Performance methodology¶

The repository includes a deterministic, dependency-light throughput harness. Run it only in release mode:

cargo run --release -p xtask -- bench

Each case warms up first, then records repeated wall-clock samples and reports median, median absolute deviation, minimum, maximum, and MiB/s. Inputs are deterministically generated so runs are reproducible. Representative codec, hash, cipher, compression, Magic, and Scan paths are included.

The harness measures in-process execution. It deliberately excludes CLI startup, filesystem and pipe I/O, allocator profiling, cross-machine comparisons, and claims of statistical significance. Compare results only on the same machine, toolchain, power profile, and build configuration.