feat: runtime manager

rtmanager/internal/service/restartruntime/service.go

@@ -0,0 +1,482 @@
+// Package restartruntime implements the `restart` lifecycle operation
+// owned by Runtime Manager. Restart is a recreate: under one outer
+// per-game lease the service runs the stop service, removes the
+// container with `docker rm`, and runs the start service with the
+// runtime's current `image_ref`. The hostname / engine endpoint stays
+// stable across the recreate; `container_id` changes.
+//
+// Lifecycle and failure-mode semantics follow `rtmanager/README.md
+// §Lifecycles → Restart`. Design rationale is captured in
+// `rtmanager/docs/services.md`, in particular the lease-sharing
+// pattern with `startruntime.Service.Run` / `stopruntime.Service.Run`,
+// the correlation-id reuse on `source_ref`, and the
+// inner-stop-then-rm-failure recovery rule.
+package restartruntime
+
+import (
+	"context"
+	"crypto/rand"
+	"encoding/base64"
+	"errors"
+	"fmt"
+	"log/slog"
+	"strings"
+	"time"
+
+	"galaxy/rtmanager/internal/config"
+	"galaxy/rtmanager/internal/domain/operation"
+	"galaxy/rtmanager/internal/domain/runtime"
+	"galaxy/rtmanager/internal/logging"
+	"galaxy/rtmanager/internal/ports"
+	"galaxy/rtmanager/internal/service/startruntime"
+	"galaxy/rtmanager/internal/service/stopruntime"
+	"galaxy/rtmanager/internal/telemetry"
+)
+
+// leaseReleaseTimeout bounds the deferred lease-release call.
+const leaseReleaseTimeout = 5 * time.Second
+
+// Input stores the per-call arguments for one restart operation.
+type Input struct {
+	// GameID identifies the platform game to restart.
+	GameID string
+
+	// OpSource classifies how the request entered Runtime Manager.
+	// Required: every operation_log entry carries an op_source.
+	OpSource operation.OpSource
+
+	// SourceRef stores the optional opaque per-source reference (REST
+	// request id, admin user id). When non-empty it is reused as the
+	// correlation id linking the outer restart entry to the inner stop
+	// and start log entries.
+	SourceRef string
+}
+
+// Validate reports whether input carries the structural invariants the
+// service requires.
+func (input Input) Validate() error {
+	if strings.TrimSpace(input.GameID) == "" {
+		return fmt.Errorf("game id must not be empty")
+	}
+	if !input.OpSource.IsKnown() {
+		return fmt.Errorf("op source %q is unsupported", input.OpSource)
+	}
+	return nil
+}
+
+// Result stores the deterministic outcome of one Handle call.
+type Result struct {
+	// Record carries the runtime record installed by the inner start on
+	// success; zero on failure.
+	Record runtime.RuntimeRecord
+
+	// Outcome reports whether the operation completed (success) or
+	// produced a stable failure code.
+	Outcome operation.Outcome
+
+	// ErrorCode stores the stable error code on failure. Empty for
+	// success.
+	ErrorCode string
+
+	// ErrorMessage stores the operator-readable detail on failure.
+	// Empty for success.
+	ErrorMessage string
+}
+
+// Dependencies groups the collaborators required by Service.
+type Dependencies struct {
+	// RuntimeRecords reads the runtime record at the start of restart
+	// to capture the current image_ref and container_id.
+	RuntimeRecords ports.RuntimeRecordStore
+
+	// OperationLogs records the outer restart audit entry. Inner stop
+	// and start services append their own entries through their own
+	// stores.
+	OperationLogs ports.OperationLogStore
+
+	// Docker drives the docker rm step between the inner stop and
+	// inner start.
+	Docker ports.DockerClient
+
+	// Leases serialises operations against the same game id. The outer
+	// lease is held for the entire stop + rm + start sequence.
+	Leases ports.GameLeaseStore
+
+	// StopService runs the inner stop step under the outer lease.
+	StopService *stopruntime.Service
+
+	// StartService runs the inner start step under the outer lease.
+	StartService *startruntime.Service
+
+	// Coordination supplies the per-game lease TTL.
+	Coordination config.CoordinationConfig
+
+	// Telemetry records restart outcomes and lease latency. Required.
+	Telemetry *telemetry.Runtime
+
+	// Logger records structured service-level events. Defaults to
+	// `slog.Default()` when nil.
+	Logger *slog.Logger
+
+	// Clock supplies the wall-clock used for operation timestamps.
+	// Defaults to `time.Now` when nil.
+	Clock func() time.Time
+
+	// NewToken supplies a unique opaque token. Used both for the lease
+	// and for the correlation id when Input.SourceRef is empty.
+	// Defaults to a 32-byte random base64url string when nil.
+	NewToken func() string
+}
+
+// Service executes the restart lifecycle operation.
