diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.md
index 36e46b1..2e54f05 100644
--- a/docs/ARCHITECTURE.md
+++ b/docs/ARCHITECTURE.md
@@ -283,10 +283,13 @@ automatic** (D-021).
   `grounding`, `done`). It fails with `no_regulars`, `ai_failed` (the
   skill or `normalize_recipe` failed; no Side quest is made up instead,
   D-017), `search_not_configured` or `search_failed`, and the batch is set
-  `failed`; Side quests stored before the failure stay in the deck. At
-  boot, after the job runner fails the jobs a previous process left
-  running, every batch still `pending` or `running` with no job queued or
-  running is set `failed`, so it does not block the next.
+  `failed`; Side quests stored before the failure stay in the deck. If
+  recording the batch's end fails, its job fails; the job runner retries
+  a job's final status write with backoff until it lands. Each time a
+  batch is started, before the active batch is selected, every batch
+  still `pending` or `running` with no job queued or running (the process
+  stopped, or recording its end failed) is set `failed`, so it does not
+  block the next.
   `GET /api/v1/suggestions/deck` returns `{suggestions: [{id, dinner_id,
   name, photo_url, difficulty, method, minutes_total, cuisine, reason,
   cited_regular_ids, status, dinner_status, decided_at, created_at}]}`:
diff --git a/internal/jobs/jobs.go b/internal/jobs/jobs.go
index 2d2c865..b3eb2a6 100644
--- a/internal/jobs/jobs.go
+++ b/internal/jobs/jobs.go
@@ -54,12 +54,13 @@ const (
 	CodeUnknownKind = "unknown_kind"
 )
 
-// Claim retry backoff. A worker whose claim fails with a database error
-// retries after claimRetryMin, doubling up to claimRetryMax, so a job queued
-// when the error clears is claimed without waiting for a wake token.
+// Retry backoff. A worker whose claim or final job write fails with a
+// database error retries after retryMin, doubling up to retryMax, so a job
+// queued when the error clears is claimed without waiting for a wake token,
+// and a job the worker has finished is never left running.
 const (
-	claimRetryMin = 50 * time.Millisecond
-	claimRetryMax = 5 * time.Second
+	retryMin = 50 * time.Millisecond
+	retryMax = 5 * time.Second
 )
 
 // codePattern is the form of every failure code: lower-case snake case.
@@ -387,16 +388,16 @@ func (r *Runner) Get(ctx context.Context, id string) (Job, error) {
 // already queued is not stranded by a transient error.
 func (r *Runner) work(ctx context.Context) {
 	defer r.wg.Done()
-	backoff := claimRetryMin
+	backoff := retryMin
 	for ctx.Err() == nil {
 		job, err := r.claim(ctx, sql.NullString{String: store.Now().String(), Valid: true})
 		switch {
 		case err == nil:
-			backoff = claimRetryMin
+			backoff = retryMin
 			r.run(ctx, job)
 			continue
 		case errors.Is(err, sql.ErrNoRows):
-			backoff = claimRetryMin
+			backoff = retryMin
 			select {
 			case <-r.wake:
 			case <-ctx.Done():
@@ -412,7 +413,7 @@ func (r *Runner) work(ctx context.Context) {
 		case <-ctx.Done():
 			timer.Stop()
 		}
-		backoff = min(2*backoff, claimRetryMax)
+		backoff = min(2*backoff, retryMax)
 	}
 }
 
@@ -420,14 +421,13 @@ func (r *Runner) work(ctx context.Context) {
 // ctx has been cancelled, so a handler that finishes during shutdown is
 // recorded.
 func (r *Runner) run(ctx context.Context, job db.Job) {
-	write := context.WithoutCancel(ctx)
 	logger := r.logger.With("job_id", job.ID, "kind", job.Kind)
 	q := db.New(r.st.Writer)
 
 	h, ok := r.handlers[job.Kind] // read-only once started
 	if !ok {
 		logger.ErrorContext(ctx, "job failed", "code", CodeUnknownKind, "reason", "no handler registered for the kind")
-		r.fail(write, logger, job.ID, CodeUnknownKind)
+		r.fail(ctx, logger, job.ID, CodeUnknownKind)
 		return
 	}
 
@@ -449,7 +449,7 @@ func (r *Runner) run(ctx context.Context, job db.Job) {
 	result, code, reason := h.run(ctx, job.Payload, publish)
 	if code != "" {
 		if ctx.Err() != nil {
-			logger.WarnContext(write, "job interrupted by shutdown; it fails at the next start")
+			logger.WarnContext(context.WithoutCancel(ctx), "job interrupted by shutdown; it fails at the next start")
 			return
 		}
 		level := slog.LevelInfo
@@ -457,35 +457,58 @@ func (r *Runner) run(ctx context.Context, job db.Job) {
 			level = slog.LevelError
 		}
 		logger.Log(ctx, level, "job failed", "code", code, "reason", reason)
-		r.fail(write, logger, job.ID, code)
+		r.fail(ctx, logger, job.ID, code)
 		return
 	}
 
