- obs.reward now always returns current grader score (clamped to 0.01-0.99)
instead of raw shaped reward — ensures every step reward is strictly in (0,1)
- raw shaped reward preserved in obs.metadata['raw_reward'] for debugging
- [STEP] log format changed to :.4f for precision
- health endpoint returns {"status": "healthy"} per OpenEnv validator spec
- grader clamp widened from (0.001, 0.999) to (0.01, 0.99) for :.2f safety
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
