How can representation dimension dominate structurally pruned LLMs?