So what's the point of getOCRText? None of document implementations support it.

It does say "not implemented yet". :-) As you probably noticed it says this where it's designed to use it:

My guess is it was intended as a follow-up to koreader/libk2pdfopt@537d0a5

Based on ocrtess_ocrwords_from_bmp8 rather than ocrtess_single_word_from_bmp8.
https://github.com/koreader/libk2pdfopt/blob/6e79f026b044beef0015f3ddedc818a6d4a01f44/willuslib/willus.h#L1692-L1694

I'll note that this has already changed in k2pdfopt_tocr_single_word in the meantime as part of your recent update, though of course it doesn't take sboxes. I suspect OCR output might be better if it had more context to work with? Or maybe it'd just be for performance reasons.

So that comment is wrong and misleading ("is not implemented in some document backends").

It would seem so.

OK, so another libk2pdfopt change and corresponding base update is going to be needed… In no particular order:

• ocr_type, passed to ocrtess_ocrwords_from_bmp8 from k2pdfopt_tocr_single_word, was previously unused, but should be changed to -1 (the default: assume a single uniform block of text): this seem to make a big difference for Arabic.
• for the getOCRText / getOCRWord situation: "implement" it in koptinterface (move the workaround from ReaderHighlight:lookup) (this will also fix ReaderHighlight:translate when OCR is involved).
• there's an issue with text selection when the document language is Chinese: if the tesseract API as not been initialized, then k2pdfopt_get_reflowed_word_boxes & k2pdfopt_get_native_word_boxes will fail when k2pdfopt_get_word_boxes_from_tesseract is involved…

All sounds good. 👍

• there's an issue with text selection when the document language is Chinese: if the tesseract API as not been initialized, then k2pdfopt_get_reflowed_word_boxes & k2pdfopt_get_native_word_boxes will fail when k2pdfopt_get_word_boxes_from_tesseract is involved…

Possibly related: I never looked into it, but there's potentially a fair number of old open issues about crashes on CJK when OCR and/or reflow are involved (issues that, IIRC, may have started since the previous mupdf bump ages ago ;)).