[autoantohaki]

Requested by mau

Before: serve only listened for Ctrl+C. docker stop, kill, and service managers send SIGTERM, which skipped the graceful drain entirely. In the container the server runs as PID 1, where an unhandled SIGTERM is ignored, so docker stop hung for its full grace period and then SIGKILLed the process, abandoning in-flight requests and leaving the llama.cpp child to be killed without the bounded shutdown path.

After: on Unix, SIGTERM triggers the same bounded shutdown as Ctrl+C: HTTP and gRPC admission stops, in-flight work drains within --shutdown-timeout-ms, the private runtime is terminated, and the process exits 0.

How: a small shutdown_requested() helper in main.rs waits for either Ctrl+C or SIGTERM (Unix only; Windows keeps Ctrl+C). A new Unix-only binary test starts the real server, sends kill -TERM, and asserts a clean exit. The README container section mentions the behavior.

CI needs #1 merged first (it restores the privacy-scan and supply-chain checks on main).