diff --git a/Dockerfile b/Dockerfile
index 853c2544..9a4a077f 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -41,4 +41,7 @@ LABEL org.opencontainers.image.title="ember" \
       org.opencontainers.image.description="low-latency distributed cache" \
       org.opencontainers.image.source="https://github.com/kacy/ember"
 
+HEALTHCHECK --interval=10s --timeout=3s --start-period=5s --retries=3 \
+  CMD wget -qO- http://localhost:9100/health || exit 1
+
 ENTRYPOINT ["ember-server"]
diff --git a/crates/ember-cluster/src/topology.rs b/crates/ember-cluster/src/topology.rs
index b309d6da..51846bed 100644
--- a/crates/ember-cluster/src/topology.rs
+++ b/crates/ember-cluster/src/topology.rs
@@ -350,7 +350,8 @@ impl std::fmt::Display for ClusterHealth {
 
 impl ClusterState {
     /// Creates a new cluster state for a single-node cluster.
-    pub fn single_node(local_node: ClusterNode) -> Self {
+    pub fn single_node(mut local_node: ClusterNode) -> Self {
+        local_node.config_epoch = 1;
         let local_id = local_node.id;
         let slot_map = SlotMap::single_node(local_id);
         let mut nodes = HashMap::new();
diff --git a/crates/ember-server/src/cluster.rs b/crates/ember-server/src/cluster.rs
index 63d0e86e..c91b235f 100644
--- a/crates/ember-server/src/cluster.rs
+++ b/crates/ember-server/src/cluster.rs
@@ -123,12 +123,16 @@ impl ClusterCoordinator {
         })?;
         let gossip_addr = SocketAddr::new(bind_addr.ip(), gossip_port);
 
-        let gossip = GossipEngine::new(local_id, gossip_addr, gossip_config, event_tx);
+        let mut gossip = GossipEngine::new(local_id, gossip_addr, gossip_config, event_tx);
 
         let state = if bootstrap {
             let mut node = ClusterNode::new_primary_with_offset(local_id, bind_addr, port_offset);
             node.set_myself();
-            ClusterState::single_node(node)
+            let cs = ClusterState::single_node(node);
+            // Populate local_slots so Welcome replies correctly advertise all owned slots
+            // instead of sending an empty list and triggering a stale SlotsChanged event.
+            gossip.set_local_slots(cs.slot_map.slots_for_node(local_id));
+            cs
         } else {
             let mut cs = ClusterState::new(local_id);
             let mut node = ClusterNode::new_primary_with_offset(local_id, bind_addr, port_offset);
@@ -431,6 +435,16 @@ impl ClusterCoordinator {
             }
         };
 
+        // Insert placeholder before sending UDP so the gossip receive task always
+        // finds an entry during MemberJoined resolution. Without this, the Welcome
+        // reply can arrive and be processed before we add the placeholder, leaving
+        // both a real entry and a stale placeholder in state.nodes.
+        {
+            let mut state = self.state.write().await;
+            let node = ClusterNode::new_primary_with_offset(new_id, addr, self.gossip_port_offset);
+            state.add_node(node);
+        }
+
         // send join message via UDP
         {
             let socket = self.udp_socket.lock().await;
@@ -444,14 +458,6 @@ impl ClusterCoordinator {
             }
         }
 
-        // add a placeholder to the routing table so MOVED redirects work
-        // immediately — the real node ID arrives via gossip and replaces this
-        {
-            let mut state = self.state.write().await;
-            let node = ClusterNode::new_primary_with_offset(new_id, addr, self.gossip_port_offset);
-            state.add_node(node);
-        }
-
         self.save_config().await;
 
         Frame::Simple("OK".into())
@@ -1617,6 +1623,11 @@ impl ClusterCoordinator {
                             false
                         }
                         GossipEvent::SlotsChanged(id, slots) => {
+                            // The local node is authoritative for its own slot ownership;
+                            // external gossip about it must never overwrite canonical state.
+                            if id == coordinator.local_id {
+                                false
+                            } else {
                             debug!(
                                 "cluster: node {} slots changed ({} ranges)",
                                 id,
@@ -1640,6 +1651,7 @@ impl ClusterCoordinator {
                             }
                             state.update_health();
                             true
+                            }
                         }
                         GossipEvent::RoleChanged(id, is_primary, replicates) => {
                             debug!(
diff --git a/scripts/dev-cluster.sh b/scripts/dev-cluster.sh
index 0fd3d4f5..449aa502 100755
--- a/scripts/dev-cluster.sh
+++ b/scripts/dev-cluster.sh
@@ -103,12 +103,12 @@ cmd_start() {
     "$CLI" -p "$p1" cluster meet 127.0.0.1 "$p3" >/dev/null
 
     echo "assigning slots..."
-    "$CLI" -p "$p1" cluster addslotsrange 0 5460 >/dev/null
+    # node-1 already owns slots 0-5460 via bootstrap; nodes 2 and 3 take their ranges
     "$CLI" -p "$p2" cluster addslotsrange 5461 10922 >/dev/null
     "$CLI" -p "$p3" cluster addslotsrange 10923 16383 >/dev/null
 
-    # give gossip a moment to converge
-    sleep 0.5
+    # give gossip time to fully converge before printing status
+    sleep 1.5
 
     echo ""
     "$CLI" -p "$p1" cluster info