Create a diagnostic visualization library for Vision Transformers that computes ranking consistency and off-diagonal correlations. This tool would help researchers identify exactly when their token compression models will collapse during inference.