Make vGPU retention records fully delete-only

A fork or snapshot of a failed-create retention stub could never boot:
the stub has no boot configuration, and clearing the delete-only marker
on the child would only produce a startable-but-broken record that
recreates a vGPU from GPUProfile with incomplete metadata. Reject fork
and snapshot of retention stubs with the same invalid_state guidance as
start, so delete (which retries the release) is the only action on
them.

Author: yummybomb

lib/instances/fork.go

@@ -219,6 +219,12 @@ func (m *manager) forkInstanceFromStoppedOrStandby(ctx context.Context, id strin
 	default:
 		return nil, false, fmt.Errorf("%w: cannot fork from state %s (must be Stopped or Standby)", ErrInvalidState, source.State)
 	}
+	if stored.GPURetainedForCleanup {
+		// A delete-only retention stub from a failed create has no boot
+		// configuration, so a fork of it could never boot. Delete the stub to
+		// release its retained vGPU assignment.
+		return nil, false, fmt.Errorf("%w: instance retains a vGPU assignment from a failed create and has no boot configuration; delete it to release the assignment", ErrInvalidState)
+	}
 
 	if !supportValidated {
 		if err := m.validateForkSupport(ctx, stored.HypervisorType); err != nil {

lib/instances/fork_test.go

@@ -63,6 +63,31 @@ func TestForkInstanceClearsVGPUAssignment(t *testing.T) {
 	assert.Equal(t, "/sys/bus/pci/devices/0000:82:00.4", source.GPUDevicePath)
 }
 
+func TestForkInstanceRejectsVGPURetentionRecord(t *testing.T) {
+	manager, _ := setupTestManager(t)
+	ctx := context.Background()
+	hvType := hypervisor.Type("fork-vgpu-retention-test")
+	hypervisor.RegisterCapabilities(hvType, hypervisor.Capabilities{SupportsConcurrentForkPrepare: true})
+	manager.vmStarters[hvType] = concurrentForkPrepareTestStarter{}
+
+	sourceID := "fork-vgpu-retention-source"
+	createStoppedSnapshotSourceFixture(t, manager, sourceID, sourceID, hvType)
+
+	meta, err := manager.loadMetadata(sourceID)
+	require.NoError(t, err)
+	meta.GPUProfile = "NVIDIA L40S-2Q"
+	meta.GPUFramework = devices.VGPUFramework("future-framework")
+	meta.GPUDevicePath = "/sys/bus/pci/devices/0000:82:00.4"
+	meta.GPURetainedForCleanup = true
+	require.NoError(t, manager.saveMetadata(meta))
+
+	// The delete-only retention stub has no boot configuration, so a fork of
+	// it could never boot; only delete may act on it.
+	_, err = manager.ForkInstance(ctx, sourceID, ForkInstanceRequest{Name: "fork-vgpu-retention-copy"})
+	require.ErrorIs(t, err, ErrInvalidState)
+	require.ErrorContains(t, err, "delete it to release the assignment")
+}
+
 func TestForkInstance_VZStoppedSourceSupported(t *testing.T) {
 	t.Parallel()
 	manager, _ := setupTestManager(t)

lib/instances/snapshot.go

@@ -66,6 +66,13 @@ func (m *manager) createSnapshot(ctx context.Context, id string, req CreateSnaps
 	inst := m.toInstance(ctx, meta)
 	stored := &meta.StoredMetadata
 
+	if stored.GPURetainedForCleanup {
+		// A delete-only retention stub from a failed create has no boot
+		// configuration, so a snapshot of it could never be restored or
+		// forked into a bootable instance. Delete the stub to release its
+		// retained vGPU assignment.
+		return nil, fmt.Errorf("%w: instance retains a vGPU assignment from a failed create and has no boot configuration; delete it to release the assignment", ErrInvalidState)
+	}
 	if err := validateForkVolumeSafety(stored.Volumes); err != nil {
 		return nil, fmt.Errorf("%w: snapshot requires readonly volume attachments: %v", ErrNotSupported, err)
 	}

lib/instances/snapshot_test.go

@@ -52,6 +52,31 @@ func TestForkSnapshotClearsVGPUAssignment(t *testing.T) {
 	assert.Equal(t, "/sys/bus/pci/devices/0000:82:00.4", source.GPUDevicePath)
 }
 
+func TestCreateSnapshotRejectsVGPURetentionRecord(t *testing.T) {
+	mgr, _ := setupTestManager(t)
+	ctx := context.Background()
+
+	sourceID := "snapshot-vgpu-retention"
+	createStoppedSnapshotSourceFixture(t, mgr, sourceID, sourceID, mgr.defaultHypervisor)
+
+	meta, err := mgr.loadMetadata(sourceID)
+	require.NoError(t, err)
+	meta.GPUProfile = "NVIDIA L40S-2Q"
+	meta.GPUFramework = devices.VGPUFramework("future-framework")
+	meta.GPUDevicePath = "/sys/bus/pci/devices/0000:82:00.4"
+	meta.GPURetainedForCleanup = true
+	require.NoError(t, mgr.saveMetadata(meta))
+
+	// The delete-only retention stub has no boot configuration, so a snapshot
+	// of it could never be restored or forked into a bootable instance.
+	_, err = mgr.CreateSnapshot(ctx, sourceID, CreateSnapshotRequest{
+		Kind: SnapshotKindStopped,
+		Name: "snapshot-vgpu-retention",
+	})
+	require.ErrorIs(t, err, ErrInvalidState)
+	require.ErrorContains(t, err, "delete it to release the assignment")
+}
+
 func TestRestoreSnapshotDoesNotResurrectStaleVGPUAssignment(t *testing.T) {
 	mgr, _ := setupTestManager(t)
 	ctx := context.Background()