>This seems to completely stump the algorithm, even when your eyes are plainly visible That's a false positive. This uses quite simple algos, decisition trees,

>This seems to completely stump the algorithm, even when your eyes are plainly visible That's a false positive. This uses quite simple algos, decisition trees, which have more false outcomes (both positive and negative) on this pupil detection task that state of the art models, which would be deep convolutional networks. The decision trees are much less computationally expensive though - running a modern deep CNN at 30 fps needs a dedicated GPU. The decision trees will run just fine on a CPU in a browser!