agent/recover: only treat running containers as up

Recover built its up-set from every listed container, so a stopped/exited
container or a dead unikernel VM (whose state dir is still listed) counted
as already-running and was skipped — the exact reason dead VMs stayed in
drift instead of being recovered. Filter the set to State==running.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

internal/agent/recover.go

@@ -21,16 +21,21 @@ func (a *Agent) Recover(ctx context.Context) error {
 		return fmt.Errorf("list services: %w", err)
 	}
 
-	// Get the list of currently running containers from podman.
+	// Get the list of containers across runtimes. Only those actually in the
-	running, err := a.listAllContainers(ctx)
+	// "running" state count as up — a stopped/exited container or a dead
+	// unikernel VM (whose state dir still exists, so it is listed) must be
+	// recovered, not skipped.
+	listed, err := a.listAllContainers(ctx)
 	if err != nil {
 		a.Logger.Warn("cannot list containers, assuming none running", "err", err)
-		running = nil
+		listed = nil
 	}
 	runningSet := make(map[string]bool)
-	for _, c := range running {
+	for _, c := range listed {
+		if c.State == "running" {
 			runningSet[c.Name] = true
 		}
+	}
 
 	var recovered, skipped, already int