For agents, on top of README.md, which they read first.

• Rerun this before trusting packages/replay after any change to the core (a jgenesis subtree pull, a new snes-core feature, a toolchain move that touches float maths). A diverged core makes seek, rebuild and the disk cap silently produce a wrong machine; this probe is the only thing that says so.
• Measure here before and after changing a replay knob (RECORD_EVERY, the tier retention constants, the storage cap): its numbers are what those constants were sized by.
• It starts a real zbanks::Bot, so it is one bot per process like everything else; do not add a second pass that starts another bot.
• Output is run/replay-probe/, removed and recreated on every run. Never point it at anything that matters.