[nick2781]

The native Asio connection benchmark previously collected socket state only on its IO worker. A stalled worker and a socket with no consumable bytes could therefore look similar. Add a default-off Unix controller-thread observer that records non-consuming poll/FIONREAD/MSG_PEEK results, bounded socket histories, strand probes and validated Apple TCP information.

Pin descriptors with generation-checked leases and retire them before cancellation; cap storage at 16 sessions and eight samples per session. Include bracketed steady/wall-clock anchors and cancellation timestamps so later local sequence/ACK captures can be aligned without assuming clocks are interchangeable. Reject the option on Windows and keep library transport, workload deadlines and default benchmark behavior unchanged.

Validation at fe83d005b5c318da34938e06a1384f9d4f84a3f9: six observer cases / 45 assertions, existing kernel-sample checks, native self-test, a controlled one-connection timeout/cleanup integration check, and clang-format 18 passed. English and Chinese benchmark instructions are included.

Three bounded local 5,000-connection trials established every connection and ended with 1, 3 and 0 stalled peer pairs. The retained pre-cancellation observations showed no consumable receive-socket bytes; that narrows investigation but does not establish a transport defect, root cause, fix or deployment capacity. Raw logs and measurements remain local. Follow-up sequence/ACK investigation is tracked by #61; this diagnostic PR does not resolve that issue.