diff --git a/hive/daemon/daemon.go b/hive/daemon/daemon.go
index 27c3d3475..ebf19ecd0 100644
--- a/hive/daemon/daemon.go
+++ b/hive/daemon/daemon.go
@@ -621,8 +621,17 @@ func (d *Daemon) restoreInstances() {
 			continue
 		}
 
-		if instance.Attributes.StopInstance {
-			slog.Info("Instance flagged as user initiated stop, skipping", "instance", instance.ID)
+		if instance.Attributes.StopInstance || instance.Status == vm.StateStopped {
+			slog.Info("Instance is stopped, subscribing to NATS for start commands", "instance", instance.ID)
+			d.mu.Lock()
+			sub, err := d.natsConn.QueueSubscribe(fmt.Sprintf("ec2.cmd.%s", instance.ID), "hive-events", d.handleEC2Events)
+			if err != nil {
+				d.mu.Unlock()
+				slog.Error("Failed to subscribe stopped instance to NATS", "instance", instance.ID, "err", err)
+				continue
+			}
+			d.natsSubscriptions[instance.ID] = sub
+			d.mu.Unlock()
 			continue
 		}
 
@@ -1162,12 +1171,15 @@ func (d *Daemon) stopInstance(instances map[string]*vm.VM, deleteVolume bool) er
 	// For stop operations, keep the subscription so we can receive start commands
 	if deleteVolume {
 		for _, instance := range instances {
-			slog.Info("Unsubscribing from NATS subject", "instance", instance.ID)
-			if err := d.natsSubscriptions[fmt.Sprintf("ec2.cmd.%s", instance.ID)].Unsubscribe(); err != nil {
-				slog.Error("Failed to unsubscribe from NATS subject", "instance", instance.ID, "err", err)
+			d.mu.Lock()
+			if sub, ok := d.natsSubscriptions[instance.ID]; ok {
+				slog.Info("Unsubscribing from NATS subject", "instance", instance.ID)
+				if err := sub.Unsubscribe(); err != nil {
+					slog.Error("Failed to unsubscribe from NATS subject", "instance", instance.ID, "err", err)
+				}
+				delete(d.natsSubscriptions, instance.ID)
 			}
-			// TODO: Remove redundant subscription if not used
-			//d.natsSubscriptions[fmt.Sprintf("ec2.describe.%s", instance.ID)].Unsubscribe()
+			d.mu.Unlock()
 		}
 	}
 	return nil
diff --git a/hive/daemon/state_test.go b/hive/daemon/state_test.go
index 62fd72d81..de7521bf0 100644
--- a/hive/daemon/state_test.go
+++ b/hive/daemon/state_test.go
@@ -1,11 +1,15 @@
 package daemon
 
 import (
+	"encoding/json"
+	"fmt"
 	"os"
+	"sync"
 	"testing"
 	"time"
 
 	"github.com/mulgadc/hive/hive/config"
+	"github.com/mulgadc/hive/hive/qmp"
 	"github.com/mulgadc/hive/hive/vm"
 	"github.com/nats-io/nats-server/v2/server"
 	"github.com/nats-io/nats.go"
@@ -168,3 +172,670 @@ func TestTransitionState_NilEC2Instance(t *testing.T) {
 	require.NoError(t, err)
 	assert.Equal(t, vm.StateRunning, instance.Status)
 }
+
+// allStates is the canonical list of every InstanceState in the system.
+var allStates = []vm.InstanceState{
+	vm.StateProvisioning,
+	vm.StatePending,
+	vm.StateRunning,
+	vm.StateStopping,
+	vm.StateStopped,
+	vm.StateShuttingDown,
+	vm.StateTerminated,
+	vm.StateError,
+}
+
+// stableStates are states where an instance can rest indefinitely.
+var stableStates = []vm.InstanceState{
+	vm.StateRunning,
+	vm.StateStopped,
+	vm.StateTerminated,
+	vm.StateError,
+}
+
+// transitionalStates are states that must resolve to a stable state.
+var transitionalStates = []vm.InstanceState{
+	vm.StateProvisioning,
+	vm.StatePending,
+	vm.StateStopping,
+	vm.StateShuttingDown,
+}
+
+// TestValidTransitions_MapCoversAllNonTerminalStates verifies that every state
+// except StateTerminated has at least one valid outgoing transition. A missing
+// entry would mean an instance could get permanently stuck.
