docs(perf): record the SQPOLL send-ring A/B as a measured negative

[gg582]

Summary

• Records the completed A/B for the audit's Add doxygen style comments #1 ranked opportunity — zero-syscall response write via the SQPOLL send ring — as a measured negative, following the doc's existing convention for the reverted multishot-accept (1.1) and listen-NODELAY (1.3) experiments.
• Without this, the doc's ranked list keeps telling the next reader (human or agent) to chase the biggest "headroom" first.

What was measured (3 interleaved rounds per leg, cwist_bench, base-vs-base noise RPS ±2–4% / P99 ±8% / P99.999 ±25%)

The mechanism worked exactly as designed — strace shows sendmsg dropping from 1.00 → 0.00 per request — but:

leg	RPS	P99.999
c256 base → candidate	352k → 178k (−49.5%)	46.5 → 77.0 ms
c400 base → candidate	386k → 195k (−49.5%)	56.7 → 77.6 ms

Structural cause, same shape as the reverted multishot accept: the synchronous path re-arms the connection for reads immediately after sendmsg returns; the async path pays a full extra reactor round-trip per request (SEND CQE → ring-fd poll wake → reap → re-arm → next read wake) plus ~65% more io_uring_enter calls. Not merged, no knob shipped; code discarded per the meaningful-change-only rule.

The branch-implementation notes (per-worker ring after fork, CQ reader via reactor-registered ring fd, obuf lifetime gated on CQE, -EPIPE semantics, the IORING_SQ_NEED_WAKEUP idle-stall gotcha) are summarized in the doc entry for whoever considers the pipelining-across-the-gap architecture that might actually win.

Test plan

• Doc-only change; no code touched