Port the post-freeze childcare / Tax-Free Childcare inputs and targets, re-pin the Chronicle feed, and activate the DfT bus targets (#834, #789)

[juaristi22]

Closes #834 and #789 (WS-E of #145, epic #665). Plan: repos/uk-834-childcare-tfc-port-plan.md (approved 2026-09-04, re-based on main 96faa5d5); Codex implemented increments I0–I2 under the approved-plan protocol (deviations listed below), followed by the licensed lane (L0–L5) whose receipts are in experiments/834-childcare-tfc-receipts.md. Stacked on #849 (A11: the v2/v3 artifact loader this feed needs): #849's 22 commits ride in this diff until it merges, after which this branch rebases and the diff is #834's own. Base is main so CI runs on the merge ref. Draft — María reviews everything at the end; never merge without her go.

What the issue asked for, re-measured against main

Item	State on main	This PR
tax_free_childcare_spend_routed_share (uk-data #472)	Absent; policyengine-uk ≥2.93 reads it and defaults to 1.0 (the un-routed spend the June cut shipped).	Generic assign_period_constant op in frs_person_draws, reading a new stochastic entry of take_up_contract.json at the build year (2024-04-01 → 0.593; series 2017–2025 from HMRC's account-activity ODS, Chronicle package pinned by sha). Engine floor ≥2.93, lock 2.94.0.
Five childcare targets (uk-data #473/#474, pe-uk #1830/#1834)	None on the surface.	hmrc.tfc.government_top_up, hmrc.tfc.children_with_used_accounts, dfe.funded_childcare.working_parent_children_2_to_4, dfe.funded_childcare.early_learning_2_year_olds, dfe.funded_childcare.universal_only_children — declared once, resolved by period (Part E): the issue's five numbers at 2024, the FY2025-26 / Jan-2025 counterparts at the 2025 calibration year.
Take-up rates re-fit (uk-data #473)	June rates.	L5: re-fit at the 2024 base on the L2 column twin with the repaired fitter (A6). Fitted on the L2 column twin with an expected-count objective (see Deviations): extended 0.812 → 0.6054, universal 0.563 → 0.4539 landed; TFC (0.997, at its ceiling: with spend routed at 0.593 the two HMRC rows conflict) and targeted (1.0, a ~157-child eligible base) held at 0.88 / 0.597 for María's ruling (plan A22/A23).
DfT bus targets (#789, Chronicle #202/#231)	Facts absent from the pinned feed.	dft.bus_fare_receipts.england and dft.bus_net_support.england active at household grain; London / outside-London / UK rows declared with model-side filters and signed-excluded until the regional measure lands (target_reference_signed_exclusions.json).
Chronicle feed pin	226358e7 feed (v1 schema); 12 references broken on the current feed.	Re-pinned to chronicle main 6fb700e (128 717 rows, facts sha 6ae49d7d…, manifest sha dcda51d6…, v2 schema via #849); contract repaired by declaration only (Part A / plan §5c).

Contract repairs carried by the re-pin (I0)

The previous feed pin (226358e7, schema v1) broke 12 references on the current Chronicle main and moved 101 values. The repair is declaration-only: UC household rows re-bound to the dwp.uc_households family-type/children/youngest-child series with dimensions (exact key set) + dimension_values (subset) selectors and 2025 calendar-year averages (April–December, the months the feed carries); the caseload row keeps DWP deductions Table 1 total_units with a tightened selector (A16; chronicle#247 filed for the Stat-Xplore family-type Total row); the two CGT totals bound to Table 1's Total line (Table 5 duplicates excluded by source_measure_id); the stale PIPR deferral retired and the available local rent rows activated on the same temporal basis (A18); the runner now requires --ledger-manifest-sha256 and refuses a feed that is not the committed pin unless --allow-unpinned-feed is recorded. One I0 defect surfaced by the L4-pre calibration is fixed in this PR's second commit: the Scotland youngest-child row keeps its household-grain binding (the measure provider has no categorical household-to-benefit-unit broadcast). Receipts Part A; docs/uk-chronicle-feed-repin.md.

