RFM/AGOP Steering Direction Extraction

Extracts a per-concept steering direction for diffusion-model U-Net/DiT activations by training a Recursive Feature Machine (RFM) as a target-vs-rest classifier on forward-noised (not reverse-sampled) labeled activations at a fixed low-noise timestep, then taking the sign-corrected, eigenvalue-weighted top-k eigenvectors of the RFM's Average Gradient Outer Product (AGOP) matrix as the direction to add back into activations during sampling.