feat: upload audio, get a private Swedish transcript, discuss it (#33, ADR-033) (#39)

Upload a recording (audio, or video with the camera off), get a private Swedish transcript, and chat about it.

- ADR-033: RLS-scoped private transcript store; ffprobe/ffmpeg intake; gateway transcription (iguana kb-whisper, berget fallback); in-process single-replica worker; chat over the transcript (local models only).
- ADR-034: optional two-engine mode (pianissimo + kb-whisper) reconciled by a LOCAL model; off unless TAPIR_RECONCILE_BASE_MODEL is set.
- ADR-035: real progress (phases, running/quiet/lost), phase logging, every judged spot stored, merged segments.
- Hardening from four adversarial reviews: errors logged and stored as kinds never text; per-engine deadlines; converted-length check; bounded queue; client hang-ups are not errors; strict local-alias check.
- Dockerfile: static ffmpeg/ffprobe 7.0.2 pinned by sha256 (#34). Built and inspected by hand; CI does not build the image on a PR, so its first CI build is this merge.

Uploads stay OFF until TAPIR_UPLOADS_ENABLED is set in the deployment.

Closes #33, #34, #42, #43, #44.

Co-authored-by: Mathias <mthbqv@gmail.com>

internal/adapters/transcribe/chain.go

@@ -0,0 +1,53 @@
+package transcribe
+
+import (
+	"context"
+	"errors"
+	"fmt"
+
+	"git.d-ma.be/mathias/tapir/internal/domain"
+)
+
+// Backend is one transcription endpoint. *Gateway satisfies it; Name is the value
+// recorded as the serving endpoint.
+type Backend interface {
+	Name() string
+	Transcribe(ctx context.Context, wavPath string) (domain.Transcription, error)
+}
+
+// Chain tries a primary (local) backend and, if it fails for ANY reason, a
+// fallback. The fallback exists because iguana can wedge permanently on a Metal
+// OOM (infra#273) — which is also why a fallback is a data-egress decision: the
+// berget endpoint is a cloud service, so the result records which one served it.
+type Chain struct {
+	primary  Backend
+	fallback Backend // nil disables the fallback
+}
+
+// NewChain builds a Chain. A nil fallback means a primary failure is the result.
+func NewChain(primary, fallback Backend) *Chain {
+	return &Chain{primary: primary, fallback: fallback}
+}
+
+// Transcribe returns the first backend's success. It falls back on any primary
+// error — transport, timeout, non-2xx, empty transcript — and never inspects a
+// response body to decide. A cancelled context stops the chain: a job the user or
+// shutdown cancelled must not ship its audio to a cloud endpoint as a last act.
+func (c *Chain) Transcribe(ctx context.Context, wavPath string) (domain.Transcription, error) {
+	t, err := c.primary.Transcribe(ctx, wavPath)
+	if err == nil {
+		t.Endpoint = c.primary.Name()
+		return t, nil
+	}
+	if c.fallback == nil || ctx.Err() != nil {
+		return domain.Transcription{}, fmt.Errorf("%s: %w", c.primary.Name(), err)
+	}
+	t, ferr := c.fallback.Transcribe(ctx, wavPath)
+	if ferr != nil {
+		return domain.Transcription{}, errors.Join(
+			fmt.Errorf("%s: %w", c.primary.Name(), err),
+			fmt.Errorf("%s: %w", c.fallback.Name(), ferr))
+	}
+	t.Endpoint = c.fallback.Name()
+	return t, nil
+}