+type Service struct {
+	runtimeRecords ports.RuntimeRecordStore
+	operationLogs  ports.OperationLogStore
+	docker         ports.DockerClient
+	leases         ports.GameLeaseStore
+	stopService    *stopruntime.Service
+	startService   *startruntime.Service
+
+	leaseTTL time.Duration
+
+	telemetry *telemetry.Runtime
+	logger    *slog.Logger
+
+	clock    func() time.Time
+	newToken func() string
+}
+
+// NewService constructs one Service from deps.
+func NewService(deps Dependencies) (*Service, error) {
+	switch {
+	case deps.RuntimeRecords == nil:
+		return nil, errors.New("new restart runtime service: nil runtime records")
+	case deps.OperationLogs == nil:
+		return nil, errors.New("new restart runtime service: nil operation logs")
+	case deps.Docker == nil:
+		return nil, errors.New("new restart runtime service: nil docker client")
+	case deps.Leases == nil:
+		return nil, errors.New("new restart runtime service: nil lease store")
+	case deps.StopService == nil:
+		return nil, errors.New("new restart runtime service: nil stop service")
+	case deps.StartService == nil:
+		return nil, errors.New("new restart runtime service: nil start service")
+	case deps.Telemetry == nil:
+		return nil, errors.New("new restart runtime service: nil telemetry runtime")
+	}
+	if err := deps.Coordination.Validate(); err != nil {
+		return nil, fmt.Errorf("new restart runtime service: coordination config: %w", err)
+	}
+
+	clock := deps.Clock
+	if clock == nil {
+		clock = time.Now
+	}
+	logger := deps.Logger
+	if logger == nil {
+		logger = slog.Default()
+	}
+	logger = logger.With("service", "rtmanager.restartruntime")
+
+	newToken := deps.NewToken
+	if newToken == nil {
+		newToken = defaultTokenGenerator()
+	}
+
+	return &Service{
+		runtimeRecords: deps.RuntimeRecords,
+		operationLogs:  deps.OperationLogs,
+		docker:         deps.Docker,
+		leases:         deps.Leases,
+		stopService:    deps.StopService,
+		startService:   deps.StartService,
+		leaseTTL:       deps.Coordination.GameLeaseTTL,
+		telemetry:      deps.Telemetry,
+		logger:         logger,
+		clock:          clock,
+		newToken:       newToken,
+	}, nil
+}
+
+// Handle executes one restart operation end-to-end. The Go-level error
+// return is reserved for non-business failures (nil context, nil
+// receiver). Every business outcome — success or any of the stable
+// failure codes — flows through Result.
+func (service *Service) Handle(ctx context.Context, input Input) (Result, error) {
+	if service == nil {
+		return Result{}, errors.New("restart runtime: nil service")
+	}
+	if ctx == nil {
+		return Result{}, errors.New("restart runtime: nil context")
+	}
+
+	opStartedAt := service.clock().UTC()
+
+	if err := input.Validate(); err != nil {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeInvalidRequest,
+			errorMessage: err.Error(),
+		}), nil
+	}
+
+	token := service.newToken()
+	leaseStart := service.clock()
+	acquired, err := service.leases.TryAcquire(ctx, input.GameID, token, service.leaseTTL)
+	service.telemetry.RecordLeaseAcquireLatency(ctx, service.clock().Sub(leaseStart))
+	if err != nil {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeServiceUnavailable,
+			errorMessage: fmt.Sprintf("acquire game lease: %s", err.Error()),
+		}), nil
+	}
+	if !acquired {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeConflict,
+			errorMessage: "another lifecycle operation is in progress for this game",
+		}), nil
+	}
+	defer service.releaseLease(ctx, input.GameID, token)
+
+	return service.runUnderLease(ctx, input, opStartedAt)
+}
+
+// runUnderLease executes the lease-protected restart sequence. Loads
+// the runtime record, runs inner stop, removes the container, runs
+// inner start.
