Publish a Populace UK local variant with rowwise H5 and long weights #146

Parent: #144.

Why

The dense areas x households matrix approach from policyengine-uk-data is deprecated. Populace now has the primitives to replace it with:

• a row-wise UK H5 where cloned households carry finest available geography, and
• a long local-geography weights sidecar with (area, household, weight) rows and source-household support diagnostics.

Existing Populace foundation

• Build UK geography crosswalk from official sources #142: official-source UK crosswalk for England/Wales, Scotland, and Northern Ireland.
• Add UK rowwise dataset build driver #143: reproducible tools/build_uk_rowwise_dataset.py driver.
• populace.build.uk.local_geography: long weights, sparse stacked matrix, zero-support support summaries.
• populace.build.uk.local_runner: local candidate outputs: local_geography_weights.csv.gz, diagnostics, support summary, solve summary.

Migrated uk-data context

• Add longwise local geography weights policyengine-uk-data#437: bridge longwise sidecar and local H5 migration note.
• Constituency weights produce unrealistic rent estimates for Scotland and Wales policyengine-uk-data#316 and Add guarded US area artifact release path #206: constituency/local estimates, especially rent and Wales/Scotland credibility.
• Calibrate 2024-29 inclusive years policyengine-uk-data#158/Add 7 state-program benchmarks and a federal-EITC-by-state cross-check suite #345/Informed-init L0 selection: warm-start reselection for pool changes (ACS/ASEC ingest forcing function) #346: whether multi-year or pooled FRS years should feed the longer local artifact.

Acceptance criteria

• Publish a variant="local" UK artifact in populace-data alongside the compact fast national artifact.
• Release manifest declares the rowwise H5, official crosswalk, long weights, solve diagnostics, support summary, and hashes.
• PolicyEngine/policyengine.py can consume the longwise sidecar for constituency/local-authority outputs without dense matrix H5s.
• The local artifact can scale in controlled steps, e.g. compact 1x, 2x clone, pooled-year input, then L0 local fit.
• Gates report weakest support by constituency and local authority, including source-household support and effective sample size.