fix: clamp reward to [0.01,0.99] so .2f never rounds to 0.00 or 1.00

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

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

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

Rewrite inference.py for strict OpenEnv parsing + add httpx

Minimal patch: standalone proxy ping + reward clamped to (0,1)

fix: use API_BASE_URL/API_KEY for LiteLLM proxy — always make API call (Phase 2)

fix: make inference.py crash-proof when OPENAI_API_KEY is missing (Phase 2)

Initial commit