Now, it's possible that the test is failing to touch files in the
cache the same way that they are used in the test, because of all the
pathname normalizations that may happen, but that the discrepancy
somehow only shows up on Windows.
Ahem. Maybe I could normalize this
cache by using the namestring or native-namestring as the key, instead
of the pathname object? Probably.
OK, done, please try again (sigh).