perf(http3): one send per response - pure C# h3 now beats nghttp3 at every size

[MDA2AV]

ioxide.http3 lost to ioxide.nghttp3 on small responses and won on large ones, so the gap was a fixed per-request cost. Profiling it meant finding what every response repeats regardless of body size.

Two suspects, one answer

Caching the encoded head — QPACK field section, content-length and frame headers, on a response object a hot handler reuses for every request — was worth 2.8% (371,250 → 381,480 req/s). Kept, but not the cause.

The cause was two SendStream calls per response: one for the ~40-byte header frame, one for the body. Every short response made two trips through the QUIC send path. nghttp3 never did — its writev hands back one buffer carrying both frames, and the shim sends it once. Copying a small body in behind the head to make a single call is far cheaper than the second call.

13-byte body, 2 reactors, 16 conns, 8 streams
  before   371250 req/s   5.37 us cpu/req
  after    897217 req/s   2.22 us cpu/req     2.4x

Result

That was the only size it lost at, and it is now the size it wins by most:

body	nghttp3	pure C#
13 B	580,919	897,217	1.54× (was 0.64×)
50 KiB	38,738	49,727	1.28×
1 MiB	1,986	2,612	1.32×

Bodies over 4 KiB still send separately — copying them would cost more than the second call saves. That threshold is a reasoned guess, not a swept optimum.

Also here

Playground/Http3/Managed — the pure-C# sample these numbers were measured with. The package shipped with no sample at all, only an E2E test, which is why it had never been benchmarked against its native counterpart.

Http3/Nghttp3 gains the bodyBytes knob every other sample of its family already had, so the two can be compared at the same size instead of 13 bytes against 8 KiB.

Verification

Solution builds; Unit 29, E2E 46, Http 35, Tls 15, File 4, Chaos 37 pass. All benchmark cells reported zero failed requests.