[gouyt13]

Summary

HNSW's random_seed seeds the graph-level generator, but construction currently gives the FHT/Kac rotator a fresh random-device seed. Identical inputs and seeds can therefore produce different serialized indexes and search results even with one construction thread.

Pass the seed's low 32 bits to the rotator. Add C++ L2/IP and Python padded-dimension regressions that compare serialized bytes and search results from repeated serial builds, and check that another seed changes the index. Document the seed scope and the remaining scheduling/toolchain limits.

Newly constructed indexes now use a reproducible rotation. Existing files still restore their saved rotations; public signatures and the saved-index format are unchanged. Parallel insertion order remains scheduling-dependent.

Testing

Validated this branch independently of the NumPy alignment fix and IVF append optimization on Linux x86-64, using GCC 14.3, GoogleTest 1.14.0, Python 3.13, and RABITQ_ENABLE_NATIVE_OPTIMIZATION=OFF:

• Full C++ suite, including historical persistence compatibility: 406 passed, 1 conditional OpenMP skip. The separate OpenMPBinding test passes.
• Built and installed this branch's wheel into an isolated test prefix using the existing Python environment; verified the imported extension bytes against the wheel. Python suite excluding C++ example-binary tests: 700 passed, 2 platform-specific wheel skips.
• The new Python seed regression fails against the retained unmodified-main extension and passes on this branch.
• Full production/bindings clang-tidy 23, C++ formatting, Python format/lint, strict MkDocs build, and git diff --check pass.

No performance improvement is claimed: this changes seed plumbing, not the search kernels. Windows, macOS, and ARM64 execution remain for CI.