[AlxcBTW]

Handed over to @hypermoist for T41; built, fixed and tested locally (below).

What

The engine side of the virtual LAN (bridge protocol, Broadcasts and the virtual LAN). Everything is behind the bridge's virtual address (state.address != 0 and a connected net service); with no session the host path is exactly as before.

• Interface. sceNetCtlGetInfo shows the virtual address (code 14), 255.254.0.0 (15), no gateway (16) and MTU 1200 (3) while a session is active. The virtual interface is also reported when the host has none.
• Sockets. A plain AF_INET SOCK_DGRAM socket bound to 0.0.0.0, the virtual address or a broadcast address, or one that sends/connects before a bind, is carried through the net service. Per virtual port one loopback group socket is registered with bind; sends leave through it with the 8-byte header, and a small forwarder thread copies each delivered datagram to every game socket on that port (so SO_REUSEADDR/SO_REUSEPORT sharing binds once and delivers to all). Port 0 takes a free virtual port, and getsockname shows the game's port.
• Out. 255.255.255.255 and 198.19.255.255 need SO_BROADCAST (else EACCES) and go out as broadcast frames; 198.18.0.0/15 unicast goes to that peer; multicast is dropped without an error; every other destination fails at once with ENETUNREACH. A socket that stays on the host during a session (bound to a real or loopback address, or bound before the session) does not send to a broadcast or virtual address either: such a send or connect fails with ENETUNREACH, and its multicast is dropped. Nothing is ever sent to the host network. connect/send/getpeername work on these sockets and a connected socket only hears its peer.
• In. Frames arrive with the peer's virtual address and the sender's port (recv, recvfrom, recvmsg).
• Ports are data. No port numbers or title knowledge in code; the tests use arbitrary ports.
• Tests: net_bridge_lan against FakeNetService covers broadcast out (both addresses, no SO_BROADCAST, multicast, off-subnet), broadcast in, unicast, an inactive session (before and after), the interface answers, bind to the virtual address, ephemeral and implicit binds, connect, shared ports with SO_REUSEADDR and SO_REUSEPORT, a cleared reuse option, host sockets during a session, two first sends at once on one unbound socket, and sceNetRecvmsg (vectors, MSG_TRUNC, sender name).
• Docs: new entry in docs/dev/debt/libSceNet.md.

Spec questions this design settles

• Answer by broadcast or by unicast? Both work: a unicast reply to the sender's address and port is an ordinary datagram, and a reply to a broadcast address is carried as a broadcast.
• Limited or subnet broadcast? Both are carried; the game's own computed subnet broadcast equals 198.19.255.255 because the interface netmask is 255.254.0.0.
• Which interface call? sceNetCtlGetInfo (the call seen at boot) is answered. sceNetGetIfList* and sceNetConfigGetIfaddr still throw (not implemented today) and stay open.

Still open

• Multicast (mDNS): dropped, not carried, as the spec says; group joins are not implemented.
• Interface list/ifaddr calls (broadcast address of an interface).
• Sockets bound to a real address, to 127.0.0.0/8 or before the session started stay on the host (their broadcast and virtual sends fail, see Out); IPv6 sockets are untouched.
• A socket keeps its mode when the session ends (sends are then dropped by the net service).
• Nothing measured on a PS5 or with a game; no game was started.
• Overlap: feat(bridge): observe the network libraries through the import thunks, not in every export #35 touches the same exports in libSceNet/Export.cpp (observation hook); whichever lands second merges main and resolves it.

Tested

Built and run locally on Linux and on Windows (MinGW cross build, run under Wine), commit 5ede8a54d, main merged in.

• Linux (Release): ctest -R 'net|bridge|^public_headers' --timeout 120 passes, 124 of 124 that run; bridge_e2e_game skips as designed without a live net service. net_bridge_lan also passed 30 runs in a row and 16 runs 8 at a time.
• Windows under Wine (Release): libSceNet.prx, libSceNetCtl.prx and every net/bridge test built. net_bridge_lan, net_bridge_p2p, net_bridge, guest_net_ctl and the rest pass (net_bridge_lan passed 50 serial runs). guest_sce_net and net_ipv6only_patched fail under Wine on unchanged main sources too, so they are not from this PR. public_headers_compile_0 hit its 30 s limit on a cold first run and passes when run alone.
• Mutation checks: the tests fail when any of these is undone: clearing SO_REUSEADDR, the host-socket guard, the fix for two first sends at once, MSG_TRUNC in sceNetRecvmsg, the SO_BROADCAST check, and connected-peer filtering.
• test_public_headers.py --audit-only and tools/check_conventions.py pass.

Fixed while testing: host sockets could still send broadcasts to the real network during a session; two threads sending first on one unbound socket bound it twice and leaked a virtual port (they could also fail with EINVAL); leaving a shared port dropped every membership of that socket; a cleared reuse option stayed set; the bind request held the lock that the forwarder and every send need.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Fixes: none
• Unblocks: none (hub task T41)
• GPUs tested: n/a
• AI-assisted: yes

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}