I've thought about this off and on for a while, and I have yet to come to API design that works well. Conceptually, with extreme control over niche optimization (that I'm pretty sure rustc doesn't have at the moment) it is possible to have a Result<FiniteF32, FpError> that is identical in representation to f32 and furthermore can implement Add<Result<FiniteF32, FpError>> in such a way that it compiles down to exactly the same as f32 addition (as infinities and NaNs in IEEE 754 are generally s...
