Spectral Consensus (Global Evolutionary) Steering

Rectifies a noisy per-layer contrastive steering vector by projecting it onto the dominant principal component of a cross-layer, cross-sample matrix of tangent semantic differences — treating that first PC as a single globally-consistent 'evolutionary direction' shared across all layers, with a Wedin/Davis-Kahan perturbation-theory guarantee that this PC1 estimate converges to the true direction as sample x layer count grows.