fix: final reward calibration and persistence cleanup for Phase 2

fix: adjust reward range to (0, 1) to pass Phase 2 evaluation