Run checks WebGPU, GPU device, WebLLM import, and API shape. Load fetches the model only when you ask.