3.5. The per-spawn floor🔗

Every result includes per-spawn floor (harness self-measurement): X ms. This is the harness's own startup cost, measured at the start of each run by spawning a no-op child against the first registered benchmark.

Any data point whose total_nanos is smaller than ~10× the floor is noise. The auto-tuner inside the child usually drives the inner-repeat count up enough that this isn't a problem (each batch targets ~500ms, which is much larger than typical floors), but for very fast operations on small param you may see flat or non-monotone per-call times until the param grows past where startup dominates.