Length-Dependent Processing Modes in Transformer Attention: Evidence from Multi-Architecture Ablation Studies | Synapse