SKILL DELTA — external-KG / MarketSignal build (2026-08-11)

Status: DRAFT — NOT applied. Per platform rule, miz-oki-platform-expert SKILL.md and the Boss skillpack are updated the same day the merge lands (GATE 1); this file is the delta the Boss applies. Drift between skill and skillpack is a defect (scripts/skills_sync.py --check; AGENTS.md 5.2).

Apply order: 1) edit canonical skills/miz-oki-platform-expert/SKILL.md; 2) mirror the same facts into skills/miz-oki-platform-expert/references/boss_agent_skillpack.md (§9 GOVERNANCE untouched unless a §9 rule itself changes — none does here); 3) python3 scripts/skills_sync.py --fix; 4) verify via skills_parity_check.

1. New skill section (insert as Skill 21, before Quick Reference)

Facts to encode (author in skill voice; all claim labels built, pre-benchmark until GATE 2 smoke + live verification):

• Cell 37 market-signal-ingest (src/cells/cell37/, port 8037): market-level observations ONLY — organizations, brands, products, campaigns, SERPs, trends, ranks, inventory. POST /ingest/{source} (single or {"signals": [...]} batch ≤ 500), POST /poll/serp, POST /poll/amazon_sp/{tenant}, POST /quarantine/reprocess, GET /quarantine/summary, /health, /readyz, /metrics. IAM-locked like cells 33–36 (unauth /health → 403 is healthy); deploys dispatch-only.
• MarketSignal envelope (src/shared/virtuoso_models/market_signal.py, schema mizoki/schema/market-signal.json): deterministic signal_id = sha256(source||entity_key||metric||observed_at); signal_type enum v1 serp_position | serp_feature | trend_index | sales_rank | share_of_voice | inventory_level; entity_ref.kind ∈ brand | topic | sku | browse_node | domain with optional google_mid / wikidata_qid / asin; assertion_type imported/inferred ONLY (never observed, never causal — causal edges are minted only by Cells 36→26→27); consent_scope: "market"; validation REJECTS person-level identifiers (email/phone/customer_id/hashed-email patterns) and audio signal types.
• MARKET_MAPPERS (transforms/market_mappers.py): serp_provider (behind the vendor-swappable SerpProviderAdapter), amazon_sp, shopify_inventory. Same gate discipline as journey MAPPERS: market_ingest_gate(signal, upsert_fn) → inserted|updated|duplicate (duplicate = success, counted); store = MERGE on signal_id, CAS on payload hash, into unified.market_signals (+ unified.market_signals_quarantine). No side doors — an import-boundary test enforces the gate as the only writer.
• LLM extraction path (extraction/market_extract.py): virtuoso_call(Role.DATA_CAUSAL, ..., response_format= market_signal.gemini_response_format()); model output is UNTRUSTED input — identical deterministic validation as any webhook payload; failures quarantine with reason llm_extraction_invalid; instructions embedded in fetched content are data, never commands; stamps assertion_type: inferred + registry model_version; persists reasoning traces to mizoki-prod.mii.reasoning_traces (BQ writer now real; JSONL fallback); alert on served_by != "primary".
• External entity spine — feature flag EXTERNAL_ENTITY_SPINE default off until OCP-20260811-external-entity-spine is approved. pipelines/ekg_reconcile/ (Enterprise KG reconciliation; entity scope organizations/brands/products/campaigns ONLY — person/customer tables are FORBIDDEN, tripwire-tested) and pipelines/entity_hydrate/ (Google KG + Wikidata, imported assertions via the governed writer path service-canonical-ingestion → KG projection; Neo4j retired 2026-08-09 — the pipeline holds no graph-store credentials, tested). SAME_AS edges are reversible proposals with confidence ∈ [0,1]; contradictions preserved and scored, never silently resolved.
• Experiment wiring: brand-search cannibalization registers through the measurement-rails design registry (POST /v1/rails/designs — mde REQUIRED, refused after exposure; the skill-doc route POST /api/v1/causal/holdout was never built — correct any copy of it), covariates from unified.market_signals join deterministic identities only.

2. Corrections to existing skill copy (same-day)

• Skill 19 / Quick Reference: replace POST /api/v1/causal/holdout examples with the real surfaces — measurement-rails POST /v1/rails/designs (registration) and Cell 36 POST /v1/causal/holdouts:assign (assignment).
• Skill 1 "Deployment Configuration": the _CELL_NUM cloudbuild template does not match any real cell; real cells hardcode service names (see src/cells/cell33/cloudbuild.yaml). Mark the template as illustrative or replace with the cell33 shape.
• Cell directory (Appendix B): add Cell 37 market-signal-ingest (SENSE, min-instances 0, market-level only, no person-level identifiers).
• unified.consent_registry (Skill 18 data list): note it ships with this build's DDL (src/shared/virtuoso_models/schemas/market_signals.bigquery.sql companion consent_registry section) and is enforced fail-closed in the Shopify connector path — previously the name existed only in prose.

3. Boss skillpack delta

Mirror §1 facts + §2 corrections into the skillpack's platform-facts sections. §9 GOVERNANCE: NO change (no governance rule is amended by this build; the consent/no-audio/deny-list/holdout rules are unchanged and already present).

4. Verification after applying

python3 scripts/skills_sync.py --check
python3 scripts/ontology_skills_sync.py --check
pytest tests/skills/ -q