Apply Housing Benefit accommodation disregards from their start date

[anth-volk]

Fixes #2233

Implements approved Housing Benefit plan B2 only: 13 files, no new inputs. The accommodation disregard already exists through merged #2123, but the fiscal-year loader's 30-April sampling gives none of it in 2026-27 despite its 5 October 2026 commencement. This PR retains the seven monetary parameters' actual dates and uses the established fiscal_year_segments helper to calculate complete awards before and after each change.

Calculation

For each date segment, use the full applicable accommodation amount; cap earnings disregards and childcare against remaining earnings/tax credits; test the additional disregard against those full amounts; then apply the income floor, applicable-amount threshold, 65% taper and zero-award floor. Only then combine the completed awards by elapsed fiscal-year days. The baseline 2026-27 calculation has 182 days before commencement and 183 days afterwards. A fractional disregard applied before the nonlinear rules is not equivalent.

Example with annual earnings £1,000, standard disregard £260, applicable amount £500 and eligible rent £10,400: the pre-commencement annual-equivalent award is £10,244; afterwards earnings are fully disregarded and the award is £10,400. Annual entitlement is (10,244 × 182 + 10,400 × 183) / 365 = £10,322.21, not £10,400 from averaging income before applying its threshold.

The calculation reads reformed parameter schedules and clones only the leaves used to construct a temporary flat schedule. It does not change model parameters, parent links, supplied inputs or computed caches. Supplied annual income/disregard/childcare values and replacement formulas retain their override semantics. Tests cover zero parameter reforms, exact dated ranges, one-day reforms, fiscal-year-only reforms, supplied-period attribution, later set_input, replacement formulas and neutralization.

Sources and scope

Government sources are embedded in the parameter metadata and formula references: GB commencement, SI 2026/753 regulation 1, GB amounts and earnings-test amendment, SI 2026/978 regulation 2, NI commencement and amounts, SR 2026/157 regulations 1/2, GB award taper, regulation 71, and childcare earnings limits, regulation 27, with their NI counterparts.

Annual earnings, rent, ages, household composition and legal status remain constant across segments. Other parameters retain the existing annual convention. This is date integration for the accommodation provision, not a weekly household-history model. The annual income/childcare intermediates are not period-specific assessments; entitlement recomputes the period-specific amounts when needed. Pension-age and passported assessments retain their existing treatment. This PR does not change accommodation eligibility, working-age closure, rent valuation, take-up or datasets. C1 child allowances are expressly deferred under #2232.

Dependencies and validation

Stacked on #2229: #2211 → #2225 → #2219 → #2229 → this PR. #1925 owns the fixed childcare caps and should precede the childcare work; this PR does not duplicate those changes. Coordinate with #2131 when rebasing its shared entitlement file. Upstream main advanced while this batch was in progress; its new commits do not change these HB files. Rebase dependent branches as their bases merge.

The 13-file scope uses the existing accommodation YAML file rather than adding the separately named YAML file proposed in the plan. Its old intentional-omission expectation is replaced, and its childcare fixture uses the qualifying-charge input introduced by #2229.

• uv run --no-sync pytest policyengine_uk/tests/test_housing_benefit_accommodation_dates.py -q: 17 passed.
• uv run --no-sync policyengine-core test policyengine_uk/tests/policy/baseline/finance/benefit/family/housing_benefit/applicable_income/housing_benefit_specified_or_temporary_accommodation_disregard.yaml -c policyengine_uk: 49 passed.
• uv run --no-sync make format: formatting and lint passed.
• git diff --check: passed.

Commands ran sequentially in the existing Python 3.13/core 3.32.13 environment, using the implementation worktree on PYTHONPATH. The first test run reproduced eight failures before implementation. A date-only reform fixture was corrected to use an explicit range for a multi-day override, and a separate one-day regression retains that distinction. No full suite, microsimulation, documentation build or CI polling; hosted CI is not confirmed.

axiom: needed