Warm WebGPU before first prompt

Use manifest context limits and handle memory exhaustion safely

Set 8K context defaults and add artifact ZIP export

Use 4096-token release defaults