I should say, I don't actually care about this bug at all: I only ran into it because of trying to recreate the random API and testing that my recreation worked sensibly. I thought I'd do the good citizen thing and report it, but I'm totally fine with any resolution and given that that code has been unchanged for at least 18 years I doubt anyone else has ever run into this problem.

I do think it would be beneficial to document the range of reasonable alpha on paretovariate - the lack of such documentation is the only reason I hit this at all - but I think leaving the behaviour as is would be fine.