NICs salary sacrifice cap: carried, scoreable, and not assumption-neutral (#109)

[vahid-ahmadi]

Closes #109. Branches from main — independent of the other UK PRs.

NICs is one of the best-modelled areas of policyengine-uk (class 1/2/3/4, employer and employee legs, salary sacrifice), the Budget's biggest NICs measure is already in the certified baseline, and no scorecard lane touched any of it.

The measure is carried

gov.hmrc.national_insurance.salary_sacrifice_pension_cap:

year	value
2026–2028	uncapped
2029 onward	£2,000

Correct gov.uk references on the parameter. That answers the #99/#101 baseline-integrity question for this measure: carried, with the dated step in the right place.

Checked and not a finding: the step is dated 2029-01-01 while the measure starts April 2029. Every parameter under gov.hmrc.national_insurance uses 01-01 — module-wide convention, not an anomaly. Recorded in the spec so nobody re-raises it as a defect later.

The finding

A default PE-UK score of this measure is not assumption-neutral, and the two assumptions point opposite ways:

parameter	default
salary_sacrifice_broad_base_haircut_rate	0.0016	ON
employee_salary_sacrifice_reduction_rate	0	OFF

salary_sacrifice_broad_base_haircut reduces employment income for all workers — not just sacrificers — on the reasoning that employers spread the added NI cost across the whole workforce. It's gated on the cap being finite (if np.isinf(cap): return 0), so it switches on exactly when the measure bites, and it cites a PolicyEngine research page as its methodology. Meanwhile the employee response — people reducing salary sacrifice once it's capped — is switched off.

So the default carries a PolicyEngine incidence assumption that is neither HMRC's nor OBR's. Published without disclosure, it shows up in a comparison as arithmetic disagreement.

This is #97 sharpened. There, the engine's default asserted no behaviour, so a static leg was honest without qualification. Here the default asserts some.

So the validator enforces disclosure, not just citation

• A leg running the haircut may be publishable, but only while marked disclosure_required — otherwise it publishes an incidence assumption as if it were arithmetic.
• An assumption-free leg must exist and be publishable, or a static official costing has no comparable PE leg to be compared against.
• The sweep is refused if the haircut default ever reads zero, or if the asymmetry disappears. Both would mean this lane's finding no longer holds, and failing beats continuing to assert it.
• The schedule must show both states and must never lapse — a one-state schedule can't evidence "carried".
• No staged value claims: IFS's published assessment and the official costing belong to their originators (Ingest the two harvested-but-unused UK think-tank families: IFS (268) + Resolution Foundation (71) #86/Ingest the two harvested-but-unused UK think-tank families: IFS + Resolution Foundation (#86) #91) and are harvested separately from primary documents.

Three divergence axes are named up front (#59): broad-base incidence, employee response, and employer-NI incidence (PE assumes full employee incidence, 1.0).

Not computed here

Specifies and resolves; computes no counterpart. UK compute runs through #51.

Gates

282 tests (19 new), ruff format --check clean, two build_db runs agree on content_hash, no drift.