I think the real answer is to test a lot of middleboxes, come up with some fallback behaviour that will at least work with 99% of them, and then just break the last 1%.

The same happened with Max Segment Size detection - and that mostly works now.