/multi_edit_single_commit.jsonl",
+      "byte_offset": 743,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0006multiedit",
+    "seq": 4,
+    "event_id": "call_apply_patch_multi_001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T19:00:23.8Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f60718293a4b5c6d7e8f90010203040506070809",
+        "branch": "feat/rename-client"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_apply_patch_multi_001",
+      "name": "apply_patch",
+      "args": {
+        "input": "*** Begin Patch\n*** Update File: src/client.rs\n@@\n-pub struct HttpClient {\n+pub struct ApiClient {\n*** Update File: src/handlers.rs\n@@\n-use crate::client::HttpClient;\n+use crate::client::ApiClient;\n@@\n-    client: HttpClient,\n+    client: ApiClient,\n*** Update File: src/main.rs\n@@\n-    let client = HttpClient::new();\n+    let client = ApiClient::new();\n*** End Patch\n"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/multi_edit_single_commit.jsonl",
+      "byte_offset": 989,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0006multiedit",
+    "seq": 5,
+    "event_id": "c7c8034ea95e5b6d9d949ff28ca832f5af41d3a1d57a7b79f4d301572d3a874c",
+    "parent_id": "call_apply_patch_multi_001",
+    "timestamp": "2026-06-22T19:00:23.8Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f60718293a4b5c6d7e8f90010203040506070809",
+        "branch": "feat/rename-client"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_apply_patch_multi_001",
+      "diff": {
+        "path": "src/client.rs",
+        "old": null,
+        "new": null,
+        "unified": "@@\n-pub struct HttpClient {\n+pub struct ApiClient {",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/multi_edit_single_commit.jsonl",
+      "byte_offset": 989,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0006multiedit",
+    "seq": 6,
+    "event_id": "a78e661cf73019f97ee5377b512c4c40d698904bd85e2e908a984c800f1530a8",
+    "parent_id": "call_apply_patch_multi_001",
+    "timestamp": "2026-06-22T19:00:23.8Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f60718293a4b5c6d7e8f90010203040506070809",
+        "branch": "feat/rename-client"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_apply_patch_multi_001",
+      "diff": {
+        "path": "src/handlers.rs",
+        "old": null,
+        "new": null,
+        "unified": "@@\n-use crate::client::HttpClient;\n+use crate::client::ApiClient;\n@@\n-    client: HttpClient,\n+    client: ApiClient,",
+        "added_lines": 2,
+        "removed_lines": 2
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/multi_edit_single_commit.jsonl",
+      "byte_offset": 989,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0006multiedit",
+    "seq": 7,
+    "event_id": "532c44875f5babcab68075e2a223d64f808d2eab90fb46c50eb82e9dc8e1c9eb",
+    "parent_id": "call_apply_patch_multi_001",
+    "timestamp": "2026-06-22T19:00:23.8Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f60718293a4b5c6d7e8f90010203040506070809",
+        "branch": "feat/rename-client"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_apply_patch_multi_001",
+      "diff": {
+        "path": "src/main.rs",
+        "old": null,
+        "new": null,
+        "unified": "@@\n-    let client = HttpClient::new();\n+    let client = ApiClient::new();",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/multi_edit_single_commit.jsonl",
+      "byte_offset": 989,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0006multiedit",
+    "seq": 8,
+    "event_id": "call_apply_patch_multi_001:output",
+    "parent_id": "call_apply_patch_multi_001",
+    "timestamp": "2026-06-22T19:00:24.1Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f60718293a4b5c6d7e8f90010203040506070809",
+        "branch": "feat/rename-client"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_apply_patch_multi_001",
+      "ok": true,
+      "output": {
+        "output": "Applied patch to 3 files: src/client.rs, src/handlers.rs, src/main.rs",
+        "metadata": {
+          "exit_code": 0,
+          "duration_ms": 53
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/multi_edit_single_commit.jsonl",
+      "byte_offset": 1582,
+      "line_no": 6
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0006multiedit",
+    "seq": 9,
+    "event_id": "54481de1bb93e9cd51309f2f39ce4afa4df4f2c7bfbed36b151b608dfe458b35",
+    "parent_id": null,
+    "timestamp": "2026-06-22T19:00:30Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f60718293a4b5c6d7e8f90010203040506070809",
+        "branch": "feat/rename-client"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Done. HttpClient is now ApiClient across all three files in a single commit-sized change.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Done. HttpClient is now ApiClient across all three files in a single commit-sized change."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/multi_edit_single_commit.jsonl",
+      "byte_offset": 1863,
+      "line_no": 7
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__codex__v2__no_commitment_marker.snap b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__no_commitment_marker.snap
new file mode 100644
index 0000000..a165246
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__no_commitment_marker.snap
@@ -0,0 +1,238 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0009nomarker",
+    "seq": 0,
+    "event_id": "session_meta:01JZ8codexsess0009nomarker",
+    "parent_id": null,
+    "timestamp": "2026-06-22T22:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "293a4b5c6d7e8f90010203040506070809101112",
+        "branch": "chore/tidy"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme-api",
+      "git": {
+        "sha": "293a4b5c6d7e8f90010203040506070809101112",
+        "branch": "chore/tidy"
+      },
+      "model": null,
+      "tool_version": "0.6.2"
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/no_commitment_marker.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0009nomarker",
+    "seq": 1,
+    "event_id": "8427bd17b926f36de096fe9ec0f2c947b5e5c9d3e60a5676dded6e6a8aa06125",
+    "parent_id": null,
+    "timestamp": "2026-06-22T22:00:05.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "293a4b5c6d7e8f90010203040506070809101112",
+        "branch": "chore/tidy"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Thanks, this looks great. Go ahead and tidy up that stray import while you're in there.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Thanks, this looks great. Go ahead and tidy up that stray import while you're in there."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/no_commitment_marker.jsonl",
+      "byte_offset": 266,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0009nomarker",
+    "seq": 2,
+    "event_id": "3cbdc3874cda67270f46a4287c0301107b1473ce7db1cdc38e84c62478c3a9e6",
+    "parent_id": null,
+    "timestamp": "2026-06-22T22:00:10.1Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "293a4b5c6d7e8f90010203040506070809101112",
+        "branch": "chore/tidy"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Sure, removing the unused import.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Sure, removing the unused import."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/no_commitment_marker.jsonl",
+      "byte_offset": 504,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0009nomarker",
+    "seq": 3,
+    "event_id": "call_apply_patch_nomark_001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T22:00:16.7Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "293a4b5c6d7e8f90010203040506070809101112",
+        "branch": "chore/tidy"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_apply_patch_nomark_001",
+      "name": "apply_patch",
+      "args": {
+        "input": "*** Begin Patch\n*** Update File: src/handlers.rs\n@@\n-use std::collections::HashMap;\n use std::collections::BTreeMap;\n*** End Patch\n"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/no_commitment_marker.jsonl",
+      "byte_offset": 694,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0009nomarker",
+    "seq": 4,
+    "event_id": "8ef38fd78b06b1090ccda79dd885ee48367107b7cf758dffa5c3cf6dd6177335",
+    "parent_id": "call_apply_patch_nomark_001",
+    "timestamp": "2026-06-22T22:00:16.7Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "293a4b5c6d7e8f90010203040506070809101112",
+        "branch": "chore/tidy"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_apply_patch_nomark_001",
+      "diff": {
+        "path": "src/handlers.rs",
+        "old": null,
+        "new": null,
+        "unified": "@@\n-use std::collections::HashMap;\n use std::collections::BTreeMap;",
+        "added_lines": 0,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/no_commitment_marker.jsonl",
+      "byte_offset": 694,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0009nomarker",
+    "seq": 5,
+    "event_id": "call_apply_patch_nomark_001:output",
+    "parent_id": "call_apply_patch_nomark_001",
+    "timestamp": "2026-06-22T22:00:16.95Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "293a4b5c6d7e8f90010203040506070809101112",
+        "branch": "chore/tidy"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_apply_patch_nomark_001",
+      "ok": true,
+      "output": {
+        "output": "Applied patch to 1 file: src/handlers.rs",
+        "metadata": {
+          "exit_code": 0,
+          "duration_ms": 21
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/no_commitment_marker.jsonl",
+      "byte_offset": 1028,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0009nomarker",
+    "seq": 6,
+    "event_id": "fbfc46416ccf1d37161895fbd96998c80528e2781ef82a32efeef53ac3d6c3d5",
+    "parent_id": null,
+    "timestamp": "2026-06-22T22:00:21Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "293a4b5c6d7e8f90010203040506070809101112",
+        "branch": "chore/tidy"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Removed the stray HashMap import.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Removed the stray HashMap import."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/no_commitment_marker.jsonl",
+      "byte_offset": 1281,
+      "line_no": 6
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__codex__v2__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__rejected_alternative.snap
new file mode 100644
index 0000000..49d515a
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__rejected_alternative.snap
@@ -0,0 +1,245 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0002rejected",
+    "seq": 0,
+    "event_id": "session_meta:01JZ8codexsess0002rejected",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:10:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "b2c3d4e5f60718293a4b5c6d7e8f900102030405",
+        "branch": "feat/api-shape"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme-api",
+      "git": {
+        "sha": "b2c3d4e5f60718293a4b5c6d7e8f900102030405",
+        "branch": "feat/api-shape"
+      },
+      "model": null,
+      "tool_version": "0.6.2"
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rejected_alternative.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0002rejected",
+    "seq": 1,
+    "event_id": "4cc3a5a14d57aef39b3895570a423315b21225f45615db2678e544457c35c08c",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:10:08.22Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "b2c3d4e5f60718293a4b5c6d7e8f900102030405",
+        "branch": "feat/api-shape"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "We considered GraphQL for the public API, but let's reject that and stick with REST. GraphQL adds caching and schema-evolution overhead we don't want right now.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "We considered GraphQL for the public API, but let's reject that and stick with REST. GraphQL adds caching and schema-evolution overhead we don't want right now."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rejected_alternative.jsonl",
+      "byte_offset": 270,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0002rejected",
+    "seq": 2,
+    "event_id": "3ce6b4af1624b655cabaf85186396254f498f8f9479214779aa838d954bcd5be",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:10:14.5Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "b2c3d4e5f60718293a4b5c6d7e8f900102030405",
+        "branch": "feat/api-shape"
+      }
+    },
+    "kind": {
+      "event": "unknown",
+      "raw_type": "response_item",
+      "raw": {
+        "timestamp": "2026-06-22T15:10:14.500Z",
+        "type": "response_item",
+        "payload": {
+          "type": "reasoning",
+          "summary": [
+            "User rejected GraphQL in favor of REST. No code change requested beyond confirming the decision."
+          ]
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rejected_alternative.jsonl",
+      "byte_offset": 581,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0002rejected",
+    "seq": 3,
+    "event_id": "9a01123691b704c3e1385c04c8590204e819d5829e09502bc32ac08cbb23dcd3",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:10:16.9Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "b2c3d4e5f60718293a4b5c6d7e8f900102030405",
+        "branch": "feat/api-shape"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Understood. We'll keep the REST design and drop the GraphQL exploration.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Understood. We'll keep the REST design and drop the GraphQL exploration."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rejected_alternative.jsonl",
+      "byte_offset": 787,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0002rejected",
+    "seq": 4,
+    "event_id": "call_grep_001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:10:30.1Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "b2c3d4e5f60718293a4b5c6d7e8f900102030405",
+        "branch": "feat/api-shape"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_grep_001",
+      "name": "shell",
+      "args": {
+        "command": [
+          "grep",
+          "-r",
+          "graphql",
+          "src/"
+        ]
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rejected_alternative.jsonl",
+      "byte_offset": 1016,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0002rejected",
+    "seq": 5,
+    "event_id": "call_grep_001:output",
+    "parent_id": "call_grep_001",
+    "timestamp": "2026-06-22T15:10:30.45Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "b2c3d4e5f60718293a4b5c6d7e8f900102030405",
+        "branch": "feat/api-shape"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_grep_001",
+      "ok": false,
+      "output": {
+        "output": "",
+        "metadata": {
+          "exit_code": 1,
+          "duration_ms": 18
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rejected_alternative.jsonl",
+      "byte_offset": 1223,
+      "line_no": 6
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0002rejected",
+    "seq": 6,
+    "event_id": "d747cb47466f8632d7838f19a8d0a462ba6e7c5542d40b585a75e4e30a5dc2a5",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:10:33.7Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "b2c3d4e5f60718293a4b5c6d7e8f900102030405",
+        "branch": "feat/api-shape"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "No GraphQL references exist in the codebase, so nothing to remove.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "No GraphQL references exist in the codebase, so nothing to remove."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rejected_alternative.jsonl",
+      "byte_offset": 1422,
+      "line_no": 7
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__codex__v2__rewind_compaction.snap b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__rewind_compaction.snap
new file mode 100644
index 0000000..f79b72b
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__rewind_compaction.snap
@@ -0,0 +1,378 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0007rewindcompact",
+    "seq": 0,
+    "event_id": "session_meta:01JZ8codexsess0007rewindcompact",
+    "parent_id": null,
+    "timestamp": "2026-06-22T20:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0718293a4b5c6d7e8f9001020304050607080910",
+        "branch": "feat/persistence-pivot"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme-api",
+      "git": {
+        "sha": "0718293a4b5c6d7e8f9001020304050607080910",
+        "branch": "feat/persistence-pivot"
+      },
+      "model": null,
+      "tool_version": "0.6.2"
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rewind_compaction.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0007rewindcompact",
+    "seq": 1,
+    "event_id": "af025c062b71e782f169243c8043a4886b78c35e3733948a94bf11fbbef2bef9",
+    "parent_id": null,
+    "timestamp": "2026-06-22T20:00:06.3Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0718293a4b5c6d7e8f9001020304050607080910",
+        "branch": "feat/persistence-pivot"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use MongoDB for the document store. Set up the driver and the connection config.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's use MongoDB for the document store. Set up the driver and the connection config."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rewind_compaction.jsonl",
+      "byte_offset": 283,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0007rewindcompact",
+    "seq": 2,
+    "event_id": "addc21e1f941adcf5be1ffceb57cbd0c01536079dafd82918ec11c1b64418914",
+    "parent_id": null,
+    "timestamp": "2026-06-22T20:00:11.7Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0718293a4b5c6d7e8f9001020304050607080910",
+        "branch": "feat/persistence-pivot"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Understood, I'll wire up MongoDB for the document store.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Understood, I'll wire up MongoDB for the document store."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rewind_compaction.jsonl",
+      "byte_offset": 520,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0007rewindcompact",
+    "seq": 3,
+    "event_id": "e72ff9d6c487fbc532610e0e072afeb42130cda02a0e2bdb0d841196b715cf84",
+    "parent_id": null,
+    "timestamp": "2026-06-22T20:00:30Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0718293a4b5c6d7e8f9001020304050607080910",
+        "branch": "feat/persistence-pivot"
+      }
+    },
+    "kind": {
+      "event": "unknown",
+      "raw_type": "event_msg",
+      "raw": {
+        "timestamp": "2026-06-22T20:00:30.000Z",
+        "type": "event_msg",
+        "payload": {
+          "type": "compacted",
+          "reason": "auto_compact",
+          "message": "Conversation history compacted to fit the context window. The MongoDB exploration above was summarized and the verbatim early turns were dropped from the working context.",
+          "replaced_response_ids": [
+            "msg-mongo-user",
+            "msg-mongo-assistant"
+          ]
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rewind_compaction.jsonl",
+      "byte_offset": 733,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0007rewindcompact",
+    "seq": 4,
+    "event_id": "b27e2518664e1c5b4d0b31d9546ce984031f843158f292857b626cf2bbfcdaae",
+    "parent_id": null,
+    "timestamp": "2026-06-22T20:00:31Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0718293a4b5c6d7e8f9001020304050607080910",
+        "branch": "feat/persistence-pivot"
+      }
+    },
+    "kind": {
+      "event": "unknown",
+      "raw_type": "turn_context",
+      "raw": {
+        "timestamp": "2026-06-22T20:00:31.000Z",
+        "type": "turn_context",
+        "payload": {
+          "summary": "Earlier the user explored MongoDB for the document store; that turn was compacted out of the active context.",
+          "cwd": "/Users/dev/acme-api"
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rewind_compaction.jsonl",
+      "byte_offset": 1096,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0007rewindcompact",
+    "seq": 5,
+    "event_id": "42174043fd82823604e1d3d21305cb37e8514aa685387e04fb425e445c217c18",
+    "parent_id": null,
+    "timestamp": "2026-06-22T20:00:40.2Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0718293a4b5c6d7e8f9001020304050607080910",
+        "branch": "feat/persistence-pivot"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Actually let's use Postgres with JSONB for the document store instead. Switch to Postgres and drop the MongoDB plan.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Actually let's use Postgres with JSONB for the document store instead. Switch to Postgres and drop the MongoDB plan."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rewind_compaction.jsonl",
+      "byte_offset": 1320,
+      "line_no": 6
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0007rewindcompact",
+    "seq": 6,
+    "event_id": "9285d994a655a6d45fd0b220a02c6bb65917130f7cde43d9a9472de376c6a0c4",
+    "parent_id": null,
+    "timestamp": "2026-06-22T20:00:45.9Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0718293a4b5c6d7e8f9001020304050607080910",
+        "branch": "feat/persistence-pivot"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Switching the document store to Postgres with JSONB columns.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Switching the document store to Postgres with JSONB columns."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rewind_compaction.jsonl",
+      "byte_offset": 1587,
+      "line_no": 7
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0007rewindcompact",
+    "seq": 7,
+    "event_id": "call_apply_patch_rewind_001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T20:00:53.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0718293a4b5c6d7e8f9001020304050607080910",
+        "branch": "feat/persistence-pivot"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_apply_patch_rewind_001",
+      "name": "apply_patch",
+      "args": {
+        "input": "*** Begin Patch\n*** Update File: src/store.rs\n@@\n-use mongodb::Client;\n+use sqlx::PgPool;\n@@\n-    let client = Client::with_uri_str(\"mongodb://localhost:27017\").await?;\n+    let pool = PgPool::connect(\"postgres://localhost:5432/acme\").await?;\n*** End Patch\n"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rewind_compaction.jsonl",
+      "byte_offset": 1804,
+      "line_no": 8
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0007rewindcompact",
+    "seq": 8,
+    "event_id": "db5adb3aa4e0469ccf0f56928b9e00f8244ceab7e40f5a258608725c1c82cd44",
+    "parent_id": "call_apply_patch_rewind_001",
+    "timestamp": "2026-06-22T20:00:53.4Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0718293a4b5c6d7e8f9001020304050607080910",
+        "branch": "feat/persistence-pivot"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_apply_patch_rewind_001",
+      "diff": {
+        "path": "src/store.rs",
+        "old": null,
+        "new": null,
+        "unified": "@@\n-use mongodb::Client;\n+use sqlx::PgPool;\n@@\n-    let client = Client::with_uri_str(\"mongodb://localhost:27017\").await?;\n+    let pool = PgPool::connect(\"postgres://localhost:5432/acme\").await?;",
+        "added_lines": 2,
+        "removed_lines": 2
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rewind_compaction.jsonl",
+      "byte_offset": 1804,
+      "line_no": 8
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0007rewindcompact",
+    "seq": 9,
+    "event_id": "call_apply_patch_rewind_001:output",
+    "parent_id": "call_apply_patch_rewind_001",
+    "timestamp": "2026-06-22T20:00:53.7Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0718293a4b5c6d7e8f9001020304050607080910",
+        "branch": "feat/persistence-pivot"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_apply_patch_rewind_001",
+      "ok": true,
+      "output": {
+        "output": "Applied patch to 1 file: src/store.rs",
+        "metadata": {
+          "exit_code": 0,
+          "duration_ms": 36
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rewind_compaction.jsonl",
+      "byte_offset": 2282,
+      "line_no": 9
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0007rewindcompact",
+    "seq": 10,
+    "event_id": "f59c29634cb1b5cfb532af0db58b5d37c91a212cf5805c1552ecb837351e4907",
+    "parent_id": null,
+    "timestamp": "2026-06-22T20:00:58Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0718293a4b5c6d7e8f9001020304050607080910",
+        "branch": "feat/persistence-pivot"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Done. The document store now uses Postgres with JSONB.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Done. The document store now uses Postgres with JSONB."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/rewind_compaction.jsonl",
+      "byte_offset": 2532,
+      "line_no": 10
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__codex__v2__subagent_thread.snap b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__subagent_thread.snap
new file mode 100644
index 0000000..3b144e5
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__subagent_thread.snap
@@ -0,0 +1,307 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0008subagent",
+    "seq": 0,
+    "event_id": "session_meta:01JZ8codexsess0008subagent",
+    "parent_id": null,
+    "timestamp": "2026-06-22T21:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "18293a4b5c6d7e8f900102030405060708091011",
+        "branch": "feat/worker-split"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme-api",
+      "git": {
+        "sha": "18293a4b5c6d7e8f900102030405060708091011",
+        "branch": "feat/worker-split"
+      },
+      "model": null,
+      "tool_version": "0.6.2"
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/subagent_thread.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0008subagent",
+    "seq": 1,
+    "event_id": "a0394bf72b7f9078fa0e05bad27253d6fb473d12a11841f6ac2e1ba5c8923e27",
+    "parent_id": null,
+    "timestamp": "2026-06-22T21:00:06.5Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "18293a4b5c6d7e8f900102030405060708091011",
+        "branch": "feat/worker-split"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use a dedicated background worker for the email queue. Extract the send loop into a worker module.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's use a dedicated background worker for the email queue. Extract the send loop into a worker module."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/subagent_thread.jsonl",
+      "byte_offset": 273,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0008subagent",
+    "seq": 2,
+    "event_id": "799050f69f84d972c36b2b0e7a637ad4b99b0fc5386a34f0416a0a68415ab644",
+    "parent_id": null,
+    "timestamp": "2026-06-22T21:00:11.8Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "18293a4b5c6d7e8f900102030405060708091011",
+        "branch": "feat/worker-split"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "I'll extract the send loop into a dedicated worker module.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "I'll extract the send loop into a dedicated worker module."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/subagent_thread.jsonl",
+      "byte_offset": 528,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0008subagent",
+    "seq": 3,
+    "event_id": "df3c77013575ccaa14b9500d127f476b6e8be60caf3d96093f960ad1e0723140",
+    "parent_id": null,
+    "timestamp": "2026-06-22T21:00:14Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "18293a4b5c6d7e8f900102030405060708091011",
+        "branch": "feat/worker-split"
+      }
+    },
+    "kind": {
+      "event": "unknown",
+      "raw_type": "turn_context",
+      "raw": {
+        "timestamp": "2026-06-22T21:00:14.000Z",
+        "type": "turn_context",
+        "payload": {
+          "thread_id": "thread-email-worker",
+          "parent_thread_id": "01JZ8codexsess0008subagent",
+          "summary": "Nested task thread for the email worker extraction. Codex has no first-class subagent; this is a secondary thread id carried in turn_context."
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/subagent_thread.jsonl",
+      "byte_offset": 743,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0008subagent",
+    "seq": 4,
+    "event_id": "call_apply_patch_sub_001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T21:00:22.3Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "18293a4b5c6d7e8f900102030405060708091011",
+        "branch": "feat/worker-split"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_apply_patch_sub_001",
+      "name": "apply_patch",
+      "args": {
+        "input": "*** Begin Patch\n*** Add File: src/worker/email.rs\n+pub async fn run_email_worker() {\n+    loop { drain_email_queue().await; }\n+}\n*** Update File: src/main.rs\n@@\n-    send_emails_inline();\n+    tokio::spawn(worker::email::run_email_worker());\n*** End Patch\n"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/subagent_thread.jsonl",
+      "byte_offset": 1054,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0008subagent",
+    "seq": 5,
+    "event_id": "091aea6456fde260cda2dbeb1891a25c746c0e32b07c7a14a94c2d033a288e27",
+    "parent_id": "call_apply_patch_sub_001",
+    "timestamp": "2026-06-22T21:00:22.3Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "18293a4b5c6d7e8f900102030405060708091011",
+        "branch": "feat/worker-split"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_apply_patch_sub_001",
+      "diff": {
+        "path": "src/worker/email.rs",
+        "old": null,
+        "new": null,
+        "unified": "+pub async fn run_email_worker() {\n+    loop { drain_email_queue().await; }\n+}",
+        "added_lines": 3,
+        "removed_lines": 0
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/subagent_thread.jsonl",
+      "byte_offset": 1054,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0008subagent",
+    "seq": 6,
+    "event_id": "98298da2eba010bc3a26b90a25051691e033a2aebd3f49a44d88ac26be87535b",
+    "parent_id": "call_apply_patch_sub_001",
+    "timestamp": "2026-06-22T21:00:22.3Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "18293a4b5c6d7e8f900102030405060708091011",
+        "branch": "feat/worker-split"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_apply_patch_sub_001",
+      "diff": {
+        "path": "src/main.rs",
+        "old": null,
+        "new": null,
+        "unified": "@@\n-    send_emails_inline();\n+    tokio::spawn(worker::email::run_email_worker());",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/subagent_thread.jsonl",
+      "byte_offset": 1054,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0008subagent",
+    "seq": 7,
+    "event_id": "call_apply_patch_sub_001:output",
+    "parent_id": "call_apply_patch_sub_001",
+    "timestamp": "2026-06-22T21:00:22.6Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "18293a4b5c6d7e8f900102030405060708091011",
+        "branch": "feat/worker-split"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_apply_patch_sub_001",
+      "ok": true,
+      "output": {
+        "output": "Applied patch to 2 files: src/worker/email.rs, src/main.rs",
+        "metadata": {
+          "exit_code": 0,
+          "duration_ms": 44
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/subagent_thread.jsonl",
+      "byte_offset": 1518,
+      "line_no": 6
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0008subagent",
+    "seq": 8,
+    "event_id": "8c8565566e6b4bebb1e675c4ab955cab7a09d0be0fca2c85aadaa7662b7e2b04",
+    "parent_id": null,
+    "timestamp": "2026-06-22T21:00:28Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "18293a4b5c6d7e8f900102030405060708091011",
+        "branch": "feat/worker-split"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Done. The email send loop now runs in a dedicated worker module spawned from main.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "Done. The email send loop now runs in a dedicated worker module spawned from main."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/subagent_thread.jsonl",
+      "byte_offset": 1786,
+      "line_no": 7
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__codex__v2__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__tool_failure.snap
new file mode 100644
index 0000000..0d2ccff
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__codex__v2__tool_failure.snap
@@ -0,0 +1,238 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0004toolfail",
+    "seq": 0,
+    "event_id": "session_meta:01JZ8codexsess0004toolfail",
+    "parent_id": null,
+    "timestamp": "2026-06-22T17:05:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "d4e5f60718293a4b5c6d7e8f9001020304050607",
+        "branch": "fix/config-loader"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme-api",
+      "git": {
+        "sha": "d4e5f60718293a4b5c6d7e8f9001020304050607",
+        "branch": "fix/config-loader"
+      },
+      "model": null,
+      "tool_version": "0.6.2"
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/tool_failure.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0004toolfail",
+    "seq": 1,
+    "event_id": "dc572a23e3174d126ae33ccdf71ce7f52af5276037f1c84a8bc51210b031263b",
+    "parent_id": null,
+    "timestamp": "2026-06-22T17:05:06.41Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "d4e5f60718293a4b5c6d7e8f9001020304050607",
+        "branch": "fix/config-loader"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Switch the config loader to read from config.toml instead of config.yaml.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Switch the config loader to read from config.toml instead of config.yaml."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/tool_failure.jsonl",
+      "byte_offset": 273,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0004toolfail",
+    "seq": 2,
+    "event_id": "ce147e972f19728adb38a24bd4f737c41e866b99e69dae6a187658827e2ecd92",
+    "parent_id": null,
+    "timestamp": "2026-06-22T17:05:11.9Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "d4e5f60718293a4b5c6d7e8f9001020304050607",
+        "branch": "fix/config-loader"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "I'll update the loader to point at config.toml.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "I'll update the loader to point at config.toml."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/tool_failure.jsonl",
+      "byte_offset": 497,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0004toolfail",
+    "seq": 3,
+    "event_id": "call_apply_patch_fail_001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T17:05:18.22Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "d4e5f60718293a4b5c6d7e8f9001020304050607",
+        "branch": "fix/config-loader"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_apply_patch_fail_001",
+      "name": "apply_patch",
+      "args": {
+        "input": "*** Begin Patch\n*** Update File: src/config.rs\n@@\n-let path = \"config.yaml\";\n+let path = \"config.toml\";\n*** End Patch\n"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/tool_failure.jsonl",
+      "byte_offset": 701,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0004toolfail",
+    "seq": 4,
+    "event_id": "c50dc0619d2743e40efdfaee3b103fe8a36632bd4f57face44e5a6bcdac91e82",
+    "parent_id": "call_apply_patch_fail_001",
+    "timestamp": "2026-06-22T17:05:18.22Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "d4e5f60718293a4b5c6d7e8f9001020304050607",
+        "branch": "fix/config-loader"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_apply_patch_fail_001",
+      "diff": {
+        "path": "src/config.rs",
+        "old": null,
+        "new": null,
+        "unified": "@@\n-let path = \"config.yaml\";\n+let path = \"config.toml\";",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/tool_failure.jsonl",
+      "byte_offset": 701,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0004toolfail",
+    "seq": 5,
+    "event_id": "call_apply_patch_fail_001:output",
+    "parent_id": "call_apply_patch_fail_001",
+    "timestamp": "2026-06-22T17:05:18.56Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "d4e5f60718293a4b5c6d7e8f9001020304050607",
+        "branch": "fix/config-loader"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_apply_patch_fail_001",
+      "ok": false,
+      "output": {
+        "output": "error: patch does not apply: src/config.rs: context line \"let path = \\\"config.yaml\\\";\" not found",
+        "metadata": {
+          "exit_code": 1,
+          "duration_ms": 27
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/tool_failure.jsonl",
+      "byte_offset": 1032,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "codex",
+    "session_id": "01JZ8codexsess0004toolfail",
+    "seq": 6,
+    "event_id": "d04c22b3915e9ed36493345e95e7de5eb0ae2d259bac3554a64c7a440fc30ef7",
+    "parent_id": null,
+    "timestamp": "2026-06-22T17:05:25.7Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "d4e5f60718293a4b5c6d7e8f9001020304050607",
+        "branch": "fix/config-loader"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "The patch failed to apply because the context didn't match. Let me re-read the file before retrying.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": [
+        {
+          "part": "text",
+          "text": "The patch failed to apply because the context didn't match. Let me re-read the file before retrying."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/codex/v2/tool_failure.jsonl",
+      "byte_offset": 1345,
+      "line_no": 6
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__copilot__v1__ban.snap b/crates/memscribe-testkit/tests/snapshots/events__copilot__v1__ban.snap
new file mode 100644
index 0000000..1ed823a
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__copilot__v1__ban.snap
@@ -0,0 +1,126 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-003",
+    "seq": 0,
+    "event_id": "e4f6ec5b1adc5e1f2c8ef8d65fd15c556bd7bcc56f8b47b42112e65b69cb912d",
+    "parent_id": null,
+    "timestamp": "2026-06-22T12:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/core",
+      "repo_root": null,
+      "git": {
+        "sha": "beadfeed9001",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/projects/core",
+      "git": {
+        "sha": "beadfeed9001",
+        "branch": "main"
+      },
+      "model": "gpt-4o",
+      "tool_version": "copilot-cli 1.4.0"
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/ban.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-003",
+    "seq": 1,
+    "event_id": "msg-1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T12:00:11Z",
+    "project": {
+      "cwd": "/Users/dev/projects/core",
+      "repo_root": null,
+      "git": {
+        "sha": "beadfeed9001",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "We will never add a dependency on left-pad. Do not introduce it under any circumstances.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/ban.jsonl",
+      "byte_offset": 215,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-003",
+    "seq": 2,
+    "event_id": "msg-2",
+    "parent_id": "msg-1",
+    "timestamp": "2026-06-22T12:00:16Z",
+    "project": {
+      "cwd": "/Users/dev/projects/core",
+      "repo_root": null,
+      "git": {
+        "sha": "beadfeed9001",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Acknowledged. I won't add left-pad; I'll inline a small padding helper instead.",
+      "thinking": null,
+      "model": "gpt-4o",
+      "usage": {
+        "input_tokens": 640,
+        "output_tokens": 120,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/ban.jsonl",
+      "byte_offset": 419,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-003",
+    "seq": 3,
+    "event_id": "410bfd80ed4be1266fec0d8bf94388faa287cb388794958c6ceb6abb3d2e54d2",
+    "parent_id": null,
+    "timestamp": "2026-06-22T12:00:50Z",
+    "project": {
+      "cwd": "/Users/dev/projects/core",
+      "repo_root": null,
+      "git": {
+        "sha": "beadfeed9001",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_end",
+      "reason": "user_closed"
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/ban.jsonl",
+      "byte_offset": 674,
+      "line_no": 4
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__copilot__v1__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/events__copilot__v1__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..174fd9f
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__copilot__v1__happy_path_decision_then_edits.snap
@@ -0,0 +1,221 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-001",
+    "seq": 0,
+    "event_id": "375265a8449751cea57ef19c191458299045da00883ff3c6c796bc1fd40a1135",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/projects/orbit",
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      },
+      "model": "gpt-4o",
+      "tool_version": "copilot-cli 1.4.0"
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-001",
+    "seq": 1,
+    "event_id": "msg-1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:05Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use Postgres instead of MySQL for the persistence layer.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 216,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-001",
+    "seq": 2,
+    "event_id": "msg-2",
+    "parent_id": "msg-1",
+    "timestamp": "2026-06-22T10:00:09Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Good call. I'll switch the connection config and the schema module to Postgres.",
+      "thinking": null,
+      "model": "gpt-4o",
+      "usage": {
+        "input_tokens": 1200,
+        "output_tokens": 340,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 394,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-001",
+    "seq": 3,
+    "event_id": "msg-2:edit:0:src/db/config.rs",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T10:00:09Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "src/db/config.rs",
+        "old": "let url = \"mysql://localhost/orbit\";",
+        "new": "let url = \"postgres://localhost/orbit\";",
+        "unified": "@@ -1 +1 @@\n-let url = \"mysql://localhost/orbit\";\n+let url = \"postgres://localhost/orbit\";",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 394,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-001",
+    "seq": 4,
+    "event_id": "msg-2:edit:1:src/db/schema.rs",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T10:00:09Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "src/db/schema.rs",
+        "old": "// MySQL dialect\n",
+        "new": "// Postgres dialect\nuse sqlx::postgres::PgPool;\n",
+        "unified": null,
+        "added_lines": 2,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 394,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-001",
+    "seq": 5,
+    "event_id": "msg-3",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T10:01:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Looks right, thanks.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1065,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-001",
+    "seq": 6,
+    "event_id": "49ba239a7c5289521d9533277d60ea535ddb39cb92b3b7b7477d410f2326aae9",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:01:30Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_end",
+      "reason": "user_closed"
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1204,
+      "line_no": 5
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__copilot__v1__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/events__copilot__v1__rejected_alternative.snap
new file mode 100644
index 0000000..a6827d5
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__copilot__v1__rejected_alternative.snap
@@ -0,0 +1,245 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-002",
+    "seq": 0,
+    "event_id": "581518c238196ba553d0c09adc5331811d219219050e1a0bbb90c2a46c158414",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/checkout",
+      "repo_root": null,
+      "git": {
+        "sha": "f00dcafe1234",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/projects/checkout",
+      "git": {
+        "sha": "f00dcafe1234",
+        "branch": "feat/payments"
+      },
+      "model": "gpt-4o",
+      "tool_version": "copilot-cli 1.4.0"
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/rejected_alternative.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-002",
+    "seq": 1,
+    "event_id": "msg-1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:00:08Z",
+    "project": {
+      "cwd": "/Users/dev/projects/checkout",
+      "repo_root": null,
+      "git": {
+        "sha": "f00dcafe1234",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use Stripe instead of PayPal for the payment provider.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/rejected_alternative.jsonl",
+      "byte_offset": 228,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-002",
+    "seq": 2,
+    "event_id": "msg-2",
+    "parent_id": "msg-1",
+    "timestamp": "2026-06-22T11:00:14Z",
+    "project": {
+      "cwd": "/Users/dev/projects/checkout",
+      "repo_root": null,
+      "git": {
+        "sha": "f00dcafe1234",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Understood. I'll wire up the Stripe SDK and drop the PayPal client.",
+      "thinking": null,
+      "model": "gpt-4o",
+      "usage": {
+        "input_tokens": 980,
+        "output_tokens": 210,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/rejected_alternative.jsonl",
+      "byte_offset": 404,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-002",
+    "seq": 3,
+    "event_id": "msg-2:call:call-a1",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T11:00:14Z",
+    "project": {
+      "cwd": "/Users/dev/projects/checkout",
+      "repo_root": null,
+      "git": {
+        "sha": "f00dcafe1234",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-a1",
+      "name": "apply_patch",
+      "args": {
+        "path": "src/payments/provider.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/rejected_alternative.jsonl",
+      "byte_offset": 404,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-002",
+    "seq": 4,
+    "event_id": "msg-2:result:call-a1",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T11:00:14Z",
+    "project": {
+      "cwd": "/Users/dev/projects/checkout",
+      "repo_root": null,
+      "git": {
+        "sha": "f00dcafe1234",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-a1",
+      "ok": true,
+      "output": "patch applied (1 file changed)"
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/rejected_alternative.jsonl",
+      "byte_offset": 404,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-002",
+    "seq": 5,
+    "event_id": "msg-2:edit:0:src/payments/provider.rs",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T11:00:14Z",
+    "project": {
+      "cwd": "/Users/dev/projects/checkout",
+      "repo_root": null,
+      "git": {
+        "sha": "f00dcafe1234",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-a1",
+      "diff": {
+        "path": "src/payments/provider.rs",
+        "old": "use paypal_sdk::Client;",
+        "new": "use stripe::Client;",
+        "unified": "@@ -1 +1 @@\n-use paypal_sdk::Client;\n+use stripe::Client;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/rejected_alternative.jsonl",
+      "byte_offset": 404,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-002",
+    "seq": 6,
+    "event_id": "msg-3",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T11:01:20Z",
+    "project": {
+      "cwd": "/Users/dev/projects/checkout",
+      "repo_root": null,
+      "git": {
+        "sha": "f00dcafe1234",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Perfect.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/rejected_alternative.jsonl",
+      "byte_offset": 1051,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-002",
+    "seq": 7,
+    "event_id": "9691e9aeae7c817e723481665f6c429992006c9b4953a5ab9f687e0f6a8db66d",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:01:40Z",
+    "project": {
+      "cwd": "/Users/dev/projects/checkout",
+      "repo_root": null,
+      "git": {
+        "sha": "f00dcafe1234",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "session_end",
+      "reason": "user_closed"
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/rejected_alternative.jsonl",
+      "byte_offset": 1178,
+      "line_no": 5
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__copilot__v1__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/events__copilot__v1__tool_failure.snap
new file mode 100644
index 0000000..fc2e7db
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__copilot__v1__tool_failure.snap
@@ -0,0 +1,245 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-004",
+    "seq": 0,
+    "event_id": "bbed3938fda757d9109de017f2ac87f79b2361e5b46f31793415b4cf6f5e5bb2",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0badc0de7777",
+        "branch": "fix/migration"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/projects/api",
+      "git": {
+        "sha": "0badc0de7777",
+        "branch": "fix/migration"
+      },
+      "model": "gpt-4o",
+      "tool_version": "copilot-cli 1.4.0"
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/tool_failure.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-004",
+    "seq": 1,
+    "event_id": "msg-1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:00:07Z",
+    "project": {
+      "cwd": "/Users/dev/projects/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0badc0de7777",
+        "branch": "fix/migration"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Add a NOT NULL constraint to the users.email column.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/tool_failure.jsonl",
+      "byte_offset": 223,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-004",
+    "seq": 2,
+    "event_id": "msg-2",
+    "parent_id": "msg-1",
+    "timestamp": "2026-06-22T13:00:13Z",
+    "project": {
+      "cwd": "/Users/dev/projects/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0badc0de7777",
+        "branch": "fix/migration"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "I'll apply the migration patch.",
+      "thinking": null,
+      "model": "gpt-4o",
+      "usage": {
+        "input_tokens": 720,
+        "output_tokens": 90,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/tool_failure.jsonl",
+      "byte_offset": 391,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-004",
+    "seq": 3,
+    "event_id": "msg-2:call:call-z9",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T13:00:13Z",
+    "project": {
+      "cwd": "/Users/dev/projects/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0badc0de7777",
+        "branch": "fix/migration"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-z9",
+      "name": "apply_patch",
+      "args": {
+        "path": "migrations/0007_email_not_null.sql"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/tool_failure.jsonl",
+      "byte_offset": 391,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-004",
+    "seq": 4,
+    "event_id": "msg-2:result:call-z9",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T13:00:13Z",
+    "project": {
+      "cwd": "/Users/dev/projects/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0badc0de7777",
+        "branch": "fix/migration"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-z9",
+      "ok": false,
+      "output": "error: patch did not apply cleanly: hunk #1 FAILED at line 3"
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/tool_failure.jsonl",
+      "byte_offset": 391,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-004",
+    "seq": 5,
+    "event_id": "msg-2:edit:0:migrations/0007_email_not_null.sql",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T13:00:13Z",
+    "project": {
+      "cwd": "/Users/dev/projects/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0badc0de7777",
+        "branch": "fix/migration"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-z9",
+      "diff": {
+        "path": "migrations/0007_email_not_null.sql",
+        "old": "email TEXT",
+        "new": "email TEXT NOT NULL",
+        "unified": "@@ -3 +3 @@\n-email TEXT\n+email TEXT NOT NULL",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/tool_failure.jsonl",
+      "byte_offset": 391,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-004",
+    "seq": 6,
+    "event_id": "msg-3",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T13:01:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0badc0de7777",
+        "branch": "fix/migration"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Leave it for now, I'll handle the conflict manually.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/tool_failure.jsonl",
+      "byte_offset": 1026,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "copilot",
+    "session_id": "copilot-thread-004",
+    "seq": 7,
+    "event_id": "eb9e62c623409b2d1005ea8f13d6d892c05a82982374192db59b3875e6a576a8",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:01:20Z",
+    "project": {
+      "cwd": "/Users/dev/projects/api",
+      "repo_root": null,
+      "git": {
+        "sha": "0badc0de7777",
+        "branch": "fix/migration"
+      }
+    },
+    "kind": {
+      "event": "session_end",
+      "reason": "user_closed"
+    },
+    "provenance": {
+      "file": "fixtures/copilot/v1/tool_failure.jsonl",
+      "byte_offset": 1197,
+      "line_no": 5
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__cursor__v1__ban.snap b/crates/memscribe-testkit/tests/snapshots/events__cursor__v1__ban.snap
new file mode 100644
index 0000000..b1083a2
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__cursor__v1__ban.snap
@@ -0,0 +1,202 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-003",
+    "seq": 0,
+    "event_id": "183398368baef6a254adbb482303071b26c377d8cda59bb52fcad8dc37602d9a",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/payments-core",
+      "repo_root": null,
+      "git": {
+        "sha": "a17de4490b22",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/projects/payments-core",
+      "git": {
+        "sha": "a17de4490b22",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "0.43.2"
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/ban.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-003",
+    "seq": 1,
+    "event_id": "msg-201",
+    "parent_id": null,
+    "timestamp": "2026-06-22T16:20:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/payments-core",
+      "repo_root": null,
+      "git": {
+        "sha": "a17de4490b22",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "We will never add a dependency on left-pad. Use the standard library padding instead, always.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "We will never add a dependency on left-pad. Use the standard library padding instead, always."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/ban.jsonl",
+      "byte_offset": 162,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-003",
+    "seq": 2,
+    "event_id": "msg-202",
+    "parent_id": "msg-201",
+    "timestamp": "2026-06-22T16:20:11Z",
+    "project": {
+      "cwd": "/Users/dev/projects/payments-core",
+      "repo_root": null,
+      "git": {
+        "sha": "a17de4490b22",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Understood — no left-pad dependency, ever. I'll use String::repeat / format width specifiers for any padding we need.",
+      "thinking": null,
+      "model": "cursor-fast",
+      "usage": {
+        "input_tokens": 140,
+        "output_tokens": 58,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Understood — no left-pad dependency, ever. I'll use String::repeat / format width specifiers for any padding we need."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/ban.jsonl",
+      "byte_offset": 367,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-003",
+    "seq": 3,
+    "event_id": "msg-202#toolcall:call-edit-3",
+    "parent_id": "msg-202",
+    "timestamp": "2026-06-22T16:20:11Z",
+    "project": {
+      "cwd": "/Users/dev/projects/payments-core",
+      "repo_root": null,
+      "git": {
+        "sha": "a17de4490b22",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-edit-3",
+      "name": "edit_file",
+      "args": {
+        "path": "src/format.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/ban.jsonl",
+      "byte_offset": 367,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-003",
+    "seq": 4,
+    "event_id": "msg-202#toolresult:call-edit-3",
+    "parent_id": "msg-202",
+    "timestamp": "2026-06-22T16:20:11Z",
+    "project": {
+      "cwd": "/Users/dev/projects/payments-core",
+      "repo_root": null,
+      "git": {
+        "sha": "a17de4490b22",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-edit-3",
+      "ok": true,
+      "output": "applied 1 edit"
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/ban.jsonl",
+      "byte_offset": 367,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-003",
+    "seq": 5,
+    "event_id": "msg-202#edit:0",
+    "parent_id": "msg-202",
+    "timestamp": "2026-06-22T16:20:11Z",
+    "project": {
+      "cwd": "/Users/dev/projects/payments-core",
+      "repo_root": null,
+      "git": {
+        "sha": "a17de4490b22",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "src/format.rs",
+        "old": "use left_pad::left_pad;\nlet s = left_pad(id, 8);",
+        "new": "let s = format!(\"{id:>8}\");",
+        "unified": "@@ -1,2 +1 @@\n-use left_pad::left_pad;\n-let s = left_pad(id, 8);\n+let s = format!(\"{id:>8}\");\n",
+        "added_lines": 1,
+        "removed_lines": 2
+      }
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/ban.jsonl",
+      "byte_offset": 367,
+      "line_no": 3
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__cursor__v1__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/events__cursor__v1__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..b7a2bab
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__cursor__v1__happy_path_decision_then_edits.snap
@@ -0,0 +1,360 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-001",
+    "seq": 0,
+    "event_id": "7ee7998bc4dde9e040d2d988935adf29ea005f43044b1881a44758cbd2941442",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f2c1ad4b7e0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/projects/orders-api",
+      "git": {
+        "sha": "9f2c1ad4b7e0",
+        "branch": "main"
+      },
+      "model": "cursor-fast",
+      "tool_version": "0.43.2"
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-001",
+    "seq": 1,
+    "event_id": "msg-001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:01:10Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f2c1ad4b7e0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use Postgres instead of MySQL for the orders service. Update the config and the connection module.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's use Postgres instead of MySQL for the orders service. Update the config and the connection module."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 181,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-001",
+    "seq": 2,
+    "event_id": "msg-002",
+    "parent_id": "msg-001",
+    "timestamp": "2026-06-22T14:01:18Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f2c1ad4b7e0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Good call — Postgres gives us better JSONB support. I'll switch the engine in the config and update the DSN in the connection module.",
+      "thinking": null,
+      "model": "cursor-fast",
+      "usage": {
+        "input_tokens": 214,
+        "output_tokens": 98,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Good call — Postgres gives us better JSONB support. I'll switch the engine in the config and update the DSN in the connection module."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 397,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-001",
+    "seq": 3,
+    "event_id": "msg-002#toolcall:call-edit-1",
+    "parent_id": "msg-002",
+    "timestamp": "2026-06-22T14:01:18Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f2c1ad4b7e0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-edit-1",
+      "name": "edit_file",
+      "args": {
+        "path": "config/database.toml"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 397,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-001",
+    "seq": 4,
+    "event_id": "msg-002#toolresult:call-edit-1",
+    "parent_id": "msg-002",
+    "timestamp": "2026-06-22T14:01:18Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f2c1ad4b7e0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-edit-1",
+      "ok": true,
+      "output": "applied 1 edit"
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 397,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-001",
+    "seq": 5,
+    "event_id": "msg-002#edit:0",
+    "parent_id": "msg-002",
+    "timestamp": "2026-06-22T14:01:18Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f2c1ad4b7e0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "config/database.toml",
+        "old": "[database]\nengine = \"mysql\"\nport = 3306\n",
+        "new": "[database]\nengine = \"postgres\"\nport = 5432\n",
+        "unified": "@@ -1,3 +1,3 @@\n [database]\n-engine = \"mysql\"\n-port = 3306\n+engine = \"postgres\"\n+port = 5432\n",
+        "added_lines": 2,
+        "removed_lines": 2
+      }
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 397,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-001",
+    "seq": 6,
+    "event_id": "msg-003",
+    "parent_id": "msg-002",
+    "timestamp": "2026-06-22T14:01:31Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f2c1ad4b7e0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Now updating the connection module to use the Postgres driver.",
+      "thinking": null,
+      "model": "cursor-fast",
+      "usage": {
+        "input_tokens": 120,
+        "output_tokens": 64,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Now updating the connection module to use the Postgres driver."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1172,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-001",
+    "seq": 7,
+    "event_id": "msg-003#toolcall:call-edit-2",
+    "parent_id": "msg-003",
+    "timestamp": "2026-06-22T14:01:31Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f2c1ad4b7e0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-edit-2",
+      "name": "edit_file",
+      "args": {
+        "path": "src/db/connect.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1172,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-001",
+    "seq": 8,
+    "event_id": "msg-003#toolresult:call-edit-2",
+    "parent_id": "msg-003",
+    "timestamp": "2026-06-22T14:01:31Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f2c1ad4b7e0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-edit-2",
+      "ok": true,
+      "output": "applied 1 edit"
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1172,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-001",
+    "seq": 9,
+    "event_id": "msg-003#edit:0",
+    "parent_id": "msg-003",
+    "timestamp": "2026-06-22T14:01:31Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f2c1ad4b7e0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "src/db/connect.rs",
+        "old": "let url = \"mysql://localhost:3306/orders\";",
+        "new": "let url = \"postgres://localhost:5432/orders\";",
+        "unified": "@@ -1 +1 @@\n-let url = \"mysql://localhost:3306/orders\";\n+let url = \"postgres://localhost:5432/orders\";\n",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1172,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-001",
+    "seq": 10,
+    "event_id": "1ecb1f98d7a0c61ff270e8562653c49c1c04e633abb70be4b9d428913ee5baaa",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:02:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f2c1ad4b7e0",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_end",
+      "reason": "user_closed"
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1873,
+      "line_no": 5
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__cursor__v1__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/events__cursor__v1__rejected_alternative.snap
new file mode 100644
index 0000000..e74c652
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__cursor__v1__rejected_alternative.snap
@@ -0,0 +1,274 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-002",
+    "seq": 0,
+    "event_id": "befa06a9687c173e6e592e40b16848d455058e85933537a6d917996582ab6742",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/web-dashboard",
+      "repo_root": null,
+      "git": {
+        "sha": "3b81fe09cc12",
+        "branch": "feat/state-mgmt"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/projects/web-dashboard",
+      "git": {
+        "sha": "3b81fe09cc12",
+        "branch": "feat/state-mgmt"
+      },
+      "model": null,
+      "tool_version": "0.43.2"
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/rejected_alternative.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-002",
+    "seq": 1,
+    "event_id": "msg-101",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:10:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/web-dashboard",
+      "repo_root": null,
+      "git": {
+        "sha": "3b81fe09cc12",
+        "branch": "feat/state-mgmt"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Should we use Redux or Zustand for the new dashboard state?",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Should we use Redux or Zustand for the new dashboard state?"
