fix(cluster): skip persistence scheduling until initialize has run

schedulePersist and scheduleControlPersist can fire from configure() and
the public scheduling paths before initialize() has completed. Without a
guard, those queued microtasks call persistState/persistControlState,
which try to mkdir PATHS.DATA_DIR and write state files from tests and
short-lived scripts that never meant to touch the data directory. That
produced async-leak warnings in the Cluster manager unit tests and
left orphan directories on hosts that only constructed a ClusterManager
to inspect it.

Add an `initialized` flag set at the end of initialize() and early-return
from both schedulers when it is false. Real runtime paths always call
initialize() during Daemon startup, so this changes no production
behavior.

ts/cluster/cluster-manager.ts

@@ -13,6 +13,7 @@ import type {
 import { CLUSTER, PATHS } from '../constants.ts';
 
 export class ClusterManager {
+  private initialized = false;
   private config: IClusterConfig = {
     enabled: false,
     nodeName: 'modelgrid-local',
@@ -63,6 +64,8 @@ export class ClusterManager {
     } catch {
       // No persisted control state yet.
     }
+
+    this.initialized = true;
   }
 
   public configure(config: IClusterConfig): void {
@@ -384,6 +387,10 @@ export class ClusterManager {
   }
 
   private schedulePersist(): void {
+    if (!this.initialized) {
+      return;
+    }
+
     if (this.persistQueued) {
       return;
     }
@@ -396,6 +403,10 @@ export class ClusterManager {
   }
 
   private scheduleControlPersist(): void {
+    if (!this.initialized) {
+      return;
+    }
+
     if (this.controlPersistQueued) {
       return;
     }