[anttiviljami]

Summary

Field-name and type rules cannot see personal data inside values of unclassified custom attributes, e.g. opportunity link_gebaeudeansicht holding a Google Maps link with the customer's street, house number and coordinates (found while verifying STABLE360-13580). The catch-all scrub of kept strings (scrubText) now also detects, replacing only the matched span:

• URLs: components (path segments, query values, hash, userinfo) are decoded and scrubbed. Plain URLs without PII are returned byte-identical.
  • Map links (Google incl. maps.google.*, Apple, Bing, OpenStreetMap, goo.gl/maps / maps.app.goo.gl): place name / address query → street_xxxx pseudonym + postal code and city; place ids (cid, ftid, place_id, short-link ids) pseudonymized; Google's data= segment (feature / knowledge-graph ids, precise pins) dropped; coordinates rounded to 2 decimals (~1 km).
  • Phones are not matched in path segments (/030-1234567 is a slug far more often), but they are in query values.
• German street addresses: capitalized street word ending in straße/strasse/str./weg/platz/allee/gasse/ring/damm (compound or standalone after up to two capitalized words) + 1-3 digit house number (12a, 12-14, 10/1). Postal code and city stay, consistent with address items.
• IBAN checksum: IBAN matches are mod-97 validated (shorter prefixes ending at a space are tried, so DE89 … 00 BIC still masks the IBAN).
• Gate: strings without @, / or a digit skip all detectors; street matching is behind a cheap suffix+digit prefilter.
• data_classification: 'public' and system keys stay untouched, as before.

Classification fixes:

• internal_user → user strategy (template created_by / updated_by embed user objects).
• Curated _title defaults (text) for billing_account (= contact name), calendar_event, contract, lead, request, ticket.
• data_classification: 'pii' now honours curated and type strategies (pii first_name → person, salutation → redact) instead of falling back to value, and a pii attribute is never kept via a keep override. This unblocks marking names pii in the entity-api blueprints (entity-api MR !1733).

False-positive tradeoffs

• Compound nouns ending in a street suffix followed by a small number are masked (Arbeitsplatz 2, Lösungsweg 3 → street_…). Years don't match (house numbers ≤ 3 digits).
• Streets without a suffix (An der Kirche 3, Platz der Republik 1) and names in free text are still not detected.
• Map links become non-functional for the exact place (by design).

Test plan

• npm test: 99 passed (53 new in value-detection.test.ts: real link_gebaeudeansicht value on an entity and in an activity payload/operations, all map providers, encoded URLs, streets, IBAN checksum, idempotency, false positives incl. product codes, OBIS, dates, prices, UUIDs, plain URLs)
• perf case: 2,000 entities × 8 mixed strings well under budget (whole suite 76 ms)
• npm run lint (tsc) and npm run build
• Release: version bump commit after merge, as for 0.3.0. Not published.

🤖 Generated with Claude Code