Fill in per-task breakdown for the 5 peer models (independently reproduced, lm-eval 0.4.13)

v1: 6-model same-harness arch→score panel (lm-eval 0.4.12, 0-shot, seed 1234, decontaminated)