fix: Word-Validation auf Segment-Woerter beschraenken

Woerter aus Sub-Header-Bereichen ueberlappten korrekte Spaltenluecken
und liessen die Word-Validation faelschlich Gaps verwerfen. Jetzt werden
nur Woerter aus dem gewaehlten Segment fuer die Validation verwendet.

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

klausur-service/backend/cv_vocab_pipeline.py

@@ -2240,11 +2240,25 @@ def detect_column_geometry(ocr_img: np.ndarray, dewarped_bgr: np.ndarray) -> Opt
                 f"{[(g[0]+left_x, g[1]+left_x, g[1]-g[0]) for g in raw_gaps]}")
 
     # --- Step 5: Validate gaps against word bounding boxes ---
+    # When using a segment for projection, only validate against words
+    # inside that segment — words from sub-headers or other sections
+    # would incorrectly overlap with real column gaps.
+    if segments and len(segments) > 1:
+        seg_top_abs = best_seg[0]  # relative to content strip
+        seg_bot_abs = best_seg[1]
+        segment_words = [wd for wd in word_dicts
+                         if wd['top'] >= seg_top_abs
+                         and wd['top'] + wd['height'] <= seg_bot_abs]
+        logger.info(f"ColumnGeometry: filtering words to segment: "
+                    f"{len(segment_words)}/{len(word_dicts)} words")
+    else:
+        segment_words = word_dicts
+
     validated_gaps = []
     for gap_start_rel, gap_end_rel in raw_gaps:
         # Check if any word overlaps with this gap region
         overlapping = False
-        for wd in word_dicts:
+        for wd in segment_words:
             word_left = wd['left']
             word_right = wd['left'] + wd['width']
             if word_left < gap_end_rel and word_right > gap_start_rel:
@@ -2258,7 +2272,7 @@ def detect_column_geometry(ocr_img: np.ndarray, dewarped_bgr: np.ndarray) -> Opt
             # Find the tightest word boundaries within the gap region
             min_word_left = content_w
             max_word_right = 0
-            for wd in word_dicts:
+            for wd in segment_words:
                 word_left = wd['left']
                 word_right = wd['left'] + wd['width']
                 if word_left < gap_end_rel and word_right > gap_start_rel:
@@ -2287,7 +2301,7 @@ def detect_column_geometry(ocr_img: np.ndarray, dewarped_bgr: np.ndarray) -> Opt
     if len(validated_gaps) < 2:
         logger.info("ColumnGeometry: < 2 pixel-gaps, trying word-coverage gaps")
         word_coverage = np.zeros(content_w, dtype=np.int32)
-        for wd in word_dicts:
+        for wd in segment_words:
             wl = max(0, wd['left'])
             wr = min(wd['left'] + wd['width'], content_w)
             if wr > wl: