Fix the capital gains realisation response

[MaxGhenis]

Why

The capital gains realisation response has never affected a costing. Setting gov.simulation.capital_gains_responses.elasticity to any value leaves revenue unchanged, because relative_capital_gains_mtr_change measured the baseline marginal rate on simulation.get_branch("baseline"). Simulations hold their baseline as a separately constructed simulation rather than a branch, so that call clones the reform simulation and both sides of the comparison carry reform parameters. The measured rate change is exactly zero for every reform, for every taxpayer.

Verified on 2.89.3 before the fix — equalising CGT rates with income tax rates on the enhanced FRS returns £36.97bn of CGT at elasticities 0, 1.0, 1.5 and 2.0 alike, with a £0.00bn response in all four runs. Same result via scenario= and reform=.

#1319 reported this symptom and fixed the input plumbing it identified (move_values for capital_gains_before_response, which does now populate). The measurement itself still reported no change, so the symptom survived the fix.

What changed

• Measure the baseline rate in simulation.baseline rather than in a branch of the reform simulation. Extracted to policyengine_uk/utils/capital_gains.py, shared by both change variables.
• Clone the tax-benefit system for the measurement branches. They neutralise capital_gains_behavioural_response to stop the measurement recursing, and cloning keeps that neutralisation off the simulation being measured.
• Define the elasticity against the retention rate rather than the tax rate, and add relative_capital_gains_retention_rate_change. Every published estimate — Agersnap and Zidar, Lavecchia and Tazhitdinova, the CenTax package — is reported against the retention rate, so a published value can now be entered as published. Under the previous convention a positive elasticity raised realisations when rates rose; nothing could depend on that, since no elasticity had any effect.

The default elasticity stays 0, so every existing costing is unchanged.

Effect

Equalising CGT rates with income tax rates (20/40/45), enhanced FRS, 2026:

Elasticity	Gains	CGT	vs baseline	vs static
0 (static)	£89.6bn	£36.97bn	+£15.73bn	—
0.5	£76.9bn	£31.40bn	+£10.17bn	−£5.57bn
1.0	£66.0bn	£26.66bn	+£5.42bn	−£10.31bn
1.5	£56.7bn	£22.62bn	+£1.39bn	−£14.35bn
2.0	£48.7bn	£19.20bn	−£2.04bn	−£17.78bn

Read the shape rather than the levels: the imputed gains distribution has almost no top tail (99 taxpayers above £2m against roughly 5,000 in HMRC's 2023-24 statistics, and none above £5m), so the level of any CGT costing is understated where most of the tax sits. That is a data issue, tracked separately.

Tests

New policyengine_uk/tests/test_capital_gains_responses.py, six situation-based tests that run without dataset access: the rate change registers against the baseline, realisations fall further at higher elasticities, dynamic revenue lands below static, a zero elasticity keeps costings static, a no-reform simulation reports no response, and the measurement leaves the response variable active.

pytest -m "not microsimulation": 128 passed, 15 skipped. pytest -m microsimulation on a local enhanced FRS: 18 passed, 8 failed — an identical failure set to origin/main on the same dataset (salary sacrifice columns and a VAT reform target), so unrelated to this change.

🤖 Generated with Claude Code

[MaxGhenis]

Lint fails here for a reason unrelated to this change: uvx ruff format --check . resolves the current ruff release and four docs pages have drifted out of format on main, so the check is red on every branch cut from it. #1804 fixes that on its own; this PR goes green once it lands.