The issue is trivially reproductible in any 32 bits platform, simply
allocating objects until you go up the 2GB mark.

Since __hash__() wants to take advantage of every bit in a 32 bit
platform, and we don't have unsigned integers in python, I vote for
"invalid" too.

There is no promise of "id(obj)==hash(obj)": you can overload
"__hash__()" anytime, and that is already done for strings, integers,
tuples, etc.

The speed advantage is interesting, though.