Capture the offline parity oracle + decide the country↔grid id system

[Polichinel]

Epic: #31 · Type: spike / decision — first story, gates the rest.

Background

The frozen views-reporting reconciliation groups grid cells by VIEWS country_id, which build_country_to_grids_cache fetches from viewser over the network (views_reporting/metadata/entity_metadata.py). Our GAUL lookup (views_postprocessing/data/gaul_lookup.parquet) carries admin1_gaul0_code — a different country-id system. To prove parity we must group by the same VIEWS country_id, and we want the parity test to run offline (no viewser, no torch in CI).

Work

1. Build a small but realistic cm + pgm sample (several countries, multiple grid cells each, multiple months, ≥1 target, a real-ish sample count) — richer than views-reporting/tests/test_reconciliation_module.py fixtures.
2. Run the untouched views-reporting reconciliation on it with an explicitly injected country→grids mapping (bypass viewser), and freeze (cm, pgm, mapping, reconciled_pgm) to a committed fixture (tests/fixtures/reconciliation_e2e_parity.npz) via a generator script (mirror scripts/gen_reconciliation_parity_fixture.py).
3. Decide & record (short ADR or risk-register note): the frames-native module groups by an injected VIEWS country_id mapping for parity; sourcing the mapping from the GAUL lookup (different id system) is a separate later decision.

Acceptance criteria

• Committed fixture with cm, pgm, the injected mapping, and the oracle's reconciled pgm output.
• Generator needs torch + views-reporting only at generation time; the fixture is consumed offline.
• The id-system decision is written down (ADR/register) and linked from the epic Epic: Migrate forecast reconciliation into views-postprocessing (frames-native, parity-proven) #31.

Validation

• Fixture loads with numpy alone; a smoke check confirms the slice-1 reconcile_proportional reproduces per-country slices of the oracle output.

Dependencies

First story — none.