internal/adapters/transcribe/chain_test.go

@@ -0,0 +1,99 @@
+package transcribe_test
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/transcribe"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+)
+
+type fakeBackend struct {
+	name  string
+	text  string
+	err   error
+	calls int
+}
+
+func (f *fakeBackend) Name() string { return f.name }
+func (f *fakeBackend) Transcribe(_ context.Context, _ string) (domain.Transcription, error) {
+	f.calls++
+	if f.err != nil {
+		return domain.Transcription{}, f.err
+	}
+	return domain.Transcription{Text: f.text, Endpoint: f.name}, nil
+}
+
+func TestChain_PrimarySuccessNeverTouchesTheFallback(t *testing.T) {
+	primary := &fakeBackend{name: "iguana/kb-whisper", text: "primary text"}
+	fallback := &fakeBackend{name: "berget/kb-whisper", text: "fallback text"}
+
+	got, err := transcribe.NewChain(primary, fallback).Transcribe(context.Background(), "/a.wav")
+
+	require.NoError(t, err)
+	assert.Equal(t, "primary text", got.Text)
+	assert.Equal(t, "iguana/kb-whisper", got.Endpoint)
+	assert.Zero(t, fallback.calls, "audio must not leave the homelab when the local endpoint worked")
+}
+
+func TestChain_AnyPrimaryFailureFallsBackAndRecordsTheFallbackEndpoint(t *testing.T) {
+	tests := []struct {
+		name string
+		err  error
+	}{
+		{"server error", &transcribe.StatusError{Code: 500}},
+		{"gateway timeout", &transcribe.StatusError{Code: 504}},
+		{"client error", &transcribe.StatusError{Code: 400}},
+		{"transport failure", errors.New("dial tcp: connection refused")},
+		{"empty transcript", transcribe.ErrEmptyTranscript},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			primary := &fakeBackend{name: "iguana/kb-whisper", err: tt.err}
+			fallback := &fakeBackend{name: "berget/kb-whisper", text: "fallback text"}
+
+			got, err := transcribe.NewChain(primary, fallback).Transcribe(context.Background(), "/a.wav")
+
+			require.NoError(t, err)
+			assert.Equal(t, "fallback text", got.Text)
+			assert.Equal(t, "berget/kb-whisper", got.Endpoint,
+				"the row must say berget served it: that means the audio left the homelab")
+		})
+	}
+}
+
+func TestChain_BothFailingReportsBothErrors(t *testing.T) {
+	primary := &fakeBackend{name: "iguana/kb-whisper", err: errors.New("iguana down")}
+	fallback := &fakeBackend{name: "berget/kb-whisper", err: errors.New("berget down")}
+
+	_, err := transcribe.NewChain(primary, fallback).Transcribe(context.Background(), "/a.wav")
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "iguana down")
+	assert.Contains(t, err.Error(), "berget down")
+}
+
+func TestChain_WithoutAFallbackAPrimaryFailureIsTheResult(t *testing.T) {
+	primary := &fakeBackend{name: "iguana/kb-whisper", err: errors.New("iguana down")}
+
+	_, err := transcribe.NewChain(primary, nil).Transcribe(context.Background(), "/a.wav")
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "iguana down")
+}
+
+func TestChain_ACancelledJobDoesNotFallBack(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	primary := &fakeBackend{name: "iguana/kb-whisper", err: context.Canceled}
+	fallback := &fakeBackend{name: "berget/kb-whisper", text: "x"}
+
+	_, err := transcribe.NewChain(primary, fallback).Transcribe(ctx, "/a.wav")
+
+	assert.ErrorIs(t, err, context.Canceled)
+	assert.Zero(t, fallback.calls, "a cancelled job must not ship the audio to a cloud endpoint")
+}

