Hi there,
This workaround Walter wrote is indeed the more efficient way to preserve the track...
I've just answered another question about MM by suggesting at leat trying the MM node in the Fusion page instead, to compare. Yes, no "person" choice but the "object" one works great on people too.
And there, it's more usual and therefore handy and comfortable to feed the MM node by a very native node - in the beginning of the node flow.
Because yes, you actually want the MM node to remain alone or almost alone in its chain to finally pipe it in the end (you had understood, I'd prefer a merge node in Fusion than a key mixer in Color for this).
It's annoying in the beginning because not really natural sometimes to work this way, but keeping it from any changes prevent you from the boring retracking necessity ("regenerate tracking" in Fusion).
My guess -- to answer my own question -- is I was grading in HD just for speed, and then rendered in 4K. And my bet is that it had to retrack in 4K to make an accurate mask.
I don't think so but previously thought that too, until I experienced like written above that not only the change of resolution you wrote about was needing to regenerate the track, but also a lot (any ?) change in its chain of nodes.
For me it was my playing with a NR node activation/deactivation in the chain, also a color correction node tect etc...
Keeping it isolated from other common nodes and working on it in the end (the most you can) is the best option so far imho.