> Maybe it's considered undefined behavior for a subclass to use 
> a method of one of its bases which it has overriden.

In general, it's true that if OrderedDict is a subclass of dict, then it would have no defense against someone making a direct call to the dict base class.  Such a call should be expected to violate the OrderedDicts invariants.

> it's a complete coin toss on whether one of the many internal
> API function will set an attribute or name via PyDict_SetItem 
> or PyObject_SetItem. 

Not really.  The CPython source is supposed to only call PyDict_SetItem when the target is known to be an exact dictionary.  If you find a case where that isn't true, please file a bug report and we'll fix it.

> It might be more appropriate to start a new issue for this, but I'll > leave that decision to somehow who would know for sure.

No need.  We've known about this sort of problem for years.  See https://bugs.python.org/issue10977 for example.  There isn't really much we could do about it without causing other issues that would be worse.

FWIW, this doesn't seem to be a problem in practice.  Further, OrderedDict is expected to become less relevant now that regular dicts are order preserving.