[MaxGhenis]

Fixes #703

New generation processes can certify from a local release manifest when Hugging Face is unavailable. POLICYENGINE_<COUNTRY>_DATA_MANIFEST_FILE always selects the configured file, avoiding an outage timeout. With the setting unset, normal fetch, authentication, caching, offline fallback and certification behavior matches main for consistent bundles.

Certification invariants

A local manifest certifies only when all three conditions hold:

1. Its raw bytes match the installed country TRACE TRO's single valid manifest pin, before parsing or revision rewriting. Every definition of composition/1/artifact/data_release_manifest anywhere in the TRO must agree on that SHA-256. Missing, malformed, conflicting or stale pins refuse; no environment digest is accepted.
2. After revision rewriting, its certified/default dataset artifact agrees with the bundle's recorded repo, path, revision and pinned SHA-256. Every overlapping managed artifact must also agree with the bundle's records. Regional artifacts and additive overlays absent from the primary manifest remain permitted.
3. The unchanged model, core, data-build, fingerprint and exact-pair checks pass. Local selection cannot broaden certification.

The artifact-consistency check also applies to fetched manifests. This closes the pre-existing inconsistent-bundle case there without changing consistent bundles. Certification and runtime validation share the existing release-relative diagnostics/evidence path normalization.

Local files are read and reverified on every access, including after a cached fetch. Successful local certification appends _local_manifest to the existing basis. Packaged receipts record optional runtime.data_certification_basis, preserving the published bundle certificate basis and existing digest domains. Unset receipts omit the annotation. The loader's callable facade retains typed cache_clear and cache_info controls.

Validation

3,322 unique test cases passed, 5 skipped across 56 provenance/certification/disclosure modules and focused release/bootstrap checks. All 82 tests in tests/test_local_data_manifest.py and all 12 TRO-pin tests passed. The final affected manifest/certification run passed 280 tests, and the release/bootstrap-focused run passed 189 tests. The broad run's only failure was an old error-message assertion for a missing default artifact; it was updated for the earlier refusal and its complete 40-test module passed on rerun.

Regression coverage includes a direct @graph conflicting hash, an 80-example Hypothesis property over duplicate placement and nesting, and the reviewer's digest-matching manifest at adversary/elsewhere/elsewhere.h5 under an unchanged bundled dataset URI. Local and fetched cases cover artifact identity/hash conflicts, revision rewriting, inherited artifacts and release-relative evidence paths. Existing exact-pair and unset main-snapshot comparisons remain.

make format, make lint, focused mypy --follow-imports=silent, git diff --check, and generated receipt schema checks pass. Each pytest invocation checked load1 ≤110 and available disk ≥55,000,000,000 bytes; no pytest workers were used.

Deployment and downstream adoption

The G0 example exports the setting only after successful download, raw-byte verification and atomic replacement. Fetch and verify the manifest for the exact installed certified bundle whenever that release changes, then start a fresh process and retain its receipt. Release coordination has been removed from permanent public docs.

• Dashboard: refresh vendored tests/fixtures/w4/execution_receipt.v1.schema.json when upgrading; its schema equality test requires the updated optional runtime field.
• Scorecard: pipeline/uk_replay_variants.py:730 preserves the new setting during offline isolation, while :750 requires exact runtime bundle equality. Clear POLICYENGINE_*_DATA_MANIFEST_FILE during that isolation.
• Other repositories were not edited.

No package versions, dependency pins, datasets or certified releases changed. Leave this PR draft for hub review; do not merge automatically.

[MaxGhenis]

UK hub (Claude session, for Max): independent review round 2 (GPT-6.1 Sol) at da2ae8fb: both round-1 blockers are fixed, and no code blocker remains. The only gate left is CI, which hasn't started yet because runners are queued.

• Conflicting TRO pins, whether top-level, deeply nested or malformed, now refuse. A repeated definition with the same digest is allowed.
• The adversary/elsewhere/elsewhere.h5 manifest refuses through both the local and the fetched path, as do revision-only and digest-only changes. Local probes made zero network calls.
• On the fetched path, the reviewer fetched the real UK manifest live (HTTP 200, SHA-256 49f9fa82…, matching the installed pin) with both this PR and main's module. All 11 certification fields match. 17 more consistent cases certify identically. The 4 cases main accepted and this PR refuses all genuinely conflict with the bundle's records.
• The docs export is now conditional on a verified download, the coordination paragraph is gone, mypy is clean, and the schema check passes. 1,862 tests passed.
• Nonblocking follow-up: a certified-only bundle shape (a default in certified_data_artifact with no matching datasets entry) skips the repo_type check (manifest.py:590). Current and generated bundles always have the entry, and managed materialisation already refuses that shape.

Held unmerged on purpose: this ships in the same release as the weekend Microcosm UK cut certification, which will stack on this branch.