-	n, err := q.SucceedJob(write, db.SucceedJobParams{
-		Result: sql.NullString{String: string(result), Valid: true},
-		Now:    sql.NullString{String: store.Now().String(), Valid: true},
-		ID:     job.ID,
+	r.record(ctx, logger, "record job success", func(write context.Context) (int64, error) {
+		return q.SucceedJob(write, db.SucceedJobParams{
+			Result: sql.NullString{String: string(result), Valid: true},
+			Now:    sql.NullString{String: store.Now().String(), Valid: true},
+			ID:     job.ID,
+		})
 	})
-	if err == nil && n != 1 {
-		err = errors.New("job is not running")
-	}
-	if err != nil {
-		logger.ErrorContext(write, "record job success", "error", err)
-	}
 }
 
 // fail records the job as failed with code. The record errors it logs come
 // from the database, never from a handler.
 func (r *Runner) fail(ctx context.Context, logger *slog.Logger, id, code string) {
-	n, err := db.New(r.st.Writer).FailJob(ctx, db.FailJobParams{
-		Error: sql.NullString{String: code, Valid: true},
-		Now:   sql.NullString{String: store.Now().String(), Valid: true},
-		ID:    id,
+	r.record(ctx, logger, "record job failure", func(write context.Context) (int64, error) {
+		return db.New(r.st.Writer).FailJob(write, db.FailJobParams{
+			Error: sql.NullString{String: code, Valid: true},
+			Now:   sql.NullString{String: store.Now().String(), Valid: true},
+			ID:    id,
+		})
 	})
-	if err == nil && n != 1 {
-		err = errors.New("job is not running")
-	}
-	if err != nil {
-		logger.ErrorContext(ctx, "record job failure", "error", err)
+}
+
+// record makes a job's final write, which must change the one running job.
+// A database error is retried with backoff until the write lands, so a
+// worker never moves on from a job it has finished while leaving it
+// running. Each attempt is made even when ctx has been cancelled; once it
+// is, a failed attempt is not retried and the job stays running, so the
+// next Start fails it as interrupted.
+func (r *Runner) record(ctx context.Context, logger *slog.Logger, what string, write func(context.Context) (int64, error)) {
+	backoff := retryMin
+	for {
+		n, err := write(context.WithoutCancel(ctx))
+		if err == nil {
+			if n != 1 {
+				logger.ErrorContext(ctx, what, "error", "job is not running")
+			}
+			return
+		}
+		if ctx.Err() != nil {
+			logger.ErrorContext(context.WithoutCancel(ctx), what+" failed during shutdown; it fails at the next start", "error", err)
+			return
+		}
+		logger.ErrorContext(ctx, what+" failed; retrying", "error", err, "retry_in", backoff)
+		timer := time.NewTimer(backoff)
+		select {
+		case <-timer.C:
+		case <-ctx.Done():
+			timer.Stop()
+		}
+		backoff = min(2*backoff, retryMax)
 	}
 }
diff --git a/internal/jobs/jobs_test.go b/internal/jobs/jobs_test.go
index b1b6752..1727d18 100644
--- a/internal/jobs/jobs_test.go
+++ b/internal/jobs/jobs_test.go
@@ -609,6 +609,39 @@ func TestClaimErrorIsRetried(t *testing.T) {
 	}
 }
 
+func TestFinalWriteErrorIsRetried(t *testing.T) {
+	st := migratedStore(t)
+	logs := &logBuffer{}
+	r := newRunnerLogging(t, st, 1, logs)
+	k := register(t, r, "count", func(ctx context.Context, p countPayload, publish func(countResult) error) (countResult, error) {
+		return countResult{Done: p.Steps}, nil
+	})
+	if _, err := st.Writer.ExecContext(t.Context(), `CREATE TRIGGER refuse_end BEFORE UPDATE OF status ON jobs WHEN NEW.status IN ('succeeded', 'failed') BEGIN SELECT RAISE(ABORT, 'end refused'); END`); err != nil {
+		t.Fatal(err)
+	}
+	start(t, r)
+	id := enqueue(t, k, countPayload{Steps: 3})
+
+	deadline := time.Now().Add(5 * time.Second)
+	for !strings.Contains(logs.String(), "record job success") {
+		if time.Now().After(deadline) {
+			t.Fatal("job success never recorded")
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+	if _, err := st.Writer.ExecContext(t.Context(), `DROP TRIGGER refuse_end`); err != nil {
+		t.Fatal(err)
+	}
+
+	job := waitFor(t, r, id, finished)
+	if job.Status != StatusSucceeded || string(job.Result) != `{"done":3}` {
+		t.Fatalf("job = %+v (result %s), want succeeded with {\"done\":3}", job, job.Result)
+	}
+	if out := logs.String(); !strings.Contains(out, "record job success failed; retrying") {
+		t.Fatalf("final write error not logged as retried: %s", out)
+	}
+}
+
 func TestRunnerEnqueueByKindName(t *testing.T) {
 	t.Parallel()
 	st := migratedStore(t)
diff --git a/internal/suggest/suggest_test.go b/internal/suggest/suggest_test.go
index 43eef70..7f89c88 100644
--- a/internal/suggest/suggest_test.go
+++ b/internal/suggest/suggest_test.go
@@ -44,6 +44,9 @@ type env struct {
 	regulars map[string]string
 	grounder *recipes.Grounder
 
