feat: runtime manager

rtmanager/internal/service/cleanupcontainer/service.go

@@ -0,0 +1,442 @@
+// Package cleanupcontainer implements the `cleanup_container` lifecycle
+// operation owned by Runtime Manager. The service removes the Docker
+// container of an already-stopped runtime and transitions the record
+// to `removed`. It refuses to operate on a still-running runtime —
+// callers must stop first.
+//
+// Two callers exercise this surface: the administrative
+// `DELETE /api/v1/internal/runtimes/{game_id}/container` endpoint, and
+// the periodic container-cleanup worker that walks
+// `runtime_records.status='stopped'` rows older than
+// `RTMANAGER_CONTAINER_RETENTION_DAYS`. Both paths flow through Handle.
+//
+// Lifecycle and failure-mode semantics follow `rtmanager/README.md
+// §Lifecycles → Cleanup`. Design rationale is captured in
+// `rtmanager/docs/services.md`.
+package cleanupcontainer
+
+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/telemetry"
+)
+
+const leaseReleaseTimeout = 5 * time.Second
+
+// Input stores the per-call arguments for one cleanup operation.
+type Input struct {
+	// GameID identifies the platform game whose container is removed.
+	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). Empty for the periodic auto-TTL
+	// caller.
+	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 updated runtime record on success and on
+	// idempotent replay; 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, or
+	// `replay_no_op` on idempotent replay. Empty for fresh successes.
+	ErrorCode string
+
+	// ErrorMessage stores the operator-readable detail on failure.
+	ErrorMessage string
+}
+
+// Dependencies groups the collaborators required by Service.
+type Dependencies struct {
+	RuntimeRecords ports.RuntimeRecordStore
+	OperationLogs  ports.OperationLogStore
+	Docker         ports.DockerClient
+	Leases         ports.GameLeaseStore
+
+	Coordination config.CoordinationConfig
+
+	Telemetry *telemetry.Runtime
+	Logger    *slog.Logger
+	Clock     func() time.Time
+	NewToken  func() string
+}
+
+// Service executes the cleanup_container lifecycle operation.
+type Service struct {
+	runtimeRecords ports.RuntimeRecordStore
+	operationLogs  ports.OperationLogStore
+	docker         ports.DockerClient
+	leases         ports.GameLeaseStore
+
+	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 cleanup container service: nil runtime records")
+	case deps.OperationLogs == nil:
+		return nil, errors.New("new cleanup container service: nil operation logs")
+	case deps.Docker == nil:
+		return nil, errors.New("new cleanup container service: nil docker client")
+	case deps.Leases == nil:
+		return nil, errors.New("new cleanup container service: nil lease store")
+	case deps.Telemetry == nil:
+		return nil, errors.New("new cleanup container service: nil telemetry runtime")
+	}
+	if err := deps.Coordination.Validate(); err != nil {
+		return nil, fmt.Errorf("new cleanup container 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.cleanupcontainer")
+
+	newToken := deps.NewToken
+	if newToken == nil {
+		newToken = defaultTokenGenerator()
+	}
+
+	return &Service{
+		runtimeRecords: deps.RuntimeRecords,
+		operationLogs:  deps.OperationLogs,
+		docker:         deps.Docker,
+		leases:         deps.Leases,
+		leaseTTL:       deps.Coordination.GameLeaseTTL,
+		telemetry:      deps.Telemetry,
+		logger:         logger,
+		clock:          clock,
+		newToken:       newToken,
+	}, nil
+}
+
+// Handle executes one cleanup operation end-to-end. The Go-level error
+// return is reserved for non-business failures (nil context, nil
+// receiver). Every business outcome — success, idempotent replay, or
+// any of the stable failure modes — flows through Result.
+func (service *Service) Handle(ctx context.Context, input Input) (Result, error) {
+	if service == nil {
+		return Result{}, errors.New("cleanup container: nil service")
+	}
+	if ctx == nil {
+		return Result{}, errors.New("cleanup container: 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 cleanup steps.
