-Dsimd defaulted to true on the assumption that vectorizing a weighted sum would beat the loop. `zig build bench` says otherwise: on this machine the ISBN-10 checksum measures 2.5 ns/op vectorized against 2.1 scalar, and the ISBN-13 one is level within noise at about 2.4. Nine or twelve lanes of multiply-and-add is not enough work to earn back widening the digits to u16, padding to sixteen lanes, and assembling the vector -- and the scalar loop is nine iterations that a compiler already unrolls. So the default is now false, and the faster path is the one a consumer gets without knowing the option exists. The option itself stays: it is how the two implementations are compared, and a target where the answer is different can turn it back on. Nothing about the SIMD code changes, and both settings keep their test run. The benchmark is unaffected either way, since it calls computeCheckDigitWith directly rather than following the build option. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NJz47vhFE5XcSQUDbfP1N3