refactor: remove llama-cpp entirely, use deterministic mock-only mode

fix: eagerly load LLM at startup so /api/health surfaces the real error