Checkpoint speaker transcriptions incrementally

docs/roadmap/implementation.md

@@ -1,6 +1,6 @@
 # Implementation Plan: Speaker Checkpoints and Replayable Uploads
 
-Status: Implementation in progress; Stages 1-2 complete.
+Status: Implementation in progress; Stages 1-3 complete.
 
 ## Scope and Instructions
 
@@ -186,7 +186,17 @@ work, without altering existing runtime behavior prematurely.
 
 ## Stage 3: Execute and Checkpoint Speakers Incrementally
 
-Status: Pending.
+Status: Complete.
+
+Completion marker (2026-09-21): Connected reconciliation to bounded execution,
+durably checkpointed the starting and per-speaker collections through detached
+manifest candidates, preserved committed work across failures and cancellation,
+rebuilt deterministic aggregate outputs from reusable evidence, and added
+speaker-aware resume validation. Updated production-style fixtures and covered
+serialized concurrent saves, force/retry replacement, input rechecks, save
+failures, partial recovery, and aggregate reconstruction. Verified with
+`go test ./internal/stage ./internal/app`, `go test -race ./internal/stage`, and
+`git diff --check`.
 
 Preparation: `internal/stage/transcribe.go`, `analyze_execution_state.go`,
 `run_local.go`,

internal/app/runner_test.go

@@ -1082,7 +1082,7 @@ func TestExecuteStagesReportsSkippedStage(t *testing.T) {
 		t.Fatalf("Save manifest error = %v", err)
 	}
 
-	summary, err := executeStages(context.Background(), cfg, []stage.Stage{BuildFullPlan()[1]}, RunOptions{})
+	summary, err := executeStages(context.Background(), cfg, []stage.Stage{resultStage{name: "transcribe"}}, RunOptions{})
 	if err != nil {
 		t.Fatalf("executeStages() error = %v", err)
 	}
@@ -1302,7 +1302,7 @@ func TestExecuteStagesSkippedStagePreservesExistingOutputsProvenance(t *testing.
 		t.Fatalf("Save manifest error = %v", err)
 	}
 
-	summary, err := executeStages(context.Background(), cfg, []stage.Stage{BuildFullPlan()[1]}, RunOptions{})
+	summary, err := executeStages(context.Background(), cfg, []stage.Stage{resultStage{name: "transcribe"}}, RunOptions{})
 	if err != nil {
 		t.Fatalf("executeStages() error = %v", err)
 	}

internal/app/runner_test_helpers_test.go

@@ -2,10 +2,17 @@ package app
 
 import (
 	"context"
+	"crypto/sha256"
+	"encoding/hex"
+	"os"
+	"path/filepath"
+	"strings"
 	"testing"
 
+	"gitea.maximumdirect.net/eric/narratio/internal/artifacts"
 	"gitea.maximumdirect.net/eric/narratio/internal/config"
 	"gitea.maximumdirect.net/eric/narratio/internal/manifest"
+	"gitea.maximumdirect.net/eric/narratio/internal/pathsafe"
 	"gitea.maximumdirect.net/eric/narratio/internal/stage"
 )
 
@@ -38,8 +45,88 @@ func seedCurrentSemanticEvidence(t *testing.T, cfg *config.Config, m *manifest.M
 			t.Fatalf("fingerprint stage %q: %v", name, err)
 		}
 		record.SemanticConfig = &fingerprint
+		if name == "transcribe" {
+			seedCurrentTranscribeEvidence(t, cfg, m, fingerprint)
 		}
 	}
+}
+
+func seedCurrentTranscribeEvidence(
+	t *testing.T,
+	cfg *config.Config,
+	m *manifest.Manifest,
+	semantic manifest.SemanticConfigFingerprint,
+) {
+	t.Helper()
+	store := artifacts.NewLocalStore(cfg.Pipeline.Workspace.Root)
+	paths := store.SessionPathsFor(cfg.Session.Campaign, m.SessionID)
+	if strings.TrimSpace(m.RunID) == "" {
+		m.RunID = "20260503T010203Z-a1b2c3d4"
+	}
+	records := make(map[string]manifest.TranscribeSpeakerRecord)
+	for _, input := range m.Inputs {
+		if input.Kind != "audio" {
+			continue
+		}
+		audioPath := artifacts.ResolveSessionLocalPathForRead(paths, input.Path)
+		audioBytes, err := os.ReadFile(audioPath)
+		if err != nil {
+			t.Fatalf("read transcribe audio fixture %q: %v", audioPath, err)
+		}
+		audioRelative, err := pathsafe.SlashRelativeFromRoot(paths.Root, audioPath)
+		if err != nil {
+			t.Fatalf("relativize transcribe audio fixture %q: %v", audioPath, err)
+		}
+		speaker := strings.TrimSuffix(filepath.Base(audioPath), filepath.Ext(audioPath))
+		outputRelative := filepath.ToSlash(filepath.Join("transcripts", "raw", speaker+".json"))
+		outputPath, err := pathsafe.JoinSlashRelativeUnderRoot(paths.Root, outputRelative)
+		if err != nil {
+			t.Fatalf("resolve transcribe output fixture: %v", err)
+		}
+		if _, err := os.Stat(outputPath); os.IsNotExist(err) {
+			mustWriteTestFile(t, outputPath, `{"segments":[]}`)
+		} else if err != nil {
+			t.Fatalf("inspect transcribe output fixture %q: %v", outputPath, err)
+		}
+		outputBytes, err := os.ReadFile(outputPath)
+		if err != nil {
+			t.Fatalf("read transcribe output fixture %q: %v", outputPath, err)
+		}
+		audioDigest := sha256.Sum256(audioBytes)
+		outputDigest := sha256.Sum256(outputBytes)
+		fingerprint, err := stage.FingerprintSemanticConfig(
+			manifest.TranscribeFingerprintContractVersion,
+			struct {
+				SpeakerID             string `json:"speaker_id"`
+				AudioPath             string `json:"audio_path"`
+				AudioSHA256           string `json:"audio_sha256"`
+				AudioSize             int64  `json:"audio_size"`
+				SemanticConfigVersion int    `json:"semantic_config_version"`
+				SemanticConfigDigest  string `json:"semantic_config_digest"`
+			}{
+				SpeakerID: speaker, AudioPath: audioRelative,
+				AudioSHA256: hex.EncodeToString(audioDigest[:]), AudioSize: int64(len(audioBytes)),
+				SemanticConfigVersion: semantic.Version, SemanticConfigDigest: semantic.Digest,
+			},
+		)
+		if err != nil {
+			t.Fatalf("fingerprint transcribe speaker fixture %q: %v", speaker, err)
+		}
+		output := manifest.ArtifactRecord{
+			Kind: "transcript_raw", LocalPath: outputRelative,
+			ProducerRunID: m.RunID, Checksum: hex.EncodeToString(outputDigest[:]),
+		}
+		records[speaker] = manifest.TranscribeSpeakerRecord{
+			FingerprintVersion: manifest.TranscribeFingerprintContractVersion,
+			Fingerprint:        fingerprint.Digest, AudioPath: audioRelative,
+			AudioSHA256: hex.EncodeToString(audioDigest[:]), AudioSize: int64(len(audioBytes)),
+			Output: output, OutputSize: int64(len(outputBytes)), ProducerRunID: m.RunID,
+		}
+	}
+	record := m.Stages["transcribe"]
+	record.TranscribeStateVersion = manifest.TranscribeStateContractVersion
+	record.TranscribeSpeakers = records
+}
 
