fix(data): add source citations for 7086 website records

[Seungpyo1007]

Summary

Upgrades 7,086 data/website/ records from uncited (Wikidata-only,
verified: false) to source-verified by adding an independent, checkable
authoritative citation and flipping verified: true. No stored values were
fabricated, guessed, or altered.

Addresses the uncited-website backlog. Closes #296.

Methodology

Every website record previously carried exactly one source_urls entry — its
Wikidata item (wikidata.org/wiki/Q…) — and ~32k were verified: false. A
Wikidata item alone is a database cross-reference, not an independent
confirmation, so those records were effectively uncited.

For each verified: false, Wikidata-only record:

1. Resolved its Wikidata QID and queried the Wikidata API
   (wbgetentities, props=sitelinks|claims|labels).
2. If the item has an English Wikipedia article (enwiki sitelink), that
   article is a real, human-checkable authoritative source describing the same
   entity. The exact article URL returned by the API was appended to
   source_urls and verified set to true.
3. Cross-checked stored launch_date against Wikidata inception (P571).
4. If the item has no enwiki article, the record was left untouched
   (still verified: false, Wikidata-only) — no citation was invented.

Every added URL is the genuine Wikipedia sitelink Wikidata reports for that
exact entity; verified: true is only ever set together with a real article
URL. This is fully reproducible against the live Wikidata API.

Counts

Outcome	Count
Cited-and-verified (added enwiki Wikipedia URL, verified: true)	7,086
Value-corrected	0
Left uncited (no enwiki article; not fabricated)	25,346
Left uncited — launch_date vs Wikidata P571 year mismatch, flagged for manual review	17
Already verified: true before this pass (untouched)	7,635
Total website records	40,084

On the 0 value corrections

17 records had a launch_date whose year disagreed with the Wikidata inception
date (P571). These were not auto-corrected: a website's launch date and the
inception of its subject/parent organisation are distinct facts (e.g. Netflix the
company, 1997, vs. the streaming service, 2007). Overwriting one with the other
would introduce wrong data, which the task forbids. They are left exactly as-is
and flagged for manual review instead.

Verification

• python -m app.validate → passed.
• TechEngine python integrity_check.py <worktree>/data --strict → exit 0, no
  hard anomalies (empty structural section; only pre-existing advisory
  CPU/GPU benchmark ratio outliers printed, unrelated to this change).
• Static dump refreshed from TechEngine
  (python -m app.dump, fresh scratch DATABASE_URL, TECHAPI_DATA_DIR
  pointing at this worktree) — regenerated site/public/v1/websites for the
  affected records (last commit). Unrelated pre-existing dump drift in other
  categories was reverted to keep this PR scoped to data/website/ only.

Scope

data/website/ only. 7,086 data files + 7,086 regenerated dump files.

This is one batch; the remaining uncited records (those without an English
Wikipedia article) are intentionally left for future batches per the task's
"batch into multiple PRs" guidance.