feat(ocr): Word-based image deskew for Ground Truth pipeline

Begradigt schiefe Scans vor der OCR-Extraktion anhand der linksbuendigen
Wortanfaenge der Vokabelspalte. Tesseract liefert achsenparallele Boxen,
die bei ~2-3 Grad Schraege in Nachbarzeilen bluten — der Deskew behebt das.

- Neue Funktion deskew_image_by_word_alignment() in cv_vocab_pipeline.py
- Deskew-Integration im extract-with-boxes Endpoint (vor OCR)
- Neuer GET Endpoint /deskewed-image/{page} fuer begradigtes Seitenbild
- Frontend: GroundTruthPanel wechselt nach Extraktion auf deskewed Image
- ~1s Overhead durch schnellen Tesseract-Pass auf halbiertem Bild

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

klausur-service/backend/vocab_worksheet_api.py

@@ -2134,7 +2134,22 @@ async def extract_with_boxes(session_id: str, page_number: int):
     # Convert page to hires image
     image_data = await convert_pdf_page_to_image(pdf_data, page_number, thumbnail=False)
 
-    # Extract entries with boxes
+    # Deskew image before OCR
+    deskew_angle = 0.0
+    try:
+        from cv_vocab_pipeline import deskew_image_by_word_alignment, CV2_AVAILABLE
+        if CV2_AVAILABLE:
+            image_data, deskew_angle = deskew_image_by_word_alignment(image_data)
+            logger.info(f"Deskew: {deskew_angle:.2f}° for page {page_number}")
+    except Exception as e:
+        logger.warning(f"Deskew failed for page {page_number}: {e}")
+
+    # Cache deskewed image in session for later serving
+    if "deskewed_images" not in session:
+        session["deskewed_images"] = {}
+    session["deskewed_images"][str(page_number)] = image_data
+
+    # Extract entries with boxes (now on deskewed image)
     result = await extract_entries_with_boxes(image_data)
 
     # Cache in session
@@ -2148,9 +2163,35 @@ async def extract_with_boxes(session_id: str, page_number: int):
         "entry_count": len(result["entries"]),
         "image_width": result["image_width"],
         "image_height": result["image_height"],
+        "deskew_angle": round(deskew_angle, 2),
+        "deskewed": abs(deskew_angle) > 0.05,
     }
 
 
+@router.get("/sessions/{session_id}/deskewed-image/{page_number}")
+async def get_deskewed_image(session_id: str, page_number: int):
+    """Return the deskewed page image as PNG.
+
+    Falls back to the original hires image if no deskewed version is cached.
+    """
+    if session_id not in _sessions:
+        raise HTTPException(status_code=404, detail="Session not found")
+
+    session = _sessions[session_id]
+    deskewed = session.get("deskewed_images", {}).get(str(page_number))
+
+    if deskewed:
+        return StreamingResponse(io.BytesIO(deskewed), media_type="image/png")
+
+    # Fallback: render original hires image
+    pdf_data = session.get("pdf_data")
+    if not pdf_data:
+        raise HTTPException(status_code=400, detail="No PDF uploaded for this session")
+
+    image_data = await convert_pdf_page_to_image(pdf_data, page_number, thumbnail=False)
+    return StreamingResponse(io.BytesIO(image_data), media_type="image/png")
+
+
 @router.post("/sessions/{session_id}/ground-truth/{page_number}")
 async def save_ground_truth(session_id: str, page_number: int, data: dict = Body(...)):
     """Save ground truth labels for a page.