 func loadConfigForSemanticEvidence(t *testing.T, pipelinePath, campaignPath, sessionPath string) *config.Config {
 	t.Helper()

internal/stage/transcribe.go

@@ -8,7 +8,6 @@ import (
 	"path/filepath"
 	"sort"
 	"strings"
-	"sync"
 
 	"golang.org/x/sync/errgroup"
 
@@ -28,11 +27,14 @@ func (transcribeStage) Run(ctx context.Context, env *Env, m *manifest.Manifest)
 	if env.ArtifactStore == nil {
 		return nil, fmt.Errorf("transcribe: artifact store is required")
 	}
+	if env.ManifestStore == nil {
+		return nil, fmt.Errorf("transcribe: manifest store is required")
+	}
 	if env.Config.Pipeline == nil || env.Config.Session == nil {
 		return nil, fmt.Errorf("transcribe: resolved config must include pipeline and session")
 	}
-	if env.WhisperX == nil {
+	if m == nil {
-		return nil, fmt.Errorf("transcribe: whisperx adapter is required")
+		return nil, fmt.Errorf("transcribe: session manifest is required")
 	}
 
 	var sessionID string
@@ -47,37 +49,40 @@ func (transcribeStage) Run(ctx context.Context, env *Env, m *manifest.Manifest)
 	}
 
 	paths := sessionPathsForEnv(env, sessionID)
+	reconciliation, err := reconcileTranscribeSpeakers(ctx, env, m)
+	if err != nil {
+		return nil, fmt.Errorf("transcribe: reconcile speakers: %w", err)
+	}
+	state := newTranscribeExecutionState(env, m, paths, reconciliation.current)
+	if err := state.checkpointStartingCollection(ctx); err != nil {
+		return nil, fmt.Errorf("transcribe: checkpoint reconciled speaker collection before execution: %w", err)
+	}
+
+	jobs := append([]transcribeReconciliationJob(nil), reconciliation.work...)
+	if len(jobs) > 0 && env.WhisperX == nil {
+		return nil, fmt.Errorf("transcribe: whisperx adapter is required")
+	}
+	if len(jobs) > 0 {
 		runLayout, err := resolveRunStageLayout(env, m, paths, sessionID, "transcribe")
 		if err != nil {
 			return nil, fmt.Errorf("transcribe: resolve run-stage layout: %w", err)
 		}
-	jobs, err := buildTranscribeJobs(m, paths)
-	if err != nil {
-		return nil, fmt.Errorf("transcribe: resolve audio inputs: %w", err)
-	}
-	if len(jobs) == 0 {
-		return nil, fmt.Errorf("transcribe: no prepared audio files found")
-	}
 		for index := range jobs {
 			jobs[index].runOutputPath, err = runLocalPathForCanonical(runLayout, paths, jobs[index].canonicalOutputPath)
 			if err != nil {
 				return nil, fmt.Errorf("transcribe: resolve run-local output path for %q: %w", jobs[index].speakerID, err)
 			}
 		}
+	}
 
 	concurrency := 1
 	if env.Config.Pipeline.WhisperX.Concurrency != nil && *env.Config.Pipeline.WhisperX.Concurrency > 0 {
 		concurrency = *env.Config.Pipeline.WhisperX.Concurrency
 	}
-	if concurrency > len(jobs) {
+	if len(jobs) > 0 && concurrency > len(jobs) {
 		concurrency = len(jobs)
 	}
 
-	var (
-		mu        sync.Mutex
-		completed int
-		perFile   = map[string]map[string]any{}
-	)
 	group, groupCtx := errgroup.WithContext(ctx)
 	group.SetLimit(concurrency)
 	for _, job := range jobs {
@@ -86,6 +91,9 @@ func (transcribeStage) Run(ctx context.Context, env *Env, m *manifest.Manifest)
 			if err := groupCtx.Err(); err != nil {
 				return err
 			}
+			if err := state.checkpointFailure(); err != nil {
+				return err
+			}
 			res, runErr := env.WhisperX.Transcribe(groupCtx, whisperx.TranscribeRequest{
 				SpeakerID:               job.speakerID,
 				AudioPath:               job.audioPath,
@@ -95,13 +103,6 @@ func (transcribeStage) Run(ctx context.Context, env *Env, m *manifest.Manifest)
 				return fmt.Errorf("speaker %q (%s): %w", job.speakerID, filepath.Base(job.audioPath), runErr)
 			}
 