+func (service *Service) runUnderLease(ctx context.Context, input Input, opStartedAt time.Time) (Result, error) {
+	existing, err := service.runtimeRecords.Get(ctx, input.GameID)
+	if errors.Is(err, runtime.ErrNotFound) {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeNotFound,
+			errorMessage: fmt.Sprintf("runtime record for game %q does not exist", input.GameID),
+		}), nil
+	}
+	if err != nil {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeInternal,
+			errorMessage: fmt.Sprintf("load runtime record: %s", err.Error()),
+		}), nil
+	}
+	if existing.Status == runtime.StatusRemoved {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeConflict,
+			errorMessage: fmt.Sprintf("runtime for game %q is removed; cannot restart", input.GameID),
+			imageRef:     existing.CurrentImageRef,
+		}), nil
+	}
+	if strings.TrimSpace(existing.CurrentImageRef) == "" {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeInternal,
+			errorMessage: fmt.Sprintf("runtime record for game %q has no image_ref to restart with", input.GameID),
+		}), nil
+	}
+
+	correlationRef := input.SourceRef
+	if correlationRef == "" {
+		correlationRef = service.newToken()
+	}
+	containerID := existing.CurrentContainerID
+	imageRef := existing.CurrentImageRef
+
+	stopResult, err := service.stopService.Run(ctx, stopruntime.Input{
+		GameID:    input.GameID,
+		Reason:    stopruntime.StopReasonAdminRequest,
+		OpSource:  input.OpSource,
+		SourceRef: correlationRef,
+	})
+	if err != nil {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeInternal,
+			errorMessage: fmt.Sprintf("inner stop: %s", err.Error()),
+			imageRef:     imageRef,
+			containerID:  containerID,
+		}), nil
+	}
+	if stopResult.Outcome == operation.OutcomeFailure {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    stopResult.ErrorCode,
+			errorMessage: fmt.Sprintf("inner stop failed: %s", stopResult.ErrorMessage),
+			imageRef:     imageRef,
+			containerID:  containerID,
+		}), nil
+	}
+
+	if containerID != "" {
+		if err := service.docker.Remove(ctx, containerID); err != nil {
+			return service.recordFailure(ctx, failureCtx{
+				opStartedAt:  opStartedAt,
+				input:        input,
+				errorCode:    startruntime.ErrorCodeServiceUnavailable,
+				errorMessage: fmt.Sprintf("docker remove: %s", err.Error()),
+				imageRef:     imageRef,
+				containerID:  containerID,
+			}), nil
+		}
+	}
+
+	startResult, err := service.startService.Run(ctx, startruntime.Input{
+		GameID:    input.GameID,
+		ImageRef:  imageRef,
+		OpSource:  input.OpSource,
+		SourceRef: correlationRef,
+	})
+	if err != nil {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeInternal,
+			errorMessage: fmt.Sprintf("inner start: %s", err.Error()),
+			imageRef:     imageRef,
+		}), nil
+	}
+	if startResult.Outcome == operation.OutcomeFailure {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startResult.ErrorCode,
+			errorMessage: fmt.Sprintf("inner start failed: %s", startResult.ErrorMessage),
+			imageRef:     imageRef,
+		}), nil
+	}
+
+	finishedAt := service.clock().UTC()
+	service.bestEffortAppend(ctx, operation.OperationEntry{
+		GameID:      input.GameID,
+		OpKind:      operation.OpKindRestart,
+		OpSource:    input.OpSource,
+		SourceRef:   correlationRef,
+		ImageRef:    imageRef,
+		ContainerID: startResult.Record.CurrentContainerID,
+		Outcome:     operation.OutcomeSuccess,
+		StartedAt:   opStartedAt,
+		FinishedAt:  &finishedAt,
+	})
+	service.telemetry.RecordRestartOutcome(ctx, string(operation.OutcomeSuccess), "")
+
+	logArgs := []any{
+		"game_id", input.GameID,
+		"prev_container_id", containerID,
+		"new_container_id", startResult.Record.CurrentContainerID,
+		"image_ref", imageRef,
+		"op_source", string(input.OpSource),
+	}
+	logArgs = append(logArgs, logging.ContextAttrs(ctx)...)
+	service.logger.InfoContext(ctx, "runtime restarted", logArgs...)
+
+	return Result{
+		Record:  startResult.Record,
+		Outcome: operation.OutcomeSuccess,
+	}, nil
+}
+
+// failureCtx groups the inputs to recordFailure.
+type failureCtx struct {
+	opStartedAt  time.Time
+	input        Input
+	errorCode    string
+	errorMessage string
+	imageRef     string
+	containerID  string
+}
+
+// recordFailure records the outer failure operation_log entry and emits
+// telemetry. Inner stop / start services have already recorded their
+// own entries; this is the outer summary.
+func (service *Service) recordFailure(ctx context.Context, fc failureCtx) Result {
+	finishedAt := service.clock().UTC()
+	service.bestEffortAppend(ctx, operation.OperationEntry{
+		GameID:       fc.input.GameID,
+		OpKind:       operation.OpKindRestart,
+		OpSource:     fc.input.OpSource,
+		SourceRef:    correlationRefOrEmpty(fc.input),
+		ImageRef:     fc.imageRef,
+		ContainerID:  fc.containerID,
+		Outcome:      operation.OutcomeFailure,
+		ErrorCode:    fc.errorCode,
+		ErrorMessage: fc.errorMessage,
+		StartedAt:    fc.opStartedAt,
+		FinishedAt:   &finishedAt,
+	})
+	service.telemetry.RecordRestartOutcome(ctx, string(operation.OutcomeFailure), fc.errorCode)
+
+	logArgs := []any{
+		"game_id", fc.input.GameID,
+		"image_ref", fc.imageRef,
+		"op_source", string(fc.input.OpSource),
+		"error_code", fc.errorCode,
+		"error_message", fc.errorMessage,
+	}
+	logArgs = append(logArgs, logging.ContextAttrs(ctx)...)
+	service.logger.WarnContext(ctx, "runtime restart failed", logArgs...)
