I just finished doing a clean test run across those 3 releases of Python with each of the 8 major config flags on/off (i.e. 48 runs of the test suite). As expected, almost all of the failures were in Python 3.x, and mostly related to unicode, where COPY_ON_ERROR generated failures when a unicode string was passed in; but those were almost all *usage* errors (i.e. errors in the test suite).
However, a nasty real bug in the non-accelerated buffer handler was found and is now fixed (the accelerated code uses the C API directly, so it worked everywhere, but the Python version is using ctypes wrapped calls).