I'm not sure about this. It would seem like a separable concern, and I tend to be very supportive of concern separation.
If one were to specify such an API that is defined in terms of a stream of VideoFrame, that would offer instant pluggability without any spec linkage needed (and a native implementation could save some memory copies by operating on the internal representation and only surface the actual tracking result).

But this may be out of webrtc scope - is anything like this being pursued in (for instance) the AR/XR groups?

I don't beleive it is separable. This proposed API claims to enable Funny Hats. If delivered without a non-discriminatory face tracking implementation it will cause lazy/discriminatory sites to proliferate.

Have a listen to the first 5 mins of a recent podcast I co-host https://distributedfutu.re/#episode56 for a sense of the scale of the problem that already exists: 'not working for darker skin tones .... Has become the norm now for me' says my co-host.

@steely-glint I have added this to the agenda for Wednesday Dec 2's WG meeting. I'd like to get a firm grasp of "next steps".

Closing, since my understanding is that we have decided to work on this and other potential platform transformations in a separate document (maybe mediacapture-extensions initially).