I don't have much of a horse in the race, but since the test has historically been skipped on Windows, and the test hasn't and doesn't work on Windows, modifications to restore the skip behavior seem reasonable to me.  The trigger for this issue was Windows adding support for underscore in locale names (like tr_TR) so the test began executing.  But it's not a regression or new issue, it's just existing reality becoming exposed.

The user machine and buildbot discrepancy can be attributed to version differences, as the buildbot hadn't yet received the same underscore locale name support.

I'd be fine with removing the test entirely - always skipping on a failure just seems pointless.  Then again, issue1813 created the test for a purpose on other systems, though even back then it appears it was complicated.  Leaving the test but skipping known failing systems (I guess at least Windows and OpenBSD) might be slightly less intrusive of a change, assuming the test is still serving a purpose elsewhere.

Separately, there's a lot of useful/interesting detail here which could inform any eventual normalization changes on Windows, should the underlying issue be deemed worthy of addressing.  But that seems like something that could be a distinct operation from clearing up the test issue.