Replace the non-blocking accept + 50ms poll with a plain blocking
accept(). A signal-hook thread waits for SIGTERM/SIGINT, sets a flag,
and wakes accept() with a throwaway connection to the socket.
On shutdown the listener is closed and the socket file unlinked before
the grace period, so clients that connect late fail fast with ENOENT
instead of queueing and being reset. In-flight handlers are tracked by
a counter + Condvar rather than a Vec of JoinHandles.
Accepted streams now get a 5s read/write timeout so a silent client
cannot pin a handler thread or stretch every shutdown to the full
grace period. Transient accept() errors are logged and retried instead
of exiting the daemon. A second SIGTERM/SIGINT during the grace period
forces an immediate exit with status 1.
Tests: wrap the child in the kill-on-drop guard immediately after
spawn, drop the redundant wall-clock assertion, and add a case for a
client that connects but never sends.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Kmrgkpp9YBJ6AMaWYPPZuD
Running as PID 1 in a container, the daemon silently dropped SIGTERM,
so `docker stop` always hit the 10s timeout and SIGKILL, and host
reboots waited on it.
The accept loop now polls a non-blocking listener and a shutdown flag
registered via signal-hook. On SIGTERM or SIGINT it stops accepting,
gives in-flight requests up to 3s to finish, removes the socket file,
and exits with status 0. Log lines mark signal receipt and completion.
Adds integration tests that spawn the binary, perform an auth
round-trip, signal it, and assert a clean exit and socket removal.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Kmrgkpp9YBJ6AMaWYPPZuD