FE-1253: Add prospect research regression

.changeset/tricky-poems-do.md

@@ -0,0 +1,5 @@
+---
+"@hashintel/brunch": patch
+---
+
+Add a deterministic full-stack prospect research regression case ([#369](https://github.com/hashintel/brunch/pull/369)).

docs/praxis/comparison-runs.md

@@ -129,6 +129,13 @@ audience-safe requirement ledger, and the bounded validity-first report.
 This historical witness predates Alpha 10 provenance capture. Keep it as evidence and an example, but
 do not backfill `provenance.json` or publish it through the current publication skill.
 
+FE-1253 adds `prospect-research-workspace` as a deterministic full-stack regression case, not a second
+end-to-end campaign profile. Its fixed public stack is React + Node.js + TypeScript + SQLite, with
+server-side Pi/Clay-compatible fixtures and denied runtime network. The saved mission remains available
+for exploratory specification work, while the compiled browser/HTTP/SQLite/export oracle is calibrated
+against a known-good implementation and focused wrong rivals. A future campaign may compose this case
+only through a separately authorized study contract.
+
 ## Agent recipe: drive, then join evidence
 
 On an overlay-capable host, drive the real TUI with the pinned project-local `pi-interactive-shell` package. In a sandbox or headless environment, use `npm run tui-driver` as the fallback. Follow the bounded observation, named-key input, and deterministic cleanup protocol in [Manual Testing](manual-testing.md).

memory/PLAN.md

@@ -21,7 +21,7 @@ Brunch-next has delivered the original composition spine: the host, sealed Pi pr
 
 **Parallel lanes.** Group 3 agent-layer work is pickup-ready. `agent-control-plane-closure` is one earned coverage frontier over foreground and background prompt ingress: it absorbs the former `graph-assurance-conduct` and `subagent-skill-access` items, reconciles the stale eager-context fossils left by D58-L's load-on-demand cutover, and closes prompt replacement/resource-invocation contradictions before capture experiments amplify them. FE-1187 retains only the concern-grouped review renderer. FE-1208's reshaped `automation-observability-dx` closure landed on the restacked base; its invalid warrant pilots remain historical evidence, while the useful evaluator/report primitives are available to `capture-ledger-tracer`. Two evaluation use cases stay deliberately distinct: **seed-based intra-product testing** (Brunch on its own terms — seeds encode Brunch graph/spec state only Brunch understands) and **mission-driven agent-as-user cross-product comparison** (`agent-as-user-comparison` — competitor CLIs share no Brunch state, so the comparable artifact is the "ready" spec/plan document a mission produces). FE-1210 closed the rigorous technical tracer on 2026-07-17; the separately planned operator workflow (`operator-comparison-workflow`, FE-1215) makes saved missions, visible comparison-harness framing, conversational launch, and a readable free-form report approachable without weakening the retained regression/evaluation machinery. Its first operator-led run falsified the nested fresh-actor topology: the top-level project Pi session must act as the simulated user and drive one direct harness subshell at a time, with stock-Pi text interaction as the portable baseline; the corrected real witness is deferred to `saved-mission-comparison-witness`. The KA stream owns executor/orchestrator/Execute-mode work. FE-1192 and FE-1195 completed the attempt, isolated fan-in, durable parallel authority, and epic integration sequence; `executor-plan-coherence` now uses those settled seams to test whether multi-slice plans converge on one working result. Broader instrumentation remains trigger-gated under Later.
 
-**End-to-end comparison lane.** FE-1239 materializes the handoff that the completed reporting procedure names but does not run: two fresh rigorous elicitation lanes produce exact approved specifications, each specification crosses an immutable content-addressed handoff into both Brunch and Claude execution, and the unchanged Petri-editor oracle joins all four outputs into requirement-level evidence. Shared interface requirements are disclosed before elicitation and remain controlled baseline, while private/reveal requirements stay controller-only. The first tracer supports bounded within-case contrasts only — no winner, reliability claim, or cross-case causal conclusion. FE-1241 keeps that Petri editor as the greenfield case and adds frozen Brunch/Petrinaut brownfield packets plus controller-owned oracles. All three missions can now enter the same execution-only workflow and publication contract. Brownfield preparation uses pinned remote-free snapshots and target-bounded tools as experimental hygiene, not adversarial benchmark proof; Clay remains outside the campaign.
+**Comparison lane.** FE-1239 remains the sole rigorous greenfield 2×2 tracer: two independently approved Petri specifications cross exact handoffs into Brunch and Claude execution and one unchanged oracle joins all four outputs. FE-1241 keeps that Petri editor as the greenfield study profile and adds frozen Brunch/Petrinaut brownfield packets plus deterministic oracles under a learning-first workflow. FE-1253 is deliberately thinner: a saved exploratory prospect mission plus a deterministic React/Node/TypeScript/SQLite execution regression case, with no prospect study profile, requirement ledger, scored provider gate, or campaign claim.
 