internal/adapters/transcribe/ensemble.go

@@ -0,0 +1,114 @@
+package transcribe
+
+import (
+	"context"
+	"errors"
+	"fmt"
+
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/errkind"
+	"git.d-ma.be/mathias/tapir/internal/progress"
+	"git.d-ma.be/mathias/tapir/internal/reconcile"
+)
+
+// Ensemble transcribes the same WAV with two engines, one after the other, and has a
+// local model reconcile their disagreements (ADR-034). It satisfies uploads.Transcriber,
+// so the worker neither knows nor cares that two engines were involved.
+//
+// It degrades in steps and never turns a recoverable problem into a failed job:
+// one engine down → the other's text, said so on the result; the model down → the
+// base text unreconciled; both engines down → the cloud fallback, if configured, and
+// recorded as such. A cancelled job stops everything and never reaches the cloud.
+type Ensemble struct {
+	base, alt Backend
+	fallback  Backend // nil = no cloud fallback
+	adj       reconcile.Adjudicator
+	opts      reconcile.Options
+}
+
+// NewEnsemble builds an Ensemble. base is the engine whose text stands where the two
+// disagree and is the faster one, so it runs first; alt supplies the other reading.
+func NewEnsemble(base, alt, fallback Backend, adj reconcile.Adjudicator, opts reconcile.Options) *Ensemble {
+	return &Ensemble{base: base, alt: alt, fallback: fallback, adj: adj, opts: opts}
+}
+
+// Plan implements uploads.Planner: the phases this ensemble reports, so the page can
+// show them before they run. The fallback is a phase only when one is configured.
+func (e *Ensemble) Plan() []progress.Planned {
+	plan := []progress.Planned{
+		{Key: "engine_a", Label: "Transcribing with " + e.base.Name()},
+		{Key: "engine_b", Label: "Transcribing with " + e.alt.Name()},
+		{Key: "compare", Label: "Comparing the two transcripts"},
+		{Key: "decide", Label: "Deciding the differences"},
+	}
+	if e.fallback != nil {
+		plan = append(plan, progress.Planned{Key: "fallback", Label: "Cloud fallback (" + e.fallback.Name() + ")"})
+	}
+	return plan
+}
+
+// Transcribe implements uploads.Transcriber. Each phase is reported through the
+// progress.Reporter in ctx (a no-op when nobody listens).
+func (e *Ensemble) Transcribe(ctx context.Context, wavPath string) (domain.Transcription, error) {
+	rep := progress.From(ctx)
+	b, berr := e.run(ctx, rep, "engine_a", e.base, wavPath)
+	if err := ctx.Err(); err != nil {
+		return domain.Transcription{}, err
+	}
+	a, aerr := e.run(ctx, rep, "engine_b", e.alt, wavPath)
+	if err := ctx.Err(); err != nil {
+		return domain.Transcription{}, err
+	}
+
+	switch {
+	case berr == nil && aerr == nil:
+		out, err := reconcile.Reconcile(ctx, b, a, e.adj, e.opts)
+		if err != nil {
+			return domain.Transcription{}, err
+		}
+		out.Endpoint = fmt.Sprintf("%s + %s, reconciled by %s", e.base.Name(), e.alt.Name(), e.opts.Reconciler)
+		if rc := out.Reconciliation; rc != nil {
+			rep.Note("reconciled", "status", string(rc.Status), "disputes", rc.Disputes, "auto_resolved", rc.AutoResolved,
+				"adjudicated", rc.Adjudicated, "changed", rc.Changed, "unsure", rc.Unsure, "errors", rc.Errors)
+		}
+		return out, nil
+	case berr == nil:
+		return e.single(b, e.base.Name(), fmt.Sprintf("%s failed (%s)", e.alt.Name(), errkind.Of(aerr))), nil
+	case aerr == nil:
+		return e.single(a, e.alt.Name(), fmt.Sprintf("%s failed (%s)", e.base.Name(), errkind.Of(berr))), nil
+	}
+
+	both := errors.Join(fmt.Errorf("%s: %w", e.base.Name(), berr), fmt.Errorf("%s: %w", e.alt.Name(), aerr))
+	if e.fallback == nil {
+		return domain.Transcription{}, both
+	}
+	f, ferr := e.run(ctx, rep, "fallback", e.fallback, wavPath)
+	if ferr != nil {
+		return domain.Transcription{}, errors.Join(both, fmt.Errorf("%s: %w", e.fallback.Name(), ferr))
+	}
+	return e.single(f, e.fallback.Name(), "both local engines failed"), nil
+}
+
+// run is one engine call, reported as a phase. A job that was stopped is not a broken
+// engine, so a cancelled context is never reported as a failed phase.
+func (e *Ensemble) run(ctx context.Context, rep progress.Reporter, key string, be Backend, wav string) (domain.Transcription, error) {
+	rep.Begin(key, 0)
+	t, err := be.Transcribe(ctx, wav)
+	switch {
+	case err == nil:
+		rep.Done(key)
+	case ctx.Err() == nil:
+		rep.Fail(key, err)
+	}
+	return t, err
+}
+
+// single is a transcript from one engine only: nothing to reconcile, and the result
+// says why, because a reader should know the second opinion was missing.
+func (e *Ensemble) single(t domain.Transcription, engine, reason string) domain.Transcription {
+	t.Endpoint = engine
+	t.Reconciliation = &domain.Reconciliation{
+		Status: domain.ReconcileSingle, Reason: reason, Base: e.opts.BaseEngine, Alt: e.opts.AltEngine, Reconciler: e.opts.Reconciler,
+	}
+	return t
+}

