test(cluster): gate test execution on stable Raft leader election

Add `NodeHandle::metadata_group_leader()` which reads the observed
metadata-group leader id from the node's local Raft state, returning
`0` while an election is still in progress.

Wire a leader-stability barrier into `ClusterHarness::spawn_three()`
that polls until every node reports the same non-zero leader id before
returning control to the test. This closes the race where CPU pressure
delayed the initial Raft heartbeats past topology convergence,
producing `not leader (leader hint: None)` errors on the first DDL or
descriptor-lease call issued by the test.

Also add the `pgwire_gateway_migration` binary to the nextest cluster
test group so it runs serialised with the other 3-node integration
tests and does not compete for threads.

Author: farhan-syah

.config/nextest.toml

@@ -45,6 +45,7 @@ binary(/cluster/)
 | binary(descriptor_versioning_cross_node)
 | binary(prepared_cache_invalidation)
 | binary(sql_cluster_cross_node_dml)
+| binary(pgwire_gateway_migration)
 '''
 test-group = 'cluster'
 threads-required = 'num-test-threads'

nodedb/tests/common/cluster_harness/cluster.rs

@@ -86,6 +86,39 @@ impl TestCluster {
         )
         .await;
 
+        // CRITICAL: wait for the metadata Raft group to elect a leader
+        // and for every node's local view to agree on the same leader id.
+        //
+        // Topology convergence + rolling-upgrade exit only guarantees
+        // membership and wire version are agreed; they say nothing about
+        // election state. Under heavy host load (e.g. running this test
+        // immediately after another full-suite cluster test exits and
+        // the unit-test pool ramps back up), the initial Raft heartbeat
+        // window can be missed and the first `acquire`/`propose` issued
+        // by the test races a re-election — surfacing as
+        // `raft error: not leader (leader hint: None)` from a
+        // descriptor-lease or DDL call.
+        //
+        // Waiting until every node reports the same non-zero leader id
+        // closes the window deterministically. Symmetric to the
+        // rolling-upgrade wait above: no retries, no flakes, no
+        // wasted CI minutes on cleanup of a doomed cluster bringup.
+        wait_for(
+            "metadata group has stable leader visible on every node",
+            Duration::from_secs(10),
+            Duration::from_millis(20),
+            || {
+                let leaders: Vec<u64> = cluster
+                    .nodes
+                    .iter()
+                    .map(|n| n.metadata_group_leader())
+                    .collect();
+                let first = leaders[0];
+                first != 0 && leaders.iter().all(|&l| l == first)
+            },
+        )
+        .await;
+
         Ok(cluster)
     }
 

nodedb/tests/common/cluster_harness/node.rs

@@ -320,6 +320,25 @@ impl TestClusterNode {
             .unwrap_or(0)
     }
 
+    /// Observed metadata-group leader id from this node's local Raft
+    /// state, or `0` if no leader is known yet (election in progress).
+    /// Polled by the cluster harness `spawn_three()` to gate test
+    /// execution on a stable leader — otherwise tests racing the first
+    /// election see `not leader (leader hint: None)` errors when CPU
+    /// pressure delays the initial heartbeats past topology convergence.
+    pub fn metadata_group_leader(&self) -> u64 {
+        let Some(observer) = self.shared.cluster_observer.get() else {
+            return 0;
+        };
+        observer
+            .group_status
+            .group_statuses()
+            .into_iter()
+            .find(|g| g.group_id == nodedb_cluster::METADATA_GROUP_ID)
+            .map(|g| g.leader_id)
+            .unwrap_or(0)
+    }
+
     /// Number of active collections visible on this node (read through
     /// the local `SystemCatalog` redb — populated by the
     /// `MetadataCommitApplier` on every node via