Post-0.5.0 review pass. The crate's public surface stays source-compatible — every fix below tightens existing semantics (correctness, performance, robustness against kernel edge cases) without breaking the API. The minor bump rather than a patch reflects the breadth of behavioral changes (Linux RPDB-aware best-local selection, lazy nexthop dump, BSD source-route filtering, Windows forwarding-table walk, etc.) that callers exercising edge cases may notice.

• Lazy RTM_GETNEXTHOP dump in netlink_walk_routes, netlink_best_local_addrs_into, and rt_generic_addrs (the gateway walker). The dump only runs if the route walk actually encounters an RTA_NH_ID reference. Hosts without ip nexthop-managed routes (the common case) skip the round-trip entirely; transient EINTR / NLM_F_DUMP_INTR from unrelated nexthop-subsystem churn no longer fails ordinary route_table() / best_local_*() / gateway_addrs*() calls.
• Per-family route walk — route_table_by_filter issues separate AF_INET and AF_INET6 dumps instead of a single AF_UNSPEC request. Some kernel versions return only IPv4 for AF_UNSPEC route dumps; the per-family walk avoids silent IPv6 loss on dual-stack hosts.
• RT_TABLE_DEFAULT (253) accepted alongside RT_TABLE_MAIN (254) and RT_TABLE_LOCAL (255). Hosts with a fallback default route installed via ip route add default ... table default now surface it correctly.
• RPDB precedence in best-local selection — candidate key is now (table_rank, metric, pref_rank) lexicographic. Lower table rank wins regardless of metric (matching the kernel's rule chain local < main < default); within a table, lower metric wins; RFC 4191 router preference (RTA_PREF) breaks equal-metric IPv6 ties with HIGH < MEDIUM < LOW.
• ECMP correctness — equal-key default candidates now extend best_oifs instead of replacing it, so addresses from every winning interface surface in best_local_*.
• RTA_VIA cross-family gateways are dropped, documented — IpRoute only models same-family (dest, gateway) pairs, so a route whose next-hop family differs from the destination family used to silently emit as gateway = None (on-link). They now skip cleanly; the omission is noted in route_table's rustdoc.
• Source-prefix and policy-table filtering — routes with rtm_src_len != 0 or RTA_SRC set, plus routes from custom policy tables, are filtered out. route_table / best_local_* rustdoc carries a "best-effort" caveat for policy-routed hosts.
• Nexthop-group missing-member handling — resolve_nh_id now returns None when a group leaf is missing from the nexthop snapshot (triggering the deferred-retry → EINTR path), instead of silently dropping the leg.
• Malformed-attribute defence — dump_nexthops filters nexthops with malformed NHA_GATEWAY payloads instead of emitting them as gw = None. Same defence for RTA_DST in best-local selection.
• Receive-buffer size — route / nexthop walkers now use a 32 KiB buffer (matching iproute2) instead of one OS page. Routes with large RTA_MULTIPATH ECMP lists or deep NHA_GROUP payloads no longer silently truncate.
• NLM_F_DUMP_INTR / family-unavailable handling in rt_generic_addrs — the gateway walker now mirrors the route walker. Interrupted snapshots return EINTR; unsupported-family errors collapse to an empty result.
• Big-endian byte order — rt_generic_addrs's IPv4 gateway decode uses Ipv4Addr::from([u8; 4]) instead of u32::from_ne_bytes(...).swap_bytes(). The previous form happened to work on little-endian Linux but produced byte-reversed addresses on big-endian targets (powerpc64, riscv64gc, etc.) that recently joined the cross-target CI matrix.

• fetch() truncation fix — the sysctl(NET_RT_*) wrapper truncates the buffer to the kernel-written length before parsing. NetBSD and OpenBSD NET_RT_IFLIST dumps no longer surface as Err(InvalidData "invalid message") on the trailing zero-padding.
• Source-specific routes filtered — NetBSD's RTF_SRC flag and OpenBSD's RTAX_SRC / RTAX_SRCMASK slot bits are respected.
• OpenBSD rtm_priority for best-local — used as the best-default selection key. Other BSDs have no documented per-route priority; on those targets every default route ties and addresses from all default-route interfaces are returned (documented as best-effort).
• Per-family union APIs — best_local_addrs / route_table_by_filter walk AF_INET and AF_INET6 separately so single-stack hosts don't lose the populated family on EAFNOSUPPORT / EPROTONOSUPPORT / EOPNOTSUPP.
• End-of-stream sentinel — walk_route_table, best_local_addrs_in, and rt_generic_addrs_in treat a zero-length record header as end-of-stream padding instead of an error.
• Defensive netmask parsing — interface_addr_table_into skips individual addresses with non-canonical RTAX_NETMASK instead of failing the whole walk (helps point-to-point / tunnel interfaces that emit peer-address bytes in the mask slot).
• Big-endian IPv4 gateway fix — same swap_bytes() byte-order bug as the Linux walker, fixed identically.
• compat::RtMsghdr / RtMetrics layout assertions — every BSD RtMsghdr and per-platform rt_metrics struct now has compile-time size_of and offset_of! checks against the kernel ABI for every field we read. Catches version-skew silently breaking selection on a future libc bump.
• DragonFly multicast — multicast_addrs() returns Err(ErrorKind::Unsupported) instead of pretending success with an empty list (the DragonFly kernel has no NET_RT_IFMALIST).

