We got the deletion half moving by putting a hard size cap on the file: a commit-gate test fails any commit that pushes it past a fixed byte budget, so nothing lands until something moves out. The why still gets written down - each rule change adds a changelog entry with the date and the instruction behind it, and the file itself carries only current rules, so a policy change deletes the contradicting old text instead of layering on top of it. What I don't know is whether notes survive contact with a real file over months - have you kept them current on one of yours?