Make the focused skills marketplace-aware: add Shopee Brasil

[norightsforlis]

Why

JoomPulse now serves Shopee Brasil alongside Mercado Livre (Brasil), and seller-copilot was already taught both (#28). The focused skills were not: every one of them scoped itself to Mercado Livre, so a Shopee question either went unanswered or came back with the other marketplace's numbers — a wrong answer that looks right.

Each of the 17 focused skills was analysed individually against the live Shopee data before anything was written, and the disposition below follows what the data actually supports.

What changed

13 skills now cover both marketplaces. Each decides the marketplace first — from what the seller says, from a link, or from the identifier shape (MLB… vs a bare 10–11 digit number) — and never guesses or defaults. Each gained the same two blocks, inserted byte-identically so the behaviour cannot drift: a marketplace-decision step, and a statement of how Shopee data differs.

The Shopee output contract differs where the data does:

Mercado Livre	Shopee
weekly sales windows	30-day windows (relabelled — a 30-day figure must never sit under a weekly heading)
seller medals (platina/ouro/prata)	Official store / Preferred (Indicado) / Common — three tiers, no ladder
free shipping, Mercado Envios Full, listing tier	no equivalent → —
JoomPulse dashboard link	Shopee item link
catalogue and buy-box	no such mechanic
seasonality, 12-month trend	history starts May 2026 → stated as unavailable, not drawn empty

4 stay Mercado Livre only, because the Shopee data does not exist. They now say so and offer the nearest real alternative instead of answering wrongly:

• top-keywords-in-my-category — Shopee has no search-demand data at all;
• my-product-vs-catalog — Shopee has no catalogue and no buy-box;
• top-brand-position-tracker — Shopee brand coverage is incomplete by construction, so a ranking would imply a completeness the data cannot support;
• top-sellers-in-category — no per-seller revenue within a Shopee category yet (left unchanged in this PR; the gap is documented below).

Two traps found against live data

Both were caught while verifying, and both are now encoded in the skills:

1. An unreviewed Shopee item reports a rating of zero, not a blank. In a probe of the "worst rated, still selling" items, 4 of the top 5 had rating 0 with 0 reviews (2 000 / 890 / 820 / 610 units sold). Without a guard the skill would brand them as high-demand, terrible-quality products. The low-rating search now requires at least one review before the rating threshold applies.
2. Shopee price and review history are step functions — a value is recorded only when it changes. So "a week ago" carries the last record forward, and an absent record means nothing changed (delta 0), not "data unavailable". This replaces the Mercado Livre rule of giving up after a two-week lookback.

Also confirmed and encoded: Shopee sold counters are rounded into buckets (the top items all report exactly 10 000), so ranking on small gaps is meaningless.

top-sellers-in-category — why it is untouched

Shopee has no per-seller revenue within a category, and the shop profile carries no monthly rollups, so the skill's ranking axis does not exist there. Seven of its twelve columns would be empty: Receita média (mês), Vendas méd. (mês), Vendas 365d, Cancel rate, Sales trend, Classic, Premium. Adding Shopee would mean rebuilding the ranking from per-item figures summed client-side — a separate change, deliberately not bundled here.

Also in this PR

• seller-copilot/references/which-marketplace.md claimed every focused skill was Mercado Livre only. That is no longer true, so it now points at the skills that cover both and names the four that do not.
• README: a short note on marketplace coverage above the skills table.

Verification

• Every description is ≤1024 characters with no angle brackets — the two gates that have broken uploads before (Fix: shorten 7 SKILL.md descriptions to the 1024-char limit #24, Fix: remove XML-like placeholders from new-growing description #25). The check caught two over-limit descriptions during the work; both fixed.
• No internal cube, field or tool names anywhere in skills/ — the public-safety rule from CONTRIBUTING.
• Both shared blocks verified byte-identical across all 13 skills.
• All 18 skill .zips build with <name>/SKILL.md at the root.
• Live Shopee runs confirm the new recipes return real data; a Mercado Livre run confirms the existing path is unchanged (seasonality, platinum counts and the 12-month trend still resolve).
• All 13 ## Visualization sections retained — cards and charts adapted per marketplace, none dropped.

🤖 Generated with Claude Code