VAD: Attributing Visual Evidence for Target Reconstruction in Multimodal On-Policy Distillation — ThinkLLM