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/rejected_alternative.jsonl",
+      "byte_offset": 173,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-002",
+    "seq": 2,
+    "event_id": "msg-102",
+    "parent_id": "msg-101",
+    "timestamp": "2026-06-22T15:10:09Z",
+    "project": {
+      "cwd": "/Users/dev/projects/web-dashboard",
+      "repo_root": null,
+      "git": {
+        "sha": "3b81fe09cc12",
+        "branch": "feat/state-mgmt"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "I considered Redux, but for this dashboard it's heavier than we need — too much boilerplate for the small store. We'll go with Zustand instead.",
+      "thinking": null,
+      "model": "cursor-pro",
+      "usage": {
+        "input_tokens": 180,
+        "output_tokens": 72,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "I considered Redux, but for this dashboard it's heavier than we need — too much boilerplate for the small store. We'll go with Zustand instead."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/rejected_alternative.jsonl",
+      "byte_offset": 344,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-002",
+    "seq": 3,
+    "event_id": "msg-103",
+    "parent_id": "msg-102",
+    "timestamp": "2026-06-22T15:10:40Z",
+    "project": {
+      "cwd": "/Users/dev/projects/web-dashboard",
+      "repo_root": null,
+      "git": {
+        "sha": "3b81fe09cc12",
+        "branch": "feat/state-mgmt"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Agreed, reject Redux. Let's go with Zustand and wire up the store.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Agreed, reject Redux. Let's go with Zustand and wire up the store."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/rejected_alternative.jsonl",
+      "byte_offset": 666,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-002",
+    "seq": 4,
+    "event_id": "msg-104",
+    "parent_id": "msg-103",
+    "timestamp": "2026-06-22T15:10:52Z",
+    "project": {
+      "cwd": "/Users/dev/projects/web-dashboard",
+      "repo_root": null,
+      "git": {
+        "sha": "3b81fe09cc12",
+        "branch": "feat/state-mgmt"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Creating the Zustand store now.",
+      "thinking": null,
+      "model": "cursor-pro",
+      "usage": {
+        "input_tokens": 96,
+        "output_tokens": 40,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Creating the Zustand store now."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/rejected_alternative.jsonl",
+      "byte_offset": 849,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-002",
+    "seq": 5,
+    "event_id": "msg-104#toolcall:call-write-1",
+    "parent_id": "msg-104",
+    "timestamp": "2026-06-22T15:10:52Z",
+    "project": {
+      "cwd": "/Users/dev/projects/web-dashboard",
+      "repo_root": null,
+      "git": {
+        "sha": "3b81fe09cc12",
+        "branch": "feat/state-mgmt"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-write-1",
+      "name": "write_file",
+      "args": {
+        "path": "src/store/dashboard.ts"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/rejected_alternative.jsonl",
+      "byte_offset": 849,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-002",
+    "seq": 6,
+    "event_id": "msg-104#toolresult:call-write-1",
+    "parent_id": "msg-104",
+    "timestamp": "2026-06-22T15:10:52Z",
+    "project": {
+      "cwd": "/Users/dev/projects/web-dashboard",
+      "repo_root": null,
+      "git": {
+        "sha": "3b81fe09cc12",
+        "branch": "feat/state-mgmt"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-write-1",
+      "ok": true,
+      "output": "created file"
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/rejected_alternative.jsonl",
+      "byte_offset": 849,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-002",
+    "seq": 7,
+    "event_id": "msg-104#edit:0",
+    "parent_id": "msg-104",
+    "timestamp": "2026-06-22T15:10:52Z",
+    "project": {
+      "cwd": "/Users/dev/projects/web-dashboard",
+      "repo_root": null,
+      "git": {
+        "sha": "3b81fe09cc12",
+        "branch": "feat/state-mgmt"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "src/store/dashboard.ts",
+        "old": "",
+        "new": "import { create } from 'zustand';\nexport const useDashboard = create((set) => ({ panels: [] }));\n",
+        "unified": "@@ -0,0 +1,2 @@\n+import { create } from 'zustand';\n+export const useDashboard = create((set) => ({ panels: [] }));\n",
+        "added_lines": 2,
+        "removed_lines": 0
+      }
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/rejected_alternative.jsonl",
+      "byte_offset": 849,
+      "line_no": 5
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__cursor__v1__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/events__cursor__v1__tool_failure.snap
new file mode 100644
index 0000000..ea3d501
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__cursor__v1__tool_failure.snap
@@ -0,0 +1,202 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-004",
+    "seq": 0,
+    "event_id": "5531b6e6cb1be2222fffe011757a5d4b59d8cd80dcd98fa8521b94f47fdcc7b0",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/infra-scripts",
+      "repo_root": null,
+      "git": {
+        "sha": "c0ffee123456",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/projects/infra-scripts",
+      "git": {
+        "sha": "c0ffee123456",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "0.43.2"
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/tool_failure.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-004",
+    "seq": 1,
+    "event_id": "msg-301",
+    "parent_id": null,
+    "timestamp": "2026-06-22T17:05:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/infra-scripts",
+      "repo_root": null,
+      "git": {
+        "sha": "c0ffee123456",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Add a retry wrapper around the deploy step in deploy.sh.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Add a retry wrapper around the deploy step in deploy.sh."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/tool_failure.jsonl",
+      "byte_offset": 162,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-004",
+    "seq": 2,
+    "event_id": "msg-302",
+    "parent_id": "msg-301",
+    "timestamp": "2026-06-22T17:05:14Z",
+    "project": {
+      "cwd": "/Users/dev/projects/infra-scripts",
+      "repo_root": null,
+      "git": {
+        "sha": "c0ffee123456",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "I'll wrap the deploy step in a retry loop.",
+      "thinking": null,
+      "model": "cursor-fast",
+      "usage": {
+        "input_tokens": 110,
+        "output_tokens": 52,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "I'll wrap the deploy step in a retry loop."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/tool_failure.jsonl",
+      "byte_offset": 330,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-004",
+    "seq": 3,
+    "event_id": "msg-302#toolcall:call-edit-4",
+    "parent_id": "msg-302",
+    "timestamp": "2026-06-22T17:05:14Z",
+    "project": {
+      "cwd": "/Users/dev/projects/infra-scripts",
+      "repo_root": null,
+      "git": {
+        "sha": "c0ffee123456",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-edit-4",
+      "name": "edit_file",
+      "args": {
+        "path": "deploy.sh"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/tool_failure.jsonl",
+      "byte_offset": 330,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-004",
+    "seq": 4,
+    "event_id": "msg-302#toolresult:call-edit-4",
+    "parent_id": "msg-302",
+    "timestamp": "2026-06-22T17:05:14Z",
+    "project": {
+      "cwd": "/Users/dev/projects/infra-scripts",
+      "repo_root": null,
+      "git": {
+        "sha": "c0ffee123456",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-edit-4",
+      "ok": false,
+      "output": "error: permission denied: deploy.sh is read-only (EACCES)"
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/tool_failure.jsonl",
+      "byte_offset": 330,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "cursor",
+    "session_id": "cur-sess-004",
+    "seq": 5,
+    "event_id": "msg-302#edit:0",
+    "parent_id": "msg-302",
+    "timestamp": "2026-06-22T17:05:14Z",
+    "project": {
+      "cwd": "/Users/dev/projects/infra-scripts",
+      "repo_root": null,
+      "git": {
+        "sha": "c0ffee123456",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-edit-4",
+      "diff": {
+        "path": "deploy.sh",
+        "old": "run_deploy",
+        "new": "for i in 1 2 3; do run_deploy && break; done",
+        "unified": "@@ -1 +1 @@\n-run_deploy\n+for i in 1 2 3; do run_deploy && break; done\n",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/cursor/v1/tool_failure.jsonl",
+      "byte_offset": 330,
+      "line_no": 3
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__gemini__legacy_json__version_tolerance.snap b/crates/memscribe-testkit/tests/snapshots/events__gemini__legacy_json__version_tolerance.snap
new file mode 100644
index 0000000..54e98dd
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__gemini__legacy_json__version_tolerance.snap
@@ -0,0 +1,153 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "legacy-0001",
+    "seq": 0,
+    "event_id": "a53bda02853957d7db23852ba98d761aedae9ca2557bef565f2938da8b5ff2b9",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/shop-api",
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      },
+      "model": "gemini-1.5-pro",
+      "tool_version": "0.2.0"
+    },
+    "provenance": {
+      "file": "fixtures/gemini/legacy_json/version_tolerance.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "legacy-0001",
+    "seq": 1,
+    "event_id": "leg-0001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T08:00:00Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use Postgres instead of MySQL for the orders service.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/legacy_json/version_tolerance.jsonl",
+      "byte_offset": 183,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "legacy-0001",
+    "seq": 2,
+    "event_id": "leg-0002",
+    "parent_id": null,
+    "timestamp": "2026-06-22T08:00:03.12Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Good call. Switching the connection string to Postgres.",
+      "thinking": "swap the dsn",
+      "model": null,
+      "usage": null,
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/legacy_json/version_tolerance.jsonl",
+      "byte_offset": 326,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "legacy-0001",
+    "seq": 3,
+    "event_id": "f601b26ea26e10557f9c9d993a63bcf87dfc45a8a18eb466bebf4124bfd02aed",
+    "parent_id": null,
+    "timestamp": "2026-06-22T08:00:04Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "unknown",
+      "raw_type": "unknown",
+      "raw": {
+        "id": "leg-0003",
+        "role": "system",
+        "content": "Context window compacted.",
+        "timestamp": "2026-06-22T08:00:04.000Z"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/legacy_json/version_tolerance.jsonl",
+      "byte_offset": 492,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "legacy-0001",
+    "seq": 4,
+    "event_id": "9cfc1acb8e98f1936b2258065acc5dab59566f23d22e07daec8aa9babf0d7812",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "rewind",
+      "to_event": "leg-0001"
+    },
+    "provenance": {
+      "file": "fixtures/gemini/legacy_json/version_tolerance.jsonl",
+      "byte_offset": 603,
+      "line_no": 5
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__ban.snap b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__ban.snap
new file mode 100644
index 0000000..c9e90b1
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__ban.snap
@@ -0,0 +1,100 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0003",
+    "seq": 0,
+    "event_id": "9ba17a73a3ea97a439e4153d2471a5de56dbae63bfed30b23280a651d43ec27d",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/core-lib",
+      "repo_root": null,
+      "git": {
+        "sha": "0042deadbe",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/core-lib",
+      "git": {
+        "sha": "0042deadbe",
+        "branch": "main"
+      },
+      "model": "gemini-2.5-pro",
+      "tool_version": "0.4.1"
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/ban.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0003",
+    "seq": 1,
+    "event_id": "msg-0201",
+    "parent_id": null,
+    "timestamp": "2026-06-22T12:30:00Z",
+    "project": {
+      "cwd": "/home/dev/core-lib",
+      "repo_root": null,
+      "git": {
+        "sha": "0042deadbe",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Important policy: we will never add a dependency on left-pad. Use the std library instead.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/ban.jsonl",
+      "byte_offset": 150,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0003",
+    "seq": 2,
+    "event_id": "msg-0202",
+    "parent_id": null,
+    "timestamp": "2026-06-22T12:30:02.7Z",
+    "project": {
+      "cwd": "/home/dev/core-lib",
+      "repo_root": null,
+      "git": {
+        "sha": "0042deadbe",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Acknowledged. I won't add left-pad; I'll implement padding with the standard library.",
+      "thinking": "This is a ban on the left-pad dependency.",
+      "model": null,
+      "usage": {
+        "input_tokens": 110,
+        "output_tokens": 58,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/ban.jsonl",
+      "byte_offset": 321,
+      "line_no": 3
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..6562932
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__happy_path_decision_then_edits.snap
@@ -0,0 +1,329 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0001",
+    "seq": 0,
+    "event_id": "ea5baebfc8d80eba00b2965eca2ed473a364cf9bffc744e531d05fbac3a064e0",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/shop-api",
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      },
+      "model": "gemini-2.5-pro",
+      "tool_version": "0.4.1"
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0001",
+    "seq": 1,
+    "event_id": "msg-0001",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:00Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use Postgres instead of MySQL for the orders service. Update the config and the migration runner.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 185,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0001",
+    "seq": 2,
+    "event_id": "msg-0002",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:03.12Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Good call. I'll switch the connection string to Postgres and update the migration runner to use the pg driver.",
+      "thinking": "Replace the mysql DSN with a postgres URL, then swap the driver import in the runner.",
+      "model": null,
+      "usage": {
+        "input_tokens": 182,
+        "output_tokens": 97,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 369,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0001",
+    "seq": 3,
+    "event_id": "call-cfg-1:call",
+    "parent_id": "msg-0002",
+    "timestamp": "2026-06-22T10:00:03.12Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-cfg-1",
+      "name": "write_file",
+      "args": {
+        "path": "config/database.toml"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 369,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0001",
+    "seq": 4,
+    "event_id": "call-cfg-1:result",
+    "parent_id": "msg-0002",
+    "timestamp": "2026-06-22T10:00:03.12Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-cfg-1",
+      "ok": true,
+      "output": {
+        "fileName": "config/database.toml",
+        "originalContent": "[db]\ndriver = \"mysql\"\nurl = \"mysql://localhost:3306/orders\"\n",
+        "newContent": "[db]\ndriver = \"postgres\"\nurl = \"postgres://localhost:5432/orders\"\n",
+        "fileDiff": "@@ -1,3 +1,3 @@\n [db]\n-driver = \"mysql\"\n-url = \"mysql://localhost:3306/orders\"\n+driver = \"postgres\"\n+url = \"postgres://localhost:5432/orders\"\n",
+        "diffStat": {
+          "added": 2,
+          "removed": 2
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 369,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0001",
+    "seq": 5,
+    "event_id": "call-cfg-1:edit",
+    "parent_id": "msg-0002",
+    "timestamp": "2026-06-22T10:00:03.12Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-cfg-1",
+      "diff": {
+        "path": "config/database.toml",
+        "old": "[db]\ndriver = \"mysql\"\nurl = \"mysql://localhost:3306/orders\"\n",
+        "new": "[db]\ndriver = \"postgres\"\nurl = \"postgres://localhost:5432/orders\"\n",
+        "unified": "@@ -1,3 +1,3 @@\n [db]\n-driver = \"mysql\"\n-url = \"mysql://localhost:3306/orders\"\n+driver = \"postgres\"\n+url = \"postgres://localhost:5432/orders\"\n",
+        "added_lines": 2,
+        "removed_lines": 2
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 369,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0001",
+    "seq": 6,
+    "event_id": "call-run-1:call",
+    "parent_id": "msg-0002",
+    "timestamp": "2026-06-22T10:00:03.12Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-run-1",
+      "name": "replace",
+      "args": {
+        "path": "src/migrate.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 369,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0001",
+    "seq": 7,
+    "event_id": "call-run-1:result",
+    "parent_id": "msg-0002",
+    "timestamp": "2026-06-22T10:00:03.12Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-run-1",
+      "ok": true,
+      "output": {
+        "filePath": "src/migrate.rs",
+        "originalContent": "use mysql::Pool;\n",
+        "newContent": "use postgres::Client;\n",
+        "fileDiff": "@@ -1 +1 @@\n-use mysql::Pool;\n+use postgres::Client;\n",
+        "diffStat": {
+          "added": 1,
+          "removed": 1
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 369,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0001",
+    "seq": 8,
+    "event_id": "call-run-1:edit",
+    "parent_id": "msg-0002",
+    "timestamp": "2026-06-22T10:00:03.12Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-run-1",
+      "diff": {
+        "path": "src/migrate.rs",
+        "old": "use mysql::Pool;\n",
+        "new": "use postgres::Client;\n",
+        "unified": "@@ -1 +1 @@\n-use mysql::Pool;\n+use postgres::Client;\n",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 369,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0001",
+    "seq": 9,
+    "event_id": "msg-0003",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:30Z",
+    "project": {
+      "cwd": "/home/dev/shop-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f1c0aa2b3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Looks good, thanks.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1529,
+      "line_no": 4
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__interleaved_arcs.snap b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__interleaved_arcs.snap
new file mode 100644
index 0000000..00c6d16
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__interleaved_arcs.snap
@@ -0,0 +1,593 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 0,
+    "event_id": "1ac40ee870055ec85e0b876e5ce400e5a6585dea3c29af969c697ee2b0cb19fb",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/orders-api",
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      },
+      "model": "gemini-2.5-pro",
+      "tool_version": "0.4.1"
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 1,
+    "event_id": "msg-ia-01",
+    "parent_id": null,
+    "timestamp": "2026-06-22T09:00:00Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use Postgres for the orders store. Update the connection module and the schema.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 194,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 2,
+    "event_id": "msg-ia-02",
+    "parent_id": null,
+    "timestamp": "2026-06-22T09:00:02Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Switching the orders store to Postgres now.",
+      "thinking": "Rewrite db.rs to the pg driver and adjust schema.rs.",
+      "model": null,
+      "usage": {
+        "input_tokens": 120,
+        "output_tokens": 61,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 361,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 3,
+    "event_id": "call-ia-db1:call",
+    "parent_id": "msg-ia-02",
+    "timestamp": "2026-06-22T09:00:02Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-ia-db1",
+      "name": "replace",
+      "args": {
+        "path": "src/db.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 361,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 4,
+    "event_id": "call-ia-db1:result",
+    "parent_id": "msg-ia-02",
+    "timestamp": "2026-06-22T09:00:02Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-ia-db1",
+      "ok": true,
+      "output": {
+        "fileName": "src/db.rs",
+        "originalContent": "use mysql::Pool;\n",
+        "newContent": "use postgres::Client;\n",
+        "fileDiff": "@@ -1 +1 @@\n-use mysql::Pool;\n+use postgres::Client;\n",
+        "diffStat": {
+          "added": 1,
+          "removed": 1
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 361,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 5,
+    "event_id": "call-ia-db1:edit",
+    "parent_id": "msg-ia-02",
+    "timestamp": "2026-06-22T09:00:02Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-ia-db1",
+      "diff": {
+        "path": "src/db.rs",
+        "old": "use mysql::Pool;\n",
+        "new": "use postgres::Client;\n",
+        "unified": "@@ -1 +1 @@\n-use mysql::Pool;\n+use postgres::Client;\n",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 361,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 6,
+    "event_id": "call-ia-schema1:call",
+    "parent_id": "msg-ia-02",
+    "timestamp": "2026-06-22T09:00:02Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-ia-schema1",
+      "name": "replace",
+      "args": {
+        "path": "src/schema.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 361,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 7,
+    "event_id": "call-ia-schema1:result",
+    "parent_id": "msg-ia-02",
+    "timestamp": "2026-06-22T09:00:02Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-ia-schema1",
+      "ok": true,
+      "output": {
+        "fileName": "src/schema.rs",
+        "originalContent": "-- mysql schema\n",
+        "newContent": "-- postgres schema\n",
+        "fileDiff": "@@ -1 +1 @@\n--- mysql schema\n+-- postgres schema\n",
+        "diffStat": {
+          "added": 1,
+          "removed": 1
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 361,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 8,
+    "event_id": "call-ia-schema1:edit",
+    "parent_id": "msg-ia-02",
+    "timestamp": "2026-06-22T09:00:02Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-ia-schema1",
+      "diff": {
+        "path": "src/schema.rs",
+        "old": "-- mysql schema\n",
+        "new": "-- postgres schema\n",
+        "unified": "@@ -1 +1 @@\n--- mysql schema\n+-- postgres schema\n",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 361,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 9,
+    "event_id": "msg-ia-03",
+    "parent_id": null,
+    "timestamp": "2026-06-22T09:05:00Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Now switch to a bounded connection pool in the same db module instead of a single client.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 1194,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 10,
+    "event_id": "msg-ia-04",
+    "parent_id": null,
+    "timestamp": "2026-06-22T09:05:02Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Adding a bounded pool to db.rs and a new pool config.",
+      "thinking": "Re-edit db.rs to use a pool; add pool.rs.",
+      "model": null,
+      "usage": {
+        "input_tokens": 98,
+        "output_tokens": 52,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 1365,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 11,
+    "event_id": "call-ia-db2:call",
+    "parent_id": "msg-ia-04",
+    "timestamp": "2026-06-22T09:05:02Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-ia-db2",
+      "name": "replace",
+      "args": {
+        "path": "src/db.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 1365,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 12,
+    "event_id": "call-ia-db2:result",
+    "parent_id": "msg-ia-04",
+    "timestamp": "2026-06-22T09:05:02Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-ia-db2",
+      "ok": true,
+      "output": {
+        "fileName": "src/db.rs",
+        "originalContent": "use postgres::Client;\n",
+        "newContent": "use deadpool_postgres::Pool;\n",
+        "fileDiff": "@@ -1 +1 @@\n-use postgres::Client;\n+use deadpool_postgres::Pool;\n",
+        "diffStat": {
+          "added": 1,
+          "removed": 1
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 1365,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 13,
+    "event_id": "call-ia-db2:edit",
+    "parent_id": "msg-ia-04",
+    "timestamp": "2026-06-22T09:05:02Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-ia-db2",
+      "diff": {
+        "path": "src/db.rs",
+        "old": "use postgres::Client;\n",
+        "new": "use deadpool_postgres::Pool;\n",
+        "unified": "@@ -1 +1 @@\n-use postgres::Client;\n+use deadpool_postgres::Pool;\n",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 1365,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 14,
+    "event_id": "call-ia-pool1:call",
+    "parent_id": "msg-ia-04",
+    "timestamp": "2026-06-22T09:05:02Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-ia-pool1",
+      "name": "write_file",
+      "args": {
+        "path": "src/pool.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 1365,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 15,
+    "event_id": "call-ia-pool1:result",
+    "parent_id": "msg-ia-04",
+    "timestamp": "2026-06-22T09:05:02Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-ia-pool1",
+      "ok": true,
+      "output": {
+        "fileName": "src/pool.rs",
+        "originalContent": "",
+        "newContent": "pub struct PoolCfg { pub max: usize }\n",
+        "fileDiff": "@@ -0,0 +1 @@\n+pub struct PoolCfg { pub max: usize }\n",
+        "diffStat": {
+          "added": 1,
+          "removed": 0
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 1365,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 16,
+    "event_id": "call-ia-pool1:edit",
+    "parent_id": "msg-ia-04",
+    "timestamp": "2026-06-22T09:05:02Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-ia-pool1",
+      "diff": {
+        "path": "src/pool.rs",
+        "old": null,
+        "new": "pub struct PoolCfg { pub max: usize }\n",
+        "unified": "@@ -0,0 +1 @@\n+pub struct PoolCfg { pub max: usize }\n",
+        "added_lines": 1,
+        "removed_lines": 0
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 1365,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-interleaved-0001",
+    "seq": 17,
+    "event_id": "msg-ia-05",
+    "parent_id": null,
+    "timestamp": "2026-06-22T09:05:30Z",
+    "project": {
+      "cwd": "/home/dev/orders-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1a1a1a1a1",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Great, that covers it.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+      "byte_offset": 2222,
+      "line_no": 6
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__multi_edit_single_commit.snap b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__multi_edit_single_commit.snap
new file mode 100644
index 0000000..470ce7d
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__multi_edit_single_commit.snap
@@ -0,0 +1,430 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-multiedit-0001",
+    "seq": 0,
+    "event_id": "f70c126012255c073120846675da0f73ac103b2bca790a07b348bc69e1d82623",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/auth-svc",
+      "repo_root": null,
+      "git": {
+        "sha": "b2b2b2b2b2",
+        "branch": "feat/argon2"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/auth-svc",
+      "git": {
+        "sha": "b2b2b2b2b2",
+        "branch": "feat/argon2"
+      },
+      "model": "gemini-2.5-pro",
+      "tool_version": "0.4.1"
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/multi_edit_single_commit.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-multiedit-0001",
+    "seq": 1,
+    "event_id": "msg-me-01",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:00:00Z",
+    "project": {
+      "cwd": "/home/dev/auth-svc",
+      "repo_root": null,
+      "git": {
+        "sha": "b2b2b2b2b2",
+        "branch": "feat/argon2"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "We will use Argon2 for password hashing across the auth service. Update the hasher, the verifier, and the config in one pass.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/multi_edit_single_commit.jsonl",
+      "byte_offset": 195,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-multiedit-0001",
+    "seq": 2,
+    "event_id": "msg-me-02",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:00:03Z",
+    "project": {
+      "cwd": "/home/dev/auth-svc",
+      "repo_root": null,
+      "git": {
+        "sha": "b2b2b2b2b2",
+        "branch": "feat/argon2"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Rolling Argon2 through the hasher, verifier, and config as a single change.",
+      "thinking": "Three coordinated edits landing as one logical commit.",
+      "model": null,
+      "usage": {
+        "input_tokens": 160,
+        "output_tokens": 88,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/multi_edit_single_commit.jsonl",
+      "byte_offset": 402,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-multiedit-0001",
+    "seq": 3,
+    "event_id": "call-me-hash:call",
+    "parent_id": "msg-me-02",
+    "timestamp": "2026-06-22T14:00:03Z",
+    "project": {
+      "cwd": "/home/dev/auth-svc",
+      "repo_root": null,
+      "git": {
+        "sha": "b2b2b2b2b2",
+        "branch": "feat/argon2"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-me-hash",
+      "name": "replace",
+      "args": {
+        "path": "src/auth/hasher.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/multi_edit_single_commit.jsonl",
+      "byte_offset": 402,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-multiedit-0001",
+    "seq": 4,
+    "event_id": "call-me-hash:result",
+    "parent_id": "msg-me-02",
+    "timestamp": "2026-06-22T14:00:03Z",
+    "project": {
+      "cwd": "/home/dev/auth-svc",
+      "repo_root": null,
+      "git": {
+        "sha": "b2b2b2b2b2",
+        "branch": "feat/argon2"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-me-hash",
+      "ok": true,
+      "output": {
+        "fileName": "src/auth/hasher.rs",
+        "originalContent": "use bcrypt::hash;\n",
+        "newContent": "use argon2::Argon2;\n",
+        "fileDiff": "@@ -1 +1 @@\n-use bcrypt::hash;\n+use argon2::Argon2;\n",
+        "diffStat": {
+          "added": 1,
+          "removed": 1
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/multi_edit_single_commit.jsonl",
+      "byte_offset": 402,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-multiedit-0001",
+    "seq": 5,
+    "event_id": "call-me-hash:edit",
+    "parent_id": "msg-me-02",
+    "timestamp": "2026-06-22T14:00:03Z",
+    "project": {
+      "cwd": "/home/dev/auth-svc",
+      "repo_root": null,
+      "git": {
+        "sha": "b2b2b2b2b2",
+        "branch": "feat/argon2"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-me-hash",
+      "diff": {
+        "path": "src/auth/hasher.rs",
+        "old": "use bcrypt::hash;\n",
+        "new": "use argon2::Argon2;\n",
+        "unified": "@@ -1 +1 @@\n-use bcrypt::hash;\n+use argon2::Argon2;\n",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/multi_edit_single_commit.jsonl",
+      "byte_offset": 402,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-multiedit-0001",
+    "seq": 6,
+    "event_id": "call-me-verify:call",
+    "parent_id": "msg-me-02",
+    "timestamp": "2026-06-22T14:00:03Z",
+    "project": {
+      "cwd": "/home/dev/auth-svc",
+      "repo_root": null,
+      "git": {
+        "sha": "b2b2b2b2b2",
+        "branch": "feat/argon2"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-me-verify",
+      "name": "replace",
+      "args": {
+        "path": "src/auth/verifier.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/multi_edit_single_commit.jsonl",
+      "byte_offset": 402,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-multiedit-0001",
+    "seq": 7,
+    "event_id": "call-me-verify:result",
+    "parent_id": "msg-me-02",
+    "timestamp": "2026-06-22T14:00:03Z",
+    "project": {
+      "cwd": "/home/dev/auth-svc",
+      "repo_root": null,
+      "git": {
+        "sha": "b2b2b2b2b2",
+        "branch": "feat/argon2"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-me-verify",
+      "ok": true,
+      "output": {
+        "fileName": "src/auth/verifier.rs",
+        "originalContent": "use bcrypt::verify;\n",
+        "newContent": "use argon2::PasswordVerifier;\n",
+        "fileDiff": "@@ -1 +1 @@\n-use bcrypt::verify;\n+use argon2::PasswordVerifier;\n",
+        "diffStat": {
+          "added": 1,
+          "removed": 1
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/multi_edit_single_commit.jsonl",
+      "byte_offset": 402,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-multiedit-0001",
+    "seq": 8,
+    "event_id": "call-me-verify:edit",
+    "parent_id": "msg-me-02",
+    "timestamp": "2026-06-22T14:00:03Z",
+    "project": {
+      "cwd": "/home/dev/auth-svc",
+      "repo_root": null,
+      "git": {
+        "sha": "b2b2b2b2b2",
+        "branch": "feat/argon2"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-me-verify",
+      "diff": {
+        "path": "src/auth/verifier.rs",
+        "old": "use bcrypt::verify;\n",
+        "new": "use argon2::PasswordVerifier;\n",
+        "unified": "@@ -1 +1 @@\n-use bcrypt::verify;\n+use argon2::PasswordVerifier;\n",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/multi_edit_single_commit.jsonl",
+      "byte_offset": 402,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-multiedit-0001",
+    "seq": 9,
+    "event_id": "call-me-config:call",
+    "parent_id": "msg-me-02",
+    "timestamp": "2026-06-22T14:00:03Z",
+    "project": {
+      "cwd": "/home/dev/auth-svc",
+      "repo_root": null,
+      "git": {
+        "sha": "b2b2b2b2b2",
+        "branch": "feat/argon2"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-me-config",
+      "name": "replace",
+      "args": {
+        "path": "config/auth.toml"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/multi_edit_single_commit.jsonl",
+      "byte_offset": 402,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-multiedit-0001",
+    "seq": 10,
+    "event_id": "call-me-config:result",
+    "parent_id": "msg-me-02",
+    "timestamp": "2026-06-22T14:00:03Z",
+    "project": {
+      "cwd": "/home/dev/auth-svc",
+      "repo_root": null,
+      "git": {
+        "sha": "b2b2b2b2b2",
+        "branch": "feat/argon2"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-me-config",
+      "ok": true,
+      "output": {
+        "fileName": "config/auth.toml",
+        "originalContent": "hash = \"bcrypt\"\n",
+        "newContent": "hash = \"argon2\"\n",
+        "fileDiff": "@@ -1 +1 @@\n-hash = \"bcrypt\"\n+hash = \"argon2\"\n",
+        "diffStat": {
+          "added": 1,
+          "removed": 1
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/multi_edit_single_commit.jsonl",
+      "byte_offset": 402,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-multiedit-0001",
+    "seq": 11,
+    "event_id": "call-me-config:edit",
+    "parent_id": "msg-me-02",
+    "timestamp": "2026-06-22T14:00:03Z",
+    "project": {
+      "cwd": "/home/dev/auth-svc",
+      "repo_root": null,
+      "git": {
+        "sha": "b2b2b2b2b2",
+        "branch": "feat/argon2"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-me-config",
+      "diff": {
+        "path": "config/auth.toml",
+        "old": "hash = \"bcrypt\"\n",
+        "new": "hash = \"argon2\"\n",
+        "unified": "@@ -1 +1 @@\n-hash = \"bcrypt\"\n+hash = \"argon2\"\n",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/multi_edit_single_commit.jsonl",
+      "byte_offset": 402,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-multiedit-0001",
+    "seq": 12,
+    "event_id": "msg-me-03",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:01:00Z",
+    "project": {
+      "cwd": "/home/dev/auth-svc",
+      "repo_root": null,
+      "git": {
+        "sha": "b2b2b2b2b2",
+        "branch": "feat/argon2"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Perfect, ship it.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/multi_edit_single_commit.jsonl",
+      "byte_offset": 1634,
+      "line_no": 4
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__no_commitment_marker.snap b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__no_commitment_marker.snap
new file mode 100644
index 0000000..549d18a
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__no_commitment_marker.snap
@@ -0,0 +1,201 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-nomarker-0001",
+    "seq": 0,
+    "event_id": "b491fb245e8b756ef8349d16223411a08529d53ad664e8faa7930d35c70a8df4",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/widgets",
+      "repo_root": null,
+      "git": {
+        "sha": "e5e5e5e5e5",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/widgets",
+      "git": {
+        "sha": "e5e5e5e5e5",
+        "branch": "main"
+      },
+      "model": "gemini-2.5-pro",
+      "tool_version": "0.4.1"
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/no_commitment_marker.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-nomarker-0001",
+    "seq": 1,
+    "event_id": "msg-nm-01",
+    "parent_id": null,
+    "timestamp": "2026-06-22T17:00:00Z",
+    "project": {
+      "cwd": "/home/dev/widgets",
+      "repo_root": null,
+      "git": {
+        "sha": "e5e5e5e5e5",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Thanks, this looks good to me.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/no_commitment_marker.jsonl",
+      "byte_offset": 185,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-nomarker-0001",
+    "seq": 2,
+    "event_id": "msg-nm-02",
+    "parent_id": null,
+    "timestamp": "2026-06-22T17:00:02Z",
+    "project": {
+      "cwd": "/home/dev/widgets",
+      "repo_root": null,
+      "git": {
+        "sha": "e5e5e5e5e5",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Glad it works. I'll tidy the trailing whitespace in the widget while I'm here.",
+      "thinking": "A small cleanup edit with no governing decision turn.",
+      "model": null,
+      "usage": {
+        "input_tokens": 70,
+        "output_tokens": 38,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/no_commitment_marker.jsonl",
+      "byte_offset": 297,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-nomarker-0001",
+    "seq": 3,
+    "event_id": "call-nm-tidy:call",
+    "parent_id": "msg-nm-02",
+    "timestamp": "2026-06-22T17:00:02Z",
+    "project": {
+      "cwd": "/home/dev/widgets",
+      "repo_root": null,
+      "git": {
+        "sha": "e5e5e5e5e5",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-nm-tidy",
+      "name": "replace",
+      "args": {
+        "path": "src/widget.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/no_commitment_marker.jsonl",
+      "byte_offset": 297,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-nomarker-0001",
+    "seq": 4,
+    "event_id": "call-nm-tidy:result",
+    "parent_id": "msg-nm-02",
+    "timestamp": "2026-06-22T17:00:02Z",
+    "project": {
+      "cwd": "/home/dev/widgets",
+      "repo_root": null,
+      "git": {
+        "sha": "e5e5e5e5e5",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-nm-tidy",
+      "ok": true,
+      "output": {
+        "fileName": "src/widget.rs",
+        "originalContent": "fn render() { }  \n",
+        "newContent": "fn render() {}\n",
+        "fileDiff": "@@ -1 +1 @@\n-fn render() { }  \n+fn render() {}\n",
+        "diffStat": {
+          "added": 1,
+          "removed": 1
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/no_commitment_marker.jsonl",
+      "byte_offset": 297,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-nomarker-0001",
+    "seq": 5,
+    "event_id": "call-nm-tidy:edit",
+    "parent_id": "msg-nm-02",
+    "timestamp": "2026-06-22T17:00:02Z",
+    "project": {
+      "cwd": "/home/dev/widgets",
+      "repo_root": null,
+      "git": {
+        "sha": "e5e5e5e5e5",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-nm-tidy",
+      "diff": {
+        "path": "src/widget.rs",
+        "old": "fn render() { }  \n",
+        "new": "fn render() {}\n",
+        "unified": "@@ -1 +1 @@\n-fn render() { }  \n+fn render() {}\n",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/no_commitment_marker.jsonl",
+      "byte_offset": 297,
+      "line_no": 3
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__rejected_alternative.snap
new file mode 100644
index 0000000..b470095
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__rejected_alternative.snap
@@ -0,0 +1,201 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0002",
+    "seq": 0,
+    "event_id": "cb223821b21cf2939e2a719a2690520c3c7e8b649dced8b12fdb43c3ae8b7642",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/billing",
+      "repo_root": null,
+      "git": {
+        "sha": "77ab12cd",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/billing",
+      "git": {
+        "sha": "77ab12cd",
+        "branch": "feat/payments"
+      },
+      "model": "gemini-2.5-pro",
+      "tool_version": "0.4.1"
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rejected_alternative.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0002",
+    "seq": 1,
+    "event_id": "msg-0101",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:15:00Z",
+    "project": {
+      "cwd": "/home/dev/billing",
+      "repo_root": null,
+      "git": {
+        "sha": "77ab12cd",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "For the checkout flow, use Stripe instead of PayPal. Wire up the Stripe client.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rejected_alternative.jsonl",
+      "byte_offset": 156,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0002",
+    "seq": 2,
+    "event_id": "msg-0102",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:15:04.5Z",
+    "project": {
+      "cwd": "/home/dev/billing",
+      "repo_root": null,
+      "git": {
+        "sha": "77ab12cd",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Understood — Stripe it is. Adding the Stripe client and removing the PayPal stub.",
+      "thinking": "User explicitly rejected PayPal in favor of Stripe.",
+      "model": null,
+      "usage": {
+        "input_tokens": 140,
+        "output_tokens": 72,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rejected_alternative.jsonl",
+      "byte_offset": 316,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0002",
+    "seq": 3,
+    "event_id": "call-stripe-1:call",
+    "parent_id": "msg-0102",
+    "timestamp": "2026-06-22T11:15:04.5Z",
+    "project": {
+      "cwd": "/home/dev/billing",
+      "repo_root": null,
+      "git": {
+        "sha": "77ab12cd",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-stripe-1",
+      "name": "write_file",
+      "args": {
+        "path": "src/payments/stripe.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rejected_alternative.jsonl",
+      "byte_offset": 316,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0002",
+    "seq": 4,
+    "event_id": "call-stripe-1:result",
+    "parent_id": "msg-0102",
+    "timestamp": "2026-06-22T11:15:04.5Z",
+    "project": {
+      "cwd": "/home/dev/billing",
+      "repo_root": null,
+      "git": {
+        "sha": "77ab12cd",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-stripe-1",
+      "ok": true,
+      "output": {
+        "fileName": "src/payments/stripe.rs",
+        "originalContent": "",
+        "newContent": "pub struct StripeClient { key: String }\n",
+        "fileDiff": "@@ -0,0 +1 @@\n+pub struct StripeClient { key: String }\n",
+        "diffStat": {
+          "added": 1,
+          "removed": 0
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rejected_alternative.jsonl",
+      "byte_offset": 316,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0002",
+    "seq": 5,
+    "event_id": "call-stripe-1:edit",
+    "parent_id": "msg-0102",
+    "timestamp": "2026-06-22T11:15:04.5Z",
+    "project": {
+      "cwd": "/home/dev/billing",
+      "repo_root": null,
+      "git": {
+        "sha": "77ab12cd",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-stripe-1",
+      "diff": {
+        "path": "src/payments/stripe.rs",
+        "old": null,
+        "new": "pub struct StripeClient { key: String }\n",
+        "unified": "@@ -0,0 +1 @@\n+pub struct StripeClient { key: String }\n",
+        "added_lines": 1,
+        "removed_lines": 0
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rejected_alternative.jsonl",
+      "byte_offset": 316,
+      "line_no": 3
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__rewind_compaction.snap b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__rewind_compaction.snap
new file mode 100644
index 0000000..39bcdb7
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__rewind_compaction.snap
@@ -0,0 +1,444 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-rewind-0001",
+    "seq": 0,
+    "event_id": "39cf7ba48601766d8753e2ec3d314b4fc5141ef9cb0849f168f3f0829e2b8564",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/catalog",
+      "repo_root": null,
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/catalog",
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      },
+      "model": "gemini-2.5-pro",
+      "tool_version": "0.4.1"
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-rewind-0001",
+    "seq": 1,
+    "event_id": "msg-rc-00",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:00:00Z",
+    "project": {
+      "cwd": "/home/dev/catalog",
+      "repo_root": null,
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's plan the storage layer for the catalog.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+      "byte_offset": 183,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-rewind-0001",
+    "seq": 2,
+    "event_id": "msg-rc-01",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:01:00Z",
+    "project": {
+      "cwd": "/home/dev/catalog",
+      "repo_root": null,
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Use MongoDB for the catalog store. Wire up the document client.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+      "byte_offset": 310,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-rewind-0001",
+    "seq": 3,
+    "event_id": "msg-rc-02",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:01:03Z",
+    "project": {
+      "cwd": "/home/dev/catalog",
+      "repo_root": null,
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Setting up the MongoDB document client.",
+      "thinking": "Replace the placeholder store with a mongo client.",
+      "model": null,
+      "usage": {
+        "input_tokens": 110,
+        "output_tokens": 54,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+      "byte_offset": 455,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-rewind-0001",
+    "seq": 4,
+    "event_id": "call-rc-mongo:call",
+    "parent_id": "msg-rc-02",
+    "timestamp": "2026-06-22T15:01:03Z",
+    "project": {
+      "cwd": "/home/dev/catalog",
+      "repo_root": null,
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-rc-mongo",
+      "name": "write_file",
+      "args": {
+        "path": "src/store.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+      "byte_offset": 455,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-rewind-0001",
+    "seq": 5,
+    "event_id": "call-rc-mongo:result",
+    "parent_id": "msg-rc-02",
+    "timestamp": "2026-06-22T15:01:03Z",
+    "project": {
+      "cwd": "/home/dev/catalog",
+      "repo_root": null,
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-rc-mongo",
+      "ok": true,
+      "output": {
+        "fileName": "src/store.rs",
+        "originalContent": "pub struct Store;\n",
+        "newContent": "use mongodb::Client;\npub struct Store(Client);\n",
+        "fileDiff": "@@ -1 +1,2 @@\n-pub struct Store;\n+use mongodb::Client;\n+pub struct Store(Client);\n",
+        "diffStat": {
+          "added": 2,
+          "removed": 1
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+      "byte_offset": 455,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-rewind-0001",
+    "seq": 6,
+    "event_id": "call-rc-mongo:edit",
+    "parent_id": "msg-rc-02",
+    "timestamp": "2026-06-22T15:01:03Z",
+    "project": {
+      "cwd": "/home/dev/catalog",
+      "repo_root": null,
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-rc-mongo",
+      "diff": {
+        "path": "src/store.rs",
+        "old": "pub struct Store;\n",
+        "new": "use mongodb::Client;\npub struct Store(Client);\n",
+        "unified": "@@ -1 +1,2 @@\n-pub struct Store;\n+use mongodb::Client;\n+pub struct Store(Client);\n",
+        "added_lines": 2,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+      "byte_offset": 455,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-rewind-0001",
+    "seq": 7,
+    "event_id": "c17d6e2538f83b5f0eab90658c1408979002f15d9e3bdb62a7d20272e16c82e6",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/catalog",
+      "repo_root": null,
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "rewind",
+      "to_event": "msg-rc-00"
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+      "byte_offset": 1052,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-rewind-0001",
+    "seq": 8,
+    "event_id": "msg-rc-03",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:10:00Z",
+    "project": {
+      "cwd": "/home/dev/catalog",
+      "repo_root": null,
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "On reflection, use Postgres with JSONB for the catalog store instead of MongoDB.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+      "byte_offset": 1078,
+      "line_no": 6
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-rewind-0001",
+    "seq": 9,
+    "event_id": "msg-rc-04",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:10:03Z",
+    "project": {
+      "cwd": "/home/dev/catalog",
+      "repo_root": null,
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Replacing the store with a Postgres JSONB client.",
+      "thinking": "Swap the mongo client for a postgres client.",
+      "model": null,
+      "usage": {
+        "input_tokens": 105,
+        "output_tokens": 58,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+      "byte_offset": 1240,
+      "line_no": 7
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-rewind-0001",
+    "seq": 10,
+    "event_id": "call-rc-pg:call",
+    "parent_id": "msg-rc-04",
+    "timestamp": "2026-06-22T15:10:03Z",
+    "project": {
+      "cwd": "/home/dev/catalog",
+      "repo_root": null,
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-rc-pg",
+      "name": "replace",
+      "args": {
+        "path": "src/store.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+      "byte_offset": 1240,
+      "line_no": 7
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-rewind-0001",
+    "seq": 11,
+    "event_id": "call-rc-pg:result",
+    "parent_id": "msg-rc-04",
+    "timestamp": "2026-06-22T15:10:03Z",
+    "project": {
+      "cwd": "/home/dev/catalog",
+      "repo_root": null,
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-rc-pg",
+      "ok": true,
+      "output": {
+        "fileName": "src/store.rs",
+        "originalContent": "use mongodb::Client;\npub struct Store(Client);\n",
+        "newContent": "use postgres::Client;\npub struct Store(Client);\n",
+        "fileDiff": "@@ -1,2 +1,2 @@\n-use mongodb::Client;\n+use postgres::Client;\n pub struct Store(Client);\n",
+        "diffStat": {
+          "added": 1,
+          "removed": 1
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+      "byte_offset": 1240,
+      "line_no": 7
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-rewind-0001",
+    "seq": 12,
+    "event_id": "call-rc-pg:edit",
+    "parent_id": "msg-rc-04",
+    "timestamp": "2026-06-22T15:10:03Z",
+    "project": {
+      "cwd": "/home/dev/catalog",
+      "repo_root": null,
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-rc-pg",
+      "diff": {
+        "path": "src/store.rs",
+        "old": "use mongodb::Client;\npub struct Store(Client);\n",
+        "new": "use postgres::Client;\npub struct Store(Client);\n",
+        "unified": "@@ -1,2 +1,2 @@\n-use mongodb::Client;\n+use postgres::Client;\n pub struct Store(Client);\n",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+      "byte_offset": 1240,
+      "line_no": 7
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-rewind-0001",
+    "seq": 13,
+    "event_id": "msg-rc-05",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:10:30Z",
+    "project": {
+      "cwd": "/home/dev/catalog",
+      "repo_root": null,
+      "git": {
+        "sha": "c3c3c3c3c3",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Looks right now, thanks.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+      "byte_offset": 1872,
+      "line_no": 8
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__subagent_thread.snap b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__subagent_thread.snap
new file mode 100644
index 0000000..b0de5fa
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__subagent_thread.snap
@@ -0,0 +1,290 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-subagent-0001",
+    "seq": 0,
+    "event_id": "468b47cea3585c2fc97471d5794be0385a078c7a064281b568615c844e3654ae",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/pipeline",
+      "repo_root": null,
+      "git": {
+        "sha": "d4d4d4d4d4",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/pipeline",
+      "git": {
+        "sha": "d4d4d4d4d4",
+        "branch": "main"
+      },
+      "model": "gemini-2.5-pro",
+      "tool_version": "0.4.1"
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/subagent_thread.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-subagent-0001",
+    "seq": 1,
+    "event_id": "msg-sa-01",
+    "parent_id": null,
+    "timestamp": "2026-06-22T16:00:00Z",
+    "project": {
+      "cwd": "/home/dev/pipeline",
+      "repo_root": null,
+      "git": {
+        "sha": "d4d4d4d4d4",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Kick off the ingest refactor; delegate the parser rewrite to a focused sub-thread.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/subagent_thread.jsonl",
+      "byte_offset": 187,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-subagent-0001",
+    "seq": 2,
+    "event_id": "msg-sa-02",
+    "parent_id": null,
+    "timestamp": "2026-06-22T16:00:02Z",
+    "project": {
+      "cwd": "/home/dev/pipeline",
+      "repo_root": null,
+      "git": {
+        "sha": "d4d4d4d4d4",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Spawning a secondary thread to handle the parser rewrite.",
+      "thinking": "Gemini has no real subagents; modeling the sub-task as a nested threadId on the same session.",
+      "model": null,
+      "usage": {
+        "input_tokens": 90,
+        "output_tokens": 40,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/subagent_thread.jsonl",
+      "byte_offset": 369,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-subagent-0001",
+    "seq": 3,
+    "event_id": "msg-sa-03",
+    "parent_id": null,
+    "timestamp": "2026-06-22T16:01:00Z",
+    "project": {
+      "cwd": "/home/dev/pipeline",
+      "repo_root": null,
+      "git": {
+        "sha": "d4d4d4d4d4",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "In this sub-thread, use a streaming pull-parser for the ingest format instead of the buffered one.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/subagent_thread.jsonl",
+      "byte_offset": 668,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-subagent-0001",
+    "seq": 4,
+    "event_id": "msg-sa-04",
+    "parent_id": null,
+    "timestamp": "2026-06-22T16:01:03Z",
+    "project": {
+      "cwd": "/home/dev/pipeline",
+      "repo_root": null,
+      "git": {
+        "sha": "d4d4d4d4d4",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Rewriting the parser to a streaming pull-parser.",
+      "thinking": "Replace the buffered reader in parser.rs.",
+      "model": null,
+      "usage": {
+        "input_tokens": 102,
+        "output_tokens": 57,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/subagent_thread.jsonl",
+      "byte_offset": 896,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-subagent-0001",
+    "seq": 5,
+    "event_id": "call-sa-parser:call",
+    "parent_id": "msg-sa-04",
+    "timestamp": "2026-06-22T16:01:03Z",
+    "project": {
+      "cwd": "/home/dev/pipeline",
+      "repo_root": null,
+      "git": {
+        "sha": "d4d4d4d4d4",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-sa-parser",
+      "name": "replace",
+      "args": {
+        "path": "src/ingest/parser.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/subagent_thread.jsonl",
+      "byte_offset": 896,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-subagent-0001",
+    "seq": 6,
+    "event_id": "call-sa-parser:result",
+    "parent_id": "msg-sa-04",
+    "timestamp": "2026-06-22T16:01:03Z",
+    "project": {
+      "cwd": "/home/dev/pipeline",
