[gouyt13]

Summary

Graph construction in two search tests repeatedly reconstructs vectors, adding substantial ASan/UBSan runtime. Enable the existing construction cache in those fixtures and reduce the 1025/4097-dimensional scratch-search cases from 257 to 65 points, retaining a non-complete graph with degree 32 and search window 1.

The 64/65-dimensional cases retain 257 points to cover IDs beyond 255. All dimensions, L2/IP metrics, raw/4-bit/8-bit modes, query counts, hints, ties, and assertions remain covered. Cached-versus-uncached construction parity and maximum-dimension tests are unchanged. No production code, sanitizer flags, APIs, formats, or dependencies change.

Testing

Local ASan/UBSan timings (one measurement per version, build excluded):

Test	Before	After
ScratchSearchMatchesRegularSearch	406.26 s	59.98 s
SingleResultMatchesTopKWithTiesAndHints	25.88 s	16.14 s
Combined	432.14 s	76.12 s

These two tests take 82.4% less time. Measurements used an Intel Xeon Gold 6418H, GCC 14.3.0, Linux x86-64 with AVX2/AVX-512 available and runtime dispatch, Debug builds with native optimization off, ASan leak detection and UBSan halt-on-error enabled, serial CTest pinned to CPUs 0–3, and OMP_NUM_THREADS=2.

• Unmodified sanitizer baseline: 244/244 tests passed (1092.47 s wall time).
• Modified sanitizer tests: 2/2 passed (76.22 s wall time); the full sanitizer suite was not rerun after these test-only edits.
• Release tests: both modified tests and cached-versus-fresh construction parity passed (3/3).
• Formatting, focused clang-tidy for the changed translation unit, and git diff --check passed. Full repository clang-tidy was not rerun.