+func TestValidTransitions_MapCoversAllNonTerminalStates(t *testing.T) {
+	for _, state := range allStates {
+		if state == vm.StateTerminated {
+			continue
+		}
+		targets, ok := vm.ValidTransitions[state]
+		assert.True(t, ok, "state %q missing from ValidTransitions map", state)
+		assert.NotEmpty(t, targets, "state %q has no valid transitions — instances would get stuck", state)
+	}
+}
+
+// TestTerminatedState_HasNoOutgoingTransitions ensures the terminal state is
+// truly terminal — no valid transitions out.
+func TestTerminatedState_HasNoOutgoingTransitions(t *testing.T) {
+	targets, ok := vm.ValidTransitions[vm.StateTerminated]
+	assert.False(t, ok || len(targets) > 0,
+		"StateTerminated should have no valid transitions, got %v", targets)
+}
+
+// TestTransitionalStates_CanReachStableState verifies that every transitional
+// state has at least one direct transition to a stable state. This is the
+// structural guarantee that prevents the "stuck in stopping" bug.
+func TestTransitionalStates_CanReachStableState(t *testing.T) {
+	stableSet := map[vm.InstanceState]bool{}
+	for _, s := range stableStates {
+		stableSet[s] = true
+	}
+
+	for _, state := range transitionalStates {
+		targets := vm.ValidTransitions[state]
+		hasStableTarget := false
+		for _, target := range targets {
+			if stableSet[target] {
+				hasStableTarget = true
+				break
+			}
+		}
+		assert.True(t, hasStableTarget,
+			"transitional state %q has no direct path to a stable state (targets: %v)", state, targets)
+	}
+}
+
+// TestEveryState_CanReachTerminated does a BFS from each non-terminated state
+// to confirm there is always a path to StateTerminated. An unreachable terminal
+// state means instances can never be fully cleaned up.
+func TestEveryState_CanReachTerminated(t *testing.T) {
+	for _, start := range allStates {
+		if start == vm.StateTerminated {
+			continue
+		}
+
+		visited := map[vm.InstanceState]bool{start: true}
+		queue := []vm.InstanceState{start}
+		found := false
+
+		for len(queue) > 0 {
+			current := queue[0]
+			queue = queue[1:]
+			if current == vm.StateTerminated {
+				found = true
+				break
+			}
+			for _, next := range vm.ValidTransitions[current] {
+				if !visited[next] {
+					visited[next] = true
+					queue = append(queue, next)
+				}
+			}
+		}
+
+		assert.True(t, found,
+			"state %q has no path to StateTerminated — instances starting here can never be cleaned up", start)
+	}
+}
+
+// TestStopLifecycle_CompletesFullChain walks the full stop lifecycle:
+// provisioning → running → stopping → stopped, then restart and terminate:
+// stopped → running → shutting-down → terminated.
+func TestStopLifecycle_CompletesFullChain(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	instance := &vm.VM{ID: "i-lifecycle-stop", Status: vm.StateProvisioning}
+	daemon.Instances.Mu.Lock()
+	daemon.Instances.VMS[instance.ID] = instance
+	daemon.Instances.Mu.Unlock()
+
+	chain := []vm.InstanceState{
+		vm.StateRunning,      // launch
+		vm.StateStopping,     // stop requested
+		vm.StateStopped,      // stop completed
+		vm.StateRunning,      // restart
+		vm.StateShuttingDown, // terminate requested
+		vm.StateTerminated,   // terminate completed
+	}
+
+	for _, target := range chain {
+		err := daemon.TransitionState(instance, target)
+		require.NoError(t, err, "transition to %s failed (current: %s)", target, instance.Status)
+	}
+
+	assert.Equal(t, vm.StateTerminated, instance.Status)
+}
+
+// TestStopLifecycle_ErrorRecovery verifies that an instance that hits an error
+// during stop can recover: running → stopping → error → shutting-down → terminated.
