feat: improve paraphrase recall coverage

[davidrobertson]

Summary

• retry lcm_grep with up to two same-scope query reformulations after an empty result
• expand retrieval benchmark corpus v2 to 39 sessions and 38 queries
• report development and holdout metrics across 32 paraphrase cases

The old eight-query corpus overstated retrieval quality. Corpus v2 keeps exact, cross-session, and temporal checks while exposing the lexical paraphrase gap without changing the search engine.

Results

• overall Recall@1: 42.11%
• paraphrase Recall@1: 31.25%
• exact, cross-session, and temporal Recall@1: 100%

Validation

• npm test (184 tests passed)
• npm run smoke
• npx tsc --noEmit
• npm --cache /tmp/codex-lcm-npm-cache pack --dry-run
• retrieval benchmark corpus v2
• skill validation
• independent VoltFlow review

Risk

Search ranking and storage are unchanged. The behavioral change is limited to the shipped recall skill; the larger benchmark records a stricter baseline.