Expert-routing centroid steering

Causally biases which expert(s) a mixture-of-experts router selects by shifting a token's routing-space position toward (or away from) a specific expert's learned centroid, or by directly rewriting an expert's own output-projection weight vector — a routing-geometry analogue of ordinary activation steering, operating on router coordinates or expert weights rather than the residual stream.