+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
+	}
+
+	switch existing.Status {
+	case runtime.StatusRemoved:
+		return service.recordReplayNoOp(ctx, opStartedAt, input, existing), nil
+	case runtime.StatusRunning:
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeConflict,
+			errorMessage: fmt.Sprintf("runtime for game %q is running; stop the runtime first", input.GameID),
+			containerID:  existing.CurrentContainerID,
+			imageRef:     existing.CurrentImageRef,
+		}), nil
+	case runtime.StatusStopped:
+		// proceed
+	default:
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeInternal,
+			errorMessage: fmt.Sprintf("runtime record has unsupported status %q", existing.Status),
+		}), nil
+	}
+
+	if existing.CurrentContainerID != "" {
+		if err := service.docker.Remove(ctx, existing.CurrentContainerID); err != nil {
+			return service.recordFailure(ctx, failureCtx{
+				opStartedAt:  opStartedAt,
+				input:        input,
+				errorCode:    startruntime.ErrorCodeServiceUnavailable,
+				errorMessage: fmt.Sprintf("docker remove: %s", err.Error()),
+				containerID:  existing.CurrentContainerID,
+				imageRef:     existing.CurrentImageRef,
+			}), nil
+		}
+	}
+
+	updateNow := service.clock().UTC()
+	err = service.runtimeRecords.UpdateStatus(ctx, ports.UpdateStatusInput{
+		GameID:              input.GameID,
+		ExpectedFrom:        runtime.StatusStopped,
+		ExpectedContainerID: existing.CurrentContainerID,
+		To:                  runtime.StatusRemoved,
+		Now:                 updateNow,
+	})
+	if errors.Is(err, runtime.ErrConflict) {
+		// CAS race: another caller (reconciler dispose, concurrent admin)
+		// already moved the record. The desired terminal state was
+		// reached by another path.
+		return service.recordReplayNoOp(ctx, opStartedAt, input, existing), nil
+	}
+	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 vanished mid-cleanup", input.GameID),
+			containerID:  existing.CurrentContainerID,
+			imageRef:     existing.CurrentImageRef,
+		}), nil
+	}
+	if err != nil {
+		return service.recordFailure(ctx, failureCtx{
+			opStartedAt:  opStartedAt,
+			input:        input,
+			errorCode:    startruntime.ErrorCodeInternal,
+			errorMessage: fmt.Sprintf("update runtime status: %s", err.Error()),
+			containerID:  existing.CurrentContainerID,
+			imageRef:     existing.CurrentImageRef,
+		}), nil
+	}
+
+	finishedAt := service.clock().UTC()
+	service.bestEffortAppend(ctx, operation.OperationEntry{
+		GameID:      input.GameID,
+		OpKind:      operation.OpKindCleanupContainer,
+		OpSource:    input.OpSource,
+		SourceRef:   input.SourceRef,
+		ImageRef:    existing.CurrentImageRef,
+		ContainerID: existing.CurrentContainerID,
+		Outcome:     operation.OutcomeSuccess,
+		StartedAt:   opStartedAt,
+		FinishedAt:  &finishedAt,
+	})
+	service.telemetry.RecordCleanupOutcome(ctx, string(operation.OutcomeSuccess), string(input.OpSource))
+
+	record := existing
+	record.Status = runtime.StatusRemoved
+	record.CurrentContainerID = ""
+	removedAt := updateNow
+	record.RemovedAt = &removedAt
+	record.LastOpAt = updateNow
+
+	logArgs := []any{
+		"game_id", input.GameID,
+		"container_id", existing.CurrentContainerID,
+		"op_source", string(input.OpSource),
+	}
+	logArgs = append(logArgs, logging.ContextAttrs(ctx)...)
+	service.logger.InfoContext(ctx, "runtime container cleaned up", logArgs...)
