Performance

Each kernel's run costs over the cases: their time, CPU time and memory, and how well they use many cores. A time here is the kernel's own on its shortest path to a B-rep, building the solid and writing it as STEP where it can, and neither meshing nor measuring it (the harness reads the STEP file afterwards, untimed): the runner's start-up is left out, since it is far longer for a Python or .NET runner than a Rust one, and is shown on its own. None of it is graded, and all of it depends on the machine of the last run.