+      "repo_root": null,
+      "git": {
+        "sha": "d4d4d4d4d4",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-sa-parser",
+      "ok": true,
+      "output": {
+        "fileName": "src/ingest/parser.rs",
+        "originalContent": "let buf = read_all(input);\n",
+        "newContent": "let mut reader = PullParser::new(input);\n",
+        "fileDiff": "@@ -1 +1 @@\n-let buf = read_all(input);\n+let mut reader = PullParser::new(input);\n",
+        "diffStat": {
+          "added": 1,
+          "removed": 1
+        }
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/subagent_thread.jsonl",
+      "byte_offset": 896,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-subagent-0001",
+    "seq": 7,
+    "event_id": "call-sa-parser:edit",
+    "parent_id": "msg-sa-04",
+    "timestamp": "2026-06-22T16:01:03Z",
+    "project": {
+      "cwd": "/home/dev/pipeline",
+      "repo_root": null,
+      "git": {
+        "sha": "d4d4d4d4d4",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-sa-parser",
+      "diff": {
+        "path": "src/ingest/parser.rs",
+        "old": "let buf = read_all(input);\n",
+        "new": "let mut reader = PullParser::new(input);\n",
+        "unified": "@@ -1 +1 @@\n-let buf = read_all(input);\n+let mut reader = PullParser::new(input);\n",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/subagent_thread.jsonl",
+      "byte_offset": 896,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "g-subagent-0001",
+    "seq": 8,
+    "event_id": "msg-sa-05",
+    "parent_id": null,
+    "timestamp": "2026-06-22T16:02:00Z",
+    "project": {
+      "cwd": "/home/dev/pipeline",
+      "repo_root": null,
+      "git": {
+        "sha": "d4d4d4d4d4",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Sub-thread done, looks good.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/subagent_thread.jsonl",
+      "byte_offset": 1556,
+      "line_no": 6
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__tool_failure.snap
new file mode 100644
index 0000000..4ce295a
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__gemini__v1__tool_failure.snap
@@ -0,0 +1,237 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0004",
+    "seq": 0,
+    "event_id": "7036cef4e9e98fd546db9000e4dc951075772cb75bcfcbd8a57b5ec99f50d4d5",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/service",
+      "repo_root": null,
+      "git": {
+        "sha": "abc0001234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/service",
+      "git": {
+        "sha": "abc0001234",
+        "branch": "main"
+      },
+      "model": "gemini-2.5-pro",
+      "tool_version": "0.4.1"
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/tool_failure.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0004",
+    "seq": 1,
+    "event_id": "msg-0301",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:45:00Z",
+    "project": {
+      "cwd": "/home/dev/service",
+      "repo_root": null,
+      "git": {
+        "sha": "abc0001234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Add a rate limiter to the gateway handler.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/tool_failure.jsonl",
+      "byte_offset": 149,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0004",
+    "seq": 2,
+    "event_id": "msg-0302",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:45:03Z",
+    "project": {
+      "cwd": "/home/dev/service",
+      "repo_root": null,
+      "git": {
+        "sha": "abc0001234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "I'll add a token-bucket limiter to the handler.",
+      "thinking": "Edit gateway.rs to insert the limiter.",
+      "model": null,
+      "usage": {
+        "input_tokens": 95,
+        "output_tokens": 40,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/tool_failure.jsonl",
+      "byte_offset": 272,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0004",
+    "seq": 3,
+    "event_id": "call-edit-fail:call",
+    "parent_id": "msg-0302",
+    "timestamp": "2026-06-22T13:45:03Z",
+    "project": {
+      "cwd": "/home/dev/service",
+      "repo_root": null,
+      "git": {
+        "sha": "abc0001234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-edit-fail",
+      "name": "replace",
+      "args": {
+        "path": "src/gateway.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/tool_failure.jsonl",
+      "byte_offset": 272,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0004",
+    "seq": 4,
+    "event_id": "call-edit-fail:result",
+    "parent_id": "msg-0302",
+    "timestamp": "2026-06-22T13:45:03Z",
+    "project": {
+      "cwd": "/home/dev/service",
+      "repo_root": null,
+      "git": {
+        "sha": "abc0001234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-edit-fail",
+      "ok": false,
+      "output": {
+        "fileName": "src/gateway.rs",
+        "originalContent": "fn handle() {}\n",
+        "newContent": "fn handle() { limiter.check(); }\n",
+        "fileDiff": "@@ -1 +1 @@\n-fn handle() {}\n+fn handle() { limiter.check(); }\n",
+        "diffStat": {
+          "added": 1,
+          "removed": 1
+        },
+        "error": "write failed: read-only file system"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/tool_failure.jsonl",
+      "byte_offset": 272,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0004",
+    "seq": 5,
+    "event_id": "call-edit-fail:edit",
+    "parent_id": "msg-0302",
+    "timestamp": "2026-06-22T13:45:03Z",
+    "project": {
+      "cwd": "/home/dev/service",
+      "repo_root": null,
+      "git": {
+        "sha": "abc0001234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-edit-fail",
+      "diff": {
+        "path": "src/gateway.rs",
+        "old": "fn handle() {}\n",
+        "new": "fn handle() { limiter.check(); }\n",
+        "unified": "@@ -1 +1 @@\n-fn handle() {}\n+fn handle() { limiter.check(); }\n",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/tool_failure.jsonl",
+      "byte_offset": 272,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "gemini",
+    "session_id": "a1b2c3d4-0004",
+    "seq": 6,
+    "event_id": "msg-0303",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:45:05Z",
+    "project": {
+      "cwd": "/home/dev/service",
+      "repo_root": null,
+      "git": {
+        "sha": "abc0001234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "The edit failed because the file system is read-only. Please re-mount with write access and I'll retry.",
+      "thinking": null,
+      "model": null,
+      "usage": {
+        "input_tokens": 60,
+        "output_tokens": 33,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/gemini/v1/tool_failure.jsonl",
+      "byte_offset": 872,
+      "line_no": 4
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__otel__genai__ban.snap b/crates/memscribe-testkit/tests/snapshots/events__otel__genai__ban.snap
new file mode 100644
index 0000000..2420688
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__otel__genai__ban.snap
@@ -0,0 +1,110 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-ban-001",
+    "seq": 0,
+    "event_id": "031db5d30a89efdce1f49029d63da3e76da287a91621e52aa669404d6378b5c3:session_start:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T12:00:00Z",
+    "project": {
+      "cwd": "/home/dev/core",
+      "repo_root": "/home/dev/core",
+      "git": {
+        "sha": "9abcdef",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/core",
+      "git": {
+        "sha": "9abcdef",
+        "branch": "main"
+      },
+      "model": "gemini-2.5-pro",
+      "tool_version": null
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/ban.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-ban-001",
+    "seq": 1,
+    "event_id": "c1c3f679b52d8ea3b11db0c3cadce544f3d942b89683293b2e19d15cda5b170c:user_prompt:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T12:00:06Z",
+    "project": {
+      "cwd": "/home/dev/core",
+      "repo_root": "/home/dev/core",
+      "git": {
+        "sha": "9abcdef",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "We will never add a dependency on left-pad. Keep the tree clean.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "We will never add a dependency on left-pad. Keep the tree clean."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/ban.jsonl",
+      "byte_offset": 272,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-ban-001",
+    "seq": 2,
+    "event_id": "f24ec653cf8d2ee6ae5656ccfdd758c6ac37ee7cfa3e99235fde118f7893a5fc:output_msg:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T12:00:10Z",
+    "project": {
+      "cwd": "/home/dev/core",
+      "repo_root": "/home/dev/core",
+      "git": {
+        "sha": "9abcdef",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Acknowledged. No left-pad dependency will be added.",
+      "thinking": null,
+      "model": "gemini-2.5-pro",
+      "usage": {
+        "input_tokens": 90,
+        "output_tokens": 18,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Acknowledged. No left-pad dependency will be added."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/ban.jsonl",
+      "byte_offset": 492,
+      "line_no": 3
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__otel__genai__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/events__otel__genai__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..d048032
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__otel__genai__happy_path_decision_then_edits.snap
@@ -0,0 +1,260 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-happy-001",
+    "seq": 0,
+    "event_id": "59d021101e643dbc72408bc413a3c432a2cf36fd7fec2c942a3435d363614c07:session_start:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:00Z",
+    "project": {
+      "cwd": "/home/dev/svc",
+      "repo_root": "/home/dev/svc",
+      "git": {
+        "sha": "abc1234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/svc",
+      "git": {
+        "sha": "abc1234",
+        "branch": "main"
+      },
+      "model": "claude-opus-4-8",
+      "tool_version": null
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-happy-001",
+    "seq": 1,
+    "event_id": "0dbbdc53b19ef8a702bfe04f12e1b336afcf36ab3e87cdc1a15a4293edc80b80:user_prompt:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:05Z",
+    "project": {
+      "cwd": "/home/dev/svc",
+      "repo_root": "/home/dev/svc",
+      "git": {
+        "sha": "abc1234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use Postgres instead of MySQL for the persistence layer.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Let's use Postgres instead of MySQL for the persistence layer."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 276,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-happy-001",
+    "seq": 2,
+    "event_id": "d55e8dfa088302e936e4b88aa6f43ca39bcb197b29cc9c099fe3598a82c9fe11:output_msg:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:09Z",
+    "project": {
+      "cwd": "/home/dev/svc",
+      "repo_root": "/home/dev/svc",
+      "git": {
+        "sha": "abc1234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Agreed. I'll switch the persistence layer to Postgres and update the config.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 420,
+        "output_tokens": 88,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Agreed. I'll switch the persistence layer to Postgres and update the config."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 499,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-happy-001",
+    "seq": 3,
+    "event_id": "c7cc03ee310af61eb6f93be5d7f0d1b483711c6ca26269b1592be2bc851a34a8:tool_call:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:12Z",
+    "project": {
+      "cwd": "/home/dev/svc",
+      "repo_root": "/home/dev/svc",
+      "git": {
+        "sha": "abc1234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-edit-1",
+      "name": "edit_file",
+      "args": null
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 874,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-happy-001",
+    "seq": 4,
+    "event_id": "c7cc03ee310af61eb6f93be5d7f0d1b483711c6ca26269b1592be2bc851a34a8:tool_result:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:12Z",
+    "project": {
+      "cwd": "/home/dev/svc",
+      "repo_root": "/home/dev/svc",
+      "git": {
+        "sha": "abc1234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-edit-1",
+      "ok": true,
+      "output": "ok"
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 874,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-happy-001",
+    "seq": 5,
+    "event_id": "c7cc03ee310af61eb6f93be5d7f0d1b483711c6ca26269b1592be2bc851a34a8:file_edit:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:12Z",
+    "project": {
+      "cwd": "/home/dev/svc",
+      "repo_root": "/home/dev/svc",
+      "git": {
+        "sha": "abc1234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-edit-1",
+      "diff": {
+        "path": "db/config.rs",
+        "old": null,
+        "new": null,
+        "unified": null,
+        "added_lines": 12,
+        "removed_lines": 4
+      }
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 874,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-happy-001",
+    "seq": 6,
+    "event_id": "610a6620f9def7e4d0ff0e62caa2ca8d9b27d751fa2f17f4db7198e831228656:file_edit:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:15Z",
+    "project": {
+      "cwd": "/home/dev/svc",
+      "repo_root": "/home/dev/svc",
+      "git": {
+        "sha": "abc1234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "db/pool.rs",
+        "old": null,
+        "new": null,
+        "unified": null,
+        "added_lines": 30,
+        "removed_lines": 9
+      }
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1182,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-happy-001",
+    "seq": 7,
+    "event_id": "5f057a5bb00511120a31401a5018da0493b454b21f45003f34ca98a9ec076aa3:session_end:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:20Z",
+    "project": {
+      "cwd": "/home/dev/svc",
+      "repo_root": "/home/dev/svc",
+      "git": {
+        "sha": "abc1234",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_end",
+      "reason": "completed"
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1395,
+      "line_no": 6
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__otel__genai__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/events__otel__genai__rejected_alternative.snap
new file mode 100644
index 0000000..1cfeb60
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__otel__genai__rejected_alternative.snap
@@ -0,0 +1,144 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-reject-001",
+    "seq": 0,
+    "event_id": "d84a04e91845802b1f222093afb3cd2b8885a085f17c7545b7162dedbe057855:session_start:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:00:00Z",
+    "project": {
+      "cwd": "/home/dev/checkout",
+      "repo_root": "/home/dev/checkout",
+      "git": {
+        "sha": "def5678",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/checkout",
+      "git": {
+        "sha": "def5678",
+        "branch": "feat/payments"
+      },
+      "model": "gpt-5",
+      "tool_version": null
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/rejected_alternative.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-reject-001",
+    "seq": 1,
+    "event_id": "9e46ceb387c7779c4193b18b69f41e09a4bb4583dd17a3402f31dbc169944c3c:input_msg:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:00:04Z",
+    "project": {
+      "cwd": "/home/dev/checkout",
+      "repo_root": "/home/dev/checkout",
+      "git": {
+        "sha": "def5678",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "We should use Stripe instead of PayPal for the checkout integration.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "We should use Stripe instead of PayPal for the checkout integration."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/rejected_alternative.jsonl",
+      "byte_offset": 283,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-reject-001",
+    "seq": 2,
+    "event_id": "c2f76d0e15483645951e6e80c296d723ead4a32266243ec3d73ae34486aa5883:output_msg:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:00:08Z",
+    "project": {
+      "cwd": "/home/dev/checkout",
+      "repo_root": "/home/dev/checkout",
+      "git": {
+        "sha": "def5678",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Understood, I'll wire up Stripe and drop the PayPal path.",
+      "thinking": null,
+      "model": "gpt-5",
+      "usage": {
+        "input_tokens": 210,
+        "output_tokens": 40,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": [
+        {
+          "part": "text",
+          "text": "Understood, I'll wire up Stripe and drop the PayPal path."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/rejected_alternative.jsonl",
+      "byte_offset": 537,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-reject-001",
+    "seq": 3,
+    "event_id": "8ecf501b3fff25d6cae667ef82fe84d7fb440e37f6fb7eab0a77c810e84de7e7:file_edit:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:00:12Z",
+    "project": {
+      "cwd": "/home/dev/checkout",
+      "repo_root": "/home/dev/checkout",
+      "git": {
+        "sha": "def5678",
+        "branch": "feat/payments"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "checkout/stripe.rs",
+        "old": null,
+        "new": null,
+        "unified": null,
+        "added_lines": 48,
+        "removed_lines": 0
+      }
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/rejected_alternative.jsonl",
+      "byte_offset": 881,
+      "line_no": 4
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__otel__genai__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/events__otel__genai__tool_failure.snap
new file mode 100644
index 0000000..ee79e90
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__otel__genai__tool_failure.snap
@@ -0,0 +1,160 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-fail-001",
+    "seq": 0,
+    "event_id": "1ca08f1a290b7a7e923e39b32ef242b470dcf22e5b4e55f2fdcb4c59bfb85588:session_start:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:00:00Z",
+    "project": {
+      "cwd": "/home/dev/svc",
+      "repo_root": "/home/dev/svc",
+      "git": {
+        "sha": "badf00d",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/svc",
+      "git": {
+        "sha": "badf00d",
+        "branch": "main"
+      },
+      "model": "claude-opus-4-8",
+      "tool_version": null
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/tool_failure.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-fail-001",
+    "seq": 1,
+    "event_id": "941ec2df1d06a67528a2ba8fc5fb3dcbe732d81b5b840db9716e5d5591d3b956:user_prompt:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:00:05Z",
+    "project": {
+      "cwd": "/home/dev/svc",
+      "repo_root": "/home/dev/svc",
+      "git": {
+        "sha": "badf00d",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Switch the cache layer over to Redis.",
+      "parts": [
+        {
+          "part": "text",
+          "text": "Switch the cache layer over to Redis."
+        }
+      ]
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/tool_failure.jsonl",
+      "byte_offset": 275,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-fail-001",
+    "seq": 2,
+    "event_id": "765089f0ab456073510a7df920d4f6a28c8a1df57be18a52c1214d1ab8950873:tool_call:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:00:09Z",
+    "project": {
+      "cwd": "/home/dev/svc",
+      "repo_root": "/home/dev/svc",
+      "git": {
+        "sha": "badf00d",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-edit-fail",
+      "name": "edit_file",
+      "args": null
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/tool_failure.jsonl",
+      "byte_offset": 472,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-fail-001",
+    "seq": 3,
+    "event_id": "765089f0ab456073510a7df920d4f6a28c8a1df57be18a52c1214d1ab8950873:tool_result:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:00:09Z",
+    "project": {
+      "cwd": "/home/dev/svc",
+      "repo_root": "/home/dev/svc",
+      "git": {
+        "sha": "badf00d",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-edit-fail",
+      "ok": false,
+      "output": "hunk failed to apply: context mismatch"
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/tool_failure.jsonl",
+      "byte_offset": 472,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "otel",
+    "session_id": "sess-fail-001",
+    "seq": 4,
+    "event_id": "765089f0ab456073510a7df920d4f6a28c8a1df57be18a52c1214d1ab8950873:file_edit:0",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:00:09Z",
+    "project": {
+      "cwd": "/home/dev/svc",
+      "repo_root": "/home/dev/svc",
+      "git": {
+        "sha": "badf00d",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-edit-fail",
+      "diff": {
+        "path": "cache/redis.rs",
+        "old": null,
+        "new": null,
+        "unified": null,
+        "added_lines": 20,
+        "removed_lines": 2
+      }
+    },
+    "provenance": {
+      "file": "fixtures/otel/genai/tool_failure.jsonl",
+      "byte_offset": 472,
+      "line_no": 3
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__ban.snap b/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__ban.snap
new file mode 100644
index 0000000..f14a3e4
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__ban.snap
@@ -0,0 +1,134 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-3",
+    "seq": 0,
+    "event_id": "5a358f6bc035fac8f141845d388596f8a927b30b8efdc8f97f7d3af247899d18",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0011223344556677",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/acme-api",
+      "git": {
+        "sha": "0011223344556677",
+        "branch": "main"
+      },
+      "model": "gpt-4o",
+      "tool_version": "1.92.0"
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/ban.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-3",
+    "seq": 1,
+    "event_id": "b1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T12:00:00Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0011223344556677",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "We will never add a dependency on left-pad. Roll our own padding helper.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/ban.jsonl",
+      "byte_offset": 166,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-3",
+    "seq": 2,
+    "event_id": "b2",
+    "parent_id": "b1",
+    "timestamp": "2026-06-22T12:00:06Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0011223344556677",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Understood — no left-pad. I'll add a tiny internal pad() utility.",
+      "thinking": null,
+      "model": "gpt-4o",
+      "usage": {
+        "input_tokens": 640,
+        "output_tokens": 120,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/ban.jsonl",
+      "byte_offset": 343,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-3",
+    "seq": 3,
+    "event_id": "b2:edit:0:src/util/pad.ts",
+    "parent_id": "b2",
+    "timestamp": "2026-06-22T12:00:06Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "0011223344556677",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "src/util/pad.ts",
+        "old": null,
+        "new": "export const pad = (s, n) => s.padStart(n, '0');",
+        "unified": null,
+        "added_lines": 1,
+        "removed_lines": 0
+      }
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/ban.jsonl",
+      "byte_offset": 343,
+      "line_no": 3
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..b2daf18
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__happy_path_decision_then_edits.snap
@@ -0,0 +1,221 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-1",
+    "seq": 0,
+    "event_id": "5cc7d42f773a8bb1e8d79d3f015b036fdb1fdc271110aa72cb1edbab564edc6c",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/acme-api",
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      },
+      "model": "gpt-4o",
+      "tool_version": "1.92.0"
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-1",
+    "seq": 1,
+    "event_id": "m1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:00Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use Postgres instead of MySQL for the persistence layer.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 162,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-1",
+    "seq": 2,
+    "event_id": "m2",
+    "parent_id": "m1",
+    "timestamp": "2026-06-22T10:00:05Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Good call. I'll switch the connection string and update the schema module.",
+      "thinking": null,
+      "model": "gpt-4o",
+      "usage": {
+        "input_tokens": 1200,
+        "output_tokens": 340,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 329,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-1",
+    "seq": 3,
+    "event_id": "m2:edit:0:src/db/connect.ts",
+    "parent_id": "m2",
+    "timestamp": "2026-06-22T10:00:05Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "src/db/connect.ts",
+        "old": "const url = 'mysql://localhost/acme';",
+        "new": "const url = 'postgres://localhost/acme';",
+        "unified": null,
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 329,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-1",
+    "seq": 4,
+    "event_id": "m2:edit:1:src/db/schema.sql",
+    "parent_id": "m2",
+    "timestamp": "2026-06-22T10:00:05Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "src/db/schema.sql",
+        "old": "ENGINE=InnoDB",
+        "new": "-- postgres has no storage engines",
+        "unified": "@@ -1 +1 @@\n-ENGINE=InnoDB\n+-- postgres has no storage engines",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 329,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-1",
+    "seq": 5,
+    "event_id": "m3",
+    "parent_id": "m2",
+    "timestamp": "2026-06-22T10:01:00Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Looks good, thanks.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 928,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-1",
+    "seq": 6,
+    "event_id": "67f82833d61938bb88349e0a2aaa0d31882cf90d17c1d9099f6dcdf2720b9927",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:02:00Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_end",
+      "reason": "user_closed"
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1052,
+      "line_no": 5
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__native_chatsession.snap b/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__native_chatsession.snap
new file mode 100644
index 0000000..ea5ba94
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__native_chatsession.snap
@@ -0,0 +1,57 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-native-1",
+    "seq": 0,
+    "event_id": "2c588cc1bef79d0e417388725b98f9ba899e146fc6f96692cd7af92dade2324a:req:0:user",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": ".",
+      "repo_root": null,
+      "git": null
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Add a health check endpoint",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/native_chatsession.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-native-1",
+    "seq": 1,
+    "event_id": "2c588cc1bef79d0e417388725b98f9ba899e146fc6f96692cd7af92dade2324a:req:0:asst",
+    "parent_id": "2c588cc1bef79d0e417388725b98f9ba899e146fc6f96692cd7af92dade2324a:req:0:user",
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": ".",
+      "repo_root": null,
+      "git": null
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "I'll add a GET /healthz route that returns 200.Done.",
+      "thinking": null,
+      "model": "GitHub Copilot",
+      "usage": null,
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/native_chatsession.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__rejected_alternative.snap
new file mode 100644
index 0000000..0889891
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__rejected_alternative.snap
@@ -0,0 +1,164 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-2",
+    "seq": 0,
+    "event_id": "c9969e0271967c46507a9b377be675b3d20feefc7ce35269609e82873e75ba51",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f6e5d4c3b2a1",
+        "branch": "feat/cache"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/acme-api",
+      "git": {
+        "sha": "f6e5d4c3b2a1",
+        "branch": "feat/cache"
+      },
+      "model": "gpt-4o",
+      "tool_version": "1.92.0"
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/rejected_alternative.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-2",
+    "seq": 1,
+    "event_id": "r1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:00:00Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f6e5d4c3b2a1",
+        "branch": "feat/cache"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Should we use Redis or just an in-memory LRU for the cache?",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/rejected_alternative.jsonl",
+      "byte_offset": 168,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-2",
+    "seq": 2,
+    "event_id": "r2",
+    "parent_id": "r1",
+    "timestamp": "2026-06-22T11:00:08Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f6e5d4c3b2a1",
+        "branch": "feat/cache"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "We considered Redis but rejected it: it adds an external dependency we don't want for this small cache. We'll go with an in-memory LRU instead.",
+      "thinking": null,
+      "model": "gpt-4o",
+      "usage": {
+        "input_tokens": 900,
+        "output_tokens": 210,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/rejected_alternative.jsonl",
+      "byte_offset": 332,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-2",
+    "seq": 3,
+    "event_id": "r3",
+    "parent_id": "r2",
+    "timestamp": "2026-06-22T11:00:20Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f6e5d4c3b2a1",
+        "branch": "feat/cache"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Implementing the LRU now.",
+      "thinking": null,
+      "model": "gpt-4o",
+      "usage": null,
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/rejected_alternative.jsonl",
+      "byte_offset": 637,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-2",
+    "seq": 4,
+    "event_id": "r3:edit:0:src/cache/lru.ts",
+    "parent_id": "r3",
+    "timestamp": "2026-06-22T11:00:20Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "f6e5d4c3b2a1",
+        "branch": "feat/cache"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "src/cache/lru.ts",
+        "old": null,
+        "new": "export class Lru { /* ... */ }",
+        "unified": null,
+        "added_lines": 1,
+        "removed_lines": 0
+      }
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/rejected_alternative.jsonl",
+      "byte_offset": 637,
+      "line_no": 4
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__tool_failure.snap
new file mode 100644
index 0000000..6d2c637
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__vscode__v1__tool_failure.snap
@@ -0,0 +1,222 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-4",
+    "seq": 0,
+    "event_id": "fa460e6ac6dd971ca3eff097faf3675689623725973b8a001e36e1f11a4c8a62",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "99aabbccddeeff00",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/home/dev/acme-api",
+      "git": {
+        "sha": "99aabbccddeeff00",
+        "branch": "main"
+      },
+      "model": "gpt-4o",
+      "tool_version": "1.92.0"
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/tool_failure.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-4",
+    "seq": 1,
+    "event_id": "t1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:00:00Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "99aabbccddeeff00",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Rename the config flag enableV2 to enableV2Beta everywhere.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/tool_failure.jsonl",
+      "byte_offset": 166,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-4",
+    "seq": 2,
+    "event_id": "t2",
+    "parent_id": "t1",
+    "timestamp": "2026-06-22T13:00:07Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "99aabbccddeeff00",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "I'll apply the edit to src/config.ts.",
+      "thinking": null,
+      "model": "gpt-4o",
+      "usage": {
+        "input_tokens": 700,
+        "output_tokens": 160,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/tool_failure.jsonl",
+      "byte_offset": 330,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-4",
+    "seq": 3,
+    "event_id": "t2:toolcall:0:call-edit-1",
+    "parent_id": "t2",
+    "timestamp": "2026-06-22T13:00:07Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "99aabbccddeeff00",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-edit-1",
+      "name": "applyEdit",
+      "args": {
+        "path": "src/config.ts"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/tool_failure.jsonl",
+      "byte_offset": 330,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-4",
+    "seq": 4,
+    "event_id": "t2:edit:0:src/config.ts",
+    "parent_id": "t2",
+    "timestamp": "2026-06-22T13:00:07Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "99aabbccddeeff00",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-edit-1",
+      "diff": {
+        "path": "src/config.ts",
+        "old": "enableV2",
+        "new": "enableV2Beta",
+        "unified": null,
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/tool_failure.jsonl",
+      "byte_offset": 330,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-4",
+    "seq": 5,
+    "event_id": "t3",
+    "parent_id": "t2",
+    "timestamp": "2026-06-22T13:00:09Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "99aabbccddeeff00",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/tool_failure.jsonl",
+      "byte_offset": 741,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "vs_code",
+    "session_id": "vsc-sess-4",
+    "seq": 6,
+    "event_id": "t3:toolresult:0:call-edit-1",
+    "parent_id": "t3",
+    "timestamp": "2026-06-22T13:00:09Z",
+    "project": {
+      "cwd": "/home/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "99aabbccddeeff00",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-edit-1",
+      "ok": false,
+      "output": "EDIT FAILED: file src/config.ts is read-only on disk"
+    },
+    "provenance": {
+      "file": "fixtures/vscode/v1/tool_failure.jsonl",
+      "byte_offset": 741,
+      "line_no": 4
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__windsurf__v1__ban.snap b/crates/memscribe-testkit/tests/snapshots/events__windsurf__v1__ban.snap
new file mode 100644
index 0000000..33bce25
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__windsurf__v1__ban.snap
@@ -0,0 +1,192 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-ban-003",
+    "seq": 0,
+    "event_id": "65236230bca6bd72086a65e34bf3c1decb637d8522440c75fff8eec82e5c86e2",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "77c0de01",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme-api",
+      "git": {
+        "sha": "77c0de01",
+        "branch": "main"
+      },
+      "model": "cascade-base",
+      "tool_version": "1.8.2"
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/ban.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-ban-003",
+    "seq": 1,
+    "event_id": "u1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T16:20:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "77c0de01",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "We will never add a dependency on left-pad. Implement padding inline instead.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/ban.jsonl",
+      "byte_offset": 164,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-ban-003",
+    "seq": 2,
+    "event_id": "a1",
+    "parent_id": "u1",
+    "timestamp": "2026-06-22T16:20:09Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "77c0de01",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Understood — no left-pad dependency. I'll add a small inline pad helper.",
+      "thinking": null,
+      "model": "cascade-base",
+      "usage": {
+        "input_tokens": 300,
+        "output_tokens": 140,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/ban.jsonl",
+      "byte_offset": 346,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-ban-003",
+    "seq": 3,
+    "event_id": "a1#toolcall:call_pad_1",
+    "parent_id": "a1",
+    "timestamp": "2026-06-22T16:20:09Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "77c0de01",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_pad_1",
+      "name": "edit_file",
+      "args": {
+        "path": "src/util/pad.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/ban.jsonl",
+      "byte_offset": 346,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-ban-003",
+    "seq": 4,
+    "event_id": "a1#toolresult:call_pad_1",
+    "parent_id": "a1",
+    "timestamp": "2026-06-22T16:20:09Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "77c0de01",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_pad_1",
+      "ok": true,
+      "output": "applied"
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/ban.jsonl",
+      "byte_offset": 346,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-ban-003",
+    "seq": 5,
+    "event_id": "a1#edit:0",
+    "parent_id": "a1",
+    "timestamp": "2026-06-22T16:20:09Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "77c0de01",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_pad_1",
+      "diff": {
+        "path": "src/util/pad.rs",
+        "old": "",
+        "new": "pub fn pad_left(s: &str, width: usize) -> String { format!(\"{s:>width$}\") }",
+        "unified": "@@ -0,0 +1,1 @@\n+pub fn pad_left(s: &str, width: usize) -> String { format!(\"{s:>width$}\") }",
+        "added_lines": 1,
+        "removed_lines": 0
+      }
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/ban.jsonl",
+      "byte_offset": 346,
+      "line_no": 3
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__windsurf__v1__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/events__windsurf__v1__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..aa5702c
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__windsurf__v1__happy_path_decision_then_edits.snap
@@ -0,0 +1,311 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-happy-001",
+    "seq": 0,
+    "event_id": "426e144c4a989f842c70156c36b6777b14250cc54a02889e2e55e95831f1d4c4",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f3c1a7e",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme-api",
+      "git": {
+        "sha": "9f3c1a7e",
+        "branch": "main"
+      },
+      "model": "cascade-base",
+      "tool_version": "1.8.2"
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-happy-001",
+    "seq": 1,
+    "event_id": "u1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T14:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f3c1a7e",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use Postgres instead of MySQL for the orders service.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 166,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-happy-001",
+    "seq": 2,
+    "event_id": "a1",
+    "parent_id": "u1",
+    "timestamp": "2026-06-22T14:00:06Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f3c1a7e",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Good call. I'll swap the driver and update the connection string.",
+      "thinking": null,
+      "model": "cascade-base",
+      "usage": {
+        "input_tokens": 420,
+        "output_tokens": 188,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 332,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-happy-001",
+    "seq": 3,
+    "event_id": "a1#toolcall:call_db_1",
+    "parent_id": "a1",
+    "timestamp": "2026-06-22T14:00:06Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f3c1a7e",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_db_1",
+      "name": "edit_file",
+      "args": {
+        "path": "src/db/mod.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 332,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-happy-001",
+    "seq": 4,
+    "event_id": "a1#toolcall:call_cfg_1",
+    "parent_id": "a1",
+    "timestamp": "2026-06-22T14:00:06Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f3c1a7e",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_cfg_1",
+      "name": "edit_file",
+      "args": {
+        "path": "config/database.toml"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 332,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-happy-001",
+    "seq": 5,
+    "event_id": "a1#toolresult:call_db_1",
+    "parent_id": "a1",
+    "timestamp": "2026-06-22T14:00:06Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f3c1a7e",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_db_1",
+      "ok": true,
+      "output": "applied"
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 332,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-happy-001",
+    "seq": 6,
+    "event_id": "a1#toolresult:call_cfg_1",
+    "parent_id": "a1",
+    "timestamp": "2026-06-22T14:00:06Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f3c1a7e",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_cfg_1",
+      "ok": true,
+      "output": "applied"
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 332,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-happy-001",
+    "seq": 7,
+    "event_id": "a1#edit:0",
+    "parent_id": "a1",
+    "timestamp": "2026-06-22T14:00:06Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f3c1a7e",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_db_1",
+      "diff": {
+        "path": "src/db/mod.rs",
+        "old": "use mysql_async::Pool;",
+        "new": "use sqlx::postgres::PgPool;",
+        "unified": "@@ -1,1 +1,1 @@\n-use mysql_async::Pool;\n+use sqlx::postgres::PgPool;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 332,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-happy-001",
+    "seq": 8,
+    "event_id": "a1#edit:1",
+    "parent_id": "a1",
+    "timestamp": "2026-06-22T14:00:06Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f3c1a7e",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_cfg_1",
+      "diff": {
+        "path": "config/database.toml",
+        "old": "driver = \"mysql\"\nport = 3306",
+        "new": "driver = \"postgres\"\nport = 5432",
+        "unified": "@@ -1,2 +1,2 @@\n-driver = \"mysql\"\n-port = 3306\n+driver = \"postgres\"\n+port = 5432",
+        "added_lines": 2,
+        "removed_lines": 2
+      }
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 332,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-happy-001",
+    "seq": 9,
+    "event_id": "u2",
+    "parent_id": "a1",
+    "timestamp": "2026-06-22T14:01:10Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "9f3c1a7e",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Looks right, thanks.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1332,
+      "line_no": 4
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__windsurf__v1__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/events__windsurf__v1__rejected_alternative.snap
new file mode 100644
index 0000000..0f5b85e
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__windsurf__v1__rejected_alternative.snap
@@ -0,0 +1,254 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-reject-002",
+    "seq": 0,
+    "event_id": "8ef09b7040fcbdab2ce3675d1726c6f5e3974e2c78ac5efb08c99721edc66173",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "1b2d44aa",
+        "branch": "feature/cache"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme-api",
+      "git": {
+        "sha": "1b2d44aa",
+        "branch": "feature/cache"
+      },
+      "model": "cascade-pro",
+      "tool_version": "1.8.2"
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/rejected_alternative.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-reject-002",
+    "seq": 1,
+    "event_id": "u1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T15:10:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "1b2d44aa",
+        "branch": "feature/cache"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Should we use Redis or Memcached for the session cache?",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/rejected_alternative.jsonl",
+      "byte_offset": 175,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-reject-002",
+    "seq": 2,
+    "event_id": "a1",
+    "parent_id": "u1",
+    "timestamp": "2026-06-22T15:10:08Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "1b2d44aa",
+        "branch": "feature/cache"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "I considered Memcached, but we already run Redis for the queue, so I'll reuse Redis and skip Memcached to avoid a second dependency.",
+      "thinking": null,
+      "model": "cascade-pro",
+      "usage": {
+        "input_tokens": 510,
+        "output_tokens": 233,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/rejected_alternative.jsonl",
+      "byte_offset": 338,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-reject-002",
+    "seq": 3,
+    "event_id": "u2",
+    "parent_id": "a1",
+    "timestamp": "2026-06-22T15:11:30Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "1b2d44aa",
+        "branch": "feature/cache"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Agreed, do not add Memcached. Wire up Redis.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/rejected_alternative.jsonl",
+      "byte_offset": 640,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-reject-002",
+    "seq": 4,
+    "event_id": "a2",
+    "parent_id": "u2",
+    "timestamp": "2026-06-22T15:11:40Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "1b2d44aa",
+        "branch": "feature/cache"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Wiring Redis as the session store.",
+      "thinking": null,
+      "model": "cascade-pro",
+      "usage": {
+        "input_tokens": 540,
+        "output_tokens": 120,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/rejected_alternative.jsonl",
+      "byte_offset": 792,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-reject-002",
+    "seq": 5,
+    "event_id": "a2#toolcall:call_cache_1",
+    "parent_id": "a2",
+    "timestamp": "2026-06-22T15:11:40Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "1b2d44aa",
+        "branch": "feature/cache"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_cache_1",
+      "name": "edit_file",
+      "args": {
+        "path": "src/cache.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/rejected_alternative.jsonl",
+      "byte_offset": 792,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-reject-002",
+    "seq": 6,
+    "event_id": "a2#toolresult:call_cache_1",
+    "parent_id": "a2",
+    "timestamp": "2026-06-22T15:11:40Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "1b2d44aa",
+        "branch": "feature/cache"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_cache_1",
+      "ok": true,
+      "output": "applied"
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/rejected_alternative.jsonl",
+      "byte_offset": 792,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-reject-002",
+    "seq": 7,
+    "event_id": "a2#edit:0",
+    "parent_id": "a2",
+    "timestamp": "2026-06-22T15:11:40Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "1b2d44aa",
+        "branch": "feature/cache"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_cache_1",
+      "diff": {
+        "path": "src/cache.rs",
+        "old": "// TODO: choose cache backend",
+        "new": "let client = redis::Client::open(cfg.redis_url)?;",
+        "unified": "@@ -1,1 +1,1 @@\n-// TODO: choose cache backend\n+let client = redis::Client::open(cfg.redis_url)?;",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/rejected_alternative.jsonl",
+      "byte_offset": 792,
+      "line_no": 5
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__windsurf__v1__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/events__windsurf__v1__tool_failure.snap
new file mode 100644
index 0000000..3769d23
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__windsurf__v1__tool_failure.snap
@@ -0,0 +1,192 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-fail-004",
+    "seq": 0,
+    "event_id": "0138e418c43d8825cf1c3a3e9cd92031b0f6ecd274c355095f4c69483ae3cfe6",
+    "parent_id": null,
+    "timestamp": "1970-01-01T00:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "deadbeef",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/acme-api",
+      "git": {
+        "sha": "deadbeef",
+        "branch": "main"
+      },
+      "model": "cascade-base",
+      "tool_version": "1.8.2"
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/tool_failure.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-fail-004",
+    "seq": 1,
+    "event_id": "u1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T17:05:00Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "deadbeef",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Update the read-only vendored file to bump the version.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/tool_failure.jsonl",
+      "byte_offset": 165,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-fail-004",
+    "seq": 2,
+    "event_id": "a1",
+    "parent_id": "u1",
+    "timestamp": "2026-06-22T17:05:07Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "deadbeef",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Attempting the edit now.",
+      "thinking": null,
+      "model": "cascade-base",
+      "usage": {
+        "input_tokens": 260,
+        "output_tokens": 90,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/tool_failure.jsonl",
+      "byte_offset": 326,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-fail-004",
+    "seq": 3,
+    "event_id": "a1#toolcall:call_fail_1",
+    "parent_id": "a1",
+    "timestamp": "2026-06-22T17:05:07Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "deadbeef",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call_fail_1",
+      "name": "edit_file",
+      "args": {
+        "path": "vendor/locked.toml"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/tool_failure.jsonl",
+      "byte_offset": 326,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-fail-004",
+    "seq": 4,
+    "event_id": "a1#toolresult:call_fail_1",
+    "parent_id": "a1",
+    "timestamp": "2026-06-22T17:05:07Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "deadbeef",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call_fail_1",
+      "ok": false,
+      "output": "EACCES: permission denied, open 'vendor/locked.toml'"
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/tool_failure.jsonl",
+      "byte_offset": 326,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "windsurf",
+    "session_id": "ws-fail-004",
+    "seq": 5,
+    "event_id": "a1#edit:0",
+    "parent_id": "a1",
+    "timestamp": "2026-06-22T17:05:07Z",
+    "project": {
+      "cwd": "/Users/dev/acme-api",
+      "repo_root": null,
+      "git": {
+        "sha": "deadbeef",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call_fail_1",
+      "diff": {
+        "path": "vendor/locked.toml",
+        "old": "version = \"1.0.0\"",
+        "new": "version = \"1.0.1\"",
+        "unified": "@@ -1,1 +1,1 @@\n-version = \"1.0.0\"\n+version = \"1.0.1\"",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/windsurf/v1/tool_failure.jsonl",
+      "byte_offset": 326,
+      "line_no": 3
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__zed__v1__ban.snap b/crates/memscribe-testkit/tests/snapshots/events__zed__v1__ban.snap
new file mode 100644
index 0000000..55c1084
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__zed__v1__ban.snap
@@ -0,0 +1,188 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-003",
+    "seq": 0,
+    "event_id": "859026e7d1e45de90d1a7021888d56f346375b6a1dacc7ebb9c743b70ccbd2db",
+    "parent_id": null,
+    "timestamp": "2026-06-22T12:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "0011223344556677",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/projects/orbit",
+      "git": {
+        "sha": "0011223344556677",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "zed 0.182.0"
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/ban.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-003",
+    "seq": 1,
+    "event_id": "msg-1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T12:00:03Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "0011223344556677",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "We will never add a dependency on left-pad. Use the std library instead.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/ban.jsonl",
+      "byte_offset": 193,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-003",
+    "seq": 2,
+    "event_id": "msg-2",
+    "parent_id": "msg-1",
+    "timestamp": "2026-06-22T12:00:07Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "0011223344556677",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Understood — no left-pad dependency. I'll pad with format! from std.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 640,
+        "output_tokens": 120,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/ban.jsonl",
+      "byte_offset": 377,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-003",
+    "seq": 3,
+    "event_id": "msg-2:call:call-1",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T12:00:07Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "0011223344556677",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-1",
+      "name": "read_file",
+      "args": {
+        "path": "Cargo.toml"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/ban.jsonl",
+      "byte_offset": 377,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-003",
+    "seq": 4,
+    "event_id": "msg-3",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T12:00:10Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "0011223344556677",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Confirmed Cargo.toml has no left-pad.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/ban.jsonl",
+      "byte_offset": 706,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-003",
+    "seq": 5,
+    "event_id": "msg-3:result:call-1",
+    "parent_id": "msg-3",
+    "timestamp": "2026-06-22T12:00:10Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "0011223344556677",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-1",
+      "ok": true,
+      "output": "[dependencies]\nserde = \"1\"\n"
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/ban.jsonl",
+      "byte_offset": 706,
+      "line_no": 4
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__zed__v1__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/events__zed__v1__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..6132a58
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__zed__v1__happy_path_decision_then_edits.snap
@@ -0,0 +1,221 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-001",
+    "seq": 0,
+    "event_id": "506f78ec77e175d9a1b6bf12be8b5b84d60360505a8bd4b29545744f2f8ddebe",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/projects/orbit",
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "zed 0.182.0"
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-001",
+    "seq": 1,
+    "event_id": "msg-1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:00:05Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Let's use Postgres instead of MySQL for the persistence layer.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 189,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-001",
+    "seq": 2,
+    "event_id": "msg-2",
+    "parent_id": "msg-1",
+    "timestamp": "2026-06-22T10:00:09Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Good call. I'll switch the connection config and the schema module to Postgres.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 1200,
+        "output_tokens": 340,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 363,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-001",
+    "seq": 3,
+    "event_id": "msg-2:edit:0",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T10:00:09Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "src/db/config.rs",
+        "old": "let url = \"mysql://localhost/orbit\";",
+        "new": "let url = \"postgres://localhost/orbit\";",
+        "unified": "@@ -1 +1 @@\n-let url = \"mysql://localhost/orbit\";\n+let url = \"postgres://localhost/orbit\";",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 363,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-001",
+    "seq": 4,
+    "event_id": "msg-2:edit:1",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T10:00:09Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "src/db/schema.rs",
+        "old": "// MySQL dialect\n",
+        "new": "// Postgres dialect\nuse sqlx::postgres::PgPool;\n",
+        "unified": null,
+        "added_lines": 2,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 363,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-001",
+    "seq": 5,
+    "event_id": "msg-3",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T10:01:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Looks right, thanks.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1039,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-001",
+    "seq": 6,
+    "event_id": "c27b0f619cacff966dde4ae35ed48f50113a66178ea42cf13a04f629550fdc49",
+    "parent_id": null,
+    "timestamp": "2026-06-22T10:01:30Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "a1b2c3d4e5f6",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_end",
+      "reason": "user_closed"
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/happy_path_decision_then_edits.jsonl",
+      "byte_offset": 1174,
+      "line_no": 5
