[gouyt13]

Summary

Fix four edge cases where duplicated code had diverged between backends or callers. Each commit has its own fix and regression test.

• AVX FHT rotation with overlapping buffers. Rotators may be called in place. The generic and NEON backends copy the input with memmove, but the shared AVX2/AVX-512 template used memcpy, which is undefined for overlapping ranges. It now uses memmove. The existing overlap test only ran on aarch64; a new test covers every available backend.
• Uninitialized HNSW. On a default-constructed HierarchicalNSW, search and save dereferenced a null rotator. Both now throw std::logic_error using the initialization check resize already has. save throws before opening the output file.
• Zero-residual xy codes. When the residual norm is zero, quantize_xy_single wrote magnitude codes 2^ex_bits - 1, while the full, split and v1 encoders write 0. It now writes 0 as well. All factors are zero on this path, so distance estimates do not change; only these degenerate code bytes do.
• PiPNN point bound. QuantizedGraph and IVF accept up to kSearchBufferMaxPointCount points, but the PiPNN seed rejected a count equal to it. It now uses the same bound.

No public API, Python API or saved-index format changes.

Testing

• The new FHT, HNSW and xy tests fail before the fixes under ASan/UBSan: memcpy-param-overlap in fht_rotate_avx512, a null-pointer SEGV in HierarchicalNSW::search, and xy code 7 vs. full code 0. They pass after the fixes.
• Full portable CTest (RABITQ_ENABLE_NATIVE_OPTIMIZATION=OFF, --no-tests=error) on an AVX-512 Linux host:
  • Release: 410 passed, 1 skipped.
  • ASan/UBSan Debug: 410 passed, 1 skipped.
  • The skipped test is the placement test, which requires OpenMP binding.
• check-format.sh, full check-tidy.sh build-tidy and git diff --check passed.
• The PiPNN bound has no regression test, because reaching it needs 2^31 points; the existing PiPNN tests pass.
• Not run locally: NEON, Windows and macOS builds.