That sounds right to me.
Note that PEP 399 is following older conventions that were laid down in a time when unittest did not *have* test discovery, so this is a new paradigm we'd like to move to (for the reasons Brett mentioned), and it may take a while to get there. It applies to more than just the python/C accelerator distinction; it applies any time a base class plus specialized classes are used to construct test cases. (I do this a bunch in the email tests, for example, and that has no accelerator.)