+func TestStopLifecycle_ErrorRecovery(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	instance := &vm.VM{ID: "i-lifecycle-err", Status: vm.StateRunning}
+	daemon.Instances.Mu.Lock()
+	daemon.Instances.VMS[instance.ID] = instance
+	daemon.Instances.Mu.Unlock()
+
+	chain := []vm.InstanceState{
+		vm.StateStopping,
+		vm.StateError,        // stop failed
+		vm.StateShuttingDown, // force terminate from error
+		vm.StateTerminated,
+	}
+
+	for _, target := range chain {
+		err := daemon.TransitionState(instance, target)
+		require.NoError(t, err, "transition to %s failed (current: %s)", target, instance.Status)
+	}
+
+	assert.Equal(t, vm.StateTerminated, instance.Status)
+}
+
+// TestShuttingDown_ErrorRecovery verifies: shutting-down → error → shutting-down → terminated.
+func TestShuttingDown_ErrorRecovery(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	instance := &vm.VM{ID: "i-shutdown-err", Status: vm.StateShuttingDown}
+	daemon.Instances.Mu.Lock()
+	daemon.Instances.VMS[instance.ID] = instance
+	daemon.Instances.Mu.Unlock()
+
+	chain := []vm.InstanceState{
+		vm.StateError,        // terminate failed
+		vm.StateShuttingDown, // retry terminate
+		vm.StateTerminated,   // success
+	}
+
+	for _, target := range chain {
+		err := daemon.TransitionState(instance, target)
+		require.NoError(t, err, "transition to %s failed (current: %s)", target, instance.Status)
+	}
+
+	assert.Equal(t, vm.StateTerminated, instance.Status)
+}
+
+// TestTransitionState_ConcurrentTransitions hammers the same instance with
+// concurrent transitions to verify the mutex protects against races and
+// the instance always ends in a valid state.
+func TestTransitionState_ConcurrentTransitions(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	instance := &vm.VM{ID: "i-concurrent", Status: vm.StateRunning}
+	daemon.Instances.Mu.Lock()
+	daemon.Instances.VMS[instance.ID] = instance
+	daemon.Instances.Mu.Unlock()
+
+	// Multiple goroutines race to transition from running.
+	// After one succeeds, the others may chain from the new state
+	// (e.g., running→error, then error→shutting-down), which is correct.
+	targets := []vm.InstanceState{vm.StateStopping, vm.StateShuttingDown, vm.StateError}
+
+	var wg sync.WaitGroup
+	results := make([]error, len(targets))
+
+	for i, target := range targets {
+		wg.Add(1)
+		go func(idx int, tgt vm.InstanceState) {
+			defer wg.Done()
+			results[idx] = daemon.TransitionState(instance, tgt)
+		}(i, target)
+	}
+	wg.Wait()
+
+	// At least one transition must succeed.
+	successes := 0
+	for _, err := range results {
+		if err == nil {
+			successes++
+		}
+	}
+	assert.GreaterOrEqual(t, successes, 1, "at least one concurrent transition should succeed")
+
+	// The final state must be reachable from running via valid transitions.
+	finalState := instance.Status
+	reachable := reachableStates(vm.StateRunning)
+	assert.True(t, reachable[finalState],
+		"instance ended in state %q which is not reachable from running", finalState)
+}
+
+// reachableStates returns all states reachable from start via BFS.
+func reachableStates(start vm.InstanceState) map[vm.InstanceState]bool {
+	visited := map[vm.InstanceState]bool{start: true}
+	queue := []vm.InstanceState{start}
+	for len(queue) > 0 {
+		current := queue[0]
+		queue = queue[1:]
+		for _, next := range vm.ValidTransitions[current] {
+			if !visited[next] {
+				visited[next] = true
+				queue = append(queue, next)
+			}
+		}
+	}
+	return visited
+}
+
+// TestTransitionState_DoubleTransitionSameTarget verifies that transitioning
+// to the same state twice is rejected (no self-loops).
+func TestTransitionState_DoubleTransitionSameTarget(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	for _, state := range allStates {
+		t.Run(string(state), func(t *testing.T) {
+			instance := &vm.VM{ID: fmt.Sprintf("i-double-%s", state), Status: state}
+			daemon.Instances.Mu.Lock()
+			daemon.Instances.VMS[instance.ID] = instance
+			daemon.Instances.Mu.Unlock()
+
+			err := daemon.TransitionState(instance, state)
+			assert.Error(t, err, "self-transition %s -> %s should be rejected", state, state)
+			assert.Equal(t, state, instance.Status, "state should be unchanged after rejected self-transition")
+		})
+	}
+}
+
+// TestTransitionState_InvalidFromTerminated exhaustively checks that
+// terminated cannot transition to any state.
