Add 2022 SOI sanity-check test, retire superseded variable totals tests (closes most of #501)

[donboyd5]

Summary

This is the SOI sanity-check piece described in umbrella #501 (PR 2 in that plan), plus the cleanup deletions that depended on it. It introduces a focused test that anchors five 2022 weighted totals to externally published Statistics of Income (SOI) figures, and retires three pre-existing tests whose roles it covers.

The five checks each pass with comfortable headroom against current TMD data (gaps of 0.1–0.5% against a 1% tolerance).

What's added

tests/test_soi_sanity_2022.py — one test function, five assertions, all with 1% relative tolerance:

Check	TMD side	SOI side (tmd/storage/input/soi.csv)	Observed gap
Number of filers	Σ s006 on PUF records	count	+0.5%
Wages	Σ e00200 × s006 on PUF records	employment_income	−0.3%
AGI	Σ c00100 × s006 on PUF records	adjusted_gross_income	+0.1%
Income tax	Σ iitax × s006 (all records)	tottax	+0.35%
Partnership / S-corp net	Σ e26270 × s006 (all records)	partnership_and_s_corp_income minus partnership_and_s_corp_losses	+0.1%

A note on the income-tax comparator

The comparator for iitax is SOI tottax, not income_tax_after_credits, because both TaxCalc iitax (= c09200 - refund) and SOI tottax include NIIT in scope (and AMT, and other additional taxes except SE and Additional Medicare). Comparing against income_tax_after_credits would show a 2.3% gap that is purely a NIIT-inclusion mismatch, not anything to do with TMD.

A note on PUF-only filtering

Three checks restrict to the PUF subsample (data_source == 1) because the SOI "All filers, full population" target is a filer-population concept. CPS-only records in TMD represent the non-filer population. For the iitax and e26270 checks the mask is unnecessary because CPS records contribute zero to both; the simpler unmask form is used there.

What's removed

• tests/test_variable_totals.py + tests/taxdata_variable_totals.yaml — superseded entirely. The old test compared against pre-TY2015 taxdata PUF totals with a 45%-or-$30B tolerance — too loose to catch real regressions. The new SOI sanity test replaces that role with current SOI 2022 figures and a 1% tolerance.
• test_misc.py::test_income_tax (was @pytest.mark.skip) — subsumed by the iitax-vs-tottax check.
• test_misc.py::test_partnership_s_corp_income (was running) — folded into the new SOI test as the e26270 check, with a tightened 1% tolerance and a sourced SOI target replacing the previous unattributed $975 B / 10% tolerance.

What stays in test_misc.py

• test_no_negative_weights — narrow invariant check.
• test_population — Census-anchored population check (distinct external source from SOI; tighter 0.1% tolerance is appropriate).

Test plan

• make format — clean.
• make lint — exit 0.
• pytest tests/test_soi_sanity_2022.py -v — 1 passed in ~13s.
• make test — 59 passed, 2 skipped (the 2 skipped are the existing markers on test_imputed_variable_distribution and test_tax_revenue, both unchanged by this PR; the previous third skip on test_income_tax is gone because that test was retired).

Related

• Umbrella Plan for handling the remaining five skipped tests (follow-on to #430) #501 — most of that plan now lands with this PR. Remaining loose ends from Plan for handling the remaining five skipped tests (follow-on to #430) #501 will be tracked in a separate focused issue (or noted as resolved-by-deletion) before the umbrella can be closed.
• Design discussion: future-year revenue sanity check — how to align tax-calculator iitax / payrolltax with a CBO or Treasury comparator #502 — sister design-discussion issue on future-year revenue comparator (open, independent, awaiting input).
• Add unweighted national-file fingerprint test (closes #503) #504 — PR 1 of the umbrella (unweighted-file fingerprint), already merged.
• Remove tests superseded by #504 and orphan 2021 expected-revenue files (part of #430 / #501 cleanup) #507 — first cleanup PR, already merged.