Subspace PGA (Predictive Geometric Alignment)

Tests whether a layer's full representation-space distance structure is significantly better preserved by projecting onto the unembedding matrix's top-k readout subspace than by projecting onto a random subspace of the same dimension — a z-scored RSA variant that asks specifically whether intermediate-layer geometry is organized around the subspace the model will eventually read predictions out of.