Here's an example patch that removes any bias from randrange(n) (except for bias resulting from the imperfectness of the core MT generator).  I added a small private method to Modules/_randommodule.c to aid the computation.

This only fixes one instance of int(random() * n) in the Lib/random.py source;  the other instances should be modified accordingly.

With this patch, randrange is a touch faster than before (20-30% speedup) for small arguments.

Is this worth pursuing?