Block A B C: partial observability, LLM judge, adversarial scheduler

v3: compound tasks, hardened graders, other type, 8 tasks total

add 6 tasks, fix log format, multi-turn retry, grader improvements

fix: inference.py calls LLM proxy directly

fix: 20/20 all tasks 1.0

fix: use API_BASE_URL and API_KEY env vars for LLM proxy

fix: self-contained inference.py, no network dependency

fix: emit [START]/[STEP]/[END] structured output for Phase 2 validator

Fix inference.py to hit deployed HF Space baseline endpoint

Add inference.py for hackathon checker