fix: full-width Zeilen vor Spaltenerkennung maskieren

Farbige Sub-Header (z.B. "Unit 4: Bonnie Scotland") mit voller Breite
fuellten die Spaltenluecken im vertikalen Projektionsprofil auf und
fuehrten zu 11 statt 5 erkannten Spalten. Zeilen mit >40% Tintendichte
werden jetzt vor der Projektion maskiert.

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

klausur-service/backend/cv_vocab_pipeline.py

@@ -2131,10 +2131,46 @@ def detect_column_geometry(ocr_img: np.ndarray, dewarped_bgr: np.ndarray) -> Opt
 
     logger.info(f"ColumnGeometry: {len(left_edges)} words detected in content area")
 
-    # --- Step 3: Vertical projection profile ---
+    # --- Step 2b: Mask out full-width rows (sub-headers, colored bands) ---
+    # Rows where ink spans nearly the full content width distort the vertical
+    # projection by filling in column gaps.  Detect them via horizontal density
+    # and zero them out before computing v_proj.
     content_strip = inv[top_y:bottom_y, left_x:right_x]
-    v_proj = np.sum(content_strip, axis=0).astype(float)
-    v_proj_norm = v_proj / (content_h * 255) if content_h > 0 else v_proj
+    h_proj_row = np.sum(content_strip, axis=1).astype(float)
+    h_proj_row_norm = h_proj_row / (content_w * 255) if content_w > 0 else h_proj_row
+
+    FULLWIDTH_THRESHOLD = 0.40  # normal text ~10-25%; full-width bands 40%+
+    fullwidth_mask = h_proj_row_norm > FULLWIDTH_THRESHOLD
+
+    # Only mask contiguous bands (>=3 rows), not isolated noisy rows
+    masked_strip = content_strip.copy()
+    n_masked = 0
+    band_start = None
+    for y_idx in range(len(fullwidth_mask)):
+        if fullwidth_mask[y_idx]:
+            if band_start is None:
+                band_start = y_idx
+        else:
+            if band_start is not None:
+                band_height = y_idx - band_start
+                if band_height >= 3:
+                    masked_strip[band_start:y_idx, :] = 0
+                    n_masked += band_height
+                band_start = None
+    if band_start is not None:
+        band_height = len(fullwidth_mask) - band_start
+        if band_height >= 3:
+            masked_strip[band_start:len(fullwidth_mask), :] = 0
+            n_masked += band_height
+
+    if n_masked > 0:
+        logger.info(f"ColumnGeometry: masked {n_masked} full-width rows "
+                    f"({n_masked * 100 / content_h:.1f}% of content height)")
+
+    # --- Step 3: Vertical projection profile ---
+    effective_h = content_h - n_masked
+    v_proj = np.sum(masked_strip, axis=0).astype(float)
+    v_proj_norm = v_proj / (effective_h * 255) if effective_h > 0 else v_proj
 
     # Smooth the projection to avoid noise-induced micro-gaps
     kernel_size = max(5, content_w // 80)