feat: box-aware column detection — exclude box content from global columns

- Enrich column geometries with original full-page words (box-filtered)
  so _detect_sub_columns() finds narrow sub-columns across box boundaries
- Add inline marker guard: bullet points (1., 2., •) are not split into
  sub-columns (minimum gap check: 1.2× word height or 20px)
- Add box_rects parameter to build_grid_from_words() — words inside boxes
  are excluded from X-gap column clustering
- Pass box rects from zones to words_first grid builder
- Add 9 tests for box-aware column detection

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

klausur-service/backend/cv_layout.py

@@ -7,6 +7,7 @@ DATENSCHUTZ: Alle Verarbeitung erfolgt lokal.
 
 import logging
 import re
+import statistics
 from typing import Any, Dict, List, Optional, Tuple
 
 import numpy as np
@@ -737,6 +738,24 @@ def _detect_sub_columns(
             result.append(geo)
             continue
 
+        # --- Guard against inline markers (bullet points, numbering) ---
+        # Bullet points like "1.", "2.", "•", "-" sit close to the main
+        # column text and are part of the cell, not a separate column.
+        # Only split if the horizontal gap between the rightmost sub-word
+        # and the main column start is large enough.
+        max_sub_right = max(w['left'] + w.get('width', 0) for w in sub_words)
+        gap_to_main = col_start_bin[2] - max_sub_right  # px gap
+        median_heights = [w.get('height', 20) for w in confident]
+        med_h = statistics.median(median_heights) if median_heights else 20
+        min_gap = max(med_h * 1.2, 20)  # at least 1.2× word height or 20px
+        if gap_to_main < min_gap:
+            logger.debug(
+                "SubColumnSplit: column idx=%d skipped — gap=%dpx < min=%dpx "
+                "(likely inline markers, not a sub-column)",
+                geo.index, gap_to_main, min_gap)
+            result.append(geo)
+            continue
+
         # --- Build two sub-column geometries ---
         # Word 'left' values are relative to left_x; geo.x is absolute.
         # Convert the split position from relative to absolute coordinates.
@@ -3221,6 +3240,46 @@ def detect_column_geometry_zoned(
             g.y = abs_y
             g.height = abs_y_end - abs_y
 
+    # --- Enrich column geometries with box-filtered original words ---
+    # The combined-image Tesseract may miss words in small content strips
+    # (e.g. a single row above a box).  Use the original full-page word_dicts
+    # filtered to exclude box interiors, so that _detect_sub_columns()
+    # downstream has ALL content-zone words for left-edge clustering.
+    # This ensures narrow sub-columns (page_ref, marker) are detectable
+    # even when only a few entries exist above/below a box.
+    if word_dicts:
+        content_words = []
+        for w in word_dicts:
+            # word positions are relative to left_x / top_y
+            w_abs_cx = w['left'] + left_x + w['width'] / 2
+            w_abs_cy = w['top'] + top_y + w['height'] / 2
+            inside_box = any(
+                box.x <= w_abs_cx <= box.x + box.width
+                and box.y <= w_abs_cy <= box.y + box.height
+                for box in boxes
+            )
+            if not inside_box:
+                content_words.append(w)
+
+        target_geoms = combined_geoms if combined_result is not None else geometries
+        for g in target_geoms:
+            # Word 'left' is relative to left_x; geometry 'x' is absolute
+            g_left_rel = g.x - left_x
+            g_right_rel = g_left_rel + g.width
+            g.words = [
+                w for w in content_words
+                if g_left_rel <= w['left'] + w['width'] / 2 < g_right_rel
+            ]
+            g.word_count = len(g.words)
+
+        excluded_count = len(word_dicts) - len(content_words)
+        if excluded_count:
+            logger.info(
+                "ZonedColumns: enriched geometries with %d content words "
+                "(excluded %d box-interior words)",
+                len(content_words), excluded_count,
+            )
+
     # Build zones_data for the response
     zones_data: List[Dict] = []
     for zone in zones: