CAPSTONE: Composable Attribute‑Prompted Scene Translation for Zero‑Shot Vision–Language Reasoning