[gouyt13]

Summary

• Reduce clustering copies and assignment work with centroid-buffer swaps, workload-aware batching, and previous-winner pruning.
• Add IVF/SymphonyQG batch search, reuse SymphonyQG worker scratch, and bound IVF encoding scratch.
• Avoid redundant graph zero-initialization and apply size-aware huge-page allocation across all indexes.

Existing search APIs, saved-index formats, and dependencies remain compatible. Profiling and benchmark code/docs are excluded.

Testing

• 292 C++ tests; 700 Python tests (2 platform skips); 284 ASan/UBSan tests with leak checking passed.
• Bidirectional compatibility with main: 44 index variants, identical IDs/distances at 1, 2, and automatic threads.
• Formatting, Ruff, full clang-tidy, strict docs, examples, and installed CMake consumer passed.
• GIST1M construction: 46.9 → 34.0 s median over 3 runs (Xeon 6418H, GCC 14.3, native AVX-512, 96 threads, reproduced single-NUMA allocation stall); recall stayed within the baseline range.

PR CI passed on Linux x86-64/ARM64, Windows, and macOS ARM64, including Python wheels and AVX2/AVX-512 emulation.