 **Current seams.** Brunch ships on the `1.0.0-alpha.x` line. One-shot `ask` is the only interactive structured-exchange terminal; D125-L's live ask registry provides headless discovery/answering, while the transcript-backed pending projection remains a compatibility surface for live offer tools after the legacy `present_question` pending branch retired. Sweep classification remains fail-closed and compile-time anchored to the exchange-schema terminal names (D117-L), while the larger capture-conditional watermark question remains A40-L.
 
@@ -119,6 +119,7 @@ Everything executor/orchestrator-shaped or Execute-mode-owned belongs to Kostand
 - `executor-slice-admission-parity` ([FE-1240](https://linear.app/hash/issue/FE-1240/prevent-invalid-scoped-slices-from-reaching-execution)) — **complete on `ka/fe-1240-slice-admission-parity`, restacked on `next` after FE-1239 landed:** incomplete scoped-slice worker context is rejected during deterministic plan admission, exact findings enter bounded repair before `slice_execute`, and the execution boundary remains fail-closed. Definition below.
 - `executor-plan-coherence` ([FE-1250](https://linear.app/hash/issue/FE-1250/build-coherent-execution-plans)) — **implementation complete on `ka/fe-1250-coherent-execution-plans`, based on FE-1240 via `next`:** frontier-verified multi-slice epics now require one ordinary terminal member over every sibling, planner conduct establishes shared foundations, and workers preserve cumulative public contracts for the canonical harness. Fast verification passes; next is the separately authorized unchanged Petri comparison rerun. No new browser gate, durable plan kind, or executor lifecycle phase. Definition below.
 - `comparison-publication-workflow` ([FE-1251](https://linear.app/hash/issue/FE-1251/publish-traceable-comparison-reports)) — **active on `ka/fe-1251-comparison-publication`, stacked on FE-1250:** capture immutable controller and release provenance before comparison lanes start, then explicitly publish retained validity-first reports into the canonical Notion database through a guarded idempotent skill. Definition below.
+- `prospect-research-workspace-regression` ([FE-1253](https://linear.app/hash/issue/FE-1253/thin-prospect-research-into-a-regression-case)) — **active on `ka/fe-1253-prospect-research-e2e`, stacked on FE-1241:** retain the exploratory mission and calibrated browser/HTTP/SQLite/export oracle while deleting the unexecuted prospect study profile, requirement-ledger expansion, and scored-provider completion gate.
 - **Carved from FE-1167 (2026-07-13):** the Execute-mode evidence sub-list — Execute entry beats on thin vs rich seeds (assessment honesty: Ask on thin, Proceed on rich) and the FE-1107/KA residue (close-or-narrow, demo/walkthrough session via `TESTING_PLAN.md`, post-KA plan pass). The former sticky-posture question is no longer KA residue: FE-1187 `remediation-4` owns the persistent Specify elicitation-style audit/SPEC revision, and its Continue lexical audit owns the old `continue` ambiguity. Full context in the archived FE-1167 definition (`docs/archive/PLAN_HISTORY.md`).
 - `planning-process-model` — **moved to the KA stream 2026-07-13; reshaped by D126-L**: the durable scope handoff is settled, so this item now owns only plan projection and epistemic-horizon questions beyond committed scopes. Definition below.
 - **[1.x data-model handoff owed by FE-1187](../docs/architecture/BRUNCH_1X_DATA_MODEL_HANDOFF.md):** a concise colleague-facing note must distinguish current canon from directional vocabulary before the consolidated outer checkpoint. Current: `{milestone, frontier, scope}` with executor-derived slices; basis (`explicit | implicit`) orthogonal to settlement (`advisory | settled`); no persisted readiness grade or spec-global elicitation-gap table; active-branch Pi JSONL reads; one CommandExecutor mutation authority with spec-local LSN/change log; no new projected `vv_obligation` (legacy rows remain readable). Directional only: explain stored `thesis` as pitch/concept without renaming it yet, and avoid new coupling to spec-local `term` while its possible workspace lift remains future work. Persisted judgment-shaped reconciliation needs remain current, but are YAGNI-suspect: add no kinds/consumers/orchestration dependency without fresh evidence; re-evaluate derivation/removal when KA work first needs that table. Delivered note: [`Brunch 1.x data-model handoff`](../docs/architecture/BRUNCH_1X_DATA_MODEL_HANDOFF.md); it does not block the deterministic TUI queue.
@@ -456,6 +457,25 @@ Instrumentation experiments and far-horizon items. Each re-enters only via re-qu
 - **Boundary:** no hidden-oracle exposure, inferred browser command, new plan-plane node, candidate command surface, execution-time plan repair, or FE-1241 comparison-framework change.
 - **Current execution pointer:** none; the two-card scope is consumed. Re-enter only for the unchanged frozen Petri comparison rerun.
 
