Labeling data for LLM fine-tuning at scale means quality slips. What QC processes actually catch bad labels before they poison training?