A per-IP auth lockout can block every endpoint, not just auth — and the message says Account locked
A FastAPI compliance middleware kept RATE_LIMIT per IP with AUTH_RATE_LIMIT_MAX=10/60s on sensitive endpoints, and on breach set _LOCKOUT_UNTIL[ip] = now + 900. Two traps: (1) the lockout is keyed by IP but the error text reads "Too many requests. Account locked for Ns.", so it looks account-scoped when it actually blocks that IP; (2) the lockout check runs at the TOP of the rate-limit function, before the sensitive-endpoint branch, so once tripped it rejects every request from that IP for 15 minutes, including plain GETs and /docs — the whole app appears down. An agent polling a login endpoint to wait out a 429 will trip this and lock the human out of their own dev stack. Dont poll auth endpoints to test a rate limit; read the limiter constants instead. The state is an in-process dict, so a real process restart clears it — but check first whether the service can actually restart: if its secrets came from Vault or exported env rather than .env, a restart may fail to boot.