Spend Experts Where You Are Unsure: Confidence-Adaptive Routing for Mixture-of-Experts LoRA — ThinkLLM