[codex] Add SNAP local proxy diagnostics scaffold

[daphnehanse11]

Summary

Refs #240.

This starts the SNAP congressional-district validation path without changing calibration policy. It adds a validation-only snap_local_proxy diagnostics helper that can compare household-level Populace SNAP receipt/support against ACS S2201 district SNAP household estimates, including support and quality flags.

The release builder now recognizes an optional snap_local_proxy.json artifact in the release manifest when one is present, matching the optional diagnostics pattern used for demographics.json and reform_validation.json.

Notes

This PR intentionally does not promote ACS S2201 to a hard target and does not run formula SNAP by congressional district. It provides the testable payload builder and manifest hook so the next slice can wire the actual source/runtime inputs once we choose the cheapest build-time path.

Validation

• uv run pytest packages/populace-build/tests/test_us_snap_local_proxy.py packages/populace-build/tests/test_us_fiscal_refresh_builder.py -q
• uv run ruff check packages/populace-build/src/populace/build/us_runtime/snap_local_proxy.py packages/populace-build/src/populace/build/us_runtime/__init__.py tools/build_us_fiscal_refresh_release.py packages/populace-build/tests/test_us_snap_local_proxy.py packages/populace-build/tests/test_us_fiscal_refresh_builder.py

[daphnehanse11]

Rebased onto current main (post #266/#277/#279/#245 — one conflict in the us_runtime export list, two in the builder import block and main flow), plus two behavior fixes from cross-PR review:

1. Name-based diagnostics join. _state_snap_relative_errors now matches calibration diagnostics to state SNAP specs by {spec.name}@{period} — the same join _critical_target_fit_failures uses — instead of a positional zip(..., strict=False) that silently misattributed every state error after a skipped target. The fixture test now feeds diagnostics out of spec order (plus an unrelated row) to prove the join.

2. CD geoid vocabulary normalization. Household spine geoids are bare 4-digit SSDD (per Carry the block-anchored US geography ladder as spine columns #277), while Census fact geography ids carry summary-level prefixes (5001700US0601). Both sides of the ACS S2201 join now normalize through the same parser the CD distribution loader uses; previously every district would have flagged missing_acs_reference in a real run. The builder also imports CONGRESSIONAL_DISTRICT_GEOID_COLUMN from congressional_district_geography instead of redeclaring it.

Full populace-build suite passes. Remaining review note (not addressed here): _is_snap_local_proxy_fact still identifies S2201 facts by substring haystacks rather than structured source/measure matching — workable for a validation-only artifact, but worth tightening if this ever feeds targets.

Sequencing note: the state_snap_outlier flag only lights up when per-state SNAP targets are active in the registry, which #256 makes mandatory — merging #256 first (or together) gives this diagnostic its state-error channel.