internal/adapters/transcribe/ensemble_progress_test.go

@@ -0,0 +1,162 @@
+package transcribe_test
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"sync"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/transcribe"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/progress"
+	"git.d-ma.be/mathias/tapir/internal/reconcile"
+)
+
+// What the Ensemble reports to the page and the log (ADR-035). The reporter here is a
+// recording double carried in the context, exactly as the worker's Tracker is.
+
+type recReporter struct {
+	mu     sync.Mutex
+	events []string
+	notes  []map[string]any
+}
+
+func (r *recReporter) add(e string) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.events = append(r.events, e)
+}
+
+func (r *recReporter) Begin(key string, _ int)   { r.add("begin:" + key) }
+func (r *recReporter) Advance(key string, _ int) { r.add("advance:" + key) }
+func (r *recReporter) Done(key string)           { r.add("done:" + key) }
+func (r *recReporter) Fail(key string, _ error)  { r.add("fail:" + key) }
+func (r *recReporter) Note(msg string, kv ...any) {
+	m := map[string]any{"msg": msg}
+	for i := 0; i+1 < len(kv); i += 2 {
+		m[kv[i].(string)] = kv[i+1]
+	}
+	r.mu.Lock()
+	r.notes = append(r.notes, m)
+	r.mu.Unlock()
+	r.add("note:" + msg)
+}
+
+func (r *recReporter) joined() string {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	return strings.Join(r.events, ",")
+}
+
+func TestEnsemblePlan_DeclaresBothEnginesAndTheReconcileStepsByName(t *testing.T) {
+	r := newEnsemble("a", "a", nil, nil, false, &verdictAdj{})
+
+	plan := r.e.Plan()
+
+	require.Len(t, plan, 4)
+	assert.Equal(t, []string{"engine_a", "engine_b", "compare", "decide"}, []string{plan[0].Key, plan[1].Key, plan[2].Key, plan[3].Key})
+	assert.Contains(t, plan[0].Label, "iguana/pianissimo", "the page says which engine is running")
+	assert.Contains(t, plan[1].Label, "iguana/kb-whisper")
+}
+
+func TestEnsemblePlan_TheCloudFallbackIsAStepOnlyWhenOneIsConfigured(t *testing.T) {
+	with := newEnsemble("a", "a", nil, nil, true, &verdictAdj{}).e.Plan()
+	without := newEnsemble("a", "a", nil, nil, false, &verdictAdj{}).e.Plan()
+
+	assert.Equal(t, "fallback", with[len(with)-1].Key)
+	assert.Len(t, without, 4, "no fallback configured: the page must not promise one")
+}
+
+func TestEnsembleProgress_BothEnginesAreReportedInTheOrderTheyRun(t *testing.T) {
+	rep := &recReporter{}
+	r := newEnsemble("Totalt 1015 kronor idag.", "Totalt 1050 kronor idag.", nil, nil, false, &verdictAdj{v: reconcile.VerdictAlt})
+
+	_, err := r.e.Transcribe(progress.With(context.Background(), rep), "/a.wav")
+
+	require.NoError(t, err)
+	got := rep.joined()
+	a, aDone := strings.Index(got, "begin:engine_a"), strings.Index(got, "done:engine_a")
+	b, bDone := strings.Index(got, "begin:engine_b"), strings.Index(got, "done:engine_b")
+	require.True(t, a >= 0 && aDone > a && b > aDone && bDone > b, "events out of order: %s", got)
+	assert.NotContains(t, got, "fail:")
+}
+
+func TestEnsembleProgress_AnEngineThatFailsIsShownFailedAndTheOtherStillRuns(t *testing.T) {
