That problem seems reasonably solvable. The main problem I see is that saccades are ridiculously fast, so not only does the eye tracking have to be extremely p

That problem seems reasonably solvable. The main problem I see is that saccades are ridiculously fast, so not only does the eye tracking have to be extremely precise, but the latency for the track-update-display pipeline has to be tiny. This problem is already difficult for head motion. Doing it at the eyeball level is way more ambitious.