+
+	return Result{
+		Record:  record,
+		Outcome: operation.OutcomeSuccess,
+	}, nil
+}
+
+// recordReplayNoOp records the idempotent replay outcome and returns the
+// existing record unchanged.
+func (service *Service) recordReplayNoOp(ctx context.Context, opStartedAt time.Time, input Input, existing runtime.RuntimeRecord) Result {
+	finishedAt := service.clock().UTC()
+	service.bestEffortAppend(ctx, operation.OperationEntry{
+		GameID:      input.GameID,
+		OpKind:      operation.OpKindCleanupContainer,
+		OpSource:    input.OpSource,
+		SourceRef:   input.SourceRef,
+		ImageRef:    existing.CurrentImageRef,
+		ContainerID: existing.CurrentContainerID,
+		Outcome:     operation.OutcomeSuccess,
+		ErrorCode:   startruntime.ErrorCodeReplayNoOp,
+		StartedAt:   opStartedAt,
+		FinishedAt:  &finishedAt,
+	})
+	service.telemetry.RecordCleanupOutcome(ctx, string(operation.OutcomeSuccess), string(input.OpSource))
+
+	logArgs := []any{
+		"game_id", input.GameID,
+		"container_id", existing.CurrentContainerID,
+		"op_source", string(input.OpSource),
+	}
+	logArgs = append(logArgs, logging.ContextAttrs(ctx)...)
+	service.logger.InfoContext(ctx, "runtime cleanup replay no-op", logArgs...)
+
+	return Result{
+		Record:    existing,
+		Outcome:   operation.OutcomeSuccess,
+		ErrorCode: startruntime.ErrorCodeReplayNoOp,
+	}
+}
+
+// failureCtx groups the inputs to recordFailure.
+type failureCtx struct {
+	opStartedAt  time.Time
+	input        Input
+	errorCode    string
+	errorMessage string
+	containerID  string
+	imageRef     string
+}
+
+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.OpKindCleanupContainer,
+		OpSource:     fc.input.OpSource,
+		SourceRef:    fc.input.SourceRef,
+		ImageRef:     fc.imageRef,
+		ContainerID:  fc.containerID,
+		Outcome:      operation.OutcomeFailure,
+		ErrorCode:    fc.errorCode,
+		ErrorMessage: fc.errorMessage,
+		StartedAt:    fc.opStartedAt,
+		FinishedAt:   &finishedAt,
+	})
+	service.telemetry.RecordCleanupOutcome(ctx, string(operation.OutcomeFailure), string(fc.input.OpSource))
+
+	logArgs := []any{
+		"game_id", fc.input.GameID,
+		"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 cleanup failed", logArgs...)
+
+	return Result{
+		Outcome:      operation.OutcomeFailure,
+		ErrorCode:    fc.errorCode,
+		ErrorMessage: fc.errorMessage,
+	}
+}
+
+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(),
+		)
+	}
+}
+
+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(),
+		)
+	}
+}
+
+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/cleanupcontainer/service_test.go

@@ -0,0 +1,382 @@
+package cleanupcontainer_test
+
+import (
+	"context"
+	"errors"
+	"sync"
+	"testing"
+	"time"
+
+	"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/cleanupcontainer"
+	"galaxy/rtmanager/internal/service/startruntime"
+	"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
+	updateStatusErr error
+
+	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, _ runtime.RuntimeRecord) error {
+	return errors.New("not used in cleanup tests")
+}
+
+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
+	if input.To == 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 cleanup tests")
+}
+
+func (s *fakeRuntimeRecords) List(_ context.Context) ([]runtime.RuntimeRecord, error) {
+	return nil, errors.New("not used in cleanup 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 cleanup tests")
+}
+