+	rep := &recReporter{}
+	r := newEnsemble("text", "text", errors.New("pianissimo down"), nil, false, &verdictAdj{})
+
+	got, err := r.e.Transcribe(progress.With(context.Background(), rep), "/a.wav")
+
+	require.NoError(t, err, "one engine down is a degraded result, not a failed job")
+	assert.Equal(t, domain.ReconcileSingle, got.Reconciliation.Status)
+	ev := rep.joined()
+	assert.Contains(t, ev, "fail:engine_a")
+	assert.Contains(t, ev, "done:engine_b")
+	assert.NotContains(t, ev, "begin:compare", "nothing to compare with one engine")
+}
+
+func TestEnsembleProgress_TheFallbackIsReportedOnlyWhenBothEnginesFailed(t *testing.T) {
+	rep := &recReporter{}
+	r := newEnsemble("t", "t", errors.New("a down"), errors.New("b down"), true, &verdictAdj{})
+
+	_, err := r.e.Transcribe(progress.With(context.Background(), rep), "/a.wav")
+
+	require.NoError(t, err)
+	ev := rep.joined()
+	assert.Contains(t, ev, "fail:engine_a")
+	assert.Contains(t, ev, "fail:engine_b")
+	assert.Contains(t, ev, "begin:fallback")
+	assert.Contains(t, ev, "done:fallback")
+
+	rep2 := &recReporter{}
+	ok := newEnsemble("x", "x", nil, nil, true, &verdictAdj{})
+	_, err = ok.e.Transcribe(progress.With(context.Background(), rep2), "/a.wav")
+	require.NoError(t, err)
+	assert.NotContains(t, rep2.joined(), "fallback", "a healthy run never touches the cloud step")
+}
+
+// cancellingBackend ends the job in the middle of its own call, like a shutdown.
+type cancellingBackend struct {
+	name   string
+	cancel context.CancelFunc
+}
+
+func (c cancellingBackend) Name() string { return c.name }
+func (c cancellingBackend) Transcribe(ctx context.Context, _ string) (domain.Transcription, error) {
+	c.cancel()
+	return domain.Transcription{}, ctx.Err()
+}
+
+func TestEnsembleProgress_AShutdownMidEngineIsNotReportedAsAFailedEngine(t *testing.T) {
+	rep := &recReporter{}
+	ctx, cancel := context.WithCancel(progress.With(context.Background(), rep))
+	calls := &[]string{}
+	alt := &orderedBackend{name: "iguana/kb-whisper", text: "x", calls: calls}
+	e := transcribe.NewEnsemble(cancellingBackend{name: "iguana/pianissimo", cancel: cancel}, alt, nil, &verdictAdj{}, ensembleOpts)
+
+	_, err := e.Transcribe(ctx, "/a.wav")
+
+	require.ErrorIs(t, err, context.Canceled)
+	assert.NotContains(t, rep.joined(), "fail:", "a stopped job is not a broken engine: the page must not blame one")
+	assert.Empty(t, *calls, "and the second engine never starts")
+}
+
+func TestEnsembleProgress_TheReconcileCountsAreNotedForTheOperator(t *testing.T) {
+	rep := &recReporter{}
+	r := newEnsemble("Totalt 1015 kronor idag.", "Totalt 1050 kronor idag.", nil, nil, false, &verdictAdj{v: reconcile.VerdictAlt})
+
+	_, err := r.e.Transcribe(progress.With(context.Background(), rep), "/a.wav")
+
+	require.NoError(t, err)
+	require.Len(t, rep.notes, 1)
+	n := rep.notes[0]
+	assert.Equal(t, "reconciled", n["msg"])
+	assert.Equal(t, 1, n["disputes"])
+	assert.Equal(t, 1, n["changed"])
+	assert.Equal(t, 0, n["unsure"])
+	assert.Equal(t, 0, n["errors"])
+}

