Register item 18 — Step-0 verification + build record (2026-08-20)

Read-only verification of "Connector rollout phases 2–3 (build, after the above): Merchant Center + public Shopping; then LinkedIn / Amazon Ads / The Trade Desk" before building, and the record of what was then built. Claims are measured in-repo at file:line unless labeled otherwise. Companion to docs/reports/ITEM15_STEP0_VERIFICATION_2026-08-20.md (whose projector this item's KG leg extends) and the item-16 report.

Headline

The register line was stale in the same way items 15/16's were: every phase-2 AND phase-3 adapter has been built and mounted since 2026-08-12 (direct_connectors.py — MerchantCenterAdapter, PublicShoppingAdapter, LinkedInAdsAdapter, AmazonAdsAdapter, TradeDeskAdapter; module memory records them live on revision service-marketing-connectors-00007-52f). What was actually open was narrower and partly a defect list:

1. TENANT-001 gap — /api/v1/direct-connectors/{provider}/sync passed req.tenant_id straight to ingestion (old direct_connectors.py:441) while every sibling ingest door (ingest_batch main.py:650, shopify sync :721, bucket sync :679) resolves the tenant against the verified caller. A mapped caller could ingest into a tenant it does not hold, on all eight providers.
2. Fetch-guard gap — PublicShoppingAdapter fetched caller-supplied options.urls with no scheme/host validation: an authed caller could point the gateway at internal hosts (probe primitive) or ingest arbitrary content as public_source.
3. The KG lineage leg could not exist for ANY direct-pull provider — not just phases 2–3. The adapters emitted provider-prefixed record_types (meta_ads_insight, google_merchant_center_product_report via record_type=f"{provider}_{kind}") and raw-only payloads, while the item-26 projector maps (provider, record_type) keys against payload.normalized. Every direct-pull event would land in the canonical store and skip projection (unmapped:* / no_normalized_payload). The module's connected bar — "source→canonical→KG lineage proof passes" — was unreachable on this door.
4. Test depth: 3 in-service tests covered the eight adapters; no route, tenant, or normalization pins.

What was already right: the governed order itself. The sync route flows fetched records through _ingest_records → service-canonical-ingestion (main.py:300-322,1814-1820) — no side door, rule 1 honored.

What was built (this change set)

• TENANT-001 closed on the direct door: build_router now takes resolve_tenant and the sync route resolves BEFORE the provider pull (a refused caller costs no API quota); response exposes the resolved tenant_id. main.py wires the real mizoki_contracts.resolve_tenant.
• Public-feed fetch guard: _validate_feed_url — https only, public DNS hosts only (metadata/localhost/*.internal/single-label refused, literal IPs must be global), ≤20 urls per sync. Scope stated honestly: literal targets only, no DNS re-resolution — the route stays verify_caller-gated.
• Record contract converged with the bucket puller (item-15 shape): record_type is the CONCEPT (campaign_performance, product_report, feed_item, report_row) — provider stays in source_system and provenance; payload is {"raw": <provider-native row>, "normalized": <flat dict or {}>}. Per-provider normalizers added (tolerant numeric coercion; junk → absent key, never a crash; a non-urn LinkedIn pivot or a TTD row without campaign structure refuses to mint ids).
• Projector vocabulary, phases 2–3 (projector_kg.py): google_merchant_center:product_report → account_gmc_* PlatformAccount + product_gmc_* Product + CONTAINS; google_shopping_public:feed_item → product_gsp_{host}_{offer} Product (host-qualified so feeds cannot collide; no account parent by construction); linkedin_ads/amazon_ads campaign_performance and the_trade_desk:report_row reuse the parameterized ad hierarchy (tokens li/amz/ttd). Product is live platform vocabulary (cell02 EKIS NodeType, market mappers), not a new type. Identity/config props only — metrics stay in the canonical event. ALL_ALLOWED_PROPERTY_KEYS extended in the same commit; still disjoint from PERSON_TOKENS (tripwire test).
• Direct meta/google now project through the EXISTING mappings — the convergence means the phase-1 direct door gained its KG leg for free.
• Tests: tests/connectors/test_direct_connectors_governed.py (20 items — tenant both directions incl. refusal-before-pull, main.py wiring pin, record contract, url guard 13-case parametrize, normalizer tolerance) + 4 item-18 additions to tests/connectors/test_kg_projection.py including the cross-service contract test feeding all seven direct normalizers into project_envelope. Suites: tests/connectors 307 passed, gateway in-service 83 passed (pydantic 2.13.4 venv).

Deliberately NOT built

• GA4 direct rows stay unmapped (report_row, normalized {}): the pull is session/transaction-grained (transactionId dimension) — a different concept from the mapped aggregate audience_segment, and transaction linkage needs its own design pass. Named skip, never a guess.
• TTD rows without campaign structure stay unmapped for the same reason.
• No scheduler/creds/terraform — operator surface, below.

Honest end condition

Build half of item 18 is code-complete and pinned; nothing here is live-verified, because no phase-2/3 provider has credentials: every adapter still reports configured: false and stays ready-to-connect (the module's own bar). Remaining operator work per provider: install credentials/account ids via Secret Manager (GOOGLE_MERCHANT_ID + content-API access, PUBLIC_SHOPPING_FEED_URLS, LINKEDIN_*, AMAZON_ADS_*, TTD_*), map the calling tenant (TENANT-001 registry), then run one bounded /api/v1/direct-connectors/{provider}/sync and prove the source→canonical→KG lineage on the projector sweep — that proof, not this change, is what upgrades a provider's status.