fix(inference): clamp all reward outputs strictly within (0, 1)

chore(inference): accept codex improvements for hackathon format and add tests