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__zed__v1__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/events__zed__v1__rejected_alternative.snap
new file mode 100644
index 0000000..68393bc
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__zed__v1__rejected_alternative.snap
@@ -0,0 +1,191 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-002",
+    "seq": 0,
+    "event_id": "f31c2e32429c38bc8de7177ae9878961569e923d005b8a6d3be82f93532a7342",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "f6e5d4c3b2a1",
+        "branch": "feature/caching"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/projects/orbit",
+      "git": {
+        "sha": "f6e5d4c3b2a1",
+        "branch": "feature/caching"
+      },
+      "model": null,
+      "tool_version": "zed 0.182.0"
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/rejected_alternative.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-002",
+    "seq": 1,
+    "event_id": "msg-1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T11:00:04Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "f6e5d4c3b2a1",
+        "branch": "feature/caching"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Should we add Redis for the cache layer?",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/rejected_alternative.jsonl",
+      "byte_offset": 200,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-002",
+    "seq": 2,
+    "event_id": "msg-2",
+    "parent_id": "msg-1",
+    "timestamp": "2026-06-22T11:00:08Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "f6e5d4c3b2a1",
+        "branch": "feature/caching"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "We considered Redis but rejected it: the working set fits in process memory, so an in-process LRU is simpler. Not adding Redis.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": {
+        "input_tokens": 900,
+        "output_tokens": 210,
+        "cache_read_tokens": null,
+        "cache_creation_tokens": null
+      },
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/rejected_alternative.jsonl",
+      "byte_offset": 352,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-002",
+    "seq": 3,
+    "event_id": "msg-3",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T11:00:40Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "f6e5d4c3b2a1",
+        "branch": "feature/caching"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Agreed, go with the in-process LRU.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/rejected_alternative.jsonl",
+      "byte_offset": 660,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-002",
+    "seq": 4,
+    "event_id": "msg-4",
+    "parent_id": "msg-3",
+    "timestamp": "2026-06-22T11:00:46Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "f6e5d4c3b2a1",
+        "branch": "feature/caching"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Done — added a bounded LRU.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": null,
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/rejected_alternative.jsonl",
+      "byte_offset": 810,
+      "line_no": 5
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-002",
+    "seq": 5,
+    "event_id": "msg-4:edit:0",
+    "parent_id": "msg-4",
+    "timestamp": "2026-06-22T11:00:46Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "f6e5d4c3b2a1",
+        "branch": "feature/caching"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": null,
+      "diff": {
+        "path": "src/cache/lru.rs",
+        "old": "",
+        "new": "pub struct Lru { cap: usize }\n",
+        "unified": "@@ -0,0 +1 @@\n+pub struct Lru { cap: usize }",
+        "added_lines": 1,
+        "removed_lines": 0
+      }
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/rejected_alternative.jsonl",
+      "byte_offset": 810,
+      "line_no": 5
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/events__zed__v1__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/events__zed__v1__tool_failure.snap
new file mode 100644
index 0000000..cee9331
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/events__zed__v1__tool_failure.snap
@@ -0,0 +1,217 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: events
+---
+[
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-004",
+    "seq": 0,
+    "event_id": "f01734af6017bb88b0a67c86dfeb4f8cec1d3ac3efc38443783ef097412600f5",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:00:00Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "99aa88bb77cc",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "session_start",
+      "cwd": "/Users/dev/projects/orbit",
+      "git": {
+        "sha": "99aa88bb77cc",
+        "branch": "main"
+      },
+      "model": null,
+      "tool_version": "zed 0.182.0"
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/tool_failure.jsonl",
+      "byte_offset": 0,
+      "line_no": 1
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-004",
+    "seq": 1,
+    "event_id": "msg-1",
+    "parent_id": null,
+    "timestamp": "2026-06-22T13:00:02Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "99aa88bb77cc",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "user_turn",
+      "text": "Rename the config field timeout_ms to timeout_millis everywhere.",
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/tool_failure.jsonl",
+      "byte_offset": 189,
+      "line_no": 2
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-004",
+    "seq": 2,
+    "event_id": "msg-2",
+    "parent_id": "msg-1",
+    "timestamp": "2026-06-22T13:00:06Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "99aa88bb77cc",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "Applying the rename.",
+      "thinking": null,
+      "model": "claude-opus-4-8",
+      "usage": null,
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/tool_failure.jsonl",
+      "byte_offset": 365,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-004",
+    "seq": 3,
+    "event_id": "msg-2:call:call-9",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T13:00:06Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "99aa88bb77cc",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_call",
+      "call_id": "call-9",
+      "name": "edit_file",
+      "args": {
+        "path": "src/config.rs"
+      }
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/tool_failure.jsonl",
+      "byte_offset": 365,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-004",
+    "seq": 4,
+    "event_id": "msg-2:edit:0",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T13:00:06Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "99aa88bb77cc",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "file_edit",
+      "call_id": "call-9",
+      "diff": {
+        "path": "src/config.rs",
+        "old": "timeout_ms: u64,",
+        "new": "timeout_millis: u64,",
+        "unified": "@@ -1 +1 @@\n-timeout_ms: u64,\n+timeout_millis: u64,",
+        "added_lines": 1,
+        "removed_lines": 1
+      }
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/tool_failure.jsonl",
+      "byte_offset": 365,
+      "line_no": 3
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-004",
+    "seq": 5,
+    "event_id": "msg-3",
+    "parent_id": "msg-2",
+    "timestamp": "2026-06-22T13:00:08Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "99aa88bb77cc",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "assistant_turn",
+      "text": "The edit failed — the file was locked by another process.",
+      "thinking": null,
+      "model": null,
+      "usage": null,
+      "parts": []
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/tool_failure.jsonl",
+      "byte_offset": 794,
+      "line_no": 4
+    }
+  },
+  {
+    "schema_version": 1,
+    "source": "zed",
+    "session_id": "zed-thread-004",
+    "seq": 6,
+    "event_id": "msg-3:result:call-9",
+    "parent_id": "msg-3",
+    "timestamp": "2026-06-22T13:00:08Z",
+    "project": {
+      "cwd": "/Users/dev/projects/orbit",
+      "repo_root": null,
+      "git": {
+        "sha": "99aa88bb77cc",
+        "branch": "main"
+      }
+    },
+    "kind": {
+      "event": "tool_result",
+      "call_id": "call-9",
+      "ok": false,
+      "output": "error: could not write src/config.rs: resource temporarily unavailable"
+    },
+    "provenance": {
+      "file": "fixtures/zed/v1/tool_failure.jsonl",
+      "byte_offset": 794,
+      "line_no": 4
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__ban.snap b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__ban.snap
new file mode 100644
index 0000000..053a583
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__ban.snap
@@ -0,0 +1,61 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "sess-ban-003",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "We will never add a dependency on left-pad in this repo. Use the std string padding instead.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "Use",
+        "offset": 57
+      },
+      {
+        "rule_id": "ban.negated_use",
+        "category": "ban",
+        "matched_text": "never add",
+        "offset": 8
+      },
+      {
+        "rule_id": "imperative.must_always_never",
+        "category": "imperative",
+        "matched_text": "never",
+        "offset": 8
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/claude_code/2.0/ban.jsonl",
+        "byte_offset": 0,
+        "line_no": 1
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Use the std string padding instead.",
+    "considered_options": [
+      {
+        "text": "the",
+        "chosen": true
+      }
+    ],
+    "is_ban": true,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..7730027
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__happy_path_decision_then_edits.snap
@@ -0,0 +1,135 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "sess-happy-001",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use Postgres instead of MySQL for the orders service.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 19
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl",
+        "byte_offset": 0,
+        "line_no": 1
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use Postgres instead of MySQL for the orders service.",
+    "considered_options": [
+      {
+        "text": "MySQL",
+        "chosen": false
+      },
+      {
+        "text": "Postgres",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "/Users/dev/acme/api/src/db/config.rs",
+    "diff": {
+      "path": "/Users/dev/acme/api/src/db/config.rs",
+      "old": "mysql://localhost:3306/orders",
+      "new": "postgres://localhost:5432/orders",
+      "unified": "-    const DSN: &str = \"mysql://localhost:3306/orders\";\n+    const DSN: &str = \"postgres://localhost:5432/orders\";",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+      "branch": "main"
+    },
+    "episode_id": "831546951728ec6f3d421d1fbc91754a17375210fb27985c8222a2d36c4a7731"
+  },
+  {
+    "node": "episode",
+    "path": "/Users/dev/acme/api/src/db/schema.rs",
+    "diff": {
+      "path": "/Users/dev/acme/api/src/db/schema.rs",
+      "old": "use mysql::Pool;",
+      "new": "use postgres::Client;",
+      "unified": "-use mysql::Pool;\n+use postgres::Client;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0",
+      "branch": "main"
+    },
+    "episode_id": "8c74bf0bb292f70631373f4e1486d088586b3606df77adbf13265d936103c556"
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-happy-001:1",
+    "to": "episode:831546951728ec6f3d421d1fbc91754a17375210fb27985c8222a2d36c4a7731",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-happy-001",
+      "used_decision": "decision:sess-happy-001:1",
+      "was_generated_by_session": "sess-happy-001",
+      "t_use": "2026-06-22T10:00:00Z",
+      "t_gen": "2026-06-22T10:00:03.48Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-happy-001:1",
+    "to": "episode:8c74bf0bb292f70631373f4e1486d088586b3606df77adbf13265d936103c556",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-happy-001",
+      "used_decision": "decision:sess-happy-001:1",
+      "was_generated_by_session": "sess-happy-001",
+      "t_use": "2026-06-22T10:00:00Z",
+      "t_gen": "2026-06-22T10:00:07.21Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__interleaved_arcs.snap b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__interleaved_arcs.snap
new file mode 100644
index 0000000..978de38
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__interleaved_arcs.snap
@@ -0,0 +1,200 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "sess-interleaved-005",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use argon2 for password hashing in the auth module.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+        "byte_offset": 0,
+        "line_no": 1
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use argon2 for password hashing in the auth module.",
+    "considered_options": [
+      {
+        "text": "argon2",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "/Users/dev/acme/api/src/auth.rs",
+    "diff": {
+      "path": "/Users/dev/acme/api/src/auth.rs",
+      "old": "use bcrypt::hash;",
+      "new": "use argon2::hash;",
+      "unified": "-use bcrypt::hash;\n+use argon2::hash;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "1111111122222222333333334444444455555555",
+      "branch": "main"
+    },
+    "episode_id": "716b9f34e431f29a01f70e58c9b3688c7e78ff70d4949783e0aebef91608bded"
+  },
+  {
+    "node": "conversation",
+    "session_id": "sess-interleaved-005",
+    "turn_range": {
+      "start": 7,
+      "end": 8
+    },
+    "text": "Actually, let's switch to scrypt for the auth module and tune the cost in config.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "switch to",
+        "offset": 16
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/claude_code/2.0/interleaved_arcs.jsonl",
+        "byte_offset": 1853,
+        "line_no": 4
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Actually, let's switch to scrypt for the auth module and tune the cost in config.",
+    "considered_options": [
+      {
+        "text": "scrypt",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 7,
+      "end": 8
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "/Users/dev/acme/api/src/auth.rs",
+    "diff": {
+      "path": "/Users/dev/acme/api/src/auth.rs",
+      "old": "use argon2::hash;",
+      "new": "use scrypt::hash;",
+      "unified": "-use argon2::hash;\n+use scrypt::hash;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "1111111122222222333333334444444455555555",
+      "branch": "main"
+    },
+    "episode_id": "fae4cb4a3422578a245f8a410186e34879e2a06c7f28bdba57542da3380b6737"
+  },
+  {
+    "node": "episode",
+    "path": "/Users/dev/acme/api/src/config.rs",
+    "diff": {
+      "path": "/Users/dev/acme/api/src/config.rs",
+      "old": "const HASH_COST: u32 = 10;",
+      "new": "const HASH_COST: u32 = 15;",
+      "unified": "-const HASH_COST: u32 = 10;\n+const HASH_COST: u32 = 15;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "1111111122222222333333334444444455555555",
+      "branch": "main"
+    },
+    "episode_id": "32ccc9e296218969f0beed49e54cc8b851b9bbe421ba51d4cf319a6c95004638"
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-interleaved-005:1",
+    "to": "episode:716b9f34e431f29a01f70e58c9b3688c7e78ff70d4949783e0aebef91608bded",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-interleaved-005",
+      "used_decision": "decision:sess-interleaved-005:1",
+      "was_generated_by_session": "sess-interleaved-005",
+      "t_use": "2026-06-22T14:00:00Z",
+      "t_gen": "2026-06-22T14:00:03.5Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-interleaved-005:7",
+    "to": "episode:32ccc9e296218969f0beed49e54cc8b851b9bbe421ba51d4cf319a6c95004638",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-interleaved-005",
+      "used_decision": "decision:sess-interleaved-005:7",
+      "was_generated_by_session": "sess-interleaved-005",
+      "t_use": "2026-06-22T14:01:00Z",
+      "t_gen": "2026-06-22T14:01:08.5Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.333333,
+      "confidence": 0.5,
+      "lift": 1.5,
+      "phi": 0.5,
+      "p": 0.386476
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-interleaved-005:7",
+    "to": "episode:fae4cb4a3422578a245f8a410186e34879e2a06c7f28bdba57542da3380b6737",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-interleaved-005",
+      "used_decision": "decision:sess-interleaved-005:7",
+      "was_generated_by_session": "sess-interleaved-005",
+      "t_use": "2026-06-22T14:01:00Z",
+      "t_gen": "2026-06-22T14:01:04.5Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.333333,
+      "confidence": 0.5,
+      "lift": 0.75,
+      "phi": -0.5,
+      "p": 0.386476
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__multi_edit_single_commit.snap b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__multi_edit_single_commit.snap
new file mode 100644
index 0000000..6cfdea0
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__multi_edit_single_commit.snap
@@ -0,0 +1,163 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "sess-multiedit-006",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's migrate to anyhow for error handling across the request, handler, and router modules.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "migrate to",
+        "offset": 6
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/claude_code/2.0/multi_edit_single_commit.jsonl",
+        "byte_offset": 0,
+        "line_no": 1
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's migrate to anyhow for error handling across the request, handler, and router modules.",
+    "considered_options": [
+      {
+        "text": "anyhow",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "/Users/dev/acme/api/src/request.rs",
+    "diff": {
+      "path": "/Users/dev/acme/api/src/request.rs",
+      "old": "use crate::error::AppError;",
+      "new": "use anyhow::Result;",
+      "unified": "-use crate::error::AppError;\n+use anyhow::Result;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+      "branch": "main"
+    },
+    "episode_id": "31bfaed16633a3210f82532b4937a80934d43f836ea4ecab223a42f7a8738687"
+  },
+  {
+    "node": "episode",
+    "path": "/Users/dev/acme/api/src/handler.rs",
+    "diff": {
+      "path": "/Users/dev/acme/api/src/handler.rs",
+      "old": "use crate::error::AppError;",
+      "new": "use anyhow::Result;",
+      "unified": "-use crate::error::AppError;\n+use anyhow::Result;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+      "branch": "main"
+    },
+    "episode_id": "c3ac123fcc0421301cd349c74ef3aa15601618c45e699f9af10353ed2d3441ef"
+  },
+  {
+    "node": "episode",
+    "path": "/Users/dev/acme/api/src/router.rs",
+    "diff": {
+      "path": "/Users/dev/acme/api/src/router.rs",
+      "old": "use crate::error::AppError;",
+      "new": "use anyhow::Result;",
+      "unified": "-use crate::error::AppError;\n+use anyhow::Result;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "aaaa1111bbbb2222cccc3333dddd4444eeee5555",
+      "branch": "main"
+    },
+    "episode_id": "ebd98e0127df7899ea3298aa5e84efd3a23d93d9e774d55006f5a895ebde6458"
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-multiedit-006:1",
+    "to": "episode:31bfaed16633a3210f82532b4937a80934d43f836ea4ecab223a42f7a8738687",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-multiedit-006",
+      "used_decision": "decision:sess-multiedit-006:1",
+      "was_generated_by_session": "sess-multiedit-006",
+      "t_use": "2026-06-22T15:00:00Z",
+      "t_gen": "2026-06-22T15:00:03.4Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.333333,
+      "confidence": 0.333333,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-multiedit-006:1",
+    "to": "episode:c3ac123fcc0421301cd349c74ef3aa15601618c45e699f9af10353ed2d3441ef",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-multiedit-006",
+      "used_decision": "decision:sess-multiedit-006:1",
+      "was_generated_by_session": "sess-multiedit-006",
+      "t_use": "2026-06-22T15:00:00Z",
+      "t_gen": "2026-06-22T15:00:06.4Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.333333,
+      "confidence": 0.333333,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-multiedit-006:1",
+    "to": "episode:ebd98e0127df7899ea3298aa5e84efd3a23d93d9e774d55006f5a895ebde6458",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-multiedit-006",
+      "used_decision": "decision:sess-multiedit-006:1",
+      "was_generated_by_session": "sess-multiedit-006",
+      "t_use": "2026-06-22T15:00:00Z",
+      "t_gen": "2026-06-22T15:00:09.4Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.333333,
+      "confidence": 0.333333,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__no_commitment_marker.snap b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__no_commitment_marker.snap
new file mode 100644
index 0000000..98453d6
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__no_commitment_marker.snap
@@ -0,0 +1,23 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "episode",
+    "path": "/Users/dev/acme/api/src/handler.rs",
+    "diff": {
+      "path": "/Users/dev/acme/api/src/handler.rs",
+      "old": "    let body = read(); ",
+      "new": "    let body = read();",
+      "unified": "-    let body = read(); \n+    let body = read();",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "0a0b0c0d0e0f10111213141516171819202122ab",
+      "branch": "main"
+    },
+    "episode_id": "0137490964ce998a3a343bba5dc9af20656aef0d0fdd9cbe0626583f27cec262"
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__rejected_alternative.snap
new file mode 100644
index 0000000..20445d9
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__rejected_alternative.snap
@@ -0,0 +1,91 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "sess-reject-002",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "We'll go with Stripe instead of PayPal for the checkout flow.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "go with",
+        "offset": 6
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 21
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/claude_code/2.0/rejected_alternative.jsonl",
+        "byte_offset": 0,
+        "line_no": 1
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "We'll go with Stripe instead of PayPal for the checkout flow.",
+    "considered_options": [
+      {
+        "text": "PayPal",
+        "chosen": false
+      },
+      {
+        "text": "Stripe",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "/Users/dev/acme/web/src/checkout/provider.ts",
+    "diff": {
+      "path": "/Users/dev/acme/web/src/checkout/provider.ts",
+      "old": "",
+      "new": "export { stripe as paymentProvider } from './stripe';\n",
+      "unified": "+export { stripe as paymentProvider } from './stripe';",
+      "added_lines": 1,
+      "removed_lines": 0
+    },
+    "git": {
+      "sha": "0f1e2d3c4b5a69788796a5b4c3d2e1f00f1e2d3c",
+      "branch": "feat/payments"
+    },
+    "episode_id": "39d384a667fa2764297c714568536ab737e3e713025ea94551c3db5de2d182c6"
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-reject-002:1",
+    "to": "episode:39d384a667fa2764297c714568536ab737e3e713025ea94551c3db5de2d182c6",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-reject-002",
+      "used_decision": "decision:sess-reject-002:1",
+      "was_generated_by_session": "sess-reject-002",
+      "t_use": "2026-06-22T11:30:00Z",
+      "t_gen": "2026-06-22T11:30:04.9Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__rewind_compaction.snap b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__rewind_compaction.snap
new file mode 100644
index 0000000..133e749
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__rewind_compaction.snap
@@ -0,0 +1,156 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "sess-rewind-007",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use rustls for the TLS layer in the gateway.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+        "byte_offset": 0,
+        "line_no": 1
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use rustls for the TLS layer in the gateway.",
+    "considered_options": [
+      {
+        "text": "rustls",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "/Users/dev/acme/api/src/gateway/tls.rs",
+    "diff": {
+      "path": "/Users/dev/acme/api/src/gateway/tls.rs",
+      "old": "use native_tls::TlsConnector;",
+      "new": "use rustls::ClientConfig;",
+      "unified": "-use native_tls::TlsConnector;\n+use rustls::ClientConfig;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "9999888877776666555544443333222211110000",
+      "branch": "main"
+    },
+    "episode_id": "6e05760ec5f154e886f626f291cc328dbc8044040116dd04416710e0edd2794d"
+  },
+  {
+    "node": "conversation",
+    "session_id": "sess-rewind-007",
+    "turn_range": {
+      "start": 8,
+      "end": 9
+    },
+    "text": "On reflection, let's switch to native-tls after all for the gateway to match the rest of the fleet.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "switch to",
+        "offset": 21
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/claude_code/2.0/rewind_compaction.jsonl",
+        "byte_offset": 2128,
+        "line_no": 5
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "On reflection, let's switch to native-tls after all for the gateway to match the rest of the fleet.",
+    "considered_options": [
+      {
+        "text": "native-tls",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 8,
+      "end": 9
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "/Users/dev/acme/api/src/gateway/tls.rs",
+    "diff": {
+      "path": "/Users/dev/acme/api/src/gateway/tls.rs",
+      "old": "use rustls::ClientConfig;",
+      "new": "use native_tls::TlsConnector;",
+      "unified": "-use rustls::ClientConfig;\n+use native_tls::TlsConnector;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "9999888877776666555544443333222211110000",
+      "branch": "main"
+    },
+    "episode_id": "7087e5c0eb53ed1fee4c0b6c0a883d0fd4bb7e77d5a89d7d2dc381c52f090af9"
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-rewind-007:1",
+    "to": "episode:6e05760ec5f154e886f626f291cc328dbc8044040116dd04416710e0edd2794d",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-rewind-007",
+      "used_decision": "decision:sess-rewind-007:1",
+      "was_generated_by_session": "sess-rewind-007",
+      "t_use": "2026-06-22T16:00:00Z",
+      "t_gen": "2026-06-22T16:00:03.4Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-rewind-007:8",
+    "to": "episode:7087e5c0eb53ed1fee4c0b6c0a883d0fd4bb7e77d5a89d7d2dc381c52f090af9",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-rewind-007",
+      "used_decision": "decision:sess-rewind-007:8",
+      "was_generated_by_session": "sess-rewind-007",
+      "t_use": "2026-06-22T16:05:00Z",
+      "t_gen": "2026-06-22T16:05:04.4Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__subagent_thread.snap b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__subagent_thread.snap
new file mode 100644
index 0000000..5079514
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__subagent_thread.snap
@@ -0,0 +1,124 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "sess-main-008",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use a dedicated search subagent to migrate the indexer to tantivy.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/claude_code/2.0/subagent_thread.jsonl",
+        "byte_offset": 0,
+        "line_no": 1
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use a dedicated search subagent to migrate the indexer to tantivy.",
+    "considered_options": [
+      {
+        "text": "a",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "conversation",
+    "session_id": "sess-main-008",
+    "turn_range": {
+      "start": 4,
+      "end": 5
+    },
+    "text": "Let's adopt tantivy as the search index backend in the indexer module.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "adopt",
+        "offset": 6
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/claude_code/2.0/subagent_thread.jsonl",
+        "byte_offset": 1075,
+        "line_no": 3
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's adopt tantivy as the search index backend in the indexer module.",
+    "considered_options": [
+      {
+        "text": "tantivy",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 4,
+      "end": 5
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "/Users/dev/acme/api/src/indexer.rs",
+    "diff": {
+      "path": "/Users/dev/acme/api/src/indexer.rs",
+      "old": "use tantivy_lite::Index;",
+      "new": "use tantivy::Index;",
+      "unified": "-use tantivy_lite::Index;\n+use tantivy::Index;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "abcdefab12345678abcdefab12345678abcdefab",
+      "branch": "main"
+    },
+    "episode_id": "ecde881f1e7a645f74ddfe66f1cfbe262a0cf5cc157d85f65410ea79c370900e"
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-main-008:4",
+    "to": "episode:ecde881f1e7a645f74ddfe66f1cfbe262a0cf5cc157d85f65410ea79c370900e",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-main-008",
+      "used_decision": "decision:sess-main-008:4",
+      "was_generated_by_session": "sess-main-008",
+      "t_use": "2026-06-22T17:00:05Z",
+      "t_gen": "2026-06-22T17:00:08.4Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__tool_failure.snap
new file mode 100644
index 0000000..a825381
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_0__tool_failure.snap
@@ -0,0 +1,49 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "sess-fail-004",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use a connection pool in the orders repository.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/claude_code/2.0/tool_failure.jsonl",
+        "byte_offset": 0,
+        "line_no": 1
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use a connection pool in the orders repository.",
+    "considered_options": [
+      {
+        "text": "a",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_1__version_tolerance.snap b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_1__version_tolerance.snap
new file mode 100644
index 0000000..3b672a6
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__claude_code__2_1__version_tolerance.snap
@@ -0,0 +1,59 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "sess-v21-001",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use Postgres instead of MySQL for the orders service.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 19
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/claude_code/2.1/version_tolerance.jsonl",
+        "byte_offset": 0,
+        "line_no": 1
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use Postgres instead of MySQL for the orders service.",
+    "considered_options": [
+      {
+        "text": "MySQL",
+        "chosen": false
+      },
+      {
+        "text": "Postgres",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__codex__v1__version_tolerance.snap b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v1__version_tolerance.snap
new file mode 100644
index 0000000..e9c6f75
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v1__version_tolerance.snap
@@ -0,0 +1,5 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__ban.snap b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__ban.snap
new file mode 100644
index 0000000..1dbebc6
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__ban.snap
@@ -0,0 +1,50 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "01JZ8codexsess0003ban",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Policy: we will never add a dependency on left-pad. Do not introduce it under any circumstances, even transitively if you can avoid it.",
+    "markers": [
+      {
+        "rule_id": "ban.negated_use",
+        "category": "ban",
+        "matched_text": "never add",
+        "offset": 16
+      },
+      {
+        "rule_id": "imperative.must_always_never",
+        "category": "imperative",
+        "matched_text": "never",
+        "offset": 16
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/codex/v2/ban.jsonl",
+        "byte_offset": 255,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Policy: we will never add a dependency on left-pad.",
+    "considered_options": [],
+    "is_ban": true,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..ae41108
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__happy_path_decision_then_edits.snap
@@ -0,0 +1,135 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "01JZ8codexsess0001happypath",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use Postgres instead of MySQL for the persistence layer. Swap the driver and the connection URL.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 19
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/codex/v2/happy_path_decision_then_edits.jsonl",
+        "byte_offset": 275,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use Postgres instead of MySQL for the persistence layer.",
+    "considered_options": [
+      {
+        "text": "MySQL",
+        "chosen": false
+      },
+      {
+        "text": "Postgres",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "Cargo.toml",
+    "diff": {
+      "path": "Cargo.toml",
+      "old": null,
+      "new": null,
+      "unified": "@@\n-mysql = \"24\"\n+sqlx = { version = \"0.7\", features = [\"postgres\", \"runtime-tokio\"] }",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+      "branch": "feat/storage-layer"
+    },
+    "episode_id": "117c5fd3325566f8c4f87cea69685e577e5404c958202234ed3dd286dc2d7012"
+  },
+  {
+    "node": "episode",
+    "path": "src/db.rs",
+    "diff": {
+      "path": "src/db.rs",
+      "old": null,
+      "new": null,
+      "unified": "@@\n-let url = \"mysql://localhost:3306/acme\";\n-let pool = mysql::Pool::new(url)?;\n+let url = \"postgres://localhost:5432/acme\";\n+let pool = sqlx::PgPool::connect(url).await?;",
+      "added_lines": 2,
+      "removed_lines": 2
+    },
+    "git": {
+      "sha": "a1b2c3d4e5f60718293a4b5c6d7e8f9001020304",
+      "branch": "feat/storage-layer"
+    },
+    "episode_id": "0992089b822a803ecddbe1a0c4bd5e88206b9295597038b97162add8a3e653c5"
+  },
+  {
+    "node": "binding",
+    "from": "decision:01JZ8codexsess0001happypath:1",
+    "to": "episode:0992089b822a803ecddbe1a0c4bd5e88206b9295597038b97162add8a3e653c5",
+    "relation": "produced",
+    "prov": {
+      "used_session": "01JZ8codexsess0001happypath",
+      "used_decision": "decision:01JZ8codexsess0001happypath:1",
+      "was_generated_by_session": "01JZ8codexsess0001happypath",
+      "t_use": "2026-06-22T14:00:05.12Z",
+      "t_gen": "2026-06-22T14:00:18.64Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:01JZ8codexsess0001happypath:1",
+    "to": "episode:117c5fd3325566f8c4f87cea69685e577e5404c958202234ed3dd286dc2d7012",
+    "relation": "produced",
+    "prov": {
+      "used_session": "01JZ8codexsess0001happypath",
+      "used_decision": "decision:01JZ8codexsess0001happypath:1",
+      "was_generated_by_session": "01JZ8codexsess0001happypath",
+      "t_use": "2026-06-22T14:00:05.12Z",
+      "t_gen": "2026-06-22T14:00:18.64Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__interleaved_arcs.snap b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__interleaved_arcs.snap
new file mode 100644
index 0000000..ccc33f5
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__interleaved_arcs.snap
@@ -0,0 +1,244 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use argon2 for password hashing. Update the auth module and the shared crypto helper accordingly.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+        "byte_offset": 277,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use argon2 for password hashing.",
+    "considered_options": [
+      {
+        "text": "argon2",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/auth.rs",
+    "diff": {
+      "path": "src/auth.rs",
+      "old": null,
+      "new": null,
+      "unified": "@@\n-use bcrypt::hash;\n+use argon2::Argon2;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+      "branch": "feat/auth-overhaul"
+    },
+    "episode_id": "a4e4c0d70baa57d560d4fba5bd2236984971124d4a6a514239407c2872643ba7"
+  },
+  {
+    "node": "episode",
+    "path": "src/shared/crypto.rs",
+    "diff": {
+      "path": "src/shared/crypto.rs",
+      "old": null,
+      "new": null,
+      "unified": "@@\n-pub const HASH_COST: u32 = 12;\n+pub const ARGON2_MEM_KIB: u32 = 19456;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+      "branch": "feat/auth-overhaul"
+    },
+    "episode_id": "a42ee0d09ada744881547a8143cc30337a86f5dc1c5ee34de287a237dda53ed3"
+  },
+  {
+    "node": "conversation",
+    "session_id": "01JZ8codexsess0005interleaved",
+    "turn_range": {
+      "start": 8,
+      "end": 9
+    },
+    "text": "Now let's switch to a 30-minute session TTL. Update the session manager and bump the shared default in the crypto helper's config block.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "switch to",
+        "offset": 10
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/codex/v2/interleaved_arcs.jsonl",
+        "byte_offset": 1671,
+        "line_no": 7
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Now let's switch to a 30-minute session TTL.",
+    "considered_options": [
+      {
+        "text": "a",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 8,
+      "end": 9
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/session.rs",
+    "diff": {
+      "path": "src/session.rs",
+      "old": null,
+      "new": null,
+      "unified": "@@\n-const SESSION_TTL_SECS: u64 = 3600;\n+const SESSION_TTL_SECS: u64 = 1800;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+      "branch": "feat/auth-overhaul"
+    },
+    "episode_id": "e7f5bf0a2286c1a6c3a9ef53a97af2941ea1698b7f1cabbf1f4c5c24e94d5bd2"
+  },
+  {
+    "node": "episode",
+    "path": "src/shared/crypto.rs",
+    "diff": {
+      "path": "src/shared/crypto.rs",
+      "old": null,
+      "new": null,
+      "unified": "@@\n-pub const DEFAULT_TTL_SECS: u64 = 3600;\n+pub const DEFAULT_TTL_SECS: u64 = 1800;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "e5f60718293a4b5c6d7e8f900102030405060708",
+      "branch": "feat/auth-overhaul"
+    },
+    "episode_id": "8950849692cc4d27e72a42f4e2a82284766365a120a24be4adc958c6fc8a8cf6"
+  },
+  {
+    "node": "binding",
+    "from": "decision:01JZ8codexsess0005interleaved:1",
+    "to": "episode:a42ee0d09ada744881547a8143cc30337a86f5dc1c5ee34de287a237dda53ed3",
+    "relation": "produced",
+    "prov": {
+      "used_session": "01JZ8codexsess0005interleaved",
+      "used_decision": "decision:01JZ8codexsess0005interleaved:1",
+      "was_generated_by_session": "01JZ8codexsess0005interleaved",
+      "t_use": "2026-06-22T18:00:06.1Z",
+      "t_gen": "2026-06-22T18:00:16.45Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.25,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:01JZ8codexsess0005interleaved:1",
+    "to": "episode:a4e4c0d70baa57d560d4fba5bd2236984971124d4a6a514239407c2872643ba7",
+    "relation": "produced",
+    "prov": {
+      "used_session": "01JZ8codexsess0005interleaved",
+      "used_decision": "decision:01JZ8codexsess0005interleaved:1",
+      "was_generated_by_session": "01JZ8codexsess0005interleaved",
+      "t_use": "2026-06-22T18:00:06.1Z",
+      "t_gen": "2026-06-22T18:00:16.45Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.25,
+      "confidence": 0.5,
+      "lift": 2.0,
+      "phi": 0.57735,
+      "p": 0.248213
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:01JZ8codexsess0005interleaved:8",
+    "to": "episode:8950849692cc4d27e72a42f4e2a82284766365a120a24be4adc958c6fc8a8cf6",
+    "relation": "produced",
+    "prov": {
+      "used_session": "01JZ8codexsess0005interleaved",
+      "used_decision": "decision:01JZ8codexsess0005interleaved:8",
+      "was_generated_by_session": "01JZ8codexsess0005interleaved",
+      "t_use": "2026-06-22T18:00:40.2Z",
+      "t_gen": "2026-06-22T18:00:52.8Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.25,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:01JZ8codexsess0005interleaved:8",
+    "to": "episode:e7f5bf0a2286c1a6c3a9ef53a97af2941ea1698b7f1cabbf1f4c5c24e94d5bd2",
+    "relation": "produced",
+    "prov": {
+      "used_session": "01JZ8codexsess0005interleaved",
+      "used_decision": "decision:01JZ8codexsess0005interleaved:8",
+      "was_generated_by_session": "01JZ8codexsess0005interleaved",
+      "t_use": "2026-06-22T18:00:40.2Z",
+      "t_gen": "2026-06-22T18:00:52.8Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.25,
+      "confidence": 0.5,
+      "lift": 2.0,
+      "phi": 0.57735,
+      "p": 0.248213
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__multi_edit_single_commit.snap b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__multi_edit_single_commit.snap
new file mode 100644
index 0000000..1615f20
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__multi_edit_single_commit.snap
@@ -0,0 +1,163 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "01JZ8codexsess0006multiedit",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's rename the HttpClient type to ApiClient everywhere. Use ApiClient as the canonical name across the codebase.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "Use",
+        "offset": 58
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/codex/v2/multi_edit_single_commit.jsonl",
+        "byte_offset": 275,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Use ApiClient as the canonical name across the codebase.",
+    "considered_options": [
+      {
+        "text": "ApiClient",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/client.rs",
+    "diff": {
+      "path": "src/client.rs",
+      "old": null,
+      "new": null,
+      "unified": "@@\n-pub struct HttpClient {\n+pub struct ApiClient {",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "f60718293a4b5c6d7e8f90010203040506070809",
+      "branch": "feat/rename-client"
+    },
+    "episode_id": "b17710b8b2e82b2426e04b5ce18e0b014dcabfdb87c527e93f90b4a0f6d2d99b"
+  },
+  {
+    "node": "episode",
+    "path": "src/handlers.rs",
+    "diff": {
+      "path": "src/handlers.rs",
+      "old": null,
+      "new": null,
+      "unified": "@@\n-use crate::client::HttpClient;\n+use crate::client::ApiClient;\n@@\n-    client: HttpClient,\n+    client: ApiClient,",
+      "added_lines": 2,
+      "removed_lines": 2
+    },
+    "git": {
+      "sha": "f60718293a4b5c6d7e8f90010203040506070809",
+      "branch": "feat/rename-client"
+    },
+    "episode_id": "9d2ded5004f7b1005b1696b583e17ee11274813e4ca8fc3a03283bfc211993d1"
+  },
+  {
+    "node": "episode",
+    "path": "src/main.rs",
+    "diff": {
+      "path": "src/main.rs",
+      "old": null,
+      "new": null,
+      "unified": "@@\n-    let client = HttpClient::new();\n+    let client = ApiClient::new();",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "f60718293a4b5c6d7e8f90010203040506070809",
+      "branch": "feat/rename-client"
+    },
+    "episode_id": "a19e913dac9286bae373c8fdf23c513f5ca018d8fb50c60d56e5b2071a88cf83"
+  },
+  {
+    "node": "binding",
+    "from": "decision:01JZ8codexsess0006multiedit:1",
+    "to": "episode:9d2ded5004f7b1005b1696b583e17ee11274813e4ca8fc3a03283bfc211993d1",
+    "relation": "produced",
+    "prov": {
+      "used_session": "01JZ8codexsess0006multiedit",
+      "used_decision": "decision:01JZ8codexsess0006multiedit:1",
+      "was_generated_by_session": "01JZ8codexsess0006multiedit",
+      "t_use": "2026-06-22T19:00:07.25Z",
+      "t_gen": "2026-06-22T19:00:23.8Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.333333,
+      "confidence": 0.333333,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:01JZ8codexsess0006multiedit:1",
+    "to": "episode:a19e913dac9286bae373c8fdf23c513f5ca018d8fb50c60d56e5b2071a88cf83",
+    "relation": "produced",
+    "prov": {
+      "used_session": "01JZ8codexsess0006multiedit",
+      "used_decision": "decision:01JZ8codexsess0006multiedit:1",
+      "was_generated_by_session": "01JZ8codexsess0006multiedit",
+      "t_use": "2026-06-22T19:00:07.25Z",
+      "t_gen": "2026-06-22T19:00:23.8Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.333333,
+      "confidence": 0.333333,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:01JZ8codexsess0006multiedit:1",
+    "to": "episode:b17710b8b2e82b2426e04b5ce18e0b014dcabfdb87c527e93f90b4a0f6d2d99b",
+    "relation": "produced",
+    "prov": {
+      "used_session": "01JZ8codexsess0006multiedit",
+      "used_decision": "decision:01JZ8codexsess0006multiedit:1",
+      "was_generated_by_session": "01JZ8codexsess0006multiedit",
+      "t_use": "2026-06-22T19:00:07.25Z",
+      "t_gen": "2026-06-22T19:00:23.8Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.333333,
+      "confidence": 0.333333,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__no_commitment_marker.snap b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__no_commitment_marker.snap
new file mode 100644
index 0000000..64d3ae1
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__no_commitment_marker.snap
@@ -0,0 +1,23 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "episode",
+    "path": "src/handlers.rs",
+    "diff": {
+      "path": "src/handlers.rs",
+      "old": null,
+      "new": null,
+      "unified": "@@\n-use std::collections::HashMap;\n use std::collections::BTreeMap;",
+      "added_lines": 0,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "293a4b5c6d7e8f90010203040506070809101112",
+      "branch": "chore/tidy"
+    },
+    "episode_id": "f3dfe5b52f955a01d9e2deb6513eab2cf73c6dc2dad58f9447a4fa30abea9b9a"
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__rejected_alternative.snap
new file mode 100644
index 0000000..e9c6f75
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__rejected_alternative.snap
@@ -0,0 +1,5 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__rewind_compaction.snap b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__rewind_compaction.snap
new file mode 100644
index 0000000..c858d7b
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__rewind_compaction.snap
@@ -0,0 +1,124 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "01JZ8codexsess0007rewindcompact",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use MongoDB for the document store. Set up the driver and the connection config.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/codex/v2/rewind_compaction.jsonl",
+        "byte_offset": 283,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use MongoDB for the document store.",
+    "considered_options": [
+      {
+        "text": "MongoDB",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "conversation",
+    "session_id": "01JZ8codexsess0007rewindcompact",
+    "turn_range": {
+      "start": 5,
+      "end": 6
+    },
+    "text": "Actually let's use Postgres with JSONB for the document store instead. Switch to Postgres and drop the MongoDB plan.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 15
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/codex/v2/rewind_compaction.jsonl",
+        "byte_offset": 1320,
+        "line_no": 6
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Actually let's use Postgres with JSONB for the document store instead.",
+    "considered_options": [
+      {
+        "text": "Postgres",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 5,
+      "end": 6
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/store.rs",
+    "diff": {
+      "path": "src/store.rs",
+      "old": null,
+      "new": null,
+      "unified": "@@\n-use mongodb::Client;\n+use sqlx::PgPool;\n@@\n-    let client = Client::with_uri_str(\"mongodb://localhost:27017\").await?;\n+    let pool = PgPool::connect(\"postgres://localhost:5432/acme\").await?;",
+      "added_lines": 2,
+      "removed_lines": 2
+    },
+    "git": {
+      "sha": "0718293a4b5c6d7e8f9001020304050607080910",
+      "branch": "feat/persistence-pivot"
+    },
+    "episode_id": "4c207b03d5513ed806e41371a1f400264f7f4c230c4a90649a43dd379f6b8f35"
+  },
+  {
+    "node": "binding",
+    "from": "decision:01JZ8codexsess0007rewindcompact:5",
+    "to": "episode:4c207b03d5513ed806e41371a1f400264f7f4c230c4a90649a43dd379f6b8f35",
+    "relation": "produced",
+    "prov": {
+      "used_session": "01JZ8codexsess0007rewindcompact",
+      "used_decision": "decision:01JZ8codexsess0007rewindcompact:5",
+      "was_generated_by_session": "01JZ8codexsess0007rewindcompact",
+      "t_use": "2026-06-22T20:00:40.2Z",
+      "t_gen": "2026-06-22T20:00:53.4Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__subagent_thread.snap b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__subagent_thread.snap
new file mode 100644
index 0000000..ef8ad92
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__subagent_thread.snap
@@ -0,0 +1,125 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "01JZ8codexsess0008subagent",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use a dedicated background worker for the email queue. Extract the send loop into a worker module.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/codex/v2/subagent_thread.jsonl",
+        "byte_offset": 273,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use a dedicated background worker for the email queue.",
+    "considered_options": [
+      {
+        "text": "a",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/worker/email.rs",
+    "diff": {
+      "path": "src/worker/email.rs",
+      "old": null,
+      "new": null,
+      "unified": "+pub async fn run_email_worker() {\n+    loop { drain_email_queue().await; }\n+}",
+      "added_lines": 3,
+      "removed_lines": 0
+    },
+    "git": {
+      "sha": "18293a4b5c6d7e8f900102030405060708091011",
+      "branch": "feat/worker-split"
+    },
+    "episode_id": "20a1158d262d1464d0cc8923564d3099030e4eb76f8b8f79b2657381559f3db8"
+  },
+  {
+    "node": "episode",
+    "path": "src/main.rs",
+    "diff": {
+      "path": "src/main.rs",
+      "old": null,
+      "new": null,
+      "unified": "@@\n-    send_emails_inline();\n+    tokio::spawn(worker::email::run_email_worker());",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "18293a4b5c6d7e8f900102030405060708091011",
+      "branch": "feat/worker-split"
+    },
+    "episode_id": "b662f854f070d854d05850a268f3c9e2bd8b8b2cff5b058027ebffabf2839dd1"
+  },
+  {
+    "node": "binding",
+    "from": "decision:01JZ8codexsess0008subagent:1",
+    "to": "episode:20a1158d262d1464d0cc8923564d3099030e4eb76f8b8f79b2657381559f3db8",
+    "relation": "produced",
+    "prov": {
+      "used_session": "01JZ8codexsess0008subagent",
+      "used_decision": "decision:01JZ8codexsess0008subagent:1",
+      "was_generated_by_session": "01JZ8codexsess0008subagent",
+      "t_use": "2026-06-22T21:00:06.5Z",
+      "t_gen": "2026-06-22T21:00:22.3Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:01JZ8codexsess0008subagent:1",
+    "to": "episode:b662f854f070d854d05850a268f3c9e2bd8b8b2cff5b058027ebffabf2839dd1",
+    "relation": "produced",
+    "prov": {
+      "used_session": "01JZ8codexsess0008subagent",
+      "used_decision": "decision:01JZ8codexsess0008subagent:1",
+      "was_generated_by_session": "01JZ8codexsess0008subagent",
+      "t_use": "2026-06-22T21:00:06.5Z",
+      "t_gen": "2026-06-22T21:00:22.3Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__tool_failure.snap
new file mode 100644
index 0000000..550eaae
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__codex__v2__tool_failure.snap
@@ -0,0 +1,49 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "01JZ8codexsess0004toolfail",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Switch the config loader to read from config.toml instead of config.yaml.",
+    "markers": [
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 50
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/codex/v2/tool_failure.jsonl",
+        "byte_offset": 273,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "toml instead of config.",
+    "considered_options": [
+      {
+        "text": "config",
+        "chosen": false
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__copilot__v1__ban.snap b/crates/memscribe-testkit/tests/snapshots/nodes__copilot__v1__ban.snap
new file mode 100644
index 0000000..6ecce08
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__copilot__v1__ban.snap
@@ -0,0 +1,50 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "copilot-thread-003",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "We will never add a dependency on left-pad. Do not introduce it under any circumstances.",
+    "markers": [
+      {
+        "rule_id": "ban.negated_use",
+        "category": "ban",
+        "matched_text": "never add",
+        "offset": 8
+      },
+      {
+        "rule_id": "imperative.must_always_never",
+        "category": "imperative",
+        "matched_text": "never",
+        "offset": 8
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/copilot/v1/ban.jsonl",
+        "byte_offset": 215,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "We will never add a dependency on left-pad.",
+    "considered_options": [],
+    "is_ban": true,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__copilot__v1__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/nodes__copilot__v1__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..17266a3
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__copilot__v1__happy_path_decision_then_edits.snap
@@ -0,0 +1,135 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "copilot-thread-001",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use Postgres instead of MySQL for the persistence layer.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 19
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/copilot/v1/happy_path_decision_then_edits.jsonl",
+        "byte_offset": 216,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use Postgres instead of MySQL for the persistence layer.",
+    "considered_options": [
+      {
+        "text": "MySQL",
+        "chosen": false
+      },
+      {
+        "text": "Postgres",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/db/config.rs",
+    "diff": {
+      "path": "src/db/config.rs",
+      "old": "let url = \"mysql://localhost/orbit\";",
+      "new": "let url = \"postgres://localhost/orbit\";",
+      "unified": "@@ -1 +1 @@\n-let url = \"mysql://localhost/orbit\";\n+let url = \"postgres://localhost/orbit\";",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "a1b2c3d4e5f6",
+      "branch": "main"
+    },
+    "episode_id": "87884849abc6ce82577d208874169dd8e912a593e2cf556ce0d8e1a5f2d41171"
+  },
+  {
+    "node": "episode",
+    "path": "src/db/schema.rs",
+    "diff": {
+      "path": "src/db/schema.rs",
+      "old": "// MySQL dialect\n",
+      "new": "// Postgres dialect\nuse sqlx::postgres::PgPool;\n",
+      "unified": null,
+      "added_lines": 2,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "a1b2c3d4e5f6",
+      "branch": "main"
+    },
+    "episode_id": "a758af4d3c7aae84de30d7876c0c2f76f8ea00b3505133ecf7412c381105235e"
+  },
+  {
+    "node": "binding",
+    "from": "decision:copilot-thread-001:1",
+    "to": "episode:87884849abc6ce82577d208874169dd8e912a593e2cf556ce0d8e1a5f2d41171",
+    "relation": "produced",
+    "prov": {
+      "used_session": "copilot-thread-001",
+      "used_decision": "decision:copilot-thread-001:1",
+      "was_generated_by_session": "copilot-thread-001",
+      "t_use": "2026-06-22T10:00:05Z",
+      "t_gen": "2026-06-22T10:00:09Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:copilot-thread-001:1",
+    "to": "episode:a758af4d3c7aae84de30d7876c0c2f76f8ea00b3505133ecf7412c381105235e",
+    "relation": "produced",
+    "prov": {
+      "used_session": "copilot-thread-001",
+      "used_decision": "decision:copilot-thread-001:1",
+      "was_generated_by_session": "copilot-thread-001",
+      "t_use": "2026-06-22T10:00:05Z",
+      "t_gen": "2026-06-22T10:00:09Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__copilot__v1__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/nodes__copilot__v1__rejected_alternative.snap
new file mode 100644
index 0000000..f589c80
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__copilot__v1__rejected_alternative.snap
@@ -0,0 +1,91 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "copilot-thread-002",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use Stripe instead of PayPal for the payment provider.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 17
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/copilot/v1/rejected_alternative.jsonl",
+        "byte_offset": 228,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use Stripe instead of PayPal for the payment provider.",
+    "considered_options": [
+      {
+        "text": "PayPal",
+        "chosen": false
+      },
+      {
+        "text": "Stripe",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/payments/provider.rs",
+    "diff": {
+      "path": "src/payments/provider.rs",
+      "old": "use paypal_sdk::Client;",
+      "new": "use stripe::Client;",
+      "unified": "@@ -1 +1 @@\n-use paypal_sdk::Client;\n+use stripe::Client;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "f00dcafe1234",
+      "branch": "feat/payments"
+    },
+    "episode_id": "1b351df7a9cbcdd0092f44ca28904f99a89922befd9cf2866ff41502e8d29bb7"
+  },
+  {
+    "node": "binding",
+    "from": "decision:copilot-thread-002:1",
+    "to": "episode:1b351df7a9cbcdd0092f44ca28904f99a89922befd9cf2866ff41502e8d29bb7",
+    "relation": "produced",
+    "prov": {
+      "used_session": "copilot-thread-002",
+      "used_decision": "decision:copilot-thread-002:1",
+      "was_generated_by_session": "copilot-thread-002",
+      "t_use": "2026-06-22T11:00:08Z",
+      "t_gen": "2026-06-22T11:00:14Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__copilot__v1__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/nodes__copilot__v1__tool_failure.snap
new file mode 100644
index 0000000..e9c6f75
