Can a search hold its own bar? In-loop deflation and its exploits in autonomous strategy discovery

Preprint of a manuscript submitted for publication. The paper moves multiple-testing correction inside an autonomous strategy search: the search optimizes a deflated objective against its own running trial ledger, stops when its edge evaporates under that accounting, and spends a bounded statistical budget, and the paper measures what this in-loop accounting certifies and how it can be gamed, including a variance-dilution exploit and its defences, across a frozen 53-run experiment suite. The code and data artifact accompanying the paper is archived separately at 10.5281/zenodo.21725589; every reported number regenerates from it.