+	// logs collects the service's and job runner's log output.
+	logs *logBuffer
+
 	mu     sync.Mutex
 	inputs []skills.SideQuestsRequest
 	// stub, when set, answers in place of suggest_side_quests.
@@ -82,6 +85,7 @@ func newEnv(t *testing.T) *env {
 		model:    &llm.Fake{},
 		side:     &llm.Fake{},
 		regulars: map[string]string{},
+		logs:     &logBuffer{},
 	}
 	for _, name := range []string{"Beef Chili", "Chicken Fajitas"} {
 		e.regulars[name] = e.dinner(t, name, "regular")
@@ -108,7 +112,7 @@ func newEnv(t *testing.T) *env {
 		}
 		return skills.SuggestSideQuests(ctx, cfg, scope, in)
 	}
-	logger := slog.New(slog.NewTextHandler(io.Discard, nil))
+	logger := slog.New(slog.NewTextHandler(e.logs, nil))
 	e.runner, err = jobs.NewRunner(st, logger, 1)
 	if err != nil {
 		t.Fatal(err)
@@ -120,6 +124,24 @@ func newEnv(t *testing.T) *env {
 	return e
 }
 
+// logBuffer collects log output written by concurrent goroutines.
+type logBuffer struct {
+	mu  sync.Mutex
+	buf strings.Builder
+}
+
+func (b *logBuffer) Write(p []byte) (int, error) {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	return b.buf.Write(p)
+}
+
+func (b *logBuffer) String() string {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	return b.buf.String()
+}
+
 // start starts the env's job runner, stopping it when the test ends.
 func (e *env) start(t *testing.T) {
 	t.Helper()
@@ -951,6 +973,62 @@ func TestBatchJobFailsWhenRecordingItsEndFails(t *testing.T) {
 	}
 }
 
+func TestBatchRecoversWhenRecordingItsEndAndItsJobsEndFail(t *testing.T) {
+	e := newEnv(t)
+	e.start(t)
+	e.recipe(t, "Chili Verde")
+	e.reply(sideQuest("Chili Verde", "Your Beef Chili gone green.", "Beef Chili"))
+	for _, trigger := range []string{
+		`CREATE TRIGGER refuse_done BEFORE UPDATE OF status ON suggestion_batches WHEN NEW.status = 'done' BEGIN SELECT RAISE(ABORT, 'done refused'); END`,
+		`CREATE TRIGGER refuse_job_end BEFORE UPDATE OF status ON jobs WHEN NEW.status IN ('succeeded', 'failed') BEGIN SELECT RAISE(ABORT, 'job end refused'); END`,
+	} {
+		if _, err := e.st.Writer.ExecContext(t.Context(), trigger); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	b, err := e.svc.StartBatch(t.Context(), suggest.TriggerManual)
+	if err != nil {
+		t.Fatal(err)
+	}
+	deadline := time.Now().Add(10 * time.Second)
+	for !strings.Contains(e.logs.String(), "record job failure") {
+		if time.Now().After(deadline) {
+			t.Fatalf("job end never recorded; batch %s", e.batchStatus(t, b.ID))
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+	// While neither end is recorded the batch is still the active one.
+	again, err := e.svc.StartBatch(t.Context(), suggest.TriggerManual)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if again.Created || again.ID != b.ID {
+		t.Fatalf("batch = %+v while batch %s is running, want it", again, b.ID)
+	}
+
+	// The database recovers; the service is not restarted.
+	for _, trigger := range []string{"refuse_done", "refuse_job_end"} {
+		if _, err := e.st.Writer.ExecContext(t.Context(), "DROP TRIGGER "+trigger); err != nil {
+			t.Fatal(err)
+		}
+	}
+	job := e.waitJob(t, b.JobID)
+	if job.Status != jobs.StatusFailed {
+		t.Fatalf("job %s, want failed", job.Status)
+	}
+	next, err := e.svc.StartBatch(t.Context(), suggest.TriggerManual)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !next.Created || next.ID == b.ID {
+		t.Errorf("next batch = %+v, want a new one", next)
+	}
+	if got := e.batchStatus(t, b.ID); got != suggest.BatchFailed {
+		t.Errorf("batch status = %s with its job failed, want failed", got)
+	}
+}
+
 func TestStartBatchRefusesUnknownTrigger(t *testing.T) {
 	e := newEnv(t)
 	if _, err := e.svc.StartBatch(t.Context(), "whenever"); err == nil {