Review feedback: every round measured `before` first, so `after` always got the
second slot and any within-round drift biased the ratio. Alternate which arm
leads across an even round count, and document that the script needs tsx (it
imports cross-platform-path.ts directly, so a bare `node` shebang can't run it).
Numbers are unchanged: 8.1-8.2x on the 25x5000 row across two runs.
Co-authored-by: Orca <help@stably.ai>