fix(recall): relaxed recall rung + v0.1.1 (ADR-0102)

[Xxx91n]

What

Hotfix 0.1.1: recall_memory couldn't reach stored evidence when asked in natural language.

• Both recall layers (internal retrieval_results and the project index) build the FTS5 query as (w1* AND … AND wN*) OR "phrase" (ADR-0023 D2).
• A host's natural-language follow-up almost never has every term inside a single row.
• Host-in-loop repro (Codex 0.142.5): 111 internal rows and 85 project-index rows were stored, yet recall_memory("latest NBA champion NBA Finals MVP") → 0 hits in both layers. The same terms OR-ed → 85/111.

Fix (ADR-0102 D1): strict rung first, relaxed rung on under-fill

• The strict rung is unchanged; its ranking and priority are kept.
• When it returns fewer than limit rows, append a relaxed rung:
  • content words OR-ed with prefix match;
  • same BM25 ranking, plus freshness for the internal layer.
• Coverage floor: a relaxed row must cover a strict majority of the query's content words. This keeps the golden unanswerable (un_*) near-miss contract: a row sharing a single entity word is not evidence. The first draft without the floor broke 3 eval tests; with it they pass.
• Applied to SqliteSessionStore.searchMemory, searchMemoryMultiQuery (per variant) and ProjectIndexStore.search.
• Also noted (ADR-0102 Consequences): ProjectIndexStore's SELECT p.rowid returns the PK as id at runtime, so dedupe keys on id. The SQL and the /recall shape are unchanged.

Version

0.1.0 → 0.1.1 in one commit: 9 package.json + plugin.json + ship-gate release pin + provider clientInfo. CHANGELOG 0.1.1, ADR-0102 + index regen, CONTEXT term.

Verification (local, Windows)

• pnpm -r check → exit 0
• pnpm test → exit 0, 0 failures, including eval-abstain / eval-gate / eval-runner
• New regression tests:
  • packages/store/test/fts5-relaxed-recall.test.ts (8, including the 0-hit repro and the near-miss negative)
  • apps/plugin/test/project-index-relaxed.test.ts (2)
• node scripts/ship-gate.mjs → ship gate green — ready to tag the next release (server reports v0.1.1)

After merge (release)

gh workflow run release.yml -f runPurpose=pre-tag → when green, git tag v0.1.1 && git push origin v0.1.1 → the OIDC publish job runs.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GoFKmiw99ChCMjGAX8BVxL