[Didée]

Thanks for the pointer, though it's only a reminder. This point already was mentioned earlier.

"Chroma shift" is a slippery formulation here. (Congrat's, if intended.) It's a valid denomination for what is happening, but predominantly it suggests sort of an "incorrectness", like e.g. wrong spatial position of chroma samples, or possibly a shift of hue. Which it is not.

It is not incorrect in any way. It is a bit suboptimal, yes. But it is still correct.
All it does is to take the end result, and adjust each chroma pixel a notch towards the chroma pixel of the initial EDI interpolation. Since the initial EDI interpolation can be assumed to be technically correct, so is the result of this operation. The only loss is a tiny bit of temporal stability, at worst the result will have some minor additional bob shimmer in the chroma channels. So little, and and in such places, that nobody ever would see a difference. Set up a blind test, it'll come in close to 50%.

With the affected older MVTools versions, it was a good thing. And after MVTools's correction, it doesn't matter. TGMC is not pixel-exact / value-exact anyway, and ... wait ... Ooohh, waitwaitwait! Make a PSNR comparison! Didn't try, but I bet that this shift operation increases the PSNR score of the result! Strike!

Leave it in, kick it out, whatever. It doesn't matter.