fix: Sub-Session Zeilenerkennung — Tesseract+inv im Spalten-Schritt cachen

Bisher wurden _word_dicts, _inv und _content_bounds fuer Sub-Sessions
nicht gecacht, sodass detect_rows auf detect_column_geometry() zurueckfiel.
Das konnte bei kleinen Box-Bildern mit <5 Woertern fehlschlagen.

Jetzt laeuft Tesseract + Binarisierung direkt im Pseudo-Spalten-Block,
und die Intermediates werden gecacht. Zusaetzlich ausfuehrliche Kommentare
zur Zeilenerkennung (detect_row_geometry, _regularize_row_grid).

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

klausur-service/backend/ocr_pipeline_api.py

@@ -1209,10 +1209,55 @@ async def detect_columns(session_id: str):
     if img_bgr is None:
         raise HTTPException(status_code=400, detail="Crop or dewarp must be completed before column detection")
 
-    # Sub-sessions: skip column detection, create single pseudo-column
+    # -----------------------------------------------------------------------
+    # Sub-sessions (box crops): skip column detection entirely.
+    # Instead, create a single pseudo-column spanning the full image width.
+    # Also run Tesseract + binarization here so that the row detection step
+    # can reuse the cached intermediates (_word_dicts, _inv, _content_bounds)
+    # instead of falling back to detect_column_geometry() which may fail
+    # on small box images with < 5 words.
+    # -----------------------------------------------------------------------
     session = await get_session_db(session_id)
     if session and session.get("parent_session_id"):
         h, w = img_bgr.shape[:2]
+
+        # Binarize + invert for row detection (horizontal projection profile)
+        ocr_img = create_ocr_image(img_bgr)
+        inv = cv2.bitwise_not(ocr_img)
+
+        # Run Tesseract to get word bounding boxes.
+        # Word positions are relative to the full image (no ROI crop needed
+        # because the sub-session image IS the cropped box already).
+        # detect_row_geometry expects word positions relative to content ROI,
+        # so with content_bounds = (0, w, 0, h) the coordinates are correct.
+        try:
+            from PIL import Image as PILImage
+            pil_img = PILImage.fromarray(cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB))
+            import pytesseract
+            data = pytesseract.image_to_data(pil_img, lang='eng+deu', output_type=pytesseract.Output.DICT)
+            word_dicts = []
+            for i in range(len(data['text'])):
+                conf = int(data['conf'][i]) if str(data['conf'][i]).lstrip('-').isdigit() else -1
+                text = str(data['text'][i]).strip()
+                if conf < 30 or not text:
+                    continue
+                word_dicts.append({
+                    'text': text, 'conf': conf,
+                    'left': int(data['left'][i]),
+                    'top': int(data['top'][i]),
+                    'width': int(data['width'][i]),
+                    'height': int(data['height'][i]),
+                })
+            logger.info(f"OCR Pipeline: sub-session {session_id}: Tesseract found {len(word_dicts)} words")
+        except Exception as e:
+            logger.warning(f"OCR Pipeline: sub-session {session_id}: Tesseract failed: {e}")
+            word_dicts = []
+
+        # Cache intermediates for row detection (detect_rows reuses these)
+        cached["_word_dicts"] = word_dicts
+        cached["_inv"] = inv
+        cached["_content_bounds"] = (0, w, 0, h)
+
         column_result = {
             "columns": [{
                 "type": "column_text",