[EMNLP 2026] Beyond Token Positions: Safety Alignment Across Denoising Steps in Diffusion Language Models