perf(cyper-axum): avoid buffered TCP compatibility I/O

[Xerxes-2]

Uses readiness-based TCP I/O on every platform, where hyper writes to the socket directly instead of going through Compio's futures adapter and its intermediate buffer. Accepted sockets reuse their shared descriptor via to_poll_fd(); if that fails the buffered path is kept.

Relies on the vectored-write and half-close impls for PollFd from compio#993, released in compio-runtime 0.12.4. Cargo resolves that by default, but the compio facade still declares compio-runtime = "0.12.3", so the requirement cannot be expressed transitively — see the note below.

Closes #53.

Latency

Measured on Linux. Single connection, tokio-tungstenite client, 40 000 round trips per series (6 000 at 1 MiB).

Payload	server	p50	p90	p99	p99.9	max
64 B	master	6.77 µs	7.38 µs	8.76 µs	11.30 µs	45.8 µs
	this PR	5.99 µs	6.64 µs	7.71 µs	10.34 µs	43.0 µs
	axum	5.96 µs	6.52 µs	7.07 µs	8.43 µs	20.9 µs
16 KiB	master	12.28 µs	13.53 µs	16.88 µs	33.32 µs	53.8 µs
	this PR	9.16 µs	10.15 µs	12.52 µs	21.73 µs	42.1 µs
	axum	7.07 µs	7.60 µs	8.05 µs	10.55 µs	21.0 µs
1 MiB	master	377.96 µs	391.71 µs	427.77 µs	539.86 µs	573.6 µs
	this PR	149.29 µs	154.46 µs	203.44 µs	248.35 µs	286.4 µs
	axum	186.43 µs	190.86 µs	197.92 µs	215.55 µs	367.3 µs

Every percentile improves on master, including max. At 1 MiB the whole distribution moves left and stays ahead of Axum out to the tail. At 64 B the median is now level with Axum; at 16 KiB a gap remains, and in both cases it is concentrated in the tail rather than the median — the same upward bend shows up at every payload size, which looks more like an occasional extra wakeup than per-byte cost. Worth a follow-up.

Concurrency

A separate measurement, so small differences from the table above are run-to-run variance.

Payload × conns	round trips/conn	master p50	this PR p50	axum p50	master MiB/s	this PR MiB/s
64 B × 64	313	265 µs	224 µs	208 µs	29	35
64 B × 256	79	877 µs	304 µs	280 µs	24	28
64 B × 1024	20	396 µs	349 µs	342 µs	19	22
16 KiB × 64	125	494 µs	253 µs	233 µs	4022	7799
16 KiB × 256	32	2.21 ms	316 µs	1.09 ms	3515	6780
16 KiB × 1024	8	20.55 ms	385 µs	353 µs	1389	4325

Medians of 4 alternating runs per build. The buffered path degrades sharply past 256 connections at 16 KiB while this branch stays flat — at 16 KiB × 1024 master measures 20.2–21.5 ms against 374–398 µs here, consistent across every run on both sides. Note this compares two implementations, not two I/O models: master chunks every message through an 8 KiB buffer, and raising that buffer alone recovers much of the difference.

Only the 64-connection rows are safe to read as an axum comparison. Axum reproduces to within 1.00–1.02× there, but swings several-fold at 256, so its 1.09 ms is not meaningful. Two separate reasons: ws_latency splits a fixed round-trip budget across connections, leaving too few samples per connection to support a tail percentile at these depths (8 at 16 KiB × 1024, addressed in #71); and axum keeps swinging at 256 even with 500 samples each, which looks like real behaviour rather than a measurement artefact. The master-versus-branch gap is unaffected either way, since both sides see identical sampling. The compio client saturates around 1024 connections — 17–27 ms for both servers — so those rows say nothing about the server and are omitted.

HTTP

Vectored writes let hyper keep its Queue strategy instead of flattening the body into its own buffer, so plain HTTP benefits too: a 1 MiB keep-alive response body goes from 73.8 µs to 58.4 µs (median of 3 runs).

Benchmark environment

• AMD Ryzen 9 9950X3D (16C/32T), 92 GiB RAM, performance governor
• CachyOS, Linux 7.1.6-1-cachyos x86_64, io_uring enabled
• rustc 1.97.1, release profile
• Client and server on the same host over loopback, each single-threaded
• Both pinned to one CCD (taskset -c 0-7). A mask spanning both CCDs makes round trips bimodal — at 16 KiB, ~9.6 µs when co-resident against ~20.7 µs across the fabric — which flips run-level p50. This affects master and this branch alike.
• Closed loop, one request in flight per connection: service time at a given queue depth, not response time at a target rate.
• Concurrency numbers come from the ws_latency benchmark in bench(cyper-axum): add websocket latency benchmark #69, with CELLS extended to 256 and 1024 connections; the single-connection distributions come from a standalone dumper so every round trip is recorded individually.

[Berrysoft]

Oh, please... don't write response with LLM.

[Xerxes-2]

My bad, Claude Opus is so eager to do this, deleted

[Berrysoft]

Please make all targets use the PollFd, not only Linux and Android.

[Xerxes-2]

Done, feature gate is removed for using PollFd