diff --git a/src/commands/common.rs b/src/commands/common.rs
index fac68d86..a95ca969 100644
--- a/src/commands/common.rs
+++ b/src/commands/common.rs
@@ -1436,7 +1436,10 @@ pub async fn create_snapshot_core(
 
     // Firecracker resets all vsock connections during snapshot creation
     // (VIRTIO_VSOCK_EVENT_TRANSPORT_RESET). Bump restore-epoch in MMDS so fc-agent's
-    // background watcher detects this and remounts FUSE volumes.
+    // background watcher detects this and triggers handle_clone_restore() which
+    // re-registers the exec server's AsyncFd (stale after transport reset) and
+    // reconnects the output vsock. This MUST succeed — if the epoch isn't bumped,
+    // the exec server stays stale and health checks hang for ~60s.
     let restore_epoch = std::time::SystemTime::now()
         .duration_since(std::time::UNIX_EPOCH)
         .map(|d| d.as_secs())
@@ -1449,7 +1452,8 @@ pub async fn create_snapshot_core(
         }))
         .await
     {
-        warn!(error = %e, "failed to bump restore-epoch after snapshot (FUSE remount may be delayed)");
+        let _ = tokio::fs::remove_dir_all(&temp_snapshot_dir).await;
+        return Err(e).context("bumping restore-epoch in MMDS after snapshot");
     }
 
     if has_base {