Finalizing CodeArena RL Benchmark: frontend improvements, GRPO training scripts, and cleaned environment

Complete all tasks: Adaptive curriculum, GRPO, React frontend, LLM-as-a-judge

Upload meta

fix: reset task_id parsing, grader tuple crash fallback, and inference score output

fix: clamp scores to strict (0.001, 0.999) range

fix: openenv.yaml tasks graders + strict score bounds

fix: add safe_reward() clamp at every reward return point

Add normalize_reward(); enforce strict (0,1) on all paths

Update grading logic to strict (0,1) ratio fallback

Initial commit