Is it mandatory that the significand bits are used for that purpose, or could there be an additional dedicated bit in the type?
For example, Float80 has an extra bit which indicates, essentially, isNormal. Could the same strategy be used for distinguishing NaNs?