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__copilot__v1__tool_failure.snap
@@ -0,0 +1,5 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__cursor__v1__ban.snap b/crates/memscribe-testkit/tests/snapshots/nodes__cursor__v1__ban.snap
new file mode 100644
index 0000000..5dbb67d
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__cursor__v1__ban.snap
@@ -0,0 +1,93 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "cur-sess-003",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "We will never add a dependency on left-pad. Use the standard library padding instead, always.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "Use",
+        "offset": 44
+      },
+      {
+        "rule_id": "ban.negated_use",
+        "category": "ban",
+        "matched_text": "never add",
+        "offset": 8
+      },
+      {
+        "rule_id": "imperative.must_always_never",
+        "category": "imperative",
+        "matched_text": "never",
+        "offset": 8
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/cursor/v1/ban.jsonl",
+        "byte_offset": 162,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Use the standard library padding instead, always.",
+    "considered_options": [
+      {
+        "text": "the",
+        "chosen": true
+      }
+    ],
+    "is_ban": true,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/format.rs",
+    "diff": {
+      "path": "src/format.rs",
+      "old": "use left_pad::left_pad;\nlet s = left_pad(id, 8);",
+      "new": "let s = format!(\"{id:>8}\");",
+      "unified": "@@ -1,2 +1 @@\n-use left_pad::left_pad;\n-let s = left_pad(id, 8);\n+let s = format!(\"{id:>8}\");\n",
+      "added_lines": 1,
+      "removed_lines": 2
+    },
+    "git": {
+      "sha": "a17de4490b22",
+      "branch": "main"
+    },
+    "episode_id": "c91c60a318d5d4d7b552f4cde429d542f10ec282629508007588fcacb301e6b2"
+  },
+  {
+    "node": "binding",
+    "from": "decision:cur-sess-003:1",
+    "to": "episode:c91c60a318d5d4d7b552f4cde429d542f10ec282629508007588fcacb301e6b2",
+    "relation": "produced",
+    "prov": {
+      "used_session": "cur-sess-003",
+      "used_decision": "decision:cur-sess-003:1",
+      "was_generated_by_session": "cur-sess-003",
+      "t_use": "2026-06-22T16:20:00Z",
+      "t_gen": "2026-06-22T16:20:11Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__cursor__v1__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/nodes__cursor__v1__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..fa8f03d
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__cursor__v1__happy_path_decision_then_edits.snap
@@ -0,0 +1,135 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "cur-sess-001",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use Postgres instead of MySQL for the orders service. Update the config and the connection module.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 19
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/cursor/v1/happy_path_decision_then_edits.jsonl",
+        "byte_offset": 181,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use Postgres instead of MySQL for the orders service.",
+    "considered_options": [
+      {
+        "text": "MySQL",
+        "chosen": false
+      },
+      {
+        "text": "Postgres",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "config/database.toml",
+    "diff": {
+      "path": "config/database.toml",
+      "old": "[database]\nengine = \"mysql\"\nport = 3306\n",
+      "new": "[database]\nengine = \"postgres\"\nport = 5432\n",
+      "unified": "@@ -1,3 +1,3 @@\n [database]\n-engine = \"mysql\"\n-port = 3306\n+engine = \"postgres\"\n+port = 5432\n",
+      "added_lines": 2,
+      "removed_lines": 2
+    },
+    "git": {
+      "sha": "9f2c1ad4b7e0",
+      "branch": "main"
+    },
+    "episode_id": "d47bc20d5db7fb6070d26165756dd148e877bcd96febb9232574f6edbf6ec037"
+  },
+  {
+    "node": "episode",
+    "path": "src/db/connect.rs",
+    "diff": {
+      "path": "src/db/connect.rs",
+      "old": "let url = \"mysql://localhost:3306/orders\";",
+      "new": "let url = \"postgres://localhost:5432/orders\";",
+      "unified": "@@ -1 +1 @@\n-let url = \"mysql://localhost:3306/orders\";\n+let url = \"postgres://localhost:5432/orders\";\n",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "9f2c1ad4b7e0",
+      "branch": "main"
+    },
+    "episode_id": "6e0f5572a9ce7d0bd617a39c474ef009f7c3c70719d798309b21b470d5888f50"
+  },
+  {
+    "node": "binding",
+    "from": "decision:cur-sess-001:1",
+    "to": "episode:6e0f5572a9ce7d0bd617a39c474ef009f7c3c70719d798309b21b470d5888f50",
+    "relation": "produced",
+    "prov": {
+      "used_session": "cur-sess-001",
+      "used_decision": "decision:cur-sess-001:1",
+      "was_generated_by_session": "cur-sess-001",
+      "t_use": "2026-06-22T14:01:10Z",
+      "t_gen": "2026-06-22T14:01:31Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:cur-sess-001:1",
+    "to": "episode:d47bc20d5db7fb6070d26165756dd148e877bcd96febb9232574f6edbf6ec037",
+    "relation": "produced",
+    "prov": {
+      "used_session": "cur-sess-001",
+      "used_decision": "decision:cur-sess-001:1",
+      "was_generated_by_session": "cur-sess-001",
+      "t_use": "2026-06-22T14:01:10Z",
+      "t_gen": "2026-06-22T14:01:18Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__cursor__v1__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/nodes__cursor__v1__rejected_alternative.snap
new file mode 100644
index 0000000..2e389cf
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__cursor__v1__rejected_alternative.snap
@@ -0,0 +1,124 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "cur-sess-002",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Should we use Redux or Zustand for the new dashboard state?",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 10
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/cursor/v1/rejected_alternative.jsonl",
+        "byte_offset": 173,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Should we use Redux or Zustand for the new dashboard state?",
+    "considered_options": [
+      {
+        "text": "Redux",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "conversation",
+    "session_id": "cur-sess-002",
+    "turn_range": {
+      "start": 3,
+      "end": 4
+    },
+    "text": "Agreed, reject Redux. Let's go with Zustand and wire up the store.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "Let's go with",
+        "offset": 22
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/cursor/v1/rejected_alternative.jsonl",
+        "byte_offset": 666,
+        "line_no": 4
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's go with Zustand and wire up the store.",
+    "considered_options": [
+      {
+        "text": "Zustand",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 3,
+      "end": 4
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/store/dashboard.ts",
+    "diff": {
+      "path": "src/store/dashboard.ts",
+      "old": "",
+      "new": "import { create } from 'zustand';\nexport const useDashboard = create((set) => ({ panels: [] }));\n",
+      "unified": "@@ -0,0 +1,2 @@\n+import { create } from 'zustand';\n+export const useDashboard = create((set) => ({ panels: [] }));\n",
+      "added_lines": 2,
+      "removed_lines": 0
+    },
+    "git": {
+      "sha": "3b81fe09cc12",
+      "branch": "feat/state-mgmt"
+    },
+    "episode_id": "4b6d7e32b2056c2b250ee1571e4033f982adb608af744cfd31c5390bbdd1df44"
+  },
+  {
+    "node": "binding",
+    "from": "decision:cur-sess-002:3",
+    "to": "episode:4b6d7e32b2056c2b250ee1571e4033f982adb608af744cfd31c5390bbdd1df44",
+    "relation": "produced",
+    "prov": {
+      "used_session": "cur-sess-002",
+      "used_decision": "decision:cur-sess-002:3",
+      "was_generated_by_session": "cur-sess-002",
+      "t_use": "2026-06-22T15:10:40Z",
+      "t_gen": "2026-06-22T15:10:52Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__cursor__v1__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/nodes__cursor__v1__tool_failure.snap
new file mode 100644
index 0000000..e9c6f75
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__cursor__v1__tool_failure.snap
@@ -0,0 +1,5 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__gemini__legacy_json__version_tolerance.snap b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__legacy_json__version_tolerance.snap
new file mode 100644
index 0000000..9b3f19c
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__legacy_json__version_tolerance.snap
@@ -0,0 +1,59 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "legacy-0001",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use Postgres instead of MySQL for the orders service.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 19
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/gemini/legacy_json/version_tolerance.jsonl",
+        "byte_offset": 183,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use Postgres instead of MySQL for the orders service.",
+    "considered_options": [
+      {
+        "text": "MySQL",
+        "chosen": false
+      },
+      {
+        "text": "Postgres",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__ban.snap b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__ban.snap
new file mode 100644
index 0000000..2f2e510
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__ban.snap
@@ -0,0 +1,61 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "a1b2c3d4-0003",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Important policy: we will never add a dependency on left-pad. Use the std library instead.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "Use",
+        "offset": 62
+      },
+      {
+        "rule_id": "ban.negated_use",
+        "category": "ban",
+        "matched_text": "never add",
+        "offset": 26
+      },
+      {
+        "rule_id": "imperative.must_always_never",
+        "category": "imperative",
+        "matched_text": "never",
+        "offset": 26
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/gemini/v1/ban.jsonl",
+        "byte_offset": 150,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Use the std library instead.",
+    "considered_options": [
+      {
+        "text": "the",
+        "chosen": true
+      }
+    ],
+    "is_ban": true,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..e068460
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__happy_path_decision_then_edits.snap
@@ -0,0 +1,135 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "a1b2c3d4-0001",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use Postgres instead of MySQL for the orders service. Update the config and the migration runner.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 19
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/gemini/v1/happy_path_decision_then_edits.jsonl",
+        "byte_offset": 185,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use Postgres instead of MySQL for the orders service.",
+    "considered_options": [
+      {
+        "text": "MySQL",
+        "chosen": false
+      },
+      {
+        "text": "Postgres",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "config/database.toml",
+    "diff": {
+      "path": "config/database.toml",
+      "old": "[db]\ndriver = \"mysql\"\nurl = \"mysql://localhost:3306/orders\"\n",
+      "new": "[db]\ndriver = \"postgres\"\nurl = \"postgres://localhost:5432/orders\"\n",
+      "unified": "@@ -1,3 +1,3 @@\n [db]\n-driver = \"mysql\"\n-url = \"mysql://localhost:3306/orders\"\n+driver = \"postgres\"\n+url = \"postgres://localhost:5432/orders\"\n",
+      "added_lines": 2,
+      "removed_lines": 2
+    },
+    "git": {
+      "sha": "9f1c0aa2b3",
+      "branch": "main"
+    },
+    "episode_id": "b78ab9a156d7d598f3543b766cf27bca39c9fd43bf73511518bcd4e57f2429a0"
+  },
+  {
+    "node": "episode",
+    "path": "src/migrate.rs",
+    "diff": {
+      "path": "src/migrate.rs",
+      "old": "use mysql::Pool;\n",
+      "new": "use postgres::Client;\n",
+      "unified": "@@ -1 +1 @@\n-use mysql::Pool;\n+use postgres::Client;\n",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "9f1c0aa2b3",
+      "branch": "main"
+    },
+    "episode_id": "9e53e5876500176fb220e30f1475ff32c5e3c551142ca14e0bc990ae1db855d6"
+  },
+  {
+    "node": "binding",
+    "from": "decision:a1b2c3d4-0001:1",
+    "to": "episode:9e53e5876500176fb220e30f1475ff32c5e3c551142ca14e0bc990ae1db855d6",
+    "relation": "produced",
+    "prov": {
+      "used_session": "a1b2c3d4-0001",
+      "used_decision": "decision:a1b2c3d4-0001:1",
+      "was_generated_by_session": "a1b2c3d4-0001",
+      "t_use": "2026-06-22T10:00:00Z",
+      "t_gen": "2026-06-22T10:00:03.12Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:a1b2c3d4-0001:1",
+    "to": "episode:b78ab9a156d7d598f3543b766cf27bca39c9fd43bf73511518bcd4e57f2429a0",
+    "relation": "produced",
+    "prov": {
+      "used_session": "a1b2c3d4-0001",
+      "used_decision": "decision:a1b2c3d4-0001:1",
+      "was_generated_by_session": "a1b2c3d4-0001",
+      "t_use": "2026-06-22T10:00:00Z",
+      "t_gen": "2026-06-22T10:00:03.12Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__interleaved_arcs.snap b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__interleaved_arcs.snap
new file mode 100644
index 0000000..85c4a44
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__interleaved_arcs.snap
@@ -0,0 +1,250 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "g-interleaved-0001",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use Postgres for the orders store. Update the connection module and the schema.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+        "byte_offset": 194,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use Postgres for the orders store.",
+    "considered_options": [
+      {
+        "text": "Postgres",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/db.rs",
+    "diff": {
+      "path": "src/db.rs",
+      "old": "use mysql::Pool;\n",
+      "new": "use postgres::Client;\n",
+      "unified": "@@ -1 +1 @@\n-use mysql::Pool;\n+use postgres::Client;\n",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "a1a1a1a1a1",
+      "branch": "main"
+    },
+    "episode_id": "931069dd1718c8c61609b0e97b3f3eaf3441e987f38b65ed9aacf96e2a9bf5e6"
+  },
+  {
+    "node": "episode",
+    "path": "src/schema.rs",
+    "diff": {
+      "path": "src/schema.rs",
+      "old": "-- mysql schema\n",
+      "new": "-- postgres schema\n",
+      "unified": "@@ -1 +1 @@\n--- mysql schema\n+-- postgres schema\n",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "a1a1a1a1a1",
+      "branch": "main"
+    },
+    "episode_id": "6a95432062f10904dce08fbfbf6c5f709a5084bcd3b43b6c1302889b08678b52"
+  },
+  {
+    "node": "conversation",
+    "session_id": "g-interleaved-0001",
+    "turn_range": {
+      "start": 9,
+      "end": 10
+    },
+    "text": "Now switch to a bounded connection pool in the same db module instead of a single client.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "switch to",
+        "offset": 4
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 62
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/gemini/v1/interleaved_arcs.jsonl",
+        "byte_offset": 1194,
+        "line_no": 4
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Now switch to a bounded connection pool in the same db module instead of a single client.",
+    "considered_options": [
+      {
+        "text": "a",
+        "chosen": false
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 9,
+      "end": 10
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/db.rs",
+    "diff": {
+      "path": "src/db.rs",
+      "old": "use postgres::Client;\n",
+      "new": "use deadpool_postgres::Pool;\n",
+      "unified": "@@ -1 +1 @@\n-use postgres::Client;\n+use deadpool_postgres::Pool;\n",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "a1a1a1a1a1",
+      "branch": "main"
+    },
+    "episode_id": "83fc8ad09cfae229f6e8ddf1429a02ee6d14cf50c9ab015f639e97b9718750c8"
+  },
+  {
+    "node": "episode",
+    "path": "src/pool.rs",
+    "diff": {
+      "path": "src/pool.rs",
+      "old": null,
+      "new": "pub struct PoolCfg { pub max: usize }\n",
+      "unified": "@@ -0,0 +1 @@\n+pub struct PoolCfg { pub max: usize }\n",
+      "added_lines": 1,
+      "removed_lines": 0
+    },
+    "git": {
+      "sha": "a1a1a1a1a1",
+      "branch": "main"
+    },
+    "episode_id": "a628c0fb5fc9081b6d99e8cc6f7ea896565d08a8d43e5dec50d85fb2d04fffe0"
+  },
+  {
+    "node": "binding",
+    "from": "decision:g-interleaved-0001:1",
+    "to": "episode:6a95432062f10904dce08fbfbf6c5f709a5084bcd3b43b6c1302889b08678b52",
+    "relation": "produced",
+    "prov": {
+      "used_session": "g-interleaved-0001",
+      "used_decision": "decision:g-interleaved-0001:1",
+      "was_generated_by_session": "g-interleaved-0001",
+      "t_use": "2026-06-22T09:00:00Z",
+      "t_gen": "2026-06-22T09:00:02Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.25,
+      "confidence": 0.5,
+      "lift": 2.0,
+      "phi": 0.57735,
+      "p": 0.248213
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:g-interleaved-0001:1",
+    "to": "episode:931069dd1718c8c61609b0e97b3f3eaf3441e987f38b65ed9aacf96e2a9bf5e6",
+    "relation": "produced",
+    "prov": {
+      "used_session": "g-interleaved-0001",
+      "used_decision": "decision:g-interleaved-0001:1",
+      "was_generated_by_session": "g-interleaved-0001",
+      "t_use": "2026-06-22T09:00:00Z",
+      "t_gen": "2026-06-22T09:00:02Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.25,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:g-interleaved-0001:9",
+    "to": "episode:83fc8ad09cfae229f6e8ddf1429a02ee6d14cf50c9ab015f639e97b9718750c8",
+    "relation": "produced",
+    "prov": {
+      "used_session": "g-interleaved-0001",
+      "used_decision": "decision:g-interleaved-0001:9",
+      "was_generated_by_session": "g-interleaved-0001",
+      "t_use": "2026-06-22T09:05:00Z",
+      "t_gen": "2026-06-22T09:05:02Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.25,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:g-interleaved-0001:9",
+    "to": "episode:a628c0fb5fc9081b6d99e8cc6f7ea896565d08a8d43e5dec50d85fb2d04fffe0",
+    "relation": "produced",
+    "prov": {
+      "used_session": "g-interleaved-0001",
+      "used_decision": "decision:g-interleaved-0001:9",
+      "was_generated_by_session": "g-interleaved-0001",
+      "t_use": "2026-06-22T09:05:00Z",
+      "t_gen": "2026-06-22T09:05:02Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.25,
+      "confidence": 0.5,
+      "lift": 2.0,
+      "phi": 0.57735,
+      "p": 0.248213
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__multi_edit_single_commit.snap b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__multi_edit_single_commit.snap
new file mode 100644
index 0000000..e4d7be4
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__multi_edit_single_commit.snap
@@ -0,0 +1,163 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "g-multiedit-0001",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "We will use Argon2 for password hashing across the auth service. Update the hasher, the verifier, and the config in one pass.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 8
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/gemini/v1/multi_edit_single_commit.jsonl",
+        "byte_offset": 195,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "We will use Argon2 for password hashing across the auth service.",
+    "considered_options": [
+      {
+        "text": "Argon2",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/auth/hasher.rs",
+    "diff": {
+      "path": "src/auth/hasher.rs",
+      "old": "use bcrypt::hash;\n",
+      "new": "use argon2::Argon2;\n",
+      "unified": "@@ -1 +1 @@\n-use bcrypt::hash;\n+use argon2::Argon2;\n",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "b2b2b2b2b2",
+      "branch": "feat/argon2"
+    },
+    "episode_id": "15e4b7e951721a92cc59d23457313e646aeff9b452ffea3aec88aaa4f6b72809"
+  },
+  {
+    "node": "episode",
+    "path": "src/auth/verifier.rs",
+    "diff": {
+      "path": "src/auth/verifier.rs",
+      "old": "use bcrypt::verify;\n",
+      "new": "use argon2::PasswordVerifier;\n",
+      "unified": "@@ -1 +1 @@\n-use bcrypt::verify;\n+use argon2::PasswordVerifier;\n",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "b2b2b2b2b2",
+      "branch": "feat/argon2"
+    },
+    "episode_id": "f0315f623442fdb8c58ee4f260bb0f4c476e97c2100ba80cf94f84be1f1ae075"
+  },
+  {
+    "node": "episode",
+    "path": "config/auth.toml",
+    "diff": {
+      "path": "config/auth.toml",
+      "old": "hash = \"bcrypt\"\n",
+      "new": "hash = \"argon2\"\n",
+      "unified": "@@ -1 +1 @@\n-hash = \"bcrypt\"\n+hash = \"argon2\"\n",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "b2b2b2b2b2",
+      "branch": "feat/argon2"
+    },
+    "episode_id": "c6341552eebb634b23180453678bcf136c6e35e1e7f8b5552dc8c83bd837a01f"
+  },
+  {
+    "node": "binding",
+    "from": "decision:g-multiedit-0001:1",
+    "to": "episode:15e4b7e951721a92cc59d23457313e646aeff9b452ffea3aec88aaa4f6b72809",
+    "relation": "produced",
+    "prov": {
+      "used_session": "g-multiedit-0001",
+      "used_decision": "decision:g-multiedit-0001:1",
+      "was_generated_by_session": "g-multiedit-0001",
+      "t_use": "2026-06-22T14:00:00Z",
+      "t_gen": "2026-06-22T14:00:03Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.333333,
+      "confidence": 0.333333,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:g-multiedit-0001:1",
+    "to": "episode:c6341552eebb634b23180453678bcf136c6e35e1e7f8b5552dc8c83bd837a01f",
+    "relation": "produced",
+    "prov": {
+      "used_session": "g-multiedit-0001",
+      "used_decision": "decision:g-multiedit-0001:1",
+      "was_generated_by_session": "g-multiedit-0001",
+      "t_use": "2026-06-22T14:00:00Z",
+      "t_gen": "2026-06-22T14:00:03Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.333333,
+      "confidence": 0.333333,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:g-multiedit-0001:1",
+    "to": "episode:f0315f623442fdb8c58ee4f260bb0f4c476e97c2100ba80cf94f84be1f1ae075",
+    "relation": "produced",
+    "prov": {
+      "used_session": "g-multiedit-0001",
+      "used_decision": "decision:g-multiedit-0001:1",
+      "was_generated_by_session": "g-multiedit-0001",
+      "t_use": "2026-06-22T14:00:00Z",
+      "t_gen": "2026-06-22T14:00:03Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.333333,
+      "confidence": 0.333333,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__no_commitment_marker.snap b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__no_commitment_marker.snap
new file mode 100644
index 0000000..b437e1d
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__no_commitment_marker.snap
@@ -0,0 +1,23 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "episode",
+    "path": "src/widget.rs",
+    "diff": {
+      "path": "src/widget.rs",
+      "old": "fn render() { }  \n",
+      "new": "fn render() {}\n",
+      "unified": "@@ -1 +1 @@\n-fn render() { }  \n+fn render() {}\n",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "e5e5e5e5e5",
+      "branch": "main"
+    },
+    "episode_id": "5d3321cd409dee36fb6152017bf0bd6908b06cbb5042bf0ce0179bc47e6a0c58"
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__rejected_alternative.snap
new file mode 100644
index 0000000..2aad0d4
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__rejected_alternative.snap
@@ -0,0 +1,91 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "a1b2c3d4-0002",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "For the checkout flow, use Stripe instead of PayPal. Wire up the Stripe client.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 23
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 34
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/gemini/v1/rejected_alternative.jsonl",
+        "byte_offset": 156,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "For the checkout flow, use Stripe instead of PayPal.",
+    "considered_options": [
+      {
+        "text": "PayPal",
+        "chosen": false
+      },
+      {
+        "text": "Stripe",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/payments/stripe.rs",
+    "diff": {
+      "path": "src/payments/stripe.rs",
+      "old": null,
+      "new": "pub struct StripeClient { key: String }\n",
+      "unified": "@@ -0,0 +1 @@\n+pub struct StripeClient { key: String }\n",
+      "added_lines": 1,
+      "removed_lines": 0
+    },
+    "git": {
+      "sha": "77ab12cd",
+      "branch": "feat/payments"
+    },
+    "episode_id": "845329c8cf0c4ffc5521e8081954d2b3b2819d3cbc4972324c2966f338aaee41"
+  },
+  {
+    "node": "binding",
+    "from": "decision:a1b2c3d4-0002:1",
+    "to": "episode:845329c8cf0c4ffc5521e8081954d2b3b2819d3cbc4972324c2966f338aaee41",
+    "relation": "produced",
+    "prov": {
+      "used_session": "a1b2c3d4-0002",
+      "used_decision": "decision:a1b2c3d4-0002:1",
+      "was_generated_by_session": "a1b2c3d4-0002",
+      "t_use": "2026-06-22T11:15:00Z",
+      "t_gen": "2026-06-22T11:15:04.5Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__rewind_compaction.snap b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__rewind_compaction.snap
new file mode 100644
index 0000000..8f5036b
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__rewind_compaction.snap
@@ -0,0 +1,151 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "g-rewind-0001",
+    "turn_range": {
+      "start": 2,
+      "end": 3
+    },
+    "text": "Use MongoDB for the catalog store. Wire up the document client.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "Use",
+        "offset": 0
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+        "byte_offset": 310,
+        "line_no": 3
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Use MongoDB for the catalog store.",
+    "considered_options": [
+      {
+        "text": "MongoDB",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": "rewind:g-rewind-0001:7",
+    "confirmation": null,
+    "source_span": {
+      "start": 2,
+      "end": 3
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/store.rs",
+    "diff": {
+      "path": "src/store.rs",
+      "old": "pub struct Store;\n",
+      "new": "use mongodb::Client;\npub struct Store(Client);\n",
+      "unified": "@@ -1 +1,2 @@\n-pub struct Store;\n+use mongodb::Client;\n+pub struct Store(Client);\n",
+      "added_lines": 2,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "c3c3c3c3c3",
+      "branch": "main"
+    },
+    "episode_id": "ee23c7a84ad21ea90971c8aabd8fd1c749ba107f8e2c067a98e6213d17226fa5"
+  },
+  {
+    "node": "conversation",
+    "session_id": "g-rewind-0001",
+    "turn_range": {
+      "start": 8,
+      "end": 9
+    },
+    "text": "On reflection, use Postgres with JSONB for the catalog store instead of MongoDB.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 15
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 61
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/gemini/v1/rewind_compaction.jsonl",
+        "byte_offset": 1078,
+        "line_no": 6
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "On reflection, use Postgres with JSONB for the catalog store instead of MongoDB.",
+    "considered_options": [
+      {
+        "text": "MongoDB",
+        "chosen": false
+      },
+      {
+        "text": "Postgres",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 8,
+      "end": 9
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/store.rs",
+    "diff": {
+      "path": "src/store.rs",
+      "old": "use mongodb::Client;\npub struct Store(Client);\n",
+      "new": "use postgres::Client;\npub struct Store(Client);\n",
+      "unified": "@@ -1,2 +1,2 @@\n-use mongodb::Client;\n+use postgres::Client;\n pub struct Store(Client);\n",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "c3c3c3c3c3",
+      "branch": "main"
+    },
+    "episode_id": "97025faaf6aefb035e1cefb8ed9961b6a0edce1333b21bcbda0facbe3cd751e8"
+  },
+  {
+    "node": "binding",
+    "from": "decision:g-rewind-0001:8",
+    "to": "episode:97025faaf6aefb035e1cefb8ed9961b6a0edce1333b21bcbda0facbe3cd751e8",
+    "relation": "produced",
+    "prov": {
+      "used_session": "g-rewind-0001",
+      "used_decision": "decision:g-rewind-0001:8",
+      "was_generated_by_session": "g-rewind-0001",
+      "t_use": "2026-06-22T15:10:00Z",
+      "t_gen": "2026-06-22T15:10:03Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__subagent_thread.snap b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__subagent_thread.snap
new file mode 100644
index 0000000..3673428
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__subagent_thread.snap
@@ -0,0 +1,91 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "g-subagent-0001",
+    "turn_range": {
+      "start": 3,
+      "end": 4
+    },
+    "text": "In this sub-thread, use a streaming pull-parser for the ingest format instead of the buffered one.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 20
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 70
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/gemini/v1/subagent_thread.jsonl",
+        "byte_offset": 668,
+        "line_no": 4
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "In this sub-thread, use a streaming pull-parser for the ingest format instead of the buffered one.",
+    "considered_options": [
+      {
+        "text": "the",
+        "chosen": false
+      },
+      {
+        "text": "a",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 3,
+      "end": 4
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/ingest/parser.rs",
+    "diff": {
+      "path": "src/ingest/parser.rs",
+      "old": "let buf = read_all(input);\n",
+      "new": "let mut reader = PullParser::new(input);\n",
+      "unified": "@@ -1 +1 @@\n-let buf = read_all(input);\n+let mut reader = PullParser::new(input);\n",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "d4d4d4d4d4",
+      "branch": "main"
+    },
+    "episode_id": "88bee143d2c0e8e6adae2c7ec82782a1b5152fac140c379654ef4de9afec9dd0"
+  },
+  {
+    "node": "binding",
+    "from": "decision:g-subagent-0001:3",
+    "to": "episode:88bee143d2c0e8e6adae2c7ec82782a1b5152fac140c379654ef4de9afec9dd0",
+    "relation": "produced",
+    "prov": {
+      "used_session": "g-subagent-0001",
+      "used_decision": "decision:g-subagent-0001:3",
+      "was_generated_by_session": "g-subagent-0001",
+      "t_use": "2026-06-22T16:01:00Z",
+      "t_gen": "2026-06-22T16:01:03Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__tool_failure.snap
new file mode 100644
index 0000000..e9c6f75
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__gemini__v1__tool_failure.snap
@@ -0,0 +1,5 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__otel__genai__ban.snap b/crates/memscribe-testkit/tests/snapshots/nodes__otel__genai__ban.snap
new file mode 100644
index 0000000..9946897
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__otel__genai__ban.snap
@@ -0,0 +1,50 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "sess-ban-001",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "We will never add a dependency on left-pad. Keep the tree clean.",
+    "markers": [
+      {
+        "rule_id": "ban.negated_use",
+        "category": "ban",
+        "matched_text": "never add",
+        "offset": 8
+      },
+      {
+        "rule_id": "imperative.must_always_never",
+        "category": "imperative",
+        "matched_text": "never",
+        "offset": 8
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/otel/genai/ban.jsonl",
+        "byte_offset": 272,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "We will never add a dependency on left-pad.",
+    "considered_options": [],
+    "is_ban": true,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__otel__genai__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/nodes__otel__genai__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..102f974
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__otel__genai__happy_path_decision_then_edits.snap
@@ -0,0 +1,135 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "sess-happy-001",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use Postgres instead of MySQL for the persistence layer.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 19
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/otel/genai/happy_path_decision_then_edits.jsonl",
+        "byte_offset": 276,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use Postgres instead of MySQL for the persistence layer.",
+    "considered_options": [
+      {
+        "text": "MySQL",
+        "chosen": false
+      },
+      {
+        "text": "Postgres",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "db/config.rs",
+    "diff": {
+      "path": "db/config.rs",
+      "old": null,
+      "new": null,
+      "unified": null,
+      "added_lines": 12,
+      "removed_lines": 4
+    },
+    "git": {
+      "sha": "abc1234",
+      "branch": "main"
+    },
+    "episode_id": "693cec409d8ab5556b605faed5e2c901525b4d964e9d9a2815bfbe6a4da1bf01"
+  },
+  {
+    "node": "episode",
+    "path": "db/pool.rs",
+    "diff": {
+      "path": "db/pool.rs",
+      "old": null,
+      "new": null,
+      "unified": null,
+      "added_lines": 30,
+      "removed_lines": 9
+    },
+    "git": {
+      "sha": "abc1234",
+      "branch": "main"
+    },
+    "episode_id": "92496e034cb5c70f7f8015bef66f30b4cc4eaeb6490875947d7c32732236af8d"
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-happy-001:1",
+    "to": "episode:693cec409d8ab5556b605faed5e2c901525b4d964e9d9a2815bfbe6a4da1bf01",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-happy-001",
+      "used_decision": "decision:sess-happy-001:1",
+      "was_generated_by_session": "sess-happy-001",
+      "t_use": "2026-06-22T10:00:05Z",
+      "t_gen": "2026-06-22T10:00:12Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-happy-001:1",
+    "to": "episode:92496e034cb5c70f7f8015bef66f30b4cc4eaeb6490875947d7c32732236af8d",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-happy-001",
+      "used_decision": "decision:sess-happy-001:1",
+      "was_generated_by_session": "sess-happy-001",
+      "t_use": "2026-06-22T10:00:05Z",
+      "t_gen": "2026-06-22T10:00:15Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__otel__genai__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/nodes__otel__genai__rejected_alternative.snap
new file mode 100644
index 0000000..54dffd8
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__otel__genai__rejected_alternative.snap
@@ -0,0 +1,91 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "sess-reject-001",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "We should use Stripe instead of PayPal for the checkout integration.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 10
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 21
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/otel/genai/rejected_alternative.jsonl",
+        "byte_offset": 283,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "We should use Stripe instead of PayPal for the checkout integration.",
+    "considered_options": [
+      {
+        "text": "PayPal",
+        "chosen": false
+      },
+      {
+        "text": "Stripe",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "checkout/stripe.rs",
+    "diff": {
+      "path": "checkout/stripe.rs",
+      "old": null,
+      "new": null,
+      "unified": null,
+      "added_lines": 48,
+      "removed_lines": 0
+    },
+    "git": {
+      "sha": "def5678",
+      "branch": "feat/payments"
+    },
+    "episode_id": "05edecd3835e7699baa29183276ba03358b1020702e2adf3453c08fad31b6336"
+  },
+  {
+    "node": "binding",
+    "from": "decision:sess-reject-001:1",
+    "to": "episode:05edecd3835e7699baa29183276ba03358b1020702e2adf3453c08fad31b6336",
+    "relation": "produced",
+    "prov": {
+      "used_session": "sess-reject-001",
+      "used_decision": "decision:sess-reject-001:1",
+      "was_generated_by_session": "sess-reject-001",
+      "t_use": "2026-06-22T11:00:04Z",
+      "t_gen": "2026-06-22T11:00:12Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__otel__genai__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/nodes__otel__genai__tool_failure.snap
new file mode 100644
index 0000000..e9c6f75
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__otel__genai__tool_failure.snap
@@ -0,0 +1,5 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__ban.snap b/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__ban.snap
new file mode 100644
index 0000000..1db7869
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__ban.snap
@@ -0,0 +1,82 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "vsc-sess-3",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "We will never add a dependency on left-pad. Roll our own padding helper.",
+    "markers": [
+      {
+        "rule_id": "ban.negated_use",
+        "category": "ban",
+        "matched_text": "never add",
+        "offset": 8
+      },
+      {
+        "rule_id": "imperative.must_always_never",
+        "category": "imperative",
+        "matched_text": "never",
+        "offset": 8
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/vscode/v1/ban.jsonl",
+        "byte_offset": 166,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "We will never add a dependency on left-pad.",
+    "considered_options": [],
+    "is_ban": true,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/util/pad.ts",
+    "diff": {
+      "path": "src/util/pad.ts",
+      "old": null,
+      "new": "export const pad = (s, n) => s.padStart(n, '0');",
+      "unified": null,
+      "added_lines": 1,
+      "removed_lines": 0
+    },
+    "git": {
+      "sha": "0011223344556677",
+      "branch": "main"
+    },
+    "episode_id": "954d592eada689aae0fc35bb3a3f90af3683538ddf8702df0217ad551780c6cb"
+  },
+  {
+    "node": "binding",
+    "from": "decision:vsc-sess-3:1",
+    "to": "episode:954d592eada689aae0fc35bb3a3f90af3683538ddf8702df0217ad551780c6cb",
+    "relation": "produced",
+    "prov": {
+      "used_session": "vsc-sess-3",
+      "used_decision": "decision:vsc-sess-3:1",
+      "was_generated_by_session": "vsc-sess-3",
+      "t_use": "2026-06-22T12:00:00Z",
+      "t_gen": "2026-06-22T12:00:06Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..22fc328
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__happy_path_decision_then_edits.snap
@@ -0,0 +1,135 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "vsc-sess-1",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use Postgres instead of MySQL for the persistence layer.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 19
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/vscode/v1/happy_path_decision_then_edits.jsonl",
+        "byte_offset": 162,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use Postgres instead of MySQL for the persistence layer.",
+    "considered_options": [
+      {
+        "text": "MySQL",
+        "chosen": false
+      },
+      {
+        "text": "Postgres",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/db/connect.ts",
+    "diff": {
+      "path": "src/db/connect.ts",
+      "old": "const url = 'mysql://localhost/acme';",
+      "new": "const url = 'postgres://localhost/acme';",
+      "unified": null,
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "a1b2c3d4e5f6",
+      "branch": "main"
+    },
+    "episode_id": "09c96e7aa322489f38c27c0c54d017ea7879c80cf061d0b6febc26c500e546a6"
+  },
+  {
+    "node": "episode",
+    "path": "src/db/schema.sql",
+    "diff": {
+      "path": "src/db/schema.sql",
+      "old": "ENGINE=InnoDB",
+      "new": "-- postgres has no storage engines",
+      "unified": "@@ -1 +1 @@\n-ENGINE=InnoDB\n+-- postgres has no storage engines",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "a1b2c3d4e5f6",
+      "branch": "main"
+    },
+    "episode_id": "ed82ab347d94a38a8d9fb9b6dff2f38b09331bb4552153dd86a750e7281b0df8"
+  },
+  {
+    "node": "binding",
+    "from": "decision:vsc-sess-1:1",
+    "to": "episode:09c96e7aa322489f38c27c0c54d017ea7879c80cf061d0b6febc26c500e546a6",
+    "relation": "produced",
+    "prov": {
+      "used_session": "vsc-sess-1",
+      "used_decision": "decision:vsc-sess-1:1",
+      "was_generated_by_session": "vsc-sess-1",
+      "t_use": "2026-06-22T10:00:00Z",
+      "t_gen": "2026-06-22T10:00:05Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:vsc-sess-1:1",
+    "to": "episode:ed82ab347d94a38a8d9fb9b6dff2f38b09331bb4552153dd86a750e7281b0df8",
+    "relation": "produced",
+    "prov": {
+      "used_session": "vsc-sess-1",
+      "used_decision": "decision:vsc-sess-1:1",
+      "was_generated_by_session": "vsc-sess-1",
+      "t_use": "2026-06-22T10:00:00Z",
+      "t_gen": "2026-06-22T10:00:05Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__native_chatsession.snap b/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__native_chatsession.snap
new file mode 100644
index 0000000..e9c6f75
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__native_chatsession.snap
@@ -0,0 +1,5 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__rejected_alternative.snap
new file mode 100644
index 0000000..5a048bc
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__rejected_alternative.snap
@@ -0,0 +1,81 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "vsc-sess-2",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Should we use Redis or just an in-memory LRU for the cache?",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 10
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/vscode/v1/rejected_alternative.jsonl",
+        "byte_offset": 168,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Should we use Redis or just an in-memory LRU for the cache?",
+    "considered_options": [
+      {
+        "text": "Redis",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/cache/lru.ts",
+    "diff": {
+      "path": "src/cache/lru.ts",
+      "old": null,
+      "new": "export class Lru { /* ... */ }",
+      "unified": null,
+      "added_lines": 1,
+      "removed_lines": 0
+    },
+    "git": {
+      "sha": "f6e5d4c3b2a1",
+      "branch": "feat/cache"
+    },
+    "episode_id": "e3bb2d813ca36317fb2f39829afcdb865d3fa6e9d679bfd90d5c24e012bc4075"
+  },
+  {
+    "node": "binding",
+    "from": "decision:vsc-sess-2:1",
+    "to": "episode:e3bb2d813ca36317fb2f39829afcdb865d3fa6e9d679bfd90d5c24e012bc4075",
+    "relation": "produced",
+    "prov": {
+      "used_session": "vsc-sess-2",
+      "used_decision": "decision:vsc-sess-2:1",
+      "was_generated_by_session": "vsc-sess-2",
+      "t_use": "2026-06-22T11:00:00Z",
+      "t_gen": "2026-06-22T11:00:20Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__tool_failure.snap
new file mode 100644
index 0000000..e9c6f75
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__vscode__v1__tool_failure.snap
@@ -0,0 +1,5 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__windsurf__v1__ban.snap b/crates/memscribe-testkit/tests/snapshots/nodes__windsurf__v1__ban.snap
new file mode 100644
index 0000000..5dc2b45
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__windsurf__v1__ban.snap
@@ -0,0 +1,82 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "ws-ban-003",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "We will never add a dependency on left-pad. Implement padding inline instead.",
+    "markers": [
+      {
+        "rule_id": "ban.negated_use",
+        "category": "ban",
+        "matched_text": "never add",
+        "offset": 8
+      },
+      {
+        "rule_id": "imperative.must_always_never",
+        "category": "imperative",
+        "matched_text": "never",
+        "offset": 8
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/windsurf/v1/ban.jsonl",
+        "byte_offset": 164,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "We will never add a dependency on left-pad.",
+    "considered_options": [],
+    "is_ban": true,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/util/pad.rs",
+    "diff": {
+      "path": "src/util/pad.rs",
+      "old": "",
+      "new": "pub fn pad_left(s: &str, width: usize) -> String { format!(\"{s:>width$}\") }",
+      "unified": "@@ -0,0 +1,1 @@\n+pub fn pad_left(s: &str, width: usize) -> String { format!(\"{s:>width$}\") }",
+      "added_lines": 1,
+      "removed_lines": 0
+    },
+    "git": {
+      "sha": "77c0de01",
+      "branch": "main"
+    },
+    "episode_id": "0067aa169c782405410ae6556b605a1ebc9612798d93f98d9cd9fc2a1b247592"
+  },
+  {
+    "node": "binding",
+    "from": "decision:ws-ban-003:1",
+    "to": "episode:0067aa169c782405410ae6556b605a1ebc9612798d93f98d9cd9fc2a1b247592",
+    "relation": "produced",
+    "prov": {
+      "used_session": "ws-ban-003",
+      "used_decision": "decision:ws-ban-003:1",
+      "was_generated_by_session": "ws-ban-003",
+      "t_use": "2026-06-22T16:20:00Z",
+      "t_gen": "2026-06-22T16:20:09Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__windsurf__v1__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/nodes__windsurf__v1__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..0b3a302
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__windsurf__v1__happy_path_decision_then_edits.snap
@@ -0,0 +1,135 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "ws-happy-001",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use Postgres instead of MySQL for the orders service.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 19
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/windsurf/v1/happy_path_decision_then_edits.jsonl",
+        "byte_offset": 166,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use Postgres instead of MySQL for the orders service.",
+    "considered_options": [
+      {
+        "text": "MySQL",
+        "chosen": false
+      },
+      {
+        "text": "Postgres",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/db/mod.rs",
+    "diff": {
+      "path": "src/db/mod.rs",
+      "old": "use mysql_async::Pool;",
+      "new": "use sqlx::postgres::PgPool;",
+      "unified": "@@ -1,1 +1,1 @@\n-use mysql_async::Pool;\n+use sqlx::postgres::PgPool;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "9f3c1a7e",
+      "branch": "main"
+    },
+    "episode_id": "2c213c5b2e2f7e0864fa6a49a580d052a483bc569894b8f3c32e451fcf0b21bb"
+  },
+  {
+    "node": "episode",
+    "path": "config/database.toml",
+    "diff": {
+      "path": "config/database.toml",
+      "old": "driver = \"mysql\"\nport = 3306",
+      "new": "driver = \"postgres\"\nport = 5432",
+      "unified": "@@ -1,2 +1,2 @@\n-driver = \"mysql\"\n-port = 3306\n+driver = \"postgres\"\n+port = 5432",
+      "added_lines": 2,
+      "removed_lines": 2
+    },
+    "git": {
+      "sha": "9f3c1a7e",
+      "branch": "main"
+    },
+    "episode_id": "7cb65f7f37d541ceb39094a1e95fe86b247cfec4dab0b65a889278ab73a14273"
+  },
+  {
+    "node": "binding",
+    "from": "decision:ws-happy-001:1",
+    "to": "episode:2c213c5b2e2f7e0864fa6a49a580d052a483bc569894b8f3c32e451fcf0b21bb",
+    "relation": "produced",
+    "prov": {
+      "used_session": "ws-happy-001",
+      "used_decision": "decision:ws-happy-001:1",
+      "was_generated_by_session": "ws-happy-001",
+      "t_use": "2026-06-22T14:00:00Z",
+      "t_gen": "2026-06-22T14:00:06Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:ws-happy-001:1",
+    "to": "episode:7cb65f7f37d541ceb39094a1e95fe86b247cfec4dab0b65a889278ab73a14273",
+    "relation": "produced",
+    "prov": {
+      "used_session": "ws-happy-001",
+      "used_decision": "decision:ws-happy-001:1",
+      "was_generated_by_session": "ws-happy-001",
+      "t_use": "2026-06-22T14:00:00Z",
+      "t_gen": "2026-06-22T14:00:06Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__windsurf__v1__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/nodes__windsurf__v1__rejected_alternative.snap
new file mode 100644
index 0000000..c46c2f4
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__windsurf__v1__rejected_alternative.snap
@@ -0,0 +1,119 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "ws-reject-002",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Should we use Redis or Memcached for the session cache?",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 10
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/windsurf/v1/rejected_alternative.jsonl",
+        "byte_offset": 175,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Should we use Redis or Memcached for the session cache?",
+    "considered_options": [
+      {
+        "text": "Redis",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "conversation",
+    "session_id": "ws-reject-002",
+    "turn_range": {
+      "start": 3,
+      "end": 4
+    },
+    "text": "Agreed, do not add Memcached. Wire up Redis.",
+    "markers": [
+      {
+        "rule_id": "ban.negated_use",
+        "category": "ban",
+        "matched_text": "do not add",
+        "offset": 8
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/windsurf/v1/rejected_alternative.jsonl",
+        "byte_offset": 640,
+        "line_no": 4
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Agreed, do not add Memcached.",
+    "considered_options": [],
+    "is_ban": true,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 3,
+      "end": 4
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/cache.rs",
+    "diff": {
+      "path": "src/cache.rs",
+      "old": "// TODO: choose cache backend",
+      "new": "let client = redis::Client::open(cfg.redis_url)?;",
+      "unified": "@@ -1,1 +1,1 @@\n-// TODO: choose cache backend\n+let client = redis::Client::open(cfg.redis_url)?;",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "1b2d44aa",
+      "branch": "feature/cache"
+    },
+    "episode_id": "13a96ab757242797f8dd50af223260bbb9267d27980e7c1ee7dcaea74f169a85"
+  },
+  {
+    "node": "binding",
+    "from": "decision:ws-reject-002:3",
+    "to": "episode:13a96ab757242797f8dd50af223260bbb9267d27980e7c1ee7dcaea74f169a85",
+    "relation": "produced",
+    "prov": {
+      "used_session": "ws-reject-002",
+      "used_decision": "decision:ws-reject-002:3",
+      "was_generated_by_session": "ws-reject-002",
+      "t_use": "2026-06-22T15:11:30Z",
+      "t_gen": "2026-06-22T15:11:40Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__windsurf__v1__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/nodes__windsurf__v1__tool_failure.snap
new file mode 100644
index 0000000..e9c6f75
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__windsurf__v1__tool_failure.snap
@@ -0,0 +1,5 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__zed__v1__ban.snap b/crates/memscribe-testkit/tests/snapshots/nodes__zed__v1__ban.snap
new file mode 100644
index 0000000..91d5460
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__zed__v1__ban.snap
@@ -0,0 +1,61 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "zed-thread-003",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "We will never add a dependency on left-pad. Use the std library instead.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "Use",
+        "offset": 44
+      },
+      {
+        "rule_id": "ban.negated_use",
+        "category": "ban",
+        "matched_text": "never add",
+        "offset": 8
+      },
+      {
+        "rule_id": "imperative.must_always_never",
+        "category": "imperative",
+        "matched_text": "never",
+        "offset": 8
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/zed/v1/ban.jsonl",
+        "byte_offset": 193,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Use the std library instead.",
+    "considered_options": [
+      {
+        "text": "the",
+        "chosen": true
+      }
+    ],
+    "is_ban": true,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__zed__v1__happy_path_decision_then_edits.snap b/crates/memscribe-testkit/tests/snapshots/nodes__zed__v1__happy_path_decision_then_edits.snap
new file mode 100644
index 0000000..c95d5df
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__zed__v1__happy_path_decision_then_edits.snap
@@ -0,0 +1,135 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "zed-thread-001",
+    "turn_range": {
+      "start": 1,
+      "end": 2
+    },
+    "text": "Let's use Postgres instead of MySQL for the persistence layer.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "use",
+        "offset": 6
+      },
+      {
+        "rule_id": "rejection.instead_of",
+        "category": "rejection",
+        "matched_text": "instead of",
+        "offset": 19
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/zed/v1/happy_path_decision_then_edits.jsonl",
+        "byte_offset": 189,
+        "line_no": 2
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Let's use Postgres instead of MySQL for the persistence layer.",
+    "considered_options": [
+      {
+        "text": "MySQL",
+        "chosen": false
+      },
+      {
+        "text": "Postgres",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 1,
+      "end": 2
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/db/config.rs",
+    "diff": {
+      "path": "src/db/config.rs",
+      "old": "let url = \"mysql://localhost/orbit\";",
+      "new": "let url = \"postgres://localhost/orbit\";",
+      "unified": "@@ -1 +1 @@\n-let url = \"mysql://localhost/orbit\";\n+let url = \"postgres://localhost/orbit\";",
+      "added_lines": 1,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "a1b2c3d4e5f6",
+      "branch": "main"
+    },
+    "episode_id": "5bff86a74314fa74565a056cc22a2761228a44545dce893b102ac138d80f4e07"
+  },
+  {
+    "node": "episode",
+    "path": "src/db/schema.rs",
+    "diff": {
+      "path": "src/db/schema.rs",
+      "old": "// MySQL dialect\n",
+      "new": "// Postgres dialect\nuse sqlx::postgres::PgPool;\n",
+      "unified": null,
+      "added_lines": 2,
+      "removed_lines": 1
+    },
+    "git": {
+      "sha": "a1b2c3d4e5f6",
+      "branch": "main"
+    },
+    "episode_id": "599b265bbcf4c0163fbc1f116146bbfbdef58c4a77f1cdcee2fdaf5d2c3ac1ec"
+  },
+  {
+    "node": "binding",
+    "from": "decision:zed-thread-001:1",
+    "to": "episode:599b265bbcf4c0163fbc1f116146bbfbdef58c4a77f1cdcee2fdaf5d2c3ac1ec",
+    "relation": "produced",
+    "prov": {
+      "used_session": "zed-thread-001",
+      "used_decision": "decision:zed-thread-001:1",
+      "was_generated_by_session": "zed-thread-001",
+      "t_use": "2026-06-22T10:00:05Z",
+      "t_gen": "2026-06-22T10:00:09Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  },
+  {
+    "node": "binding",
+    "from": "decision:zed-thread-001:1",
+    "to": "episode:5bff86a74314fa74565a056cc22a2761228a44545dce893b102ac138d80f4e07",
+    "relation": "produced",
+    "prov": {
+      "used_session": "zed-thread-001",
+      "used_decision": "decision:zed-thread-001:1",
+      "was_generated_by_session": "zed-thread-001",
+      "t_use": "2026-06-22T10:00:05Z",
+      "t_gen": "2026-06-22T10:00:09Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": {
+      "support": 0.5,
+      "confidence": 0.5,
+      "lift": 1.0,
+      "phi": 0.0,
+      "p": 1.0
+    }
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__zed__v1__rejected_alternative.snap b/crates/memscribe-testkit/tests/snapshots/nodes__zed__v1__rejected_alternative.snap
new file mode 100644
index 0000000..838bfec
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__zed__v1__rejected_alternative.snap
@@ -0,0 +1,81 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[
+  {
+    "node": "conversation",
+    "session_id": "zed-thread-002",
+    "turn_range": {
+      "start": 3,
+      "end": 4
+    },
+    "text": "Agreed, go with the in-process LRU.",
+    "markers": [
+      {
+        "rule_id": "decision_verb.use",
+        "category": "decision_verb",
+        "matched_text": "go with",
+        "offset": 8
+      }
+    ],
+    "fact_status": "observed",
+    "provenance": [
+      {
+        "file": "fixtures/zed/v1/rejected_alternative.jsonl",
+        "byte_offset": 660,
+        "line_no": 4
+      }
+    ]
+  },
+  {
+    "node": "decision",
+    "epitome": "Agreed, go with the in-process LRU.",
+    "considered_options": [
+      {
+        "text": "the",
+        "chosen": true
+      }
+    ],
+    "is_ban": false,
+    "superseded_by": null,
+    "confirmation": null,
+    "source_span": {
+      "start": 3,
+      "end": 4
+    },
+    "fact_status": "observed"
+  },
+  {
+    "node": "episode",
+    "path": "src/cache/lru.rs",
+    "diff": {
+      "path": "src/cache/lru.rs",
+      "old": "",
+      "new": "pub struct Lru { cap: usize }\n",
+      "unified": "@@ -0,0 +1 @@\n+pub struct Lru { cap: usize }",
+      "added_lines": 1,
+      "removed_lines": 0
+    },
+    "git": {
+      "sha": "f6e5d4c3b2a1",
+      "branch": "feature/caching"
+    },
+    "episode_id": "85b3f3ce6b9be9a870c5a8eace5734be2f128a4dea809c91faa590a31a7a3d5b"
+  },
+  {
+    "node": "binding",
+    "from": "decision:zed-thread-002:3",
+    "to": "episode:85b3f3ce6b9be9a870c5a8eace5734be2f128a4dea809c91faa590a31a7a3d5b",
+    "relation": "produced",
+    "prov": {
+      "used_session": "zed-thread-002",
+      "used_decision": "decision:zed-thread-002:3",
+      "was_generated_by_session": "zed-thread-002",
+      "t_use": "2026-06-22T11:00:40Z",
+      "t_gen": "2026-06-22T11:00:46Z"
+    },
+    "fact_status": "deterministically_derived",
+    "correlation": null
+  }
+]
diff --git a/crates/memscribe-testkit/tests/snapshots/nodes__zed__v1__tool_failure.snap b/crates/memscribe-testkit/tests/snapshots/nodes__zed__v1__tool_failure.snap
new file mode 100644
index 0000000..e9c6f75
--- /dev/null
+++ b/crates/memscribe-testkit/tests/snapshots/nodes__zed__v1__tool_failure.snap
@@ -0,0 +1,5 @@
+---
+source: crates/memscribe-testkit/tests/golden.rs
+expression: nodes
+---
+[]
diff --git a/deny.toml b/deny.toml
new file mode 100644
index 0000000..cc3cb53
--- /dev/null
+++ b/deny.toml
@@ -0,0 +1,31 @@
+# cargo-deny: license + advisory gate (whitepaper §2, §8.9).
