A 2,400-run benchmark on how agents read docs, and why one link to llms.txt does almost all of the work.