Route UK dated reforms through policyengine-uk's statutory pipeline

[MaxGhenis]

Fixes #572

Draft: GPT-6.1 Sol approved 7a1acc0 in round 2. One commit since then (0d4e9b8, below) is under review. The numerical behaviour needs PolicyEngine/policyengine-uk#2305.

What this changes

• Simulation(policy=...) (UK). A policy's parameter values now reach policyengine-uk as an explicit-interval reform dict when the Microsimulation is constructed. Before, simulation_modifier_from_parameter_values edited the already-annualised parameters afterwards. With policyengine-uk#2305, construction writes the changes onto the statutory timeline and reprocesses them as current law: uprating, threshold indexation and 30 April fiscal-year annualisation. That happens before the UC rebalancing inputs and the parameter-gated structural reforms read them. A policy or dynamic with a custom simulation_modifier still runs it after construction, in the same order as before.
• A modifier no longer replaces its object's parameter values. A Policy or Dynamic carrying both (what Policy + Policy gives when one side has a modifier) ran the modifier instead of the values, on main as well. The order is now: policy values, policy modifier, dynamic values, dynamic modifier.
• pe.uk.calculate_household(reform=...). Date keys now get the same reading as Simulation(policy=...): YYYY-MM-DD is an effective date, running until the parameter's next date or 2100-12-31. Before, the dict went straight to policyengine-uk, where a lone date is a one-day override, so {"2027-04-06": 0.21} changed nothing.
• Dict policies accept YYYY-MM-DD.YYYY-MM-DD inclusive intervals, so a finite change can be written as a dict and the sim API can keep stop dates (Policy normalisation drops the stop date of start.stop keys, so bounded reforms never expire policyengine-sim-api#740). They reject bare years, with an explanation: policyengine-uk's dict reads "2027" as fiscal year 2027 only, while policyengine-core's Reform.from_dict reads it as onwards. The US household route keeps its country package's key handling.
• The UK child benefit household test now asserts that the reform changes the result. Before, it only checked that floats came back.

Invariants

• Representation. A UK reform reaches policyengine-uk in the same explicit-interval form from the household route and from the policy route.
• Effective dates. A schedule's order doesn't matter: each date runs until the next date for that parameter.
• Custom modifiers keep their behaviour, and run after their own object's parameter values.

Behaviour by installed policyengine-uk

	pinned 2.125.1	with policyengine-uk#2305
Policy route, January-dated reform	same results as before; both edit the annualised tree	statutory timeline, reprocessed
Household route, {"2027-04-06": 0.21}: income tax change, £30k earner, 2027	£0 (one-day override)	+£174.30
Household route, "2027-04-06.2028-04-05": 2027 / 2028	+£174.30 / £0	+£174.30 / £0

Review

GPT-6.1 Sol round 1 (on 883cfca) asked for three changes. All are fixed in 7a1acc0, and round 2 is running.

1. Custom policy before a parametric dynamic. The dynamic had joined the construction reform, so it ran before the custom policy modifier. It now runs after that modifier, as before. Tests cover all four policy/dynamic combinations on the population route.
2. Merge precedence. merge_reform_dicts now puts override periods after the base's, so a dynamic wins where it overlaps a policy period that has a different key.
3. Unpadded dates. Keys like "2027-4-6", which strptime always accepted, are accepted again.

Round 2 approved 7a1acc0 and noted that an object with both a modifier and parameter values lost the values. That is fixed in 0d4e9b8. test_population_route_applies_policy_before_dynamic now covers eight policy/dynamic combinations.

Tests

• Pinned 2.125.1: tests/test_uk_dated_reforms.py, test_dict_reforms_on_simulation.py, test_household_impact.py, test_parametric_reforms.py, test_extra_variables.py. 86 passed, 3 skipped (needs policyengine-uk#2305); these figures are from before the review fixes.
• After the review fixes, with policyengine-uk#2305 installed: test_uk_dated_reforms.py, test_dict_reforms_on_simulation.py, test_parametric_reforms.py. 51 passed.
• With policyengine-uk#2305 installed: test_uk_dated_reforms.py, test_household_impact.py, test_extra_variables.py, test_uk_analysis.py. 53 passed.
• After 0d4e9b8: test_uk_dated_reforms.py, 17 passed. CI passed at 19ba06b; it runs again on 0d4e9b8.

The Budget reform contract runner that qualification imports is in a separate PR (W1 of the UK Autumn Budget 2026 CI plan).

🤖 Generated with Claude Code