[gouyt13]

Summary

Reduce repeated CI work while making example, installed-package, and SIMD coverage explicit. This replaces the closed PR #121 and includes its graph-search fixture optimization.

• Cache graph construction in two search fixtures and use fewer points for wide vectors, preserving search assertions and the large-ID cases. Local ASan/UBSan time for these two tests fell from 432.14 s to 76.12 s (82.4%).
• Compile sanitizer CI with -O1 -g -fno-optimize-sibling-calls in Debug, retaining assertions, ASan/UBSan, leak detection and frame pointers. The original CI test step took 874.83 s; the optimized run passed in 180.17 s (79.4% less test time). Total job time fell from 16m33s to 6m45s (59.2% less), despite build time increasing from 97 s to 191 s. Both runs registered 246 tests with zero failures and the same five expected AVX-512 skips. Locally, all 246 tests passed in 157.51 s, and a same-machine comparison of the two slowest fixtures improved from 280.97 s to 54.87 s (80.5%).
• Keep full portable, platform, and sanitizer C++ suites; use 26 smoke tests for the native Ubuntu build and 14 for each Python source installation. Repaired-wheel jobs retain the full core Python suite.
• Run FAISS comparison tests in one Linux job instead of installing FAISS in every wheel test environment. Reduce repeated large-dimension Python combinations from 36 to 15 while retaining C++ boundary coverage.
• Run C++ example integration tests in the existing Ubuntu build, fail on missing executables there, exclude them from wheel tests, and route changes to the appropriate job.
• Train both clustering methods through the installed CMake package and check results against a scalar L2 reference. Add a small 65,536-dimensional RaBitQKMeans training test for approximate and exact final assignment.
• Reuse the portable test binary under checksum-verified Intel SDE 10.13.1 Haswell and Ice Lake models. Assert and report CPU capabilities so AVX2-only and AVX-512 coverage does not depend on runner hardware. No extra C++ build or SIMD job is added.

No production library code, public APIs, saved formats, or production dependencies change. The intentional clang-tidy exclusion of tests/examples remains; changed C++ files received focused local analysis. Test ownership is documented in tests/README.md.

Testing

• Final portable Release C++ suite: 246/246 passed.
• Final core Python suite after rebuilding/installing the package: 540 passed, 2 platform-specific skips.
• Native C++ smoke selection: 26 passed; source-install Python smoke selection: 14 passed.
• Dedicated FAISS tests: 4 passed; C++ example integration and missing-binary behavior: 18 passed.
• Installed CMake consumer: passed using only installed headers/library.
• Flat-clustering Release tests: 7 passed; new maximum-dimension case passed under ASan/UBSan in 0.51 s.
• Emulated Haswell: 58 passed, 5 expected AVX-512-only skips. Emulated Ice Lake: 63 passed, no skips. Deliberately requesting Ice Lake features under Haswell failed as expected.
• CI routing/release tests: 22 passed. Formatting, Python lint, focused clang-tidy, documentation links, and diff checks passed. Actionlint passed with only its pre-existing unknown windows-2025-vs2026 runner-label diagnostic excluded.

Sanitizer timing comparison used one before/after measurement on an Intel Xeon Gold 6418H, GCC 14.3.0, Linux x86-64, Debug with native tuning off, ASan leak detection and UBSan halt-on-error, serial CTest pinned to CPUs 0–3, and OMP_NUM_THREADS=2; build time excluded. The speedup applies only to the two search tests, not the full CI suite. Emulation establishes correctness coverage, not native hardware performance. The follow-up optimized sanitizer build passed the full 246-test suite locally. Temporary probes confirmed use-after-free, signed-overflow, and assertion failures still fail; all 63 first-party compile commands retain instrumentation and no NDEBUG. The focused before/after timing used the same local machine, CPUs 0–3, OMP_NUM_THREADS=2, serial CTest and sanitizer settings; one measurement per configuration. Full optimized-suite timing used CPUs 24–27. Full repository clang-tidy and remote cross-platform workflows were not rerun locally; the original PR C++ checks passed on GitHub.