Hmmm, this might have been flaky for a while. I didn't realize it then, as I thought it was related to the crashdump change and it "went away" after a sync, but it was probably a flakiness I saw. My guess is that the recent regression is caused by changed timing due to the scavenger changes, but that's just a guess. The flakiness should be fixed anyway, I'll put it on the list.
I was suspecting that something was spooky here
The exception technically comes from non-go code, so I was wondering if that's simply the reason we don't get a stack trace and - for some reason I don't understand - the test for non-go-code did not come to the same conclusion before? But I could completely misunderstand it too. When I tested, an exception I caused "directly" by e.g. reading a protected memory page gave stack dumps, but when I raised the exception by calling RaiseException, it became "flaky" somewhere between 1.17 and 1.18.