Mind the Gap: Benchmarking Spatial Reasoning in Vision-Language Models