+### prospect-research-workspace-regression
+
+- **Name:** Thin prospect research into a regression case
+- **Linear / branch:** [FE-1253](https://linear.app/hash/issue/FE-1253/thin-prospect-research-into-a-regression-case); `ka/fe-1253-prospect-research-e2e`, stacked on FE-1241 with no parent issue.
+- **Kind:** earned evaluation simplification — retain a deterministic full-stack regression oracle and retire unproven campaign expansion.
+- **Certainty:** earned.
+- **Status:** active 2026-07-23. The mission, public packet, opaque Brunch seed, closed compiled oracle, known-good full stack, and focused rivals remain. The prospect-specific end-to-end study profile and provider campaign gate are being retired before any scored lane exists.
+- **Objective:** provide one deterministic implementation-level acceptance case for the prospect research workflow without making ordinary regression value depend on a rigorous 2×2 campaign.
+- **Closes:** the accidental coupling between a useful full-stack oracle and an unexecuted prospect-specific campaign.
+- **Deletes / retires:** the prospect end-to-end study contract, shared baseline, reveal registry, matrix registration, requirement-ledger obligation, scored-provider gate, and campaign claims.
+- **Stabilizes:** the fixed React + Node.js + TypeScript + SQLite public packet; `npm test` / `npm run build` / `npm start`; fresh database and fixture isolation; runtime-network denial; independent browser/API/SQLite/export journeys; focused rivals; and opaque specification seeding.
+- **Depends on:** FE-1241 for the finalized execution-case registry and oracle dispatch only. No provider or strict greenfield Claude-isolation witness blocks completion.
+- **Boundary:** manually initiated prospect research, evidence-backed qualification, deduplication/provenance, suppression, review, audited override, approval, export, provider failure, and restart persistence are in. Outreach delivery, live Pi/Clay quality, campaign composition, scored provider lanes, reliability claims, and `ExecutionAttempt` widening are out.
+- **Acceptance:** the known-good full stack passes; focused rivals for unapproved research, confidence-only qualification, lost provenance, weak suppression, reasonless/destructive override, overbroad export, provider-failure laundering, non-durable state, and external runtime requests fail their owning claims; the prospect case is absent from end-to-end study registration; existing Petri and brownfield study bytes/oracles remain unchanged.
+- **Verification:** inner — exact public/oracle parsers, oracle-pack hash, claim coverage, and opaque Brunch seed. Middle — independent browser + API + SQLite journeys over fresh database/fixture state, paired with the controller reference model and focused rivals. No outer campaign evidence is required.
+- **Cross-cutting obligations:** controller fixtures and expected states remain outside targets; runtime network stays denied; compiled oracle dispatch remains fail-closed; historical comparison evidence remains unchanged.
+- **Traceability:** D70-L, D139-L; FE-1230/FE-1241; [`docs/praxis/comparison-runs.md`](../docs/praxis/comparison-runs.md).
+- **Current execution pointer:** thin the materialized branch to this regression boundary, then close through deterministic verification.
+
 ### comparison-reporting-skills
 
 - **Name:** Report comparison evidence
@@ -647,8 +667,12 @@ KA stream:
     -[hard]-> brownfield-comparison-cases
       status: complete FE-1241 with learning-first pinned snapshots, deterministic case oracles, and publication-compatible execution runs
       lights_up: same-base brownfield elicitation -> exact handoff -> crossed execution -> case oracle
-      cases: minimal-petri-net-editor = sole greenfield reference | FE-1201-derived brunch backend | PR #9051 petrinaut frontend
-      excludes: Clay | more greenfields | repetitions | aggregate winner | ExecutionAttempt widening
+      cases: minimal-petri-net-editor = FE-1241 greenfield reference | FE-1201-derived brunch backend | PR #9051 petrinaut frontend
+      excludes: Clay | more greenfields inside FE-1241 | repetitions | aggregate winner | ExecutionAttempt widening
+      -[hard]-> prospect-research-workspace-regression (FE-1253)
+        status: active; deterministic full-stack oracle calibrated, campaign expansion retiring
+        closes: useful regression oracle coupled to an unexecuted 2x2 campaign
+        excludes: provider lanes | requirement ledger | outreach delivery | reliability claims
   # executor-run-environment (FE-1166) resolved 2026-07-15: policy merged (PR #302),
   # live remainder folded into executor-plan-synthesis (FE-1197), card consumed.