Since we can't use HTML5 contenteditable, I've switched to XTHML 1.0 Strict, which Kompozer can edit without borking it. Generated editable files and added them to the repo. Also abandoned the idea of using Git LFS; it just doesn't seem to work properly, and once it's configured and you have a large file it's tracking, it prevents any pushes because it fails on that file. Not ready for prime time. So the uncorrected OCR output is just kept locally by me for the moment.