Stop SNAP state saturation from passing release and publish a person-compatible validation

From scorecard diagnosis batch 1, item #9 (classification: pe_gap, confidence high; full memo: PolicyEngine/policyengine-scorecard diagnosis/DIAGNOSES.md).

Divergence

AK has 107,166.399 PE eligible persons, 107,163.924 mapped participant persons, a 99.99769% rate, and a 2.476-person gap; Urban reports 86,000 eligible, 34.2%, and about 57,000 gap. The FY2024 FNS Alaska household target is hit within 6.495 households (0.0207%), yet the person statistic nearly saturates. Fifteen jurisdictions are exactly 100% (AL, AR, DC, HI, ID, IN, LA, MS, MO, ND, OK, RI, TN, UT, WY); AK, SD, and TX bring the >=99.9% count to 18. The release-stage pre-calibration diagnostic marks 30 states saturated but marks AK false, whereas AK's final-weight person rate is 99.998%; this establishes that the current stage diagnostic does not describe the final-weight person result, but post-calibration eligible-unit weights are still needed to identify AK's exact causal path.

Proposed fix

Build P accepts target-at-or-above-eligible SPM-unit weight as saturation, and its stage diagnostic uses pre-calibration weights. AK nevertheless ships as 99.998% of eligible persons although its FY2024 FNS household target is hit within 0.021%; the pre-calibration diagnostic itself marks AK unsaturated. Make unexplained final-weight saturation release-blocking; emit final-calibrated-weight eligible and caseload diagnostics; and add AK/systemic-saturation regression tests. In the scorecard, do not translate household-targeted SPM-unit receipt into an administrative-shaped person numerator by counting every member; add a SNAP assistance-unit/person counterpart or a separate FNS average-month-person validation.

Evidence

• c80b6c4:data/comparison.json, Alaska SNAP: eligible_count is held_out; participation_rate is consumed_as_target through FNS state average-month household counts. A divergence on the consumed quantity is therefore a calibration-machinery finding.
• Build P us_snap_state_take_up.json: the pre-calibration Alaska FY2024 target is 31,319.25 households, eligible weight 33,205.856, caseload 31,109.222, and saturated=false; the diagnostic explicitly uses pre-calibration design weights.
• Build P calibration_diagnostics.json, usda_snap.fy2024.state_average_monthly_households.wro.ak.average_monthly_households@2024: 31,319.25 target versus 31,325.745 final (+0.0207%), so the official household target is internally consistent and hit.
• POP@8828dee:packages/populace-build/src/populace/build/us_runtime/snap_state_take_up.py:186-211 derives target / weighted eligible and clips at one; lines 307-380 define saturation and document pre-calibration weights.
• POP@8828dee:packages/populace-build/src/populace/build/us/source_stages.json:1806 states that saturated states set all eligible units to take-up and that person targets are deliberately absent because SPM-unit membership overcounts SNAP assistance-unit participants.
• pipeline/compute_counterparts.py:300-306 nevertheless maps SPM-unit receipt and eligibility to every person for the scorecard.
• c80b6c4:data/comparison.json, state SNAP total rows: 15 jurisdictions are exactly 100%; 18 are at least 99.9% after adding AK, SD, and TX.

🤖 Generated with Claude Code

Start with packages/populace-build/src/populace/build/us_runtime/snap_state_take_up.py:186-211 and 307-380 to understand the existing saturation diagnostic and its pre-calibration inputs. Then read packages/populace-build/src/populace/build/us/source_stages.json:1806 and pipeline/compute_counterparts.py:300-306, using the cited Build P and scorecard evidence. Done means final-weight saturation is release-blocking, eligible and caseload diagnostics are emitted, and AK/systemic-saturation and person-compatible validation regressions cover the behavior.