[gouyt13]

Summary

Some regression tests could pass after the behavior they were meant to protect was broken. Strengthen the unaligned index-build oracle, verify every historical fixture manifest, enforce the tested SHA in release API mocks, and compare distinct concurrent core workloads against serial results on every iteration.

Consolidate duplicate keyword, IVF threading, popcount and XY recombination checks while retaining their effective coverage. The XY estimator test now compares against unsplit encoding. Only nine test files change; production code, APIs, persisted formats and dependencies are unchanged.

Testing

• Portable Clang 23 Release CTest, native optimization OFF: 409 passed, one conditional skip; the dedicated OpenMP binding test passed.
• Current-source Python extension: 793 passed, two platform skips, one mypy test deselected; C++ CLI example tests excluded.
• CI tooling: 48 passed. C++ formatting and Python lint passed.
• Fault injection confirms the strengthened tests reject discarded unaligned data/centroids, incomplete fixture registration, missing release SHA filtering, transient worker factor errors and conversion corruption that the original tests missed.
• Independent review of the complete diff passed, including focused Python/release checks, 20 repetitions of the focused C++ cases and reruns of the fault probes.

Mypy/Pyright are unavailable locally. C++ CLI examples, other platforms, repaired wheels and sanitizers were not rerun. Focused clang-tidy retains eight diagnostics on unchanged baseline test lines; no new diagnostics in the modified core test.