Well, it's embarrassing...
I'm still seeing (rarer) crashes under Windows, and indeed there was a second race condition (on making the GL context current to the thread being initialized) in that code...
I fully reworked the code and will publish fixed releases today.