LoRA-trained direction realization

Trains a low-rank (LoRA) weight-space adapter so that the adapter's own induced change in generated images causally realizes a target semantic direction -- either a directly-supervised contrastive-prompt-pair direction, or (in a later extension) a pre-discovered PCA direction matched via a cosine-alignment loss -- turning a direction into an actively causal, reusable weight-space edit, as distinct from inference-time activation addition.