feat: holdout split — 5 reserved scenarios for eval (Task A)

fix: exclude instruction keywords from reward overlap pool (anti-hack)

data: untrained baseline rollouts (overall 0.472)