Surfaces moved

Lane A: assign_period_constant operation kind (schema branch, allow-list, sources.yaml + byte-lockstep source_stages.json); frs_person_draws artifact pin of the take-up contract (its sha folds into the node key) and build-year lockstep check; take_up_contract.json (new stochastic entry; hours entry frozen_by_adjudication; four programs entries with corrected citations, two re-fitted with fitting_receipt); loader validation of entity; H2 parity fixture regenerated (twice: the column, then the contract sha); coverage manifest; degenerate_reviewed_exclusions.json (+1); spine_swap_signed_differences.json (net-new column; re-fit mechanism_change with L5 magnitudes); contract digests; roster pins. Lane B: uk_population_targets.json (+7 active rows, +6 signed-held bus rows, Scotland row grain); difference value operation with value_operands; expected_member_count cardinality guard; observation_basis; entity_name/period_type/period_value/layout_groupby_value_id selector keys; target_reference_signed_exclusions.json (new register, declared in the country package); regenerated national/local references, membership and compile-parity receipts (415 active / 7 no fact / 8 signed on feed 6ae49d7d…); test fixtures from Chronicle rows only (uk_bus05i_fy2025.json, feed rows appended). Fitter: tools/fit_uk_childcare_takeup.py rewritten (tool version 3). Engine: policyengine-uk>=2.93, lock 2.94.0.

Deviations from the plan

• Fitter objective (plan D6 said identity-keyed draws). The build keys each take-up flag on the spine's int64 benunit id at the frs_take_up stage and later stages clone and re-key rows, so no re-drawn realisation can sit on the persisted one (66% flag agreement re-hashing final ids; chance 51%), and the draw objective is a step function whose 1e-2 finite-difference step flipped no unit on the targeted row's 157-child base (first fit left it untouched). The fitter now takes two engine runs with the flags forced on and fits each row's expectation (linear in its own rate, bilinear with the extended rate through the engine's mutual exclusions): smooth, exact for the build, ceilings reported. Plan A24; receipts Part C L5.
• Two rates held rather than landed (plan A6 said the four land). TFC and targeted come out at their ceilings (0.997 / 1.0); landing 1.0 would make the persisted flag constant (degenerate-gate refusal) and assert a 100% take-up the evidence does not support. Held at 0.88 / 0.597 with the full fitted vector in the receipts; María's rulings A22/A23.
• A21 (CGT vintage after the re-pin): ruled 2026-09-05 — facts bound as published, the OBR miss signed on the stack register under UK CGT: translate the 2024-25 gains outturn to the 2025 calibration year (forestalling year; OBR CGT +44.2% signed on the stack register) #875 (window to 2026-10-05); the with-children deferral clock (2026-09-30) is unchanged.
• Codex implemented I0–I2 with zero deviations from the plan text; the two I0 row defects above were found by the licensed runs and fixed by hand.

Verification

• Hermetic funnel on the I1+I2 tree (María's UK + machinery set, test_contract.py, the graph package incl. test_h2_uk_spine_parity), JUnit-counted from this branch's venv with policyengine-uk 2.94.0: 2 893 tests, 2 failures, 27 skipped — the two failures are the documented licensed-cache regeneration tests (test_cached_reference_regeneration…, test_cached_candidate_regeneration…: committed evidence minted under 2.89.0 vs the installed engine, Make the UK eFRS parity re-pin one reviewed command; hold every identity mirror in lockstep (companion to #747) #749's lane), failing on main the same way. Post-port targeted set 468 passed / 1 skipped. After the L5 landing: graph package + digest-sensitive set 463 tests, the same single licensed-cache failure. ruff check clean on every touched file; tools/ci_test_groups.py --verify ok; tools/build_uk_release_input_coverage_manifest.py --check ok. CI is authoritative. Merged main c1b83241 (Portable authenticated worker identity: semantic identity plus audit-only aliases for the primary-QRF worker binding (fixes #863) #871, portable worker identity) on 2026-09-05: two conflicts (a root journal kept at main's version; the build tests' conftest keeps both Accept chronicle-era Chronicle artifact schema ids and fact-key domains; CHRONICLE_* env names (chronicle#143 consumer side) #849's Logbook isolation and Portable authenticated worker identity: semantic identity plus audit-only aliases for the primary-QRF worker binding (fixes #863) #871's worker-identity priming), the primary-QRF worker identity's approved uv.lock digest moved to the merged tree's (this branch's engine floor bump changed the lock) with no derived identity re-pin needed (spec-engine identity, legacy adapter, re-emission, field-usage and inventory-coverage suites green), and the national membership report regenerated after the uk-data citation fix. Post-merge set: 1 027 tests on the identity, pin, UK contract, graph and data-contract files with only the documented licensed-cache failure. Full-tree funnel on the L5 tree (every shard's tests, 8 233 tests): 19 failures — 13 US tests that need policyengine-us (this venv carries the uk extra only; the same files fail on a clean origin/main checkout in the same venv), the same two licensed-cache tests, and four of mine, fixed in the last commit: the live-tree guard found the incumbent package name in the new signed-exclusions register (now uk-data#230), and the monetary-binding receipt contract needed the two new reference fields (value_operands, expected_member_count) — test_monetary_targets and the calibrate shard's monetary tests green after the fix.
• Composition branch uk-publication-stack-834 (ca050866, 8c8f60d9, 9b71e0f8, 9ab60650 — register port, required artifact, seam clock, signed CGT deferral — on this branch's head): the register port's 406-test set green (pins, spec bundles, source stages, terminal gates, battery bindings, gate battery, release certification, test_contract.py).

