Measures Box::new+type-erasure (2-160ns) vs full send_to (1.9-7.2us) to
identify the real send latency bottleneck. Finding: allocation is only
1-5% of total cost; the dominant overhead is RwLock address_map lookup
and HybridChannel operations.
Authored by Claude, lovingly guided by Zachery Aaron Shores-Chmielewski
