[gouyt13]

Summary

C-contiguous NumPy arrays can still start at an address that is not aligned for float32. A view with a one-byte offset currently reaches misaligned float loads during RaBitQKMeans training.

Copy unaligned two-dimensional inputs into aligned, owned arrays at the shared binding conversion boundary. Already-aligned compatible arrays retain their existing path, and input contents remain unchanged. Add read-only offset-1/2/3 regressions for both clustering classes and all three index wrappers, and document the copy behavior.

No public signature, saved-index format, or dependency changes.

Testing

Validated this branch independently of the HNSW seed 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: 405 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: 714 passed, 2 platform-specific wheel skips.
• The retained unmodified-main extension passes the aligned-input control and fails the offset-1 reproducer with load of misaligned address. This branch passes that reproducer and 15/15 alignment regressions with -fsanitize=alignment -fno-sanitize-recover=alignment on the bindings and included template code. This is a focused alignment check, not a whole-library ASan/UBSan run.
• Full production/bindings clang-tidy 23, C++ formatting, Python format/lint, strict MkDocs build, and git diff --check pass.

This is a correctness fix; no throughput improvement is claimed. Windows, macOS, and ARM64 execution remain for CI.