Simplify graders to return plain float (OpenEnv spec)

fix: rename task ids to use underscores instead of hyphens

fix: rename 'final_score' to 'score' in inference output payload

feat: implement full OpenEnv spec compliance

fix(critical): rewrite inference.py to use in-process environment

fix(inference): add rule-based fallback when API keys are absent

fix: replace bad output JSONs with valid strictly-bounded scores