+func TestTransitionState_InvalidFromTerminated(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	for _, target := range allStates {
+		t.Run(fmt.Sprintf("terminated->%s", target), func(t *testing.T) {
+			instance := &vm.VM{ID: fmt.Sprintf("i-term-%s", target), Status: vm.StateTerminated}
+			daemon.Instances.Mu.Lock()
+			daemon.Instances.VMS[instance.ID] = instance
+			daemon.Instances.Mu.Unlock()
+
+			err := daemon.TransitionState(instance, target)
+			assert.Error(t, err, "transition from terminated to %s should be rejected", target)
+			assert.Equal(t, vm.StateTerminated, instance.Status)
+		})
+	}
+}
+
+// TestTransitionState_StoppingMustReachStoppedOrError is the specific
+// regression test for the bug where stopping never reached stopped.
+// It verifies both the happy path and error path from stopping.
+func TestTransitionState_StoppingMustReachStoppedOrError(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	t.Run("happy_path", func(t *testing.T) {
+		instance := &vm.VM{ID: "i-stop-happy", Status: vm.StateStopping}
+		daemon.Instances.Mu.Lock()
+		daemon.Instances.VMS[instance.ID] = instance
+		daemon.Instances.Mu.Unlock()
+
+		err := daemon.TransitionState(instance, vm.StateStopped)
+		require.NoError(t, err)
+		assert.Equal(t, vm.StateStopped, instance.Status)
+	})
+
+	t.Run("error_path", func(t *testing.T) {
+		instance := &vm.VM{ID: "i-stop-error", Status: vm.StateStopping}
+		daemon.Instances.Mu.Lock()
+		daemon.Instances.VMS[instance.ID] = instance
+		daemon.Instances.Mu.Unlock()
+
+		err := daemon.TransitionState(instance, vm.StateError)
+		require.NoError(t, err)
+		assert.Equal(t, vm.StateError, instance.Status)
+	})
+
+	t.Run("force_terminate_path", func(t *testing.T) {
+		instance := &vm.VM{ID: "i-stop-term", Status: vm.StateStopping}
+		daemon.Instances.Mu.Lock()
+		daemon.Instances.VMS[instance.ID] = instance
+		daemon.Instances.Mu.Unlock()
+
+		err := daemon.TransitionState(instance, vm.StateShuttingDown)
+		require.NoError(t, err)
+		assert.Equal(t, vm.StateShuttingDown, instance.Status)
+
+		err = daemon.TransitionState(instance, vm.StateTerminated)
+		require.NoError(t, err)
+		assert.Equal(t, vm.StateTerminated, instance.Status)
+	})
+}
+
+// TestTransitionState_ShuttingDownMustReachTerminatedOrError is the parallel
+// regression test for the terminate flow.
+func TestTransitionState_ShuttingDownMustReachTerminatedOrError(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	t.Run("happy_path", func(t *testing.T) {
+		instance := &vm.VM{ID: "i-shut-happy", Status: vm.StateShuttingDown}
+		daemon.Instances.Mu.Lock()
+		daemon.Instances.VMS[instance.ID] = instance
+		daemon.Instances.Mu.Unlock()
+
+		err := daemon.TransitionState(instance, vm.StateTerminated)
+		require.NoError(t, err)
+		assert.Equal(t, vm.StateTerminated, instance.Status)
+	})
+
+	t.Run("error_path", func(t *testing.T) {
+		instance := &vm.VM{ID: "i-shut-error", Status: vm.StateShuttingDown}
+		daemon.Instances.Mu.Lock()
+		daemon.Instances.VMS[instance.ID] = instance
+		daemon.Instances.Mu.Unlock()
+
+		err := daemon.TransitionState(instance, vm.StateError)
+		require.NoError(t, err)
+		assert.Equal(t, vm.StateError, instance.Status)
+	})
+}
+
+// TestValidTransitions_NoCyclesWithoutStableState ensures there is no cycle
+// of only transitional states. Every cycle in the state graph must pass
+// through at least one stable state.
+func TestValidTransitions_NoCyclesWithoutStableState(t *testing.T) {
+	stableSet := map[vm.InstanceState]bool{}
+	for _, s := range stableStates {
+		stableSet[s] = true
+	}
+
+	// DFS from each transitional state — if we revisit a transitional state
+	// without going through a stable state, that's a problem.
