[Raymond Hettinger]
... The C code I saw is covered by a BSD license -- I don't know if that's an issue or not.
That's fine, provided it doesn't have the dreaded "advertising clause". I personally don't care whether it does -- it's the FSF that has bug up their butt about that one. I expect we'd have to reproduce their copyright notice in the docs somewhere; yup: 2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. I think we *ought* to perform a similar courtesy for, e.g., the Tcl/Tk and zlib components shipped with the Python Windows installer too.
As for implementation difficulty or accuracy, the code is so short and clear that there isn't a savings from re-using the C code.
That isn't the point here. If you use Nishimura and Matsumoto's code as close to verbatim as possible, then that's the perfect answer to your earlier point:
On the minus side, random number generation is a much disputed topic, occassionly requiring full disclosure of seeds and source.
Nothing *could* be more fully disclosed than their source code: it's extremely well known to every worker in the field, and has gotten critical review from the smartest eyeballs in the world.