fix: Edge-Gaps in _split_broad_columns ignorieren + return-Tuple bei leerem Ergebnis

Gaps die den Spaltenrand beruehren (Margins) werden jetzt ausgeschlossen,
nur interne Gaps werden als Split-Kandidaten betrachtet. Behebt das
Problem dass trailing whitespace faelschlich als groesster Gap gewaehlt
wurde. Early-return in _run_ocr_pipeline_for_page gibt jetzt korrekt
([], rotation) statt [] zurueck.

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

klausur-service/backend/cv_vocab_pipeline.py

@@ -2125,27 +2125,9 @@ def _split_broad_columns(
         if cmax > 0:
             coverage /= cmax
 
-        # Find gaps where coverage < 0.5
+        # Find INTERNAL gaps where coverage < 0.5
+        # Exclude edge gaps (touching pixel 0 or geo.width) — those are margins.
         low_mask = coverage < 0.5
-        gap_start = None
-        best_gap = None  # (start, end, width)
-        for px in range(len(low_mask)):
-            if low_mask[px]:
-                if gap_start is None:
-                    gap_start = px
-            else:
-                if gap_start is not None:
-                    gw = px - gap_start
-                    if best_gap is None or gw > best_gap[2]:
-                        best_gap = (gap_start, px, gw)
-                    gap_start = None
-        # Handle trailing gap
-        if gap_start is not None:
-            gw = len(low_mask) - gap_start
-            if best_gap is None or gw > best_gap[2]:
-                best_gap = (gap_start, len(low_mask), gw)
-
-        # Log all gaps found for debugging
         all_gaps = []
         _gs = None
         for px in range(len(low_mask)):
@@ -2158,8 +2140,17 @@ def _split_broad_columns(
                     _gs = None
         if _gs is not None:
             all_gaps.append((_gs, len(low_mask), len(low_mask) - _gs))
-        logger.info(f"SplitBroadCols: col {geo.index} coverage gaps (>=5px): "
-                    f"{[g for g in all_gaps if g[2] >= 5]}, best={best_gap}")
+
+        # Filter: only internal gaps (not touching column edges)
+        _edge_margin = 10  # pixels from edge to ignore
+        internal_gaps = [g for g in all_gaps
+                         if g[0] > _edge_margin and g[1] < geo.width - _edge_margin]
+        best_gap = max(internal_gaps, key=lambda g: g[2]) if internal_gaps else None
+
+        logger.info(f"SplitBroadCols: col {geo.index} all_gaps(>=5px): "
+                    f"{[g for g in all_gaps if g[2] >= 5]}, "
+                    f"internal(>=5px): {[g for g in internal_gaps if g[2] >= 5]}, "
+                    f"best={best_gap}")
 
         if best_gap is None or best_gap[2] < _min_gap_px:
             result.append(geo)

klausur-service/backend/vocab_worksheet_api.py

@@ -1510,7 +1510,7 @@ async def _run_ocr_pipeline_for_page(
     if not is_vocab:
         logger.warning(f"  Page {page_number + 1}: layout is not vocab table "
                        f"(types: {col_types}), returning empty")
-        return []
+        return [], rotation
 
     # 8. Map cells → vocab entries
     entries = _cells_to_vocab_entries(cells, columns_meta)