M2 pre-registration (frozen 7 Nov 2026): how much of a published advantage survives matched tuning? Sampling frame, case-selection rule, pooled analysis plan and power simulation

Frozen sampling frame (arXiv cs.LG 2023-2025 papers with a code link in the abstract; two seeded draws), case-selection rule, per-case effect-size and retention definitions, the pooled estimator (median matched retention over rule-selected cases with a bootstrap over cases), its reading rule and a planted-effect power simulation. Deposited before the later cases are run; the first rule-selected case had already been run and is disclosed as not blind.