+[advisories]
+version = 2
+yanked = "deny"
+
+[licenses]
+version = 2
+# Memscribe is dual MIT OR Apache-2.0; permit the usual permissive set its
+# dependency tree pulls in.
+allow = [
+    "MIT",
+    "Apache-2.0",
+    "Apache-2.0 WITH LLVM-exception",
+    "BSD-2-Clause",
+    "BSD-3-Clause",
+    "ISC",
+    "Unicode-3.0",
+    "Unicode-DFS-2016",
+    "Zlib",
+    "MPL-2.0",
+    "CC0-1.0",
+]
+confidence-threshold = 0.9
+
+[bans]
+multiple-versions = "warn"
+wildcards = "allow"
+
+[sources]
+unknown-registry = "deny"
+unknown-git = "deny"
diff --git a/fixtures/claude_code/2.0/ban.jsonl b/fixtures/claude_code/2.0/ban.jsonl
new file mode 100644
index 0000000..0f077cb
--- /dev/null
+++ b/fixtures/claude_code/2.0/ban.jsonl
@@ -0,0 +1,2 @@
+{"type":"user","uuid":"bbbbbbb1-0000-4000-8000-000000000001","parentUuid":null,"timestamp":"2026-06-22T12:15:00.000Z","sessionId":"sess-ban-003","cwd":"/Users/dev/acme/core","gitBranch":"main","gitSha":"deadbeefdeadbeefdeadbeefdeadbeefdeadbeef","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":"We will never add a dependency on left-pad in this repo. Use the std string padding instead."}}
+{"type":"assistant","uuid":"bbbbbbb2-0000-4000-8000-000000000002","parentUuid":"bbbbbbb1-0000-4000-8000-000000000001","timestamp":"2026-06-22T12:15:03.000Z","sessionId":"sess-ban-003","cwd":"/Users/dev/acme/core","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":210,"output_tokens":48,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"content":[{"type":"text","text":"Noted. I'll keep padding on the standard library and avoid third-party micro-dependencies."}]}}
diff --git a/fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl b/fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl
new file mode 100644
index 0000000..4d76ef9
--- /dev/null
+++ b/fixtures/claude_code/2.0/happy_path_decision_then_edits.jsonl
@@ -0,0 +1,5 @@
+{"type":"user","uuid":"11111111-1111-4111-8111-111111111111","parentUuid":null,"timestamp":"2026-06-22T10:00:00.000Z","sessionId":"sess-happy-001","cwd":"/Users/dev/acme/api","gitBranch":"main","gitSha":"a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":"Let's use Postgres instead of MySQL for the orders service."}}
+{"type":"assistant","uuid":"22222222-2222-4222-8222-222222222222","parentUuid":"11111111-1111-4111-8111-111111111111","timestamp":"2026-06-22T10:00:03.120Z","sessionId":"sess-happy-001","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":482,"output_tokens":96,"cache_read_input_tokens":120,"cache_creation_input_tokens":40},"content":[{"type":"text","text":"Switching the orders service to Postgres. I'll update the connection config and the schema module."},{"type":"tool_use","id":"toolu_01edit_config","name":"Edit","input":{"file_path":"/Users/dev/acme/api/src/db/config.rs","old_string":"mysql://localhost:3306/orders","new_string":"postgres://localhost:5432/orders"}}]}}
+{"type":"user","uuid":"33333333-3333-4333-8333-333333333333","parentUuid":"22222222-2222-4222-8222-222222222222","timestamp":"2026-06-22T10:00:03.480Z","sessionId":"sess-happy-001","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_01edit_config","content":"The file /Users/dev/acme/api/src/db/config.rs has been updated.","is_error":false}]},"toolUseResult":{"filePath":"/Users/dev/acme/api/src/db/config.rs","oldString":"mysql://localhost:3306/orders","newString":"postgres://localhost:5432/orders","structuredPatch":[{"oldStart":12,"oldLines":1,"newStart":12,"newLines":1,"lines":["-    const DSN: &str = \"mysql://localhost:3306/orders\";","+    const DSN: &str = \"postgres://localhost:5432/orders\";"]}],"originalFileContents":"// db config\n"}}
+{"type":"assistant","uuid":"44444444-4444-4444-8444-444444444444","parentUuid":"33333333-3333-4333-8333-333333333333","timestamp":"2026-06-22T10:00:06.900Z","sessionId":"sess-happy-001","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":640,"output_tokens":140,"cache_read_input_tokens":480,"cache_creation_input_tokens":0},"content":[{"type":"text","text":"Now updating the schema driver import."},{"type":"tool_use","id":"toolu_02edit_schema","name":"Edit","input":{"file_path":"/Users/dev/acme/api/src/db/schema.rs","old_string":"use mysql::Pool;","new_string":"use postgres::Client;"}}]}}
+{"type":"user","uuid":"55555555-5555-4555-8555-555555555555","parentUuid":"44444444-4444-4444-8444-444444444444","timestamp":"2026-06-22T10:00:07.210Z","sessionId":"sess-happy-001","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_02edit_schema","content":"The file /Users/dev/acme/api/src/db/schema.rs has been updated.","is_error":false}]},"toolUseResult":{"filePath":"/Users/dev/acme/api/src/db/schema.rs","oldString":"use mysql::Pool;","newString":"use postgres::Client;","structuredPatch":[{"oldStart":3,"oldLines":1,"newStart":3,"newLines":1,"lines":["-use mysql::Pool;","+use postgres::Client;"]}],"originalFileContents":"// schema\nuse mysql::Pool;\n"}}
diff --git a/fixtures/claude_code/2.0/interleaved_arcs.jsonl b/fixtures/claude_code/2.0/interleaved_arcs.jsonl
new file mode 100644
index 0000000..adfb84b
--- /dev/null
+++ b/fixtures/claude_code/2.0/interleaved_arcs.jsonl
@@ -0,0 +1,8 @@
+{"type":"user","uuid":"d1aaaaa1-0000-4000-8000-000000000001","parentUuid":null,"timestamp":"2026-06-22T14:00:00.000Z","sessionId":"sess-interleaved-005","cwd":"/Users/dev/acme/api","gitBranch":"main","gitSha":"1111111122222222333333334444444455555555","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":"Let's use argon2 for password hashing in the auth module."}}
+{"type":"assistant","uuid":"d1aaaaa2-0000-4000-8000-000000000002","parentUuid":"d1aaaaa1-0000-4000-8000-000000000001","timestamp":"2026-06-22T14:00:03.000Z","sessionId":"sess-interleaved-005","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":300,"output_tokens":70,"cache_read_input_tokens":0,"cache_creation_input_tokens":120},"content":[{"type":"text","text":"Switching the hasher to argon2 in the auth module."},{"type":"tool_use","id":"toolu_arc_a_auth","name":"Edit","input":{"file_path":"/Users/dev/acme/api/src/auth.rs","old_string":"use bcrypt::hash;","new_string":"use argon2::hash;"}}]}}
+{"type":"user","uuid":"d1aaaaa3-0000-4000-8000-000000000003","parentUuid":"d1aaaaa2-0000-4000-8000-000000000002","timestamp":"2026-06-22T14:00:03.500Z","sessionId":"sess-interleaved-005","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_arc_a_auth","content":"The file /Users/dev/acme/api/src/auth.rs has been updated.","is_error":false}]},"toolUseResult":{"filePath":"/Users/dev/acme/api/src/auth.rs","oldString":"use bcrypt::hash;","newString":"use argon2::hash;","structuredPatch":[{"oldStart":1,"oldLines":1,"newStart":1,"newLines":1,"lines":["-use bcrypt::hash;","+use argon2::hash;"]}],"originalFileContents":"use bcrypt::hash;\n"}}
+{"type":"user","uuid":"d1aaaaa4-0000-4000-8000-000000000004","parentUuid":"d1aaaaa3-0000-4000-8000-000000000003","timestamp":"2026-06-22T14:01:00.000Z","sessionId":"sess-interleaved-005","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":"Actually, let's switch to scrypt for the auth module and tune the cost in config."}}
+{"type":"assistant","uuid":"d1aaaaa5-0000-4000-8000-000000000005","parentUuid":"d1aaaaa4-0000-4000-8000-000000000004","timestamp":"2026-06-22T14:01:04.000Z","sessionId":"sess-interleaved-005","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":360,"output_tokens":90,"cache_read_input_tokens":300,"cache_creation_input_tokens":0},"content":[{"type":"text","text":"Re-pointing the auth hasher to scrypt."},{"type":"tool_use","id":"toolu_arc_b_auth","name":"Edit","input":{"file_path":"/Users/dev/acme/api/src/auth.rs","old_string":"use argon2::hash;","new_string":"use scrypt::hash;"}}]}}
+{"type":"user","uuid":"d1aaaaa6-0000-4000-8000-000000000006","parentUuid":"d1aaaaa5-0000-4000-8000-000000000005","timestamp":"2026-06-22T14:01:04.500Z","sessionId":"sess-interleaved-005","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_arc_b_auth","content":"The file /Users/dev/acme/api/src/auth.rs has been updated.","is_error":false}]},"toolUseResult":{"filePath":"/Users/dev/acme/api/src/auth.rs","oldString":"use argon2::hash;","newString":"use scrypt::hash;","structuredPatch":[{"oldStart":1,"oldLines":1,"newStart":1,"newLines":1,"lines":["-use argon2::hash;","+use scrypt::hash;"]}],"originalFileContents":"use argon2::hash;\n"}}
+{"type":"assistant","uuid":"d1aaaaa7-0000-4000-8000-000000000007","parentUuid":"d1aaaaa6-0000-4000-8000-000000000006","timestamp":"2026-06-22T14:01:08.000Z","sessionId":"sess-interleaved-005","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":410,"output_tokens":85,"cache_read_input_tokens":360,"cache_creation_input_tokens":0},"content":[{"type":"text","text":"Tuning the scrypt cost parameter in config."},{"type":"tool_use","id":"toolu_arc_b_config","name":"Edit","input":{"file_path":"/Users/dev/acme/api/src/config.rs","old_string":"const HASH_COST: u32 = 10;","new_string":"const HASH_COST: u32 = 15;"}}]}}
+{"type":"user","uuid":"d1aaaaa8-0000-4000-8000-000000000008","parentUuid":"d1aaaaa7-0000-4000-8000-000000000007","timestamp":"2026-06-22T14:01:08.500Z","sessionId":"sess-interleaved-005","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_arc_b_config","content":"The file /Users/dev/acme/api/src/config.rs has been updated.","is_error":false}]},"toolUseResult":{"filePath":"/Users/dev/acme/api/src/config.rs","oldString":"const HASH_COST: u32 = 10;","newString":"const HASH_COST: u32 = 15;","structuredPatch":[{"oldStart":4,"oldLines":1,"newStart":4,"newLines":1,"lines":["-const HASH_COST: u32 = 10;","+const HASH_COST: u32 = 15;"]}],"originalFileContents":"// config\nconst HASH_COST: u32 = 10;\n"}}
diff --git a/fixtures/claude_code/2.0/multi_edit_single_commit.jsonl b/fixtures/claude_code/2.0/multi_edit_single_commit.jsonl
new file mode 100644
index 0000000..5e2ebdf
--- /dev/null
+++ b/fixtures/claude_code/2.0/multi_edit_single_commit.jsonl
@@ -0,0 +1,7 @@
+{"type":"user","uuid":"d2bbbbb1-0000-4000-8000-000000000001","parentUuid":null,"timestamp":"2026-06-22T15:00:00.000Z","sessionId":"sess-multiedit-006","cwd":"/Users/dev/acme/api","gitBranch":"main","gitSha":"aaaa1111bbbb2222cccc3333dddd4444eeee5555","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":"Let's migrate to anyhow for error handling across the request, handler, and router modules."}}
+{"type":"assistant","uuid":"d2bbbbb2-0000-4000-8000-000000000002","parentUuid":"d2bbbbb1-0000-4000-8000-000000000001","timestamp":"2026-06-22T15:00:03.000Z","sessionId":"sess-multiedit-006","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":420,"output_tokens":110,"cache_read_input_tokens":0,"cache_creation_input_tokens":180},"content":[{"type":"text","text":"I'll convert each module to anyhow::Result. Starting with request.rs."},{"type":"tool_use","id":"toolu_multi_request","name":"Edit","input":{"file_path":"/Users/dev/acme/api/src/request.rs","old_string":"use crate::error::AppError;","new_string":"use anyhow::Result;"}}]}}
+{"type":"user","uuid":"d2bbbbb3-0000-4000-8000-000000000003","parentUuid":"d2bbbbb2-0000-4000-8000-000000000002","timestamp":"2026-06-22T15:00:03.400Z","sessionId":"sess-multiedit-006","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_multi_request","content":"The file /Users/dev/acme/api/src/request.rs has been updated.","is_error":false}]},"toolUseResult":{"filePath":"/Users/dev/acme/api/src/request.rs","oldString":"use crate::error::AppError;","newString":"use anyhow::Result;","structuredPatch":[{"oldStart":2,"oldLines":1,"newStart":2,"newLines":1,"lines":["-use crate::error::AppError;","+use anyhow::Result;"]}],"originalFileContents":"// request\nuse crate::error::AppError;\n"}}
+{"type":"assistant","uuid":"d2bbbbb4-0000-4000-8000-000000000004","parentUuid":"d2bbbbb3-0000-4000-8000-000000000003","timestamp":"2026-06-22T15:00:06.000Z","sessionId":"sess-multiedit-006","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":480,"output_tokens":95,"cache_read_input_tokens":420,"cache_creation_input_tokens":0},"content":[{"type":"text","text":"Now handler.rs."},{"type":"tool_use","id":"toolu_multi_handler","name":"Edit","input":{"file_path":"/Users/dev/acme/api/src/handler.rs","old_string":"use crate::error::AppError;","new_string":"use anyhow::Result;"}}]}}
+{"type":"user","uuid":"d2bbbbb5-0000-4000-8000-000000000005","parentUuid":"d2bbbbb4-0000-4000-8000-000000000004","timestamp":"2026-06-22T15:00:06.400Z","sessionId":"sess-multiedit-006","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_multi_handler","content":"The file /Users/dev/acme/api/src/handler.rs has been updated.","is_error":false}]},"toolUseResult":{"filePath":"/Users/dev/acme/api/src/handler.rs","oldString":"use crate::error::AppError;","newString":"use anyhow::Result;","structuredPatch":[{"oldStart":2,"oldLines":1,"newStart":2,"newLines":1,"lines":["-use crate::error::AppError;","+use anyhow::Result;"]}],"originalFileContents":"// handler\nuse crate::error::AppError;\n"}}
+{"type":"assistant","uuid":"d2bbbbb6-0000-4000-8000-000000000006","parentUuid":"d2bbbbb5-0000-4000-8000-000000000005","timestamp":"2026-06-22T15:00:09.000Z","sessionId":"sess-multiedit-006","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":540,"output_tokens":95,"cache_read_input_tokens":480,"cache_creation_input_tokens":0},"content":[{"type":"text","text":"And finally router.rs."},{"type":"tool_use","id":"toolu_multi_router","name":"Edit","input":{"file_path":"/Users/dev/acme/api/src/router.rs","old_string":"use crate::error::AppError;","new_string":"use anyhow::Result;"}}]}}
+{"type":"user","uuid":"d2bbbbb7-0000-4000-8000-000000000007","parentUuid":"d2bbbbb6-0000-4000-8000-000000000006","timestamp":"2026-06-22T15:00:09.400Z","sessionId":"sess-multiedit-006","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_multi_router","content":"The file /Users/dev/acme/api/src/router.rs has been updated.","is_error":false}]},"toolUseResult":{"filePath":"/Users/dev/acme/api/src/router.rs","oldString":"use crate::error::AppError;","newString":"use anyhow::Result;","structuredPatch":[{"oldStart":2,"oldLines":1,"newStart":2,"newLines":1,"lines":["-use crate::error::AppError;","+use anyhow::Result;"]}],"originalFileContents":"// router\nuse crate::error::AppError;\n"}}
diff --git a/fixtures/claude_code/2.0/no_commitment_marker.jsonl b/fixtures/claude_code/2.0/no_commitment_marker.jsonl
new file mode 100644
index 0000000..e69f9dd
--- /dev/null
+++ b/fixtures/claude_code/2.0/no_commitment_marker.jsonl
@@ -0,0 +1,3 @@
+{"type":"user","uuid":"d5fffff1-0000-4000-8000-000000000001","parentUuid":null,"timestamp":"2026-06-22T18:00:00.000Z","sessionId":"sess-nomarker-009","cwd":"/Users/dev/acme/api","gitBranch":"main","gitSha":"0a0b0c0d0e0f10111213141516171819202122ab","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":"Thanks, that looks good to me. Please tidy the trailing whitespace in the handler."}}
+{"type":"assistant","uuid":"d5fffff2-0000-4000-8000-000000000002","parentUuid":"d5fffff1-0000-4000-8000-000000000001","timestamp":"2026-06-22T18:00:03.000Z","sessionId":"sess-nomarker-009","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":210,"output_tokens":45,"cache_read_input_tokens":0,"cache_creation_input_tokens":70},"content":[{"type":"text","text":"Tidying the trailing whitespace in the handler now."},{"type":"tool_use","id":"toolu_nomarker_fmt","name":"Edit","input":{"file_path":"/Users/dev/acme/api/src/handler.rs","old_string":"    let body = read(); ","new_string":"    let body = read();"}}]}}
+{"type":"user","uuid":"d5fffff3-0000-4000-8000-000000000003","parentUuid":"d5fffff2-0000-4000-8000-000000000002","timestamp":"2026-06-22T18:00:03.400Z","sessionId":"sess-nomarker-009","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_nomarker_fmt","content":"The file /Users/dev/acme/api/src/handler.rs has been updated.","is_error":false}]},"toolUseResult":{"filePath":"/Users/dev/acme/api/src/handler.rs","oldString":"    let body = read(); ","newString":"    let body = read();","structuredPatch":[{"oldStart":7,"oldLines":1,"newStart":7,"newLines":1,"lines":["-    let body = read(); ","+    let body = read();"]}],"originalFileContents":"// handler\n    let body = read(); \n"}}
diff --git a/fixtures/claude_code/2.0/rejected_alternative.jsonl b/fixtures/claude_code/2.0/rejected_alternative.jsonl
new file mode 100644
index 0000000..74c18f7
--- /dev/null
+++ b/fixtures/claude_code/2.0/rejected_alternative.jsonl
@@ -0,0 +1,3 @@
+{"type":"user","uuid":"aaaaaaa1-0000-4000-8000-000000000001","parentUuid":null,"timestamp":"2026-06-22T11:30:00.000Z","sessionId":"sess-reject-002","cwd":"/Users/dev/acme/web","gitBranch":"feat/payments","gitSha":"0f1e2d3c4b5a69788796a5b4c3d2e1f00f1e2d3c","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":"We'll go with Stripe instead of PayPal for the checkout flow."}}
+{"type":"assistant","uuid":"aaaaaaa2-0000-4000-8000-000000000002","parentUuid":"aaaaaaa1-0000-4000-8000-000000000001","timestamp":"2026-06-22T11:30:04.500Z","sessionId":"sess-reject-002","cwd":"/Users/dev/acme/web","gitBranch":"feat/payments","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":310,"output_tokens":72,"cache_read_input_tokens":0,"cache_creation_input_tokens":210},"content":[{"type":"text","text":"Understood. I'll wire the Stripe client into the checkout handler."},{"type":"tool_use","id":"toolu_stripe_edit","name":"Write","input":{"file_path":"/Users/dev/acme/web/src/checkout/provider.ts","content":"export { stripe as paymentProvider } from './stripe';\n"}}]}}
+{"type":"user","uuid":"aaaaaaa3-0000-4000-8000-000000000003","parentUuid":"aaaaaaa2-0000-4000-8000-000000000002","timestamp":"2026-06-22T11:30:04.900Z","sessionId":"sess-reject-002","cwd":"/Users/dev/acme/web","gitBranch":"feat/payments","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_stripe_edit","content":"File created successfully at: /Users/dev/acme/web/src/checkout/provider.ts","is_error":false}]},"toolUseResult":{"filePath":"/Users/dev/acme/web/src/checkout/provider.ts","oldString":"","newString":"export { stripe as paymentProvider } from './stripe';\n","structuredPatch":[{"oldStart":0,"oldLines":0,"newStart":1,"newLines":1,"lines":["+export { stripe as paymentProvider } from './stripe';"]}],"originalFileContents":""}}
diff --git a/fixtures/claude_code/2.0/rewind_compaction.jsonl b/fixtures/claude_code/2.0/rewind_compaction.jsonl
new file mode 100644
index 0000000..f697e36
--- /dev/null
+++ b/fixtures/claude_code/2.0/rewind_compaction.jsonl
@@ -0,0 +1,7 @@
+{"type":"user","uuid":"d3ccccc1-0000-4000-8000-000000000001","parentUuid":null,"timestamp":"2026-06-22T16:00:00.000Z","sessionId":"sess-rewind-007","cwd":"/Users/dev/acme/api","gitBranch":"main","gitSha":"9999888877776666555544443333222211110000","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":"Let's use rustls for the TLS layer in the gateway."}}
+{"type":"assistant","uuid":"d3ccccc2-0000-4000-8000-000000000002","parentUuid":"d3ccccc1-0000-4000-8000-000000000001","timestamp":"2026-06-22T16:00:03.000Z","sessionId":"sess-rewind-007","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":300,"output_tokens":60,"cache_read_input_tokens":0,"cache_creation_input_tokens":110},"content":[{"type":"text","text":"Pointing the gateway TLS layer at rustls."},{"type":"tool_use","id":"toolu_rewind_rustls","name":"Edit","input":{"file_path":"/Users/dev/acme/api/src/gateway/tls.rs","old_string":"use native_tls::TlsConnector;","new_string":"use rustls::ClientConfig;"}}]}}
+{"type":"user","uuid":"d3ccccc3-0000-4000-8000-000000000003","parentUuid":"d3ccccc2-0000-4000-8000-000000000002","timestamp":"2026-06-22T16:00:03.400Z","sessionId":"sess-rewind-007","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_rewind_rustls","content":"The file /Users/dev/acme/api/src/gateway/tls.rs has been updated.","is_error":false}]},"toolUseResult":{"filePath":"/Users/dev/acme/api/src/gateway/tls.rs","oldString":"use native_tls::TlsConnector;","newString":"use rustls::ClientConfig;","structuredPatch":[{"oldStart":1,"oldLines":1,"newStart":1,"newLines":1,"lines":["-use native_tls::TlsConnector;","+use rustls::ClientConfig;"]}],"originalFileContents":"use native_tls::TlsConnector;\n"}}
+{"type":"summary","summary":"The user chose rustls for the gateway TLS layer and the connector import was updated. Conversation compacted to free context.","leafUuid":"d3ccccc3-0000-4000-8000-000000000003"}
+{"type":"user","uuid":"d3ccccc5-0000-4000-8000-000000000005","parentUuid":"d3ccccc3-0000-4000-8000-000000000003","timestamp":"2026-06-22T16:05:00.000Z","sessionId":"sess-rewind-007","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":"On reflection, let's switch to native-tls after all for the gateway to match the rest of the fleet."}}
+{"type":"assistant","uuid":"d3ccccc6-0000-4000-8000-000000000006","parentUuid":"d3ccccc5-0000-4000-8000-000000000005","timestamp":"2026-06-22T16:05:04.000Z","sessionId":"sess-rewind-007","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":350,"output_tokens":70,"cache_read_input_tokens":0,"cache_creation_input_tokens":140},"content":[{"type":"text","text":"Reverting the gateway TLS layer back to native-tls."},{"type":"tool_use","id":"toolu_rewind_native","name":"Edit","input":{"file_path":"/Users/dev/acme/api/src/gateway/tls.rs","old_string":"use rustls::ClientConfig;","new_string":"use native_tls::TlsConnector;"}}]}}
+{"type":"user","uuid":"d3ccccc7-0000-4000-8000-000000000007","parentUuid":"d3ccccc6-0000-4000-8000-000000000006","timestamp":"2026-06-22T16:05:04.400Z","sessionId":"sess-rewind-007","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_rewind_native","content":"The file /Users/dev/acme/api/src/gateway/tls.rs has been updated.","is_error":false}]},"toolUseResult":{"filePath":"/Users/dev/acme/api/src/gateway/tls.rs","oldString":"use rustls::ClientConfig;","newString":"use native_tls::TlsConnector;","structuredPatch":[{"oldStart":1,"oldLines":1,"newStart":1,"newLines":1,"lines":["-use rustls::ClientConfig;","+use native_tls::TlsConnector;"]}],"originalFileContents":"use rustls::ClientConfig;\n"}}
diff --git a/fixtures/claude_code/2.0/subagent_thread.jsonl b/fixtures/claude_code/2.0/subagent_thread.jsonl
new file mode 100644
index 0000000..fb561bc
--- /dev/null
+++ b/fixtures/claude_code/2.0/subagent_thread.jsonl
@@ -0,0 +1,6 @@
+{"type":"user","uuid":"d4ddddd1-0000-4000-8000-000000000001","parentUuid":null,"timestamp":"2026-06-22T17:00:00.000Z","sessionId":"sess-main-008","cwd":"/Users/dev/acme/api","gitBranch":"main","gitSha":"abcdefab12345678abcdefab12345678abcdefab","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":"Let's use a dedicated search subagent to migrate the indexer to tantivy."}}
+{"type":"assistant","uuid":"d4ddddd2-0000-4000-8000-000000000002","parentUuid":"d4ddddd1-0000-4000-8000-000000000001","timestamp":"2026-06-22T17:00:03.000Z","sessionId":"sess-main-008","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":260,"output_tokens":55,"cache_read_input_tokens":0,"cache_creation_input_tokens":90},"content":[{"type":"text","text":"Dispatching a search subagent to handle the indexer migration."},{"type":"tool_use","id":"toolu_dispatch_subagent","name":"Task","input":{"description":"Migrate indexer to tantivy","subagent_type":"search"}}]}}
+{"type":"user","uuid":"d4eeeee1-0000-4000-8000-000000000011","parentUuid":null,"timestamp":"2026-06-22T17:00:05.000Z","sessionId":"sess-subagent-008b","cwd":"/Users/dev/acme/api","gitBranch":"main","gitSha":"abcdefab12345678abcdefab12345678abcdefab","version":"2.0.5","isSidechain":true,"message":{"role":"user","content":"Let's adopt tantivy as the search index backend in the indexer module."}}
+{"type":"assistant","uuid":"d4eeeee2-0000-4000-8000-000000000012","parentUuid":"d4eeeee1-0000-4000-8000-000000000011","timestamp":"2026-06-22T17:00:08.000Z","sessionId":"sess-subagent-008b","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":true,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":280,"output_tokens":65,"cache_read_input_tokens":0,"cache_creation_input_tokens":100},"content":[{"type":"text","text":"Swapping the indexer backend to tantivy."},{"type":"tool_use","id":"toolu_sub_indexer","name":"Edit","input":{"file_path":"/Users/dev/acme/api/src/indexer.rs","old_string":"use tantivy_lite::Index;","new_string":"use tantivy::Index;"}}]}}
+{"type":"user","uuid":"d4eeeee3-0000-4000-8000-000000000013","parentUuid":"d4eeeee2-0000-4000-8000-000000000012","timestamp":"2026-06-22T17:00:08.400Z","sessionId":"sess-subagent-008b","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":true,"message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_sub_indexer","content":"The file /Users/dev/acme/api/src/indexer.rs has been updated.","is_error":false}]},"toolUseResult":{"filePath":"/Users/dev/acme/api/src/indexer.rs","oldString":"use tantivy_lite::Index;","newString":"use tantivy::Index;","structuredPatch":[{"oldStart":1,"oldLines":1,"newStart":1,"newLines":1,"lines":["-use tantivy_lite::Index;","+use tantivy::Index;"]}],"originalFileContents":"use tantivy_lite::Index;\n"}}
+{"type":"assistant","uuid":"d4ddddd3-0000-4000-8000-000000000003","parentUuid":"d4ddddd2-0000-4000-8000-000000000002","timestamp":"2026-06-22T17:01:00.000Z","sessionId":"sess-main-008","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":320,"output_tokens":40,"cache_read_input_tokens":260,"cache_creation_input_tokens":0},"content":[{"type":"text","text":"The search subagent finished the tantivy migration in the indexer."}]}}
diff --git a/fixtures/claude_code/2.0/tool_failure.jsonl b/fixtures/claude_code/2.0/tool_failure.jsonl
new file mode 100644
index 0000000..c816ebb
--- /dev/null
+++ b/fixtures/claude_code/2.0/tool_failure.jsonl
@@ -0,0 +1,3 @@
+{"type":"user","uuid":"ccccccc1-0000-4000-8000-000000000001","parentUuid":null,"timestamp":"2026-06-22T13:45:00.000Z","sessionId":"sess-fail-004","cwd":"/Users/dev/acme/api","gitBranch":"main","gitSha":"feedface00000000feedface00000000feedface","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":"Let's use a connection pool in the orders repository."}}
+{"type":"assistant","uuid":"ccccccc2-0000-4000-8000-000000000002","parentUuid":"ccccccc1-0000-4000-8000-000000000001","timestamp":"2026-06-22T13:45:03.000Z","sessionId":"sess-fail-004","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"assistant","model":"claude-opus-4-8","usage":{"input_tokens":270,"output_tokens":80,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"content":[{"type":"text","text":"I'll add the pool to the repository module."},{"type":"tool_use","id":"toolu_fail_edit","name":"Edit","input":{"file_path":"/Users/dev/acme/api/src/orders/repo.rs","old_string":"let conn = Connection::new();","new_string":"let conn = pool.get()?;"}}]}}
+{"type":"user","uuid":"ccccccc3-0000-4000-8000-000000000003","parentUuid":"ccccccc2-0000-4000-8000-000000000002","timestamp":"2026-06-22T13:45:03.400Z","sessionId":"sess-fail-004","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.0.5","isSidechain":false,"message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_fail_edit","content":"Error: String to replace not found in file. The string 'let conn = Connection::new();' does not appear in /Users/dev/acme/api/src/orders/repo.rs.","is_error":true}]},"toolUseResult":{"filePath":"/Users/dev/acme/api/src/orders/repo.rs","oldString":"let conn = Connection::new();","newString":"let conn = pool.get()?;","structuredPatch":[{"oldStart":18,"oldLines":1,"newStart":18,"newLines":1,"lines":["-    let conn = Connection::new();","+    let conn = pool.get()?;"]}],"originalFileContents":"// orders repo\n"}}
diff --git a/fixtures/claude_code/2.1/version_tolerance.jsonl b/fixtures/claude_code/2.1/version_tolerance.jsonl
new file mode 100644
index 0000000..f455b56
--- /dev/null
+++ b/fixtures/claude_code/2.1/version_tolerance.jsonl
@@ -0,0 +1,4 @@
+{"type":"user","uuid":"21111111-1111-4111-8111-111111111111","parentUuid":null,"timestamp":"2026-06-22T11:00:00.000Z","sessionId":"sess-v21-001","cwd":"/Users/dev/acme/api","gitBranch":"main","gitHead":"a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0","version":"2.1.0","isSidechain":false,"clientMeta":{"surface":"cli","experiments":["streaming_v2"]},"message":{"role":"user","content":"Let's use Postgres instead of MySQL for the orders service."}}
+{"type":"assistant","uuid":"22222221-2222-4222-8222-222222222222","parentUuid":"21111111-1111-4111-8111-111111111111","timestamp":"2026-06-22T11:00:03.120Z","sessionId":"sess-v21-001","cwd":"/Users/dev/acme/api","gitBranch":"main","version":"2.1.0","message":{"role":"assistant","model":"claude-opus-4-8","stopReason":"tool_use","usage":{"input_tokens":482,"output_tokens":96,"cache_read_input_tokens":120,"cache_creation_input_tokens":40,"server_tool_use":{"web_search_requests":0}},"content":"Switching the orders service to Postgres now."}}
+{"type":"telemetry","uuid":"23333331-3333-4333-8333-333333333333","parentUuid":"22222221-2222-4222-8222-222222222222","timestamp":"2026-06-22T11:00:03.480Z","sessionId":"sess-v21-001","version":"2.1.0","payload":{"kind":"latency","ms":1200}}
+{"type":"user","uuid":"24444441-4444-4444-8444-444444444444","parentUuid":"22222221-2222-4222-8222-222222222222","timestamp":"2026-06-22T11:00:07.210Z","sessionId":"sess-v21-001","version":"2.1.0","message":{"role":"user","content":"Looks good, ship it."}}
diff --git a/fixtures/codex/v1/version_tolerance.jsonl b/fixtures/codex/v1/version_tolerance.jsonl
new file mode 100644
index 0000000..942e6af
--- /dev/null
+++ b/fixtures/codex/v1/version_tolerance.jsonl
@@ -0,0 +1,4 @@
+{"timestamp":"2026-06-22T09:00:00.000Z","type":"session_meta","payload":{"id":"01JZ8codexsess0001legacy","cwd":"/Users/dev/acme-api","git":{"sha":"a1b2c3d4e5f60718293a4b5c6d7e8f9001020304","branch":"main"},"cli_version":"0.1.0","originator":"codex_cli_rs"}}
+{"timestamp":"2026-06-22T09:00:05.120Z","record_type":"input","role":"user","text":"Let's use Postgres instead of MySQL for the persistence layer."}
+{"timestamp":"2026-06-22T09:00:09.880Z","record_type":"output","role":"assistant","text":"Switching to Postgres."}
+{"timestamp":"2026-06-22T09:00:12.400Z","kind":"state","payload":{"phase":"awaiting_input"}}
diff --git a/fixtures/codex/v2/ban.jsonl b/fixtures/codex/v2/ban.jsonl
new file mode 100644
index 0000000..01b2fa6
--- /dev/null
+++ b/fixtures/codex/v2/ban.jsonl
@@ -0,0 +1,6 @@
+{"timestamp":"2026-06-22T16:20:00.000Z","type":"session_meta","payload":{"id":"01JZ8codexsess0003ban","cwd":"/Users/dev/acme-api","git":{"sha":"c3d4e5f60718293a4b5c6d7e8f90010203040506","branch":"main"},"cli_version":"0.6.2","originator":"codex_cli_rs"}}
+{"timestamp":"2026-06-22T16:20:07.330Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"Policy: we will never add a dependency on left-pad. Do not introduce it under any circumstances, even transitively if you can avoid it."}]}}
+{"timestamp":"2026-06-22T16:20:12.770Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Acknowledged. I will never add left-pad as a dependency."}]}}
+{"timestamp":"2026-06-22T16:20:40.900Z","type":"response_item","payload":{"type":"function_call","name":"shell","arguments":"{\"command\":[\"rg\",\"left-pad\",\"package.json\",\"Cargo.toml\"]}","call_id":"call_check_001"}}
+{"timestamp":"2026-06-22T16:20:41.220Z","type":"response_item","payload":{"type":"function_call_output","call_id":"call_check_001","output":{"output":"","metadata":{"exit_code":1}}}}
+{"timestamp":"2026-06-22T16:20:44.000Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Confirmed: left-pad is not present in the manifests."}]}}
diff --git a/fixtures/codex/v2/happy_path_decision_then_edits.jsonl b/fixtures/codex/v2/happy_path_decision_then_edits.jsonl
new file mode 100644
index 0000000..23ed6ed
--- /dev/null
+++ b/fixtures/codex/v2/happy_path_decision_then_edits.jsonl
@@ -0,0 +1,8 @@
+{"timestamp":"2026-06-22T14:00:00.000Z","type":"session_meta","payload":{"id":"01JZ8codexsess0001happypath","cwd":"/Users/dev/acme-api","git":{"sha":"a1b2c3d4e5f60718293a4b5c6d7e8f9001020304","branch":"feat/storage-layer"},"cli_version":"0.6.2","originator":"codex_cli_rs"}}
+{"timestamp":"2026-06-22T14:00:05.120Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"Let's use Postgres instead of MySQL for the persistence layer. Swap the driver and the connection URL."}]}}
+{"timestamp":"2026-06-22T14:00:09.880Z","type":"response_item","payload":{"type":"reasoning","summary":["The user decided on Postgres. I'll update the Cargo deps and the connection string."]}}
+{"timestamp":"2026-06-22T14:00:12.400Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Switching to Postgres. I'll update Cargo.toml and the database module."}]}}
+{"timestamp":"2026-06-22T14:00:18.640Z","type":"response_item","payload":{"type":"function_call","name":"apply_patch","arguments":"{\"input\":\"*** Begin Patch\\n*** Update File: Cargo.toml\\n@@\\n-mysql = \\\"24\\\"\\n+sqlx = { version = \\\"0.7\\\", features = [\\\"postgres\\\", \\\"runtime-tokio\\\"] }\\n*** Update File: src/db.rs\\n@@\\n-let url = \\\"mysql://localhost:3306/acme\\\";\\n-let pool = mysql::Pool::new(url)?;\\n+let url = \\\"postgres://localhost:5432/acme\\\";\\n+let pool = sqlx::PgPool::connect(url).await?;\\n*** End Patch\\n\"}","call_id":"call_apply_patch_001"}}
+{"timestamp":"2026-06-22T14:00:19.010Z","type":"response_item","payload":{"type":"function_call_output","call_id":"call_apply_patch_001","output":{"output":"Applied patch to 2 files: Cargo.toml, src/db.rs","metadata":{"exit_code":0,"duration_ms":42}}}}
+{"timestamp":"2026-06-22T14:00:25.300Z","type":"event_msg","payload":{"type":"token_count","input_tokens":1840,"output_tokens":260}}
+{"timestamp":"2026-06-22T14:00:26.000Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Done. The persistence layer now uses Postgres via sqlx."}]}}
diff --git a/fixtures/codex/v2/interleaved_arcs.jsonl b/fixtures/codex/v2/interleaved_arcs.jsonl
new file mode 100644
index 0000000..fe13013
--- /dev/null
+++ b/fixtures/codex/v2/interleaved_arcs.jsonl
@@ -0,0 +1,11 @@
+{"timestamp":"2026-06-22T18:00:00.000Z","type":"session_meta","payload":{"id":"01JZ8codexsess0005interleaved","cwd":"/Users/dev/acme-api","git":{"sha":"e5f60718293a4b5c6d7e8f900102030405060708","branch":"feat/auth-overhaul"},"cli_version":"0.6.2","originator":"codex_cli_rs"}}
+{"timestamp":"2026-06-22T18:00:06.100Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"Let's use argon2 for password hashing. Update the auth module and the shared crypto helper accordingly."}]}}
+{"timestamp":"2026-06-22T18:00:10.300Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Switching the hasher to argon2 in the auth module and the shared crypto helper."}]}}
+{"timestamp":"2026-06-22T18:00:16.450Z","type":"response_item","payload":{"type":"function_call","name":"apply_patch","arguments":"{\"input\":\"*** Begin Patch\\n*** Update File: src/auth.rs\\n@@\\n-use bcrypt::hash;\\n+use argon2::Argon2;\\n*** Update File: src/shared/crypto.rs\\n@@\\n-pub const HASH_COST: u32 = 12;\\n+pub const ARGON2_MEM_KIB: u32 = 19456;\\n*** End Patch\\n\"}","call_id":"call_apply_patch_arc_a"}}
+{"timestamp":"2026-06-22T18:00:16.700Z","type":"response_item","payload":{"type":"function_call_output","call_id":"call_apply_patch_arc_a","output":{"output":"Applied patch to 2 files: src/auth.rs, src/shared/crypto.rs","metadata":{"exit_code":0,"duration_ms":38}}}}
+{"timestamp":"2026-06-22T18:00:24.000Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"argon2 is now wired into auth and the shared crypto helper."}]}}
+{"timestamp":"2026-06-22T18:00:40.200Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"Now let's switch to a 30-minute session TTL. Update the session manager and bump the shared default in the crypto helper's config block."}]}}
+{"timestamp":"2026-06-22T18:00:45.600Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Setting the session TTL to 30 minutes in the session manager and the shared config default."}]}}
+{"timestamp":"2026-06-22T18:00:52.800Z","type":"response_item","payload":{"type":"function_call","name":"apply_patch","arguments":"{\"input\":\"*** Begin Patch\\n*** Update File: src/session.rs\\n@@\\n-const SESSION_TTL_SECS: u64 = 3600;\\n+const SESSION_TTL_SECS: u64 = 1800;\\n*** Update File: src/shared/crypto.rs\\n@@\\n-pub const DEFAULT_TTL_SECS: u64 = 3600;\\n+pub const DEFAULT_TTL_SECS: u64 = 1800;\\n*** End Patch\\n\"}","call_id":"call_apply_patch_arc_b"}}
+{"timestamp":"2026-06-22T18:00:53.050Z","type":"response_item","payload":{"type":"function_call_output","call_id":"call_apply_patch_arc_b","output":{"output":"Applied patch to 2 files: src/session.rs, src/shared/crypto.rs","metadata":{"exit_code":0,"duration_ms":41}}}}
+{"timestamp":"2026-06-22T18:00:58.000Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Done. Session TTL is now 30 minutes in both the manager and the shared default."}]}}
diff --git a/fixtures/codex/v2/multi_edit_single_commit.jsonl b/fixtures/codex/v2/multi_edit_single_commit.jsonl
new file mode 100644
index 0000000..a4fe5b1
--- /dev/null
+++ b/fixtures/codex/v2/multi_edit_single_commit.jsonl
@@ -0,0 +1,7 @@
+{"timestamp":"2026-06-22T19:00:00.000Z","type":"session_meta","payload":{"id":"01JZ8codexsess0006multiedit","cwd":"/Users/dev/acme-api","git":{"sha":"f60718293a4b5c6d7e8f90010203040506070809","branch":"feat/rename-client"},"cli_version":"0.6.2","originator":"codex_cli_rs"}}
+{"timestamp":"2026-06-22T19:00:07.250Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"Let's rename the HttpClient type to ApiClient everywhere. Use ApiClient as the canonical name across the codebase."}]}}
+{"timestamp":"2026-06-22T19:00:12.900Z","type":"response_item","payload":{"type":"reasoning","summary":["This is a single logical rename commit touching several files. I'll do it in one apply_patch."]}}
+{"timestamp":"2026-06-22T19:00:15.400Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Renaming HttpClient to ApiClient across client.rs, handlers.rs, and main.rs in one patch."}]}}
+{"timestamp":"2026-06-22T19:00:23.800Z","type":"response_item","payload":{"type":"function_call","name":"apply_patch","arguments":"{\"input\":\"*** Begin Patch\\n*** Update File: src/client.rs\\n@@\\n-pub struct HttpClient {\\n+pub struct ApiClient {\\n*** Update File: src/handlers.rs\\n@@\\n-use crate::client::HttpClient;\\n+use crate::client::ApiClient;\\n@@\\n-    client: HttpClient,\\n+    client: ApiClient,\\n*** Update File: src/main.rs\\n@@\\n-    let client = HttpClient::new();\\n+    let client = ApiClient::new();\\n*** End Patch\\n\"}","call_id":"call_apply_patch_multi_001"}}
+{"timestamp":"2026-06-22T19:00:24.100Z","type":"response_item","payload":{"type":"function_call_output","call_id":"call_apply_patch_multi_001","output":{"output":"Applied patch to 3 files: src/client.rs, src/handlers.rs, src/main.rs","metadata":{"exit_code":0,"duration_ms":53}}}}
+{"timestamp":"2026-06-22T19:00:30.000Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Done. HttpClient is now ApiClient across all three files in a single commit-sized change."}]}}
diff --git a/fixtures/codex/v2/no_commitment_marker.jsonl b/fixtures/codex/v2/no_commitment_marker.jsonl
new file mode 100644
index 0000000..4ac0119
--- /dev/null
+++ b/fixtures/codex/v2/no_commitment_marker.jsonl
@@ -0,0 +1,6 @@
+{"timestamp":"2026-06-22T22:00:00.000Z","type":"session_meta","payload":{"id":"01JZ8codexsess0009nomarker","cwd":"/Users/dev/acme-api","git":{"sha":"293a4b5c6d7e8f90010203040506070809101112","branch":"chore/tidy"},"cli_version":"0.6.2","originator":"codex_cli_rs"}}
+{"timestamp":"2026-06-22T22:00:05.400Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"Thanks, this looks great. Go ahead and tidy up that stray import while you're in there."}]}}
+{"timestamp":"2026-06-22T22:00:10.100Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Sure, removing the unused import."}]}}
+{"timestamp":"2026-06-22T22:00:16.700Z","type":"response_item","payload":{"type":"function_call","name":"apply_patch","arguments":"{\"input\":\"*** Begin Patch\\n*** Update File: src/handlers.rs\\n@@\\n-use std::collections::HashMap;\\n use std::collections::BTreeMap;\\n*** End Patch\\n\"}","call_id":"call_apply_patch_nomark_001"}}
+{"timestamp":"2026-06-22T22:00:16.950Z","type":"response_item","payload":{"type":"function_call_output","call_id":"call_apply_patch_nomark_001","output":{"output":"Applied patch to 1 file: src/handlers.rs","metadata":{"exit_code":0,"duration_ms":21}}}}
+{"timestamp":"2026-06-22T22:00:21.000Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Removed the stray HashMap import."}]}}
diff --git a/fixtures/codex/v2/rejected_alternative.jsonl b/fixtures/codex/v2/rejected_alternative.jsonl
new file mode 100644
index 0000000..0d00c40
--- /dev/null
+++ b/fixtures/codex/v2/rejected_alternative.jsonl
@@ -0,0 +1,7 @@
+{"timestamp":"2026-06-22T15:10:00.000Z","type":"session_meta","payload":{"id":"01JZ8codexsess0002rejected","cwd":"/Users/dev/acme-api","git":{"sha":"b2c3d4e5f60718293a4b5c6d7e8f900102030405","branch":"feat/api-shape"},"cli_version":"0.6.2","originator":"codex_cli_rs"}}
+{"timestamp":"2026-06-22T15:10:08.220Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"We considered GraphQL for the public API, but let's reject that and stick with REST. GraphQL adds caching and schema-evolution overhead we don't want right now."}]}}
