Channel-scoped error can be lost while channeld exits

[tankyleo]

On master f80895bac, channeld sent a warning through peer_failed_warn_nodisconnect and began exiting. connectd then decrypted the peer's channel-scoped WIRE_ERROR and queued it to that subdaemon after lightningd had begun killing it. channeld never received the error, so the channel remained open.

BOLT #1 requires a receiving node, upon receiving a channel-scoped error, to:

MUST fail the channel referred to by channel_id, if that channel is with the sending node.

BOLT #1 requirements

Forwarding WIRE_ERROR directly from connectd to lightningd when a matching subdaemon exists fixes the race. End-to-end regtest validation confirmed that lightningd failed the channel, and broadcasted the commitment:

diff --git a/connectd/multiplex.c b/connectd/multiplex.c
index 1051645f..696e9e1b 100644
--- a/connectd/multiplex.c
+++ b/connectd/multiplex.c
@@ -1493,6 +1493,15 @@ static struct io_plan *read_body_from_peer_done(struct io_conn *peer_conn,

        /* If we don't find a subdaemon for this, create a new one. */
        subd = find_subd(peer, &channel_id);
+       if (subd && type == WIRE_ERROR) {
+               daemon_conn_send(peer->daemon->master,
+                                take(towire_connectd_peer_spoke(NULL, &peer->id,
+                                                               peer->counter,
+                                                               type,
+                                                               &channel_id,
+                                                               is_peer_error(tmpctx, decrypted))));
+               return next_read(peer_conn, peer);
+       }
        if (!subd) {
               enum peer_wire t = fromwire_peektype(decrypted);