Flexible inhaltsbasierte Spaltenerkennung (2-Phasen)

Ersetzt hardcodierte Positionsregeln durch ein zweistufiges System:
Phase A erkennt Spaltengeometrie (Clustering), Phase B klassifiziert
Typen per Inhalt (Sprache/Rolle) mit 3-stufiger Fallback-Kette.

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

klausur-service/backend/ocr_pipeline_api.py

@@ -648,8 +648,16 @@ async def detect_columns(session_id: str):
     duration = time.time() - t0
 
     columns = [asdict(r) for r in regions]
+
+    # Determine classification methods used
+    methods = list(set(
+        c.get("classification_method", "") for c in columns
+        if c.get("classification_method")
+    ))
+
     column_result = {
         "columns": columns,
+        "classification_methods": methods,
         "duration_seconds": round(duration, 2),
     }
 
@@ -742,6 +750,7 @@ async def _get_columns_overlay(session_id: str) -> Response:
         "column_en": (255, 180, 0),      # Blue
         "column_de": (0, 200, 0),         # Green
         "column_example": (0, 140, 255),  # Orange
+        "column_text": (200, 200, 0),     # Cyan/Turquoise
         "page_ref": (200, 0, 200),        # Purple
         "column_marker": (0, 0, 220),     # Red
         "header": (128, 128, 128),        # Gray
@@ -760,8 +769,11 @@ async def _get_columns_overlay(session_id: str) -> Response:
         # Solid border
         cv2.rectangle(img, (x, y), (x + w, y + h), color, 3)
 
-        # Label
+        # Label with confidence
         label = col.get("type", "unknown").replace("column_", "").upper()
+        conf = col.get("classification_confidence")
+        if conf is not None and conf < 1.0:
+            label = f"{label} {int(conf * 100)}%"
         cv2.putText(img, label, (x + 10, y + 30),
                     cv2.FONT_HERSHEY_SIMPLEX, 0.8, color, 2)