Harden: skip wildcard models, make LLM errors non-fatal per step

Fix model discovery: skip wildcard '*' model IDs from LiteLLM proxy

Fix MODEL_NAME=None: auto-discover from proxy /models endpoint, fallback to gpt-4o-mini

Critical Fix: Align internal port to 8000 to satisfy OpenEnv library requirements

Compliance Fix: Resolver setup timeout with lazy Gradio and extended 120s wait

Fix proxy test: exit with 1 on API failure so validator sees the error; fallback to HF_TOKEN if API_KEY is empty

Compliance Hardening: Remove silent fallbacks to force proxy usage

Final Polish: Task-aware fallbacks and surgical refinement

Compliance fix: strictly use API_KEY and API_BASE_URL to avoid proxy bypass

Fix structured output: ensure logging always runs and format matches validator

Fix inference.py: async OpenEnv pattern, from_docker_image, proper error handling

Final Expert Tier (0.9+) Candidate — Groq Baseline Verified

feat: add reward evolution chart to Gradio dashboard

final: comprehensive 0.9+ strategic agent upgrades and infrastructure refactor

hackathon: final submission candidate (removes binary image for HF compatibility)

fix: Synchronize inference port to 7860 for standardized local testing

Revert default LLM back to Llama-3.3-70B-Instruct for optimal general scoring

Configure NVIDIA-Nemotron-3-Super-120B as the default LLM engine

Fix imports for flat deployment structure in HF Spaces

Initial commit of PolicyEvolverEnv - Meta Hackathon