The latency matrix shared one VM across every op, making results
order-dependent (identical code drifted 2-3x across runs), letting the first
op eat one-time costs (isolate creation, bridge snapshot, wasm compilation),
and never recording WHICH sidecar binary was measured.

- Per-op VM: guest/wasm/vmCmd lanes get a fresh, prewarmed, disposed-after VM
  (host lanes never touch a VM and run up front). Ops that need special VM
  options (external listeners' loopback-exempt ports) declare a prepareVm hook.
- Documented warmup contract in prewarmBenchVm: trivial guest exec (isolate +
  snapshot), one native-baseline wasm exec (module compile), one command exec
  for command ops — before the op's own discarded warmup iterations.
  BENCH_COLD=1 skips it explicitly (loudly); BENCH_SHARED_VM=1 keeps the old
  engine as an escape hatch.
- Binary provenance: every run prints and records sidecar path, profile,
  mtime, size in results JSON; run-benchmarks.sh builds release and pins it
  via the new resolveNodeRuntimeSidecarBinary hook.

Order-independence: tcp_echo solo vs in-family now 1.5% apart (was 2-3x).