> It would be good if the test timed out rather than deadlocking in the
> face of a broken import lock.

I could do that indeed. I'm not sure it would be much better, though, because the lock will still be held and posterior tests may freeze mysteriously.

> For your #9260 testing, this looks like it may be a little probabilistic

It always freezes here (with the per-module lock). I guess on modern machines, 0.5s is more than enough to launch a thread and parse a nearly empty Python module.

I should add a comment explaining a purpose of the test, though.