feat(dv2): bv_order_canonical PostgreSQL smoke seed + verify (G1)

[brownjuly2003-code]

What

G1 (endgame plan) — the bv_order_canonical live smoke artifact. The DV2 vault's governance verify seeds and exercises only the customer side (bv_customer_mdm); the reconstruction-heavy bv_order_canonical view had no PostgreSQL-dialect order seed, so a live query against real order rows was the last unproven vault sub-step.

Changes — warehouse/agentflow/dv2/postgres/smoke/

• order_smoke_seed.sql — 8 deterministic orders (own-brand kitchen-appliance legend, RUB), one canonical row each. Mirrors the ClickHouse synthetic_seed.sql conventions ported to the PG idiom (decode(md5(bk),'hex'), ON CONFLICT DO NOTHING). Covers:
  • SCD2 latest-wins across the header UNION (Bitrix beats an older cross-source 1C header) + per-satellite pricing/marketplace collapse
  • a soft-delete tombstone (is_deleted=1) that must not win over the older active version
  • marketplace state for MSK only — the wb satellite the production seed never populated (new coverage)
  • a pricing LEFT JOIN miss (ala has a header but no 1C pricing)
  • per-jurisdiction VAT (RU 20 % / UAE 5 %) and split_part branch derivation
• verify_bv_order.sh — applies the seed and asserts 17 invariants against bv_order_canonical (PASS/FAIL, PostgreSQL-owner recipe, APPLY=1 runs apply.sh first).
• README.md — Mac run recipe, hand-derived expected output table, honest scope.

Verification

• CI (no Docker): order_smoke_seed.sql is parsed under sqlglot's PostgreSQL dialect and checked for stray ClickHouse constructs by 3 new tests in test_dv2_postgres_ddl.py — 28/28 green locally, ruff clean.
• Logic dry-run: the view reconstruction + all 17 expected values were validated on DuckDB (a real SQL engine, DISTINCT ON/split_part/LEFT JOIN) — 17/17.
• Remaining sub-step: the live PASS/FAIL run on a single-node Mac stand (the standalone no-Docker recipe from the governance verify applies unchanged). Transcript target: docs/perf/bv-order-canonical-pg-smoke-*.md.

🤖 Generated with Claude Code

[github-actions]

DORA Metrics

• Window: last 30 days
• Branch: main
• Deployment frequency: 117 total / 27.3 per week
• Lead time for changes: avg 0.44h / median 0.0h
• Change failure rate: 70.94% (83/117)
• MTTR: 0.29h across 2 incident(s)