+
+	return Result{
+		Outcome:      operation.OutcomeFailure,
+		ErrorCode:    fc.errorCode,
+		ErrorMessage: fc.errorMessage,
+	}
+}
+
+// correlationRefOrEmpty returns the original Input.SourceRef for the
+// outer entry. Outer-failure paths that did not yet generate a
+// correlation id (input validation, lease busy) keep the original
+// `source_ref` which is the actor ref.
+func correlationRefOrEmpty(input Input) string {
+	return input.SourceRef
+}
+
+// releaseLease releases the per-game lease in a fresh background context.
+func (service *Service) releaseLease(ctx context.Context, gameID, token string) {
+	cleanupCtx, cancel := context.WithTimeout(context.Background(), leaseReleaseTimeout)
+	defer cancel()
+	if err := service.leases.Release(cleanupCtx, gameID, token); err != nil {
+		service.logger.WarnContext(ctx, "release game lease",
+			"game_id", gameID,
+			"err", err.Error(),
+		)
+	}
+}
+
+// bestEffortAppend writes one outer operation_log entry. Inner ops have
+// already appended their own; a failure here only loses the outer
+// summary, which is acceptable.
+func (service *Service) bestEffortAppend(ctx context.Context, entry operation.OperationEntry) {
+	if _, err := service.operationLogs.Append(ctx, entry); err != nil {
+		service.logger.ErrorContext(ctx, "append operation log",
+			"game_id", entry.GameID,
+			"op_kind", string(entry.OpKind),
+			"outcome", string(entry.Outcome),
+			"error_code", entry.ErrorCode,
+			"err", err.Error(),
+		)
+	}
+}
+
+// defaultTokenGenerator returns a function that produces 32-byte
+// base64url-encoded tokens.
+func defaultTokenGenerator() func() string {
+	return func() string {
+		var buf [32]byte
+		if _, err := rand.Read(buf[:]); err != nil {
+			return "rtmanager-fallback-token"
+		}
+		return base64.RawURLEncoding.EncodeToString(buf[:])
+	}
+}

rtmanager/internal/service/restartruntime/service_test.go

@@ -0,0 +1,584 @@
+package restartruntime_test
+
+import (
+	"context"
+	"errors"
+	"sync"
+	"testing"
+	"time"
+
+	"galaxy/notificationintent"
+	"galaxy/rtmanager/internal/adapters/docker/mocks"
+	"galaxy/rtmanager/internal/config"
+	"galaxy/rtmanager/internal/domain/operation"
+	"galaxy/rtmanager/internal/domain/runtime"
+	"galaxy/rtmanager/internal/ports"
+	"galaxy/rtmanager/internal/service/restartruntime"
+	"galaxy/rtmanager/internal/service/startruntime"
+	"galaxy/rtmanager/internal/service/stopruntime"
+	"galaxy/rtmanager/internal/telemetry"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"go.uber.org/mock/gomock"
+)
+
+// --- shared fake doubles ----------------------------------------------
+
+type fakeRuntimeRecords struct {
+	mu sync.Mutex
+
+	stored          map[string]runtime.RuntimeRecord
+	getErr          error
+	upsertErr       error
+	updateStatusErr error
+
+	upserts []runtime.RuntimeRecord
+	updates []ports.UpdateStatusInput
+}
+
+func newFakeRuntimeRecords() *fakeRuntimeRecords {
+	return &fakeRuntimeRecords{stored: map[string]runtime.RuntimeRecord{}}
+}
+
+func (s *fakeRuntimeRecords) Get(_ context.Context, gameID string) (runtime.RuntimeRecord, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.getErr != nil {
+		return runtime.RuntimeRecord{}, s.getErr
+	}
+	record, ok := s.stored[gameID]
+	if !ok {
+		return runtime.RuntimeRecord{}, runtime.ErrNotFound
+	}
+	return record, nil
+}
+
+func (s *fakeRuntimeRecords) Upsert(_ context.Context, record runtime.RuntimeRecord) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.upsertErr != nil {
+		return s.upsertErr
+	}
+	s.upserts = append(s.upserts, record)
+	s.stored[record.GameID] = record
+	return nil
+}
+
+func (s *fakeRuntimeRecords) UpdateStatus(_ context.Context, input ports.UpdateStatusInput) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.updates = append(s.updates, input)
+	if s.updateStatusErr != nil {
+		return s.updateStatusErr
+	}
+	record, ok := s.stored[input.GameID]
+	if !ok {
+		return runtime.ErrNotFound
+	}
+	if record.Status != input.ExpectedFrom {
+		return runtime.ErrConflict
+	}
+	if input.ExpectedContainerID != "" && record.CurrentContainerID != input.ExpectedContainerID {
+		return runtime.ErrConflict
+	}
+	record.Status = input.To