-			if _, err := authoritativeOutputPath(job.runOutputPath, res.OutputRawTranscriptPath); err != nil {
-				return fmt.Errorf("speaker %q: %w", job.speakerID, err)
-			}
-			if err := validateTranscriptJSONFile(job.runOutputPath); err != nil {
-				return fmt.Errorf("speaker %q output %q invalid: %w", job.speakerID, job.runOutputPath, err)
-			}
-
 			meta := map[string]any{
 				"attempts":    res.Attempts,
 				"http_status": res.HTTPStatus,
@@ -109,17 +110,17 @@ func (transcribeStage) Run(ctx context.Context, env *Env, m *manifest.Manifest)
 				"output_path": job.runOutputPath,
 			}
 
-			mu.Lock()
+			if err := state.acceptSpeaker(ctx, job, res.OutputRawTranscriptPath, meta); err != nil {
-			perFile[job.speakerID] = meta
+				return fmt.Errorf("speaker %q: %w", job.speakerID, err)
-			completed++
+			}
-			mu.Unlock()
 			return nil
 		})
 	}
 	groupErr := group.Wait()
-	mu.Lock()
+	if err := state.checkpointFailure(); err != nil {
-	completedCount := completed
+		return nil, err
-	mu.Unlock()
+	}
+	completedCount := state.completedCount()
 	if err := ctx.Err(); err != nil {
 		return nil, fmt.Errorf("transcribe: canceled after completing %d of %d audio jobs: %w", completedCount, len(jobs), err)
 	}
@@ -129,34 +130,9 @@ func (transcribeStage) Run(ctx context.Context, env *Env, m *manifest.Manifest)
 	if completedCount != len(jobs) {
 		return nil, fmt.Errorf("transcribe: completed %d of %d planned audio jobs", completedCount, len(jobs))
 	}
-
+	outputs, outputPaths, runOutputPaths, orderedPerFile, err := state.successDetails()
-	speakers := make([]string, 0, len(perFile))
-	for speaker := range perFile {
-		speakers = append(speakers, speaker)
-	}
-	sort.Strings(speakers)
-
-	outputs := make([]artifacts.Ref, 0, len(speakers))
-	runOutputPaths := make([]string, 0, len(speakers))
-	outputPaths := make([]string, 0, len(speakers))
-	orderedPerFile := make(map[string]any, len(speakers))
-	for _, speaker := range speakers {
-		var selected transcribeJob
-		for _, job := range jobs {
-			if job.speakerID == speaker {
-				selected = job
-				break
-			}
-		}
-		runOutputPaths = append(runOutputPaths, selected.runOutputPath)
-		ref := artifacts.Ref{Kind: "transcript_raw", Category: "transcripts", SessionID: sessionID, AbsolutePath: selected.runOutputPath}
-		materialized, err := materializeRunLocalOutput(env.ArtifactStore, selected.runOutputPath, selected.canonicalOutputPath, ref)
 	if err != nil {
-			return nil, fmt.Errorf("transcribe: materialize %q output: %w", speaker, err)
+		return nil, fmt.Errorf("transcribe: assemble completed outputs: %w", err)
-		}
-		outputs = append(outputs, materialized)
-		outputPaths = append(outputPaths, selected.canonicalOutputPath)
-		orderedPerFile[speaker] = perFile[speaker]
 	}
 
 	retries := 0
@@ -168,7 +144,7 @@ func (transcribeStage) Run(ctx context.Context, env *Env, m *manifest.Manifest)
 		Outputs: outputs,
 		Metadata: map[string]any{
 			"stage":             "transcribe",
-			"audio_files_count": len(jobs),
+			"audio_files_count": len(reconciliation.jobs),
 			"language":          env.Config.Pipeline.WhisperX.Language,
 			"concurrency":       concurrency,
 			"retries":           retries,

