A Free Model Returned Markdown That NVDA Read as One Endless Sentence: Normalizing LLM Output Before It Reaches the Live Region