+	record.LastOpAt = input.Now
+	switch input.To {
+	case runtime.StatusStopped:
+		stoppedAt := input.Now
+		record.StoppedAt = &stoppedAt
+	case runtime.StatusRemoved:
+		removedAt := input.Now
+		record.RemovedAt = &removedAt
+		record.CurrentContainerID = ""
+	}
+	s.stored[input.GameID] = record
+	return nil
+}
+
+func (s *fakeRuntimeRecords) ListByStatus(_ context.Context, _ runtime.Status) ([]runtime.RuntimeRecord, error) {
+	return nil, errors.New("not used in restart tests")
+}
+
+func (s *fakeRuntimeRecords) List(_ context.Context) ([]runtime.RuntimeRecord, error) {
+	return nil, errors.New("not used in restart tests")
+}
+
+type fakeOperationLogs struct {
+	mu sync.Mutex
+
+	appendErr error
+	appends   []operation.OperationEntry
+}
+
+func (s *fakeOperationLogs) Append(_ context.Context, entry operation.OperationEntry) (int64, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.appendErr != nil {
+		return 0, s.appendErr
+	}
+	s.appends = append(s.appends, entry)
+	return int64(len(s.appends)), nil
+}
+
+func (s *fakeOperationLogs) ListByGame(_ context.Context, _ string, _ int) ([]operation.OperationEntry, error) {
+	return nil, errors.New("not used in restart tests")
+}
+
+func (s *fakeOperationLogs) byKind(kind operation.OpKind) []operation.OperationEntry {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	out := []operation.OperationEntry{}
+	for _, entry := range s.appends {
+		if entry.OpKind == kind {
+			out = append(out, entry)
+		}
+	}
+	return out
+}
+
+type fakeLeases struct {
+	mu sync.Mutex
+
+	acquired   bool
+	acquireErr error
+	releaseErr error
+
+	acquires []string
+	releases []string
+}
+
+func (l *fakeLeases) TryAcquire(_ context.Context, _, token string, _ time.Duration) (bool, error) {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	l.acquires = append(l.acquires, token)
+	if l.acquireErr != nil {
+		return false, l.acquireErr
+	}
+	return l.acquired, nil
+}
+
+func (l *fakeLeases) Release(_ context.Context, _, token string) error {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	l.releases = append(l.releases, token)
+	return l.releaseErr
+}
+
+type fakeHealthEvents struct {
+	mu sync.Mutex
+
+	publishErr error
+	envelopes  []ports.HealthEventEnvelope
+}
+
+func (h *fakeHealthEvents) Publish(_ context.Context, envelope ports.HealthEventEnvelope) error {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	if h.publishErr != nil {
+		return h.publishErr
+	}
+	h.envelopes = append(h.envelopes, envelope)
+	return nil
+}
+
+type fakeNotifications struct {
+	mu sync.Mutex
+
+	publishErr error
+	intents    []notificationintent.Intent
+}
+
+func (n *fakeNotifications) Publish(_ context.Context, intent notificationintent.Intent) error {
+	n.mu.Lock()
+	defer n.mu.Unlock()
+	if n.publishErr != nil {
+		return n.publishErr
+	}
+	n.intents = append(n.intents, intent)
+	return nil
+}
+
+type fakeLobby struct {
+	record ports.LobbyGameRecord
+	err    error
+}
+
+func (l *fakeLobby) GetGame(_ context.Context, _ string) (ports.LobbyGameRecord, error) {
+	if l.err != nil {
+		return ports.LobbyGameRecord{}, l.err
+	}
+	return l.record, nil
+}
+
+// --- harness ----------------------------------------------------------
+
+type harness struct {
+	records       *fakeRuntimeRecords
+	operationLogs *fakeOperationLogs
+	docker        *mocks.MockDockerClient
+	leases        *fakeLeases
+	healthEvents  *fakeHealthEvents
+	notifications *fakeNotifications
+	lobby         *fakeLobby
+	telemetry     *telemetry.Runtime
+
+	now      time.Time
+	stateDir string
+
+	startService *startruntime.Service
+	stopService  *stopruntime.Service
+}
+
+func newHarness(t *testing.T) *harness {
+	t.Helper()
+	ctrl := gomock.NewController(t)
+	t.Cleanup(ctrl.Finish)
+
+	telemetryRuntime, err := telemetry.NewWithProviders(nil, nil)
+	require.NoError(t, err)
+
+	h := &harness{
+		records:       newFakeRuntimeRecords(),
+		operationLogs: &fakeOperationLogs{},
+		docker:        mocks.NewMockDockerClient(ctrl),
+		leases:        &fakeLeases{acquired: true},
+		healthEvents:  &fakeHealthEvents{},
+		notifications: &fakeNotifications{},
+		lobby:         &fakeLobby{},
+		telemetry:     telemetryRuntime,
