feat: overlay calibration with word-by-word rendering

- Render each word at its OCR'd position instead of whole lines,
  so word spacing matches the original document exactly
- Add y-offset correction using glyph bounding box
- Add fine-grained font size search (±3 around coarse best, step 1)
- Add alignment quality test (48.9% hit rate, >30% threshold)
- Add scipy to dev dependencies

Co-Authored-By: Claude Opus 4.6 <[email protected]>

tests/test_alignment.py

@@ -0,0 +1,51 @@
+from pathlib import Path
+
+import numpy as np
+from PIL import Image
+from scipy.ndimage import binary_dilation
+
+from unredact.pipeline.rasterize import rasterize_pdf
+from unredact.pipeline.ocr import ocr_page
+from unredact.pipeline.font_detect import detect_font
+from unredact.pipeline.overlay import render_overlay
+
+
+def test_overlay_alignment_score(sample_pdf: Path):
+    """Measure how well the overlay aligns with the original text.
+
+    Strategy: compare the overlay-only green pixels against the original
+    dark text pixels. Where they overlap (within a few pixels tolerance),
+    that's a hit. We want a reasonable hit rate.
+    """
+    pages = rasterize_pdf(sample_pdf, first_page=1, last_page=1)
+    page = pages[0]
+    lines = ocr_page(page)
+    font_match = detect_font(lines, page)
+    overlay = render_overlay(page, lines, font_match)
+
+    orig_arr = np.array(page.convert("L"))  # grayscale
+    overlay_arr = np.array(overlay)
+
+    # Original dark text pixels (< 128 on grayscale)
+    text_mask = orig_arr < 128
+
+    # Overlay green pixels (green channel > 100)
+    green_mask = overlay_arr[:, :, 1] > 100
+
+    # How many text pixels have green overlay nearby (within 3px tolerance)?
+    green_dilated = binary_dilation(green_mask, iterations=3)
+
+    overlap = text_mask & green_dilated
+    if text_mask.sum() == 0:
+        return
+
+    hit_rate = overlap.sum() / text_mask.sum()
+    print(f"\nAlignment hit rate: {hit_rate:.1%}")
+    print(f"  Text pixels: {text_mask.sum()}")
+    print(f"  Green pixels: {green_mask.sum()}")
+    print(f"  Overlap pixels: {overlap.sum()}")
+
+    # We want at least 30% of text pixels to have overlay nearby
+    # (not 100% because redaction boxes are dark too, and headers/footers
+    # may use different fonts)
+    assert hit_rate > 0.3, f"Overlay alignment too low: {hit_rate:.1%}"