+{"timestamp":"2026-06-22T15:10:14.500Z","type":"response_item","payload":{"type":"reasoning","summary":["User rejected GraphQL in favor of REST. No code change requested beyond confirming the decision."]}}
+{"timestamp":"2026-06-22T15:10:16.900Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Understood. We'll keep the REST design and drop the GraphQL exploration."}]}}
+{"timestamp":"2026-06-22T15:10:30.100Z","type":"response_item","payload":{"type":"function_call","name":"shell","arguments":"{\"command\":[\"grep\",\"-r\",\"graphql\",\"src/\"]}","call_id":"call_grep_001"}}
+{"timestamp":"2026-06-22T15:10:30.450Z","type":"response_item","payload":{"type":"function_call_output","call_id":"call_grep_001","output":{"output":"","metadata":{"exit_code":1,"duration_ms":18}}}}
+{"timestamp":"2026-06-22T15:10:33.700Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"No GraphQL references exist in the codebase, so nothing to remove."}]}}
diff --git a/fixtures/codex/v2/rewind_compaction.jsonl b/fixtures/codex/v2/rewind_compaction.jsonl
new file mode 100644
index 0000000..8ce6e97
--- /dev/null
+++ b/fixtures/codex/v2/rewind_compaction.jsonl
@@ -0,0 +1,10 @@
+{"timestamp":"2026-06-22T20:00:00.000Z","type":"session_meta","payload":{"id":"01JZ8codexsess0007rewindcompact","cwd":"/Users/dev/acme-api","git":{"sha":"0718293a4b5c6d7e8f9001020304050607080910","branch":"feat/persistence-pivot"},"cli_version":"0.6.2","originator":"codex_cli_rs"}}
+{"timestamp":"2026-06-22T20:00:06.300Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"Let's use MongoDB for the document store. Set up the driver and the connection config."}]}}
+{"timestamp":"2026-06-22T20:00:11.700Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Understood, I'll wire up MongoDB for the document store."}]}}
+{"timestamp":"2026-06-22T20:00:30.000Z","type":"event_msg","payload":{"type":"compacted","reason":"auto_compact","message":"Conversation history compacted to fit the context window. The MongoDB exploration above was summarized and the verbatim early turns were dropped from the working context.","replaced_response_ids":["msg-mongo-user","msg-mongo-assistant"]}}
+{"timestamp":"2026-06-22T20:00:31.000Z","type":"turn_context","payload":{"summary":"Earlier the user explored MongoDB for the document store; that turn was compacted out of the active context.","cwd":"/Users/dev/acme-api"}}
+{"timestamp":"2026-06-22T20:00:40.200Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"Actually let's use Postgres with JSONB for the document store instead. Switch to Postgres and drop the MongoDB plan."}]}}
+{"timestamp":"2026-06-22T20:00:45.900Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Switching the document store to Postgres with JSONB columns."}]}}
+{"timestamp":"2026-06-22T20:00:53.400Z","type":"response_item","payload":{"type":"function_call","name":"apply_patch","arguments":"{\"input\":\"*** Begin Patch\\n*** Update File: src/store.rs\\n@@\\n-use mongodb::Client;\\n+use sqlx::PgPool;\\n@@\\n-    let client = Client::with_uri_str(\\\"mongodb://localhost:27017\\\").await?;\\n+    let pool = PgPool::connect(\\\"postgres://localhost:5432/acme\\\").await?;\\n*** End Patch\\n\"}","call_id":"call_apply_patch_rewind_001"}}
+{"timestamp":"2026-06-22T20:00:53.700Z","type":"response_item","payload":{"type":"function_call_output","call_id":"call_apply_patch_rewind_001","output":{"output":"Applied patch to 1 file: src/store.rs","metadata":{"exit_code":0,"duration_ms":36}}}}
+{"timestamp":"2026-06-22T20:00:58.000Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Done. The document store now uses Postgres with JSONB."}]}}
diff --git a/fixtures/codex/v2/subagent_thread.jsonl b/fixtures/codex/v2/subagent_thread.jsonl
new file mode 100644
index 0000000..fae31e9
--- /dev/null
+++ b/fixtures/codex/v2/subagent_thread.jsonl
@@ -0,0 +1,7 @@
+{"timestamp":"2026-06-22T21:00:00.000Z","type":"session_meta","payload":{"id":"01JZ8codexsess0008subagent","cwd":"/Users/dev/acme-api","git":{"sha":"18293a4b5c6d7e8f900102030405060708091011","branch":"feat/worker-split"},"cli_version":"0.6.2","originator":"codex_cli_rs"}}
+{"timestamp":"2026-06-22T21:00:06.500Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"Let's use a dedicated background worker for the email queue. Extract the send loop into a worker module."}]}}
+{"timestamp":"2026-06-22T21:00:11.800Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"I'll extract the send loop into a dedicated worker module."}]}}
+{"timestamp":"2026-06-22T21:00:14.000Z","type":"turn_context","payload":{"thread_id":"thread-email-worker","parent_thread_id":"01JZ8codexsess0008subagent","summary":"Nested task thread for the email worker extraction. Codex has no first-class subagent; this is a secondary thread id carried in turn_context."}}
+{"timestamp":"2026-06-22T21:00:22.300Z","type":"response_item","payload":{"type":"function_call","name":"apply_patch","arguments":"{\"input\":\"*** Begin Patch\\n*** Add File: src/worker/email.rs\\n+pub async fn run_email_worker() {\\n+    loop { drain_email_queue().await; }\\n+}\\n*** Update File: src/main.rs\\n@@\\n-    send_emails_inline();\\n+    tokio::spawn(worker::email::run_email_worker());\\n*** End Patch\\n\"}","call_id":"call_apply_patch_sub_001"}}
+{"timestamp":"2026-06-22T21:00:22.600Z","type":"response_item","payload":{"type":"function_call_output","call_id":"call_apply_patch_sub_001","output":{"output":"Applied patch to 2 files: src/worker/email.rs, src/main.rs","metadata":{"exit_code":0,"duration_ms":44}}}}
+{"timestamp":"2026-06-22T21:00:28.000Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"Done. The email send loop now runs in a dedicated worker module spawned from main."}]}}
diff --git a/fixtures/codex/v2/tool_failure.jsonl b/fixtures/codex/v2/tool_failure.jsonl
new file mode 100644
index 0000000..e5653df
--- /dev/null
+++ b/fixtures/codex/v2/tool_failure.jsonl
@@ -0,0 +1,6 @@
+{"timestamp":"2026-06-22T17:05:00.000Z","type":"session_meta","payload":{"id":"01JZ8codexsess0004toolfail","cwd":"/Users/dev/acme-api","git":{"sha":"d4e5f60718293a4b5c6d7e8f9001020304050607","branch":"fix/config-loader"},"cli_version":"0.6.2","originator":"codex_cli_rs"}}
+{"timestamp":"2026-06-22T17:05:06.410Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"Switch the config loader to read from config.toml instead of config.yaml."}]}}
+{"timestamp":"2026-06-22T17:05:11.900Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"I'll update the loader to point at config.toml."}]}}
+{"timestamp":"2026-06-22T17:05:18.220Z","type":"response_item","payload":{"type":"function_call","name":"apply_patch","arguments":"{\"input\":\"*** Begin Patch\\n*** Update File: src/config.rs\\n@@\\n-let path = \\\"config.yaml\\\";\\n+let path = \\\"config.toml\\\";\\n*** End Patch\\n\"}","call_id":"call_apply_patch_fail_001"}}
+{"timestamp":"2026-06-22T17:05:18.560Z","type":"response_item","payload":{"type":"function_call_output","call_id":"call_apply_patch_fail_001","output":{"output":"error: patch does not apply: src/config.rs: context line \"let path = \\\"config.yaml\\\";\" not found","metadata":{"exit_code":1,"duration_ms":27}}}}
+{"timestamp":"2026-06-22T17:05:25.700Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"The patch failed to apply because the context didn't match. Let me re-read the file before retrying."}]}}
diff --git a/fixtures/copilot/v1/ban.jsonl b/fixtures/copilot/v1/ban.jsonl
new file mode 100644
index 0000000..f1bb8fe
--- /dev/null
+++ b/fixtures/copilot/v1/ban.jsonl
@@ -0,0 +1,4 @@
+{"kind":"session_start","cwd":"/Users/dev/projects/core","git":{"sha":"beadfeed9001","branch":"main"},"toolVersion":"copilot-cli 1.4.0","model":"gpt-4o","sessionId":"copilot-thread-003","ts":"2026-06-22T12:00:00Z"}
+{"id":"msg-1","parentId":null,"role":"user","ts":"2026-06-22T12:00:11Z","sessionId":"copilot-thread-003","text":"We will never add a dependency on left-pad. Do not introduce it under any circumstances."}
+{"id":"msg-2","parentId":"msg-1","role":"assistant","ts":"2026-06-22T12:00:16Z","sessionId":"copilot-thread-003","text":"Acknowledged. I won't add left-pad; I'll inline a small padding helper instead.","model":"gpt-4o","usage":{"input":640,"output":120}}
+{"kind":"session_end","sessionId":"copilot-thread-003","ts":"2026-06-22T12:00:50Z","reason":"user_closed"}
diff --git a/fixtures/copilot/v1/happy_path_decision_then_edits.jsonl b/fixtures/copilot/v1/happy_path_decision_then_edits.jsonl
new file mode 100644
index 0000000..70a0894
--- /dev/null
+++ b/fixtures/copilot/v1/happy_path_decision_then_edits.jsonl
@@ -0,0 +1,5 @@
+{"kind":"session_start","cwd":"/Users/dev/projects/orbit","git":{"sha":"a1b2c3d4e5f6","branch":"main"},"toolVersion":"copilot-cli 1.4.0","model":"gpt-4o","sessionId":"copilot-thread-001","ts":"2026-06-22T10:00:00Z"}
+{"id":"msg-1","parentId":null,"role":"user","ts":"2026-06-22T10:00:05Z","sessionId":"copilot-thread-001","text":"Let's use Postgres instead of MySQL for the persistence layer."}
+{"id":"msg-2","parentId":"msg-1","role":"assistant","ts":"2026-06-22T10:00:09Z","sessionId":"copilot-thread-001","text":"Good call. I'll switch the connection config and the schema module to Postgres.","model":"gpt-4o","usage":{"input":1200,"output":340},"edits":[{"path":"src/db/config.rs","oldText":"let url = \"mysql://localhost/orbit\";","newText":"let url = \"postgres://localhost/orbit\";","diff":"@@ -1 +1 @@\n-let url = \"mysql://localhost/orbit\";\n+let url = \"postgres://localhost/orbit\";","added":1,"removed":1},{"path":"src/db/schema.rs","oldText":"// MySQL dialect\n","newText":"// Postgres dialect\nuse sqlx::postgres::PgPool;\n","added":2,"removed":1}]}
+{"id":"msg-3","parentId":"msg-2","role":"user","ts":"2026-06-22T10:01:00Z","sessionId":"copilot-thread-001","text":"Looks right, thanks."}
+{"kind":"session_end","sessionId":"copilot-thread-001","ts":"2026-06-22T10:01:30Z","reason":"user_closed"}
diff --git a/fixtures/copilot/v1/rejected_alternative.jsonl b/fixtures/copilot/v1/rejected_alternative.jsonl
new file mode 100644
index 0000000..3271a1a
--- /dev/null
+++ b/fixtures/copilot/v1/rejected_alternative.jsonl
@@ -0,0 +1,5 @@
+{"kind":"session_start","cwd":"/Users/dev/projects/checkout","git":{"sha":"f00dcafe1234","branch":"feat/payments"},"toolVersion":"copilot-cli 1.4.0","model":"gpt-4o","sessionId":"copilot-thread-002","ts":"2026-06-22T11:00:00Z"}
+{"id":"msg-1","parentId":null,"role":"user","ts":"2026-06-22T11:00:08Z","sessionId":"copilot-thread-002","text":"Let's use Stripe instead of PayPal for the payment provider."}
+{"id":"msg-2","parentId":"msg-1","role":"assistant","ts":"2026-06-22T11:00:14Z","sessionId":"copilot-thread-002","text":"Understood. I'll wire up the Stripe SDK and drop the PayPal client.","model":"gpt-4o","usage":{"input":980,"output":210},"toolCalls":[{"id":"call-a1","name":"apply_patch","args":{"path":"src/payments/provider.rs"}}],"toolResults":[{"id":"call-a1","ok":true,"output":"patch applied (1 file changed)"}],"edits":[{"path":"src/payments/provider.rs","callId":"call-a1","oldText":"use paypal_sdk::Client;","newText":"use stripe::Client;","diff":"@@ -1 +1 @@\n-use paypal_sdk::Client;\n+use stripe::Client;","added":1,"removed":1}]}
+{"id":"msg-3","parentId":"msg-2","role":"user","ts":"2026-06-22T11:01:20Z","sessionId":"copilot-thread-002","text":"Perfect."}
+{"kind":"session_end","sessionId":"copilot-thread-002","ts":"2026-06-22T11:01:40Z","reason":"user_closed"}
diff --git a/fixtures/copilot/v1/tool_failure.jsonl b/fixtures/copilot/v1/tool_failure.jsonl
new file mode 100644
index 0000000..7f7d4e4
--- /dev/null
+++ b/fixtures/copilot/v1/tool_failure.jsonl
@@ -0,0 +1,5 @@
+{"kind":"session_start","cwd":"/Users/dev/projects/api","git":{"sha":"0badc0de7777","branch":"fix/migration"},"toolVersion":"copilot-cli 1.4.0","model":"gpt-4o","sessionId":"copilot-thread-004","ts":"2026-06-22T13:00:00Z"}
+{"id":"msg-1","parentId":null,"role":"user","ts":"2026-06-22T13:00:07Z","sessionId":"copilot-thread-004","text":"Add a NOT NULL constraint to the users.email column."}
+{"id":"msg-2","parentId":"msg-1","role":"assistant","ts":"2026-06-22T13:00:13Z","sessionId":"copilot-thread-004","text":"I'll apply the migration patch.","model":"gpt-4o","usage":{"input":720,"output":90},"toolCalls":[{"id":"call-z9","name":"apply_patch","args":{"path":"migrations/0007_email_not_null.sql"}}],"toolResults":[{"id":"call-z9","ok":false,"output":"error: patch did not apply cleanly: hunk #1 FAILED at line 3"}],"edits":[{"path":"migrations/0007_email_not_null.sql","callId":"call-z9","oldText":"email TEXT","newText":"email TEXT NOT NULL","diff":"@@ -3 +3 @@\n-email TEXT\n+email TEXT NOT NULL","added":1,"removed":1}]}
+{"id":"msg-3","parentId":"msg-2","role":"user","ts":"2026-06-22T13:01:00Z","sessionId":"copilot-thread-004","text":"Leave it for now, I'll handle the conflict manually."}
+{"kind":"session_end","sessionId":"copilot-thread-004","ts":"2026-06-22T13:01:20Z","reason":"user_closed"}
diff --git a/fixtures/cursor/v1/ban.jsonl b/fixtures/cursor/v1/ban.jsonl
new file mode 100644
index 0000000..17f80cf
--- /dev/null
+++ b/fixtures/cursor/v1/ban.jsonl
@@ -0,0 +1,3 @@
+{"kind":"session_start","sessionId":"cur-sess-003","cwd":"/Users/dev/projects/payments-core","git":{"sha":"a17de4490b22","branch":"main"},"toolVersion":"0.43.2"}
+{"id":"msg-201","parentId":null,"role":"user","ts":"2026-06-22T16:20:00Z","sessionId":"cur-sess-003","text":"We will never add a dependency on left-pad. Use the standard library padding instead, always."}
+{"id":"msg-202","parentId":"msg-201","role":"assistant","ts":"2026-06-22T16:20:11Z","sessionId":"cur-sess-003","text":"Understood — no left-pad dependency, ever. I'll use String::repeat / format width specifiers for any padding we need.","model":"cursor-fast","usage":{"input":140,"output":58},"toolCalls":[{"id":"call-edit-3","name":"edit_file","args":{"path":"src/format.rs"}}],"toolResults":[{"id":"call-edit-3","ok":true,"output":"applied 1 edit"}],"edits":[{"path":"src/format.rs","oldText":"use left_pad::left_pad;\nlet s = left_pad(id, 8);","newText":"let s = format!(\"{id:>8}\");","diff":"@@ -1,2 +1 @@\n-use left_pad::left_pad;\n-let s = left_pad(id, 8);\n+let s = format!(\"{id:>8}\");\n","added":1,"removed":2}]}
diff --git a/fixtures/cursor/v1/happy_path_decision_then_edits.jsonl b/fixtures/cursor/v1/happy_path_decision_then_edits.jsonl
new file mode 100644
index 0000000..84c3176
--- /dev/null
+++ b/fixtures/cursor/v1/happy_path_decision_then_edits.jsonl
@@ -0,0 +1,5 @@
+{"kind":"session_start","sessionId":"cur-sess-001","cwd":"/Users/dev/projects/orders-api","git":{"sha":"9f2c1ad4b7e0","branch":"main"},"toolVersion":"0.43.2","model":"cursor-fast"}
+{"id":"msg-001","parentId":null,"role":"user","ts":"2026-06-22T14:01:10Z","sessionId":"cur-sess-001","text":"Let's use Postgres instead of MySQL for the orders service. Update the config and the connection module."}
+{"id":"msg-002","parentId":"msg-001","role":"assistant","ts":"2026-06-22T14:01:18Z","sessionId":"cur-sess-001","text":"Good call — Postgres gives us better JSONB support. I'll switch the engine in the config and update the DSN in the connection module.","model":"cursor-fast","usage":{"input":214,"output":98},"toolCalls":[{"id":"call-edit-1","name":"edit_file","args":{"path":"config/database.toml"}}],"toolResults":[{"id":"call-edit-1","ok":true,"output":"applied 1 edit"}],"edits":[{"path":"config/database.toml","oldText":"[database]\nengine = \"mysql\"\nport = 3306\n","newText":"[database]\nengine = \"postgres\"\nport = 5432\n","diff":"@@ -1,3 +1,3 @@\n [database]\n-engine = \"mysql\"\n-port = 3306\n+engine = \"postgres\"\n+port = 5432\n","added":2,"removed":2}]}
+{"id":"msg-003","parentId":"msg-002","role":"assistant","ts":"2026-06-22T14:01:31Z","sessionId":"cur-sess-001","text":"Now updating the connection module to use the Postgres driver.","model":"cursor-fast","usage":{"input":120,"output":64},"toolCalls":[{"id":"call-edit-2","name":"edit_file","args":{"path":"src/db/connect.rs"}}],"toolResults":[{"id":"call-edit-2","ok":true,"output":"applied 1 edit"}],"edits":[{"path":"src/db/connect.rs","oldText":"let url = \"mysql://localhost:3306/orders\";","newText":"let url = \"postgres://localhost:5432/orders\";","diff":"@@ -1 +1 @@\n-let url = \"mysql://localhost:3306/orders\";\n+let url = \"postgres://localhost:5432/orders\";\n","added":1,"removed":1}]}
+{"kind":"session_end","sessionId":"cur-sess-001","ts":"2026-06-22T14:02:00Z","reason":"user_closed"}
diff --git a/fixtures/cursor/v1/rejected_alternative.jsonl b/fixtures/cursor/v1/rejected_alternative.jsonl
new file mode 100644
index 0000000..8c6afdf
--- /dev/null
+++ b/fixtures/cursor/v1/rejected_alternative.jsonl
@@ -0,0 +1,5 @@
+{"kind":"session_start","sessionId":"cur-sess-002","cwd":"/Users/dev/projects/web-dashboard","git":{"sha":"3b81fe09cc12","branch":"feat/state-mgmt"},"toolVersion":"0.43.2"}
+{"id":"msg-101","parentId":null,"role":"user","ts":"2026-06-22T15:10:00Z","sessionId":"cur-sess-002","text":"Should we use Redux or Zustand for the new dashboard state?"}
+{"id":"msg-102","parentId":"msg-101","role":"assistant","ts":"2026-06-22T15:10:09Z","sessionId":"cur-sess-002","text":"I considered Redux, but for this dashboard it's heavier than we need — too much boilerplate for the small store. We'll go with Zustand instead.","model":"cursor-pro","usage":{"input":180,"output":72}}
+{"id":"msg-103","parentId":"msg-102","role":"user","ts":"2026-06-22T15:10:40Z","sessionId":"cur-sess-002","text":"Agreed, reject Redux. Let's go with Zustand and wire up the store."}
+{"id":"msg-104","parentId":"msg-103","role":"assistant","ts":"2026-06-22T15:10:52Z","sessionId":"cur-sess-002","text":"Creating the Zustand store now.","model":"cursor-pro","usage":{"input":96,"output":40},"toolCalls":[{"id":"call-write-1","name":"write_file","args":{"path":"src/store/dashboard.ts"}}],"toolResults":[{"id":"call-write-1","ok":true,"output":"created file"}],"edits":[{"path":"src/store/dashboard.ts","oldText":"","newText":"import { create } from 'zustand';\nexport const useDashboard = create((set) => ({ panels: [] }));\n","diff":"@@ -0,0 +1,2 @@\n+import { create } from 'zustand';\n+export const useDashboard = create((set) => ({ panels: [] }));\n","added":2,"removed":0}]}
diff --git a/fixtures/cursor/v1/tool_failure.jsonl b/fixtures/cursor/v1/tool_failure.jsonl
new file mode 100644
index 0000000..6dbead2
--- /dev/null
+++ b/fixtures/cursor/v1/tool_failure.jsonl
@@ -0,0 +1,3 @@
+{"kind":"session_start","sessionId":"cur-sess-004","cwd":"/Users/dev/projects/infra-scripts","git":{"sha":"c0ffee123456","branch":"main"},"toolVersion":"0.43.2"}
+{"id":"msg-301","parentId":null,"role":"user","ts":"2026-06-22T17:05:00Z","sessionId":"cur-sess-004","text":"Add a retry wrapper around the deploy step in deploy.sh."}
+{"id":"msg-302","parentId":"msg-301","role":"assistant","ts":"2026-06-22T17:05:14Z","sessionId":"cur-sess-004","text":"I'll wrap the deploy step in a retry loop.","model":"cursor-fast","usage":{"input":110,"output":52},"toolCalls":[{"id":"call-edit-4","name":"edit_file","args":{"path":"deploy.sh"}}],"toolResults":[{"id":"call-edit-4","ok":false,"output":"error: permission denied: deploy.sh is read-only (EACCES)"}],"edits":[{"path":"deploy.sh","callId":"call-edit-4","oldText":"run_deploy","newText":"for i in 1 2 3; do run_deploy && break; done","diff":"@@ -1 +1 @@\n-run_deploy\n+for i in 1 2 3; do run_deploy && break; done\n","added":1,"removed":1}]}
diff --git a/fixtures/gemini/legacy_json/version_tolerance.jsonl b/fixtures/gemini/legacy_json/version_tolerance.jsonl
new file mode 100644
index 0000000..32f56ca
--- /dev/null
+++ b/fixtures/gemini/legacy_json/version_tolerance.jsonl
@@ -0,0 +1,5 @@
+{"$set":{"sessionId":"legacy-0001","cwd":"/home/dev/shop-api","projectRoot":"/home/dev/shop-api","model":"gemini-1.5-pro","branch":"main","commit":"9f1c0aa2b3","cliVersion":"0.2.0"}}
+{"id":"leg-0001","role":"user","content":"Let's use Postgres instead of MySQL for the orders service.","timestamp":"2026-06-22T08:00:00.000Z"}
+{"id":"leg-0002","role":"model","content":"Good call. Switching the connection string to Postgres.","thoughts":"swap the dsn","timestamp":"2026-06-22T08:00:03.120Z"}
+{"id":"leg-0003","role":"system","content":"Context window compacted.","timestamp":"2026-06-22T08:00:04.000Z"}
+{"$rewindTo":"leg-0001"}
diff --git a/fixtures/gemini/v1/ban.jsonl b/fixtures/gemini/v1/ban.jsonl
new file mode 100644
index 0000000..3bf9d44
--- /dev/null
+++ b/fixtures/gemini/v1/ban.jsonl
@@ -0,0 +1,3 @@
+{"$set":{"sessionId":"a1b2c3d4-0003","cwd":"/home/dev/core-lib","model":"gemini-2.5-pro","branch":"main","commit":"0042deadbe","cliVersion":"0.4.1"}}
+{"id":"msg-0201","role":"user","text":"Important policy: we will never add a dependency on left-pad. Use the std library instead.","timestamp":"2026-06-22T12:30:00.000Z"}
+{"id":"msg-0202","role":"model","text":"Acknowledged. I won't add left-pad; I'll implement padding with the standard library.","thoughts":"This is a ban on the left-pad dependency.","tokens":{"input":110,"output":58},"timestamp":"2026-06-22T12:30:02.700Z"}
diff --git a/fixtures/gemini/v1/happy_path_decision_then_edits.jsonl b/fixtures/gemini/v1/happy_path_decision_then_edits.jsonl
new file mode 100644
index 0000000..4c51382
--- /dev/null
+++ b/fixtures/gemini/v1/happy_path_decision_then_edits.jsonl
@@ -0,0 +1,4 @@
+{"$set":{"sessionId":"a1b2c3d4-0001","cwd":"/home/dev/shop-api","projectRoot":"/home/dev/shop-api","model":"gemini-2.5-pro","branch":"main","commit":"9f1c0aa2b3","cliVersion":"0.4.1"}}
+{"id":"msg-0001","role":"user","text":"Let's use Postgres instead of MySQL for the orders service. Update the config and the migration runner.","timestamp":"2026-06-22T10:00:00.000Z"}
+{"id":"msg-0002","role":"model","text":"Good call. I'll switch the connection string to Postgres and update the migration runner to use the pg driver.","thoughts":"Replace the mysql DSN with a postgres URL, then swap the driver import in the runner.","tokens":{"input":182,"output":97},"timestamp":"2026-06-22T10:00:03.120Z","toolCalls":[{"callId":"call-cfg-1","name":"write_file","args":{"path":"config/database.toml"},"resultDisplay":{"fileName":"config/database.toml","originalContent":"[db]\ndriver = \"mysql\"\nurl = \"mysql://localhost:3306/orders\"\n","newContent":"[db]\ndriver = \"postgres\"\nurl = \"postgres://localhost:5432/orders\"\n","fileDiff":"@@ -1,3 +1,3 @@\n [db]\n-driver = \"mysql\"\n-url = \"mysql://localhost:3306/orders\"\n+driver = \"postgres\"\n+url = \"postgres://localhost:5432/orders\"\n","diffStat":{"added":2,"removed":2}}},{"callId":"call-run-1","name":"replace","args":{"path":"src/migrate.rs"},"resultDisplay":{"filePath":"src/migrate.rs","originalContent":"use mysql::Pool;\n","newContent":"use postgres::Client;\n","fileDiff":"@@ -1 +1 @@\n-use mysql::Pool;\n+use postgres::Client;\n","diffStat":{"added":1,"removed":1}}}]}
+{"id":"msg-0003","role":"user","text":"Looks good, thanks.","timestamp":"2026-06-22T10:00:30.000Z"}
diff --git a/fixtures/gemini/v1/interleaved_arcs.jsonl b/fixtures/gemini/v1/interleaved_arcs.jsonl
new file mode 100644
index 0000000..418ab87
--- /dev/null
+++ b/fixtures/gemini/v1/interleaved_arcs.jsonl
@@ -0,0 +1,6 @@
+{"$set":{"sessionId":"g-interleaved-0001","cwd":"/home/dev/orders-api","projectRoot":"/home/dev/orders-api","model":"gemini-2.5-pro","branch":"main","commit":"a1a1a1a1a1","cliVersion":"0.4.1"}}
+{"id":"msg-ia-01","role":"user","text":"Let's use Postgres for the orders store. Update the connection module and the schema.","timestamp":"2026-06-22T09:00:00.000Z"}
+{"id":"msg-ia-02","role":"model","text":"Switching the orders store to Postgres now.","thoughts":"Rewrite db.rs to the pg driver and adjust schema.rs.","tokens":{"input":120,"output":61},"timestamp":"2026-06-22T09:00:02.000Z","toolCalls":[{"callId":"call-ia-db1","name":"replace","args":{"path":"src/db.rs"},"resultDisplay":{"fileName":"src/db.rs","originalContent":"use mysql::Pool;\n","newContent":"use postgres::Client;\n","fileDiff":"@@ -1 +1 @@\n-use mysql::Pool;\n+use postgres::Client;\n","diffStat":{"added":1,"removed":1}}},{"callId":"call-ia-schema1","name":"replace","args":{"path":"src/schema.rs"},"resultDisplay":{"fileName":"src/schema.rs","originalContent":"-- mysql schema\n","newContent":"-- postgres schema\n","fileDiff":"@@ -1 +1 @@\n--- mysql schema\n+-- postgres schema\n","diffStat":{"added":1,"removed":1}}}]}
+{"id":"msg-ia-03","role":"user","text":"Now switch to a bounded connection pool in the same db module instead of a single client.","timestamp":"2026-06-22T09:05:00.000Z"}
+{"id":"msg-ia-04","role":"model","text":"Adding a bounded pool to db.rs and a new pool config.","thoughts":"Re-edit db.rs to use a pool; add pool.rs.","tokens":{"input":98,"output":52},"timestamp":"2026-06-22T09:05:02.000Z","toolCalls":[{"callId":"call-ia-db2","name":"replace","args":{"path":"src/db.rs"},"resultDisplay":{"fileName":"src/db.rs","originalContent":"use postgres::Client;\n","newContent":"use deadpool_postgres::Pool;\n","fileDiff":"@@ -1 +1 @@\n-use postgres::Client;\n+use deadpool_postgres::Pool;\n","diffStat":{"added":1,"removed":1}}},{"callId":"call-ia-pool1","name":"write_file","args":{"path":"src/pool.rs"},"resultDisplay":{"fileName":"src/pool.rs","originalContent":"","newContent":"pub struct PoolCfg { pub max: usize }\n","fileDiff":"@@ -0,0 +1 @@\n+pub struct PoolCfg { pub max: usize }\n","diffStat":{"added":1,"removed":0}}}]}
+{"id":"msg-ia-05","role":"user","text":"Great, that covers it.","timestamp":"2026-06-22T09:05:30.000Z"}
diff --git a/fixtures/gemini/v1/multi_edit_single_commit.jsonl b/fixtures/gemini/v1/multi_edit_single_commit.jsonl
new file mode 100644
index 0000000..64ffe2e
--- /dev/null
+++ b/fixtures/gemini/v1/multi_edit_single_commit.jsonl
@@ -0,0 +1,4 @@
+{"$set":{"sessionId":"g-multiedit-0001","cwd":"/home/dev/auth-svc","projectRoot":"/home/dev/auth-svc","model":"gemini-2.5-pro","branch":"feat/argon2","commit":"b2b2b2b2b2","cliVersion":"0.4.1"}}
+{"id":"msg-me-01","role":"user","text":"We will use Argon2 for password hashing across the auth service. Update the hasher, the verifier, and the config in one pass.","timestamp":"2026-06-22T14:00:00.000Z"}
+{"id":"msg-me-02","role":"model","text":"Rolling Argon2 through the hasher, verifier, and config as a single change.","thoughts":"Three coordinated edits landing as one logical commit.","tokens":{"input":160,"output":88},"timestamp":"2026-06-22T14:00:03.000Z","toolCalls":[{"callId":"call-me-hash","name":"replace","args":{"path":"src/auth/hasher.rs"},"resultDisplay":{"fileName":"src/auth/hasher.rs","originalContent":"use bcrypt::hash;\n","newContent":"use argon2::Argon2;\n","fileDiff":"@@ -1 +1 @@\n-use bcrypt::hash;\n+use argon2::Argon2;\n","diffStat":{"added":1,"removed":1}}},{"callId":"call-me-verify","name":"replace","args":{"path":"src/auth/verifier.rs"},"resultDisplay":{"fileName":"src/auth/verifier.rs","originalContent":"use bcrypt::verify;\n","newContent":"use argon2::PasswordVerifier;\n","fileDiff":"@@ -1 +1 @@\n-use bcrypt::verify;\n+use argon2::PasswordVerifier;\n","diffStat":{"added":1,"removed":1}}},{"callId":"call-me-config","name":"replace","args":{"path":"config/auth.toml"},"resultDisplay":{"fileName":"config/auth.toml","originalContent":"hash = \"bcrypt\"\n","newContent":"hash = \"argon2\"\n","fileDiff":"@@ -1 +1 @@\n-hash = \"bcrypt\"\n+hash = \"argon2\"\n","diffStat":{"added":1,"removed":1}}}]}
+{"id":"msg-me-03","role":"user","text":"Perfect, ship it.","timestamp":"2026-06-22T14:01:00.000Z"}
diff --git a/fixtures/gemini/v1/no_commitment_marker.jsonl b/fixtures/gemini/v1/no_commitment_marker.jsonl
new file mode 100644
index 0000000..975be21
--- /dev/null
+++ b/fixtures/gemini/v1/no_commitment_marker.jsonl
@@ -0,0 +1,3 @@
+{"$set":{"sessionId":"g-nomarker-0001","cwd":"/home/dev/widgets","projectRoot":"/home/dev/widgets","model":"gemini-2.5-pro","branch":"main","commit":"e5e5e5e5e5","cliVersion":"0.4.1"}}
+{"id":"msg-nm-01","role":"user","text":"Thanks, this looks good to me.","timestamp":"2026-06-22T17:00:00.000Z"}
+{"id":"msg-nm-02","role":"model","text":"Glad it works. I'll tidy the trailing whitespace in the widget while I'm here.","thoughts":"A small cleanup edit with no governing decision turn.","tokens":{"input":70,"output":38},"timestamp":"2026-06-22T17:00:02.000Z","toolCalls":[{"callId":"call-nm-tidy","name":"replace","args":{"path":"src/widget.rs"},"resultDisplay":{"fileName":"src/widget.rs","originalContent":"fn render() { }  \n","newContent":"fn render() {}\n","fileDiff":"@@ -1 +1 @@\n-fn render() { }  \n+fn render() {}\n","diffStat":{"added":1,"removed":1}}}]}
diff --git a/fixtures/gemini/v1/rejected_alternative.jsonl b/fixtures/gemini/v1/rejected_alternative.jsonl
new file mode 100644
index 0000000..dab4514
--- /dev/null
+++ b/fixtures/gemini/v1/rejected_alternative.jsonl
@@ -0,0 +1,3 @@
+{"$set":{"sessionId":"a1b2c3d4-0002","cwd":"/home/dev/billing","model":"gemini-2.5-pro","branch":"feat/payments","commit":"77ab12cd","cliVersion":"0.4.1"}}
+{"id":"msg-0101","role":"user","text":"For the checkout flow, use Stripe instead of PayPal. Wire up the Stripe client.","timestamp":"2026-06-22T11:15:00.000Z"}
+{"id":"msg-0102","role":"model","text":"Understood — Stripe it is. Adding the Stripe client and removing the PayPal stub.","thoughts":"User explicitly rejected PayPal in favor of Stripe.","tokens":{"input":140,"output":72},"timestamp":"2026-06-22T11:15:04.500Z","toolCalls":[{"callId":"call-stripe-1","name":"write_file","args":{"path":"src/payments/stripe.rs"},"resultDisplay":{"fileName":"src/payments/stripe.rs","originalContent":"","newContent":"pub struct StripeClient { key: String }\n","fileDiff":"@@ -0,0 +1 @@\n+pub struct StripeClient { key: String }\n","diffStat":{"added":1,"removed":0}}}]}
diff --git a/fixtures/gemini/v1/rewind_compaction.jsonl b/fixtures/gemini/v1/rewind_compaction.jsonl
new file mode 100644
index 0000000..4f92857
--- /dev/null
+++ b/fixtures/gemini/v1/rewind_compaction.jsonl
@@ -0,0 +1,8 @@
+{"$set":{"sessionId":"g-rewind-0001","cwd":"/home/dev/catalog","projectRoot":"/home/dev/catalog","model":"gemini-2.5-pro","branch":"main","commit":"c3c3c3c3c3","cliVersion":"0.4.1"}}
+{"id":"msg-rc-00","role":"user","text":"Let's plan the storage layer for the catalog.","timestamp":"2026-06-22T15:00:00.000Z"}
+{"id":"msg-rc-01","role":"user","text":"Use MongoDB for the catalog store. Wire up the document client.","timestamp":"2026-06-22T15:01:00.000Z"}
+{"id":"msg-rc-02","role":"model","text":"Setting up the MongoDB document client.","thoughts":"Replace the placeholder store with a mongo client.","tokens":{"input":110,"output":54},"timestamp":"2026-06-22T15:01:03.000Z","toolCalls":[{"callId":"call-rc-mongo","name":"write_file","args":{"path":"src/store.rs"},"resultDisplay":{"fileName":"src/store.rs","originalContent":"pub struct Store;\n","newContent":"use mongodb::Client;\npub struct Store(Client);\n","fileDiff":"@@ -1 +1,2 @@\n-pub struct Store;\n+use mongodb::Client;\n+pub struct Store(Client);\n","diffStat":{"added":2,"removed":1}}}]}
+{"$rewindTo":"msg-rc-00"}
+{"id":"msg-rc-03","role":"user","text":"On reflection, use Postgres with JSONB for the catalog store instead of MongoDB.","timestamp":"2026-06-22T15:10:00.000Z"}
+{"id":"msg-rc-04","role":"model","text":"Replacing the store with a Postgres JSONB client.","thoughts":"Swap the mongo client for a postgres client.","tokens":{"input":105,"output":58},"timestamp":"2026-06-22T15:10:03.000Z","toolCalls":[{"callId":"call-rc-pg","name":"replace","args":{"path":"src/store.rs"},"resultDisplay":{"fileName":"src/store.rs","originalContent":"use mongodb::Client;\npub struct Store(Client);\n","newContent":"use postgres::Client;\npub struct Store(Client);\n","fileDiff":"@@ -1,2 +1,2 @@\n-use mongodb::Client;\n+use postgres::Client;\n pub struct Store(Client);\n","diffStat":{"added":1,"removed":1}}}]}
+{"id":"msg-rc-05","role":"user","text":"Looks right now, thanks.","timestamp":"2026-06-22T15:10:30.000Z"}
diff --git a/fixtures/gemini/v1/subagent_thread.jsonl b/fixtures/gemini/v1/subagent_thread.jsonl
new file mode 100644
index 0000000..ee53949
--- /dev/null
+++ b/fixtures/gemini/v1/subagent_thread.jsonl
@@ -0,0 +1,6 @@
+{"$set":{"sessionId":"g-subagent-0001","cwd":"/home/dev/pipeline","projectRoot":"/home/dev/pipeline","model":"gemini-2.5-pro","branch":"main","commit":"d4d4d4d4d4","cliVersion":"0.4.1"}}
+{"id":"msg-sa-01","role":"user","threadId":"main","text":"Kick off the ingest refactor; delegate the parser rewrite to a focused sub-thread.","timestamp":"2026-06-22T16:00:00.000Z"}
+{"id":"msg-sa-02","role":"model","threadId":"main","text":"Spawning a secondary thread to handle the parser rewrite.","thoughts":"Gemini has no real subagents; modeling the sub-task as a nested threadId on the same session.","tokens":{"input":90,"output":40},"timestamp":"2026-06-22T16:00:02.000Z"}
+{"id":"msg-sa-03","role":"user","threadId":"sub-parser","parentThreadId":"main","text":"In this sub-thread, use a streaming pull-parser for the ingest format instead of the buffered one.","timestamp":"2026-06-22T16:01:00.000Z"}
+{"id":"msg-sa-04","role":"model","threadId":"sub-parser","parentThreadId":"main","text":"Rewriting the parser to a streaming pull-parser.","thoughts":"Replace the buffered reader in parser.rs.","tokens":{"input":102,"output":57},"timestamp":"2026-06-22T16:01:03.000Z","toolCalls":[{"callId":"call-sa-parser","name":"replace","args":{"path":"src/ingest/parser.rs"},"resultDisplay":{"fileName":"src/ingest/parser.rs","originalContent":"let buf = read_all(input);\n","newContent":"let mut reader = PullParser::new(input);\n","fileDiff":"@@ -1 +1 @@\n-let buf = read_all(input);\n+let mut reader = PullParser::new(input);\n","diffStat":{"added":1,"removed":1}}}]}
+{"id":"msg-sa-05","role":"user","threadId":"main","text":"Sub-thread done, looks good.","timestamp":"2026-06-22T16:02:00.000Z"}
diff --git a/fixtures/gemini/v1/tool_failure.jsonl b/fixtures/gemini/v1/tool_failure.jsonl
new file mode 100644
index 0000000..c14fc9c
--- /dev/null
+++ b/fixtures/gemini/v1/tool_failure.jsonl
@@ -0,0 +1,4 @@
+{"$set":{"sessionId":"a1b2c3d4-0004","cwd":"/home/dev/service","model":"gemini-2.5-pro","branch":"main","commit":"abc0001234","cliVersion":"0.4.1"}}
+{"id":"msg-0301","role":"user","text":"Add a rate limiter to the gateway handler.","timestamp":"2026-06-22T13:45:00.000Z"}
+{"id":"msg-0302","role":"model","text":"I'll add a token-bucket limiter to the handler.","thoughts":"Edit gateway.rs to insert the limiter.","tokens":{"input":95,"output":40},"timestamp":"2026-06-22T13:45:03.000Z","toolCalls":[{"callId":"call-edit-fail","name":"replace","args":{"path":"src/gateway.rs"},"resultDisplay":{"fileName":"src/gateway.rs","originalContent":"fn handle() {}\n","newContent":"fn handle() { limiter.check(); }\n","fileDiff":"@@ -1 +1 @@\n-fn handle() {}\n+fn handle() { limiter.check(); }\n","diffStat":{"added":1,"removed":1},"error":"write failed: read-only file system"}}]}
+{"id":"msg-0303","role":"model","text":"The edit failed because the file system is read-only. Please re-mount with write access and I'll retry.","tokens":{"input":60,"output":33},"timestamp":"2026-06-22T13:45:05.000Z"}
diff --git a/fixtures/otel/genai/ban.jsonl b/fixtures/otel/genai/ban.jsonl
new file mode 100644
index 0000000..bf3de94
--- /dev/null
+++ b/fixtures/otel/genai/ban.jsonl
@@ -0,0 +1,3 @@
+{"time":"2026-06-22T12:00:00Z","gen_ai.system":"gemini","gen_ai.operation.name":"session.start","gen_ai.conversation.id":"sess-ban-001","gen_ai.request.model":"gemini-2.5-pro","cwd":"/home/dev/core","repo_root":"/home/dev/core","git.commit":"9abcdef","git.branch":"main"}
+{"time":"2026-06-22T12:00:06Z","gen_ai.system":"gemini","gen_ai.operation.name":"chat","gen_ai.conversation.id":"sess-ban-001","gen_ai.cli.user_prompt":"We will never add a dependency on left-pad. Keep the tree clean."}
+{"time":"2026-06-22T12:00:10Z","gen_ai.system":"gemini","gen_ai.operation.name":"chat","gen_ai.conversation.id":"sess-ban-001","gen_ai.request.model":"gemini-2.5-pro","gen_ai.usage.input_tokens":90,"gen_ai.usage.output_tokens":18,"gen_ai.output.messages":[{"role":"assistant","content":"Acknowledged. No left-pad dependency will be added."}]}
diff --git a/fixtures/otel/genai/happy_path_decision_then_edits.jsonl b/fixtures/otel/genai/happy_path_decision_then_edits.jsonl
new file mode 100644
index 0000000..761bd73
--- /dev/null
+++ b/fixtures/otel/genai/happy_path_decision_then_edits.jsonl
@@ -0,0 +1,6 @@
+{"time":"2026-06-22T10:00:00Z","gen_ai.system":"anthropic","gen_ai.operation.name":"session.start","gen_ai.conversation.id":"sess-happy-001","gen_ai.request.model":"claude-opus-4-8","cwd":"/home/dev/svc","repo_root":"/home/dev/svc","git.commit":"abc1234","git.branch":"main"}
+{"time":"2026-06-22T10:00:05Z","gen_ai.system":"anthropic","gen_ai.operation.name":"chat","gen_ai.conversation.id":"sess-happy-001","gen_ai.cli.user_prompt":"Let's use Postgres instead of MySQL for the persistence layer."}
+{"time":"2026-06-22T10:00:09Z","gen_ai.system":"anthropic","gen_ai.operation.name":"chat","gen_ai.conversation.id":"sess-happy-001","gen_ai.request.model":"claude-opus-4-8","gen_ai.usage.input_tokens":420,"gen_ai.usage.output_tokens":88,"gen_ai.output.messages":[{"role":"assistant","content":"Agreed. I'll switch the persistence layer to Postgres and update the config."}]}
+{"time":"2026-06-22T10:00:12Z","gen_ai.system":"anthropic","gen_ai.operation.name":"execute_tool","gen_ai.conversation.id":"sess-happy-001","gen_ai.tool.name":"edit_file","gen_ai.tool.call.id":"call-edit-1","file.path":"db/config.rs","model_added_lines":12,"model_removed_lines":4,"gen_ai.tool.result":"ok"}
+{"time":"2026-06-22T10:00:15Z","gen_ai.system":"anthropic","gen_ai.operation.name":"file_operation","gen_ai.conversation.id":"sess-happy-001","file.path":"db/pool.rs","code.added_lines":30,"code.removed_lines":9}
+{"time":"2026-06-22T10:00:20Z","gen_ai.system":"anthropic","gen_ai.operation.name":"session.end","gen_ai.conversation.id":"sess-happy-001","reason":"completed"}
diff --git a/fixtures/otel/genai/rejected_alternative.jsonl b/fixtures/otel/genai/rejected_alternative.jsonl
new file mode 100644
index 0000000..54eda06
--- /dev/null
+++ b/fixtures/otel/genai/rejected_alternative.jsonl
@@ -0,0 +1,4 @@
+{"time":"2026-06-22T11:00:00Z","gen_ai.system":"openai","gen_ai.operation.name":"session.start","gen_ai.conversation.id":"sess-reject-001","gen_ai.request.model":"gpt-5","cwd":"/home/dev/checkout","repo_root":"/home/dev/checkout","git.commit":"def5678","git.branch":"feat/payments"}
+{"time":"2026-06-22T11:00:04Z","gen_ai.system":"openai","gen_ai.operation.name":"chat","gen_ai.conversation.id":"sess-reject-001","gen_ai.input.messages":[{"role":"user","content":"We should use Stripe instead of PayPal for the checkout integration."}]}
+{"time":"2026-06-22T11:00:08Z","gen_ai.system":"openai","gen_ai.operation.name":"chat","gen_ai.conversation.id":"sess-reject-001","gen_ai.request.model":"gpt-5","gen_ai.usage.input_tokens":210,"gen_ai.usage.output_tokens":40,"gen_ai.output.messages":[{"role":"assistant","content":"Understood, I'll wire up Stripe and drop the PayPal path."}]}
+{"time":"2026-06-22T11:00:12Z","gen_ai.system":"openai","gen_ai.operation.name":"file_operation","gen_ai.conversation.id":"sess-reject-001","file.path":"checkout/stripe.rs","code.added_lines":48,"code.removed_lines":0}
diff --git a/fixtures/otel/genai/tool_failure.jsonl b/fixtures/otel/genai/tool_failure.jsonl
new file mode 100644
index 0000000..a02236b
--- /dev/null
+++ b/fixtures/otel/genai/tool_failure.jsonl
@@ -0,0 +1,3 @@
+{"time":"2026-06-22T13:00:00Z","gen_ai.system":"anthropic","gen_ai.operation.name":"session.start","gen_ai.conversation.id":"sess-fail-001","gen_ai.request.model":"claude-opus-4-8","cwd":"/home/dev/svc","repo_root":"/home/dev/svc","git.commit":"badf00d","git.branch":"main"}
+{"time":"2026-06-22T13:00:05Z","gen_ai.system":"anthropic","gen_ai.operation.name":"chat","gen_ai.conversation.id":"sess-fail-001","gen_ai.cli.user_prompt":"Switch the cache layer over to Redis."}
+{"time":"2026-06-22T13:00:09Z","gen_ai.system":"anthropic","gen_ai.operation.name":"execute_tool","gen_ai.conversation.id":"sess-fail-001","gen_ai.tool.name":"edit_file","gen_ai.tool.call.id":"call-edit-fail","file.path":"cache/redis.rs","model_added_lines":20,"model_removed_lines":2,"gen_ai.tool.result.status":"error","error.type":"PatchConflict","gen_ai.tool.result":"hunk failed to apply: context mismatch"}
diff --git a/fixtures/vscode/v1/ban.jsonl b/fixtures/vscode/v1/ban.jsonl
new file mode 100644
index 0000000..c7302df
--- /dev/null
+++ b/fixtures/vscode/v1/ban.jsonl
@@ -0,0 +1,3 @@
+{"kind":"session_start","sessionId":"vsc-sess-3","cwd":"/home/dev/acme-api","git":{"sha":"0011223344556677","branch":"main"},"toolVersion":"1.92.0","model":"gpt-4o"}
+{"id":"b1","parentId":null,"role":"user","ts":"2026-06-22T12:00:00Z","sessionId":"vsc-sess-3","text":"We will never add a dependency on left-pad. Roll our own padding helper."}
