GRPO Phase 13: custom rollout_func for markdown JSON tool calls

Phase 10: SFT training script (Qwen2.5-Coder-3B + LoRA via TRL)

Phase 9.1: --skip-completed flag for cheap re-runs

Phase 9: hard early-submit gate at env layer (kills the exploit class)

Phase 8: SFT corpus builder with 6-filter pipeline

Phase 7: close the 'submit source unchanged' exploit Kimi-K2.5 found

Add extended arena stuff