[MaxGhenis]

Adds seven Census ACS 2020–2024 5-year county source packages, with publisher estimates and margins of error kept as separate facts. The 2024 Census API dataset exists, so no 2023 fallback was needed.

Status: implementation and validation complete; retained as a draft for maintainer review. Three commits pin the publisher artifacts, add exact source-column bound validation, and add the generated packages and county tests.

The supplied 2026-10-24 publisher JSON files were copied unchanged, with SHA-256 pins in seven data manifests and FETCH-MANIFEST.json. No county responses were re-fetched. Source URLs contain no API credentials. R2 fields are content-addressed locators; this PR does not assert that bytes were uploaded.

Table	Source counties, excluding PR	Numeric estimate cells	Numeric MOE cells
B01003	3,144	3,144	130
B11001	3,144	9,432	9,432
B19001	3,144	53,448	53,448
B19013	3,144	3,143	3,143
B19025	3,144	3,144	3,144
B19051	3,144	9,432	9,432
B19061	3,144	3,144	3,144

All seven publisher responses contain 3,222 county rows including Puerto Rico. Full source rows preserve those responses; facts exclude state 72. County GEO_ID and Census NAME are preserved. B11001 includes total, family, and nonfamily households. B19001 includes its total and all 16 brackets, with income bounds in fact constraints and package metadata; the first and last brackets have unbounded endpoints.

Publisher special cases:

• B01003 has 3,014 MOE codes -555555555 with annotation ***** (MOE not appropriate for a controlled estimate). They remain in source rows/cells and unavailable-cell metadata, and are not replaced with zero.
• B19013 suppresses De Baca County, New Mexico (35011): estimate -666666666 / annotation -, MOE -222222222 / annotation **. Those cells are preserved, with no numeric estimate or MOE fact invented. Therefore the table covers 3,144 source counties and 3,143 numeric counties.
• No table failed ingestion. There are 166,760 numeric facts and 3,016 preserved unavailable cells.

Verified properties:

• Each source table covers all 3,144 county-equivalents in the 50 states and DC; Puerto Rico is absent from facts.
• Connecticut uses exactly its nine planning regions 09110–09190, with no legacy county codes.
• Each county’s 16 B19001 brackets sum to its B19001 total.
• B19051 with earnings plus no earnings equals its total; B11001 family plus nonfamily equals its total.
• Every emitted numeric estimate/MOE is nonnegative and equals its exact publisher cell, with row/cell lineage and pinned artifact provenance.
• Facts retain end-year 2024 identity plus explicit 2020-01-01–2024-12-31 survey reference coverage, vintage acs_5_year_2020_2024, and publisher 2024-dollar income definitions.
• No aging, imputation, cross-source reconciliation, consumer target selection, or solver construction is introduced.

The generator reuses the existing full JSON-table parser. The small acceptance-validator change recognizes exact typed lower/upper bounds declared on guarded wide-table columns, matching existing row-bound evidence. It does not change schemas, fact values, identities, or periods. MOE concepts and statistic constraints distinguish them from estimates; metadata explicitly prohibits adding MOEs across categories/counties.

axiom: n/a: source data package

Approved Chronicle agent role: ledger-source-ingestor.

Deterministic checks and tests:

• python -m chronicle.sources.census_acs5_county --check: all seven packages reproduce byte-for-byte.
• County tests: 61 passed, serial, including source-package/core/consumer/agent acceptance, exhaustive numeric cell preservation, county inventories, identities, bounds, suppression, unsupported years, and malformed-source rejection.
• Source-suite tests: 80 passed.
• Bundle regressions: 24 passed; the full merged-bundle test skipped locally because unrelated R2-only artifacts are unavailable.
• Broader seven-file run with -n 3: 260 passed, one unrelated LIHEAP PDF cell-count failure with newer unpinned pypdf 6.21.2. Both LIHEAP cases passed when rerun with the repository’s frozen dependencies (pypdf 6.11.0); no unrelated code was changed.
• Repository-wide Ruff check passed using the lockfile version 0.14.10; diff checks passed.

LLM judge verdicts:

• ledger-source-fidelity: PASS. Independent exhaustive reconstruction matched all 166,760 numeric declarations and all 3,016 unavailable encodings to publisher bytes, hashes, names/GEOIDs, bounds, and period/vintage.
• ledger-boundary: PASS. Independent review includes the source-column bound validator; no computed or consumer-owned values are introduced.
• ledger-contract: n/a; no contract/schema changes.

[MaxGhenis]

Judges at 517ee9d: ledger-source-fidelity PASS, ledger-boundary PASS (Subfleet run 20261023-232551-lwi-j2-chr386). Their one should-fix was that the nine mutation tests couldn't create temp directories in the judge sandbox. I reran tests/test_chronicle_census_acs5_county.py locally with a writable TMPDIR: 65 passed, including all nine mutation tests.