+func (s *fakeOperationLogs) lastAppend() (operation.OperationEntry, bool) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if len(s.appends) == 0 {
+		return operation.OperationEntry{}, false
+	}
+	return s.appends[len(s.appends)-1], true
+}
+
+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
+}
+
+// --- harness ----------------------------------------------------------
+
+type harness struct {
+	records       *fakeRuntimeRecords
+	operationLogs *fakeOperationLogs
+	docker        *mocks.MockDockerClient
+	leases        *fakeLeases
+
+	telemetry *telemetry.Runtime
+
+	now time.Time
+}
+
+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)
+
+	return &harness{
+		records:       newFakeRuntimeRecords(),
+		operationLogs: &fakeOperationLogs{},
+		docker:        mocks.NewMockDockerClient(ctrl),
+		leases:        &fakeLeases{acquired: true},
+		telemetry:     telemetryRuntime,
+		now:           time.Date(2026, 4, 27, 12, 0, 0, 0, time.UTC),
+	}
+}
+
+func (h *harness) build(t *testing.T) *cleanupcontainer.Service {
+	t.Helper()
+	service, err := cleanupcontainer.NewService(cleanupcontainer.Dependencies{
+		RuntimeRecords: h.records,
+		OperationLogs:  h.operationLogs,
+		Docker:         h.docker,
+		Leases:         h.leases,
+		Coordination:   config.CoordinationConfig{GameLeaseTTL: time.Minute},
+		Telemetry:      h.telemetry,
+		Clock:          func() time.Time { return h.now },
+		NewToken:       func() string { return "token-A" },
+	})
+	require.NoError(t, err)
+	return service
+}
+
+func basicInput() cleanupcontainer.Input {
+	return cleanupcontainer.Input{
+		GameID:    "game-1",
+		OpSource:  operation.OpSourceAdminRest,
+		SourceRef: "rest-cleanup-1",
+	}
+}
+
+func stoppedRecord(now time.Time) runtime.RuntimeRecord {
+	startedAt := now.Add(-2 * time.Hour)
+	stoppedAt := now.Add(-time.Hour)
+	return runtime.RuntimeRecord{
+		GameID:             "game-1",
+		Status:             runtime.StatusStopped,
+		CurrentContainerID: "ctr-old",
+		CurrentImageRef:    "registry.example.com/galaxy/game:1.4.7",
+		EngineEndpoint:     "http://galaxy-game-game-1:8080",
+		StatePath:          "/var/lib/galaxy/games/game-1",
+		DockerNetwork:      "galaxy-net",
+		StartedAt:          &startedAt,
+		StoppedAt:          &stoppedAt,
+		LastOpAt:           stoppedAt,
+		CreatedAt:          startedAt,
+	}
+}
+
+// --- happy path -----------------------------------------------------
+
+func TestHandleCleanupHappyPath(t *testing.T) {
+	h := newHarness(t)
+	h.records.stored["game-1"] = stoppedRecord(h.now)
+
+	h.docker.EXPECT().Remove(gomock.Any(), "ctr-old").Return(nil)
+
+	service := h.build(t)
+	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, runtime.StatusRemoved, result.Record.Status)
+	assert.Empty(t, result.Record.CurrentContainerID)
+
+	require.Len(t, h.records.updates, 1)
+	assert.Equal(t, runtime.StatusStopped, h.records.updates[0].ExpectedFrom)
+	assert.Equal(t, runtime.StatusRemoved, h.records.updates[0].To)
+
+	require.Len(t, h.operationLogs.appends, 1)
+	last, _ := h.operationLogs.lastAppend()
+	assert.Equal(t, operation.OpKindCleanupContainer, last.OpKind)
+	assert.Equal(t, operation.OutcomeSuccess, last.Outcome)
+	assert.Empty(t, last.ErrorCode)
+}
+
+// --- replay ---------------------------------------------------------
+
+func TestHandleReplayNoOpForRemovedRecord(t *testing.T) {
+	h := newHarness(t)
+	removed := stoppedRecord(h.now)
