Fix TANF benefit cap across all 38+ states

[MaxGhenis]

Summary

• Cap all state TANF benefit formulas to prevent negative countable income from inflating benefits above the maximum payment standard
• Fixes NC TANF household size bare .sum() bug (summed across entire population instead of within SPM unit)
• Adds min_(benefit, payment_standard) caps to 42 state TANF programs across both standard-named ({st}_tanf) and non-standard-named programs (AK ATAP, AR TEA, CT TFA, FL TCA, IA FIP, LA FITAP, MA TAFDC, MD TCA, MI FIP, MN MFIP, NE ADC, NH FANF, NJ WFNJ, NM Works, OH OWF, RI Works, UT FEP, VT Reach Up, WV Works, WY POWER)
• Adds 7 YAML regression tests for negative income cap behavior

Impact

Metric	Before	After
Total TANF (microsim)	$17.9T	$32.1B
Max household TANF	$12.9M	$25,084
Max per-state TANF	$117T (FL)	$25K (CA)

The remaining gap ($32B vs ~$9B target) is a takeup issue — takes_up_tanf_if_eligible currently defaults to True for all eligible SPM units. Companion PR PolicyEngine/policyengine-us-data#542 adds a 22% TANF takeup rate to the calibration pipeline, which would bring the total to ~$5-7B pre-calibration.

Root cause

All state TANF formulas used max_(payment_standard - countable_income, 0) which correctly prevents benefits when income exceeds the payment standard, but does NOT prevent benefits from exceeding the payment standard when countable income is negative (e.g., from deductions, labor supply responses, or data issues). The fix adds min_(benefit, payment_standard) to cap the result.

Additionally, NC's nc_tanf_household_size used bare .sum() instead of spm_unit.sum(), summing eligible members across the entire population (42,874) instead of within each SPM unit.

Test plan

• 7 new YAML tests for negative income cap (CA, PA, NY, FL, NC, MN, OH)
• All existing TANF tests pass (IL 81, CO 26, DE 91, GA 99, NC 44, IN 49+, KS 50+)
• Microsim max household TANF now $25K (was $12.9M)
• Pre-takeup max per household verified at $25,084 across all states

🤖 Generated with Claude Code

[MaxGhenis]

Replacing with a clean PR that excludes unrelated CBO data files.