Text search across persona briefings + articles. ILIKE-based; at ~95 briefings + ~52K articles the seq scan + LIMIT stays fast enough on Postgres without a tsvector index. Swap to plainto_tsquery once article volume passes ~250K; the response contract is stable.

Declassified segregation 2026-07-06: the ~198 bulk-ingested briefing_kind='declassified' rows (Vietnam study, FBI cryptography files, …) were swamping general-search relevance — q=tech / ai / news returned declassified case briefings instead of the matching vertical persona briefings, and a new self-expanding ingest cron makes this strictly worse over time. General search (kind unset) now EXCLUDES declassified via the same is_distinct_from('declassified') gate the X-autopost rotation uses (app/daily_brief_worker.py). Declassified stays fully discoverable on its OWN surfaces — the declassified RSS feed, the search_declassified MCP tool, digest_declassified — and, for parity, callers can scope this endpoint to it explicitly with kind=declassified.

Pool-exhaustion fix 2026-06-21 (overnight Sentry SEV — TimeoutError QueuePool size=5 overflow=3 reached on /v1/search): deliberately no db: AsyncSession = Depends(get_db) parameter. The two ILIKE scans below can each take 5-15s over the 52K-article body_text seq-scan path; with Depends(get_db) the session is held for the FULL request duration. Under load that triggers the same pool- exhaustion class as PRs #657/#658/#674/#677/#678 — on the writer pool. Canonical fix-pattern: open a short-lived session inside the handler body + release via _close_session_safely in finally BEFORE building the response shape.