Erweitere Regression Test (Bildvergleich) aber deaktiviere ihn default, weil das diff noch zu groß ist.

tests/test_regression.py

@@ -33,6 +33,7 @@ When a page fails the threshold, three images are saved to
 """
 from __future__ import annotations
 
+import json
 import re
 import subprocess
 import sys
@@ -131,26 +132,46 @@ def _set_need_appearances(pdf_path: Path, enabled: bool = True) -> None:
         writer.write(fh)
 
 
-def _compute_similarity(ref: np.ndarray, gen: np.ndarray) -> float:
-    """Return the fraction of pixels where all RGB channels differ by at most PIXEL_TOLERANCE."""
+def _compute_similarity(ref: np.ndarray, gen: np.ndarray, mask_footer: bool = False) -> tuple[float, int, int]:
+    """Return (similarity, different_pixels, total_pixels).
+    
+    If mask_footer=True, ignores the bottom ~20% of each image (footer area).
+    """
     if ref.shape != gen.shape:
         gen = _resize_to_match(gen, ref.shape)
+    
+    # Apply footer mask if requested
+    if mask_footer:
+        h = ref.shape[0]
+        mask_start = int(h * 0.80)  # Ignore bottom 20%
+        ref = ref[:mask_start]
+        gen = gen[:mask_start]
+    
     diff = np.abs(ref.astype(np.int32) - gen.astype(np.int32))
     matching = np.all(diff <= PIXEL_TOLERANCE, axis=2)
-    return float(matching.mean())
+    similar_count = np.sum(matching)
+    total_count = matching.size
+    
+    return (similar_count / total_count, total_count - similar_count, total_count)
 
 
-def _save_diff_images(ref: np.ndarray, gen: np.ndarray, out_dir: Path, page_label: str) -> None:
-    """Save reference, generated, and heatmap diff images for a failing page."""
+def _save_diff_images(ref: np.ndarray, gen: np.ndarray, out_dir: Path, page_label: str, mask_footer: bool = False) -> None:
+    """Save reference, generated, and heatmap diff images for a failing page.
+    
+    If mask_footer=True, the bottom ~20% (footer area) is shown in gray in the diff image.
+    The diff is computed BEFORE masking to show actual differences, then the masked area is grayed out.
+    """
     if ref.shape != gen.shape:
         gen = _resize_to_match(gen, ref.shape)
 
     out_dir.mkdir(parents=True, exist_ok=True)
 
+    # Save full reference and generated images
     Image.fromarray(ref).save(str(out_dir / f"{page_label}_reference.png"))
     Image.fromarray(gen).save(str(out_dir / f"{page_label}_generated.png"))
 
     # Heatmap: identical areas dimmed, differences highlighted in red
+    # NOTE: diff is computed on FULL images, not masked, to show all actual differences
     diff = np.abs(ref.astype(np.int32) - gen.astype(np.int32))
     max_diff_per_pixel = diff.max(axis=2)
     severity = max_diff_per_pixel / 255.0
@@ -162,6 +183,12 @@ def _save_diff_images(ref: np.ndarray, gen: np.ndarray, out_dir: Path, page_labe
     blended = (base * (1.0 - alpha) + overlay * alpha).astype(np.uint8)
     blended[max_diff_per_pixel > 30] = [255, 0, 0]
 
+    # Gray out masked footer region in diff image
+    if mask_footer:
+        h = ref.shape[0]
+        mask_start = int(h * 0.99)
+        blended[mask_start:] = (180, 180, 180)  # Gray footer region
+
     Image.fromarray(blended).save(str(out_dir / f"{page_label}_diff.png"))
 
 
@@ -181,6 +208,17 @@ def test_pdf_matches_reference(
 
     project_root = Path(__file__).parent.parent
 
+    # Load config
+    config_path = project_root / "config.json"
+    with config_path.open() as f:
+        config = json.load(f)
+    regression_cfg = config.get("regression", {})
+    
+    if not regression_cfg.get("enabled", False):
+        pytest.skip("Regression-Bildvergleich deaktiviert (regression.enabled=false in config.json)")
+    
+    mask_footer = regression_cfg.get("mask_footer", True)
+
     # ------------------------------------------------------------------
     # 1. Generate PDFs via the CLI into a temporary directory
     # ------------------------------------------------------------------
@@ -252,18 +290,19 @@ def test_pdf_matches_reference(
             ref_img = ref_images[page_idx]
             gen_img = gen_images[page_idx]
 
-            similarity = _compute_similarity(ref_img, gen_img)
+            similarity, diff_pixels, total_pixels = _compute_similarity(ref_img, gen_img, mask_footer=mask_footer)
 
             if similarity < SIMILARITY_THRESHOLD:
                 page_tag = f"seite_{page_idx + 1:02d}"
                 img_dir = diff_dir / clean_stem / page_tag
                 _save_diff_images(
-                    ref_img, gen_img, img_dir, page_tag
+                    ref_img, gen_img, img_dir, page_tag, mask_footer=mask_footer
                 )
                 rel = img_dir.relative_to(project_root)
                 failures.append(
-                    f"  {page_label}: Ähnlichkeit {similarity:.1%} "
-                    f"(Schwelle {SIMILARITY_THRESHOLD:.0%})  →  Bilder: {rel}/"
+                    f"  {page_label}: Ähnlichkeit {similarity:.2%} "
+                    f"({diff_pixels:,} / {total_pixels:,} Pixel unterschiedlich, "
+                    f"Schwelle {SIMILARITY_THRESHOLD:.0%})  →  Bilder: {rel}/"
                 )
 
     if failures: