Hmm.  I hadn't considered the reproducibility problem.

Does the module aim for reproducibility across all platforms *and* all versions of Python?  Or just one of those?

For small n, I think the patched version of randrange(n) produces the same sequence as before with very high probability, but not with certainty.  Since that sounds like a recipe for hard-to-find bugs, it might be better to deliberately perturb the outputs somehow so that the sequence is obviously different from before, rather than subtly different.