Licensed lane (receipts Parts B–D; raw JSON licensed-side)

• L0 / L1 (before-receipts): spine-n read at 2.94.0 and an engine-only twin at the floor bump: benunit and time_period byte-equal, the engine bump re-draws 37 imputed household columns through its predictors; the five childcare rows read 1.07 / 1.02 / 1.28 / 0.60 / 1.10 at 2024 with the routed share absent.
• L2 (column twin) / L3 (round trip): 28 stages, 15/15 spine gates; the only between-twin difference is the new column (0.593 for every person, read back by the engine) and its consequences — 1 414 TFC families' net income moves, so the WAS/LCFS/ETB imputations re-draw for 654 / 1 196 / 1 027 households (≈80% of them TFC families), the regional property rescale and the energy IPF ripple to every owner / household at median relative 3e-4 – 1e-3; no cash, weight or benefit-unit column moves; column totals within ±0.9%. The routed share moves only the TFC spend row (1.07 → 0.70); the four count rows are identical to three decimals.
• L5 (re-fit): three fits (draw objective at eps 1e-2 and 0.05, then the expected-count objective, converged: loss 0.080); ceilings at a take-up of one — spend 0.81, children 1.14, extended 1.63, targeted 0.79, universal 1.58. Landed extended 0.6054 / universal 0.4539; rebuilt twin realises 0.999 / 1.087 (expectations 0.986 / 1.000); the two moved flags signed (max |Δshare| 0.207 / 0.110), everything else as L2.
• L4 (composition, uk-publication-stack-834): on the rebuilt twin, family_equal as L4-pre: loss 0.352 → 0.0235, 90.0% of 371 targets within 10%, ESS 5 868. The seven new rows all fit within 0.5% (targeted from 0.77× at design weights to 0.997). The pre-existing surface is where L4-pre left it to the second decimal: the same seven uk_target_fit breaches (OBR CGT +44.2% — A21; the four UC with-children cells −29…−39% signed on the stack's register to 2026-09-30; the two state-pension SPI bands +35% / +25%), UC caseload −13.4%. The port neither helps nor harms the Port UK compact source stages and 2024-25 refresh from uk-data #145 residuals. Terminal battery with the ported register (run 5, after María's A21 ruling): the four UC with-children cells and OBR CGT (+44.2%, signed 2026-09-05 → 2026-10-05 under UK CGT: translate the 2024-25 gains outturn to the 2025 calibration year (forestalling year; OBR CGT +44.2% signed on the stack register) #875, the 2024-25 → 2025 translation of the forestalling-year gains) defer under their in-force entries; uk_target_fit still fails on the two state-pension SPI bands (+34.6% / +25.3%) and reports the private-pension 100–150k deferral stale (+24.7%, back inside the bound). Those are the stack's open register questions, not Port post-freeze UK childcare/TFC spine inputs and calibration contract #834 rows.
• Bus (UK bus fare and subsidy consumer targets from Chronicle facts #789, Part D): on the L2 twin at design weights England fares 0.569× and net support 0.712× the FY2025 facts (E9 measured 0.56× / 0.71×); London 0.276×, outside London 0.760× / 0.969× — the Adjudicate the incumbent's four post-calibration aggregate scalers #790 gradient, which is why those rows are declared with filters but signed-held.

UK-specificity note

The routed share, the DfE/HMRC taxonomies and the DfT geographies are UK; assign_period_constant, difference, expected_member_count and observation_basis are generic surface machinery.

Not in this PR

The publication-stack target-fit register itself, including the new CGT entry (lives on uk-publication-stack-834, the composition branch, not for merge on its own); the #875 translation; the expected-count objective's extension to the four would_claim_* build draws (they stay identity-keyed by design); pointer comments on #789/#790 and closing #153 (each on María's go); the London / outside-London bus measure (#790's gradient); chronicle#247 (Stat-Xplore Total row); the CGT year (A21).

🤖 Generated with Claude Code