I dealt with the comments folks made here and on python-dev regarding the docs and the limited API, as well as the fact it isn't safe to call PyErr_NoMemory() when this function fails.

However, I still haven't been able to reproduce the failure seen on the buildbots. It isn't a simple cross-test interference problem, as I tried the randomization seed from the Ubuntu log and that still passed for me. I'm still trying the seed from Stefan's FreeBSD system, but I expect that will be clean here as well.