One problem with that is that it processes images lazily. It continually extrapolates cursor moments ~1 second into the future and processes those relevant part

One problem with that is that it processes images lazily. It continually extrapolates cursor moments ~1 second into the future and processes those relevant parts of relevant images. But it should be possible that after an image is processed (or even eagerly by looking up previously recognized regions from the cached OCR server), the page could be made Ctrl+F-able.