I get the render region feature and it works very well, I use it often.
Maybe my wasn't clear. It's not about single frames but animations.
But imagine having an animation where the camera might go through a room and certain objects tend to produce more noise.
If one could use a mask render output of those objects and apply it to a fictitious extra sampling channel for the renderer, it could be resolved without having to go over the frames manually.
If the adaptivity can handle it at the same performance level, then it's fine.