It's possible that PeriodicJobImpl::stop() returns after PeriodicJobImpl::_run() starts a thread. If this happens while the mutex is locked such as here, then we get the "Assertion `INTERNAL_SYSCALL_ERRNO (e, __err) != ESRCH || !robust' failed" errors (owner of lock died).

We could just let the job thread run (and go through normal shutdown/canceled procedure) and join the thread in all cases. Or, the decision making on whether PeriodicJobImpl::stop() joins the job thread could be made atomic.