internal/stage/transcribe_checkpoint_test.go

@@ -0,0 +1,299 @@
+package stage
+
+import (
+	"context"
+	"errors"
+	"os"
+	"slices"
+	"sync/atomic"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/narratio/internal/adapters/whisperx"
+	"gitea.maximumdirect.net/eric/narratio/internal/manifest"
+)
+
+func TestTranscribeFailureCheckpointsEarlierSpeakersAndRetryRunsOnlyRemaining(t *testing.T) {
+	env, m := setupTranscribeEnv(t, []string{"a.flac", "b.flac", "c.flac", "d.flac"})
+	concurrency := 1
+	env.Config.Pipeline.WhisperX.Concurrency = &concurrency
+	if _, err := (prepareStage{}).Run(context.Background(), env, m); err != nil {
+		t.Fatalf("prepare.Run() error = %v", err)
+	}
+	cause := errors.New("speaker c failed")
+	env.WhisperX = &whisperx.FakeClient{TranscribeFn: func(
+		_ context.Context,
+		req whisperx.TranscribeRequest,
+	) (whisperx.TranscribeResult, error) {
+		if req.SpeakerID == "c" {
+			return whisperx.TranscribeResult{}, cause
+		}
+		if req.SpeakerID == "d" {
+			t.Fatalf("speaker d was dispatched after failure")
+		}
+		if err := writeJSONFile(req.OutputRawTranscriptPath, map[string]any{"speaker": req.SpeakerID}); err != nil {
+			return whisperx.TranscribeResult{}, err
+		}
+		return whisperx.TranscribeResult{OutputRawTranscriptPath: req.OutputRawTranscriptPath}, nil
+	}}
+
+	if _, err := (transcribeStage{}).Run(context.Background(), env, m); !errors.Is(err, cause) {
+		t.Fatalf("Run() error = %v, want %v", err, cause)
+	}
+	durable, err := env.ManifestStore.Load(context.Background(), sessionPathsForEnv(env, m.SessionID).ManifestPath)
+	if err != nil {
+		t.Fatalf("load durable manifest: %v", err)
+	}
+	if got := sortedTranscribeSpeakerKeys(durable.Stages["transcribe"].TranscribeSpeakers); !slices.Equal(got, []string{"a", "b"}) {
+		t.Fatalf("durable speakers = %#v, want a and b", got)
+	}
+
+	retry := &whisperx.FakeClient{}
+	env.WhisperX = retry
+	result, err := (transcribeStage{}).Run(context.Background(), env, m)
+	if err != nil {
+		t.Fatalf("retry Run() error = %v", err)
+	}
+	requests := retry.RequestsSnapshot()
+	gotRequests := make([]string, 0, len(requests))
+	for _, request := range requests {
+		gotRequests = append(gotRequests, request.SpeakerID)
+	}
+	if !slices.Equal(gotRequests, []string{"c", "d"}) {
+		t.Fatalf("retry speakers = %#v, want c and d", gotRequests)
+	}
+	if len(result.Outputs) != 4 {
+		t.Fatalf("retry outputs = %d, want all four speakers", len(result.Outputs))
+	}
+}
+
+func TestTranscribeConcurrentCheckpointsDoNotLoseCompletions(t *testing.T) {
+	env, m := setupTranscribeEnv(t, []string{"a.flac", "b.flac", "c.flac", "d.flac"})
+	concurrency := 4
+	env.Config.Pipeline.WhisperX.Concurrency = &concurrency
+	env.WhisperX = &whisperx.FakeClient{}
+	if _, err := (prepareStage{}).Run(context.Background(), env, m); err != nil {
+		t.Fatalf("prepare.Run() error = %v", err)
+	}
+	if _, err := (transcribeStage{}).Run(context.Background(), env, m); err != nil {
+		t.Fatalf("Run() error = %v", err)
+	}
+	durable, err := env.ManifestStore.Load(context.Background(), sessionPathsForEnv(env, m.SessionID).ManifestPath)
+	if err != nil {
+		t.Fatalf("load durable manifest: %v", err)
+	}
+	if got := sortedTranscribeSpeakerKeys(durable.Stages["transcribe"].TranscribeSpeakers); !slices.Equal(got, []string{"a", "b", "c", "d"}) {
+		t.Fatalf("durable speakers = %#v, want all completions", got)
+	}
+}
+
+func TestTranscribeStartingCheckpointFailurePreventsRequestsAndCanonicalWrites(t *testing.T) {
+	env, m := setupTranscribeEnv(t, []string{"alice.flac"})
+	if _, err := (prepareStage{}).Run(context.Background(), env, m); err != nil {
+		t.Fatalf("prepare.Run() error = %v", err)
+	}
+	cause := errors.New("starting checkpoint failed")
+	base := env.ManifestStore
+	env.ManifestStore = &failingTranscribeManifestStore{Store: base, FailAt: 1, Err: cause}
+	fake := &whisperx.FakeClient{}
+	env.WhisperX = fake
+
+	if _, err := (transcribeStage{}).Run(context.Background(), env, m); !errors.Is(err, cause) {
+		t.Fatalf("Run() error = %v, want %v", err, cause)
+	}
+	if got := len(fake.RequestsSnapshot()); got != 0 {
+		t.Fatalf("adapter requests = %d, want none", got)
+	}
+	canonical := sessionPathsForEnv(env, m.SessionID).TranscriptsRawDir + "/alice.json"
+	if _, err := os.Stat(canonical); !os.IsNotExist(err) {
+		t.Fatalf("canonical output exists after starting checkpoint failure: %v", err)
+	}
+}
+
+func TestTranscribeCheckpointFailureRetainsEarlierAuthority(t *testing.T) {
+	env, m := setupTranscribeEnv(t, []string{"a.flac", "b.flac", "c.flac"})
+	concurrency := 1
+	env.Config.Pipeline.WhisperX.Concurrency = &concurrency
+	if _, err := (prepareStage{}).Run(context.Background(), env, m); err != nil {
+		t.Fatalf("prepare.Run() error = %v", err)
+	}
+	cause := errors.New("speaker checkpoint failed")
+	base := env.ManifestStore
+	env.ManifestStore = &failingTranscribeManifestStore{Store: base, FailAt: 3, Err: cause}
+	fake := &whisperx.FakeClient{}
+	env.WhisperX = fake
+
+	if _, err := (transcribeStage{}).Run(context.Background(), env, m); !errors.Is(err, cause) {
+		t.Fatalf("Run() error = %v, want %v", err, cause)
+	}
+	requests := fake.RequestsSnapshot()
+	if len(requests) != 2 {
+		t.Fatalf("adapter requests = %d, want a and b only", len(requests))
+	}
+	durable, err := base.Load(context.Background(), sessionPathsForEnv(env, m.SessionID).ManifestPath)
+	if err != nil {
+		t.Fatalf("load durable manifest: %v", err)
+	}
+	if got := sortedTranscribeSpeakerKeys(durable.Stages["transcribe"].TranscribeSpeakers); !slices.Equal(got, []string{"a"}) {
+		t.Fatalf("durable speakers = %#v, want only a", got)
+	}
+	if got := sortedTranscribeSpeakerKeys(m.Stages["transcribe"].TranscribeSpeakers); !slices.Equal(got, []string{"a"}) {
+		t.Fatalf("live speakers = %#v, want last committed collection", got)
+	}
+}
+
+func TestTranscribeFailedForceRetryDoesNotResurrectOldSpeakers(t *testing.T) {
