Postgres benchmark traps 4 of 5 naïve diagnoses

A new Postgres incident benchmark makes the obvious diagnosis wrong on 4 of 5 faults and adds one healthy state to catch agents that invent trouble. The MIT-licensed six-record dataset captures live Postgres 16 telemetry after scripted missing-index, stale-statistics, bloat, lock-contention, and N+1 failures; its deterministic reference ranks all five faults first and returns zero findings on the healthy case. Teams building on Supabase’s containerized regression loop for database agents can use this as a sharp seed corpus, but not as a reliability claim: six synthetic scenarios on one instance are a unit test, not an incident distribution.