[gouyt13]

Summary

CI repeats numerical reference builds across consumer jobs and uses duplicate summary jobs. Reduce producer reference builds from 20 to 5 while preserving all 60 consumer integration configurations and OpenMP OFF coverage. Run complete clang-tidy analysis in one job and remove the duplicate FAISS summary; FAISS remains covered by Python validation.

• Consolidate constructor, search-output and SIMD tests while retaining boundary inputs and backend checks.
• Reject empty clang-tidy scopes and test failure propagation. Lightweight consumer-test stubs retain all seven scenarios.
• Use -g1 for sanitizer builds, retaining assertions, source locations, frame pointers and ASan/UBSan checks with less variable/type debug information.
• Replace exhaustive IVF lifecycle combinations under SDE with a 31-to-33-point append/remove/reload test covering L2/IP, sign-only/extra-bit/raw storage and standard/HACC scans. Full combinations remain in native Release and sanitizer suites; both emulated CPU models, numerical references and HNSW checks remain. Log runner CPU information to help diagnose emulation variability.
• Validate CLI-produced indexes and benchmark append results; handle benchmark datasets smaller than 10 points and own factory results in exception tests.

No public API, persisted-format or dependency changes. Main's branch protection has removed the obsolete FAISS summary and added Documentation validation.

Testing

Latest follow-up, with GCC 14 and native optimization disabled:

• Full ASan/UBSan CTest suite: 410 passed, 1 expected skip; separate OpenMP binding test passed.
• Installed sanitizer consumer: 3 passed.
• Native main executable: 371 passed, 1 expected OpenMP binding skip.
• Complete updated SDE filter: Haswell 81 passed / 7 AVX-512 skips; Ice Lake 88 passed.
• Intentional ASan use-after-free and UBSan overflow probes both report source lines with -g1.
• CI automation: 46 passed. C++ formatting, focused clang-tidy, actionlint and git diff checks passed.

Earlier validation of the other changes included 732 Python tests passed / 2 platform-wheel skips, OpenMP ON/OFF consumer integrations, complete 46-translation-unit clang-tidy analysis and focused benchmark checks.

The first PR revision passed all hosted workflows. The latest changes still need hosted CI, including GCC 13 and Windows/macOS/ARM. Local measurements do not establish a hosted end-to-end speedup.