feat(evaluation): shared acceptance corpus + candidate-agnostic cross-review harness

[obvious-autobuild]

Human author: Gilbert Polanco (gilbertpolanco42@gmail.com)

Why

The arena cross-review needs to judge all four candidates (arena/candidate-a|b|c|d) against the same acceptance corpus, run through the same harness, or results are not comparable. Today there is no shared corpus: each review would improvise its own tests, and candidate quality would be judged on unequal ground. This PR lands that shared ground truth as data, plus a narrow adapter contract so any candidate's entry-creation and timeline-read entry points can execute it unchanged.

What

• Corpus as data — six JSON fixtures under evaluation/fixtures/, expectations independent of any candidate code:
  1. 01-multi-event-narrative — one narrative → 3 events, exact count/categories/quantities and transcript order pinned.
  2. 02-relative-time-timezones — 8 cases pinning exact absolute instants ("this morning", "yesterday 6pm", "an hour ago", "just now") across America/New_York, Pacific/Auckland (NZST and NZDT) and UTC; conventions documented in the fixture and README; expected instants machine-computed against the IANA tzdb (DST-safe).
  3. 03-malformed-extraction — hostile raw input (control chars, DEL): entry still created with 0 events (contract: extraction failure never blocks capture), raw preserved.
  4. 04-retry-double-submit — same captureId twice → Created then IdempotentReplay, exactly one entry/event on the timeline.
  5. 05-raw-fidelity — SHA-256 byte equality of the raw transcript across a validation failure and a retry (created entry, replay, timeline read); hostile fixtures include NUL/DEL/CRLF/tabs/emoji/combining marks/trailing whitespace.
  6. 06-reload-persistence — timeline identical (canonical-JSON deep-equal) across a simulated cold start.
• Adapter interface — evaluation/src/adapter.ts: CandidateAdapter = createEntry (full pipeline, settles before resolve) / readTimeline / reload, results as a discriminated union (Created | IdempotentReplay | Rejected). Narrow on purpose; wire shapes mirror the Effect v4 schema contract (unix-ms numbers, _tag literals, absent-not-null optionals). One deliberate addition: WireEntry.captureId, the duplicate-protection key.
• Fixed runner — src/runner.ts executes each fixture identically for every candidate; adapter exceptions surface as fixture failures, never crashes.
• Worked example — src/example/example-adapter.ts (in-memory, contract-shaped, DST-aware relative-time resolution) proves the corpus is executable.
• Negative control — src/example/broken-adapter.ts plants duplicate-on-retry + transcript-trim faults; the harness flags 5/6 fixtures, proving the checks bite.
• README — adapter contract, time conventions, exact commands, per-fixture evidence, what is/isn't pinned.

Scope guard: touches evaluation/** only; no candidate branches read; zero runtime dependencies (bun primary runner, tsx alternative documented).

How to Review

• Start at evaluation/README.md (contract + commands), then src/adapter.ts (the interface cross-review adapters will implement) and src/runner.ts (the fixed protocols).
• Fixture expectations to scrutinize: the eight absolute instants in 02 (verify against a tzdb if you like — each was machine-computed, notably NZDT 1790571600000 and US-EST 1793574000000).
• Intentionally excluded: candidate adapters (attached at cross-review), NLP-quality benchmarking beyond the pinned fixtures, publish/review transitions and multi-child scoping (corpus scope is one child's draft timeline), repo-level CI wiring (PR ci: add PR pipeline (pnpm via corepack, turbo typecheck, tests, build) #3 owns the pipeline).
• Deliberate choice: fixture 4/5 pin IdempotentReplay as the defined double-submit outcome; rejection on extraction failure would violate the schema contract's "events may be empty — extraction failure never blocks capture", so fixture 3 pins creation + 0 events instead.

Test Evidence

Run on this branch (bun 1.3.14, tsc 5.9.3 strict):

• bun src/run.ts → 6/6 fixtures PASS, exit 0 (worked example adapter; all six fixtures execute: 1 + 8 + 1 + 1 + 2 + 2 = 15 adapter workloads).
• bun src/run.ts --adapter=./src/example/broken-adapter.ts --expect-failure → faults caught, exit 0: duplicate-on-retry (expected IdempotentReplay, got Created), SHA-256 byte-fidelity violations on created/replayed/timeline surfaces, missing events.
• bun install && bun run typecheck → exit 0 (strict, noUncheckedIndexedAccess, verbatimModuleSyntax).

🔗 Obvious Project · 🧵 Obvious Thread