docs(discovery): evaluate bounded business name/address retrieval

[Medformatik]

Summary

• Complete Investigate ordinary business name/address retrieval using the Berlin EDEKA case #430 with a concise decision report, 13 frozen queries, compact reviewed business identities, 52 distinct raw/adapter inputs, meaningful adapter regressions and an offline replay. Move dated UI/status/timing/repeated result captures and throwaway probe/code/measurements into checksum-linked external archives; preserve existing Git history. Current PR is about 2,225 added lines, down from about 5,534.
• Evaluate a disposable local lexical query over existing Overture data, with stable GERS/provenance, address-role/closure/identity guards, candidate/deadline caps and source credits. Combined verified Hit@3 improves 4/9→8/9; Alexa association, three address-query Enter decisions and deployment/cold/regional performance remain gates. GO staged lexical development; NO-GO production adoption.
• Replay fixed provider inputs against fresh main fe6b7969979ae024655e38d3b4f605295e757f93 (fix(search): require complete business location evidence #432 included). Preserve distinct nearby businesses during autocomplete deduplication #429/fix(search): preserve distinct nearby businesses #438 is separately evaluated at dadd03d5339ad65b64de7a8240054d3befacd10c vs verified parent 5f07b70967319b4b2c86769b542b551b93f98e02; current main 9b212363ba891e97ac4a19c080a10bbc0e343cbd has identical base traces. Geocoder rank 65→78 and combined rank 75→90, unchanged recall/Enter: dedup preservation, not retrieval. Its owner/worktree/branch remain untouched. No application provider/schema, production data/cache/configuration, ingestion or deployment changes.

Report · Reusable corpus/tools · Dated capture archive · Reviewed probe/measurements

Related issues

Fixes #430

References #424,#428/#432,#429. This completes the bounded investigation and authorized local follow-up, not a production business-search feature.

How was this tested?

• 113 focused tests passed in 10 files;2 opt-in database suites skipped. Includes 14 compact real-MapTiler-adapter/term-extractor checks; no tests force future ranking to retain historical bugs or freeze transient UI/status/timing records.
• pnpm check-types: 30 cached Turbo tasks plus fresh evaluation compiler passed. pnpm lint passed with existing warnings. Documentation production build and git diff --check passed. No lockfile/dependency changes.
• Disposable local PostgreSQL probe: 10 normalization/SQL tests and 15 transactional DB checks passed, including exact house/locality role constraints, postcode/region negatives, wrong operator/house, same-address consistency, separate branches/tenants/duplicate GERS, permanent/unknown closure, remote address, caps, injection and timeout. Initial café-first and review-found postcode-as-house bugs reproduced RED→GREEN in throwaway code. No production database accessed.
• One fresh read-only reviewer verified lossless 52-candidate compaction, archive checksums/source reconstruction and layer/measurement claims. Two Important findings fixed; final probe checks passed. Pullable pinned image/container/readiness commands provided. Both archive revisions' bytes independently verified after upload.
• Current-main source exported privately; raw/adapted/geocoder-rank Hit@3 stays 4/9, lexical7/8 inspected-source positives, combined8/9. No clean production before/after or live prototype UI claim. Warm SQL/command timings have explicit cache/overhead/host-load conditions; cold/regional/provider percentiles unknown.
• Earlier full local runs encountered unchanged service timeouts; 26 isolated affected-service tests passed. Final self-review on the unchanged head ran pnpm exec vitest run --maxWorkers=2 --hookTimeout=30000: 1,619 test files / 17,805 tests passed, 30 files / 138 tests skipped, zero failures (488.09 seconds). No test sources or configured deadlines were changed. Publication uses a per-command local hook override after the earlier executed full gates and current focused/type/lint/docs/probe checks; Remote CI and docs passed on 7f0f57e9439f2716212be24afe9cb922666f835a, together with all five required checks. Existing baseline dead-code findings remain outside scope.

Checklist

• Conventional Commits title (the pr-title check enforces this)
• Local lint/types and full test suite passed — final full-suite worker/hook options and earlier timeout history recorded above
• Added / updated a changeset if a publishable package changed — not applicable
• Docs updated if behavior is user-visible (README and/or docs sources under docs/)
• New code carries the right license header / lands in the correct license tier (see LICENSING.md); external source rights documented separately
• No secrets, real API keys, or .env files committed
• I agree to the Contributor License Agreement (the CLA bot records existing signature status)