[morajabi]

Drizzle calls postgres.js unsafe() with preparation disabled, adding a Describe/Flush exchange to parameterized queries. Dialog opening also locked each owner separately and reread unchanged state. Batch owner locks in user-ID order, bind hot owner lists as one array, and skip only proven no-op readbacks. Reset replay rewind backoff after the client passes its previous high-water mark.

Keep preparation restricted to the previously tuned replay-access queries by default. Broad preparation is now an explicit INLINE_DATABASE_PREPARED_QUERIES=all-custom opt-in: startup and readiness check force_custom_plan through the query pool. This avoids introducing generic plans for parameter-sensitive queries. The application does not change database roles/settings. Operators must qualify role defaults, recycle existing pooled backends, and measure representative replay before enabling it; one sampled backend cannot prove the whole pool inherited a setting.

Cap retained signatures at 200, admit broad shapes only after a second observation, and reserve capacity for access queries by limiting broad signatures to 100. Candidate tracking is bounded. Retained signatures are deliberately not advertised as an LRU; overflow keeps the original unnamed path. Pool/per-call opt-outs and transaction/savepoint behavior remain intact.

Validation: 40 focused prepared/connection/wire tests passed, final guard/readiness integration rerun passed 12; four dialog PostgreSQL tests and five replay guard tests passed. Partial-index tests verify retained index use and zero generic plans after more than five custom executions. Benchmark harness mode/report tests and typecheck passed. Controlled 20-sample handler benchmarks at injected +6 ms show access→all-custom exchanges 43→24 for DM sends, 51→28 for public sends, and 19→11 for history; respective medians 267.54→142.41 ms, 363.07→201.41 ms, and 192.13→101.56 ms. Proxy calibration differs slightly between runs; these are synthetic handler measurements, not production API latency. User-bucket replay qualification is separate; these replay fixtures exercise chat buckets. Full server/image CI passed at 0b51ed38f; latest test-only proxy half-close fix has 11 passing wire/measurement tests and typecheck, with hosted CI pending. Prior 5dc49a0f2 global-preparation measurements are historical, not evidence for the revised default. No production deploy, role change, migration, DNS or traffic switch.

[morajabi]

Landed on main through the rebase merge of #124, which contained all five commits of this branch (now 6c12232, f98d034, 7095cfd, dc40a07, ff9fc5b).