HE-only leg decomposition is structurally zero; SA uplift labels show annual rate as cumulative

Two bugs in the leg-decomposition pipeline, found by tracing how policyengine_uk actually applies the rebalancing reform. Filing to confirm with someone closer to the data before fixing.

Bug 1: reform_he_only is identical to counterfactual

src/uc_rebalancing/pipeline.py:96-117 describes the four sims with this mental model:

"the standard allowance uplift is gated by a parameter formula, while the health element fix is baked into the uc_LCWRA_element input column of the enhanced FRS dataset"

Neither half is accurate. In policyengine_uk/scenarios/uc_reform.py, both legs are applied by the same add_universal_credit_reform simulation modifier via set_input, and both are gated by rebalancing.active(year) (lines 21 and 39 of uc_reform.py). Neither uc_standard_allowance nor uc_LCWRA_element has a parameter formula that branches on rebalancing.active — they read CPI-uprated standard_allowance.amount / elements.disabled.amount respectively.

policyengine_uk/simulation.py:114-116 calls the modifier unconditionally at the end of every Microsimulation.__init__, after parameter changes from reform= are applied. So passing reform={"gov.dwp.universal_credit.rebalancing.active": {...: False}} makes the modifier hit continue and skip both legs.

Net effect on the four sims:

Sim	active	LCWRA set?	SA set?	LCWRA stripped?	Effective
counterfactual	False	no (skipped)	no (skipped)	no-op	both off ✓
reform	True	yes	yes	no	both on ✓
reform_sa_only	True	yes → stripped	yes	yes	SA on, HE off ✓
reform_he_only	False	no (skipped)	no (skipped)	no	both off ✗

So reform_he_only ≡ counterfactual, and every HE-leg field on the dashboard (summary_he, by_decile_he, winners_losers_he) is computed as reform_he_only − counterfactual ≈ 0. The HE column is structurally zero, not a small residual.

reform − counterfactual (total) and reform_sa_only − counterfactual (SA leg) are both correct.

Fix options

• Cleanest: define a local Scenario that runs only the HE half of add_universal_credit_reform. Build reform_he_only with reform={active: False} (so the bundled modifier skips) and attach the HE-only scenario via the scenario= kwarg.
• Mirror the existing trick: leave active=True, then after construction strip uc_standard_allowance for the schedule years via holder.delete_arrays(period=year) — same pattern as _strip_rebalanced_lcwra but on the SA variable. Reverts SA to formula while keeping the modifier-set HE values.

The per-claimant grid HE path (pipeline.py:394-399) uses set_input directly on the merged situation, so the grid HE column is unaffected by this bug.

Bug 2: SA uplift labelled as cumulative, but it's annual

gov.dwp.universal_credit.rebalancing.standard_allowance_uplift:

2026-01-01: 0.023
2027-01-01: 0.031
2028-01-01: 0.040
2029-01-01: 0.048

The modifier compounds these year-over-year: previous_value * (1 + uplift(year)). So cumulative 2025→2029 is 1.023 × 1.031 × 1.040 × 1.048 ≈ 1.150 (~15%).

Sanity check against DWP IA's £725/yr single-25+ uplift by 2029/30: SINGLE_OLD 2025 SA = £400.14/mo = £4,801.68/yr; compound to £5,522/yr → +£721/yr. Matches. The parameter values are annual rates.

But src/uc_rebalancing/scenarios.py labels them as cumulative:

• scenario_label formats FY 2029/30 as "4.8%"
• policy_description says "reaching {final_pct:.1f}% cumulatively by {final_year}" → "reaching 4.8% cumulatively by 2029"

Both should reflect either the annual rate per year ("2.3% for 2026/27, ... 4.8% for 2029/30") or the cumulative figure recomputed from compounding (~15.0% by 2029/30).

Asks

1. Confirm the read of add_universal_credit_reform — are both legs really co-gated by active, or is there a separate input-column path I'm missing in the bundled enhanced FRS dataset that re-supplies uc_LCWRA_element for reform_he_only?
2. Confirm whether standard_allowance_uplift parameter values are meant to be annual rates (as the modifier and DWP £725 figure suggest) or cumulative (as the dashboard currently labels them).

Start by tracing src/uc_rebalancing/pipeline.py:96-117 and 394-399 alongside policyengine_uk/scenarios/uc_reform.py and policyengine_uk/simulation.py:114-116. Verify the four simulation paths and the standard_allowance_uplift interpretation against the scenario labels in src/uc_rebalancing/scenarios.py. Done means the HE-only simulation is distinct from the counterfactual and the SA labels accurately describe annual or compounded rates.