I'd generally limit wider to apply to the same, uh, type classes (there's probably a proper word) int8 to int16, float to double and so on.

Casting between integers and floating point is pretty risky unless you're absolutely sure about your ranges and precision. Anything that does that silently gives me The Fear.