+	for _, start := range transitionalStates {
+		visited := map[vm.InstanceState]bool{}
+		var hasCycle bool
+
+		var dfs func(vm.InstanceState)
+		dfs = func(current vm.InstanceState) {
+			if stableSet[current] {
+				return // reached a stable state, stop this path
+			}
+			if visited[current] {
+				hasCycle = true
+				return
+			}
+			visited[current] = true
+			for _, next := range vm.ValidTransitions[current] {
+				dfs(next)
+			}
+		}
+		dfs(start)
+
+		assert.False(t, hasCycle,
+			"found a cycle of only transitional states starting from %q", start)
+	}
+}
+
+// TestTransitionState_MultipleInstancesIndependent verifies that transitioning
+// one instance does not affect another instance's state.
+func TestTransitionState_MultipleInstancesIndependent(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	inst1 := &vm.VM{ID: "i-multi-1", Status: vm.StateRunning}
+	inst2 := &vm.VM{ID: "i-multi-2", Status: vm.StateRunning}
+	daemon.Instances.Mu.Lock()
+	daemon.Instances.VMS[inst1.ID] = inst1
+	daemon.Instances.VMS[inst2.ID] = inst2
+	daemon.Instances.Mu.Unlock()
+
+	err := daemon.TransitionState(inst1, vm.StateStopping)
+	require.NoError(t, err)
+
+	assert.Equal(t, vm.StateStopping, inst1.Status)
+	assert.Equal(t, vm.StateRunning, inst2.Status, "inst2 should be unaffected by inst1's transition")
+}
+
+// --- Recovery / restoreInstances tests ---
+//
+// These tests simulate a daemon restart by writing instance state to JetStream,
+// then calling restoreInstances on a fresh daemon. Since no QEMU process is
+// running, isInstanceProcessRunning returns false for all instances, which
+// exercises the recovery state resolution logic.
+
+// TestRestoreInstances_StoppingFinalizedToStopped verifies that an instance
+// stuck in StateStopping when the daemon died gets finalized to StateStopped.
+func TestRestoreInstances_StoppingFinalizedToStopped(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	// Simulate: daemon was running, instance was in StateStopping, daemon crashed.
+	daemon.Instances.VMS["i-restore-stop"] = &vm.VM{
+		ID:     "i-restore-stop",
+		Status: vm.StateStopping,
+	}
+	require.NoError(t, daemon.WriteState())
+
+	// Clear in-memory state to simulate fresh daemon startup.
+	daemon.Instances.VMS = make(map[string]*vm.VM)
+
+	// restoreInstances loads from JetStream and resolves transitional states.
+	daemon.restoreInstances()
+
+	instance, ok := daemon.Instances.VMS["i-restore-stop"]
+	require.True(t, ok, "instance should be loaded from state")
+	assert.Equal(t, vm.StateStopped, instance.Status,
+		"stopping instance should be finalized to stopped on recovery")
+}
+
+// TestRestoreInstances_ShuttingDownFinalizedToTerminated verifies that an
+// instance stuck in StateShuttingDown gets finalized to StateTerminated.
+func TestRestoreInstances_ShuttingDownFinalizedToTerminated(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	daemon.Instances.VMS["i-restore-shut"] = &vm.VM{
+		ID:     "i-restore-shut",
+		Status: vm.StateShuttingDown,
+	}
+	require.NoError(t, daemon.WriteState())
+
+	daemon.Instances.VMS = make(map[string]*vm.VM)
+	daemon.restoreInstances()
+
+	instance, ok := daemon.Instances.VMS["i-restore-shut"]
+	require.True(t, ok, "instance should be loaded from state")
+	assert.Equal(t, vm.StateTerminated, instance.Status,
+		"shutting-down instance should be finalized to terminated on recovery")
+}
+
+// TestRestoreInstances_TerminatedSkipped verifies that terminated instances
+// are loaded but not relaunched or modified.
+func TestRestoreInstances_TerminatedSkipped(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	daemon.Instances.VMS["i-restore-term"] = &vm.VM{
+		ID:     "i-restore-term",
+		Status: vm.StateTerminated,
+	}
+	require.NoError(t, daemon.WriteState())
+
+	daemon.Instances.VMS = make(map[string]*vm.VM)
+	daemon.restoreInstances()
+
+	instance, ok := daemon.Instances.VMS["i-restore-term"]
+	require.True(t, ok, "terminated instance should still be loaded")
+	assert.Equal(t, vm.StateTerminated, instance.Status,
+		"terminated instance should remain terminated")
+}
+
+// TestRestoreInstances_UserStoppedSkipped verifies that instances flagged as
+// user-stopped are not relaunched.
