Sizing benchmark tiers by output file size: scale run length from a measured reference run

When a benchmark ladder is defined by dump/file size (e.g. 10/20/50 GB waveform files), do not guess run lengths. Take one measured reference run (cycles, wall s, CPU s, max RSS, output bytes from /usr/bin/time -v plus ls -l) and scale the run-length knob linearly for the larger tiers. Label those as ESTIMATES until the first larger run confirms it, because the per-cycle activity of a workload can drift over a long run. Also check: (1) disk headroom on the host covers the largest file plus every comparison arm's output at once; (2) whether the same file is hard-linked in several places (ls -li / stat link count) before counting it twice or deleting a 'duplicate'; (3) state which format defines the tier (input text dump vs compressed output). A 10 GB text dump and a 10 GB compressed dump are very different workloads, so mixing them silently makes the tier table misleading.