Certify Build N US Populace as the certified default

[MaxGhenis]

Fixes #480.

Merging this PR is the default flip — it certifies populace-us-2024-buildn-sparse-rmloss100-c3e378a-20260722T010408Z (published 2026-07-22, immutable tag; latest.json deliberately untouched) as the certified US default, replacing Build M. The merge decision stays with the maintainer.

What changes

Same shape as the Build M certification (#475), the #470 common case — policyengine-us==1.764.6 is unchanged (compatibility_basis: built_with_model_package), so pyproject and household snapshots are untouched:

• src/policyengine/data/bundle/manifest.json — US data package rebound to the Build N release (manifest sha256 99d8b7a5…, artifact sha256 38de53a3…)
• us.trace.tro.jsonld — regenerated against the published tag
• uk.trace.tro.jsonld — two sha fields update because the sidecar binds the same (changed) bundle manifest; regenerated with the HF token, not a degraded regeneration
• three pinned test constants (US_DATA_RELEASE_ID in test_release_manifests.py, plus the URI pins in test_models.py / test_us_regions.py)
• towncrier fragment

Scorecard: Build N sparse vs Build M

Metric	Build N	Build M
Final loss	0.028242	0.02666
Within-10%	89.49%	90.21%
Compiled targets	5,672	5,667

The headline pair is not surface-comparable: Build N repaired the target surface itself. populace#486 (capital-gain-distributions level repair) and populace#488 (cross-vintage aging chain completion) corrected 518 target values — state capital gains +$298.5B, taxable interest +$21.6B — so N is scored against a strictly harder, truthful surface than M was.

Gates and receipts (published release manifest)

• All 8 build gates passed. The M-era SSI reconcile-loop gates no longer exist: populace#477 deleted the loop; its replacement one-shot Bernoulli law gate reports bernoulli_law_violation_count = 0.
• Export-mass parity: 0 failures on both legs.
• Reform-coverage smoke: 39/39 probes green, including keogh_distribution_neutralization.
• QRF tail concentration: 14-entry evidenced exclusion register, sha-recorded, fails loud if dormant (owners populace#481/LRUCache.add silently keeps the old value for an existing key #487/Remove policyengine-us-data from the certified bundle package scaffold #432/Notify consumer repos after PyPI release #451).
• Publish contract: green with medical_expense_deduction_total at the adjudicated 0.25 broad-fit bound (populace#490; published row −20.57%). Durable fix staged on populace branch loss-contract-alignment; 0.15 restores once a boosted run holds it.
• CD surface ships off per the populace#449 verdict; the parity generator carries a fail-closed CD_SURFACE_REGIME = "off" constant.

Verification

• scripts/certify_data_release.py ran against the published HF tag (revision-pinned manifest URI)
• scripts/generate_trace_tros.py regenerated both TRO sidecars with the HF token
• Full test battery: 738 passed, 0 failed (3:54) against the published tag
• scripts/bundle.py check: green (rc 0)
• Ruff format + lint clean on the changed test files
• uv.lock deliberately untouched (main's lock is stale at 4.20.1 vs pyproject 4.22.0 — pre-existing, separate concern)

🤖 Generated with Claude Code

[MaxGhenis]

CI complete: 12/12 checks pass (including both full-battery jobs against the published tag), mergeable clean. Local receipts in the description (738/738 battery, bundle check, TROs regenerated with token).

Ready for the maintainer. Merging this PR is the default-flip decision and stays with Max, as does the subsequent latest.json pointer move.