specify behaviour of floating point operations on NaNs #1711

For example, builtins like max, min, and clamp have specific behaviour on NaN inputs.
Those should be spelled out.