Leak test of test_descr fails Base2 != Base

[nascheme]

Bug report

Bug description:

This seems to be caused by GH-133114.

Command to re-produce:

./python -m test -R 3:3 -j12 test_descr

The failure output:

FAIL: test_type_lookup_mro_reference (test.test_descr.MiscTests.test_type_lookup_mro_reference)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/buildbot/buildarea/3.13.cstratak-RHEL8-x86_64.refleak/build/Lib/test/test_descr.py", line 5181, in test_type_lookup_mro_reference
    self.assertEqual(X.mykey, 'from Base')
    ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^
AssertionError: 'from Base2' != 'from Base'
- from Base2
?          -
+ from Base
----------------------------------------------------------------------
Ran 156 tests in 0.829s
FAILED (failures=1, skipped=1, expected failures=2)
test test_descr failed

This appears to be another bug in the type lookup (MRO) cache.

CPython versions tested on:

3.13

Operating systems tested on:

Linux

Linked PRs

• [3.13] gh-133122: Fix for test_type_lookup_mro_reference. #133127

[nascheme]

Some progress on debugging this. It seems to be an issue with the specialize. In specialize.c, we have analyze_descriptor() calling _PyType_Lookup() and that's causing the tricky __eq__ to be executed before the real lookup happens. So the __bases__ gets replaced before the unit test expects. At least, that's my theory.

I'm not sure if we consider the bytecode specializer buggy for doing this or the unit test buggy for expecting it to work exactly as is being test. Perhaps you can not expect deterministic results when changing __bases__ while doing a type lookup.