Publish trained GarudaCoder Qwen3.8 27B LoRA

Qwen3.8-27B v1 Kaggle T4x2: audited SFT, pinned base, holdout manifest (Trained with Unsloth)