Add IRS SOI long-term capital gains target

[MaxGhenis]

Summary

• add the IRS SOI long-term capital gains aggregate target to the legacy enhanced CPS loss matrix
• load the same aggregate target into policy_data.db through etl_national_targets.py
• add IRS Publication 1304 Table 1.4A LTCG-by-AGI-bin targets to policy_data.db
• include the aggregate and AGI-bin LTCG targets in the default calibration target_config.yaml
• cover the loss-matrix helper, target-config inclusion, and DB-build presence with tests

Notes

PR #868 already added AGI-bracket targets for net capital gains, dividends, and taxable interest. The additional source-backed targets here are:

• calibration.gov.irs.soi.long_term_capital_gains, which resolves to about $1.681T for 2026
• Publication 1304 Table 1.4A Net long-term capital gain by AGI bin, available through TY2023

I did not add short_term_capital_gains: the available parameter is PUF-derived "in the absence of a SOI target" and its 2026 value is negative, so it is not a comparable source-backed target.

Local comparison against the latest rebuilt enhanced_cps_2024.h5:

• aggregate positive LTCG: $1.755T vs $1.686T target, 1.04x
• aggregate net/default LTCG sum: $1.645T vs $1.681T uprated parameter, 0.98x
• several AGI bins still differ materially, so the Table 1.4A rows add distributional guardrails even though the aggregate is close

Tests

• uv run ruff check policyengine_us_data/db/etl_national_targets.py policyengine_us_data/db/etl_irs_soi.py policyengine_us_data/storage/calibration_targets/refresh_soi_table_targets.py policyengine_us_data/utils/loss.py tests/unit/calibration/test_loss_targets.py tests/unit/calibration/test_target_config.py tests/unit/test_etl_irs_soi_overlay.py tests/integration/test_database_build.py
• uv run pytest tests/unit/test_etl_national_targets.py tests/unit/calibration/test_loss_targets.py
• uv run pytest tests/unit/test_etl_irs_soi_overlay.py tests/unit/calibration/test_target_config.py
• uv run pytest tests/integration/test_database_build.py::test_national_ltcg_agi_targets_loaded
• subagent review: no blocking issues

[MaxGhenis]

Local check against the latest rebuilt enhanced_cps_2024.h5 suggests this PR should not materially move the current fixed dataset. For 2026, the rebuilt file has LTCG at $1.645T versus the IRS SOI LTCG target at $1.681T (0.978x, -$36.3B / -2.2%).

That target is not raw 2026 SOI; it is the PolicyEngine-US IRS SOI LTCG parameter, with a 2020 SOI base of $1.137T uprated to 2026 via calibration.gov.cbo.income_by_source.net_capital_gain.

So this PR is mainly adding an explicit guardrail to the legacy loss matrix and target DB. Since the latest rebuilt file is already close to the uprated SOI LTCG target, I would not expect much impact from this PR on the current capital-gains aggregates or reform revenue.

[MaxGhenis]

Yes, IRS Publication 1304 Table 1.4A has Net long-term capital gain by AGI bin, available through TY2023. I added those to this PR as national domain targets for adjusted_gross_income,long_term_capital_gains:

• 19 AGI-bin long_term_capital_gains amount targets
• 19 matching tax_unit_count targets for returns with LTCG in each bin

Local comparison against the latest rebuilt enhanced_cps_2024.h5 shows the aggregate positive LTCG total is close overall ($1.755T vs $1.686T target, 1.04x after uprating), but several AGI bins differ materially. So these should act as distributional guardrails rather than a large aggregate correction.