• route_table() support — route_table_by_filter walks GetIpForwardTable2 and emits IpRoute entries. Filters expired rows (ValidLifetime == 0), multicast / broadcast destinations, loopback rows, and per-subnet directed-broadcast /32 housekeeping rows.
• Interface-keyed broadcast filter — directed-broadcast suppression set is HashSet<(InterfaceIndex, Ipv4Addr)> derived from GetUnicastIpAddressTable. Multihomed hosts with a legitimate /32 host route to an address coincidentally equal to another adapter's directed broadcast no longer get dropped. RFC 3021 /31 prefixes excluded; suppression only applies to PrefixLength == 32 && gw.is_none() rows.
• best_local_* via documented forwarding-table walk — replaces the prior undocumented GetBestRoute2(NULL, 0, ...) / GetBestInterfaceEx(unspec) calls. Walks GetIpForwardTable2 for PrefixLength == 0 && ValidLifetime > 0 && !Loopback rows, joins per-interface Connected + Metric from GetIpInterfaceTable, picks smallest effective metric (route metric + interface metric).
• Equal-cost defaults preserved — best_default_route_interface returns SmallVec<u32> instead of Option<u32>. Multi-homed hosts with equal-cost defaults across two adapters now surface addresses from every winning interface (matching Linux/BSD).
• Connected = FALSE filter — stale defaults pinned to a disconnected VPN / unplugged NIC can no longer win the metric race.

• Bulk path with per-interface fallback — get_ip_mtu / get_ipv4_mtu / get_ipv6_mtu first try a single interface_addrs() dump, then fall back to per-interface iteration if the bulk dump fails. One unrelated malformed kernel message no longer aborts the whole lookup.

• New route benchmark suite (benches/route.rs) covering route_ipv4_table, route_ipv6_table, route_table, and route_table_by_filter(default_only) on Linux / macOS / Windows.
• Linux gateway operations regained their pre-branch latency after the lazy nexthop fix:
  • gateway_ipv4_addrs: 19.9 µs (was 30.4 µs mid-branch, 18.4 µs at 0.5.0)
  • gateway_ipv6_addrs: 16.1 µs (was 26.7 µs mid-branch, 14.3 µs at 0.5.0)
  • gateway_addrs: 24.1 µs (was 34.4 µs mid-branch, 22.4 µs at 0.5.0)

• Cross-target compile matrix now covers Linux glibc/musl across aarch64 / i686 / powerpc64 / riscv64gc, FreeBSD x86_64 / i686, NetBSD x86_64, and Windows MinGW / MSVC i686 / x86_64.
• Runtime BSD VMs — test-freebsd, test-netbsd, test-openbsd, test-dragonflybsd jobs run inside vmactions/*-vm@v1. NetBSD and DragonFly use cargo test --lib (doctests + integration tests skipped); per-test cfg-gates in tests/interfaces.rs and tests/filter_variants.rs document the platform-specific gaps.
• Coverage on all four BSDs — coverage-freebsd, coverage-netbsd, coverage-openbsd, coverage-dragonflybsd use cargo-tarpaulin inside vmactions VMs (replacing the prior cargo-llvm-cov plumbing).
• Dedicated MSRV CI job removed — rust-version = "1.85.0" is the declared contract; users on Rust 1.85 follow the documented cargo update -p smol_str --precise 0.3.2 / cargo update -p criterion --precise 0.7.0 workaround. The DragonFly CI job uses these pins because pkg-shipped Rust there is 1.85.1.
• criterion dev-dep range relaxed from ^0.8 to >=0.7, <0.9 so the DragonFly CI's cargo update --precise 0.7.0 satisfies the constraint while normal users still resolve to the latest 0.8.x.

