540-image phrasing-controlled benchmark exposes VLM textual-prior reliance across 11 models | HACKOBAR_