Florent’s manual reproducer is fixed by issue18383_34_2.diff, as is test___all__. And I think the new logic for changing the priority of the filter entries might be sound now. But it does seem like we are adding unnecessary complexity to the implementation just to avoid a quirk of the test suite. Maybe it would be better to save and restore the original filters in the test suite, with a comment explaining why. Or perhaps set a _warnings._options_processed flag to prevent the double initialization?

FTR Alex already hinted at this, but I only now realize issue18383_assert_warns_and_dups_v2.patch contains extra fixes for assertWarns() that are only applicable to 3.5+. They are basically a followup of Issue 16382 and Issue 23027, rather than the original issue here, although the symptom is almost the same.