Add SOI-actuals baseline backtest to reform validation

[PavelMakarchuk]

What

Extends leg-2 (external) validation with a second benchmark family: simulated baseline totals vs published IRS SOI actuals, for tax items that are not calibration targets. Unlike the JCT reform scores, these are realized tax-year data — no behavioral, fiscal-year-timing, or projection caveats.

Nine lines from the TY2023 Individual Complete Report (Pub 1304, rev. March 2026, Tables 3.3 / 1.4), values extracted from the published XLS and cross-checked against the release target set (calibrated national SOI concepts are income lines, EITC, and itemized components — none of these):

line	PE variable	SOI TY2023 actual
Net income tax (minus refundable credits)	income_tax	$2,042.0B
Alternative minimum tax	alternative_minimum_tax	$2.75B
Child & dependent care credit	cdcc	$3.47B
Education credits (nonrefundable)	education_tax_credits	$7.55B
Saver's credit	savers_credit	$2.04B
CTC+ODC (nonrefundable)	non_refundable_ctc	$84.42B
ACTC (refundable)	refundable_ctc	$34.53B
Net investment income tax	net_investment_income_tax	$39.33B
Self-employment tax	self_employment_tax	$81.09B

Scope alignment is documented per line (e.g. PE income_tax also carries retirement-penalty and unreported-payroll amounts, ~0.5%; education_tax_credits = nonrefundable AOTC + LLC, matching SOI's nonrefundable column). Dataset period is 2024 vs TY2023 actuals — a one-year vintage gap, same convention as the ty2023 calibration targets; populace should run slightly above.

How

• soi_baseline_levels.json — curated config with values + citations.
• BaselineLevelSpec + soi_baseline_level_specs(); reform_validation_payload(baseline_levels=…) emits rows (category "IRS SOI actual", window TY2023). All levels share the one baseline simulation — the whole backtest costs one sim.
• Builder passes the default levels; an unsimulated backtest sets out_of_sample_simulated=false, so the publish guard (Refuse to publish releases with skipped out-of-sample reforms #222) covers it.
• Dashboard renders these rows as-is (benchmark/populace/error/Year/Source), no consumer change needed.

Tests

Config loads with unique ids + verified income-tax value; levels share exactly one simulate(None) call; null + guard flag when unsimulated. 21 pass.

🤖 Generated with Claude Code