internal/adapters/transcribe/ensemble_reason_test.go

@@ -0,0 +1,36 @@
+package transcribe_test
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/transcribe"
+	"git.d-ma.be/mathias/tapir/internal/reconcile"
+)
+
+// When one engine is down the page says so and why; the "why" is a kind, not the
+// engine's error text, which can carry provider output (adversarial review F3).
+
+func TestEnsemble_AFailedEnginesReasonIsAKindNotItsErrorText(t *testing.T) {
+	r := newEnsemble("text", "text", nil, errors.New("kb-whisper said: zebrakaka secret words"), false, &verdictAdj{v: reconcile.VerdictAlt})
+
+	got, err := r.e.Transcribe(context.Background(), "/a.wav")
+
+	require.NoError(t, err)
+	assert.NotContains(t, got.Reconciliation.Reason, "zebrakaka")
+	assert.Contains(t, got.Reconciliation.Reason, "iguana/kb-whisper failed", "which engine failed is still said")
+}
+
+func TestEnsemble_AnHTTPFailureShowsItsStatusCodeAsTheKind(t *testing.T) {
+	r := newEnsemble("text", "text", &transcribe.StatusError{Code: 502, Snippet: "zebrakaka"}, nil, false, &verdictAdj{})
+
+	got, err := r.e.Transcribe(context.Background(), "/a.wav")
+
+	require.NoError(t, err)
+	assert.Contains(t, got.Reconciliation.Reason, "http_502")
+	assert.NotContains(t, got.Reconciliation.Reason, "zebrakaka")
+}

