Cancel a finite position bias across presentation orders without overflowing log-probability differences. Previously full rotations of finite ±1e308 logits produced non-finite averages and failed instead of yielding equal probabilities. Per-order normalization constants cancel in the final softmax, so average divided logits with math.fsum before normalization; reported raw_logits now contain the averaged logits.

Validation: the valid baseline regression fails; 67 focused contract/generation/layout tests pass with one existing skip. The available CPU suite has 660 passes, 135 skips and three dependency/fixture failures reproduced on the pinned baseline. Full collection also encounters four missing-Torch modules. Those limits are preserved; no full-suite pass, model evaluation or GPU result is claimed.