brand-profile agent resolves the WRONG Wikidata entity for same-named brands (not fixed by #359)

## Summary

The `brand-profile` agent resolves the **wrong Wikidata entity** for brands whose name collides with an unrelated same-named entity, because `findWikidataId` searches on the **brand name alone** and ignores the strongest disambiguator we already have: the **site domain** (e.g. the TLD) and the scraped site content.

PR #359 (adobe/spacecat-task-processor#359) anchors the Wikipedia *article* fetch to the resolved QID and guards on `wikibase_item === QID`. That fixes the case where the QID is correct but the article lookup went astray. It does **not** fix the case where the resolved **QID itself is wrong** — the agent then faithfully anchors to the wrong entity and still emits contaminated `products` / `sub_brands` / `competitors`.

This is the sibling of the data-quality report adobe-rnd/llmo-data-retrieval-service#3200.

## Two distinct failure classes

| Class | Cause | Example | Fixed by #359? |
|---|---|---|---|
| **A — QID correct, article wrong** | old step grabbed a same-named *article* despite a correct QID | Lovesac → `Q6690181` (LoveSac, furniture) ✅ | **Yes** — regen now returns furniture |
| **B — QID wrong** | entity *resolution* picked a same-named unrelated entity | Capella → `Q12970` (**the star Capella**) ❌ | **No** — regen still returns star data |

## Reproduction (Class B — verified on prod after #359)

`capella.edu` is Capella University (an online university; `.edu`). Its resolved QID is `Q12970`, which is the **star Capella** (constellation Auriga), not the university.

Before #359 (name-search article): products were a children's TV series ("Cappelli & Company", songs, cassettes).
After #359 regen: products became **stars** — `Capella Aa`, `Capella Ab`, `Capella H` (category "Star (primary giant in Capella system)"), with `products_metadata.brand_wikidata_id: "Q12970"` and `wikipedia_verified: true`. Still wrong — the guard passed because the article does match the (wrong) QID.

## Root cause

`src/agents/brand-profile/services/wikipedia.js` → `findWikidataId(brandName)`:
- calls `wbsearchentities` with `search: brandName`,
- returns the first result whose description contains a company term, else the first result.

There is no use of the **site domain** or scraped content to disambiguate. `capella.edu` (a `.edu`) is an unambiguous "this is a university, not a star" signal that is discarded.

## Prevalence (fleet scan)

Of **851** brands that have a stored products QID, ~**247** have a stored QID whose Wikidata entity does not match the brand (Class B; upper bound — includes some acronym false-positives such as AIB→"Allied Irish Banks", which is actually correct). Clear examples:

- AIA (insurance) → `Q25228` = **Anguilla** (a country)
- American Airlines → `Q838581` = **"comitative case"** (grammar)
- AAA Membership → `Q1823` = **Aceh**
- Capella University → `Q12970` = **the star Capella**
- OECD → polycrystalline silicon; Borussia Dortmund → BYD; Logista → Lovisa jewellery

## Blast radius

Same as #3200: `products.items[*].category` is the priority‑1 category source for synthetic-persona prompt generation, and the profile also feeds Semrush, Citation Attempt / Strategy Chat, and GSC. A wrong entity contaminates all of them.

## Suggested fixes

1. **Disambiguate with the site domain + scraped content.** Pass the base URL (esp. the registrable domain / TLD) and the industry/vertical into `findWikidataId`; prefer the candidate whose Wikidata sitelink/official-website (`P856`) or description matches the domain/industry. A `.edu` should never resolve to a star.
2. **Post-resolution guard.** After picking a QID, verify its Wikidata description/`instance of` (`P31`) is consistent with the brand's known vertical; reject on clear mismatch.
3. **Fail safe on low confidence.** When no candidate matches the domain/industry, return **no data** rather than a wrong entity (consistent with #359's "no data rather than wrong data" principle).

## Links
- Original data-quality report: adobe-rnd/llmo-data-retrieval-service#3200
- Partial fix (Class A): adobe/spacecat-task-processor#359

Bắt đầu tại src/agents/brand-profile/services/wikipedia.js ở findWikidataId(brandName), sau đó kiểm tra cách brand-profile cung cấp domain của trang web và ngành hoặc vertical được thu thập. So sánh luồng phân giải với bản sửa một phần trong PR #359. Hoàn tất khi các brand trùng tên được phân giải thành một QID nhất quán với domain và nội dung, còn các kết quả khớp có độ tin cậy thấp trả về không có dữ liệu thay vì làm nhiễm profile.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.