+		now:           time.Date(2026, 4, 27, 12, 0, 0, 0, time.UTC),
+		stateDir:      "/var/lib/galaxy/games/game-1",
+	}
+
+	containerCfg := config.ContainerConfig{
+		DefaultCPUQuota:      1.0,
+		DefaultMemory:        "512m",
+		DefaultPIDsLimit:     512,
+		StopTimeout:          30 * time.Second,
+		Retention:            30 * 24 * time.Hour,
+		EngineStateMountPath: "/var/lib/galaxy-game",
+		EngineStateEnvName:   "GAME_STATE_PATH",
+		GameStateDirMode:     0o750,
+		GameStateRoot:        "/var/lib/galaxy/games",
+	}
+	dockerCfg := config.DockerConfig{
+		Host:       "unix:///var/run/docker.sock",
+		Network:    "galaxy-net",
+		LogDriver:  "json-file",
+		PullPolicy: config.ImagePullPolicyIfMissing,
+	}
+	coordinationCfg := config.CoordinationConfig{GameLeaseTTL: time.Minute}
+
+	startService, err := startruntime.NewService(startruntime.Dependencies{
+		RuntimeRecords:  h.records,
+		OperationLogs:   h.operationLogs,
+		Docker:          h.docker,
+		Leases:          h.leases,
+		HealthEvents:    h.healthEvents,
+		Notifications:   h.notifications,
+		Lobby:           h.lobby,
+		Container:       containerCfg,
+		DockerCfg:       dockerCfg,
+		Coordination:    coordinationCfg,
+		Telemetry:       h.telemetry,
+		Clock:           func() time.Time { return h.now },
+		NewToken:        func() string { return "inner-start-token" },
+		PrepareStateDir: func(_ string) (string, error) { return h.stateDir, nil },
+	})
+	require.NoError(t, err)
+	h.startService = startService
+
+	stopService, err := stopruntime.NewService(stopruntime.Dependencies{
+		RuntimeRecords: h.records,
+		OperationLogs:  h.operationLogs,
+		Docker:         h.docker,
+		Leases:         h.leases,
+		HealthEvents:   h.healthEvents,
+		Container:      containerCfg,
+		Coordination:   coordinationCfg,
+		Telemetry:      h.telemetry,
+		Clock:          func() time.Time { return h.now },
+		NewToken:       func() string { return "inner-stop-token" },
+	})
+	require.NoError(t, err)
+	h.stopService = stopService
+
+	return h
+}
+
+func (h *harness) build(t *testing.T, tokens ...string) *restartruntime.Service {
+	t.Helper()
+	tokenIdx := 0
+	tokenGen := func() string {
+		if tokenIdx >= len(tokens) {
+			return "outer-fallback"
+		}
+		t := tokens[tokenIdx]
+		tokenIdx++
+		return t
+	}
+	service, err := restartruntime.NewService(restartruntime.Dependencies{
+		RuntimeRecords: h.records,
+		OperationLogs:  h.operationLogs,
+		Docker:         h.docker,
+		Leases:         h.leases,
+		StopService:    h.stopService,
+		StartService:   h.startService,
+		Coordination:   config.CoordinationConfig{GameLeaseTTL: time.Minute},
+		Telemetry:      h.telemetry,
+		Clock:          func() time.Time { return h.now },
+		NewToken:       tokenGen,
+	})
+	require.NoError(t, err)
+	return service
+}
+
+const imageRef = "registry.example.com/galaxy/game:1.4.7"
+
+func runningRecord(now time.Time) runtime.RuntimeRecord {
+	startedAt := now.Add(-time.Hour)
+	return runtime.RuntimeRecord{
+		GameID:             "game-1",
+		Status:             runtime.StatusRunning,
+		CurrentContainerID: "ctr-old",
+		CurrentImageRef:    imageRef,
+		EngineEndpoint:     "http://galaxy-game-game-1:8080",
+		StatePath:          "/var/lib/galaxy/games/game-1",
+		DockerNetwork:      "galaxy-net",
+		StartedAt:          &startedAt,
+		LastOpAt:           startedAt,
+		CreatedAt:          startedAt,
+	}
+}
+
+func basicInput() restartruntime.Input {
+	return restartruntime.Input{
+		GameID:    "game-1",
+		OpSource:  operation.OpSourceGMRest,
+		SourceRef: "rest-req-42",
+	}
+}
+
+func sampleRunResult(now time.Time) ports.RunResult {
+	return ports.RunResult{
+		ContainerID:    "ctr-new",
+		EngineEndpoint: "http://galaxy-game-game-1:8080",
+		StartedAt:      now,
+	}
+}
+
+func expectInnerStart(h *harness) {
+	h.docker.EXPECT().EnsureNetwork(gomock.Any(), "galaxy-net").Return(nil)
+	h.docker.EXPECT().PullImage(gomock.Any(), imageRef, gomock.Any()).Return(nil)
+	h.docker.EXPECT().InspectImage(gomock.Any(), imageRef).Return(ports.ImageInspect{Ref: imageRef}, nil)
+	h.docker.EXPECT().Run(gomock.Any(), gomock.Any()).Return(sampleRunResult(h.now), nil)
+}
+
+// --- happy path -------------------------------------------------------
