Complete selection and lock rates

Within each arm rank eligible three-seed candidates by mean epoch-20 clean top-1, then mean MRR, then lower learning rate, using unrounded metrics. Report unequal numbers of eligible candidates. Publish all 27 slots and nine candidate means/ineligibility reasons, retry lineage, actual updates and costs. An endpoint rate winner does not expand the grid.

Bind all three selected rates, complete candidate/attempt evidence and timestamp to the unchanged admitted successor digest. If an arm has no eligible three-seed candidate, stop for a documented amendment.

Verify the fresh 95,001-row v3 cache and generate paired all-100-action batch plans and initial states for seeds 7, 17 and 29. Recheck capacity and complete-run allocation limits.