feat: retrieve applicable experience by text

[fabbrik]

Summary

Story 2.2. Stored records become findable: text-matched, eligibility-filtered and ranked.

• New port. IExperienceCandidateSource in Abstractions, with a PostgreSQL implementation. IExperienceRecordStore is untouched.
• Search in SQL. Migration 0003 adds a generated tsvector column over task ID, task summary and reflection lesson, with a GIN index. Scope, eligible status and the confidence threshold are applied in the same statement; relevance comes from normalized ts_rank_cd.
• Eligibility in Core. Only Validated and Reinforced are eligible. Expiry and environment compatibility are decided over the returned candidates, and each exclusion says which check removed it.
• Ranking. Relevance 0.35, confidence 0.25, recency 0.15, status 0.15, environment 0.10 by default — nonnegative, summing to 1, validated at construction. Every returned record exposes its normalized components and effective weights; ties break on ExperienceId. A golden fixture pins the documented order.
• Bounded. A 500 ms default timeout returns an empty result carrying a timeout signal and correlation ID rather than throwing; caller cancellation stays distinct. Authorization mismatches, source failures and unreadable candidates all fail closed to an empty result.

Decisions recorded in the spec

Expiry and required environment attributes were undefined in planning, so this story defines them: MaxAge against UpdatedAt, and required attributes as key/value pairs where a missing key excludes the record and an empty set marks the result unrestricted. Recency decays by half every 30 days, which keeps the axis defined when no expiry is configured.

Review

Three reviewers raised 21 findings, all fixed. The substantive one: ranking only saw the strongest text matches, so a high-confidence recent record with a weaker match could never surface while the docs claimed every eligible record was returned. Truncation is now reported through Truncated and the ceiling is documented. Also fixed: the timeout path could exceed its own bound by cancelling Npgsql synchronously on the caller's thread, a token-source leak, unbounded task text and timeouts, and a relevance assertion that could not fail.

Test plan

• dotnet build --configuration Release: 0 warnings, 0 errors
• dotnet test --configuration Release: 523 of 523 pass (Core 307, Postgres 123). Locally with TESTCONTAINERS_RYUK_DISABLED=true.
• Container tests for the SQL: text match, scope isolation, status and confidence filtering, relevance normalization, and 0003 applied over an existing hand-migrated database.
• CI green

🤖 Generated with Claude Code