fix: normalize [END] score to (0,1) range — divide by 4.8 max score, clamp to (0.001, 0.999)

fix: wrap LLM call in try/except with fallback responses to prevent inference.py crash

fix: rewrite inference.py to use LiteLLM proxy via API_BASE_URL/API_KEY env vars

fix: add [START]/[STEP]/[END] structured output blocks to inference.py for Phase 2 validator

FastAPI OpenEnv server