• Fix the inverted ConvertInterfaceLuidToIndex check on Windows that caused the crate to discard the LUID-derived interface index and return Ipv6IfIndex on success (or 0 on failure) instead. Affected interface_table, interface_addr_table, interface_multiaddr_table, and best_local_addrs_in.
• Harden the Linux netlink attribute parser so malformed kernel messages can no longer panic: IPv4 IFA_ADDRESS payload length guard, IFLA_MTU empty-payload guard, safe bounded read for IFLA_IFNAME (replaces CStr::from_ptr which scanned past the attribute on a non-null-terminated value), and clamp of rta_align_of() against the remaining buffer so an unaligned last attribute cannot overflow. Added the missing bounds check in the RTM_NEWROUTE metric/oif parser.
• Fix FreeBSD build (#32): libc does not expose rt_msghdr, ifa_msghdr, or NET_RT_IFLIST2 on non-Apple BSDs. getifs now ships a bsd_like::compat module with local #[repr(C)] definitions for FreeBSD, DragonFly, NetBSD, and OpenBSD, and gates NET_RT_IFLIST2 / if_msghdr2 / ifma_msghdr2 to Apple only.
• Fix NetBSD build: IfaMsghdr compat for NetBSD/OpenBSD (whose layouts differ from FreeBSD), and cast if_data.ifi_mtu to u32 since NetBSD defines it as uint64_t.

• Windows: Information no longer materializes a SmallVec<IP_ADAPTER_ADDRESSES_LH> of ~400-byte struct copies. A zero-copy AdapterIter<'_> walks the kernel's linked list in place.
• BSD routing: replace the O(n²) results.contains(&addr) dedup with an O(1) HashSet<(index, IpAddr)> so large routing tables scale linearly.
• Linux: drop the per-line MediumVec<&str> allocation in /proc/net/igmp and /proc/net/igmp6 parsing — walk the whitespace-separated iterator directly.
• Linux: switch ifindex_to_name to rustix::net::netdevice::index_to_name_inlined (rustix 1.1). The returned stack-allocated InlinedName combined with SmolStr's inline-up-to-23-bytes optimization makes the call allocation-free on the happy path.
• Extract a shared adapter_index helper on Windows, deduplicating three copies of the LUID→index resolution block.

• MSRV bumped from 1.63.0 to 1.64.0.
• Bump rustix requirement from 1 to 1.1 (required for index_to_name_inlined).
• Bump smallvec-wrapper requirement from 0.3 to 0.4.
• Bump criterion dev-dependency from 0.7 to 0.8.
• Drop the unused either dependency (was declared both at the top level and under the Linux-specific target).

• Refresh the crate description.
• Add a benchmark CI workflow with Criterion benches for interfaces, local_ip_address, and gateway.
• Bump windows-sys requirement from 0.60 to 0.61.
• Bump linux-raw-sys, which, and criterion dev-dependency versions.

• Bump up versions

• Fix #8

• Fix #6

• Add ifname_to_v6_iface, ifname_to_v4_iface, and ifname_to_iface.

• Remove libc on Linux platform, implement netlink by using rustix

• Refactor IfNet and add Ifv4Net, Ifv6Net, IfAddr, Ifv4Addr and Ifv6Addr
• Add interface_addrs, interface_ipv4_addrs, interface_ipv6_addrs, interface_addrs_by_filter, interface_ipv4_addrs_by_filter, and interface_ipv6_addrs_by_filter
• Add interface_multicast_addrs, interface_multicast_ipv4_addrs, interface_multicast_ipv6_addrs, interface_multicast_addrs_by_filter, interface_multcast_ipv4_addrs_by_filter and interface_multcast_ipv6_addrs_by_filter
• Add gateway_addrs, gateway_ipv4_addrs, gateway_ipv6_addrs, gateway_addrs_by_filter, gateway_ipv4_addrs_by_filter and gateway_ipv6_addrs_by_filter
• Add local_addrs, local_ipv4_addrs, local_ipv6_addrs, local_addrs_by_filter, local_ipv4_addrs_by_filter and local_ipv6_addrs_by_filter
• Add private_addrs, private_ipv4_addrs, private_ipv6_addrs, private_addrs_by_filter, private_ipv4_addrs_by_filter, private_ipv6_addrs_by_filter
• Add public_addrs, public_ipv4_addrs, public_ipv6_addrs, public_addrs_by_filter, public_ipv4_addrs_by_filter, and public_ipv6_addrs_by_filter
• Add ifindex_to_name and ifname_to_index