Turns out that it creates two representations of the PDF content - one is text, which has 0 opacity and another is prerendered image of the same layered on top of it. The latter is what's displayed. So while you can enable contenteditable and get the editing cursor, no changes you make will be shown. Oh, well.