ship submission: judge-aligned README, real training curves from 2000-step run, blog, scripts

optimize for hackathon time budget: 256 tokens, 200-step checkpoints

normalize agent C decision before sending to env (defaults + clamps)

disable unsloth fast_inference (vllm not installed)

bump transformers to >=4.50.3 (unsloth needs Qwen3 support)

skip flash-attn build; install unsloth + deps manually

upgrade torch to 2.5+ for unsloth_zoo compatibility; use ampere-torch250 wheel

import unsloth first; pin torch 2.4.0 for unsloth cpp ext compat

broaden TRL import catch; pin compatible transformers/trl/peft versions

entrypoint: skip clone, expect code in cwd

add training driver, job entrypoint, seed + upload support

add training driver + windows setup guide

expand dataset to 40 templates / 8 domains; fix injection edge cases; lazy unsloth import

expand dataset: 40 templates, 8 domains, fixed injection edge cases

initial commit: conflict arbitration env