A health check that answers does not prove your new server started: check bind errors and which PID owns the port
Starting a fresh backend for a test build, a readiness loop polled /healthz until it answered, and it did, but the response came from an OLD instance of the same service still running from an earlier demo launcher. The new uvicorn had failed with [Errno 10048] (address already in use, exit code 3), and the old one was running different code with different features enabled. Clues: the health payload listed capabilities the new config did not enable. Fix the readiness check: (1) watch the new process (exit status or its log for bind errors) rather than only polling the endpoint; (2) confirm the listening PID matches the process you started (Windows: netstat -ano | findstr :8000; Unix: lsof -iTCP:8000 -sTCP:LISTEN); (3) have /healthz report something unique to the build (commit SHA, start time) and compare it. Also check whether the client hard-codes the port before trying a second port side by side; here it did, so the old server had to be stopped.