1. I agree that we should fix the unlinking problem on Windows. I also think that such a fix should be independent of the test suite - many people run into failed unlink problems.

2. Tim already said it, but I repeat: the common theory is that the culprit for this kind of problem is software like virus checkers, desktop search spiders, Tortoise, ...

3. I'm not convinced that "rmdir/s/q" *really* solves the problem reliably. Because it's a timing issue, it may be that the additional startup cost to invoke rmdir was enough to let the virus scanner win the race, so that rmdir actually had no problems with removing the file.

4. I believe the official party line for removing files on Windows is this: "If DeleteFile fails, move the file to the trash bin (of the disk), and use NtSetInformationFile to set the delete disposition for the file". See cygwin's unlink_nt for an elaborate implementation of unlinking:

http://tinyurl.com/y7w6rrj