Split private_pension_wealth out of corporate_wealth on the UK WAS stage (#750)

[juaristi22]

Implements the microcosm side of #750 (uk-data#452 mechanism M2) — the measured lever for the 8 UC caseload/two-child cells blocking the first certified cut (#796). Stacked on #803 (the #792 band-edge fix; its two commits lead this branch and drop out of the diff when #803 merges). Model side already released: policyengine-uk 2.92.0 (policyengine-uk#1824).

The split

The WAS stage folded WAS private pension wealth less current-employment DB (totalpenr8_aggr − dvvaldbt_scaper8_aggr) into corporate_wealth, which policyengine-uk counts as capital in every means test — on the incumbent at 2025, 46% of UC-reporting benefit-unit records fail the £16,000 test, 96% of them on corporate_wealth alone. The stage now emits private_pension_wealth as its own household output: drawn first in chain segment 2, the share-like holdings conditional on it, segment 3 conditional on both. corporate_wealth keeps employee shares/options, UK shares, unit/investment trusts, and the stocks-and-shares-ISA fold. Row identity old corporate_wealth == corporate_wealth + private_pension_wealth holds on the donor by construction and is pinned in tests, with a segment-1 containment test documenting why the first segment's draws cannot move.

Commits:

1. Engine pin bump — policyengine-uk>=2.92 in the three uk extras (locked upgrade, 2.89.0 → 2.92.0, zero transitive churn). Required first: an older engine's loader silently drops the new column, narrowing total_wealth and the exposure keys. 2.92's corporate_sector_wealth keeps the shareholding / corporate-land-value / employer-NI keys whole, so in classification terms only the capital tests move.
2. The split — stage + lockstep spec pair + support bounds (operator-measured on the pinned tab, sha 18b3eb98…: donor pensions − db_pensions min 0.00 over 15,128 rows, zero negative rows → nonnegative_outputs; outward bound [0, 8000000], byte-verified via the tool's --check; the standing Port the WAS wealth imputation and regional property uprating as declarative source stages #714 SDC adjudication extends to the one new column) + export allow-list + stage-support gate columns/zero-allowances + coverage-manifest family outputs + the moved digests (battery policy/manifest/fingerprint, spine and release_cut certification part mirrors, UK spec bundle sha). The QRF household-tail note records the 13th unarmed household column; arming stays the WS-E spine deferrals: known bugs and signed inconsistencies awaiting their increment #796 follow-up.
3. Evidence re-record under 2.92 — the bump alone (before any split) invalidates the committed licensed extraction evidence: 2.92 newly recognizes three loader inputs the pinned artifacts already carry and populate (bus_fare_spending, employment_sector, sic_industry_division — previously in the reference's "unknown export columns" list). The parity reference, the known-gaps candidate evidence, and the coverage manifest are regenerated with their own tools from the same pinned artifacts (identities unchanged — filename/revision/sha256 all identical; every pre-existing share value byte-identical): surface 145 → 148 required, counts and version strings move, nothing else. The June candidate and the spine both carry all three columns, so the widened contract is satisfiable on both sides. This is not the efrs-post-calibration input-mass reference re-pin — that stays deferred behind the uk-data mirror per the UK WAS stage: split private pension wealth out of corporate_wealth (uk-data#452 M2) #750 ordering, and the input-mass totals evidence (fd41cb5f…) is untouched.

What this deliberately does not do

• No input-mass exclusion for corporate_wealth: the gate's reviewed tolerance is 4.5218 and the post-split drift is ~−0.85, comfortably inside — and the UK wrapper fails stale exclusions, so adding one would fail the gate. The drift gets recorded in the seam receipt when the split spine runs.
• No signed-differences re-mint yet: was-wealth-qrf-incidence bounds and the private_pension_wealth net-new-column entry are re-minted from the spine-j measurement in the licensed phase (register entries unused on a compared surface fail --strict, so they cannot precede the spine that carries the column).
• No doctrine or seed changes; the stage is row-preserving.

Licensed phase (rides this branch before review completes)

Spine-j + twin build under the armed battery, partial engine-drift control (segment-1 shares vs spine-i; a full commit-1-only control spine only if segment 1 moves), strict-parity re-mints, e4–e8 ladder, acceptance re-cut, then the seam run with the #792 fix in place — the 8 blocked UC cells' new relative errors are the adjudication input for the #736 queue — and the measurement receipt experiments/452-uk-pension-wealth-split.md. The certification cut itself stays with #793/#736 adjudication per the 2026-08-27 ruling.

Verification

Per-commit: ruff, ci_test_groups.py --verify, targeted UK/spec suites; full build+data shard after the spec changes; licensed bounds --check byte-identity; the two licensed-cache regeneration tests (skipped in CI, run locally against the pinned HF-cache artifacts) green after commit 3. The branch is main-based, so PR CI runs the engine lanes and the wheels gate once, authoritatively; local heavy suites were not duplicated (per the 2026-08-27 compute ruling) — the full build+data shard ran locally at the pre-rebase tip, and the post-rebase delta (the #803 wheels-lane reorder) was covered by its three calibration test files plus a wheel metadata build of all five shards.

🤖 Generated with Claude Code