Cross-Branch Testing

Not if you're writing your unit tests correctly. Thought should go into what you're testing. It might be true that you're going to find an error in 1% of cases, but you shouldn't be shooting random shit into your test. Or in other words you should "constrain" your "random test". I find it difficult to think of a situation where you actually don't know any helpful mathematical properties of your code - since you've obviosuly written the code with an intention to perform some function. There may be good reason for doing this stupid rotations and powers etc. But at the end of the day there was some non-code motviation for writing this code and that is what you should be thinking about when writing your unit tests.

What the cross-branch testing gives you is an assurance that your code is as broken as when you started. Now, think about what is happening here when you test across branches, you have your golden reference model (which might or might not be right) and you have some python packages that you hope are doing a better job than you at constrained random testing. In fact, their odds of hitting that 1% are no better than your unit tests.

I actually think to some extent testing against your previous implementation has value above the unit test - because my experience is that the first person to write the code probably thoroughly manually tested it and didn't convert all those tests into unit tests, and therefore your confidence in the original implementation is higher than revisions, but the lesson to learn from that is to make sure when you first implement something you push as much of your verification into unit tests.