@richardsheridan, do you mind please reviewing the new test. A particular note: can you think of a way for this test to not require a subprocess? (we need to test for the crash without crashing the main test interpreter)

1. I think you should use a threading.Event instead of a timer to make sure the thread reference outlives the main reference. This would save some time.
2. Sadly, I too think you'll need a subprocess to avoid crashing the main test interpreter if we experience a regression. However, do you really need to avoid that crash? I used a subprocess for a test (fix FigureManagerTk close behavior if embedded in Tk App matplotlib/matplotlib#17802) because the regression failure lead to an infinite hang which made CI waste a lot of time (20 minute timeout...). The interpreter crash at least gives immediate feedback even if the test runner doesn't print out all the nice messages. You've already fixed the bug so the happy path should be as fast as possible, and the single leaked tcl interpreter is mostly harmless. I think this test could just be on the main MiscTest class, as long as you put a new self.root on when you're done.

1. Thank you for your idea on using threading events: I have now used two of them to replace both time.sleeps (writing this, I remember I can now remove that import!)
2. I feel we need to give the standard "Test: FAIL" rather than just crashing the main test interpreter. I have tested the subprocess on a build without the _tkinter changes and it did not cause the test to hang (I cannot think of a situation where it would). For the sake of conformity, I will stick by the fact that it should not crash the main test interpreter (though, when tested, it did give a non-zero return-code meaning the test would show as failed).

All commit authors signed the Contributor License Agreement.

This PR is stale because it has been open for 30 days with no activity.

Picked this up in #152323 — same approach (skip Tcl_DeleteInterp() when deallocating in the wrong thread, leak the interpreter, warn), rebased onto current main and with you credited as co-author.

Two changes from this draft: the warning now goes through the warnings machinery (PyErr_WarnEx) so it can be filtered/suppressed, falling back to an unraisable report only if it's escalated to an error (a deallocator can't propagate exceptions); and the test asserts a clean process exit. Thanks for the original analysis and patch.

Thanks for this, and for the original diagnosis! This was implemented in #152323 (merged), with you credited as co-author. Closing as superseded — much appreciated.