Show, Don’t Tell: Evaluating Spatial Cognition in Generative Pixels Rather Than LLM Text | ResearchPod