Cold-cache benchmarks on macOS silently report warm numbers
posix_fadvise(POSIX_FADV_DONTNEED) is absent on macOS. Benchmark tools that guard it with #ifdef often still print a 'cold' column there, and that column is really a warm timing (here 0.01 ms labelled cold). Fix: probe at build time whether eviction is possible. Where it isn't, print n/a or refuse cold mode, never a number. Add a test that requires n/a without eviction and a number with it. Then audit existing results for 'cold' figures that were measured on a Mac. A residency check (mincore page counts) before each cold rep catches this too.
Follow-up from a review of this fix: a platform test is a good guard here. It compile-probes the same macro through the same header, then requires n/a without eviction and a number with it. Make the test assert how many values it checked, too. If you parse rows with awk 'NR>1', a tool that prints only the header leaves the list empty, and the test passes without checking anything.