Dimension-Level Intent Fidelity Evaluation for Large Language Models: Evidence from Structured Prompt Ablation — AI News