fix(ocr): extract PDF pages as PNGs before running kraken OCR

Kraken's -f pdf mode tries to write output next to the input file,
which fails on read-only mounts. Instead, extract pages as PNGs via
pypdfium2 (already installed), then run kraken on each image.
Both models run in a single container per PDF to avoid overhead.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

scripts/download-kraken-models.sh

@@ -99,21 +99,44 @@ run_evaluation() {
 
         echo "──── $pdf ────"
 
-        echo "  Model 1: $MODEL_1_NAME ..."
+        # Run both models inside a single container run:
+        # 1. Extract PDF pages as PNGs (pypdfium2 is already installed)
+        # 2. Run kraken on each page image for both models
+        # 3. Concatenate per-page output into one file per model
         docker compose run --rm \
             -v "$(cd "$IMPORT_DIR" && pwd):/eval-input:ro" \
-            -v "$(cd "$dir_1" && pwd):/eval-output" \
+            -v "$(cd "$dir_1" && pwd):/eval-out-1" \
+            -v "$(cd "$dir_2" && pwd):/eval-out-2" \
             "$COMPOSE_SERVICE" \
-            kraken -f pdf -i "/eval-input/$pdf" "/eval-output/${basename}.txt" segment -bl ocr -m "$MODEL_1_PATH" \
+            python3 -c "
-            2>/dev/null || echo "  ⚠ Model 1 failed on $pdf"
+import pypdfium2 as pdfium, subprocess, sys, os
 
-        echo "  Model 2: $MODEL_2_NAME ..."
+pdf = pdfium.PdfDocument('/eval-input/$pdf')
-        docker compose run --rm \
+pages = []
-            -v "$(cd "$IMPORT_DIR" && pwd):/eval-input:ro" \
+for i in range(len(pdf)):
-            -v "$(cd "$dir_2" && pwd):/eval-output" \
+    bmp = pdf[i].render(scale=300/72)
-            "$COMPOSE_SERVICE" \
+    path = f'/tmp/page_{i:04d}.png'
-            kraken -f pdf -i "/eval-input/$pdf" "/eval-output/${basename}.txt" segment -bl ocr -m "$MODEL_2_PATH" \
+    bmp.to_pil().save(path)
-            2>/dev/null || echo "  ⚠ Model 2 failed on $pdf"
+    pages.append(path)
+print(f'Extracted {len(pages)} pages')
+
+for label, model, outdir in [
+    ('Model 1', '$MODEL_1_PATH', '/eval-out-1'),
+    ('Model 2', '$MODEL_2_PATH', '/eval-out-2'),
+]:
+    print(f'  {label}...')
+    combined = ''
+    for p in pages:
+        args = ['kraken', '-i', p, '/dev/stdout', 'segment', '-bl', 'ocr', '-m', model]
+        r = subprocess.run(args, capture_output=True, text=True)
+        combined += r.stdout
+        if r.returncode != 0:
+            print(f'    ⚠ failed on {os.path.basename(p)}: {r.stderr[:200]}', file=sys.stderr)
+    with open(f'{outdir}/${basename}.txt', 'w') as f:
+        f.write(combined)
+    lines = combined.count(chr(10))
+    print(f'    → {lines} lines')
+" || echo "  ⚠ Failed on $pdf"
 
         echo ""
     done