Switch to canonical SPM variable names; populate 2026 baseline

[PavelMakarchuk]

Summary

• Switched poverty methodology to canonical policyengine.py variable names: spm_unit_is_in_spm_poverty and spm_unit_is_in_deep_spm_poverty, both mapped to person entity. Formulas are mathematically identical to v1 API's person_in_poverty / in_deep_poverty, but the names match policyengine/outputs/poverty.py (USPovertyType.SPM / SPM_DEEP). Worth noting: the v1 API references person_in_deep_poverty, which no longer exists in policyengine-us 1.690.7.
• Added scripts/compute_local.py so we can populate data/baseline.json without Modal. Falls back to the documented HuggingFace dataset URL convention when policyengine.countries fails to import (editable installs with mismatched manifests).
• Pinned Next.js to 15.5.4 — Next 16 + Turbopack hits an RSC bundler bug ("Could not find the module global-error.js#default in the React Client Manifest").
• Populated baseline.json for 2026 (current-law, no reforms): all 52 regions, zero errors.

Baseline highlights (2026, SPM, weighted by person_weight)

Region	Poverty	Child poverty	Deep poverty	Deep child
US	28.98%	19.02%	13.45%	4.33%
AK	—	9.15% (lowest)	—	—
NE	—	9.17%	—	—
MI	—	9.21%	—	—
NY	—	19.28%	—	—
MA	—	21.10%	—	—
CA	—	25.72% (highest)	—	—

Computed with policyengine 4.3.0 / policyengine-us 1.690.7 / policyengine-us-data 1.17.0.

Observations worth digging into separately

1. Federal (29%) is higher than the state distribution (range 9–26%). The national enhanced_cps_2024.h5 and the per-state states/{XX}.h5 files appear to be calibrated differently — possibly different SPM thresholds, or the state files use state-specific thresholds while the national file uses a single national one. We're using the same datasets the API would route to for each region, so the gap reflects the data, not the methodology.
2. State person_weight totals look low (e.g. AL ~2.2M vs real ~5.1M). The state-file weight columns may not be calibrated to full state population — also a property of the data, not this code.

Test plan

• make dev — page renders 52-row table sorted by child poverty desc
• Versions strip shows the three pinned versions
• "Check latest" (after modal deploy) flags any package that has moved on PyPI
• "Upgrade & recompute" round-trip produces a downloadable JSON that, once committed, looks identical in shape

🤖 Generated with Claude Code

[PavelMakarchuk]

Superseded by #2, which contains these commits plus the deploy and senior-poverty work.