“In Transformers with standard residual connections and PreNorm, uniform aggregation of layer outputs dilutes each layer's contribution as network depth increases.”