internal/adapters/transcribe/ensemble_test.go

@@ -0,0 +1,206 @@
+package transcribe_test
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/transcribe"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/reconcile"
+)
+
+// orderedBackend records, in a shared slice, when it was called.
+type orderedBackend struct {
+	name  string
+	text  string
+	err   error
+	calls *[]string
+}
+
+func (o *orderedBackend) Name() string { return o.name }
+func (o *orderedBackend) Transcribe(_ context.Context, _ string) (domain.Transcription, error) {
+	*o.calls = append(*o.calls, o.name)
+	if o.err != nil {
+		return domain.Transcription{}, o.err
+	}
+	return domain.Transcription{Text: o.text, Endpoint: o.name, AudioSeconds: 10}, nil
+}
+
+type verdictAdj struct {
+	v     reconcile.Verdict
+	err   error
+	calls int
+}
+
+func (a *verdictAdj) Adjudicate(context.Context, reconcile.Request) (reconcile.Verdict, error) {
+	a.calls++
+	return a.v, a.err
+}
+
+var ensembleOpts = reconcile.Options{Reconciler: "iguana/gemma4-26b", BaseEngine: "iguana/pianissimo", AltEngine: "iguana/kb-whisper"}
+
+type rig struct {
+	e     *transcribe.Ensemble
+	calls *[]string
+	adj   *verdictAdj
+}
+
+func newEnsemble(baseText, altText string, baseErr, altErr error, withFallback bool, adj *verdictAdj) *rig {
+	calls := &[]string{}
+	base := &orderedBackend{name: "iguana/pianissimo", text: baseText, err: baseErr, calls: calls}
+	alt := &orderedBackend{name: "iguana/kb-whisper", text: altText, err: altErr, calls: calls}
+	var fb transcribe.Backend
+	if withFallback {
+		fb = &orderedBackend{name: "berget/kb-whisper", text: "fallback text", calls: calls}
+	}
+	return &rig{e: transcribe.NewEnsemble(base, alt, fb, adj, ensembleOpts), calls: calls, adj: adj}
+}
+
+func TestEnsemble_BothEnginesRunInOrderAndTheirDisagreementIsReconciled(t *testing.T) {
+	adj := &verdictAdj{v: reconcile.VerdictAlt}
+	r := newEnsemble("Totalt 1015 kronor idag.", "Totalt 1050 kronor idag.", nil, nil, true, adj)
+
+	got, err := r.e.Transcribe(context.Background(), "/a.wav")
+
+	require.NoError(t, err)
+	assert.Equal(t, []string{"iguana/pianissimo", "iguana/kb-whisper"}, *r.calls,
+		"the 40-second engine runs first, so a failure there is known early; the fallback is not touched")
+	assert.Equal(t, "Totalt 1050 kronor idag.", got.Text)
+	assert.Equal(t, "iguana/pianissimo + iguana/kb-whisper, reconciled by iguana/gemma4-26b", got.Endpoint,
+		"every contributor is recorded: which engines touched the audio is a quality and a privacy fact")
+	require.NotNil(t, got.Reconciliation)
+	assert.Equal(t, domain.ReconcileDone, got.Reconciliation.Status)
+}
+
+func TestEnsemble_IdenticalTextsNeverCallTheModel(t *testing.T) {
+	adj := &verdictAdj{v: reconcile.VerdictAlt}
+	r := newEnsemble("Vi går igenom läget.", "Vi går igenom läget.", nil, nil, false, adj)
+
+	got, err := r.e.Transcribe(context.Background(), "/a.wav")
+
+	require.NoError(t, err)
+	assert.Zero(t, adj.calls)
+	assert.Equal(t, "Vi går igenom läget.", got.Text)
+}
+
+func TestEnsemble_WhenOneEngineFailsTheOtherIsUsedAndItIsSaidSo(t *testing.T) {
+	tests := []struct {
+		name          string
+		baseErr       error
+		altErr        error
+		wantText      string
+		wantEndpoint  string
+		wantReasonHas string
+	}{
+		{"pianissimo down", errors.New("pianissimo wedged"), nil, "alt text", "iguana/kb-whisper", "iguana/pianissimo failed"},
+		{"kb-whisper down", nil, errors.New("metal oom"), "base text", "iguana/pianissimo", "iguana/kb-whisper failed"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			adj := &verdictAdj{v: reconcile.VerdictAlt}
+			r := newEnsemble("base text", "alt text", tt.baseErr, tt.altErr, true, adj)
+
+			got, err := r.e.Transcribe(context.Background(), "/a.wav")
+
+			require.NoError(t, err)
+			assert.Equal(t, tt.wantText, got.Text)
+			assert.Equal(t, tt.wantEndpoint, got.Endpoint)
+			assert.Zero(t, adj.calls, "there is nothing to reconcile")
+			require.NotNil(t, got.Reconciliation)
+			assert.Equal(t, domain.ReconcileSingle, got.Reconciliation.Status)
+			assert.Contains(t, got.Reconciliation.Reason, tt.wantReasonHas)
+			assert.NotContains(t, *r.calls, "berget/kb-whisper", "one local engine working means audio does not leave the homelab")
