Idle vector-cache refresh validation
The refresh guard retains a published cache only when its snapshot sequence, committed dirty generation, and admission budget match. Empty and budget-limited stores retain the same checks. No persisted data format or public API changed.cargo test --workspace: 3,972 tests and doc tests passed; 12 existing tests ignored.- Workspace Clippy, production-contract Clippy with the production coverage feature set, formatting, and 24 Docker tooling tests passed.
- Focused LLVM coverage exercised all 25 regions in
retain_if_current(44 calls), with no uncovered regions. - All 187 production coverage tests passed. Vector coverage exceeds unchanged thresholds: functions 97.35%, source lines 97.16%, regions 94.56%.
- The overall coverage gate failed: secondary lifecycle lines are 76.97% against
77.81%, and the non-vector uncovered-line fingerprint differs from its baseline
(10,566 lines; SHA-256
cb63a9aa38bfad1bffccdc63c61ff4eb7bedcab9357c7b7fafab9c323b8427a5). This change does not edit those source areas. An equivalent baseline coverage run was not performed, so the cause remains unverified. No thresholds or non-vector baselines were lowered or replaced. - The counted-SST contract verifies unchanged store identity and no additional vector-data reads across three refreshes, including empty and bounded stores.
- The ARM64 server image passed the full packaging/runtime suite. The final MinIO regression uses a metadata-only query to identify catalog ranges, then observes 17 seconds without queries: six catalog GETs and zero vector-data GETs. Search results remain correct before and after a write.
- The same regression fails against a rebuilt image from unmodified
main(2eaa844a96046808b887914679fbe4a49b7adc05), detecting repeated non-catalog ranges.