refactor: remove unused pages and backends (model-management, OCR legacy, GPU/vast.ai, video-chat, matrix)

Deleted pages:
- /ai/model-management (mock data only, no real backend)
- /ai/ocr-compare (old /vocab/ backend, replaced by ocr-kombi)
- /ai/ocr-pipeline (minimal session browser, redundant)
- /ai/ocr-overlay (legacy monolith, redundant)
- /ai/gpu (vast.ai GPU management, no longer used)
- /infrastructure/gpu (same)
- /communication/video-chat (moved to core)
- /communication/matrix (moved to core)

Deleted backends:
- backend-lehrer/infra/vast_client.py + vast_power.py
- backend-lehrer/meetings_api.py + jitsi_api.py
- website/app/api/admin/gpu/
- edu-search-service/scripts/vast_ai_extractor.py

Total: ~7,800 LOC removed. All code preserved in git history.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

klausur-service/backend/services/lighton_ocr_service.py

@@ -0,0 +1,119 @@
+"""
+LightOnOCR-2-1B Service
+
+End-to-end VLM OCR fuer gedruckten und gemischten Text.
+1B Parameter, Apple MPS-faehig (M-Serie).
+
+Modell:  lightonai/LightOnOCR-2-1B
+Lizenz:  Apache 2.0
+Quelle:  https://huggingface.co/lightonai/LightOnOCR-2-1B
+
+Unterstuetzte Dokumenttypen:
+- Buchseiten, Vokabelseiten
+- Arbeitsblaetter, Klausuren
+- Gemischt gedruckt/handschriftlich
+
+DATENSCHUTZ: Alle Verarbeitung erfolgt lokal.
+"""
+import io
+import logging
+import os
+from typing import Optional, Tuple
+
+logger = logging.getLogger(__name__)
+
+LIGHTON_MODEL_ID = os.getenv("LIGHTON_OCR_MODEL", "lightonai/LightOnOCR-2-1B")
+
+_lighton_model = None
+_lighton_processor = None
+_lighton_available: Optional[bool] = None
+
+
+def _check_lighton_available() -> bool:
+    """Check if LightOnOCR dependencies (transformers, torch) are available."""
+    global _lighton_available
+    if _lighton_available is not None:
+        return _lighton_available
+    try:
+        from transformers import AutoModelForImageTextToText, AutoProcessor  # noqa: F401
+        import torch  # noqa: F401
+        _lighton_available = True
+    except ImportError as e:
+        logger.warning(f"LightOnOCR deps not available: {e}")
+        _lighton_available = False
+    return _lighton_available
+
+
+def get_lighton_model() -> Tuple:
+    """
+    Lazy-load LightOnOCR-2-1B processor and model.
+
+    Returns (processor, model) or (None, None) on failure.
+    Device priority: MPS (Apple Silicon) > CUDA > CPU.
+    """
+    global _lighton_model, _lighton_processor
+    if _lighton_model is not None:
+        return _lighton_processor, _lighton_model
+    if not _check_lighton_available():
+        return None, None
+
+    try:
+        import torch
+        from transformers import AutoModelForImageTextToText, AutoProcessor
+
+        if torch.backends.mps.is_available():
+            device = "mps"
+        elif torch.cuda.is_available():
+            device = "cuda"
+        else:
+            device = "cpu"
+        dtype = torch.bfloat16
+
+        logger.info(f"Loading LightOnOCR-2-1B on {device} ({dtype}) from {LIGHTON_MODEL_ID} ...")
+        _lighton_processor = AutoProcessor.from_pretrained(LIGHTON_MODEL_ID)
+        _lighton_model = AutoModelForImageTextToText.from_pretrained(
+            LIGHTON_MODEL_ID, torch_dtype=dtype
+        ).to(device)
+        _lighton_model.eval()
+        logger.info("LightOnOCR-2-1B loaded successfully")
+
+    except Exception as e:
+        logger.error(f"Failed to load LightOnOCR-2-1B: {e}")
+        _lighton_model = None
+        _lighton_processor = None
+
+    return _lighton_processor, _lighton_model
+
+
+def run_lighton_ocr_sync(image_bytes: bytes) -> Optional[str]:
+    """
+    Run LightOnOCR on image bytes (synchronous).
+
+    Returns extracted text or None on error.
+    Generic — works for any document/page region.
+    """
+    processor, model = get_lighton_model()
+    if processor is None or model is None:
+        return None
+
+    try:
+        import torch
+        from PIL import Image as _PILImage
+
+        pil_img = _PILImage.open(io.BytesIO(image_bytes)).convert("RGB")
+        conversation = [{"role": "user", "content": [{"type": "image"}]}]
+
+        inputs = processor.apply_chat_template(
+            conversation, images=[pil_img],
+            add_generation_prompt=True, return_tensors="pt"
+        ).to(model.device)
+
+        with torch.no_grad():
+            output_ids = model.generate(**inputs, max_new_tokens=1024)
+
+        text = processor.decode(output_ids[0], skip_special_tokens=True)
+        return text.strip() if text else None
+
+    except Exception as e:
+        logger.error(f"LightOnOCR inference failed: {e}")
+        return None