So I'm guessing something is just borderline under 3.9 on Windows.

In some manual testing with a standalone build of 3.9 so far for me:
  -m test.test_pickle         always succeeds (executed directly)
  -m test test_pickle         always fails (executed via test module)
  -m test -w -j1 test_pickle  fails, but succeeds on retry

The failures seem to always occur in CPicklerTests.test_bad_getattr.  I'm not sure how to run that single test via the test module, but limiting to all CPicklerTests tests or all test_bad_getattr tests succeeds even through the test module.

The last scenario above (successful retry) has to use -j or else no retry (-w) takes place.  That's the path the buildbots are following.