Router-expert weight coupling analysis

Measures cosine similarity among a sparse mixture-of-experts model's learned router weight vectors, and correlates router scores with real per-token expert-gate activations, to test whether router and expert weights become geometrically aligned because they receive gradient updates proportional to the same routed hidden-state direction.