Proposal thatfixesGH-127049. The idea is to revertGH-121126 and re-fixGH-87744 in a way that doesn't introduceGH-127049. The proposed approach here is to change child watchers to behave closer to their analog on Windows (_WindowsSubprocessTransport), i.e. letsubprocess handle the PID lifetime management instead of havingsubprocess do the allocation andasyncio do the deallocation, i.e. follow the guidance of#86724 (comment).

In the_ThreadedChildWatcher case this borrows thewaitid +WNOWAIT idea from Trio. (See also#82811 (comment).) Note that in the_ThreadedChildWatcher case, while it would ideally be safe to just callPopen.wait in the thread instead of involvingwaitid +WNOWAIT,Popen currently has some thread-unsafeties, so runningPopen.wait orPopen.poll in the thread in practice resulted in unsafekill calls and brokentransport._process_exited(returncode=None) calls. See the longer commit message.

I have marked this as a draft because it does not yet have regression tests. Both cases (_PidfdChildWatcher and_ThreadedChildWatcher) can have (nearly-)deterministic (but consistent regardless) regression tests, but I am not sure how folks would want them implemented, as the reproducers I put together for the report involve monkeypatchingos.waitpid andos.kill (which is nontrivial in part becausesubprocess holds strong references toos.waitpid, and which will miss anyos.waitid calls made withoutWNOWAIT unless we patch that too).

Anyway, if this PR is pursued, I'd appreciate some help with adding tests. This is also my first PR proposed to CPython so any feedback/pointers are appreciated. :)