Add the E9 net-new UK stages: WAS debt chain segment and UC deduction attributes (#685)

[juaristi22]

Closes #685 (WS-E child E9 of #145, epic #665). Plan: repos/uk-685-net-new-stages-plan.md (approved 2026-09-02 after two review rounds); Codex implemented increments I2–I4 under the approved-plan protocol with zero deviations, followed by an adversarial review whose findings are folded in (listed below). Licensed I5 acceptance attached (Part E of the receipts); ready for review — María decides when to un-draft.

What the issue asked for, re-measured against main

Item	State on main	This PR
Bus fares	Already built by E6 (#730): both columns imputed by QRF, clipped to donor support, bounded, allow-listed. DfT anchors are Chronicle #202 facts owned by #789; the NTS gradient reshape is #790's.	No build code. Receipt Part C documents the E6 evidence; Part D (spine vs the BUS05i facts by geography) is licensed evidence handed to #789/#790. PR #153 to be closed with a pointer.
WAS debt	Not taken by E5; the two columns were already allow-listed (June rows) with no producer.	mortgage_debt (HMortGR8) and consumer_debt (HFINWR8_SUM − HFINWNTR8_exSLC_Sum, clipped at 0) as a fourth sequential segment of the was_wealth chain, conditioned on the base predictors, a declared has_mortgage_tenure flag (WAS Ten1R8 ∈ {2,3} ↔ FRS OWNED_WITH_MORTGAGE) and every prior WAS output. Fourth child of the same seed, so E5's fourteen columns are unchanged by construction (SeedSequence.spawn prefix-stable, pinned by test).
UC deduction attributes	Nothing on main.	New benunit stage uc_deduction_attributes after uc_capital_coherence, before cgt_incidence_clone.

UC deduction attributes — design (María's rulings A1–A3)

policyengine-uk 2.92.1 computes uc_has_deduction from uc_deduction_random_draw, not from the latent rate, so persisting only the two #450 attribute columns would leave the engine's flag (which every validation statistic uses) disagreeing with the held attributes on about half the units. The stage therefore persists four columns: the two identity-keyed uniforms (blake2b of benunit_id, seed 0, output-name salts, rounded to float32 and clamped to 1 − 2⁻²⁴ exactly as the engine stores them) and the two attributes derived from them by the engine's own recipe — the eleven-band inverse CDF of u / (incidence × region factor) and the normalised seven-way combination searchsorted on v — read from dwp_uc_deduction_distributions.json, a pinned mirror of gov.simulation.uc_deductions (DWP Tables 2/5/6, March–May 2025 pre-FRR, ODS sha 307ec8fa…, Chronicle package dwp-uc-deductions-march-2025-february-2026). Both of the engine's summation orders are reproduced (left-to-right for the flag threshold, NumPy for the rate). Assignment is to all benefit units (latent, award-independent; the engine gates on would_claim_uc and a positive award at run time, and universal_credit is defined_for = would_claim_uc, so a non-claimant's attributes reach nothing) and conditioned on region only, which is what DWP publishes and what keeps the mirror exact. A parallel pe-uk seam PR (flag reads the held rate) is planned; when it releases, the two draw columns retire at that floor bump.

Verification carried in the tree: engine-tree lockstep test for the resource (requires_uk), a golden mirror test that materialises the engine's three variables on held float32 draws, band-edge equality with the engine's BAND_LOWER/UPPER, closed-world operation drift asserts, permutation identity, enum names on export, and an independent I1 re-derivation of Tables 2, 4, 5 and 6 from the ODS (Tables 2 and 5 exact; Table 6 factors within 0.01 at two decimals under a denominator convention noted in the receipt).

Surfaces moved

Two generic operation kinds (map_uniform_to_banded_rate, map_uniform_to_categorical); the WAS chain op declares debt_segment_predictors; gates: uk_uc_deduction_combination_enum_domain, uk_stage_uc_deduction_attributes (stage-health latent_attribute_realization, which recomputes the binomial tolerance and holds each cell to the tighter of producer and recomputed), uk_support +uc_deduction_support_bounds.json, uk_stage_was_wealth_support +2 columns; export allow-list +4 benunit rows (the two household debt rows were the dormant June pre-declaration, reused per A5); six net_new_column register entries; coverage-manifest sha and was_wealth outputs; spec-bundle sha; contract digests; node-graph cells/consumes/kernels; the H2 fixture regenerated (synthetic WAS donor gained the three raw columns; 27-transform oracle); roster pins 28→29. The exact-k fixture also gains uc_capital_coherence, a #828 omission surfaced by the contract test.

