make base_vlm portable (HF repo id) for out-of-the-box reload

full fine-tuned openvla (Qwen3-VL-4B) on flappy fixed_l2; eval mean_return=18.52 @latency2

initial commit