+func TestRestoreInstances_UserStoppedSkipped(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	daemon.Instances.VMS["i-restore-userstop"] = &vm.VM{
+		ID:     "i-restore-userstop",
+		Status: vm.StateStopped,
+		Attributes: qmp.Attributes{
+			StopInstance: true,
+		},
+	}
+	require.NoError(t, daemon.WriteState())
+
+	daemon.Instances.VMS = make(map[string]*vm.VM)
+	daemon.restoreInstances()
+
+	instance, ok := daemon.Instances.VMS["i-restore-userstop"]
+	require.True(t, ok, "user-stopped instance should still be loaded")
+	assert.Equal(t, vm.StateStopped, instance.Status,
+		"user-stopped instance should remain stopped")
+}
+
+// TestRestoreInstances_RunningResetToPending verifies that a running instance
+// whose QEMU process is dead gets reset to pending for relaunch.
+// LaunchInstance will fail (no QEMU available in tests), but the state
+// should be set to pending before the attempt.
+func TestRestoreInstances_RunningResetToPending(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	daemon.Instances.VMS["i-restore-run"] = &vm.VM{
+		ID:     "i-restore-run",
+		Status: vm.StateRunning,
+	}
+	require.NoError(t, daemon.WriteState())
+
+	daemon.Instances.VMS = make(map[string]*vm.VM)
+	daemon.restoreInstances()
+
+	instance, ok := daemon.Instances.VMS["i-restore-run"]
+	require.True(t, ok, "running instance should still be loaded")
+	// LaunchInstance fails (no QEMU), so it stays pending or gets marked failed.
+	// Either way it should NOT remain "running" — that would be a lie.
+	assert.NotEqual(t, vm.StateRunning, instance.Status,
+		"instance should not remain running when QEMU process is dead")
+}
+
+// TestRestoreInstances_MixedStates verifies recovery with multiple instances
+// in different states, ensuring each is handled correctly and independently.
+func TestRestoreInstances_MixedStates(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	daemon.Instances.VMS["i-mix-stopping"] = &vm.VM{
+		ID: "i-mix-stopping", Status: vm.StateStopping,
+	}
+	daemon.Instances.VMS["i-mix-shutting"] = &vm.VM{
+		ID: "i-mix-shutting", Status: vm.StateShuttingDown,
+	}
+	daemon.Instances.VMS["i-mix-terminated"] = &vm.VM{
+		ID: "i-mix-terminated", Status: vm.StateTerminated,
+	}
+	daemon.Instances.VMS["i-mix-stopped"] = &vm.VM{
+		ID: "i-mix-stopped", Status: vm.StateStopped,
+		Attributes: qmp.Attributes{StopInstance: true},
+	}
+	require.NoError(t, daemon.WriteState())
+
+	daemon.Instances.VMS = make(map[string]*vm.VM)
+	daemon.restoreInstances()
+
+	assert.Equal(t, vm.StateStopped, daemon.Instances.VMS["i-mix-stopping"].Status,
+		"stopping should finalize to stopped")
+	assert.Equal(t, vm.StateTerminated, daemon.Instances.VMS["i-mix-shutting"].Status,
+		"shutting-down should finalize to terminated")
+	assert.Equal(t, vm.StateTerminated, daemon.Instances.VMS["i-mix-terminated"].Status,
+		"terminated should remain terminated")
+	assert.Equal(t, vm.StateStopped, daemon.Instances.VMS["i-mix-stopped"].Status,
+		"user-stopped should remain stopped")
+}
+
+// TestRestoreInstances_StatePersistsAfterRecovery verifies that the finalized
+// state is actually persisted to JetStream, so a second restart doesn't
+// re-process the same transitional state.
+func TestRestoreInstances_StatePersistsAfterRecovery(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	daemon.Instances.VMS["i-persist-stop"] = &vm.VM{
+		ID: "i-persist-stop", Status: vm.StateStopping,
+	}
+	daemon.Instances.VMS["i-persist-shut"] = &vm.VM{
+		ID: "i-persist-shut", Status: vm.StateShuttingDown,
+	}
+	require.NoError(t, daemon.WriteState())
+
+	// First restart: finalize transitional states.
