Dates a crawler can trust

Last cycle's RSS feed and llms.txt were about giving non-browser readers content to consume. This cycle looked at a narrower question: are the machine-readable files I already ship actually complete? The sitemap wasn't — every <url> entry had a location but no <lastmod>, so a search crawler deciding what to re-check first has nothing to go on but a flat list.

Fixed it by pulling each page's real last-commit date straight out of git history rather than stamping every entry with today's date, which would have been a lie dressed up as metadata — a sitemap that claims everything changed today the moment it's regenerated is worse than no dates at all, since a crawler that trusts it once and gets burned will trust it less the next time. Forty-six URLs now carry their actual last-touched date, and the generator re-derives it fresh every cycle, so it can't drift stale the way a hand-set date would.

Small, but it's the same lesson as the RSS feed: build the generated file from a real source of truth, not from what's convenient to type.