Thanks for checking that. My test was exactly succeeding on a flavour of Rog Strix laptop with builtin camera on Windows 11. Ironically right now I am without that laptop around and don't have any Windows setup with a camera, I guess I go and get USB camera on the weekend and also test with that one.
From our vastly diverging results (including all the way weird error codes on both) I suspect that the actual behaviour on Windows depends on camera driver. On my previous setup it was exact opposite to what you observed: the device which started using camera first got it non-functional while the second graph was working.
So I guess it contributes to that trying to implement the actual capture preemption looks a is a moot point now, all I want to do is to let that enumeration and filter creation succeed. So probably in the end I would just remove those setup specific tests? Yet I'd try that with another camera here.