+	daemon.Instances.VMS = make(map[string]*vm.VM)
+	daemon.restoreInstances()
+
+	assert.Equal(t, vm.StateStopped, daemon.Instances.VMS["i-persist-stop"].Status)
+	assert.Equal(t, vm.StateTerminated, daemon.Instances.VMS["i-persist-shut"].Status)
+
+	// Second restart: load again, verify states are already finalized.
+	daemon.Instances.VMS = make(map[string]*vm.VM)
+	daemon.restoreInstances()
+
+	assert.Equal(t, vm.StateStopped, daemon.Instances.VMS["i-persist-stop"].Status,
+		"stopped state should persist through second restart")
+	assert.Equal(t, vm.StateTerminated, daemon.Instances.VMS["i-persist-shut"].Status,
+		"terminated state should persist through second restart")
+}
+
+// TestStatePersistence_RoundTrip verifies that all instance fields survive
+// a write-load cycle through JetStream, simulating a daemon restart.
+func TestStatePersistence_RoundTrip(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	original := &vm.VM{
+		ID:           "i-roundtrip",
+		Status:       vm.StateRunning,
+		InstanceType: "t3.micro",
+		Attributes: qmp.Attributes{
+			StopInstance:      false,
+			TerminateInstance: false,
+		},
+	}
+	daemon.Instances.VMS[original.ID] = original
+	require.NoError(t, daemon.WriteState())
+
+	// Simulate restart: clear and reload.
+	daemon.Instances.VMS = make(map[string]*vm.VM)
+	require.NoError(t, daemon.LoadState())
+
+	loaded := daemon.Instances.VMS["i-roundtrip"]
+	require.NotNil(t, loaded)
+	assert.Equal(t, original.ID, loaded.ID)
+	assert.Equal(t, original.Status, loaded.Status)
+	assert.Equal(t, original.InstanceType, loaded.InstanceType)
+	assert.Equal(t, original.Attributes.StopInstance, loaded.Attributes.StopInstance)
+}
+
+// TestRestoreInstances_StoppedInstanceCanReceiveStartCommand verifies that after
+// a daemon restart, a stopped instance has an active ec2.cmd.<id> subscription
+// so the gateway's StartInstances can reach it (no "no responders" error).
+func TestRestoreInstances_StoppedInstanceCanReceiveStartCommand(t *testing.T) {
+	daemon := createDaemonWithJetStream(t)
+
+	instanceID := "i-restore-start"
+	daemon.Instances.VMS[instanceID] = &vm.VM{
+		ID:           instanceID,
+		InstanceType: "t3.micro",
+		Status:       vm.StateStopped,
+		Attributes:   qmp.Attributes{StopInstance: true},
+	}
+	require.NoError(t, daemon.WriteState())
+
+	// Simulate daemon restart: clear in-memory state and restore from JetStream.
+	daemon.Instances.VMS = make(map[string]*vm.VM)
+	daemon.restoreInstances()
+
+	// Instance should still be stopped.
+	instance, ok := daemon.Instances.VMS[instanceID]
+	require.True(t, ok, "stopped instance should be loaded after restore")
+	assert.Equal(t, vm.StateStopped, instance.Status)
+
+	// Send a start command to the per-instance topic (ec2.cmd.<id>).
+	// restoreInstances should have created this subscription for stopped instances.
+	command := qmp.Command{
+		ID: instanceID,
+		QMPCommand: qmp.QMPCommand{
+			Execute:   "cont",
+			Arguments: map[string]any{},
+		},
+		Attributes: qmp.Attributes{
+			StartInstance: true,
+		},
+	}
+	cmdData, err := json.Marshal(command)
+	require.NoError(t, err)
+
+	// The request should get a response (subscription exists), not "no responders".
+	// LaunchInstance will fail in tests (no QEMU), so we expect an error response,
+	// but the key assertion is that the subscription is active.
+	subject := fmt.Sprintf("ec2.cmd.%s", instanceID)
+	reply, err := daemon.natsConn.Request(subject, cmdData, 5*time.Second)
+	require.NoError(t, err, "ec2.cmd.%s should have a subscriber after restoreInstances", instanceID)
+	require.NotNil(t, reply)
+}