I would guess that the “foveated” region that they stream is larger than the human fovea, large enough to contain the saccades movement (with some good-enough probability).
Are you really sure overrendering the fovea region would really work?
I’m not sure what you mean by “look through the entire image to reacquire the iris”? You’re talking about the image from the eye tracking camera?
Yes. A normal trick is to search just a bit outside the last known position to make eye tracking cheap computationally and to reduce latency in the common case.