A/B benchmark: PerfEventQueue backends, gate 4 of docs/rust-port-plan.html.

Workload: 16 ordered streams plus ~3% unordered events, pushes and pops
interleaved as PerfEventProcessor does. 2,048,000 events per run, three
runs per backend, one binary.

  cpp   63.0 / 68.9 / 72.0 ns per event
  rust  77.4 / 78.0 / 79.1 ns per event   (~15-20% slower)

Checksums are identical in every run: the two backends pop the events in
exactly the same order. The gap is architectural: three FFI calls per
event lifecycle that cannot inline, plus the slab that keeps events on
the C++ side. Prefetching pops in batches would remove the calls but
breaks the semantics: a push may legally deliver an event older than
anything prefetched, and popped events cannot be pushed back.

Per gate 4, the default therefore stays cpp. The Rust implementation is
correctness-validated (identical checksums, all three test modes, the
sabotage check) and stays in tree behind ORBIT_PERF_MERGE_BACKEND.