Commenting on the latest patch here, since the Rietveld integration isn't coping with the "hg extdiff" output (I would guess that the revision numbers in the pathnames are confusing the script that attempts to determine the base revision).

Aside from the note above about needing to restore the special case handling for non-type metaclasses, the only other recommendation I have is for the new metaclass invocation tests to add a second metaclass to the hierarchy and explicitly check the order of the __new__ calls, as in the original examples above that invoked the wrong M_A/M_B/M_A sequence. Those tests will then also be applicable to Python 2.7 (which doesn't have __prepare___).