With these changes, refleaks checks now pass:

 cpython gh-116731/importlib-refleaks @ ./python.exe -m test test_importlib -R 3:3
Using random seed: 894808194
0:00:00 load avg: 1.17 Run 1 test sequentially
0:00:00 load avg: 1.17 [1/1] test_importlib
beginning 6 repetitions. Showing number of leaks (. for 0 or less, X for 10 or more)
123:456
XX. 12.
test_importlib leaked [1, 2, 0] references, sum=3 (this is fine)
test_importlib leaked [1, 1, 0] memory blocks, sum=2 (this is fine)

== Tests result: SUCCESS ==

1 test OK.

Total duration: 14.0 sec
Total tests: run=1,274 skipped=6
Total test files: run=1/1
Result: SUCCESS

🤖 New build scheduled with the buildbot fleet by @jaraco for commit b33e4d5 🤖

If you want to schedule another build, you need to add the 🔨 test-with-buildbots label again.

🤖 New build scheduled with the buildbot fleet by @jaraco for commit b33e4d5 🤖

If you want to schedule another build, you need to add the 🔨 test-with-refleak-buildbots label again.

I believe this build was the one triggered for refleaks. It looks like it may have failed to start.

This is superseded by a cleaner approach in #116805. I still think MetadataPathFinder.invalidate_caches should be called from PathFinder.invalidate_caches, as that will match the behavior that would happen if MetadataPathFinder were on sys.meta_path. I'll track that in a separate issue.