UK compute campaign, run 3: paired runs to size the OBR divergence axes (employee incidence, CY interpolation, announcement baselines)

[vahid-ahmadi]

The divergence registry (#67, approved) carries its biggest components as unsized with sizing recipes — each recipe is a paired managed run. This issue executes them, converting the employer-NICs picture (observed gap −£7.36bn, like-for-like disagreement −£12.67bn) from labelled-honest to decomposed.

The paired runs

• employee_incidence_and_base (the dominant driver): re-run the employer-NICs measures with the incidence parameter at OBR-comparable settings vs PE's employee_incidence = 1 construction; the delta between paired runs sizes the component. This is the −£8.5bn income-tax-head question.
• cy_proxies_fy: compute each measure at both adjacent CYs and interpolate to the FY — sizes the axis exactly, per the registry recipe.
• baseline_vintage: re-run against announcement-time baselines where the Baselines registry + descriptive-register gate (issues #13, #9) #71 baselines registry (53 registered worlds) carries the leg; where it does not, register the world first — the registry exists precisely so these runs are attributable.
• Re-emit the decomposition artifacts: components re-classified explains_gap/masks_gap with values, divergence_understated updated, residuals shrunk or honestly kept.
• Any residual that survives the sized axes files as a policyengine-uk issue with the Scorecard rows as evidence — the Decompose the OBR costings divergences: PE/OBR 1.40 on the PA/HRT freezes, 1.99 on the additional-rate cut #59 model-development loop.

Depends on #67 merging and shares the managed session with #76/#77-era runs.