feat(serder)!: #142 strict canonical read-path parser — offset-based SAID verification

[joeldsouzax]

Closes #142. The read-path half of the #79 seam (design §3.3/§3.5).

What

• Strict single-pass canonical parser for the five fixed event grammars (serder::deserialize::canonical): compact JSON, spec field order, no escapes/duplicates/whitespace — rejected by construction with the new typed SerderError::NonCanonical { offset, expected, found }.
• Offset-based SAID verification: one scratch copy + span fill + hash, replacing 2–3 Value trees and a full re-render per ingested event.
• Borrowed &str field views (ParsedIcp<'a> …) — the substrate for C-a C-a · Zero-copy event input — borrow-ify KeriEvent across serder #129's borrow-ified KeriEvent<'a>.
• keripy intive=True integer kt/nt/bt still accepted (their SAIDs are computed over the integer form; rejecting them is a parity gap).
• Bonus fix, pre-existing DoS on main: VersionString::parse panicked on multi-byte UTF-8 straddling its fixed slice offsets (non-char-boundary &str slicing), reachable through deserialize_event via the old tolerant path. Fixed at the root with failing-first regression tests.

Breaking (MINOR under 0.x)

• New SerderError::NonCanonical variant; non-canonical inputs now fail with it instead of assorted Json/MissingField errors (the strict path never returns MissingField).
• said::verify_said is strict and returns Result<(), SerderError> (was Result<bool, _>); its digest-code trust contract is now documented.
• Per-ilk deserializers require their exact ilk (deserialize_rotation no longer accepts drt bytes; deserialize_inception no longer silently accepts dip bytes and drops the delegator; deserialize_interaction no longer accepts rot/drt bytes as a truncated interaction).

Verification

• Strict-vs-reference differential proptests over the builder-reachable event space (all five ilks), byte-identical re-serialization; mutation-subset property (anything strict accepts, the tolerant oracle accepts identically).
• Deterministic variant × invariant matrix (26 tests): every Seal/Tholder/Count/Config/Identifier variant, both SAID branches (single d≠i, double d==i), all five ilk-dispatch arms, and one reachability probe per reachable read-path error variant — guaranteed coverage, not probabilistic.
• Rejection probes: whitespace (size-consistent), duplicate keys, reordered fields, escapes, trailing bytes, every-prefix truncation, oversized/unknown ilks, non-ASCII version windows.
• keripy corpus through the strict path via keri/tests/differential.rs.
• Allocation count pinned (38 allocs per icp deserialize, exact, deterministic); CodSpeed serder_deserialize/icp bench added; bolero+AFL fuzz target serder_deserialize_event (idempotence oracle) wired into the deep-fuzz matrices.
• nix flake check green (wasm + no_std included).

Discovered during this work (filed separately)

• serder write path forces i==d on inception — cannot emit basic-derivation identifiers keripy produces #144 — the write path unconditionally forces i == d on inception (predates this branch; surfaced by the variant matrix trying to construct a basic-derivation icp). Read→write of a keripy basic inception corrupts d and i. Read path in this PR handles those events correctly.
• Event-tier keripy differential: full-breadth corpus + bidirectional byte-identity #145 — event-tier keripy differential corpus is one 3-event line, parse-only; needs full-breadth + byte-identity treatment like the primitive tier already has.

The old serde_json::Value read path survives only as the #[cfg(test)] differential oracle (deserialize::reference). serde demotion to dev-dependencies is the follow-up card per design §6 staging.

🤖 Generated with Claude Code

[codspeed-hq]

Merging this PR will not alter performance

✅ 25 untouched benchmarks
🆕 1 new benchmark
⏩ 21 skipped benchmarks¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
🆕	icp	N/A	38.2 µs	N/A

---

_{Comparing feat/142-strict-read-path (dedb299) with main (d8d1cf3)}

Footnotes

1. 21 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