Release reform-validation tail derives UNKNOWN county for records with valid county_fips (blocks manifest)

Reform-validation tail derives county = UNKNOWN for records whose exported artifact carries valid county_fips

The release tool's reform-validation tail (tools/build_us_fiscal_refresh_release.py, the batched zero-variable-reform simulation built from the frame) crashes at the 2025 period with ParameterNotFoundError: 'gov.local.md.flat_rate.UNKNOWN' — even though the exported artifact's MD households all carry valid county_fips. This blocks release_manifest.json (rc=1) after all certification gates + the export-mass gate + the H5 have already been written.

Root cause of the raise itself is a policyengine-us fragility (md_withheld_income_tax has no UNKNOWN-county guard — see cross-linked issue). But the reason an UNKNOWN county appears at all is harness-side: the exported dataset is geographically clean, so the batched reform-validation construction is losing/mis-deriving county for some MD record.

Three-way clean-artifact evidence (Build I sparse, rmloss100)

1. Exported H5 via the tail's own loader (USSingleYearDataset(file_path=…) → Microsimulation, _write_reform_validation line ~4691): all 1,091 MD households map to valid MD counties at both 2024 and 2025; ZERO UNKNOWN counties anywhere in the 57,240 households; md_withheld_income_tax@2025 computes cleanly ($28.13B). The H5 persists valid county_fips for all 24 MD jurisdictions (24001…24510).
2. Frame path is clean: the target-frame checkpoint carries county_fips/state_fips/place_fips; _dataset_from_frame preserves them; the full-frame reform-validation dataset also computes md_withheld with 0 UNKNOWN.
3. Full-frame reforms are clean: neutralizing salt_deduction, interest_deduction, and qualified_business_income_deduction and computing income_tax on the full frame all succeed. The crash appears only in the batched construction (_reform_household_income_tax / _batched_reform_validation_simulate_factory_from_frame, ~lines 2357–2431: _select_households_by_position(base_frame, positions) → _dataset_from_frame(batch_frame) → Microsimulation(dataset=reformed_dataset, reform=reform) at 2025).

So a per-batch frame slice appears to drop or fail to carry county_fips (or its downstream county) for at least one MD household, which then trips the unguarded MD flat-rate lookup.

The live default is more exposed, not less

The current live/default artifact (c2065b64) persists no county_fips at all (only state_fips), so its reform-validation would derive UNKNOWN for every MD household and hit this row immediately if the tail reached it. Build I actually persists correct county geography; earlier runs' tails were interrupted before reaching this row, which is why it had not surfaced.

Traceback (verbatim)

File ".../salt_deduction.py", line 16, in formula
File ".../state_and_local_sales_or_income_tax.py", line 13, in formula
File ".../gov/states/md/tax/income/md_withheld_income_tax.py", line 28, in formula
    flat_rate_withheld = p_local.flat_rate[safe_county] * reduced_agi
...
ParameterNotFoundError: The parameter 'gov.local.md.flat_rate.UNKNOWN' was not found in the 2025-01-01 tax and benefit system.

Workaround (used for Build I)

--skip-reform-validation completes the release and writes release_manifest.json; the reform-validation report is optional and all certification output (gates, export-mass parity, H5) is produced before the tail. The dataset is unaffected.

Suggested fix

Ensure the batched reform-validation frame slices carry county_fips/county end-to-end (so the reformed dataset matches the exported artifact's geography), and/or adopt the policyengine-us UNKNOWN-county guard (cross-linked issue) so the validation report is robust to any dataset lacking county detail (e.g. the current PUF-derived default).

Cross-reference (root-cause formula fragility): https://github.com/PolicyEngine/policyengine-us/issues/8975

Start in tools/build_us_fiscal_refresh_release.py around _reform_household_income_tax and _batched_reform_validation_simulate_factory_from_frame, then inspect _select_households_by_position and _dataset_from_frame. Run the reform-validation tail on the clean Build I frame and compare county_fips/county through each batch slice. Done means the batched validation preserves valid Maryland county geography and completes without an UNKNOWN county lookup.