By: ValentinHirschi <valentin.hirschi@gmail.com>
test: qualify bounded rule portfolios and record four-loop screens Add saved-program source replay and worker determinism controls. Record exact baseline/candidate qualification and the negative branch-first/rank-first measurements. Keep the separately reviewed P3 engine patch unstaged until matched performance validation. Co-authored-by: Codex <noreply@openai.com>