feat(ocr-pipeline): word-based 5-column detection for vocabulary pages

Replace projection-profile layout analysis with Tesseract word bounding
box clustering to detect 5-column vocabulary layouts (page_ref, EN, DE,
markers, examples). Falls back to projection profiles when < 3 clusters.

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

klausur-service/backend/ocr_pipeline_api.py

@@ -29,6 +29,7 @@ from pydantic import BaseModel
 
 from cv_vocab_pipeline import (
     analyze_layout,
+    analyze_layout_by_words,
     create_ocr_image,
     deskew_image,
     deskew_image_by_word_alignment,
@@ -639,15 +640,11 @@ async def detect_columns(session_id: str):
 
     t0 = time.time()
 
-    # Prepare images for analyze_layout
-    gray = cv2.cvtColor(dewarped_bgr, cv2.COLOR_BGR2GRAY)
-    # CLAHE-enhanced for layout analysis
-    clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
-    layout_img = clahe.apply(gray)
-    # Binarized for text density
+    # Binarized image for layout analysis
     ocr_img = create_ocr_image(dewarped_bgr)
 
-    regions = analyze_layout(layout_img, ocr_img)
+    # Word-based detection (with automatic fallback to projection profiles)
+    regions = analyze_layout_by_words(ocr_img, dewarped_bgr)
     duration = time.time() - t0
 
     columns = [asdict(r) for r in regions]
@@ -740,11 +737,13 @@ async def _get_columns_overlay(session_id: str) -> Response:
     if img is None:
         raise HTTPException(status_code=500, detail="Failed to decode image")
 
-    # Color map for region types
+    # Color map for region types (BGR)
     colors = {
-        "column_en": (255, 180, 0),      # Blue (BGR)
+        "column_en": (255, 180, 0),      # Blue
         "column_de": (0, 200, 0),         # Green
         "column_example": (0, 140, 255),  # Orange
+        "page_ref": (200, 0, 200),        # Purple
+        "column_marker": (0, 0, 220),     # Red
         "header": (128, 128, 128),        # Gray
         "footer": (128, 128, 128),        # Gray
     }