How we measured AI writing across arXiv, and where the measurement breaks