Adopt a site (or repair an adoption)

Movar adopts a site in two phases, in order. Most sites stop after Phase 1. This skill is the orchestration; the invariants live in the canonical docs it points at — read them at the step that needs them, never restate them.

Ground rules (both phases)

• Live proof is mandatory. Every claim about a site's markup or behavior is verified against the live site in a browser, with the probes in references/probes.md. Prefer Claude-in-Chrome (real sessions); fall back to the in-app Browser pane (logged-out — note the difference in your report). Logged-out caveats: feeds/grids sometimes stall hydration — that is transient, retry the page before concluding "empty"; survey mobile via the mobile viewport preset (UA emulation) against the m. host; NEVER solve a CAPTCHA — fall back to corpus evidence and say so in the report.
• Say what you could not observe. A surface you could not reach is reported as unverified, never silently assumed covered.
• Cheapest lever wins: never start Phase 2 work a Phase 1 rule would solve.

Step 0 — Surface inventory (before either phase)

Adoption starts with an inventory, not a survey: enumerate every page type the site serves and classify each row before touching any code. Two axes:

• Hosts — everything the host predicate will match (a suffix predicate like *.youtube.com drags music/kids/studio subdomains into scope whether intended or not; each matched host is a row).
• Routes — home/feed, search + every vertical it serves, item pages (alone and with list/queue context), section/channel/category tabs, list pages, posts/UGC surfaces, hashtag/topic pages, live hubs, the mobile host's mirrors of all of the above. Use the route-inventory probe (references/probes.md) plus the site's own nav.

Classify every row as one of: switch (Phase 1 acts), model (Phase 2 extracts), ignore (deliberately out of scope — record the stance, e.g. "the player itself", "user's own queue"), or unverified (could not observe — say why). The inventory is the adoption's contract: the phases below work through its switch/model rows, the done-bar requires every row classified, and the extractor header + PR carry the final table. A row nobody classified is a gap, not a stance.

Phase 1 — whole-site language switching

Goal: the site itself serves the target language.

1. Survey the site's own levers live — what does it already offer? A picker/gate (@movar/lang-pickers may already detect it — check packages/lang-pickers/AGENTS.md), <link rel="alternate" hreflang>, a cookie/localStorage key, a path segment or subdomain, query params. Use the census probe; confirm each lever by DRIVING it live and watching the URL/cookie/render.
2. Author the SiteRule. The strategy menu, wiring steps (registry, SAMPLE_HOSTS, behavioural test), and risk notes are canonical in apps/extension/src/sites/CONTRIBUTING-A-SITE.md. Follow its wiring list verbatim.
3. enforce: true needs two extra proofs (search-engine-shaped sites): the strategy is no-op-safe at target (re-run changes nothing), and the SPA-navigation safety holds — run the nav-timing and guard-dance probes and read docs/pitfalls.md §5 before touching anything event-driven.
4. Corpus fixture. Capture the picker/redirect surface per packages/page-content/fixtures/README.md into fixtures/pickers/ or fixtures/redirect-sites/.
5. Live-drive the finished switch on the real site as the closing proof.

Phase gate — does this site need Phase 2?

Graduate to Phase 2 only when the correctly-switched site still shows blocked-language content items: search results, feeds, recommendations, user-generated content (Google and YouTube are the two residents today). A site that fully re-renders in the chosen language stops at Phase 1 — a Phase-2 model is a budgeted content-script chunk plus a corpus and a standing maintenance surface; it must earn that cost.

Phase 2 — page-structure model

Goal: a @movar/page-content extractor that turns every hot-path page into per-card ContentNodes.

1. Hot-path survey. Enumerate the surfaces before modelling: home/feed, search, item page (and item page with list/queue context), section and channel/category pages, list pages, the mobile host. Run the census and card-anatomy probes on each; record which surfaces you could not observe.
2. Write the shapes. Selector and sampling discipline are canonical in packages/page-content/AGENTS.md and docs/pitfalls.md §1 — durable anchors only, allow-list text sampling, fail open. Kind/hideMode semantics live in packages/page-content/src/types.ts; nested cards get outermost-wins dedup (see google.ts/youtube.ts for the pattern).
3. Wire the chunk. apps/extension/src/sites/<site>/ model descriptor + lazy models/<site>.js chunk (copy ./youtube/), inside the byte budget (CAPABILITY_BUDGETS_KB, enforced at build — see apps/extension/AGENTS.md).
4. Tests + corpus. Synthetic unit tests per shape; one real-capture fixture per new/changed surface per the corpus README recipe. Verify every verdict against classifyBySnippet before pinning it (probe the classifier with the exact serialized samples); a case the pipeline gets wrong is encoded loudly as an asserted false-keep with a note, never dropped.
5. Live shape-mirror, before and after. Mirror the shipped shapes in-page (probes.md) on EVERY claimed surface: expected node counts, expected nested-dedup count, zero empty samples (except documented fail-open cells). The before-run is the drift evidence; the after-run is the proof.
6. Record provenance + non-goals in the extractor header: the dated survey table and the surfaces deliberately NOT extracted, with the stance — so the next audit can tell stance from gap. Update the package AGENTS.md youtube/google-style bullet for the new extractor.

Definition of done (all of it, both phases)

• Surface inventory complete — every host × route row classified switch / model / ignore (with stance) / unverified (with reason)
• Package-scoped tests, lint, typecheck green (pnpm --filter <pkg> …)
• Corpus fixture(s) per changed surface, classifier-verified verdicts
• Live proof captured: Phase-1 switch driven / Phase-2 mirror clean on every surface (or the unreached surfaces named as unverified)
• Extractor header carries the dated survey + non-goals; AGENTS.md updated
• Extension build green — chunk budgets hold
• pnpm test:coverage && pnpm gen:readme --refresh (metrics-gate snapshot)
• pnpm nx reset, then a plain git push (hooks must run; never --no-verify)
• PR carries the hot-path coverage matrix, the non-goals, and named follow-ups