#3 AI SEO payload

[arhimede]

3. Markdown twins — for every generated page/index.html, also write the source markdown to page.md (or page/index.md). You already have the source in hand; this is a copy plus a front-matter strip. Add a # Title + canonical URL line at the top of each twin so it stands alone as a document.
4. llms.txt — generate from the aggregation manifest, not by hand. One per documentation version, plus a root llms.txt that links the versions and marks the current stable one.
5. llms-full.txt — concatenated markdown per version. Watch the size; if it's over a few MB, split per component (api, light, admin, frontend, dot-* packages) and index them from llms.txt.
6. sitemap.xml — regenerate with real lastmod from the source repo's git commit date for that file, not build time. Build-time timestamps make every page look churned on every deploy.
7. Canonical + version metadata — <link rel="canonical"> per page, and a machine-readable version marker (<meta name="docs-version">) so agents don't mix v5 and v6 answers.
8. robots.txt — served from the new host, pointing at the sitemap; per-crawler rules go here now that you control the whole surface.

The above list needs subtasks