A missing/stale CEA cache must FAIL FAST where in-process builds are disabled.
Root cause of a production crash: switching to a propellant whose cache was not
committed (kerolox) made a request handler start a cold CEA build -- ~n_points^3
rocketcea calls whose growing in-process cache took the host to >12 GB. Builds
are an offline job (regenerate-cea-cache.yml); the deployed server disables them
by setting ED_ALLOW_CEA_BUILD=0 (docker-compose.yml) and must instead raise
immediately on a missing cache.
The gate is environment-based (env var), not code-path based, because local dev
runs the same backend over localhost and must keep building on demand. These
tests pin both directions.