US county administrative calibration packages

[MaxGhenis]

First sub-state calibration facts through the provenance-classed pipeline: IRS SOI county returns and AGI (administrative, 3,143 counties, values cross-checked against the raw publisher CSV), Census PEP Vintage 2024 county population (census, 3,144 county equivalents), and USDA SNAP FY2025 monthly state caseloads (administrative, 636 publisher-backed facts). Real publisher bytes fetched per the offline fetch manifest and hash-pinned; packages deterministically regenerated by chronicle.sources.admin_packages (fixture markers removed; the 3,000-4,000 county inventory guard enforced); consumer contract valid with content-addressed R2 provenance declared. The fixture masquerade guards now build their fixture inputs in-test, so they hold regardless of the repo db/data state.

Rebased over the wave-3 merges: kept main's header_row preamble machinery in sources/rows.py and swapped in the cp1252-tolerant decode_delimited_text (Census county files are ANSI); merged-bundle pins re-measured from a fresh build_bundle run over the union (155,288 facts).

Note: the live FNS SNAP publisher URL now serves different bytes than the archived vintage - recorded as note_new_vintage_sha256 in FETCH-MANIFEST.json for a future vintage package; archived bytes retained. Sibling of the Eurostat pilot PR #168.

🤖 Generated with Claude Code

[MaxGhenis]

Adversarial review (gpt-5.6-sol, ultra) blocked the first push with three findings, all confirmed against the code and now fixed:

1. Connecticut boundary identity (high). The IRS county file reports Connecticut's nine planning-region county equivalents (09110-09190), which only exist from the 2022 boundary vintage - but the generator stamped every row 2020_census, giving 18 facts a nonexistent boundary identity inside their stable keys. Fixed: the IRS county record set and rows now carry geography_vintage: 2022 (matching the PEP county convention of vintage-year labeling), and the production test pins the Capitol Planning Region facts against raw publisher bytes (492,650 returns; $47,178,097,000 AGI) plus asserts the CT geoid set is exactly the nine planning regions.

2. Permissive _county_fips (medium/high). int()+zfill accepted signed and overwide inputs (("100","001") -> "100001", ("-1","001") -> "-1001"). Now structurally validated: 1-2 digit state, 1-5 digit county, canonical five-digit result; eight malformed shapes covered by parametrized rejection tests. Also added the marker-removal masquerade test: stripping fixture: true from a small synthetic source trips the 3,000-4,000 inventory guard.

3. Broken/integrity-hostile reproduction metadata (medium). generated_by and the fetch workflow referenced the removed ledger.sources.admin_packages module - fixed to chronicle.sources.admin_packages throughout, packages regenerated (zero drift). The county post-download steps no longer instruct replacing established hash pins from freshly downloaded mutable bytes; they now verify against the pin and STOP on mismatch (new bytes = new reviewed artifact). Stale SOL-D2A scaffolding notes removed from the tree.

Full suite green locally after the fixes.