Re-bind the DWP UC composition and TCL disability-children targets to the publisher's grain (#808, #809)

[juaristi22]

Closes #808 and closes #809 (mechanism 3). Part of the #796 capstone queue — these bindings own 5 of the 6 cells still blocking uk_target_fit.

The defect

DWP's "household" is the UC unit of assessment — the benefit unit — and the spine carries FRS benunit ids natively. But the nine composition targets (households_children_1..5_or_more + the four family-type cells) were bound at household grain: "children == 2" summed is_child over the whole dwelling (other benunits' children included), and "SINGLE" meant "any benunit in the household". A single childless claimant sharing a dwelling with another family's child was deleted from single_no_children (−26.5%); a 1-child claim beside another benunit's child was counted as 2-children (children_2 −36.9% while 1/3/4 fit — the signature of mass leaking across adjacent counts). The headline dwp.uc.households and the payment_distribution_* targets were already benunit-grain, so the composition cells weren't even a partition of the headline. The #757 audit patched this same overcount class on the ONS composition cells (lone_households_under_65, couple_no_children) but missed the UC ones. This is also the same defect class as #791 — households_single_no_children was literally the union of the two #791-excluded ONS cells minus their disambiguating predicates — but the UC fix needs no relationship-to-head column, just the grain the spine already carries.

The two TCL disability children cells (#808, −56.5%/−40.2%) had a second binding bug: they summed uc_is_child_limit_affected (~1.27 affected children per household) against DWP sheet 04B facts that count every child in the affected household — the 3- and 4-children sibling rows divide to exactly 3.00 and 4.00 children/household, and these cells run 3.6–3.7. A structural ~2.8× shortfall no reweighting or disability-flagging could close; every sibling from the same sheet already sums is_child.

The change

• Nine composition targets → benunit_count / from_entity: benunit with per-benunit filters over universal_credit, num_children (model BenUnit variable, age<18), and family_type; the single_with_children SINGLE∪LONE_PARENT compound collapses to family_type == LONE_PARENT at benunit grain. The QYP residue (DWP counts 16–19 qualifying young persons; age<18 doesn't) is declared in each binding's notes.
• Two TCL children targets → value_variable/value_reduction sum is_child, structurally identical to the proven children_in_affected_households sibling.
• Nine target_references.json rows flip entity to benunit per the authoring derivation rule. Membership, parity fixtures, and signed-differences resources untouched (comparators key on name/period/value).
• Tests pin the new contract shape and the discriminating materialization cases (a LONE_PARENT claim counts its own child; a childless UC single counts despite a housemate's child; claimant-PIP materializes all-children counts, 4.0 not 3.0, beside the untouched children_affected pin).

Implemented by Codex from a reviewed plan (.codex-work/PLAN.md protocol), zero deviations; independently verified (129 delta-surface tests + ruff green).

Expected measurement (seam run follows on this branch)

Initial estimates for the two TCL cells should rise ~2.5–3× toward 235,270 / 479,460; the composition cells re-grain; the four family-type estimates must now sum exactly to the dwp.uc.households estimate — a coherence identity the old grain could never satisfy. Must-not-move: dwp.uc.households, payment_distribution_*, the TCL household cells and children_in_* (<0.4%), all non-UC families. The frozen scoring-register version flips; the candidate-eval pin re-freezes after acceptance.

Same-defect follow-ups deliberately out of scope: dwp.uc.scotland_households_child_under_1 (fits at −0.1% under the old grain; re-bind with its own measurement) and the local constituency matrix uc_hh_*_children (uk_runtime/local_targets.py:218-227, local-lane PR).

🤖 Generated with Claude Code

[juaristi22]

v13 seam measurement (spine-j, this branch, first run under the floored policyengine-uk 2.92.1) — the re-bind does what the diagnosis said. Receipts: runs/uk-623-first-calibrated/spine-assessment-v13/.

#808 cells clear exactly:

cell	v12	v13	initial support
children_disabled_child_element	−56.5%	+0.1%	72,118 → 190,414
children_claimant_pip	−40.2%	−0.3%	72,943 → 182,933

The coherence identity holds to the household: the four family-type estimates sum to dwp.uc.households exactly, at initial (4,242,978) and final (5,501,928) — the composition cells are a true partition of the headline for the first time.

The honest composition picture (v12 → v13): single_no_children −26.5% → −7.4%, couple_no_children −0.0%, couple_with_children −5.9%, headline −20.6% → −18.6%, loss 0.030 → 0.0254. The COUPLE_WITH_CHILDREN payment band (−46.4%) also cleared — no uc_payment_dist cell is past the gate.

What the honest grain exposes — the residual concentrates in claims with children: single_with_children (LONE_PARENT) −35.7%, children_1 −29.6%, children_2 −35.4%, children_5_or_more −29.1% (cells that "fit" before were fitting a mis-measured concept). This is the real #809 residual and the take-up lever re-measurement's target.

Must-not-move held: all TCL household cells and children_in_* within 0.4%; children_affected +0.0%; payment distribution inside; non-UC families flat.

Watch item: the three #807-lifted hmrc cells drift up as the solver presses on genuine UC composition — private_pension_income_count_100_000_to_150_000 +23.3% → +25.5% (0.5pp past the gate), state_pension 50–70k +24.2%, self_employment 50–70k +20.3%. Marginal; expected to relax if the lever raise closes the UC gap from support rather than weight mass. v13 blocks on 5 cells: the 4 genuine composition cells plus that marginal hmrc count.