>>> class MyDict(dict):
...     def keys(self): print "keys"
...     def __getitem__(self, n): print "__getitem__"
...
>>> myDict = MyDict(a=1, b=2)
>>> dict(myDict)
{'a': 1, 'b': 2}

PyDict_Merge accesses the items of the dict to be merged directly rather than going through the interface for any dict instance--even a dict derivative--by virtue of using PyDict_Check rather than PyDict_CheckExact. I believe the logic needs to be:

if type(d).__getitem__ is dict.__getitem__ and type(d).keys is dict.keys:
    ...okay to access items directly...
else:
    ...go through the methods...

IIRC, the same behavior is present for subclasses of other builtin types.

It is documented that access to special methods is more direct than normal path. This is not going to change. A similar example was discussed on python-list earlier this year with the above conclusion.

I concur with Terry.
This is just a fact of life when subclassing builtins.
The objects are "open-for-extension, closed-for-modification".
If you want more direct control use UserDict.