fix: provide task graders, correct step return signature, remove meta headers

fix: clamp final score strictly between 0 and 1

fix: fallback API_KEY to HF_TOKEN for evaluation runs and correct local port to 7860

fix: use correct base url target variables for LLM proxy vs OpenEnv client

fix: use openai client configured with API_BASE_URL to call LLM proxy for inference

fix: output correct structured tags for validation

fix: add try-except blocks for GenericEnvClient initialization and stepping

fix: use config URL to avoid ConnectionRefusedError

Initial clean commit