I did look at the code.  Indeed list and tuple compare x to e, while dict, set, and frozenset (well, I didn't check each one, just list and set) compare the e to x, where e is they key stored at hash(x).

Steve has a good point about the iteration.  And a user class can do anything it wants in contains.  I wonder if it would be even more accurate to say "conceptually equal" rather than "approximately equal", given that hash table 'in' doesn't do iteration at all.