Retire the legacy UK ETLs onto publisher-backed facts (#135 PR A)

[juaristi22]

Summary

PR A of #135: every hardcoded series in the legacy db/etl_obr.py / db/etl_ons.py dicts is now covered, ported, or registered as an explicit gap — the audit doctrine is: map to a package concept; else find and port the publisher source; else register the gap. No silent drops.

Ported (new facts):

• obr-efo-economy-march-2026 — EFO March 2026 detailed forecast tables, economy: nominal GDP (Table 1.4), ILO unemployment rate 16+ (1.6), CPI + RPI year-on-year (1.7), Bank Rate (market-derived assumptions, 1.9), employment 16+ (1.6). Workbook committed (obr.uk 403s CI runners, same rationale as the receipts/expenditure packages); manifest sha256 + deterministic R2 keys.
• obr-efo-aggregates-march-2026 — EFO March 2026 detailed forecast tables, aggregates: public sector current receipts (6.2), TME (6.1), public sector net borrowing (6.5), PSND as per-cent-of-GDP (6.13).
• ons-families-households-2025 gains Table 5 record sets: households total + average household size (already-committed workbook, no new artifact).

Registered gaps (first-class rows in the checklist audit table; the publisher does not directly assert these in the staged workbooks):

• real_gdp_growth — the economy tables publish real GDP levels (ABMI); deriving growth is consumer work per the facts-only ADR.
• public_sector_net_debt in GBP — sheet 6.13 publishes PSND only as per cent of GDP; converting would be Chronicle-computed.
• FY2024-25 rows for receipts/TME/PSNB — the aggregates sheets start those rows at FY2025-26.
• db/etl_hmrc.py — partial adjacent coverage only; kept, with its series recorded as explicit PR A gaps (its own audit is follow-up).

Deleted: db/etl_obr.py, db/etl_ons.py, their tests, the load obr/load ons CLI branches + choices entries (argparse error beats silent no-op), and the chronicle/targets/loaders.py / chronicle/jurisdictions/uk re-exports. DataSource.OBR/ONS enum members stay (existing DBs carry the strings). Negative-existence + audit-registration assertions added to tests/test_chronicle_boundaries.py. CI's load all surface shrinks by the two removed sources.

Bundle expectations re-measured on the post-#173 base: fact_count 145,295 (+73), source_package_count 122, obr 253, ons 65,646.

Publish-raw note: the two new OBR workbook artifacts carry deterministic R2 URIs in their manifests; the credentialed publish-raw upload is pending (same protocol as prior waves).

Chronicle Governance

• Approved Chronicle agent role: ledger-source-ingestor (new packages + record sets); the db/** deletions are the maintainer cleanup UK enablement tooling: source-inventory de-hardcode, UK consumer-artifact suite, legacy ETL cleanup #135 sanctions.
• Deterministic checks run: source-package validation, source-cell preservation, consumer-contract validation, raw-facts boundary validation — full suite uv run pytest -q.
• LLM judge verdicts:
  • ledger-source-fidelity: pending review (every new fact traces to the committed workbook cells; guard cells transcribed from the sheets)
  • ledger-target-profile: n/a
  • ledger-contract: n/a
  • ledger-boundary: pending review (derivation refusals above are the boundary holding)

Tests

Implemented by Codex under a test-iteration protocol; two recorded deviations (real-GDP-growth portability, aggregates FY2024-25 start) were independently verified against the workbook sheets before shipping. Verification, twice (Codex's run and an independent re-run): ruff clean; full suite 626 passed, 1 skipped (the count drops from 633 with the two deleted ETL test files; package tests +98 targeted assertions).

🤖 Generated with Claude Code

[vahid-ahmadi]

The audit doctrine — every hardcoded series covered, ported, or registered as an explicit gap, with the gaps as first-class checklist rows rather than silence — is the right way to retire an ETL, and it's the same shape as the microcosm side's reviewed-absence pattern: a fail-closed accounting of what didn't move, which is where retirement PRs usually hide their losses. Deleting etl_obr.py/etl_ons.py with their tests (−1,010 lines) against +3,150 lines of publisher-backed package declarations is the trade this repo exists to make.

Committing the EFO workbooks because obr.uk 403s CI runners is consistent with the receipts/expenditure precedent, and the manifest sha256 + deterministic R2 keys keep the committed copies verifiable rather than trusted. The registered gaps I read (e.g. real_gdp_growth, where the publisher asserts real growth only via tables the package doesn't stage) are correctly gaps rather than derivations — deriving them consumer-side from staged nominal/deflator facts is exactly the kind of decision that belongs downstream of the facts boundary, per the #166 ruling.

The targeted suites pass for me (172 tests across the consumer, source-package, alias-drift, boundaries, and source-file suites). One process note rather than a finding: this wave re-measured the bundle constants by hand again — the cost #157 exists to remove. Since wave 4 (NZ) is next and this PR's own churn is the freshest evidence, I'd sequence the #157 per-source expectations refactor before the NZ waves start rather than after; the proposal has been endorsed since the 17th and every wave landed without it pays the ~13-minute re-measurement tax again. Nothing blocking here.

[juaristi22]

Codex adversarial review completed for this PR. It found no concerning or material findings: the retired UK ETL paths are no longer referenced at runtime, the publisher-backed package coordinates match the committed artifacts, and the documented parity gaps are explicit Chronicle-boundary exclusions. The review was read-only; its sandbox could not run the Python test suite because the environment required writable dependency state.