rk_quic 0.2.1

QUIC and WebTransport for Dart over a native library, so a Dart process can serve browser clients directly instead of waiting for HTTP/3 in the SDK.

0.2.1

• An endpoint bound to an IPv6 address now asks for dual stack explicitly, instead of leaving IPV6_V6ONLY at whatever the operating system defaults to. [::]:4433 used to mean every address on Linux and only the IPv6 half of them on Windows, with nothing on either side saying which — and the missing half is the one a browser reaches first, because Windows resolves a machine name and localhost to IPv6 ahead of IPv4.
• The failure this fixes was silent in both directions. A client on the family the socket did not carry sent packets and heard nothing: Chrome reports QUIC_NETWORK_IDLE_TIMEOUT with num_undecryptable_packets: 0, and a wtransport client simply times out after thirty seconds. curl is no help at all — it answers 200 through an address the browser cannot use, because it picks a family differently.
• tests/dual_stack.rs holds it there: one listener on [::]:0, a real WebTransport client over [::1] and over 127.0.0.1, both required to open a session. Measured failing before the change on Windows 11 (the IPv4 half timed out), passing after.
• No ABI change: generation 2, and QuicServerConfig is untouched. A caller that was passing 0.0.0.0 gets exactly what it did before.

0.2.0

• Bidirectional streams: an exchange can now be a question and its answer, a subscription, or a run reporting progress. C ABI generation 2.
• server.events gains streamOpened, streamData and streamClosed, each carrying a streamId alongside the session. A session is not an exchange: a browser may have several questions in flight at once, and only the stream says which answer belongs to which.
• QuicServer.sendOn(sessionId, streamId, message) writes back into the stream a peer opened without ending it; QuicServer.closeStream() ends it. Which kind of exchange it is belongs to the caller, not to the transport.
• StreamClosed is not an unsubscribe. The endpoint reads a message whole before reporting it, so it learns of a request only once the client has finished writing — and the event therefore arrives immediately behind StreamData, identically for a one-off question and for a subscription meant to last an hour. Treating it as "the client went away" cancels every subscription the moment it is created. A peer that really has gone is learned from a write refused with peerGone, and from sessionClosed.
• sessionClosed now leaves the queue only after the session has been forgotten. The previous order was racy: under concurrent sessions a send could succeed after its own close event had already been read.
• An event field of the wrong type no longer throws. The reader used a cast, which threw instead of yielding null, on the isolate where nothing would have reported it — so one oddly-typed field from a newer library looked like "events stopped arriving" rather than like one odd event.
• C ABI generation is 2: rk_quic_stream_send and rk_quic_stream_close were added. A Dart side written against generation 2 refuses a generation-1 library by name rather than dying on a missing symbol.

0.1.0

The first version with a native part — and a QUIC endpoint that speaks first.

• QuicServer.start() brings up an HTTP/3 endpoint and accepts WebTransport sessions; server.events yields sessionOpened, sessionClosed, datagram and streamMessage; server.send() writes into a session over a stream (ordered, retransmitted) or as a datagram. Implemented on quinn — see the README for why not quiche.
• Everything that blocks — waiting for an event and sending reliably — is on helper isolates: the interface isolate never enters the native part.
• idleTimeoutMs is mandatory and has no "never" value. Measured: without it a client that vanished without warning is not noticed at all — the server had still not seen sessionClosed twenty seconds after the client left.
• The package does not issue certificates: that is rk_pki's job. The chain has one representation across both packages — rustls_pki_types::CertificateDer.

Breaking. rkQuicVersion was a const String and is now a String?, read out of the loaded library. A constant would go on reporting the right version while a year-old library sat next to it; that is what the version was raised for.

• hasNativeTransport is now a real probe: it opens the library and checks the ABI generation. No longer a stub returning false.
• probeNativeLibrary() returns a NativeProbe: loaded, unsupportedPlatform, libraryMissing, symbolMissing, abiMismatch. Not one of these cases throws an exception or brings the process down.
• Statuses cross the boundary by name, not by number; an unfamiliar name becomes RkQuicStatus.unrecognised rather than somebody else's branch.
• The native part builds and reaches the application on Windows, Linux and Android. For macOS and iOS the build files are written but not confirmed by a build — there is no Mac available.
• The package is safe to import in a browser: dart:ffi is behind a conditional import.

0.0.1

• Name claimed. No content yet: the package proves the publishing pipeline, it does not solve the problem.