Pool size config bottleneck

To reduce DB connection pressure, an SRE lowers the per-pod max DB pool size from 20 to 5 via a dynamic config change that propagates fleet-wide over ~2 minutes at 14:30 (no deploy, no image change). Immediately the API's p99 latency on write-heavy endpoints triples and a pool checkout timeout error rate appears, but the DB's own metrics show it healthy and now UNDER-utilized (fewer active connections than before). The pods have spare CPU. Triage, explain the mechanism, then mitigate.