fix(xmtp_mls): stabilize should_reconnect flaky test in d14n suite

crates/xmtp_api/src/mls.rs

@@ -278,10 +278,24 @@ where
         // _NOTE_:
         // Default ID Cursor should be one
         // else we miss welcome messages
-        self.api_client
-            .subscribe_welcome_messages(&[installation_key])
-            .await
-            .map_err(crate::dyn_err)
+        //
+        // Retry the subscription setup so callers are insulated from
+        // transient transport errors — for example, after a network blip
+        // the underlying tonic `Channel` reconnects lazily on the next RPC
+        // and the first post-reconnect call can observe a
+        // `Status::Cancelled` / "connection closed" while the channel is
+        // still transitioning out of its transient-failure state
+        // (see issue #3438). Non-retryable errors short-circuit via
+        // `RetryableError`.
+        retry_async!(
+            self.retry_strategy,
+            (async {
+                self.api_client
+                    .subscribe_welcome_messages(&[installation_key])
+                    .await
+            })
+        )
+        .map_err(crate::dyn_err)
     }
 
     pub async fn publish_commit_log(&self, requests: Vec<PublishCommitLogRequest>) -> Result<()> {