[mdebbar]

Flutter's Mac web_tool_tests has been flaking for months because the extension backend can't start:

SocketException: Failed to create server socket (OS Error: Address already in use, errno = 48), address = ::, port = 49158
  #4 startHttpServer (package:dwds/src/utilities/server.dart:49:20)
  #5 ExtensionBackend.start (package:dwds/src/servers/extension_backend.dart:63:20)

Flutter passes hostname any by default, and HttpMultiServer.bind('any', port) binds [::] dual-stack. But the port comes from findUnusedPort(), which only checks [::1] with v6Only: true. That port can still be held over IPv4, so the dual-stack bind fails. Every retry uses the same probe, and in all of these failures all 5 attempts failed (~500ms before the error).

Example CI failures:

• https://ci.chromium.org/b/8668518017083050129
• https://ci.chromium.org/b/8669996444518535409
• https://ci.chromium.org/b/8673423200118982481

This PR binds port 0 on the actual address, so the OS picks a port that's free for that exact bind. A requested port (e.g. the debug service's 44456) is still tried first, and the retry loop stays. #1451 moved away from port 0 because a single bind attempt failed without retrying (flutter/flutter#93256), so the retries still cover that case.

I confirmed the mismatch on Linux: with 127.0.0.1:P or 0.0.0.0:P held, a [::1] v6Only bind on P succeeds and a [::] bind on P fails with EADDRINUSE. There's no deterministic regression test because the failure depends on which port the kernel hands out.

Related to flutter/flutter#185452

[bkonyi]

This LGTM, but would you mind moving this over into the Dart SDK's pkg/dwds? We're in the middle of migrating DWDS development into the SDK and we're treating that version as the source of truth.

[mdebbar]

@bkonyi Done: dart-lang/sdk#64580