feat: add AI Lab with capability detection, local LLM, and MCP tools

[simonhagger]

Summary

• */ai/capabilities* — stdlib-only sidecar probe: NVIDIA GPUs via nvidia-smi (name/VRAM/driver), CPU/RAM, importable inference backends (llama_cpp/torch/onnxruntime/transformers via find_spec), local .gguf model scan, plain-language verdict + setup notes
• */ai/generate* — local inference through llama-cpp-python when backend and a GGUF model exist (instance cache across requests, GPU layers enabled when CUDA driver detected); structured unavailable+guidance otherwise. Server upgraded to ThreadingHTTPServer so long generations cannot starve /health monitoring
• *\POST /mcp* — JSON-RPC 2.0 Model-Context-Protocol surface in pure stdlib: initialize / notifications/initialized / tools/list / tools/call with echo, system_info, time_now tools; -32601/-32602 error semantics
• Validated IPC channels \�i:capabilities/\�i:generate/\�i:mcp-invoke, preload \�i\ API namespace, \DesktopAiApi\
• Renderer AI Lab route: capability matrix, local generation panel, interactive MCP tool explorer

Validation

• Python sidecar suite: 20/20 passed at 98% coverage (incl. stubbed llama_cpp happy path, cache reuse, load-failure, validation errors, malformed nvidia-smi rows)
• contracts/desktop-api/preload/desktop-main builds green; desktop-main TS tests green; renderer dev build emits 28.57 kB ai-lab chunk; renderer tests + repo lint green

Security Checklist

• Security review completed
• Threat model updated or N/A explained
• Confirmed no secrets/sensitive data present in committed files