Period-aware transformation needed for SOI EITC-by-AGI calibration targets

Problem

The 2024 Populace calibration appears to include IRS SOI TY2022 EITC-by-AGI/qualifying-child targets as hard 2024 calibration targets without re-expressing the source-period target surface for the 2024 target period.

One concrete bad target from populace-us-2024-incumbent-improved-996401a-20260618:

• irs_soi.ty2022.table_2_5.eitc_by_agi_children.one_qualifying_child.under_1.eitc_total@2024
• Source period: 2022
• Target period: 2024
• Target: $62.88M
• Final estimate: $407.67M
• Final relative error: +548.3%

A stronger signal is the sibling one-child 50k_plus rows:

• TY2022 Ledger target for one-child 50k_plus EITC returns/amount is 0
• 2024 model estimate is about 918.7K returns and $274.7M

This is consistent with nominal 2022 AGI bins and 2022-law/source distributions being evaluated against a 2024 PolicyEngine simulation, where EITC thresholds and filing behavior are period-sensitive.

Why this looks like a Populus target-compilation issue

Current code appears to carry period metadata and literal source constraints through to materialization, but not transform the target value or bucket thresholds:

• compile_us_fiscal_target_registry chooses the latest eligible Ledger fact at or before the target period.
• SOI references set source_period, target_period, agi_lower_bound, and agi_upper_bound metadata.
• target_spec_from_ledger_reference uses the raw Ledger fact value as TargetSpec.value.
• _agi_bounds copies AGI constraints directly from Ledger as float strings.
• The fiscal materializer then applies those literal AGI bounds to the 2024 simulation.

Relevant code paths:

• packages/populace-build/src/populace/build/us/fiscal_targets.py::_soi_reference_from_fact
• packages/populace-build/src/populace/build/us/fiscal_targets.py::_agi_bounds
• packages/populace-build/src/populace/build/ledger_targets.py::target_spec_from_ledger_reference
• tools/build_us_fiscal_refresh_release.py::_materialize_target_frame

The manifest also currently allows only identity value resolution from Ledger facts, so the diagnostics artifact's compiled_target remains the raw source fact value.

Expected behavior

For period-sensitive nominal bucket targets, Populus should not silently calibrate a 2024 model to an untransformed TY2022 target surface.

Possible acceptable fixes:

1. Transform source-period targets to target-period semantics before calibration, including both value and nominal income-bin thresholds where appropriate.
2. Evaluate these targets at the source period instead of the target period.
3. Mark/source-period-specific EITC-by-AGI facts as diagnostics-only unless a declared transformation exists.
4. Add a compile-time gate requiring explicit transformation metadata when source_period != target_period for nominal income-bucketed targets.

Acceptance criteria

• SOI EITC-by-AGI/qualifying-child targets with source_period != target_period have explicit transformation metadata, or are excluded/downgraded from hard calibration.
• Diagnostics artifacts expose enough information to verify the transformation: source period, target period, original fact value, compiled target value, original bucket bounds, compiled bucket bounds, and transformation method.
• The one-child 50k_plus EITC rows are no longer hard-calibrated to a 2022 zero target in a 2024 build unless that is explicitly justified by transformation metadata.
• Add tests covering dynamic SOI EITC target compilation across source and target periods.

Notes

This does not look like a dashboard formatting issue or a simple missing child-count filter. The EITC child-count filter is present in the fiscal materializer, and aggregate one-child EITC return counts are close overall. The largest miss appears concentrated in AGI-bin distribution, which points to period/bucket alignment.