Raw evaluation results: coverage audits, token budget sweep, 17-config sensitivity sweep, n=300 head-to-head

initial commit