Activation-cluster functional typing

Partitions a population of tokens into discrete functional categories by clustering their activation vectors (via a norm threshold plus cross-token cosine similarity of the resulting cluster centroids), then validates each category's functional distinctness via downstream ablation — used to show that a nominally homogeneous set of tokens (e.g. all visual tokens fed into a multimodal LLM) actually splits into a small number of geometrically near-collapsed sub-populations with sharply different causal roles.