> Given that extra info, I'd say we're fine to document that our timeouts 
> can't do any better than the OS, which "for example, is typically 
> around 15ms on Windows", and recommend using non-blocking calls 
> instead.

The 15.625 ms resolution limit is fine, as long as performance is predictable. I don't like the random inconsistency introduced by extending only certain waits, in different ways, to support SIGINT and/or waits longer than 49.7 days. For example, time.sleep() doesn't ignore WAIT_TIMEOUT to recompute the remaining time, so it's not subject to the resolution limit that's imposed by GetTickCount64().

I'd prefer a common implementation of _Py_Sleep, _Py_WaitForSingleObject, and _Py_WaitForMultiple objects in order to be able to definitively state that all wait timeouts are unconditionally limited to the resolution reported by time.get_clock_info('monotonic').resolution; are not limited to 49.7 days; and can be interrupted by Ctrl+C in the main thread -- except for waiting on I/O. (There's an open issue to enable Ctrl+C to cancel synchronous I/O in the main thread -- such as reading from a pipe.)