Redesign reward for discrimination: efficiency multiplier, strict penalties, stretch bonus, start at level 1

refactor: extract agent_utils.py (shared prompt/commands/server utils), simplify reward to env+format, add audit.py

Data-Centric AI RL Environment — OpenEnv Hackathon Submission