+{"id":"b2","parentId":"b1","role":"assistant","ts":"2026-06-22T12:00:06Z","sessionId":"vsc-sess-3","text":"Understood — no left-pad. I'll add a tiny internal pad() utility.","model":"gpt-4o","usage":{"input":640,"output":120},"edits":[{"path":"src/util/pad.ts","oldText":"","newText":"export const pad = (s, n) => s.padStart(n, '0');","added":1,"removed":0}]}
diff --git a/fixtures/vscode/v1/happy_path_decision_then_edits.jsonl b/fixtures/vscode/v1/happy_path_decision_then_edits.jsonl
new file mode 100644
index 0000000..41a4ebc
--- /dev/null
+++ b/fixtures/vscode/v1/happy_path_decision_then_edits.jsonl
@@ -0,0 +1,5 @@
+{"kind":"session_start","sessionId":"vsc-sess-1","cwd":"/home/dev/acme-api","git":{"sha":"a1b2c3d4e5f6","branch":"main"},"toolVersion":"1.92.0","model":"gpt-4o"}
+{"id":"m1","parentId":null,"role":"user","ts":"2026-06-22T10:00:00Z","sessionId":"vsc-sess-1","text":"Let's use Postgres instead of MySQL for the persistence layer."}
+{"id":"m2","parentId":"m1","role":"assistant","ts":"2026-06-22T10:00:05Z","sessionId":"vsc-sess-1","text":"Good call. I'll switch the connection string and update the schema module.","model":"gpt-4o","usage":{"input":1200,"output":340},"edits":[{"path":"src/db/connect.ts","oldText":"const url = 'mysql://localhost/acme';","newText":"const url = 'postgres://localhost/acme';","added":1,"removed":1},{"path":"src/db/schema.sql","oldText":"ENGINE=InnoDB","newText":"-- postgres has no storage engines","diff":"@@ -1 +1 @@\n-ENGINE=InnoDB\n+-- postgres has no storage engines","added":1,"removed":1}]}
+{"id":"m3","parentId":"m2","role":"user","ts":"2026-06-22T10:01:00Z","sessionId":"vsc-sess-1","text":"Looks good, thanks."}
+{"kind":"session_end","sessionId":"vsc-sess-1","ts":"2026-06-22T10:02:00Z","reason":"user_closed"}
diff --git a/fixtures/vscode/v1/native_chatsession.jsonl b/fixtures/vscode/v1/native_chatsession.jsonl
new file mode 100644
index 0000000..6d5ba2b
--- /dev/null
+++ b/fixtures/vscode/v1/native_chatsession.jsonl
@@ -0,0 +1 @@
+{"version":3,"sessionId":"vsc-native-1","requesterUsername":"dev","responderUsername":"GitHub Copilot","requests":[{"message":{"text":"Add a health check endpoint","parts":[{"kind":"text","text":"Add a health check endpoint"}]},"response":[{"kind":"text","text":"I'll add a GET /healthz route that returns 200."},{"kind":"text","text":"Done."}]}]}
diff --git a/fixtures/vscode/v1/rejected_alternative.jsonl b/fixtures/vscode/v1/rejected_alternative.jsonl
new file mode 100644
index 0000000..cf42060
--- /dev/null
+++ b/fixtures/vscode/v1/rejected_alternative.jsonl
@@ -0,0 +1,4 @@
+{"kind":"session_start","sessionId":"vsc-sess-2","cwd":"/home/dev/acme-api","git":{"sha":"f6e5d4c3b2a1","branch":"feat/cache"},"toolVersion":"1.92.0","model":"gpt-4o"}
+{"id":"r1","parentId":null,"role":"user","ts":"2026-06-22T11:00:00Z","sessionId":"vsc-sess-2","text":"Should we use Redis or just an in-memory LRU for the cache?"}
+{"id":"r2","parentId":"r1","role":"assistant","ts":"2026-06-22T11:00:08Z","sessionId":"vsc-sess-2","text":"We considered Redis but rejected it: it adds an external dependency we don't want for this small cache. We'll go with an in-memory LRU instead.","model":"gpt-4o","usage":{"input":900,"output":210}}
+{"id":"r3","parentId":"r2","role":"assistant","ts":"2026-06-22T11:00:20Z","sessionId":"vsc-sess-2","text":"Implementing the LRU now.","model":"gpt-4o","edits":[{"path":"src/cache/lru.ts","oldText":"","newText":"export class Lru { /* ... */ }","added":1,"removed":0}]}
diff --git a/fixtures/vscode/v1/tool_failure.jsonl b/fixtures/vscode/v1/tool_failure.jsonl
new file mode 100644
index 0000000..334dae7
--- /dev/null
+++ b/fixtures/vscode/v1/tool_failure.jsonl
@@ -0,0 +1,4 @@
+{"kind":"session_start","sessionId":"vsc-sess-4","cwd":"/home/dev/acme-api","git":{"sha":"99aabbccddeeff00","branch":"main"},"toolVersion":"1.92.0","model":"gpt-4o"}
+{"id":"t1","parentId":null,"role":"user","ts":"2026-06-22T13:00:00Z","sessionId":"vsc-sess-4","text":"Rename the config flag enableV2 to enableV2Beta everywhere."}
+{"id":"t2","parentId":"t1","role":"assistant","ts":"2026-06-22T13:00:07Z","sessionId":"vsc-sess-4","text":"I'll apply the edit to src/config.ts.","model":"gpt-4o","usage":{"input":700,"output":160},"toolCalls":[{"id":"call-edit-1","name":"applyEdit","args":{"path":"src/config.ts"}}],"edits":[{"path":"src/config.ts","callId":"call-edit-1","oldText":"enableV2","newText":"enableV2Beta","added":1,"removed":1}]}
+{"id":"t3","parentId":"t2","role":"assistant","ts":"2026-06-22T13:00:09Z","sessionId":"vsc-sess-4","text":"","toolResults":[{"id":"call-edit-1","ok":false,"output":"EDIT FAILED: file src/config.ts is read-only on disk"}]}
diff --git a/fixtures/windsurf/v1/ban.jsonl b/fixtures/windsurf/v1/ban.jsonl
new file mode 100644
index 0000000..76905ed
--- /dev/null
+++ b/fixtures/windsurf/v1/ban.jsonl
@@ -0,0 +1,3 @@
+{"kind":"session_start","sessionId":"ws-ban-003","cwd":"/Users/dev/acme-api","git":{"sha":"77c0de01","branch":"main"},"toolVersion":"1.8.2","model":"cascade-base"}
+{"id":"u1","parentId":null,"role":"user","ts":"2026-06-22T16:20:00Z","sessionId":"ws-ban-003","text":"We will never add a dependency on left-pad. Implement padding inline instead."}
+{"id":"a1","parentId":"u1","role":"assistant","ts":"2026-06-22T16:20:09Z","sessionId":"ws-ban-003","text":"Understood — no left-pad dependency. I'll add a small inline pad helper.","model":"cascade-base","usage":{"input":300,"output":140},"toolCalls":[{"id":"call_pad_1","name":"edit_file","args":{"path":"src/util/pad.rs"}}],"toolResults":[{"id":"call_pad_1","ok":true,"output":"applied"}],"edits":[{"id":"call_pad_1","path":"src/util/pad.rs","oldText":"","newText":"pub fn pad_left(s: &str, width: usize) -> String { format!(\"{s:>width$}\") }","diff":"@@ -0,0 +1,1 @@\n+pub fn pad_left(s: &str, width: usize) -> String { format!(\"{s:>width$}\") }","added":1,"removed":0}]}
diff --git a/fixtures/windsurf/v1/happy_path_decision_then_edits.jsonl b/fixtures/windsurf/v1/happy_path_decision_then_edits.jsonl
new file mode 100644
index 0000000..608c392
--- /dev/null
+++ b/fixtures/windsurf/v1/happy_path_decision_then_edits.jsonl
@@ -0,0 +1,4 @@
+{"kind":"session_start","sessionId":"ws-happy-001","cwd":"/Users/dev/acme-api","git":{"sha":"9f3c1a7e","branch":"main"},"toolVersion":"1.8.2","model":"cascade-base"}
+{"id":"u1","parentId":null,"role":"user","ts":"2026-06-22T14:00:00Z","sessionId":"ws-happy-001","text":"Let's use Postgres instead of MySQL for the orders service."}
+{"id":"a1","parentId":"u1","role":"assistant","ts":"2026-06-22T14:00:06Z","sessionId":"ws-happy-001","text":"Good call. I'll swap the driver and update the connection string.","model":"cascade-base","usage":{"input":420,"output":188},"toolCalls":[{"id":"call_db_1","name":"edit_file","args":{"path":"src/db/mod.rs"}},{"id":"call_cfg_1","name":"edit_file","args":{"path":"config/database.toml"}}],"toolResults":[{"id":"call_db_1","ok":true,"output":"applied"},{"id":"call_cfg_1","ok":true,"output":"applied"}],"edits":[{"id":"call_db_1","path":"src/db/mod.rs","oldText":"use mysql_async::Pool;","newText":"use sqlx::postgres::PgPool;","diff":"@@ -1,1 +1,1 @@\n-use mysql_async::Pool;\n+use sqlx::postgres::PgPool;","added":1,"removed":1},{"id":"call_cfg_1","path":"config/database.toml","oldText":"driver = \"mysql\"\nport = 3306","newText":"driver = \"postgres\"\nport = 5432","diff":"@@ -1,2 +1,2 @@\n-driver = \"mysql\"\n-port = 3306\n+driver = \"postgres\"\n+port = 5432","added":2,"removed":2}]}
+{"id":"u2","parentId":"a1","role":"user","ts":"2026-06-22T14:01:10Z","sessionId":"ws-happy-001","text":"Looks right, thanks."}
diff --git a/fixtures/windsurf/v1/rejected_alternative.jsonl b/fixtures/windsurf/v1/rejected_alternative.jsonl
new file mode 100644
index 0000000..476d707
--- /dev/null
+++ b/fixtures/windsurf/v1/rejected_alternative.jsonl
@@ -0,0 +1,5 @@
+{"kind":"session_start","sessionId":"ws-reject-002","cwd":"/Users/dev/acme-api","git":{"sha":"1b2d44aa","branch":"feature/cache"},"toolVersion":"1.8.2","model":"cascade-pro"}
+{"id":"u1","parentId":null,"role":"user","ts":"2026-06-22T15:10:00Z","sessionId":"ws-reject-002","text":"Should we use Redis or Memcached for the session cache?"}
+{"id":"a1","parentId":"u1","role":"assistant","ts":"2026-06-22T15:10:08Z","sessionId":"ws-reject-002","text":"I considered Memcached, but we already run Redis for the queue, so I'll reuse Redis and skip Memcached to avoid a second dependency.","model":"cascade-pro","usage":{"input":510,"output":233}}
+{"id":"u2","parentId":"a1","role":"user","ts":"2026-06-22T15:11:30Z","sessionId":"ws-reject-002","text":"Agreed, do not add Memcached. Wire up Redis."}
+{"id":"a2","parentId":"u2","role":"assistant","ts":"2026-06-22T15:11:40Z","sessionId":"ws-reject-002","text":"Wiring Redis as the session store.","model":"cascade-pro","usage":{"input":540,"output":120},"toolCalls":[{"id":"call_cache_1","name":"edit_file","args":{"path":"src/cache.rs"}}],"toolResults":[{"id":"call_cache_1","ok":true,"output":"applied"}],"edits":[{"id":"call_cache_1","path":"src/cache.rs","oldText":"// TODO: choose cache backend","newText":"let client = redis::Client::open(cfg.redis_url)?;","diff":"@@ -1,1 +1,1 @@\n-// TODO: choose cache backend\n+let client = redis::Client::open(cfg.redis_url)?;","added":1,"removed":1}]}
diff --git a/fixtures/windsurf/v1/tool_failure.jsonl b/fixtures/windsurf/v1/tool_failure.jsonl
new file mode 100644
index 0000000..645a4b2
--- /dev/null
+++ b/fixtures/windsurf/v1/tool_failure.jsonl
@@ -0,0 +1,3 @@
+{"kind":"session_start","sessionId":"ws-fail-004","cwd":"/Users/dev/acme-api","git":{"sha":"deadbeef","branch":"main"},"toolVersion":"1.8.2","model":"cascade-base"}
+{"id":"u1","parentId":null,"role":"user","ts":"2026-06-22T17:05:00Z","sessionId":"ws-fail-004","text":"Update the read-only vendored file to bump the version."}
+{"id":"a1","parentId":"u1","role":"assistant","ts":"2026-06-22T17:05:07Z","sessionId":"ws-fail-004","text":"Attempting the edit now.","model":"cascade-base","usage":{"input":260,"output":90},"toolCalls":[{"id":"call_fail_1","name":"edit_file","args":{"path":"vendor/locked.toml"}}],"toolResults":[{"id":"call_fail_1","ok":false,"output":"EACCES: permission denied, open 'vendor/locked.toml'"}],"edits":[{"id":"call_fail_1","path":"vendor/locked.toml","oldText":"version = \"1.0.0\"","newText":"version = \"1.0.1\"","diff":"@@ -1,1 +1,1 @@\n-version = \"1.0.0\"\n+version = \"1.0.1\"","added":1,"removed":1}]}
diff --git a/fixtures/zed/v1/ban.jsonl b/fixtures/zed/v1/ban.jsonl
new file mode 100644
index 0000000..594cd2e
--- /dev/null
+++ b/fixtures/zed/v1/ban.jsonl
@@ -0,0 +1,4 @@
+{"kind":"session_start","cwd":"/Users/dev/projects/orbit","git":{"sha":"0011223344556677","branch":"main"},"toolVersion":"zed 0.182.0","sessionId":"zed-thread-003","ts":"2026-06-22T12:00:00Z"}
+{"id":"msg-1","parentId":null,"role":"user","ts":"2026-06-22T12:00:03Z","sessionId":"zed-thread-003","text":"We will never add a dependency on left-pad. Use the std library instead."}
+{"id":"msg-2","parentId":"msg-1","role":"assistant","ts":"2026-06-22T12:00:07Z","sessionId":"zed-thread-003","text":"Understood — no left-pad dependency. I'll pad with format! from std.","model":"claude-opus-4-8","usage":{"input":640,"output":120},"toolCalls":[{"id":"call-1","name":"read_file","args":{"path":"Cargo.toml"}}]}
+{"id":"msg-3","parentId":"msg-2","role":"assistant","ts":"2026-06-22T12:00:10Z","sessionId":"zed-thread-003","text":"Confirmed Cargo.toml has no left-pad.","toolResults":[{"id":"call-1","ok":true,"output":"[dependencies]\nserde = \"1\"\n"}]}
diff --git a/fixtures/zed/v1/happy_path_decision_then_edits.jsonl b/fixtures/zed/v1/happy_path_decision_then_edits.jsonl
new file mode 100644
index 0000000..aab0f37
--- /dev/null
+++ b/fixtures/zed/v1/happy_path_decision_then_edits.jsonl
@@ -0,0 +1,5 @@
+{"kind":"session_start","cwd":"/Users/dev/projects/orbit","git":{"sha":"a1b2c3d4e5f6","branch":"main"},"toolVersion":"zed 0.182.0","sessionId":"zed-thread-001","ts":"2026-06-22T10:00:00Z"}
+{"id":"msg-1","parentId":null,"role":"user","ts":"2026-06-22T10:00:05Z","sessionId":"zed-thread-001","text":"Let's use Postgres instead of MySQL for the persistence layer."}
+{"id":"msg-2","parentId":"msg-1","role":"assistant","ts":"2026-06-22T10:00:09Z","sessionId":"zed-thread-001","text":"Good call. I'll switch the connection config and the schema module to Postgres.","model":"claude-opus-4-8","usage":{"input":1200,"output":340},"edits":[{"path":"src/db/config.rs","oldText":"let url = \"mysql://localhost/orbit\";","newText":"let url = \"postgres://localhost/orbit\";","diff":"@@ -1 +1 @@\n-let url = \"mysql://localhost/orbit\";\n+let url = \"postgres://localhost/orbit\";","added":1,"removed":1},{"path":"src/db/schema.rs","oldText":"// MySQL dialect\n","newText":"// Postgres dialect\nuse sqlx::postgres::PgPool;\n","added":2,"removed":1}]}
+{"id":"msg-3","parentId":"msg-2","role":"user","ts":"2026-06-22T10:01:00Z","sessionId":"zed-thread-001","text":"Looks right, thanks."}
+{"kind":"session_end","sessionId":"zed-thread-001","ts":"2026-06-22T10:01:30Z","reason":"user_closed"}
diff --git a/fixtures/zed/v1/rejected_alternative.jsonl b/fixtures/zed/v1/rejected_alternative.jsonl
new file mode 100644
index 0000000..8bf29ce
--- /dev/null
+++ b/fixtures/zed/v1/rejected_alternative.jsonl
@@ -0,0 +1,5 @@
+{"kind":"session_start","cwd":"/Users/dev/projects/orbit","git":{"sha":"f6e5d4c3b2a1","branch":"feature/caching"},"toolVersion":"zed 0.182.0","sessionId":"zed-thread-002","ts":"2026-06-22T11:00:00Z"}
+{"id":"msg-1","parentId":null,"role":"user","ts":"2026-06-22T11:00:04Z","sessionId":"zed-thread-002","text":"Should we add Redis for the cache layer?"}
+{"id":"msg-2","parentId":"msg-1","role":"assistant","ts":"2026-06-22T11:00:08Z","sessionId":"zed-thread-002","text":"We considered Redis but rejected it: the working set fits in process memory, so an in-process LRU is simpler. Not adding Redis.","model":"claude-opus-4-8","usage":{"input":900,"output":210}}
+{"id":"msg-3","parentId":"msg-2","role":"user","ts":"2026-06-22T11:00:40Z","sessionId":"zed-thread-002","text":"Agreed, go with the in-process LRU."}
+{"id":"msg-4","parentId":"msg-3","role":"assistant","ts":"2026-06-22T11:00:46Z","sessionId":"zed-thread-002","text":"Done — added a bounded LRU.","model":"claude-opus-4-8","edits":[{"path":"src/cache/lru.rs","oldText":"","newText":"pub struct Lru { cap: usize }\n","diff":"@@ -0,0 +1 @@\n+pub struct Lru { cap: usize }","added":1,"removed":0}]}
diff --git a/fixtures/zed/v1/tool_failure.jsonl b/fixtures/zed/v1/tool_failure.jsonl
new file mode 100644
index 0000000..559afb2
--- /dev/null
+++ b/fixtures/zed/v1/tool_failure.jsonl
@@ -0,0 +1,4 @@
+{"kind":"session_start","cwd":"/Users/dev/projects/orbit","git":{"sha":"99aa88bb77cc","branch":"main"},"toolVersion":"zed 0.182.0","sessionId":"zed-thread-004","ts":"2026-06-22T13:00:00Z"}
+{"id":"msg-1","parentId":null,"role":"user","ts":"2026-06-22T13:00:02Z","sessionId":"zed-thread-004","text":"Rename the config field timeout_ms to timeout_millis everywhere."}
+{"id":"msg-2","parentId":"msg-1","role":"assistant","ts":"2026-06-22T13:00:06Z","sessionId":"zed-thread-004","text":"Applying the rename.","model":"claude-opus-4-8","toolCalls":[{"id":"call-9","name":"edit_file","args":{"path":"src/config.rs"}}],"edits":[{"path":"src/config.rs","oldText":"timeout_ms: u64,","newText":"timeout_millis: u64,","diff":"@@ -1 +1 @@\n-timeout_ms: u64,\n+timeout_millis: u64,","added":1,"removed":1}]}
+{"id":"msg-3","parentId":"msg-2","role":"assistant","ts":"2026-06-22T13:00:08Z","sessionId":"zed-thread-004","text":"The edit failed — the file was locked by another process.","toolResults":[{"id":"call-9","ok":false,"output":"error: could not write src/config.rs: resource temporarily unavailable"}]}
diff --git a/fuzz/.gitignore b/fuzz/.gitignore
new file mode 100644
index 0000000..8e1edae
--- /dev/null
+++ b/fuzz/.gitignore
@@ -0,0 +1,6 @@
+# cargo-fuzz build/run artifacts.
+target
+corpus
+artifacts
+coverage
+Cargo.lock
diff --git a/fuzz/Cargo.toml b/fuzz/Cargo.toml
new file mode 100644
index 0000000..7f20a13
--- /dev/null
+++ b/fuzz/Cargo.toml
@@ -0,0 +1,113 @@
+# Memscribe cargo-fuzz crate (whitepaper §8.4).
+#
+# This is a standalone crate, EXCLUDED from the root workspace (see the root
+# Cargo.toml `[workspace] exclude = ["fuzz"]`). It is never built by a plain
+# `cargo build`/`cargo test` of the workspace; you run it explicitly:
+#
+#     cargo +nightly fuzz run claude_code
+#
+# One target per adapter parser. Each target builds the adapter, wraps the fuzz
+# bytes in a `RawRecord`, and calls `parse()` with a fresh `ParseCtx`, asserting
+# the parser never panics and terminates (the libFuzzer harness catches panics).
+#
+# The targets also compile under a plain stable `cargo build --manifest-path
+# fuzz/Cargo.toml` (no nightly required) so CI can smoke-test that the structure
+# is intact: outside cargo-fuzz (`--cfg fuzzing` unset) each `[[bin]]` is a tiny
+# stub `main` that exercises the same shared `run()` once. The real libFuzzer
+# entry point is only wired up under `cfg(fuzzing)`, which cargo-fuzz sets.
+
+[package]
+name = "memscribe-fuzz"
+version = "0.0.0"
+edition = "2021"
+rust-version = "1.96"
+license = "MIT OR Apache-2.0"
+publish = false
+
+[package.metadata]
+cargo-fuzz = true
+
+[dependencies]
+libfuzzer-sys = "0.4"
+arbitrary = { version = "1", features = ["derive"] }
+memscribe-adapters = { path = "../crates/memscribe-adapters" }
+memscribe-core = { path = "../crates/memscribe-core" }
+
+# Prevent this crate from being mistaken for a workspace root of its own and
+# accidentally pulling the parent in.
+[workspace]
+
+[profile.release]
+debug = 1
+
+# `cargo-fuzz` compiles these targets with `--cfg fuzzing`; a plain stable
+# `cargo build` does not. Declare the cfg as known so the non-fuzzing build is
+# warning-clean instead of flagging it as an `unexpected_cfgs` condition.
+[lints.rust]
+unexpected_cfgs = { level = "warn", check-cfg = ['cfg(fuzzing)'] }
+
+# One fuzz target per adapter parser. `test = false` / `doc = false` keep these
+# binaries out of `cargo test`/`cargo doc` (they have no unit tests or docs).
+
+[[bin]]
+name = "claude_code"
+path = "fuzz_targets/claude_code.rs"
+test = false
+doc = false
+bench = false
+
+[[bin]]
+name = "codex"
+path = "fuzz_targets/codex.rs"
+test = false
+doc = false
+bench = false
+
+[[bin]]
+name = "gemini"
+path = "fuzz_targets/gemini.rs"
+test = false
+doc = false
+bench = false
+
+[[bin]]
+name = "otel"
+path = "fuzz_targets/otel.rs"
+test = false
+doc = false
+bench = false
+
+[[bin]]
+name = "cursor"
+path = "fuzz_targets/cursor.rs"
+test = false
+doc = false
+bench = false
+
+[[bin]]
+name = "windsurf"
+path = "fuzz_targets/windsurf.rs"
+test = false
+doc = false
+bench = false
+
+[[bin]]
+name = "zed"
+path = "fuzz_targets/zed.rs"
+test = false
+doc = false
+bench = false
+
+[[bin]]
+name = "vscode"
+path = "fuzz_targets/vscode.rs"
+test = false
+doc = false
+bench = false
+
+[[bin]]
+name = "copilot"
+path = "fuzz_targets/copilot.rs"
+test = false
+doc = false
+bench = false
diff --git a/fuzz/README.md b/fuzz/README.md
new file mode 100644
index 0000000..9b62f3b
--- /dev/null
+++ b/fuzz/README.md
@@ -0,0 +1,96 @@
+# memscribe-fuzz
+
+Coverage-guided fuzz targets for every Memscribe adapter parser (implementation
+whitepaper §8.4). The contract every adapter must uphold is simple and strict:
+
+> A parser **must never panic** and must **terminate**. Unrecognized-but-well-
+> formed records route to `EventKind::Unknown`; only genuinely malformed bytes
+> return `ParseError`. Neither outcome may crash the stream.
+
+These targets exercise exactly that: each one builds its adapter, wraps the raw
+fuzz bytes in a `RawRecord`, and calls `parse()` (and `schema_fingerprint()`)
+with a fresh `ParseCtx`. libFuzzer turns any panic, hang, or OOM into a crash
+artifact you can replay.
+
+## Targets
+
+One target per adapter parser:
+
+| target        | adapter                                      |
+| ------------- | -------------------------------------------- |
+| `claude_code` | `memscribe_adapters::claude_code::ClaudeCodeAdapter` |
+| `codex`       | `memscribe_adapters::codex::CodexAdapter`            |
+| `gemini`      | `memscribe_adapters::gemini::GeminiAdapter`          |
+| `otel`        | `memscribe_adapters::otel::OtelAdapter`              |
+| `cursor`      | `memscribe_adapters::cursor::CursorAdapter`          |
+| `windsurf`    | `memscribe_adapters::windsurf::WindsurfAdapter`      |
+| `zed`         | `memscribe_adapters::zed::ZedAdapter`                |
+| `vscode`      | `memscribe_adapters::vscode::VsCodeAdapter`          |
+| `copilot`     | `memscribe_adapters::copilot::CopilotAdapter`        |
+
+## Layout
+
+This crate is **excluded** from the root Cargo workspace (`[workspace] exclude =
+["fuzz"]` in the repo-root `Cargo.toml`). That keeps `libfuzzer-sys` — which
+needs a nightly toolchain and sanitizer flags — out of a plain `cargo build` /
+`cargo test` of the workspace.
+
+## Prerequisites
+
+`cargo-fuzz` runs on a **nightly** toolchain:
+
+```sh
+rustup toolchain install nightly
+cargo install cargo-fuzz
+```
+
+## Running
+
+From the repository root:
+
+```sh
+# Build all targets (compile-only smoke test).
+cargo +nightly fuzz build
+
+# Fuzz a single adapter parser.
+cargo +nightly fuzz run claude_code
+
+# Time-box a run (recommended for CI).
+cargo +nightly fuzz run codex -- -max_total_time=60
+
+# List every target.
+cargo +nightly fuzz list
+```
+
+Crash-reproducing inputs land in `fuzz/artifacts/<target>/`; replay one with:
+
+```sh
+cargo +nightly fuzz run claude_code fuzz/artifacts/claude_code/crash-<hash>
+```
+
+## Building without nightly (CI structural check)
+
+The targets are written so that **plain stable `cargo build` also compiles
+them**, which lets CI verify the structure is intact even on a runner without
+nightly or `cargo-fuzz`:
+
+```sh
+cargo build --manifest-path fuzz/Cargo.toml
+```
+
+Outside `cargo-fuzz`, the `fuzzing` cfg is unset, so each `[[bin]]` becomes a
+tiny stub `main` that exercises the same shared `run()` helper once on an empty
+input instead of wiring up the libFuzzer runtime. The actual `fuzz_target!`
+entry point is only compiled under `#[cfg(fuzzing)]`, which `cargo-fuzz` sets.
+
+## Relationship to the non-nightly robustness suite
+
+`crates/memscribe-testkit/tests/robustness.rs` is the workspace-resident, no-
+nightly counterpart: a `proptest` that feeds mutated and adversarial bytes
+(random, truncated JSON, deeply nested JSON, huge numbers, invalid UTF-8) to
+every adapter and asserts no panic, bounded time, and that a malformed line is
+skipped/`Unknown` rather than aborting the stream. Run it with:
+
+```sh
+cargo test -p memscribe-testkit --test robustness
+```
diff --git a/fuzz/fuzz_targets/claude_code.rs b/fuzz/fuzz_targets/claude_code.rs
new file mode 100644
index 0000000..0b23053
--- /dev/null
+++ b/fuzz/fuzz_targets/claude_code.rs
@@ -0,0 +1,37 @@
+//! Fuzz target for the Claude Code adapter parser (whitepaper §8.4).
+//!
+//! Feeds arbitrary bytes to `ClaudeCodeAdapter::parse` through a `RawRecord`
+//! with a fresh `ParseCtx`, asserting the parser never panics and terminates.
+//! The libFuzzer harness turns any panic into a crash artifact.
+#![cfg_attr(fuzzing, no_main)]
+
+use memscribe_adapters::claude_code::ClaudeCodeAdapter;
+use memscribe_core::model::SourceLocation;
+use memscribe_core::{ParseCtx, RawRecord, TranscriptAdapter};
+
+/// Drive one fuzz input through the adapter. A parser is allowed to return
+/// `Ok(events)` or `Err(ParseError)`; the only contract a fuzz run enforces is
+/// that it neither panics nor diverges. We deliberately ignore the result.
+#[inline]
+fn run(data: &[u8]) {
+    let adapter = ClaudeCodeAdapter;
+    let loc = SourceLocation::new("fuzz://claude_code", 0, 1);
+    let raw = RawRecord::new(data.to_vec(), loc);
+    let mut ctx = ParseCtx::new();
+    let _ = adapter.parse(&raw, &mut ctx);
+    // Fingerprinting shares the same parse-tolerance contract.
+    let _ = adapter.schema_fingerprint(&raw);
+}
+
+#[cfg(fuzzing)]
+libfuzzer_sys::fuzz_target!(|data: &[u8]| {
+    run(data);
+});
+
+// Plain `cargo build` (no `--cfg fuzzing`): a tiny stub `main` so the target
+// compiles and links on stable without the libFuzzer runtime, and exercises the
+// same code path once on an empty input.
+#[cfg(not(fuzzing))]
+fn main() {
+    run(b"");
+}
diff --git a/fuzz/fuzz_targets/codex.rs b/fuzz/fuzz_targets/codex.rs
new file mode 100644
index 0000000..e8b0cae
--- /dev/null
+++ b/fuzz/fuzz_targets/codex.rs
@@ -0,0 +1,30 @@
+//! Fuzz target for the Codex adapter parser (whitepaper §8.4).
+//!
+//! Feeds arbitrary bytes to `CodexAdapter::parse` through a `RawRecord` with a
+//! fresh `ParseCtx`, asserting the parser never panics and terminates. The
+//! libFuzzer harness turns any panic into a crash artifact.
+#![cfg_attr(fuzzing, no_main)]
+
+use memscribe_adapters::codex::CodexAdapter;
+use memscribe_core::model::SourceLocation;
+use memscribe_core::{ParseCtx, RawRecord, TranscriptAdapter};
+
+#[inline]
+fn run(data: &[u8]) {
+    let adapter = CodexAdapter;
+    let loc = SourceLocation::new("fuzz://codex", 0, 1);
+    let raw = RawRecord::new(data.to_vec(), loc);
+    let mut ctx = ParseCtx::new();
+    let _ = adapter.parse(&raw, &mut ctx);
+    let _ = adapter.schema_fingerprint(&raw);
+}
+
+#[cfg(fuzzing)]
+libfuzzer_sys::fuzz_target!(|data: &[u8]| {
+    run(data);
+});
+
+#[cfg(not(fuzzing))]
+fn main() {
+    run(b"");
+}
diff --git a/fuzz/fuzz_targets/copilot.rs b/fuzz/fuzz_targets/copilot.rs
new file mode 100644
index 0000000..f3dabbc
--- /dev/null
+++ b/fuzz/fuzz_targets/copilot.rs
@@ -0,0 +1,30 @@
+//! Fuzz target for the GitHub Copilot adapter parser (whitepaper §8.4).
+//!
+//! Feeds arbitrary bytes to `CopilotAdapter::parse` through a `RawRecord` with
+//! a fresh `ParseCtx`, asserting the parser never panics and terminates. The
+//! libFuzzer harness turns any panic into a crash artifact.
+#![cfg_attr(fuzzing, no_main)]
+
+use memscribe_adapters::copilot::CopilotAdapter;
+use memscribe_core::model::SourceLocation;
+use memscribe_core::{ParseCtx, RawRecord, TranscriptAdapter};
+
+#[inline]
+fn run(data: &[u8]) {
+    let adapter = CopilotAdapter;
+    let loc = SourceLocation::new("fuzz://copilot", 0, 1);
+    let raw = RawRecord::new(data.to_vec(), loc);
+    let mut ctx = ParseCtx::new();
+    let _ = adapter.parse(&raw, &mut ctx);
+    let _ = adapter.schema_fingerprint(&raw);
+}
+
+#[cfg(fuzzing)]
+libfuzzer_sys::fuzz_target!(|data: &[u8]| {
+    run(data);
+});
+
+#[cfg(not(fuzzing))]
+fn main() {
+    run(b"");
+}
diff --git a/fuzz/fuzz_targets/cursor.rs b/fuzz/fuzz_targets/cursor.rs
new file mode 100644
index 0000000..3057b17
--- /dev/null
+++ b/fuzz/fuzz_targets/cursor.rs
@@ -0,0 +1,30 @@
+//! Fuzz target for the Cursor adapter parser (whitepaper §8.4).
+//!
+//! Feeds arbitrary bytes to `CursorAdapter::parse` through a `RawRecord` with a
+//! fresh `ParseCtx`, asserting the parser never panics and terminates. The
+//! libFuzzer harness turns any panic into a crash artifact.
+#![cfg_attr(fuzzing, no_main)]
+
+use memscribe_adapters::cursor::CursorAdapter;
+use memscribe_core::model::SourceLocation;
+use memscribe_core::{ParseCtx, RawRecord, TranscriptAdapter};
+
+#[inline]
+fn run(data: &[u8]) {
+    let adapter = CursorAdapter;
+    let loc = SourceLocation::new("fuzz://cursor", 0, 1);
+    let raw = RawRecord::new(data.to_vec(), loc);
+    let mut ctx = ParseCtx::new();
+    let _ = adapter.parse(&raw, &mut ctx);
+    let _ = adapter.schema_fingerprint(&raw);
+}
+
+#[cfg(fuzzing)]
+libfuzzer_sys::fuzz_target!(|data: &[u8]| {
+    run(data);
+});
+
+#[cfg(not(fuzzing))]
+fn main() {
+    run(b"");
+}
diff --git a/fuzz/fuzz_targets/gemini.rs b/fuzz/fuzz_targets/gemini.rs
new file mode 100644
index 0000000..1fc9721
--- /dev/null
+++ b/fuzz/fuzz_targets/gemini.rs
@@ -0,0 +1,30 @@
+//! Fuzz target for the Gemini adapter parser (whitepaper §8.4).
+//!
+//! Feeds arbitrary bytes to `GeminiAdapter::parse` through a `RawRecord` with a
+//! fresh `ParseCtx`, asserting the parser never panics and terminates. The
+//! libFuzzer harness turns any panic into a crash artifact.
+#![cfg_attr(fuzzing, no_main)]
+
+use memscribe_adapters::gemini::GeminiAdapter;
+use memscribe_core::model::SourceLocation;
+use memscribe_core::{ParseCtx, RawRecord, TranscriptAdapter};
+
+#[inline]
+fn run(data: &[u8]) {
+    let adapter = GeminiAdapter;
+    let loc = SourceLocation::new("fuzz://gemini", 0, 1);
+    let raw = RawRecord::new(data.to_vec(), loc);
+    let mut ctx = ParseCtx::new();
+    let _ = adapter.parse(&raw, &mut ctx);
+    let _ = adapter.schema_fingerprint(&raw);
+}
+
+#[cfg(fuzzing)]
+libfuzzer_sys::fuzz_target!(|data: &[u8]| {
+    run(data);
+});
+
+#[cfg(not(fuzzing))]
+fn main() {
+    run(b"");
+}
diff --git a/fuzz/fuzz_targets/otel.rs b/fuzz/fuzz_targets/otel.rs
new file mode 100644
index 0000000..fe83088
--- /dev/null
+++ b/fuzz/fuzz_targets/otel.rs
@@ -0,0 +1,30 @@
+//! Fuzz target for the OTel adapter parser (whitepaper §8.4).
+//!
+//! Feeds arbitrary bytes to `OtelAdapter::parse` through a `RawRecord` with a
+//! fresh `ParseCtx`, asserting the parser never panics and terminates. The
+//! libFuzzer harness turns any panic into a crash artifact.
+#![cfg_attr(fuzzing, no_main)]
+
+use memscribe_adapters::otel::OtelAdapter;
+use memscribe_core::model::SourceLocation;
+use memscribe_core::{ParseCtx, RawRecord, TranscriptAdapter};
+
+#[inline]
+fn run(data: &[u8]) {
+    let adapter = OtelAdapter;
+    let loc = SourceLocation::new("fuzz://otel", 0, 1);
+    let raw = RawRecord::new(data.to_vec(), loc);
+    let mut ctx = ParseCtx::new();
+    let _ = adapter.parse(&raw, &mut ctx);
+    let _ = adapter.schema_fingerprint(&raw);
+}
+
+#[cfg(fuzzing)]
+libfuzzer_sys::fuzz_target!(|data: &[u8]| {
+    run(data);
+});
+
+#[cfg(not(fuzzing))]
+fn main() {
+    run(b"");
+}
diff --git a/fuzz/fuzz_targets/vscode.rs b/fuzz/fuzz_targets/vscode.rs
new file mode 100644
index 0000000..9d5979f
--- /dev/null
+++ b/fuzz/fuzz_targets/vscode.rs
@@ -0,0 +1,30 @@
+//! Fuzz target for the VS Code adapter parser (whitepaper §8.4).
+//!
+//! Feeds arbitrary bytes to `VsCodeAdapter::parse` through a `RawRecord` with a
+//! fresh `ParseCtx`, asserting the parser never panics and terminates. The
+//! libFuzzer harness turns any panic into a crash artifact.
+#![cfg_attr(fuzzing, no_main)]
+
+use memscribe_adapters::vscode::VsCodeAdapter;
+use memscribe_core::model::SourceLocation;
+use memscribe_core::{ParseCtx, RawRecord, TranscriptAdapter};
+
+#[inline]
+fn run(data: &[u8]) {
+    let adapter = VsCodeAdapter;
+    let loc = SourceLocation::new("fuzz://vscode", 0, 1);
+    let raw = RawRecord::new(data.to_vec(), loc);
+    let mut ctx = ParseCtx::new();
+    let _ = adapter.parse(&raw, &mut ctx);
+    let _ = adapter.schema_fingerprint(&raw);
+}
+
+#[cfg(fuzzing)]
+libfuzzer_sys::fuzz_target!(|data: &[u8]| {
+    run(data);
+});
+
+#[cfg(not(fuzzing))]
+fn main() {
+    run(b"");
+}
diff --git a/fuzz/fuzz_targets/windsurf.rs b/fuzz/fuzz_targets/windsurf.rs
new file mode 100644
index 0000000..c2a2790
--- /dev/null
+++ b/fuzz/fuzz_targets/windsurf.rs
@@ -0,0 +1,30 @@
+//! Fuzz target for the Windsurf adapter parser (whitepaper §8.4).
+//!
+//! Feeds arbitrary bytes to `WindsurfAdapter::parse` through a `RawRecord` with
+//! a fresh `ParseCtx`, asserting the parser never panics and terminates. The
+//! libFuzzer harness turns any panic into a crash artifact.
+#![cfg_attr(fuzzing, no_main)]
+
+use memscribe_adapters::windsurf::WindsurfAdapter;
+use memscribe_core::model::SourceLocation;
+use memscribe_core::{ParseCtx, RawRecord, TranscriptAdapter};
+
+#[inline]
+fn run(data: &[u8]) {
+    let adapter = WindsurfAdapter;
+    let loc = SourceLocation::new("fuzz://windsurf", 0, 1);
+    let raw = RawRecord::new(data.to_vec(), loc);
+    let mut ctx = ParseCtx::new();
+    let _ = adapter.parse(&raw, &mut ctx);
+    let _ = adapter.schema_fingerprint(&raw);
+}
+
+#[cfg(fuzzing)]
+libfuzzer_sys::fuzz_target!(|data: &[u8]| {
+    run(data);
+});
+
+#[cfg(not(fuzzing))]
+fn main() {
+    run(b"");
+}
diff --git a/fuzz/fuzz_targets/zed.rs b/fuzz/fuzz_targets/zed.rs
new file mode 100644
index 0000000..8853514
--- /dev/null
+++ b/fuzz/fuzz_targets/zed.rs
@@ -0,0 +1,30 @@
+//! Fuzz target for the Zed adapter parser (whitepaper §8.4).
+//!
+//! Feeds arbitrary bytes to `ZedAdapter::parse` through a `RawRecord` with a
+//! fresh `ParseCtx`, asserting the parser never panics and terminates. The
+//! libFuzzer harness turns any panic into a crash artifact.
+#![cfg_attr(fuzzing, no_main)]
+
+use memscribe_adapters::zed::ZedAdapter;
+use memscribe_core::model::SourceLocation;
+use memscribe_core::{ParseCtx, RawRecord, TranscriptAdapter};
+
+#[inline]
+fn run(data: &[u8]) {
+    let adapter = ZedAdapter;
+    let loc = SourceLocation::new("fuzz://zed", 0, 1);
+    let raw = RawRecord::new(data.to_vec(), loc);
+    let mut ctx = ParseCtx::new();
+    let _ = adapter.parse(&raw, &mut ctx);
+    let _ = adapter.schema_fingerprint(&raw);
+}
+
+#[cfg(fuzzing)]
+libfuzzer_sys::fuzz_target!(|data: &[u8]| {
+    run(data);
+});
+
+#[cfg(not(fuzzing))]
+fn main() {
+    run(b"");
+}
diff --git a/memscribe.example.toml b/memscribe.example.toml
new file mode 100644
index 0000000..369981d
--- /dev/null
+++ b/memscribe.example.toml
@@ -0,0 +1,186 @@
+# Memscribe example configuration (whitepaper §10).
+#
+# This file documents the INTENDED config surface for the `memscribe` daemon
+# (`memscribe watch --config memscribe.toml`). It mirrors, one-to-one, the
+# runtime types the pipeline already exposes:
+#
+#   [tools.*.overrides]  -> memscribe_core::DiscoverCfg.overrides
+#   [[gate.rules]]       -> CommitmentGate::from_triples (id, category, pattern)
+#   [[redact.patterns]]  -> Redactor::from_patterns      (label, pattern)
+#   [sink]               -> ndjson | sqlite | memdb
+#
+# NOTE: the TOML *loader* itself is a follow-up — today the daemon runs with the
+# compiled defaults (CommitmentGate::default_table, Redactor::default, NdjsonSink)
+# and `watch` is not yet wired. Every value below is the documented default, so
+# this file doubles as a reference for what the loader will accept.
+
+# ---------------------------------------------------------------------------
+# Capture: which tools to tail, and where their transcripts live.
+# ---------------------------------------------------------------------------
+[capture]
+# The set of adapters to enable. Slugs match `SourceKind::parse` (and the CLI
+# `--as` / `--tools` values): claude_code, codex, gemini, otel, cursor,
+# windsurf, zed, vscode, copilot.
+tools = ["claude_code", "codex", "gemini"]
+
+# Restrict discovery to a single project root (maps to
+# DiscoverCfg.project_filter). Omit to capture every project the tools wrote.
+# project_filter = "/Users/me/work/myrepo"
+
+# Override $HOME for discovery (used by sandboxes/tests; DiscoverCfg.home).
+# home = "/Users/me"
+
+# Per-tool path overrides. Each key is the tool's native environment variable;
+# the value replaces the default `<home>/.<tool>` discovery root. These land in
+# DiscoverCfg.overrides verbatim, so only keys the adapter reads have an effect.
+[tools.claude_code.overrides]
+# Default discovery root is <home>/.claude; transcripts under projects/<slug>/.
+CLAUDE_CONFIG_DIR = "/Users/me/.claude"
+
+[tools.codex.overrides]
+# Default is <home>/.codex; rollouts under sessions/YYYY/MM/DD/rollout-*.jsonl[.zst].
+CODEX_HOME = "/Users/me/.codex"
+
+# Other adapters resolve relative to `home` with no override needed:
+#   gemini   -> <home>/.gemini
+#   cursor / windsurf / zed / vscode / copilot -> their editor state dirs
+#   otel     -> a network OTLP receiver, not a path
+
+# ---------------------------------------------------------------------------
+# The commitment-marker gate (whitepaper Appendix B).
+#
+# A small, inspectable rule table over user turns. Each rule is an `id`, a
+# `category`, and a case-insensitive regex `pattern` — the exact triple shape
+# CommitmentGate::from_triples consumes. Categories (MarkerCategory):
+#   decision_verb | rejection | ban | imperative | memory | confirmation
+#
+# Leaving [[gate.rules]] out entirely uses CommitmentGate::default_table()
+# (the 8 default rules below). Provide rules to REPLACE that table.
+# ---------------------------------------------------------------------------
+[[gate.rules]]
+id       = "decision_verb.use"
+category = "decision_verb"
+pattern  = '''\b(?:use|using|adopt|adopts|go with|let'?s go with|switch to|migrate to)\b'''
+
+[[gate.rules]]
+id       = "decision_verb.decide"
+category = "decision_verb"
+pattern  = '''\b(?:decide(?:d)?|we(?:'ll| will) choose|choose|chose|settle on|going to use|pick(?:ed)?)\b'''
+
+[[gate.rules]]
+id       = "rejection.instead_of"
+category = "rejection"
+pattern  = '\binstead of\b'
+
+[[gate.rules]]
+id       = "rejection.rather_than"
+category = "rejection"
+pattern  = '\b(?:rather than|as opposed to|in favor of)\b'
+
+[[gate.rules]]
+id       = "ban.negated_use"
+category = "ban"
+pattern  = '''\b(?:never|do not|don'?t|won'?t|will not|must not|should not|shouldn'?t|no longer)\s+(?:use|add|adopt|depend|introduce|rely)\b'''
+
+[[gate.rules]]
+id       = "ban.no_dependency"
+category = "ban"
+pattern  = '\bno (?:new )?dependenc(?:y|ies)\b'
+
+[[gate.rules]]
+id       = "imperative.must_always_never"
+category = "imperative"
+pattern  = '\b(?:must|always|never|shall|required to|need to)\b'
+
+[[gate.rules]]
+id       = "memory.remember"
+category = "memory"
+pattern  = '''\b(?:remember that|keep in mind|note that|for future reference|don'?t forget)\b'''
+
+# ---------------------------------------------------------------------------
+# Redaction (whitepaper §8.6, §11).
+#
+# A deterministic pass that strips known secrets from node text BEFORE the sink.
+# Each entry is a `(label, pattern)` pair (Redactor::from_patterns); a match is
+# replaced with `[REDACTED:<label>]`, applied in listed order. Redaction is ON
+# by default. Omit [[redact.patterns]] to use Redactor::default() (the patterns
+# below).
+# ---------------------------------------------------------------------------
+[redact]
+# Structure-only mode: elide ALL verbatim text, keep only node structure
+# (the CLI `--no-content` flag). Default false.
+no_content = false
+
+[[redact.patterns]]
+label   = "anthropic_key"
+pattern = 'sk-ant-[A-Za-z0-9_-]{16,}'
+
+[[redact.patterns]]
+label   = "openai_key"
+pattern = 'sk-[A-Za-z0-9]{20,}'
+
+[[redact.patterns]]
+label   = "aws_access_key"
+pattern = 'AKIA[0-9A-Z]{16}'
+
+[[redact.patterns]]
+label   = "github_token"
+pattern = 'gh[pousr]_[A-Za-z0-9]{20,}'
+
+[[redact.patterns]]
+label   = "slack_token"
+pattern = 'xox[baprs]-[A-Za-z0-9-]{10,}'
+
+[[redact.patterns]]
+label   = "google_api_key"
+pattern = 'AIza[0-9A-Za-z_-]{35}'
+
+[[redact.patterns]]
+label   = "bearer_token"
+pattern = '(?i)bearer\s+[A-Za-z0-9._~+/-]{16,}=*'
+
+[[redact.patterns]]
+label   = "assignment_secret"
+pattern = '''(?i)\b(?:api[_-]?key|secret|token|password|passwd|access[_-]?key)\b\s*[=:]\s*[^\s'"]{6,}'''
+
+[[redact.patterns]]
+label   = "private_key_block"
+pattern = '-----BEGIN (?:RSA |EC |OPENSSH |DSA |PGP )?PRIVATE KEY-----[\s\S]*?-----END (?:RSA |EC |OPENSSH |DSA |PGP )?PRIVATE KEY-----'
+
+# ---------------------------------------------------------------------------
+# Retention & ingest cadence.
+#
+# Memscribe is append-only over the captured history; retention controls how
+# much is re-scanned on each pass, never deletion of source transcripts.
+# ---------------------------------------------------------------------------
+[ingest]
+# How often the tailer re-scans discovered transcripts for new bytes (seconds).
+poll_interval_secs = 5
+# Resume each file from its last byte offset instead of re-reading from the top.
+resume_from_offset = true
+# Only ingest records newer than this lookback window on a cold start (days).
+# Omit or set 0 to ingest the full available history.
+backfill_days = 30
+
+# ---------------------------------------------------------------------------
+# Sink: where prepared nodes go (whitepaper §6).
+#
+# target = "ndjson" | "sqlite" | "memdb". NDJSON is the canonical default and
+# needs no external service. `memdb` requires building memscribe-sink with
+# `--features memdb`.
+# ---------------------------------------------------------------------------
+[sink]
+target = "ndjson"
+
+# NDJSON: one JSON node per line. Omit `path` to write to stdout.
+[sink.ndjson]
+# path = "/var/log/memscribe/nodes.ndjson"
+
+# SQLite: a queryable local store, zero external services.
+[sink.sqlite]
+# path = "/var/lib/memscribe/memscribe.db"
+
+# MemDB: bi-temporal write for Memtrace (feature `memdb`, off by default).
+# valid_at = turn/episode time, transaction_at = ingest time, episode_id = arc.
+[sink.memdb]
+# endpoint = "127.0.0.1:7878"
diff --git a/rust-toolchain.toml b/rust-toolchain.toml
new file mode 100644
index 0000000..b653fac
--- /dev/null
+++ b/rust-toolchain.toml
@@ -0,0 +1,5 @@
+# Pinned toolchain for reproducible builds (whitepaper §2).
+[toolchain]
+channel = "1.96.0"
+components = ["rustfmt", "clippy", "rust-src"]
+profile = "minimal"