fix(ecstore): fence decommission cancellation

crates/ecstore/src/core/pools.rs

@@ -3475,6 +3475,8 @@ impl ECStore {
     async fn decommission_cancel_with_owner(&self, idx: usize, owner: Option<&DecommissionCanceler>) -> Result<()> {
         ensure_decommission_terminal_operation_supported(self.single_pool(), "cancel decommission")?;
         let _start_guard = self.start_gate.lock().await;
+        let operation_gate = self.ctx.decommission_operation_gate();
+        let operation_guard = operation_gate.write().await;
 
         // Lock order: decommission_cancelers before pool_meta. Holding both makes
         // owner validation and the terminal transition one atomic operation.
@@ -3533,8 +3535,6 @@ impl ECStore {
             );
         }
 
-        self.wait_for_decommission_side_effects().await;
-
         if should_save_pool_meta && let Err(err) = self.save_current_pool_meta().await {
             if let Some(previous_pool_meta) = previous_pool_meta {
                 let mut pool_meta = self.pool_meta.write().await;
@@ -3542,6 +3542,7 @@ impl ECStore {
             }
             return Err(err);
         }
+        drop(operation_guard);
 
         if let Some(canceler) = terminal_canceler.as_ref() {
             self.release_decommission_canceler_slot(idx, canceler).await;

crates/ecstore/src/store/init.rs

@@ -4616,11 +4616,24 @@ mod tests {
             async move { store.decommission_cancel(0).await }
         });
         assert!(
-            tokio::time::timeout(StdDuration::from_millis(100), &mut cancel)
+            tokio::time::timeout(Duration::from_millis(100), &mut cancel).await.is_err(),
-                .await
-                .is_err(),
             "cancel must wait for the final sweep source cleanup"
         );
+        {
+            let pool_meta = store.pool_meta.read().await;
+            let decommission = pool_meta.pools[0]
+                .decommission
+                .as_ref()
+                .expect("decommission state should remain present");
+            assert!(
+                !decommission.canceled,
+                "cancel must not publish terminal state before the final sweep drains"
+            );
+            assert!(
+                decommission.start_time.is_some(),
+                "cancel must preserve the run identity until the final sweep drains"
+            );
+        }
 
         barrier.release();
         final_sweep
@@ -4631,12 +4644,13 @@ mod tests {
             .await
             .expect("cancel task should not panic")
             .expect("cancel should complete after the final sweep releases the operation gate");
-        assert!(
+        let pool_meta = store.pool_meta.read().await;
-            store.pool_meta.read().await.pools[0]
+        let decommission = pool_meta.pools[0]
-                .decommission
+            .decommission
-                .as_ref()
+            .as_ref()
-                .is_some_and(|info| info.canceled)
+            .expect("decommission state should remain present");
-        );
+        assert!(decommission.canceled);
+        assert!(decommission.start_time.is_none());
     }
 
     #[cfg(feature = "test-util")]