Cross-device dial ignores the multiaddr TCP port — dials a random port → `Connection refused, errno=111`

[cloudabe]

---

Title

Cross-device dial ignores the multiaddr TCP port — dials a random port → Connection refused, errno=111

Environment

• dart_libp2p: 1.0.3 (latest on pub.dev; no newer release exists)
• Dart/Flutter: Flutter 3.44 / Dart 3.12
• Devices: two real Android phones on the same Wi-Fi
  • SM A566B (Android 16), SM N975F (Android 12)
  • LAN IPs 192.168.0.104 / 192.168.0.105
• Transport: TCP (/ip4/.../tcp/.../p2p/<id>)

Summary

When dialing a peer by an explicit, correct multiaddr, BasicHost.connect /
Swarm dials a different (random) TCP port than the one in the multiaddr,
so the connection is refused. The multiaddr we supply is correct — a standalone
Multiaddr parse proves it — so the wrong port is introduced inside the
library between our addAddrs/bootstrap and the actual tcp_transport.dial.

Reproduction

1. Start host B listening on a fixed TCP port (e.g. 9002) with a fixed peer
   id (stable ed25519 seed). It advertises
   /ip4/192.168.0.105/tcp/9002/p2p/<B-peer-id>.
2. On host A, bootstrap/dial B with that exact multiaddr:
   /ip4/192.168.0.105/tcp/9002/p2p/<B-peer-id>
3. Observe the dial attempt.

Expected: TCP connection to 192.168.0.105:9002.
Actual: connection attempt to 192.168.0.105:41982 (a different,
apparently random port) → Connection refused, errno=111.

Evidence the input is correct

A standalone parse of the exact multiaddr we pass:

final ma = Multiaddr('/ip4/192.168.0.105/tcp/9002/p2p/12D3KooW...');
print(ma.valueForProtocol('tcp')); // -> 9002

So our input carries tcp: 9002. The library dials 41982.

Where the wrong port is introduced (source trace)

• tcp_transport.dart reads valueForProtocol('tcp') from the multiaddr it is
  given; if that were 0 it would throw ArgumentError. It does not throw,
  meaning the multiaddr reaching the transport already has the wrong (valid) port
  41982. So the corruption happens before the transport.
• Swarm.dialPeer reads candidate addresses from
  _peerstore.addrBook.addrs(peerId). AddressFilter.filterReachable only
  filters by IP type — it never touches the port.
• BasicHost.connect applies _addrsFactory(pi.addrs) to the peer addrs before
  storing them. The default factory (defaultAddrsFactory /
  _defaultAddrsFactoryInternal) only filters loopback / unspecified IPs and
  does not rewrite ports — so the factory is not the culprit either.

Net: the port is rewritten somewhere in the swarm/identify/address-book path
between our bootstrap addAddrs and the tcp_transport.dial call. The
same-process (loopback) "mesh" unit test passes only because both hosts resolve
the real port from in-process host state, not the cross-process multiaddr.

Impact

Cross-device / cross-process libp2p mesh is unusable on real networks with
explicit bootstrap multiaddrs. mDNS discovery is also broken on these Android
devices (Bad state: Cannot add event after closing inside the library), so
explicit multiaddr dialing is the only viable path — and it fails for the same
port reason.

Suggested fix area (for maintainer)

Inspect the address-book / Identify observed-addr handling in Swarm /
BasicHost that produces the dial candidate list. The candidate port should
equal the port in the bootstrap multiaddr; instead a random/ephemeral port is
substituted.