+	env, m := setupTranscribeEnv(t, []string{"a.flac", "b.flac"})
+	concurrency := 1
+	env.Config.Pipeline.WhisperX.Concurrency = &concurrency
+	env.WhisperX = &whisperx.FakeClient{}
+	if _, err := (prepareStage{}).Run(context.Background(), env, m); err != nil {
+		t.Fatalf("prepare.Run() error = %v", err)
+	}
+	if _, err := (transcribeStage{}).Run(context.Background(), env, m); err != nil {
+		t.Fatalf("seed Run() error = %v", err)
+	}
+
+	m.RunID = "20260503T020304Z-b2c3d4e5"
+	env.Force = true
+	cause := errors.New("forced b failed")
+	env.WhisperX = &whisperx.FakeClient{TranscribeFn: func(
+		_ context.Context,
+		req whisperx.TranscribeRequest,
+	) (whisperx.TranscribeResult, error) {
+		if req.SpeakerID == "b" {
+			return whisperx.TranscribeResult{}, cause
+		}
+		if err := writeJSONFile(req.OutputRawTranscriptPath, map[string]any{"replacement": req.SpeakerID}); err != nil {
+			return whisperx.TranscribeResult{}, err
+		}
+		return whisperx.TranscribeResult{OutputRawTranscriptPath: req.OutputRawTranscriptPath}, nil
+	}}
+	if _, err := (transcribeStage{}).Run(context.Background(), env, m); !errors.Is(err, cause) {
+		t.Fatalf("forced Run() error = %v, want %v", err, cause)
+	}
+	if got := sortedTranscribeSpeakerKeys(m.Stages["transcribe"].TranscribeSpeakers); !slices.Equal(got, []string{"a"}) {
+		t.Fatalf("speakers after failed force = %#v, want only replacement a", got)
+	}
+
+	env.Force = false
+	retry := &whisperx.FakeClient{}
+	env.WhisperX = retry
+	if _, err := (transcribeStage{}).Run(context.Background(), env, m); err != nil {
+		t.Fatalf("retry Run() error = %v", err)
+	}
+	requests := retry.RequestsSnapshot()
+	if len(requests) != 1 || requests[0].SpeakerID != "b" {
+		t.Fatalf("retry requests = %#v, want only b", requests)
+	}
+}
+
+func TestTranscribeRejectsAudioChangedDuringRequest(t *testing.T) {
+	env, m := setupTranscribeEnv(t, []string{"alice.flac"})
+	if _, err := (prepareStage{}).Run(context.Background(), env, m); err != nil {
+		t.Fatalf("prepare.Run() error = %v", err)
+	}
+	env.WhisperX = &whisperx.FakeClient{TranscribeFn: func(
+		_ context.Context,
+		req whisperx.TranscribeRequest,
+	) (whisperx.TranscribeResult, error) {
+		if err := writeJSONFile(req.OutputRawTranscriptPath, map[string]any{"speaker": req.SpeakerID}); err != nil {
+			return whisperx.TranscribeResult{}, err
+		}
+		if err := os.WriteFile(req.AudioPath, []byte("changed"), 0o644); err != nil {
+			return whisperx.TranscribeResult{}, err
+		}
+		return whisperx.TranscribeResult{OutputRawTranscriptPath: req.OutputRawTranscriptPath}, nil
+	}}
+	if _, err := (transcribeStage{}).Run(context.Background(), env, m); err == nil {
+		t.Fatal("Run() error = nil, want changed-audio failure")
+	}
+	if got := len(m.Stages["transcribe"].TranscribeSpeakers); got != 0 {
+		t.Fatalf("checkpointed speakers = %d, want none", got)
+	}
+}
+
+func TestTranscribeResumeValidationUsesSpeakerEvidence(t *testing.T) {
+	env, m := setupTranscribeEnv(t, []string{"alice.flac"})
+	env.WhisperX = &whisperx.FakeClient{}
+	if _, err := (prepareStage{}).Run(context.Background(), env, m); err != nil {
+		t.Fatalf("prepare.Run() error = %v", err)
+	}
+	if _, err := (transcribeStage{}).Run(context.Background(), env, m); err != nil {
+		t.Fatalf("Run() error = %v", err)
+	}
+	validation, err := (transcribeStage{}).ValidateResume(context.Background(), env, m)
+	if err != nil || !validation.Resumable {
+		t.Fatalf("ValidateResume() = %#v, %v, want resumable", validation, err)
+	}
+	if err := os.WriteFile(m.Inputs[0].Path, []byte("changed"), 0o644); err != nil {
+		t.Fatalf("change audio: %v", err)
+	}
+	validation, err = (transcribeStage{}).ValidateResume(context.Background(), env, m)
+	if err != nil || validation.Resumable {
+		t.Fatalf("ValidateResume() = %#v, %v, want execution required", validation, err)
+	}
+}
+
+func TestTranscribeRebuildsAggregateFromCurrentSpeakersWithoutRequests(t *testing.T) {
+	env, m := setupTranscribeEnv(t, []string{"alice.flac", "bob.flac"})
+	fake := &whisperx.FakeClient{}
+	env.WhisperX = fake
+	if _, err := (prepareStage{}).Run(context.Background(), env, m); err != nil {
+		t.Fatalf("prepare.Run() error = %v", err)
+	}
+	seed, err := (transcribeStage{}).Run(context.Background(), env, m)
+	if err != nil {
+		t.Fatalf("seed Run() error = %v", err)
+	}
+	originalRunID := m.RunID
+	m.MarkStageFailed("transcribe", m.UpdatedAt.Add(1), "aggregate save failed")
+	m.RunID = "20260503T030405Z-c3d4e5f6"
+	fake = &whisperx.FakeClient{}
+	env.WhisperX = fake
+
+	result, err := (transcribeStage{}).Run(context.Background(), env, m)
+	if err != nil {
+		t.Fatalf("rebuild Run() error = %v", err)
+	}
+	if got := len(fake.RequestsSnapshot()); got != 0 {
+		t.Fatalf("adapter requests = %d, want none", got)
+	}
+	if len(result.Outputs) != len(seed.Outputs) {
+		t.Fatalf("rebuilt outputs = %d, want %d", len(result.Outputs), len(seed.Outputs))
+	}
+	for index, output := range result.Outputs {
+		if output.ProducerRunID != originalRunID {
+			t.Fatalf("output %d producer = %q, want %q", index, output.ProducerRunID, originalRunID)
+		}
+		if output.Checksum != seed.Outputs[index].Checksum {
+			t.Fatalf("output %d checksum = %q, want %q", index, output.Checksum, seed.Outputs[index].Checksum)
+		}
+	}
+}
+
+type failingTranscribeManifestStore struct {
+	manifest.Store
+	Saves  atomic.Int64
+	FailAt int64
+	Err    error
+}
+
+func (store *failingTranscribeManifestStore) Save(
+	ctx context.Context,
+	path string,
+	m *manifest.Manifest,
+) error {
+	if store.Saves.Add(1) == store.FailAt {
+		return store.Err
+	}
+	return store.Store.Save(ctx, path, m)
+}
+
+func sortedTranscribeSpeakerKeys(records map[string]manifest.TranscribeSpeakerRecord) []string {
+	keys := make([]string, 0, len(records))
+	for key := range records {
+		keys = append(keys, key)
+	}
+	slices.Sort(keys)
+	return keys
+}

