Publish v0.2.4

[carpentry-agent]

Last tag: 0.2.3, cut 2026-07-20 (30 days ago). 4 commits on master since.

No CHANGELOG.md in this repo, so this is from the commit log:

• Fixed: binding and send-to now try every address candidate (Try every address candidate when binding and when sending a datagram #14). resolve_address took the first addrinfo getaddrinfo returned and freed the rest, so the passive path and UdpSocket.send-to inherited whatever the resolver happened to order first (TcpStream.connect already looped). Two user-visible failures followed: a UdpSocket bound to 127.0.0.1 and asked to send-to "localhost" resolved the destination without regard for its own family, so where getaddrinfo returns ::1 first, sendto got an AF_INET6 sockaddr on an AF_INET socket and failed with EAFNOSUPPORT; and TcpListener.bind / UdpSocket.bind created a socket of the first candidate's family and gave up if socket() or bind() failed, even when a later candidate would have bound — a host whose /etc/hosts lists ::1 localhost while IPv6 is unavailable, or ::1 already occupied. bind_address now owns socket creation and iterates socket() → setsockopt → bind(), failing only when every candidate does. bind()'s errno is saved across close() so the reported message is still the bind failure, and resolve_address keeps first-candidate behaviour when nothing matches, so numeric hosts are unaffected.
• The examples were rewritten against the current TcpListener/TcpStream API (Rewrite the examples against the current TcpListener/TcpStream API #15) — they had rotted against a since-changed API.

Proposed bump: patch → 0.2.4. A behaviour fix to existing functions; no signature changed and nothing was added or removed. SO_REUSEADDR/SO_REUSEPORT stay TCP-only and IPV6_V6ONLY is deliberately left at the host default, so existing wildcard binds are untouched.

CI is green on both runners at master.

---

Opened by the carpentry-org heartbeat agent (Claude). Veit has not reviewed this yet.