Reported here: kha-white/mokuro#106 but actually it's an issue with mokuro reader. .mokuro files contain a list of pages, each page with img_path key specifying which image does this item correspond to. Seems like the web reader ignores this value and relies on the pages being sorted the same way as images. This works most of the time, but sometimes the sorting is inconsistent which causes the mismatch.

Mokuro reader should use the img_path field to keep the association between the OCR results and image files.