feat(klausur): Handschrift entfernen + Klausur-HTR implementiert

Feature 1: Handschrift entfernen via OCR-Pipeline Session
- services/handwriting_detection.py: _detect_pencil() + target_ink Parameter
  ("all" | "colored" | "pencil") für gezielte Tinten-Erkennung
- ocr_pipeline_session_store.py: clean_png + handwriting_removal_meta Spalten
  (idempotentes ALTER TABLE in init_ocr_pipeline_tables)
- ocr_pipeline_api.py: POST /sessions/{id}/remove-handwriting Endpoint
  + "clean" zu valid_types für Image-Serving hinzugefügt

Feature 2: Klausur-HTR (Hochwertige Handschriftenerkennung)
- handwriting_htr_api.py: Neuer Router /api/v1/htr/recognize + /recognize-session
  Primary: qwen2.5vl:32b via Ollama, Fallback: trocr-large-handwritten
- services/trocr_service.py: size Parameter (base | large) für get_trocr_model()
  + run_trocr_ocr() - unterstützt jetzt trocr-large-handwritten
- main.py: HTR Router registriert

Config:
- docker-compose.yml: OLLAMA_HTR_MODEL, HTR_FALLBACK_MODEL
- .env.example: HTR Env-Vars dokumentiert

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

klausur-service/backend/ocr_pipeline_api.py

@@ -168,6 +168,13 @@ class RowGroundTruthRequest(BaseModel):
     notes: Optional[str] = None
 
 
+class RemoveHandwritingRequest(BaseModel):
+    method: str = "auto"       # "auto" | "telea" | "ns"
+    target_ink: str = "all"    # "all" | "colored" | "pencil"
+    dilation: int = 2          # mask dilation iterations (0-5)
+    use_source: str = "auto"   # "original" | "deskewed" | "auto"
+
+
 # ---------------------------------------------------------------------------
 # Session Management Endpoints
 # ---------------------------------------------------------------------------
@@ -309,7 +316,7 @@ async def delete_session(session_id: str):
 @router.get("/sessions/{session_id}/image/{image_type}")
 async def get_image(session_id: str, image_type: str):
     """Serve session images: original, deskewed, dewarped, binarized, columns-overlay, or rows-overlay."""
-    valid_types = {"original", "deskewed", "dewarped", "binarized", "columns-overlay", "rows-overlay", "words-overlay"}
+    valid_types = {"original", "deskewed", "dewarped", "binarized", "columns-overlay", "rows-overlay", "words-overlay", "clean"}
     if image_type not in valid_types:
         raise HTTPException(status_code=400, detail=f"Unknown image type: {image_type}")
 
@@ -1906,3 +1913,90 @@ async def _get_words_overlay(session_id: str) -> Response:
         raise HTTPException(status_code=500, detail="Failed to encode overlay image")
 
     return Response(content=result_png.tobytes(), media_type="image/png")
+
+
+# ---------------------------------------------------------------------------
+# Handwriting Removal Endpoint
+# ---------------------------------------------------------------------------
+
+@router.post("/sessions/{session_id}/remove-handwriting")
+async def remove_handwriting_endpoint(session_id: str, req: RemoveHandwritingRequest):
+    """
+    Remove handwriting from a session image using inpainting.
+
+    Steps:
+    1. Load source image (auto → deskewed if available, else original)
+    2. Detect handwriting mask (filtered by target_ink)
+    3. Dilate mask to cover stroke edges
+    4. Inpaint the image
+    5. Store result as clean_png in the session
+
+    Returns metadata including the URL to fetch the clean image.
+    """
+    import time as _time
+    t0 = _time.monotonic()
+
+    from services.handwriting_detection import detect_handwriting
+    from services.inpainting_service import inpaint_image, dilate_mask as _dilate_mask, InpaintingMethod, image_to_png
+
+    session = await get_session_db(session_id)
+    if not session:
+        raise HTTPException(status_code=404, detail=f"Session {session_id} not found")
+
+    # 1. Determine source image
+    source = req.use_source
+    if source == "auto":
+        deskewed = await get_session_image(session_id, "deskewed")
+        source = "deskewed" if deskewed else "original"
+
+    image_bytes = await get_session_image(session_id, source)
+    if not image_bytes:
+        raise HTTPException(status_code=404, detail=f"Source image '{source}' not available")
+
+    # 2. Detect handwriting mask
+    detection = detect_handwriting(image_bytes, target_ink=req.target_ink)
+
+    # 3. Convert mask to PNG bytes and dilate
+    import io
+    from PIL import Image as _PILImage
+    mask_img = _PILImage.fromarray(detection.mask)
+    mask_buf = io.BytesIO()
+    mask_img.save(mask_buf, format="PNG")
+    mask_bytes = mask_buf.getvalue()
+
+    if req.dilation > 0:
+        mask_bytes = _dilate_mask(mask_bytes, iterations=req.dilation)
+
+    # 4. Inpaint
+    method_map = {
+        "telea": InpaintingMethod.OPENCV_TELEA,
+        "ns": InpaintingMethod.OPENCV_NS,
+        "auto": InpaintingMethod.AUTO,
+    }
+    inpaint_method = method_map.get(req.method, InpaintingMethod.AUTO)
+
+    result = inpaint_image(image_bytes, mask_bytes, method=inpaint_method)
+    if not result.success:
+        raise HTTPException(status_code=500, detail="Inpainting failed")
+
+    elapsed_ms = int((_time.monotonic() - t0) * 1000)
+
+    meta = {
+        "method_used": result.method_used.value if hasattr(result.method_used, "value") else str(result.method_used),
+        "handwriting_ratio": round(detection.handwriting_ratio, 4),
+        "detection_confidence": round(detection.confidence, 4),
+        "target_ink": req.target_ink,
+        "dilation": req.dilation,
+        "source_image": source,
+        "processing_time_ms": elapsed_ms,
+    }
+
+    # 5. Persist clean image (convert BGR ndarray → PNG bytes)
+    clean_png_bytes = image_to_png(result.image)
+    await update_session_db(session_id, clean_png=clean_png_bytes, handwriting_removal_meta=meta)
+
+    return {
+        **meta,
+        "image_url": f"/api/v1/ocr-pipeline/sessions/{session_id}/image/clean",
+        "session_id": session_id,
+    }