eval_stack: point REPLAY_ROOT at the in-repo negmesh8 bank

eval_stack: add the negmesh8 160-episode replay bank (8x20) so transfer is self-contained

eval_stack: make the 160-episode bank self-identifying — object list, ep_id mapping, and the incompatible sibling bank

eval_stack: warn that two different VACE checkpoint sets exist with different norm stats

eval_stack: document precedence — flag that upstream EVAL_SEEDS.md action_horizon=10 is wrong (should be 50) and state norm-stat rules

eval_stack: pi0.5 negmesh8 160-episode replay eval harness, job scripts, protocol docs