The post-deploy health check probed only the static landing (`GET /`) and the
gateway-served SPA shell (`GET /app/`) through the edge caddy. Neither needs the
backend, so a crash-looping backend passed the deploy green — which masked the
seed-drift crash until it was inspected by hand.
Add a third probe, `http://backend:8080/readyz` on the internal network, to the
same retry loop, and dump the backend logs on failure. A backend that cannot
become ready now fails the deploy loudly instead of going green.
Ilia Denisov