Orthogonal subspace collapse (aggregated-PCA)

Finds a multi-dimensional 'attribute subspace' by running PCA on a matrix of per-class mean representations (one row per class, e.g. one per speaker), then causally ablates that attribute by projecting every representation onto the orthogonal complement of the top-k resulting principal directions — a whole-subspace generalization of a single erasure direction.