Fast frame-consistent video models as "world" models