Support scanned PDFs: OCR fallback, figure/table cropping, text cleanup

Make the reflow engine work on scanned papers (e.g. the MIT Hackman/Oldham
1976 scan) that ship a page bitmap behind an OCR text layer.

- Drop full-page scan "backdrops" so the embedded text layer reads through
  instead of coming back as one giant image.
- Recover real figures/tables baked into the scan by their caption
  ("FIG. 1", "TABLE 2"): infer the artwork box, suppress the gibberish labels
  inside it, and crop it as an image. Skip the table's title block by width so
  a same-font title doesn't collapse the crop.
- Render crops in the page's native (unrotated) orientation so landscape
  tables on rotated pages come out upright and uncut.
- Judge "needs OCR" from the raw text-layer size (not post-suppression), so a
  table whose text sits inside its figure box isn't force-OCR'd into garbage.
- Tesseract.js OCR fallback for pages with no text layer at all; OCR output
  is run through the same gibberish/letter-spacing cleaning.
- Text quality: collapse per-glyph spacing ("o f j o b" -> "of job"), drop
  OCR gibberish, stitch sentences across page breaks, char-weight + width-gate
  body-size detection and add a length guard so body paragraphs aren't styled
  as headings on size-jittery scans.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: Eliott Gandiolle

README.md

@@ -29,7 +29,20 @@ The default URL is the Deci/Olafsen/Ryan SDT paper — just press **Make readabl
     1- vs 2-column layouts to recover reading order;
   - groups lines into paragraphs/headings (de-hyphenating line breaks);
   - walks the page operator list while tracking the transform matrix to find and
-    rasterise each embedded figure to a PNG.
+    rasterise each embedded figure to a PNG;
+  - ignores full-page **scan backdrops** (a page-sized image sitting behind an
+    invisible text layer) so the real text reads through instead of coming back
+    as one giant image — while recovering the real **figures/tables** baked into
+    that scan by their captions ("FIG. 1", "TABLE 2") and cropping them as
+    images;
+  - suppresses **garbled OCR debris** that scanned diagrams paint over their
+    artwork (and crops full-page scanned **tables** as images instead of leaking
+    their cells as text), stitches sentences that **continue across a page
+    break** back into one paragraph, repairs **letter-spaced** OCR words
+    ("o f j o b" → "of job"), and avoids mistaking a large-font body paragraph
+    for a heading on size-jittery scans;
+  - falls back to **OCR** ([tesseract.js](https://github.com/naptha/tesseract.js),
+    in-browser) for pages that have no text layer at all.
 - **`src/App.tsx` / `src/Lightbox.tsx`** — re-typesets the result and provides a
   full-screen zoomable image viewer (click a figure, click again for actual
   size, `Esc` to close).
@@ -47,8 +60,12 @@ Pages.
 
 ## Notes / limits
 
-- Works on PDFs that have a real text layer (most digital papers). Pure scans
-  with no text layer would need OCR — not included here.
+- Works best on PDFs with a real text layer (most digital papers), including
+  scanned papers that carry an invisible OCR text layer behind a full-page image
+  (the backdrop is dropped and the text reads through).
+- Pages with no text layer at all are OCR'd in the browser with tesseract.js.
+  This is slow (a few seconds per page) and downloads the language model from a
+  CDN the first time; accuracy depends on scan quality.
 - Column detection is heuristic; unusual layouts may interleave oddly.
-- The pdf.js worker is loaded from jsDelivr at the exact installed version, so
+- The pdf.js worker (and the tesseract.js core/model) are loaded from a CDN, so
   the page needs internet access the first time.

bun.lock

@@ -9,9 +9,10 @@
     "build": "bun build ./index.html --outdir=dist --minify"
   },
   "dependencies": {
+    "pdfjs-dist": "4.8.69",
     "react": "^18.3.1",
     "react-dom": "^18.3.1",
-    "pdfjs-dist": "4.8.69"
+    "tesseract.js": "^7.0.0"
   },
   "devDependencies": {
     "@types/bun": "^1.3.14",