Queryable Is Not Grounded: Reference-Anchored Faithfulness Evaluation for Foundation Vision-Language Models