feat: breite Spalten per Word-Gap splitten + gedrehte Scans im Frontend anzeigen

_split_broad_columns() erkennt EN/DE-Gemisch in breiten Spalten via
Word-Coverage-Analyse und trennt sie am groessten Luecken-Gap.
Thumbnails und Page-Images werden serverseitig per fitz rotiert,
Frontend laedt Thumbnails nach OCR-Processing neu.

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

klausur-service/backend/vocab_worksheet_api.py

@@ -70,6 +70,7 @@ try:
         detect_column_geometry, analyze_layout_by_words, analyze_layout, create_layout_image,
         detect_row_geometry, build_cell_grid_v2,
         _cells_to_vocab_entries, _detect_sub_columns, _detect_header_footer_gaps,
+        _split_broad_columns,
         expand_narrow_columns, positional_column_regions, llm_review_entries,
         detect_and_fix_orientation,
         _fix_phonetic_brackets,
@@ -1182,6 +1183,9 @@ async def upload_pdf_get_info(
 async def get_pdf_thumbnail(session_id: str, page_number: int, hires: bool = Query(False)):
     """Get a thumbnail image of a specific PDF page.
 
+    Uses fitz for rendering so that page_rotations (from OCR orientation
+    detection) are applied consistently.
+
     Args:
         hires: If True, return full-resolution image (zoom=2.0) instead of thumbnail (zoom=0.5).
     """
@@ -1194,10 +1198,25 @@ async def get_pdf_thumbnail(session_id: str, page_number: int, hires: bool = Que
     if not pdf_data:
         raise HTTPException(status_code=400, detail="No PDF uploaded for this session")
 
-    thumbnail = await convert_pdf_page_to_image(pdf_data, page_number, thumbnail=not hires)
+    try:
+        import fitz
+        zoom = 2.0 if hires else 0.5
+        pdf_document = fitz.open(stream=pdf_data, filetype="pdf")
+        page = pdf_document[page_number]
+        # Apply orientation correction detected during OCR processing
+        rot = session.get("page_rotations", {}).get(page_number, 0)
+        if rot:
+            page.set_rotation(rot)
+        mat = fitz.Matrix(zoom, zoom)
+        pix = page.get_pixmap(matrix=mat)
+        png_data = pix.tobytes("png")
+        pdf_document.close()
+    except Exception as e:
+        logger.error(f"PDF thumbnail failed: {e}")
+        raise HTTPException(status_code=500, detail=f"PDF rendering failed: {str(e)}")
 
     return StreamingResponse(
-        io.BytesIO(thumbnail),
+        io.BytesIO(png_data),
         media_type="image/png",
     )
 
@@ -1226,11 +1245,15 @@ async def get_pdf_page_image(session_id: str, page_number: int, zoom: float = Qu
         import fitz
         pdf_document = fitz.open(stream=pdf_data, filetype="pdf")
         page = pdf_document[page_number]
+        # Apply orientation correction detected during OCR processing
+        rot = session.get("page_rotations", {}).get(page_number, 0)
+        if rot:
+            page.set_rotation(rot)
         mat = fitz.Matrix(zoom, zoom)
         pix = page.get_pixmap(matrix=mat)
         png_data = pix.tobytes("png")
         pdf_document.close()
-        logger.info(f"PDF page {page_number} rendered at zoom={zoom}: {len(png_data)} bytes")
+        logger.info(f"PDF page {page_number} rendered at zoom={zoom} rot={rot}: {len(png_data)} bytes")
     except Exception as e:
         logger.error(f"PDF page image failed: {e}")
         raise HTTPException(status_code=500, detail=f"PDF rendering failed: {str(e)}")
@@ -1272,10 +1295,11 @@ async def process_single_page(
         raise HTTPException(status_code=400, detail=f"Invalid page number. PDF has {page_count} pages (0-indexed).")
 
     # --- OCR Pipeline path (use same render_pdf_high_res as admin OCR pipeline) ---
+    rotation_deg = 0
     if OCR_PIPELINE_AVAILABLE:
         try:
             img_bgr = render_pdf_high_res(pdf_data, page_number, zoom=3.0)
-            page_vocabulary = await _run_ocr_pipeline_for_page(
+            page_vocabulary, rotation_deg = await _run_ocr_pipeline_for_page(
                 img_bgr, page_number, session_id,
             )
         except Exception as e:
@@ -1317,6 +1341,9 @@ async def process_single_page(
 
     logger.info(f"Page {page_number + 1}: {len(page_vocabulary)} Vokabeln extrahiert")
 
+    # Store rotation for this page (used by image/thumbnail endpoints)
+    session.setdefault("page_rotations", {})[page_number] = rotation_deg
+
     # Add to session's vocabulary (append, don't replace)
     existing_vocab = session.get("vocabulary", [])
     # Remove any existing entries from this page (in case of re-processing)
@@ -1334,6 +1361,7 @@ async def process_single_page(
         "vocabulary_count": len(page_vocabulary),
         "total_vocabulary_count": len(existing_vocab),
         "extraction_confidence": 0.9,
+        "rotation": rotation_deg,
     }
 
 
@@ -1341,7 +1369,7 @@ async def _run_ocr_pipeline_for_page(
     img_bgr: np.ndarray,
     page_number: int,
     vocab_session_id: str,
-) -> list:
+) -> tuple:
     """Run the full OCR pipeline on a single page image and return vocab entries.
 
     Uses the same pipeline as the admin OCR pipeline (ocr_pipeline_api.py).
@@ -1352,7 +1380,8 @@ async def _run_ocr_pipeline_for_page(
         vocab_session_id: Vocab session ID for logging.
 
     Steps: deskew → dewarp → columns → rows → words → (LLM review)
-    Returns list of dicts with keys: id, english, german, example_sentence, source_page
+    Returns (entries, rotation_deg) where entries is a list of dicts and
+    rotation_deg is the orientation correction applied (0, 90, 180, 270).
     """
     import time as _time
 
@@ -1418,6 +1447,7 @@ async def _run_ocr_pipeline_for_page(
         header_y, footer_y = _detect_header_footer_gaps(inv, w, h) if inv is not None else (None, None)
         geometries = _detect_sub_columns(geometries, content_w, left_x=left_x,
                                           top_y=top_y, header_y=header_y, footer_y=footer_y)
+        geometries = _split_broad_columns(geometries, content_w, left_x=left_x)
         geometries = expand_narrow_columns(geometries, content_w, left_x, word_dicts)
         content_h = bottom_y - top_y
         regions = positional_column_regions(geometries, content_w, content_h, left_x)
@@ -1534,7 +1564,7 @@ async def _run_ocr_pipeline_for_page(
     logger.info(f"OCR Pipeline page {page_number + 1}: "
                 f"{len(page_vocabulary)} vocab entries in {total_duration:.1f}s")
 
-    return page_vocabulary
+    return page_vocabulary, rotation
 
 
 @router.post("/sessions/{session_id}/process-pages")