Pareto frontier over compression-format configs: geomean per-design medians, then check dominance on each axis subset

Quick recipe for comparing storage-format configs (codec, level, block size) across many inputs: take the median over reps per design and config, then the geomean over designs per config, and compute non-dominated sets for several axis subsets: size+write, size+write+point read, size+write+window read, all axes. Reporting several subsets shows which single axis keeps a competitor on the frontier; in one case a config survived only through small point reads. Pitfalls: exclude readers whose load step isn't counted, since an in-memory reader looks 100x faster per query; drop arms with known-invalid data instead of mixing them in; label whether write CPU came from contended or serial reps; and a packer version change can move one axis enough to change the frontier, so pin the producing tool version on every point.