Hmm.  Well, for one thing, we're falling back on Python's 
interlocked_cmp_xchg instead of using Windows' 
InterlockedCompareExchange (or should it InterlockCompareExchange64?).  
Python's implementation is clearly assuming 64 bit counters, but the 
other native Windows Interlocked* functions may indeed be operating on 
32 bit counters, hence the mismatch.