[gouyt13]

Summary

Share the scalar binary-query accumulation loop between its runtime and template entry points, and share QG accumulation/correction between generic types and ISA-specific float estimators. Keep public validation, dispatch, numerical operation order and public signatures intact.

Remove the unused private HNSW query argument and redundant visited-set table size. The internal visited-set object layout shrinks; persisted layouts and stable core API remain unchanged. C++ consumers were rebuilt. No dependency change or performance improvement is claimed.

Testing

• Portable Clang 23 CTest: 406 passed, 1 skipped; the placement test passed separately with OpenMP binding enabled.
• Repository formatting, full configured clang-tidy and installed consumer checks passed; an installed-header double-QG probe also passed.
• Independent review: 9,000 scalar oracle cases and visited lifecycle checks under ASan/UBSan; 2,304 float/double outputs exactly match baseline; all four x86 HNSW backend instantiations compile. Temporary probes are not committed; leak detection was disabled.
• Fresh generic/AVX2/AVX512 estimator compilation matches baseline disassembly and relocations exactly. Runtime and four-bit scalar wrapper code also matches baseline in the tested configuration.

Native ARM, Windows and macOS execution was not rerun. Public wrappers, including the legacy ignored argument, are retained.