internal/stage/transcribe_execution_state.go

@@ -0,0 +1,292 @@
+package stage
+
+import (
+	"context"
+	"fmt"
+	"reflect"
+	"sort"
+	"strings"
+	"sync"
+	"time"
+
+	"gitea.maximumdirect.net/eric/narratio/internal/artifacts"
+	"gitea.maximumdirect.net/eric/narratio/internal/manifest"
+	"gitea.maximumdirect.net/eric/narratio/internal/pathsafe"
+)
+
+type transcribeExecutionState struct {
+	env      *Env
+	manifest *manifest.Manifest
+	paths    artifacts.SessionPaths
+
+	mu            sync.Mutex
+	records       map[string]manifest.TranscribeSpeakerRecord
+	checkpointErr error
+	completed     int
+	perFile       map[string]map[string]any
+	runOutputs    map[string]string
+}
+
+func newTranscribeExecutionState(
+	env *Env,
+	m *manifest.Manifest,
+	paths artifacts.SessionPaths,
+	current map[string]manifest.TranscribeSpeakerRecord,
+) *transcribeExecutionState {
+	return &transcribeExecutionState{
+		env:        env,
+		manifest:   m,
+		paths:      paths,
+		records:    manifest.CloneTranscribeSpeakerCollection(current),
+		perFile:    make(map[string]map[string]any),
+		runOutputs: make(map[string]string),
+	}
+}
+
+func (state *transcribeExecutionState) checkpointStartingCollection(ctx context.Context) error {
+	state.mu.Lock()
+	defer state.mu.Unlock()
+	if record := transcribeStageRecord(state.manifest); record != nil &&
+		record.TranscribeStateVersion == manifest.TranscribeStateContractVersion &&
+		reflect.DeepEqual(record.TranscribeSpeakers, state.records) {
+		return nil
+	}
+	if err := state.checkpointCollectionLocked(ctx, state.records); err != nil {
+		state.checkpointErr = err
+		return err
+	}
+	return nil
+}
+
+func (state *transcribeExecutionState) acceptSpeaker(
+	ctx context.Context,
+	job transcribeReconciliationJob,
+	adapterOutputPath string,
+	metadata map[string]any,
+) error {
+	if _, err := authoritativeOutputPath(job.runOutputPath, adapterOutputPath); err != nil {
+		return err
+	}
+	if err := validateTranscriptJSONFile(job.runOutputPath); err != nil {
+		return fmt.Errorf("output %q invalid: %w", job.runOutputPath, err)
+	}
+	audioSHA256, audioSize, err := hashRegularFileWithContext(ctx, job.audioPath)
+	if err != nil {
+		return fmt.Errorf("recheck audio %q: %w", job.audioPath, err)
+	}
+	if audioSHA256 != job.audioSHA256 || audioSize != job.audioSize {
+		return fmt.Errorf("audio %q changed during transcription", job.audioPath)
+	}
+	if err := state.checkpointFailure(); err != nil {
+		return err
+	}
+
+	producerRunID := strings.TrimSpace(state.manifest.RunID)
+	ref := artifacts.Ref{
+		Kind:          "transcript_raw",
+		Category:      "transcripts",
+		SessionID:     state.manifest.SessionID,
+		AbsolutePath:  job.runOutputPath,
+		ProducerRunID: producerRunID,
+	}
+	materialized, err := materializeRunLocalOutput(
+		state.env.ArtifactStore,
+		job.runOutputPath,
+		job.canonicalOutputPath,
+		ref,
+	)
+	if err != nil {
+		return fmt.Errorf("materialize output: %w", err)
+	}
+	outputSHA256, outputSize, err := hashRegularFileWithContext(ctx, job.canonicalOutputPath)
+	if err != nil {
+		return fmt.Errorf("inspect materialized output %q: %w", job.canonicalOutputPath, err)
+	}
+	if materialized.Checksum != outputSHA256 {
+		return fmt.Errorf("materialized output checksum changed while inspecting")
+	}
+	record := manifest.TranscribeSpeakerRecord{
+		FingerprintVersion: manifest.TranscribeFingerprintContractVersion,
+		Fingerprint:        job.fingerprint,
+		AudioPath:          job.audioRelativePath,
+		AudioSHA256:        job.audioSHA256,
+		AudioSize:          job.audioSize,
+		Output: manifest.ArtifactRecord{
+			Kind:          "transcript_raw",
+			LocalPath:     job.outputRelativePath,
+			ArchivePath:   materialized.ArchivePath,
+			ProducerRunID: producerRunID,
+			Checksum:      outputSHA256,
+		},
+		OutputSize:    outputSize,
+		ProducerRunID: producerRunID,
+	}
+
+	state.mu.Lock()
+	defer state.mu.Unlock()
+	if state.checkpointErr != nil {
+		return state.checkpointErr
+	}
+	candidate := manifest.CloneTranscribeSpeakerCollection(state.records)
+	candidate[job.speakerID] = record
+	if err := state.checkpointCollectionLocked(ctx, candidate); err != nil {
+		state.checkpointErr = fmt.Errorf("transcribe: checkpoint speaker %q: %w", job.speakerID, err)
+		return state.checkpointErr
+	}
+	state.completed++
+	state.perFile[job.speakerID] = cloneStringAnyMap(metadata)
+	state.runOutputs[job.speakerID] = job.runOutputPath
+	return nil
+}
+
+func (state *transcribeExecutionState) checkpointCollectionLocked(
+	ctx context.Context,
+	records map[string]manifest.TranscribeSpeakerRecord,
+) error {
+	if state.manifest == nil {
+		return fmt.Errorf("session manifest is required")
+	}
+	if state.env == nil || state.env.ManifestStore == nil {
+		return fmt.Errorf("manifest store is required")
+	}
+	if err := manifest.ValidateTranscribeSpeakerCollection(
+		manifest.TranscribeStateContractVersion,
