RL v3 training + eval results

RL v3 adapter (fresh LoRA, REINFORCE, 30 episodes, heuristic override)

Update README with real training evidence

Debug trial results — verify enrollment works

Debug trial results — verify enrollment works

Upload eval_results.json with huggingface_hub

v2: 10K+ diverse traces, heuristic ID eval

Upload exploit-aware SFT LoRA adapter (Qwen3-4B, r=16, 3 epochs on L40S)

initial commit