+
+func TestHandleRestartFromRunning(t *testing.T) {
+	h := newHarness(t)
+	h.records.stored["game-1"] = runningRecord(h.now)
+
+	h.docker.EXPECT().Stop(gomock.Any(), "ctr-old", 30*time.Second).Return(nil)
+	h.docker.EXPECT().Remove(gomock.Any(), "ctr-old").Return(nil)
+	expectInnerStart(h)
+
+	service := h.build(t, "outer-token")
+	result, err := service.Handle(context.Background(), basicInput())
+	require.NoError(t, err)
+	assert.Equal(t, operation.OutcomeSuccess, result.Outcome)
+	assert.Empty(t, result.ErrorCode)
+	assert.Equal(t, "ctr-new", result.Record.CurrentContainerID)
+	assert.Equal(t, imageRef, result.Record.CurrentImageRef)
+	assert.Equal(t, runtime.StatusRunning, result.Record.Status)
+
+	stops := h.operationLogs.byKind(operation.OpKindStop)
+	starts := h.operationLogs.byKind(operation.OpKindStart)
+	restarts := h.operationLogs.byKind(operation.OpKindRestart)
+	require.Len(t, stops, 1, "inner stop appended its own entry")
+	require.Len(t, starts, 1, "inner start appended its own entry")
+	require.Len(t, restarts, 1, "outer restart appended one summary entry")
+
+	assert.Equal(t, "rest-req-42", stops[0].SourceRef, "correlation id propagated to inner stop")
+	assert.Equal(t, "rest-req-42", starts[0].SourceRef, "correlation id propagated to inner start")
+	assert.Equal(t, "rest-req-42", restarts[0].SourceRef, "correlation id stored on outer restart")
+	assert.Equal(t, "ctr-new", restarts[0].ContainerID)
+	assert.Equal(t, imageRef, restarts[0].ImageRef)
+
+	assert.Equal(t, []string{"outer-token"}, h.leases.acquires)
+	assert.Equal(t, []string{"outer-token"}, h.leases.releases)
+}
+
+func TestHandleRestartFromStopped(t *testing.T) {
+	h := newHarness(t)
+	stoppedRecord := runningRecord(h.now)
+	stoppedRecord.Status = runtime.StatusStopped
+	stoppedAt := h.now.Add(-30 * time.Minute)
+	stoppedRecord.StoppedAt = &stoppedAt
+	h.records.stored["game-1"] = stoppedRecord
+
+	// No docker.Stop because inner stop short-circuits via replay no-op.
+	h.docker.EXPECT().Remove(gomock.Any(), "ctr-old").Return(nil)
+	expectInnerStart(h)
+
+	service := h.build(t, "outer-token")
+	result, err := service.Handle(context.Background(), basicInput())
+	require.NoError(t, err)
+	assert.Equal(t, operation.OutcomeSuccess, result.Outcome)
+	assert.Equal(t, "ctr-new", result.Record.CurrentContainerID)
+}
+
+// --- correlation id fallback -----------------------------------------
+
+func TestHandleGeneratesCorrelationWhenSourceRefEmpty(t *testing.T) {
+	h := newHarness(t)
+	h.records.stored["game-1"] = runningRecord(h.now)
+
+	h.docker.EXPECT().Stop(gomock.Any(), "ctr-old", 30*time.Second).Return(nil)
+	h.docker.EXPECT().Remove(gomock.Any(), "ctr-old").Return(nil)
+	expectInnerStart(h)
+
+	input := basicInput()
+	input.SourceRef = ""
+
+	// First newToken call yields the lease token, second yields the
+	// correlation id fallback.
+	service := h.build(t, "outer-token", "correlation-fallback")
+	result, err := service.Handle(context.Background(), input)
+	require.NoError(t, err)
+	assert.Equal(t, operation.OutcomeSuccess, result.Outcome)
+
+	stops := h.operationLogs.byKind(operation.OpKindStop)
+	starts := h.operationLogs.byKind(operation.OpKindStart)
+	restarts := h.operationLogs.byKind(operation.OpKindRestart)
+	require.Len(t, stops, 1)
+	require.Len(t, starts, 1)
+	require.Len(t, restarts, 1)
+	assert.Equal(t, "correlation-fallback", stops[0].SourceRef)
+	assert.Equal(t, "correlation-fallback", starts[0].SourceRef)
+	assert.Equal(t, "correlation-fallback", restarts[0].SourceRef)
+}
+
+// --- failure paths ---------------------------------------------------
+
+func TestHandleNotFoundForMissingRecord(t *testing.T) {
+	h := newHarness(t)
+	service := h.build(t, "outer-token")
+
+	result, err := service.Handle(context.Background(), basicInput())
+	require.NoError(t, err)
+	assert.Equal(t, operation.OutcomeFailure, result.Outcome)
+	assert.Equal(t, startruntime.ErrorCodeNotFound, result.ErrorCode)
+	assert.Empty(t, h.operationLogs.byKind(operation.OpKindStop))
+	assert.Empty(t, h.operationLogs.byKind(operation.OpKindStart))
+	require.Len(t, h.operationLogs.byKind(operation.OpKindRestart), 1)
+}
+
