Read scorecard#183's base-side grid record and apply known gaps

[MaxGhenis]

Merge hold: cos d1613 (security). Until 16 Oct 2026 this branch's history carried a pre-suppression aggregate-impact report (tests/fixtures/uk_metrics/budget_metrics_report.json, blobs c79f1a06 from 3ea0b19 and 6b599fb5 from b424ffb). That report published Disabled Students' Allowance, which comes from fewer than 10 records, unsuppressed. The tip now carries W3's suppressed 1.58.1 report at tool commit 228a05a3e. Max is deciding whether to rewrite the history. Don't merge, force-push or delete anything for it until d1613 is ruled.

Fixes #636

Stacked on #635, which is stacked on #574. It merges after both.

policyengine-scorecard#183 merged as 71cad306. Every replay receipt now carries:

• diff.base_events.<event>, the comparison base's own grid record, taken from the base's verified staging and run manifests;
• the candidate's missing_measure_years and unsupported_measure_years lists.

Our vendored schema and mirrors were closed, so until this lands every new receipt is refused and nothing can qualify.

Change

• Vendored schema: replay_receipt.v1.json from scorecard main at 71cad306. The README records the commit; the request and status schemas are unchanged.
• Mirrors:
  • MeasureYear replaces InertMeasureYear, which remains as an alias, and now requires a non-empty measure_key.
  • The measure-year lists and supported years must have unique items, as the schema says. The supported-years rule was already in the schema but missing from the mirror.
  • EventResult gains the optional grid lists, and Diff gains the optional base_events, of the new type BaseEvent.
• The known-gap test from the d1344 decision record, exactly:
  • A candidate measure-year that is inert, or missing and supported, is a known gap only if base_events.<event> exists and lists the same pair as inert, or as missing and supported.
  • Each known gap is recorded with the base's staging manifest SHA-256.
  • With no entry (a legacy base, or an event the base did not replay), no gap is known.
• Stricter around it:
  • An incomplete event fails unless every gap in it is known. It also fails if its receipt doesn't list its missing measure-years, or lists none.
  • A complete event that lists a gap contradicts itself.
  • A known gap never excuses a £0 total against a material base value, since a base with an inert record can't hold one.
• Policy text: replay_policy.md states the rule. Its digest changes; the policy id stays d1344-replay-regressions-v1, since this implements the agreed rule.

Invariants and tests

• Schema/mirror differential: seven new rejection cases, such as an empty key, a duplicate measure-year, a duplicate year, a base event missing a field, a short digest, or an extra field. One acceptance case shows a valid record round-trips with its digest.
• Known-gap table: inert in both, missing in both, inert here but missing in the base, the base computed it, missing but unsupported in the base, a legacy base, another event's record, another year, and nothing listed.
• Two edge tests: a receipt from before Enhance entity mapping with flexible aggregation methods #183 fails even with base evidence, and a known gap never excuses a material base value.
• The existing property test still holds: the gate fails exactly when a regression is recorded, no flagged row passes, and adding a row never improves the verdict.
• The first real receipt (6.4.1) still fails on its four inert UC measure-years. It predates Enhance entity mapping with flexible aggregation methods #183, so it has no base record.
• Mutation: eight mutants are killed.
• Qualification suites: 444 passed, 52 skipped.

axiom: n/a: qualification CI infrastructure, no policy encoded.

🤖 Generated with Claude Code

[MaxGhenis]

Merged at 00627f3 with a merge commit. Gates: gh pr checks exit 0 (12/12 pass, run against main after #635), mergeable, no change requests. Independent review (GPT-6.1 Sol): round 6 APPROVE at 3decffb and confirmation APPROVE at 656eede of this PR's own commits. Every later commit is a merge of #635 (reviewed separately, 13 rounds) or of main; the merged tree passed 1,724 tests locally before the last push. Replay gate policy: coordinator's call under cos d1344, with independent review agreement.

🤖 Generated with Claude Code