fix(search): preserve distinct nearby businesses

[Medformatik]

Summary

• Preserve distinct nearby same-name businesses and co-located tenants in autocomplete. The controlled two-branch REWE fixture now produces two selectable rows with their original IDs, addresses and coordinates.
• Keep explicit shared identity decisive. Ordinary POIs need independent source-record namespaces and equivalent concrete address evidence; conflicting Wikidata identities, addresses or categories prevent heuristic merging. Preserve city, station/entrance, station-versus-square and captured landmark reconciliation.
• Add core and real SearchBar regressions, update the existing search/identity policy and architecture docs, and add a core patch changeset. Address matching is deliberately conservative; unsupported formats or same-source records can remain duplicates until explicit identity is available.

Related issues

Fixes #429

#428 was completed by merged #432 before this work. This branch is rebased onto main 5f07b70967319b4b2c86769b542b551b93f98e02 and changes only deduplication, its tests and documentation.

How was this tested?

• RED/GREEN: reproduced the exact fictional two-branch issue fixture before fixing it. Added negative branch/tenant/address/category/entity cases and positive cross-source/shared-identity controls. One independent final review found postcode-only evidence, catalog contradiction and Eiffel Tower reconciliation defects; each was reproduced in a failing regression and corrected.
• 266 focused core ranking, captured search-evaluation and actual SearchBar tests pass after rebasing, including click and keyboard selection of both branches and one corroborated cross-source duplicate.
• pnpm lint, pnpm check-types, pnpm check:policy, pnpm check-openapi, and pnpm -C docs build pass on current main. Lint reports existing non-gating warnings. No API contract changes are included.
• Full suite: pnpm test --maxWorkers=2 --hookTimeout=30000 passes 17,426 tests / 1,530 files, with 135 tests / 28 files skipped by existing repository conditions. The initial default-concurrency run hit unchanged API setup/publisher timeouts; both suites passed an isolated rerun (81 tests). The final invocation adjusts local concurrency and hook headroom without changing repository configuration or skipping crypto. Publication uses a per-command hook override after these gates to avoid repeating the full suite with the known local 10-second setup deadline.
• Manual: actual local Next app with the exact controlled fixture, empty aggregate providers and a fixed Berlin camera at zoom 15. Personally inspected readable before/after result-list screenshots; clicked branch B and selected branch A with the keyboard, then verified original IDs, addresses and coordinates. Repeated selections after rebasing in Chromium. This is algorithm/UI evidence using fictional records; it does not verify live provider retrieval, datasets, deployment caches, enrichment or map coverage. The stub serves empty tiles and unavailable unrelated integrations.

Checklist

• Conventional Commits title (the pr-title check enforces this)
• pnpm lint && pnpm check-types && pnpm test pass locally — full test invocation and local timeout limitations recorded above
• Added / updated a changeset if a publishable package changed
• Docs updated if behavior is user-visible (sources under docs/)
• New code carries the right license header / lands in the correct license tier (see LICENSING.md)
• No secrets, real API keys, or .env files committed
• I agree to the Contributor License Agreement (the CLA bot will prompt on your first PR) — maintainer is allowlisted by the existing CLA workflow

Readable controlled before/after screenshots are attached directly to this PR; no review images are committed.

[Medformatik]

Controlled fictional #429 REWE fixture in the actual local SearchBar, at Berlin [13.4, 52.52], zoom 15. Before: only Street A 1 survives. After: Street A 1 and Street B 2 (338 m apart) are separate choices; click/keyboard selection retains each original ID, address and coordinates. The API is a controlled stub with empty tiles, so this does not establish live provider coverage. Both cropped result-list screenshots were personally inspected; no review images are committed.

Before

After