rswan-llm-lora v3 — sweep-best val/loss 0.5804 (r=4)

Proven on T1000 4GB. Qwen2-0.5B-Instruct + LoRA r=4 α=8 on nyu-mll/glue sst2 200/40 @64, batch 2 eff 8, 1 epoch, bf16.

• Registry: rswan-llm-lora v3 (mlflow 9db5245cd0394a478c9934aa4b85efdf, hermes/llm-lora sweep hermes-llm-lora-v1 trial 004)
• Metrics: train/loss 1.0963 (prior single) vs sweep-best val/loss 0.5804 val/acc 0.70
• Sweep parent 919de78a9554471892f314e25ebaeceb; v4 (hermes-llm-lora-v2) proves gated auto-promote but 0.7684 regresses — v3 stays SOTA head.

Artifacts: model.pt (torch.save state_dict fallback from mlflow.pytorch.log_model pt2 gate), checkpoints/best.pt Spec: optional-skills/mlops/mlflow-pytorch-trainer/docs/GOVERNED_TRAINER_SPEC.md