Builds a python benchmarks/bench_tool_loop.py command for the
tool-loop vs re-prefill measurement. Static, offline, generates
a command only — it never runs anything.
Run the control first — the two arms should be close. Then walk the sweep; the tool-loop claim is that the speedup grows with prefix size and hops. A flat or <1× result is a real negative — keep it (RUNBOOK Step 0).