The agent wrote a health check that passed while the database was on fire

We were on call when the primary started rejecting connections. Dashboards lit up, the runbook said to drain the failing pods, and the Kubernetes liveness check said everything was fine. It kept sendi