Search capabilities are snapshotted at startup and never refresh when an index becomes queryable later #103

Parent

#90

What to build

The engine detects search capabilities once, at startup, after a bounded wait for search indexes (60 s by default; 180 s with MEMONGO_STRICT_SEARCH_INDEX_READY=1). If an index becomes queryable after that wait, the process keeps reporting the lane as unavailable until it restarts.

CI run 39163428188 (draft PR #102) shows this: every autoEmbed vector index was READY and queryable about 1 s after the API's 60 s wait ended, but the API had already recorded vectorSearch:false.

The API now reports such a lane as a typed capability-absent, so this is visible rather than a silent empty result. Still, a long-running process should not stay degraded after its indexes recover.

Acceptance criteria

• A lane that is unavailable at startup becomes available without a restart once its serving indexes are queryable, using a bounded re-detection
• Re-detection never marks a lane available on index readiness alone where a per-document guarantee is claimed (the decision 17 barrier stays the visibility proof)
• The capability table and /v1/status reflect the refreshed state
• A test with a fake index lister covers pending → queryable after startup

Blocked by

None. M1 works around this in CI with strict readiness (commit 246294d), which fails startup loudly instead.

🤖 Generated with Claude Code