feat: Sprint 2 — TrOCR ONNX, PP-DocLayout, Model Management

D2: TrOCR ONNX export script (printed + handwritten, int8 quantization)
D3: PP-DocLayout ONNX export script (download or Docker-based conversion)
B3: Model Management admin page (PyTorch vs ONNX status, benchmarks, config)
A4: TrOCR ONNX service with runtime routing (auto/pytorch/onnx via TROCR_BACKEND)
A5: PP-DocLayout ONNX detection with OpenCV fallback (via GRAPHIC_DETECT_BACKEND)
B4: Structure Detection UI toggle (OpenCV vs PP-DocLayout) with class color coding
C3: TrOCR-ONNX.md documentation
C4: OCR-Pipeline.md ONNX section added
C5: mkdocs.yml nav updated, optimum added to requirements.txt

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

admin-lehrer/lib/navigation.ts

@@ -200,6 +200,15 @@ export const navigation: NavCategory[] = [
         audience: ['Entwickler', 'QA'],
         subgroup: 'KI-Werkzeuge',
       },
+      {
+        id: 'model-management',
+        name: 'Model Management',
+        href: '/ai/model-management',
+        description: 'ONNX & PyTorch Modell-Verwaltung',
+        purpose: 'Verfuegbare ML-Modelle verwalten (PyTorch vs ONNX), Backend umschalten, Benchmark-Vergleiche ausfuehren und RAM/Performance-Metriken einsehen.',
+        audience: ['Entwickler', 'DevOps'],
+        subgroup: 'KI-Werkzeuge',
+      },
       {
         id: 'agents',
         name: 'Agent Management',