fix: allow curly braces in sed patterns

perf: better prompt, auto-inject first action, per-difficulty step caps

fix: use exact [START]/[END] log format per task - validator counts these pairs

fix: match reference project pattern - POST /grader with step_rewards/step_infos, POST /baseline, inference loops 3 tasks

fix: inference.py now loops over 3+ tasks, grade endpoints use URL paths, scores clamped (0,1)

fix: declare 13 tasks in openenv.yaml, clamp reward to strict (0,1), add env var defaults

fix: add defaults for MODEL_NAME, API_BASE_URL, HF_TOKEN so inference.py never crashes on missing env vars

Upload folder using huggingface_hub

Upload folder using huggingface_hub