G2 S3: spec/seed number consistency — audit minors m4/m5/m6/m7/n2

[brownjuly2003-code]

Closes the five minor G2-audit findings against docs/generator-spec.md (S3 of new_plen_05_07_26.md). Each finding got an explicit fix-vs-relax decision:

m5 — seed daily rate vs §1 (fixed the data)

satellite_seed.sql + satellite_seed_all_branches.sql spread 10,000 orders over 21 days (~476/day) while §1 pins ~1,965/day. Nothing else in the repo referenced the 21-day window (git grep '24 \* 21' — only the 5 seed lines), so compressing was the low-risk path: window is now (number * 7) % 122 — a ~122-hour (≈5.1-day) flat window, ≈1,965 orders/day, matching §11's "10,000 ≈ 5.1 baseline days" exactly. §11 gained a paragraph stating why §4's monthly seasonality is deliberately not encoded (baseline days are multiplier-1.0 by definition; a 5-day snapshot can't express a 12-month curve).

m6 — 12-digit GTINs failing is_valid_gtin13 (fixed the data)

synthetic_seed.sql's gs1_gtin stems now append the genuine GS1 mod-10 check digit. Implementation: a pinned 160-char check-digit string (one char per SKU slot) rather than inline SQL arithmetic — deliberately, so correctness is machine-verified locally: new test test_invariant_7_seed_sql_gtin_check_digit_string_is_genuine recomputes the string from the stem formula via gs1.gtin13_check_digit and asserts the SQL carries exactly it (plus structural guards on the stem formula). §12#7 now states the vault-seed literals are covered alongside the reference catalog.

m7 — equal-range amounts skewing averages high (fixed the data)

Old bands' midpoints ran above §1's average checks (mp 2,250 vs 2,150; B2B RU 55k vs 52k; dxb realized 101.3k vs 90k — the small dxb/ala slices didn't even wrap their modulus, so realized means drifted further than the midpoint). New bands are centered on §1 means and multipliers re-chosen (empirically, over the exact number ranges) to equidistribute: realized means mp 2,149 / d2c 3,301 / B2B msk 52,062 / spb 52,034 / ekb 51,920 / dxb 89,827 / ala 44,941. postgres_oltp/seed.sql mirrors. No new distribution scheme — same lo + (n * mult) % mod technique.

m4 — §12#4 vs §1 contradiction + weak/unpinned tests (fixed spec text + tightened tests)

§12#4's flat "Avg B2B check ∈ [30k, 80k]" excluded dxb's §1-mandated 90k. The data is right, the sentence was wrong: §12#4 now claims the order-weighted aggregate (§1 ⇒ ≈54.9k) and explicitly names dxb's 90k as the by-design outlier. Tests: test_invariant_4 now asserts the order-weighted B2B and marketplace aggregates (not just per-branch proxies); test_invariant_5_per_sku/test_invariant_7 pin len(products) == 160 / len(suppliers) == 30 so their loops can't pass vacuously; new test_build_reference_default_counts_match_spec pins the default catalog shape incl. the §6 country split 22 CN + 5 RU + 2 AE + 1 KZ (also catches silent tax-id-collision drops in _make_suppliers).

n2 — §10 FX constants declared but never applied (relaxed the spec)

Wiring AED/KZT into the seeds would be a currency-model rebuild, far beyond this finding's scope. §1/§10 now say what is actually true: every branch is seeded in ₽, no generator or seed converts FX at runtime, and the pinned AED/KZT/CNY constants live in reference/legend.py solely as the fixed basis for doc/evidence-level conversions (and for any future revision that does store branch-local currencies).

Verification

• ruff check / ruff format --check on src/ tests/ scripts/ — clean
• mypy src/ --ignore-missing-imports — clean (124 files)
• pytest tests/unit/ tests/property/ — 1385 passed; 1 env-only failure (test_sdk_wheel_contains_py_typed_marker: local machine lacks the hatchling build backend for --no-isolation; unrelated to this diff, green in CI)
• Amount-mean and window math verified numerically over the exact seed number ranges (values quoted above)

Reviewer sanity-checks: the 160-char check-digit string (test recomputes it — trust the test, not eyeballs); substring(..., number + 1, 1) 1-based indexing in ClickHouse; note docs/dv2-multi-branch/demo_evidence.md remains a dated capture of the old seeds — evidence regeneration is its own later plan step on the Mac stand.

🤖 Generated with Claude Code

[github-actions]

DORA Metrics

• Window: last 30 days
• Branch: main
• Deployment frequency: 84 total / 19.6 per week
• Lead time for changes: avg 0.58h / median 0.0h
• Change failure rate: 55.95% (47/84)
• MTTR: n/a across 0 incident(s)

MTTR note: No failed mainline CI runs in the selected window.