Technical decisions - MyST Sites as Skills

Findings that changed our design assumptions, newest last.

Tests run against fixtures captured from live MyST sites (tests/fixtures/, saved by scripts/recon.ts) rather than synthetic data or live-network calls: offline and deterministic, while still proving the code works on what real deployments actually serve. scripts/recon.ts is temporary and goes away once the CLI can inspect sites itself.
The slash→dot URL→data derivation rule fails for folder-index pages (e.g. /community/jobs → /community.jobs.index.json, not /community.jobs.json); the xref data field must always be used as the source of truth, never derived.
xref heading records include an implicit boolean field beyond the assumed kind/url/data shape.
myst.search.json records are Algolia DocSearch-style (hierarchy, type, url, position, and content on content-type rows), richer than the bare page/heading pointer shape assumed in the design spec.
next.jupyterbook.org was unreachable (connection timeout) during Phase 0 recon; https://jupyterbook.org was surveyed instead as the second required site.
The bare mystmd.org domain (no subpath) 404s on myst.xref.json: each documentation section (/guide, /spec, /jtex) is its own independent MyST site deployment.
scripts/recon.ts has been retired now that the CLI can inspect sites directly.
myst.search.json is deliberately unused by search: it only exists on static-export hosting, and deriving records from page JSONs gives identical behavior on every site.
search and get define sections differently on purpose: search treats every anchored heading as a flat boundary, while get follows heading depth so a section includes its subsections. A search hit points at the nearest anchor, and get on that anchor retrieves it with its subtree.
All myst.xref.json interpretation lives in src/site.ts (openSite): nothing outside it reads xref records or joins data URLs, so a change in the index format touches one file. The same duplication kept reappearing in commands before this rule.