Recent works (e.g., [Li et al., 2023]) employ cross‑modal attention, but they target limited‑scale datasets (≤ 1 M pages).