I think this definitely conflicts with the behaviour we've been working on for the last week over on issue37834 (the PR is going to conflict for sure).

I am making a change to rmtree() that will cause it to delete junction point without trying to recursively remove their contents, which should avoid the listdir() that was failing here. That should be enough to cover the original issue.

We also ended up settling on isdir/islink semantics for directory junctions that don't quite match those discussed here. I'm inclined to agree that we should figure out the best way to distinguish volume mounts from directory junctions, but that is going to be best layered on top of the other changes that I'm working on getting merged right now. 

The commit messages or doc changes in PR 15231 are the best place to see where we ended up.