This makes very little sense to me...

Two separate successful AppVeyor runs confirm it's related, though.

Combined with the fact that adding -v to the tests seems to move the segfault to the end and that @cgohlke reports that this is dependent on a python mirco version makes me hope that this is something upstream in how BytesIO or bytes are handled.

It looks like there is a path (which we don't test?) that saves everything into memory and then embeds the frames. I think the code path we are testing is writing to a temporary directory.

It's now the third time around and still passing...