fix: review follow-up — rag adapter crash, count/link false positives, hygiene

[silversurfer562]

What

The remaining findings from the library review (follow-up to #6). One turned out to be a live crash, not a comment problem.

rag adapter crashed on every call

The adapter read result.is_faithful — a field attune-rag's FaithfulnessResult has never had (verified against installed attune-rag 0.1.23 and the current attune-rag source; its verdict is score/claims-based). Every semantic run raised AttributeError and degraded to a warning. faithful is now derived from unsupported_claims, the translation lives in a unit-tested _to_verdict(), and calling the sync judge from inside a running event loop raises a clear, actionable error. The stale Phase-3 TODO and the wrong "verified against rag 0.2.0" protocol claim are gone.

Count checker false positives

• "Released 2026 versions of the widgets." flagged 2026 against widgets=12 (error severity → fails a clean doc). Values in 1900–2099 are now compared only when the source keyword directly follows the number — "2026 widgets" is still checked, so no recall loss (guarded by a hallucinated corpus case).
• Keyword matching now requires a leading word boundary: test no longer matches inside latest; plural drift (widget ~ widgets) still matches.
• check_counts docstring no longer promises warnings the code never produced.

Link checker root escape

../-traversal targets that happen to hit a real file outside project_root (e.g. /etc/passwd) previously passed silently. Escapes now yield a warning (unverifiable as a project link — same "never a silent pass" rule as flags), and site-absolute targets (/docs/page.md) resolve under project_root instead of the filesystem root.

Hygiene

• FindingKind.CHECKER_ERROR for checker infra failures (was repurposing unresolved_import)
• Import resolution cached per call — one subprocess per distinct module
• VerifyContext.judge typed Optional[Judge]
• README status unstaled (was "implementation in progress" at 0.2.1)
• black enforced in CI Lint alongside ruff; repo formatted

Tests

• 2 new corpus cases (year-FP clean guard + year-valued-count recall guard)
• 8 new behavioral/semantic tests, incl. year-guard boundary values (1899/1900/2099/2100) for mutation killing
• pytest: 83 passed · ruff + black --check: clean

🤖 Generated with Claude Code