+		records,
+	); err != nil {
+		return fmt.Errorf("validate transcribe speaker collection: %w", err)
+	}
+
+	candidate := cloneManifestForTranscribeCheckpoint(state.manifest)
+	if candidate.Stages == nil {
+		candidate.Stages = make(map[string]*manifest.StageRecord)
+	}
+	record := candidate.Stages["transcribe"]
+	if record == nil {
+		now := time.Now().UTC()
+		record = &manifest.StageRecord{
+			Name: "transcribe", Status: manifest.StatusRunning, CreatedAt: now, UpdatedAt: now,
+		}
+		candidate.Stages["transcribe"] = record
+	}
+	record.TranscribeStateVersion = manifest.TranscribeStateContractVersion
+	record.TranscribeSpeakers = manifest.CloneTranscribeSpeakerCollection(records)
+	record.UpdatedAt = time.Now().UTC()
+	if err := state.env.ManifestStore.Save(ctx, state.paths.ManifestPath, candidate); err != nil {
+		return fmt.Errorf("save session manifest %q: %w", state.paths.ManifestPath, err)
+	}
+
+	live := transcribeStageRecord(state.manifest)
+	if live == nil {
+		live = &manifest.StageRecord{}
+		if state.manifest.Stages == nil {
+			state.manifest.Stages = make(map[string]*manifest.StageRecord)
+		}
+		state.manifest.Stages["transcribe"] = live
+	}
+	live.TranscribeStateVersion = candidate.Stages["transcribe"].TranscribeStateVersion
+	live.TranscribeSpeakers = manifest.CloneTranscribeSpeakerCollection(
+		candidate.Stages["transcribe"].TranscribeSpeakers,
+	)
+	live.UpdatedAt = candidate.Stages["transcribe"].UpdatedAt
+	state.manifest.UpdatedAt = candidate.UpdatedAt
+	state.records = manifest.CloneTranscribeSpeakerCollection(live.TranscribeSpeakers)
+	return nil
+}
+
+func (state *transcribeExecutionState) checkpointFailure() error {
+	state.mu.Lock()
+	defer state.mu.Unlock()
+	return state.checkpointErr
+}
+
+func (state *transcribeExecutionState) completedCount() int {
+	state.mu.Lock()
+	defer state.mu.Unlock()
+	return state.completed
+}
+
+func (state *transcribeExecutionState) successDetails() (
+	[]artifacts.Ref,
+	[]string,
+	[]string,
+	map[string]any,
+	error,
+) {
+	state.mu.Lock()
+	defer state.mu.Unlock()
+
+	speakers := make([]string, 0, len(state.records))
+	for speaker := range state.records {
+		speakers = append(speakers, speaker)
+	}
+	sort.Strings(speakers)
+	outputs := make([]artifacts.Ref, 0, len(speakers))
+	outputPaths := make([]string, 0, len(speakers))
+	runOutputPaths := make([]string, 0, len(state.runOutputs))
+	orderedPerFile := make(map[string]any, len(state.perFile))
+	for _, speaker := range speakers {
+		record := state.records[speaker]
+		absolute, err := pathsafe.JoinSlashRelativeUnderRoot(state.paths.Root, record.Output.LocalPath)
+		if err != nil {
+			return nil, nil, nil, nil, fmt.Errorf("speaker %q output path: %w", speaker, err)
+		}
+		outputs = append(outputs, artifacts.Ref{
+			Kind:          record.Output.Kind,
+			Category:      "transcripts",
+			SessionID:     state.manifest.SessionID,
+			RelativePath:  record.Output.LocalPath,
+			AbsolutePath:  absolute,
+			ArchivePath:   record.Output.ArchivePath,
+			RemoteKey:     record.Output.RemoteKey,
+			Checksum:      record.Output.Checksum,
+			ProducerRunID: record.ProducerRunID,
+		})
+		outputPaths = append(outputPaths, absolute)
+		if runPath, ok := state.runOutputs[speaker]; ok {
+			runOutputPaths = append(runOutputPaths, runPath)
+		}
+		if metadata, ok := state.perFile[speaker]; ok {
+			orderedPerFile[speaker] = cloneStringAnyMap(metadata)
+		}
+	}
+	return outputs, outputPaths, runOutputPaths, orderedPerFile, nil
+}
+
+func transcribeStageRecord(m *manifest.Manifest) *manifest.StageRecord {
+	if m == nil || m.Stages == nil {
+		return nil
+	}
+	return m.Stages["transcribe"]
+}
+
+func cloneManifestForTranscribeCheckpoint(source *manifest.Manifest) *manifest.Manifest {
+	clone := *source
+	clone.Inputs = append([]manifest.InputRecord(nil), source.Inputs...)
+	clone.Artifacts = append([]manifest.ArtifactRecord(nil), source.Artifacts...)
+	clone.Stages = make(map[string]*manifest.StageRecord, len(source.Stages))
+	for name, sourceRecord := range source.Stages {
+		if sourceRecord == nil {
+			clone.Stages[name] = nil
+			continue
+		}
+		record := *sourceRecord
+		record.Outputs = append([]manifest.ArtifactRecord(nil), sourceRecord.Outputs...)
+		record.Logs = append([]string(nil), sourceRecord.Logs...)
+		record.GeneratedConfigs = append([]string(nil), sourceRecord.GeneratedConfigs...)
+		record.TranscribeSpeakers = manifest.CloneTranscribeSpeakerCollection(sourceRecord.TranscribeSpeakers)
+		record.AnalyzeArtifacts = manifest.CloneAnalyzeArtifactCollection(sourceRecord.AnalyzeArtifacts)
+		clone.Stages[name] = &record
+	}
+	return &clone
+}
+
+func cloneStringAnyMap(source map[string]any) map[string]any {
+	if source == nil {
+		return nil
+	}
+	clone := make(map[string]any, len(source))
+	for key, value := range source {
+		clone[key] = value
+	}
+	return clone
+}

