Recalibrate SNAP take-up per state to FNS household caseloads

[daphnehanse11]

Implements #372. Stacked on #371 (base branch is snap-caseload-targets) — the take-up calibration consumes the same compiled snap_households target specs that PR adds, so the seed and the weight-calibration objective agree by construction.

Why

#294's national-rate fill bakes in CPS state underreporting: the buildi-sparse release undershoots ten state SNAP benefit targets by 7–43% while a per-record feasibility audit shows every state but California reachable under the 5× weight cap, and state taker-household counts err from −38% to +325% against FNS caseloads (full numbers in #372 and the #371 preview comment). The states that are short cannot be fixed by reweighting — they lack taker households; this stage supplies them.

What

A snap_state_take_up source stage following the medicaid_take_up pattern (#331) at SPM-unit grain, using the shared binary-assignment ops (#338):

1. Anchor — units with reported ASEC receipt (SPM_SNAPSUB > 0) always take up, identical to the national stage's anchor.
2. Fill — non-reporters draw at an in-build state rate (FNS average-monthly household count over weighted modeled-eligible units; an assignment prior, never cited as provenance per the contract's model-relative-ratio rule), then calibrate_binary_assignment greedily calibrates among eligible non-anchored units so each unsaturated state hits its FNS FY2024 count within unit-weight granularity.

The national snap_take_up stage remains as the prior; this stage recomputes the final surface after target compilation (it needs is_snap_eligible from the engine and the compiled targets, mirroring the Medicaid builder placement).

Deliberate carry-overs from the Medicaid stage: off-domain units keep a draw-based propensity (no hard-coded zero reform response), and saturation is recorded rather than failed (CA's ceiling is 1.05× its caseload — an eligibility-undercount symptom, not an assignment bug). Persons caseloads stay untargeted: the SNAP assistance unit is often a subset of the SPM unit, so member counts overcount FNS participants ~50% (see #371's second commit).

Gates and diagnostics

• us_snap_state_take_up_gate (release-blocking): every state must carry an FNS target; anchors must survive calibration in every state (unit-level check); unsaturated states must land on their count within max(2%, one unit weight); positive count with zero eligible weight fails as an eligibility-feed collapse; caseload==eligibility with a reachable floor below it fails as the universal-take-up landmine.
• us_snap_state_take_up.json release artifact + staging telemetry: per-state eligible/anchored/caseload weights, targets, saturation, plus national rollup.
• Take-up contract: SNAP entry stays out_of_scope for Bernoulli seeding (it keeps a sourced national rate, which count_calibrated forbids) with scope_owner updated to name this stage as the final surface owner.

Tests

• New test_us_snap_state_take_up.py (14 tests): manifest declaration, anchor derivation from raw subsidy, state fill-rate prior, exact count calibration in unsaturated states, saturation, anchor survival under downward pressure, empty-target refusal, and gate failure modes (missing state target, eligibility collapse, dropped anchor, count miss, universal-take-up landmine).
• Builder main-path test extended with the new stage's monkeypatches.
• Full populace-build suite passes (517 tests).

🤖 Generated with Claude Code