+	removed.Status = runtime.StatusRemoved
+	removed.CurrentContainerID = ""
+	removedAt := h.now.Add(-30 * time.Minute)
+	removed.RemovedAt = &removedAt
+	h.records.stored["game-1"] = removed
+
+	service := h.build(t)
+	result, err := service.Handle(context.Background(), basicInput())
+	require.NoError(t, err)
+	assert.Equal(t, operation.OutcomeSuccess, result.Outcome)
+	assert.Equal(t, startruntime.ErrorCodeReplayNoOp, result.ErrorCode)
+	assert.Empty(t, h.records.updates)
+
+	last, _ := h.operationLogs.lastAppend()
+	assert.Equal(t, startruntime.ErrorCodeReplayNoOp, last.ErrorCode)
+}
+
+func TestHandleReplayNoOpOnUpdateStatusConflict(t *testing.T) {
+	h := newHarness(t)
+	h.records.stored["game-1"] = stoppedRecord(h.now)
+	h.records.updateStatusErr = runtime.ErrConflict
+
+	h.docker.EXPECT().Remove(gomock.Any(), "ctr-old").Return(nil)
+
+	service := h.build(t)
+	result, err := service.Handle(context.Background(), basicInput())
+	require.NoError(t, err)
+	assert.Equal(t, operation.OutcomeSuccess, result.Outcome)
+	assert.Equal(t, startruntime.ErrorCodeReplayNoOp, result.ErrorCode)
+}
+
+// --- failure paths --------------------------------------------------
+
+func TestHandleConflictOnRunningRecord(t *testing.T) {
+	h := newHarness(t)
+	running := stoppedRecord(h.now)
+	running.Status = runtime.StatusRunning
+	startedAt := h.now.Add(-time.Hour)
+	running.StartedAt = &startedAt
+	running.StoppedAt = nil
+	h.records.stored["game-1"] = running
+
+	service := h.build(t)
+	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.Contains(t, result.ErrorMessage, "stop the runtime first")
+}
+
+func TestHandleNotFoundForMissingRecord(t *testing.T) {
+	h := newHarness(t)
+
+	service := h.build(t)
+	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)
+}
+
+func TestHandleServiceUnavailableOnDockerRemoveFailure(t *testing.T) {
+	h := newHarness(t)
+	h.records.stored["game-1"] = stoppedRecord(h.now)
+
+	h.docker.EXPECT().Remove(gomock.Any(), "ctr-old").Return(errors.New("disk i/o"))
+
+	service := h.build(t)
+	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.Empty(t, h.records.updates, "no record mutation on docker remove failure")
+}
+
+func TestHandleInternalErrorOnGenericUpdateError(t *testing.T) {
+	h := newHarness(t)
+	h.records.stored["game-1"] = stoppedRecord(h.now)
+	h.records.updateStatusErr = errors.New("postgres down")
+
+	h.docker.EXPECT().Remove(gomock.Any(), "ctr-old").Return(nil)
+
+	service := h.build(t)
+	result, err := service.Handle(context.Background(), basicInput())
+	require.NoError(t, err)
+	assert.Equal(t, operation.OutcomeFailure, result.Outcome)
+	assert.Equal(t, startruntime.ErrorCodeInternal, result.ErrorCode)
+}
+
+func TestHandleConflictWhenLeaseBusy(t *testing.T) {
+	h := newHarness(t)
+	h.leases.acquired = false
+
+	service := h.build(t)
+	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)
+}
+
+// --- input validation ----------------------------------------------
+
+func TestHandleRejectsInvalidInput(t *testing.T) {
+	h := newHarness(t)
+	service := h.build(t)
+
+	cases := []cleanupcontainer.Input{
+		{GameID: "", OpSource: operation.OpSourceAdminRest},
+		{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 := cleanupcontainer.Dependencies{
+		Coordination: config.CoordinationConfig{GameLeaseTTL: time.Minute},
+		Telemetry:    h.telemetry,
+	}
+	_, err := cleanupcontainer.NewService(deps)
+	require.Error(t, err)
+}