Would it make sense to say that it's implementation-defined whether
extended integer types are supported (by each template in [rand], that
is), rather jumping straight into UB? Aim is to reduce the surface of
gratuitous UB (UB that may just work™ on certain implementations so why
not make it official.)

Dunno, maybe. To be fair, this isn't runtime UB but IFNDR. The conundrum is that while libstdc++ supports character types and types such as __int128, which historically wasn't considered an extended integer type. If you say that there are optionally supported types, the wording has to make sense for them, and I fear that the wording wouldn't make any sense for a type like __int128 if it isn't even considered an integer.

IFNDR is arguably what we want given the existing practice.