release: 0.16.281

[MDA2AV]

All twelve published packages share one version, as always. The bump is only the twelve Version elements; IoxideRuntime.Version is generated from ioxide.csproj.

Carries everything since 0.16.268:

• Reactor:
  • reactor: the wait's timeout replaces the 250 ms timer, and wakes for a due QUIC deadline #272: the loop's wait is bounded by the next ticker run and the earliest QUIC deadline, so an idle reactor fires a retransmit, ACK or pacing step when it is due instead of at the next 250 ms tick. The timer that drove the tickers is gone.
  • reactor: a ring the kernel refuses reaches OnFault instead of ending the process #273: a ring the kernel refuses (a kernel older than 6.1, too many entries, io_uring disabled by policy) reaches OnFault instead of ending the process, and the message names the cause (reactor: a failed io_uring_setup (kernel older than 6.1) is thrown on the reactor thread and kills the process - the host cannot observe it #270).
• HTTP/2 and HTTP/3: http2, http3: a streamed response learns from FlushAsync that its peer has gone, on all four stacks #274: a streamed response's FlushAsync reports a peer that has gone (FlushResult.IsCompleted), on all four stacks (http2, http3: a streamed response cannot learn that its peer has gone - after RST_STREAM the writer keeps sending DATA, then parks forever #269).
• HTTP/3:
  • http3: a buffered response body goes out under backpressure, not into retention whole #276: a large buffered response on the pure-C# stack goes out under backpressure. Before, 32 one-MiB responses on a connection crossed the send backstop and every such connection was closed.
  • nghttp3: a streamed response at the send-retention high-water waits for acks instead of spinning #277: a streamed nghttp3 response at the send-retention high-water waits for acks. Before, it spun on the reactor thread and nothing completed.
• QUIC:
  • quic: the send-retention limits are QuicOptions, set when building the reactor #275: the send-retention high-water and backstop are QuicOptions, set when the reactor is built.
  • quic: a bare FIN counts as sent only when ngtcp2 wrote it, not on any packet #278: a bare FIN counts as sent only once ngtcp2 has written it. With the window full, an ACK-only packet swallowed it and the stream never ended (about 20 in 1.5M streamed responses).
  • quic: a stream's writes in one engine cycle go out as one STREAM frame #280: a stream's writes within one engine cycle go out as one STREAM frame, so a streamed response's headers, body and end take one packet instead of three: streamed 1 KiB responses 2.7x faster on ioxide.http3 and 1.6x on nghttp3.
• Bench: bench: drive HTTP/3 with h2load built with ngtcp2/nghttp3, not h3x #279: HTTP/3 is driven with h2load, and runs where requests or connections completed nothing are discarded.

Breaking

• Return type: the streamed response writers' FlushAsync returns ValueTask<FlushResult> (was ValueTask) on ioxide.http2, ioxide.nghttp2, ioxide.http3 and ioxide.nghttp3. await writer.FlushAsync() still compiles; code built against 0.16.268 has to be recompiled.
• Removed: the maxSendRetentionBytes parameter of the QuicEngine constructor. Use QuicOptions.SendRetentionBytes; the backstop, which was twice that, is QuicOptions.SendRetentionCeilingBytes.

Native

• libioxide_ngtcp2: ABI 6 (iq_conn_writev added), built with scripts/build-ngtcp2-native.sh from the same pinned refs.

On the merged main's tree (dcf1a62): E2E 250, Unit 63, Http 44, Tls 151, Chaos 47 and File 4 pass. The 7 kTLS tests were skipped locally because the tls module isn't loaded; CI loads it. One Http run ended early in an intermittent test-process crash that also happens on 0.16.268; it is being investigated separately, and the rerun passed.