fix: Re-dispatch SIGTERM to the handler add_signal_handler displaces

[selimacerbas]

What happens

asgi_app installs its SIGTERM handler on the first request it serves:

loop.add_signal_handler(signal.SIGTERM, _on_sigterm)

Underneath, in both uvloop and stdlib asyncio, add_signal_handler installs a dummy handler through signal.signal. That silently replaces whatever the hosting ASGI server registered. uvicorn registers Server.handle_exit in Server.capture_signals(), so once the SDK has served one request, uvicorn no longer has a SIGTERM handler.

_on_sigterm notifies the active receive channels and returns. Nothing tells the host server to shut down, so a uvicorn-hosted endpoint stops responding to SIGTERM at all.

What I measured

Endpoint served by uvicorn in a container with a 45 second stop grace period:

• docker kill --signal=SIGTERM leaves the process running. Still serving at T+72s.
• /proc/1/status shows SIGTERM caught, not blocked and not ignored, so the signal does arrive.
• signal.getsignal(SIGTERM) returns asyncio's _sighandler_noop rather than uvicorn's handle_exit.
• The listening socket stays open. uvicorn never starts its drain.
• Docker force-kills at the end of the grace period. Exit 137 on every stop and every deploy.

The change

Remember the handler that add_signal_handler displaces, then call it from _on_sigterm once the channels have been notified. The re-dispatch runs on the event loop rather than in signal context, which is why frame is None.

If there was no previous handler, meaning SIG_DFL or SIG_IGN, nothing is called and the standalone case behaves as before.

I kept add_signal_handler instead of switching to signal.signal, since _on_sigterm touches asyncio objects and wants the loop-deferred dispatch that add_signal_handler gives it.

Tests

Two tests in tests/server.py:

• a host handler installed through signal.signal gets re-dispatched after a real raise_signal(SIGTERM)
• raising SIGTERM with no previous handler installed does not error

Notes

I ran into this on Restate workers that were being force-killed on every deploy. We worked around it downstream by re-asserting uvicorn's handler from a poll loop. That holds, but the displacement seemed like something the SDK should deal with rather than each host.