Build and tune a page triage

The triage is three numbers per page and four rules. Character count says whether there is a text layer at all. Image coverage says whether the page is a scan (over half covered), a typed page with an embedded map (15–50%) or a typed page. For scans that do carry a text layer, the plausible-word ratio decides between "trust it" and "re-read the image". The work is in the page pack: it has to contain a scan with a good OCR layer next to the corrupted ones, or the report can't tell you whether your rule separates them.