feat(eval): eval harness + first failure-mode fixtures

[officialCodeWork]

Summary

TRACKER Step 0.2 — the measurement loop. Every later phase's "are we going the right way" question is now answered by pnpm eval: fixtures encode failure modes, goldens encode both the right answer and the specific poison answer, and thresholds gate CI.

What's in it

• @coderadar/eval — scans every eval/fixtures/*/, diffs against golden.json, prints a per-failure-mode scorecard, writes eval/scorecard.json, appends to eval/history.jsonl with --record, exits non-zero on eval/thresholds.json violations.
• xfail semantics (per check) — expectedFail markers let us assert the future behavior today: the C1 per-instance attributions are red-but-expected until prop-flow (2.2); the moment they start passing, the run fails with unexpected-pass until the stale marker is removed — capability arrival is always an explicit, reviewed event. Forbidden (poison) checks never xfail.
• Fixtures:
  • c1-shared-datatable — the headline case: shared DataTable, two pages, two APIs. Poison assertions (cross-page endpoint bleed) are active now.
  • a4-generic-text — three components sharing "Save": lone generic term must yield ambiguous, one distinctive term must resolve it.
  • demo-app — baseline, sharing the app source with examples/ (no duplicate copy).

Current scorecard

25 pass · 0 fail · 2 xfail · 0 unexpected-pass
lineage precision=1.000 recall=0.714 · match accuracy=1.000
eval gate: OK

(recall 0.714 is the xfail gap — the two C1 instance attributions — and is exactly the number that jumps when Phase 2.2 lands)

Worth noting: the harness caught its first design bug during this PR — the original per-kind expectedFail marker flagged the already-passing definition-level attributions as unexpected-pass, forcing the switch to per-check markers. The gate works.

Acceptance criteria (from TRACKER 0.2)

• Runner with golden diffing, forbidden entries, scorecard.json, thresholds gate
• expected-fail support in the runner
• ≥ 3 fixtures including demo-app
• pnpm eval green (C1 attribution assertions xfail as planned)

Documentation

• docs/testing-strategy.md — golden format updated to the implemented contract (per-check expectedFail, exact-set endpoints); history recording clarified
• TRACKER.md — 0.2 done, Status → 0.3

🤖 Generated with Claude Code