I haven’t made any further progress. My suggestion would be to file a bug. I’m not really convinced that it is a bug. I have investigated issues like this in the past with persistent kernels along with other “simultaneous” activity, and eventually went to the dev teams because I couldn’t understand the behavior. In at least one case, the response that I got back was that it was expected behavior and the only documentation support was the note that any CUDA API call can have variable latency.
I’m not saying this is that, or anything really, except that it might be a bug or it might not. But by filing a bug, the dev team will usually look at it.