The attached patch seems to correct this issue.  It contains the test attached yesterday, and it passes now.

I factored out the winner calculation from type_new to a new _PyType_CalculateWinner function, and type_new calls this.  I've put the declaration of this function into object.h, so __build_class__ can also call it, instead of using the metaclass of the first base.  (Am I correct in thinking that the underscore prefix keeps it out of the public API?)

A slight problem may be, that in some cases this function will be called twice.  But it is quite simple, so I don't think it matters much:

Without patch:
$ ./python -m timeit -- "class A(type): pass
class B: pass
class C(metaclass=A): pass
class D(B, C): pass
"
10000 loops, best of 3: 371 usec per loop

With patch:
$ ./python -m timeit -- "class A(type): pass
class B: pass
class C(metaclass=A): pass
class D(B, C): pass
"
10000 loops, best of 3: 381 usec per loop

(Note, that I generated the patch with hg extdiff, because the output of hg qdiff was much more unreadable than simple diff.  I can provide an equivalent patch generated by hg if needed.)