perf(ui): fewer Gradio steps + live STEPS_PER_EPISODE, tighter token cap

Space T4 + real HF path: split requirements, GPU defaults, API reset behavior

Finalize eval-friendly defaults: offline baseline, deterministic API reset, docs cleanup

feat: inference router (HF/NIM/OpenRouter/mock); README aligned with judge rubric; chart axis labels

Final hardening for 100% submission readiness

Final Structural Fix: Preserved code in candidate rankings for authentic DPO export

Fix DPO Dataset Generation: Real preference pairs now exported from adversarial loop

Real RL Training Layer: DPO Dataset Export + Unsloth Trainer + Training Report

Final High-Impact Upgrade: Real LLM Inference, Candidate Ranking, and HF API Support

Final cleanup for FORGE-v4: Colab entrypoint, OpenEnv API, 10x optimization, and Judge Narrative generation

Run app.py, generate logs/charts, and update guide.md