What other holes are there that allow my pure functions to change their answer on different machines?

Floating point poses potential surprises. Library functions required by Haskell, but obtained from outside, can differ. IEEE FP modes may not be consistently set across all or even consistently at all times in one system if control ever leaaves Haskell code. Haskell's second-rate IO conversion functions cry out for replacement. I can imagine them being replaced in installations that really care about floating-point. Doug McIlroy