Add training log with full metrics

Savant M0 LoRA adapters — Qwen3.5-9B, r=32, eval loss 0.0673

initial commit