Nigel has been letting me know things are not 100% yet. He has a very
rare scenario of a crash, and I have some issues with launching
firefox whilst tracing all system calls.
Looks like at least one isnt totally correct for some reason.
May take a while to debug these scenarios (syscall tracing should be
deterministically easy), but 1:1,000,000,000 is more tricky to
diagnose. Maybe I can add enforced long delays to force interrupt
traffic to jam up across the cpus.