Review findings folded in (adversarial round, 2026-09-02)

Engine summation order mirrored for both formulas; has_mortgage_tenure declared in the manifest and drift-asserted (it was executed but undeclared); the boolean predictor cast like is_renting; stage_health.py restored to main's bytes plus the new gate only (Codex had reformatted the whole module); the new gate no longer trusts the producer's tolerance; fail-path tests for every branch of the gate; a hand-edit indentation drift in source_stages.json repaired; receipts and changelog wording.

Verification

• Hermetic funnel (María's UK + machinery set, test_contract.py, the graph package incl. test_h2_uk_spine_parity in the engine tier), JUnit-counted from this branch's venv with policyengine-uk 2.92.1 present: targeted post-review set 857 passed / 0 failed / 1 expected skip (licensed WAS tab absent); rebased state on main 47c74225 (with UK spine: benefit-unit-grain SPI UC reporter draw with pre-take-up screen (#832); exact capital sentinel (#833) #835) 863 passed / 0 failed across the graph, roster, pin, contract and E9 suites. CI is authoritative.
• ruff check clean on every touched file; tools/ci_test_groups.py --verify ok; tools/build_uk_release_input_coverage_manifest.py --check ok; fixture regeneration byte-identical to the committed directory.
• The two known licensed-cache tests (test_cached_reference_regeneration…, test_cached_candidate_regeneration…) fail on this machine exactly as on main: committed evidence minted under 2.89.0 vs installed 2.92.1 (PR Make the UK eFRS parity re-pin one reviewed command; hold every identity mirror in lockstep (companion to #747) #749's lane); untouched here.
• Licensed I5 (done; receipts in experiments/685-net-new-stages-receipts.md, raw JSON licensed-side): two full builds of the rebased branch (main 47c74225, 28 stages) are payload-identical and pass the spine battery 15/15 (as did the pre-rebase pair). Against a control built from that main with the same inputs the only new columns are the six E9 columns, no common column differs, and E5's fourteen columns are byte-equal. The e9 identity receipt passes (30,455 unique benunits recomputed in original and permuted order equal the stored columns). Engine round-trip on 2.92.1 at 2024 and 2025 (rebased tree, with UK spine: benefit-unit-grain SPI UC reporter draw with pre-take-up screen (#832); exact capital sentinel (#833) #835's reporter redraw upstream): the flag, the attributes and the fallback recomputed from the held draws agree on every claimant (0 mismatches), non-claimants carry zero deductions; mean monthly deduction £66.81 / £50.66 (DWP £67–68 / £51–54), at-cap share 0.145 / 0.263 (pe-uk 0.134 / 0.264), unique-row incidence within three sigma (z 1.8 / 2.0). Debt: donor-faithful shares (0.307 vs 0.323; 0.526 vs 0.530), levels 1.2–1.3 times the donor means, inherited from E5's wealth surface (signed observation). Strict parity: the twin's unsigned set is identical to the control's (15 pre-existing columns from the 2.92.1 recognition re-record, Make the UK eFRS parity re-pin one reviewed command; hold every identity mirror in lockstep (companion to #747) #749's lane); E9 adds no unsigned difference.
• Gate lesson folded in (commit 0d2e4e9d): the first twin was refused by the new stage-health gate on one band cell because the receipt gated a weighted share against an unweighted binomial band; the gate now holds the unweighted realization (identity-keyed draws are weight-blind) at four sigma and reports the weighted share with its effective sample size.
• Rebased onto main after Fix the two production-path breakages the node graph left in the UK spine driver (#836) #844 and then UK spine: benefit-unit-grain SPI UC reporter draw with pre-take-up screen (#832); exact capital sentinel (#833) #835 merged (four E9 commits on 47c74225); the licensed acceptance below was re-run on the rebased tree against a control built from that main, and the six E9 columns are byte-identical to the pre-rebase twins (identity-keyed draws; UK spine: benefit-unit-grain SPI UC reporter draw with pre-take-up screen (#832); exact capital sentinel (#833) #835 does not move them).

UK-specificity note

DWP taxonomies (band edges, combination enum, region factors) and the WAS/FRS column names are UK; both new operation kinds are generic, and the debt segment reuses E5's generic chain machinery.

Not in this PR

Bus targets (#789), scalers (#790), the pe-uk seam PR (I7), closing PR #153 and the pointer comments on #789/#790 (each on María's go), Chronicle facts for Tables 2/5/6.

🤖 Generated with Claude Code