This is weird. In most itemdecls we never say the class name. This "random" business is very idiosyncratic, though. But I'd call this change a step backwards.

It is weird, but the constructors have A::A.

I haven't seen this pattern of listing requirements by declaring the member functions elsewhere. Usually we have a list of expressions and required semantics, e.g. a.seed() instead of void seed();. The function templates are easier to spell by showing the declarations though.

I can drop this commit if it's not good and just remove the void.

tuple, pair and vector don't use this pattern.

Yeah, maybe drop this commit except the void fix. Feel free to open an issue if you like.

I don't think tuple, pair and vector are really comparable. They are standard types defined with a synopsis and a detailed description of each function like a lot of others.

This section lists "adapter requirements" that any type can meet, and is more similar to the iterator categories, Allocator requirements, etc., which we specify with a table of expressions that must be valid and have some required semantics. In fact the surrounding subclauses [rand.req.seedseq], [rand.req.urng], [rand.req.eng] and [rand.req.dist] all use this approach too. I have no idea why [rand.req.adapt] is different like this. For now I'll leave it alone, just removing the void.

I see. OK, let's revisit that at some point.