🧪 test(fork): fork once the event loop has closed

[gaborbernat]

Closes #707.

test_cancelled_async_acquire_unregisters_reused_descriptor forks from inside the coroutine, one statement after await loop.shutdown_default_executor(). That await joins the executor worker the canceled acquire left behind, but Thread.join() returns as soon as the interpreter releases the thread state, while the OS thread is still unwinding. NetBSD deadlocks a child forked from a process that still carries one, so the fork wedges, os.waitpid never returns, and the 10 second subprocess timeout kills the interpreter with SIGKILL. That is the returncode: -9 in the report.

The descriptor swap and the fork move to module scope, after asyncio.run() has closed the loop and the runner has shut the executor down. The script then leaves the worker time to go and checks threading.active_count(), exiting 6 rather than forking on a thread that is still counted. The subprocess timeout goes to 30 seconds so a slow builder is not the failure mode either.

I have no NetBSD machine, so this is reasoned from the report rather than reproduced: the run reaches the fork and hangs there, and the window between the join and the fork is the only NetBSD-specific thing in the script. @0-wiz-0, could you run the suite with this branch? If it still fails, exit code 6 says a thread outlived the loop, and any other outcome puts the deadlock elsewhere, where a backtrace of the wedged process would tell us where.

[0-wiz-0]

I've tested 3.32.3 with this patch applied on top, and get 1248 passed, 58 skipped - the test failure is fixed. Thank you!