Broker: make aborting a pending TCP connect on shutdown a guest-neutral platform operation #1771

On the ulitebox branch, shutting down a TCP socket whose connect is pending follows Linux's tcp_disconnect, but the broker implements it rather than the Linux shim. Each socket provider aborts the connect and fails it with SocketError::ConnectionReset: the smoltcp provider (litebox_broker_smoltcp/src/tcp.rs, "As Linux's tcp_disconnect"), and, with #1752, the userland provider (abort_pending_connect in litebox_broker_userland/src/socket/tcp.rs). The userland provider has to do it itself because Winsock doesn't abort a pending connect on shutdown. Broker core's shutdown_stream (litebox_broker_core/src/socket.rs) infers the abort: when a shutdown completes and status() then reports Failed, core resets the socket's connection instead of recording the shutdown. #1752 states this requirement on PlatformSocket::shutdown, but it is still guest semantics in a guest-neutral layer: the broker, not the shim, decides that the connect aborts and which error the guest sees.

Consequences

• A guest with different shutdown semantics can't get them without changing every provider.
• Linux lets a socket connect again after tcp_disconnect aborts its connect. The broker's socket stays failed, so the shim can't offer that.

Proposed direction

• Make cancelling a pending connect an explicit, guest-neutral operation, in one of two ways:
  • a stated PlatformSocket::shutdown requirement whose outcome carries no guest error, for example a distinct aborted status;
  • a separate cancel-connect operation.
• Have the Linux shim decide when a shutdown cancels a pending connect. It should report ECONNRESET once, as tcp_disconnect does, and could let the socket connect again.
• Change core, smoltcp and the userland provider together. tcp_broker_states.c should pass unchanged on both the Linux and Windows brokers.