I agree that the protocol should be constrained to class types, but having the generator not be generic would prevent the optimizer from specializing and inlining it, and I'd say efficiency is very important here.
And more generally, I still think that there hasn’t been nearly enough (hardly any?) actual practical usage/testing/experimentation using the proposed API in order to know how, or even whether or not, it can be successfully extended with custom (seedable) PRNGs, and eg Gaussian distributions etc, and IMHO it sounds like a lot of people here are OK with just assuming that this API is a good starting point and that everything will just work out even if we haven't actually tried to do anything real with it (ie try to use it for some little game with procedurally generated graphics and levels, or a little raytracer (which could use 2d gaussian values for multii-sampling) or something).
I haven’t looked into the playground of @nnnnnnnn which seems to be the only effort that has been made in this direction. But for example, it's impossible to measure efficiency in a playground settings.
My guess/feeling is that I will have to continue using my own little Random library (that I have been using, testing and modifying for a couple of years) even If this proposal is accepted. My requirements are PRNG, and the generators and the stuff based on them (conversions into various floating point ranges, other distributions, etc) have to be as fast as is possible, and I've sometimes needed different generators for different scenarious, sometimes trading speed for quality, perhaps a generator with a state that is of suitable size for a particular task, etc. I have never needed a non-seedable cryptographically secure RNG, but I understand that there might be others that have, and I understand the rationale of having it be the default. My use cases have been graphics/audio/signal processing (which would include games with procedurally generated graphics, levels and stuff like that).
I don’t have the time to dig deeper in this (trying out and/or providing example code etc) but maybe these negative ramblings of mine can provide some sort of non-negative value anyway.