+func TestHandleConflictForRemovedRecord(t *testing.T) {
+	h := newHarness(t)
+	removed := runningRecord(h.now)
+	removed.Status = runtime.StatusRemoved
+	removed.CurrentContainerID = ""
+	removedAt := h.now.Add(-time.Hour)
+	removed.RemovedAt = &removedAt
+	h.records.stored["game-1"] = removed
+
+	service := h.build(t, "outer-token")
+	result, err := service.Handle(context.Background(), basicInput())
+	require.NoError(t, err)
+	assert.Equal(t, operation.OutcomeFailure, result.Outcome)
+	assert.Equal(t, startruntime.ErrorCodeConflict, result.ErrorCode)
+}
+
+func TestHandleConflictWhenLeaseBusy(t *testing.T) {
+	h := newHarness(t)
+	h.leases.acquired = false
+
+	service := h.build(t, "outer-token")
+	result, err := service.Handle(context.Background(), basicInput())
+	require.NoError(t, err)
+	assert.Equal(t, operation.OutcomeFailure, result.Outcome)
+	assert.Equal(t, startruntime.ErrorCodeConflict, result.ErrorCode)
+	assert.Empty(t, h.leases.releases, "release must not run when acquire returned false")
+}
+
+func TestHandlePropagatesInnerStopFailure(t *testing.T) {
+	h := newHarness(t)
+	h.records.stored["game-1"] = runningRecord(h.now)
+
+	h.docker.EXPECT().Stop(gomock.Any(), "ctr-old", 30*time.Second).Return(errors.New("daemon unreachable"))
+
+	service := h.build(t, "outer-token")
+	result, err := service.Handle(context.Background(), basicInput())
+	require.NoError(t, err)
+	assert.Equal(t, operation.OutcomeFailure, result.Outcome)
+	assert.Equal(t, startruntime.ErrorCodeServiceUnavailable, result.ErrorCode)
+	assert.Contains(t, result.ErrorMessage, "inner stop failed")
+}
+
+func TestHandleServiceUnavailableOnDockerRemoveFailure(t *testing.T) {
+	h := newHarness(t)
+	h.records.stored["game-1"] = runningRecord(h.now)
+
+	h.docker.EXPECT().Stop(gomock.Any(), "ctr-old", 30*time.Second).Return(nil)
+	h.docker.EXPECT().Remove(gomock.Any(), "ctr-old").Return(errors.New("disk i/o"))
+
+	service := h.build(t, "outer-token")
+	result, err := service.Handle(context.Background(), basicInput())
+	require.NoError(t, err)
+	assert.Equal(t, operation.OutcomeFailure, result.Outcome)
+	assert.Equal(t, startruntime.ErrorCodeServiceUnavailable, result.ErrorCode)
+	assert.Contains(t, result.ErrorMessage, "docker remove")
+	// inner stop did succeed and write its log entry; outer restart records failure.
+	require.Len(t, h.operationLogs.byKind(operation.OpKindStop), 1)
+	require.Len(t, h.operationLogs.byKind(operation.OpKindRestart), 1)
+}
+
+func TestHandlePropagatesInnerStartFailure(t *testing.T) {
+	h := newHarness(t)
+	h.records.stored["game-1"] = runningRecord(h.now)
+
+	h.docker.EXPECT().Stop(gomock.Any(), "ctr-old", 30*time.Second).Return(nil)
+	h.docker.EXPECT().Remove(gomock.Any(), "ctr-old").Return(nil)
+	h.docker.EXPECT().EnsureNetwork(gomock.Any(), "galaxy-net").Return(nil)
+	h.docker.EXPECT().PullImage(gomock.Any(), imageRef, gomock.Any()).Return(errors.New("manifest unknown"))
+
+	service := h.build(t, "outer-token")
+	result, err := service.Handle(context.Background(), basicInput())
+	require.NoError(t, err)
+	assert.Equal(t, operation.OutcomeFailure, result.Outcome)
+	assert.Equal(t, startruntime.ErrorCodeImagePullFailed, result.ErrorCode)
+	assert.Contains(t, result.ErrorMessage, "inner start failed")
+}
+
+// --- input validation ------------------------------------------------
+
+func TestHandleRejectsInvalidInput(t *testing.T) {
+	h := newHarness(t)
+	service := h.build(t, "outer-token")
+
+	cases := []restartruntime.Input{
+		{GameID: "", OpSource: operation.OpSourceGMRest},
+		{GameID: "g", OpSource: operation.OpSource("bogus")},
+	}
+	for _, input := range cases {
+		result, err := service.Handle(context.Background(), input)
+		require.NoError(t, err)
+		assert.Equal(t, startruntime.ErrorCodeInvalidRequest, result.ErrorCode)
+	}
+}
+
+// --- constructor -----------------------------------------------------
+
+func TestNewServiceRejectsMissingDependencies(t *testing.T) {
+	h := newHarness(t)
+	deps := restartruntime.Dependencies{
+		Coordination: config.CoordinationConfig{GameLeaseTTL: time.Minute},
+		Telemetry:    h.telemetry,
+	}
+	_, err := restartruntime.NewService(deps)
+	require.Error(t, err)
+}