Vision token pruning can accelerate vision language model inference by removing redundant visual tokens. However, we observe that ViT attention based pruning drops sharply on Qwen3-VL and Qwen3.5. Empirically, we identify the root cause: the last-block MLP of the vision encoder significantly alters the spectral structure of hidden states and induces spectral collapse. Thus, pre-MLP attention rankings no longer reflect downstream token importance. Through further analysis of vision token selection when ViT attention based pruning still succeeds, we find that selected tokens strongly anti-correlate with raw L2 norm in the vision encoder output space S_out , while favoring tokens with larger residual energy. Motivated by shared components mask token-specific importance, we propose GeoPrune, a new importance-based pruning method that operates in S_out , applies deflation to remove dominant components, and ranks tokens by residual L2 norm. At aggressive pruning rates, GeoPrune is more robust and surpasses ViT attention based pruning by +4.4/+6.3 recovery points on Qwen3-VL and Qwen3.5. We further integrate GeoPrune into the production-grade inference engine vLLM, delivering 5.4%-10.5% mean E2E latency reductions with only slight accuracy drops, advancing vision token pruning toward practical deployment.