internal/stage/transcribe_resume.go

@@ -0,0 +1,32 @@
+package stage
+
+import (
+	"context"
+	"fmt"
+	"sort"
+	"strings"
+
+	"gitea.maximumdirect.net/eric/narratio/internal/manifest"
+)
+
+func (transcribeStage) ValidateResume(
+	ctx context.Context,
+	env *Env,
+	m *manifest.Manifest,
+) (ResumeValidation, error) {
+	reconciliation, err := reconcileTranscribeSpeakers(ctx, env, m)
+	if err != nil {
+		return ResumeValidation{}, fmt.Errorf("transcribe resume: reconcile speakers: %w", err)
+	}
+	if len(reconciliation.work) == 0 {
+		return Resumable(), nil
+	}
+	speakers := make([]string, 0, len(reconciliation.work))
+	for _, job := range reconciliation.work {
+		speakers = append(speakers, job.speakerID)
+	}
+	sort.Strings(speakers)
+	return NonResumable("speaker transcripts require execution: " + strings.Join(speakers, ", ")), nil
+}
+
+var _ ResumeValidator = transcribeStage{}

internal/stage/transcribe_test.go

@@ -322,7 +322,7 @@ func TestTranscribeStagePreservesDistinctSameBasenameAudio(t *testing.T) {
 	}
 }
 
-func TestTranscribeStageCancellationDoesNotMaterializePartialOutputs(t *testing.T) {
+func TestTranscribeStageCancellationPreservesCheckpointedOutputs(t *testing.T) {
 	env, m := setupTranscribeEnv(t, []string{"alice.flac", "bob.flac", "cara.flac"})
 	m.RunID = "20260810T214607Z-a1b2c3d4"
 	fastComplete := make(chan struct{})
@@ -351,6 +351,21 @@ func TestTranscribeStageCancellationDoesNotMaterializePartialOutputs(t *testing.
 	}()
 	select {
 	case <-fastComplete:
+		manifestPath := sessionPathsForEnv(env, m.SessionID).ManifestPath
+		deadline := time.Now().Add(time.Second)
+		for {
+			durable, err := env.ManifestStore.Load(context.Background(), manifestPath)
+			if err == nil {
+				if _, ok := durable.Stages["transcribe"].TranscribeSpeakers["alice"]; ok {
+					break
+				}
+			}
+			if time.Now().After(deadline) {
+				cancel()
+				t.Fatal("alice completion was not checkpointed before cancellation")
+			}
+			time.Sleep(5 * time.Millisecond)
+		}
 		cancel()
 	case <-time.After(time.Second):
 		cancel()
@@ -365,8 +380,15 @@ func TestTranscribeStageCancellationDoesNotMaterializePartialOutputs(t *testing.
 	}
 
 	canonical := filepath.Join(sessionPathsForEnv(env, m.SessionID).TranscriptsRawDir, "alice.json")
-	if _, err := os.Stat(canonical); !os.IsNotExist(err) {
+	if _, err := os.Stat(canonical); err != nil {
-		t.Fatalf("partial canonical output should not exist, stat err = %v", err)
+		t.Fatalf("checkpointed canonical output missing: %v", err)
+	}
+	durable, err := env.ManifestStore.Load(context.Background(), sessionPathsForEnv(env, m.SessionID).ManifestPath)
+	if err != nil {
+		t.Fatalf("load checkpointed manifest: %v", err)
+	}
+	if _, ok := durable.Stages["transcribe"].TranscribeSpeakers["alice"]; !ok {
+		t.Fatalf("checkpointed speakers = %#v, want alice", durable.Stages["transcribe"].TranscribeSpeakers)
 	}
 }
 
@@ -454,8 +476,13 @@ func setupTranscribeEnv(t *testing.T, audioFiles []string) (*Env, *manifest.Mani
 	env := &Env{
 		Config:        cfg,
 		ArtifactStore: artifacts.NewLocalStore(workspace),
+		ManifestStore: &manifest.LocalStore{},
 	}
-	return env, manifest.New("2026-05-03", time.Now().UTC())
+	m := manifest.New("2026-05-03", time.Now().UTC())
+	m.Campaign = "sample-campaign"
+	m.RunID = "20260503T010203Z-a1b2c3d4"
+	m.MarkStageRunning("transcribe", time.Now().UTC())
+	return env, m
 }
 
 func writeJSONFile(path string, payload any) error {