If time gets short, priority is: ensemble, then main experiment, then demo, then public-dataset track, then ablations. Compute may spill past one night, but the freeze in section 3 holds: anything ...