I think his point was that different fields should be treated differently. I.e. if you have two objects with the same ID but different descriptions then you can assume that it's the same object but with a changed description; but if you have two objects with different IDs but the same description then you should assume that the new object is completely different and the identical description is coincidental.

I don't agree that these are always the correct interpretations though. IDs could be reused (especially in a DVCS) or mistaken IDs could be corrected. This ambiguity is a fundamental limitation of the entire concept of diffing, that is reconstructing a set of operations to go from one state to another - you simply don't have the information to deduce the correct logical steps in all cases.