The issue is that a later mzero in the transformer cannot undo an earlier action in a lower monad.

Precisely. Another example, for fun:

>>> print (runMaybeT $ lift [1,2,3] >> mzero :: [Maybe Int])

[Nothing, Nothing, Nothing]

>>> print (runMaybeT mzero :: [Maybe Int])

[Nothing]

Actually, I'd say the problem isn't that a transformer "cannot undo earlier action in a lower monad." I'd just say that most transformers happen to be implemented in this "forgetful" manner. It's conceivable that we could implement some of them differently, in a way that would obey the MonadPlus laws. Whether or not obedience to this particular law is worthwhile... well, that's debatable.

-- Dan Burton