+		})
+	}
+}
+
+func TestEnsemble_OnlyWhenBothLocalEnginesFailDoesTheCloudFallbackRun(t *testing.T) {
+	r := newEnsemble("", "", errors.New("a"), errors.New("b"), true, &verdictAdj{})
+
+	got, err := r.e.Transcribe(context.Background(), "/a.wav")
+
+	require.NoError(t, err)
+	assert.Equal(t, "fallback text", got.Text)
+	assert.Equal(t, "berget/kb-whisper", got.Endpoint, "the row must say berget served it: the audio left the homelab")
+	assert.Equal(t, domain.ReconcileSingle, got.Reconciliation.Status)
+	assert.Contains(t, got.Reconciliation.Reason, "both local engines failed")
+}
+
+func TestEnsemble_BothFailingWithNoFallbackReportsBothErrors(t *testing.T) {
+	r := newEnsemble("", "", errors.New("pianissimo down"), errors.New("whisper down"), false, &verdictAdj{})
+
+	_, err := r.e.Transcribe(context.Background(), "/a.wav")
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "pianissimo down")
+	assert.Contains(t, err.Error(), "whisper down")
+}
+
+func TestEnsemble_ACancelledJobNeverReachesTheCloud(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	r := newEnsemble("", "", context.Canceled, context.Canceled, true, &verdictAdj{})
+
+	_, err := r.e.Transcribe(ctx, "/a.wav")
+
+	assert.ErrorIs(t, err, context.Canceled)
+	assert.NotContains(t, *r.calls, "berget/kb-whisper", "a cancelled job must not ship its audio to a cloud endpoint as a last act")
+}
+
+func TestEnsemble_AModelThatIsDownGivesTheBaseTextUnreconciledNotAFailedJob(t *testing.T) {
+	adj := &verdictAdj{err: errors.New("gateway down")}
+	// Four separate disagreements, each separated by two shared words, so the
+	// breaker trips after three consecutive failures.
+	r := newEnsemble("a1 s t b1 s t c1 s t d1", "a2 s t b2 s t c2 s t d2", nil, nil, false, adj)
+
+	got, err := r.e.Transcribe(context.Background(), "/a.wav")
+
+	require.NoError(t, err)
+	assert.Equal(t, "a1 s t b1 s t c1 s t d1", got.Text)
+	assert.Equal(t, domain.ReconcileSkipped, got.Reconciliation.Status)
+	assert.Equal(t, 3, adj.calls)
+}
+
+// cancelling is a backend that cancels the job's context while it runs, the way a
+// shutdown arrives in the middle of a long transcription.
+type cancelling struct {
+	name   string
+	cancel context.CancelFunc
+	calls  *[]string
+}
+
+func (c *cancelling) Name() string { return c.name }
+func (c *cancelling) Transcribe(_ context.Context, _ string) (domain.Transcription, error) {
+	*c.calls = append(*c.calls, c.name)
+	c.cancel()
+	return domain.Transcription{}, context.Canceled
+}
+
+func TestEnsemble_AShutdownDuringTheFirstEngineStopsBeforeTheSecondOrTheCloud(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	calls := &[]string{}
+	base := &cancelling{name: "iguana/pianissimo", cancel: cancel, calls: calls}
+	alt := &orderedBackend{name: "iguana/kb-whisper", text: "alt", calls: calls}
+	fb := &orderedBackend{name: "berget/kb-whisper", text: "fb", calls: calls}
+
+	_, err := transcribe.NewEnsemble(base, alt, fb, &verdictAdj{}, ensembleOpts).Transcribe(ctx, "/a.wav")
+
+	assert.ErrorIs(t, err, context.Canceled)
+	assert.Equal(t, []string{"iguana/pianissimo"}, *calls, "no second engine and no cloud once the job is cancelled")
+}
+
+func TestEnsemble_AShutdownDuringTheSecondEngineIsAnErrorNotASingleEngineSuccess(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	calls := &[]string{}
+	base := &orderedBackend{name: "iguana/pianissimo", text: "base", calls: calls}
+	alt := &cancelling{name: "iguana/kb-whisper", cancel: cancel, calls: calls}
+
+	got, err := transcribe.NewEnsemble(base, alt, nil, &verdictAdj{}, ensembleOpts).Transcribe(ctx, "/a.wav")
+
+	assert.ErrorIs(t, err, context.Canceled, "a cancelled job must not be recorded as 'kb-whisper failed'")
+	assert.Empty(t, got.Text)
+}

internal/adapters/transcribe/gateway.go

@@ -0,0 +1,188 @@
+// Package transcribe turns a WAV file into a timed Swedish transcript by calling
+// the LiteLLM gateway's OpenAI-compatible /audio/transcriptions endpoint
+// (ADR-033). A Gateway is one model alias; a Chain tries a local one and falls
+// back to a second.
+package transcribe
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"mime/multipart"
+	"net/http"
+	"os"
+	"strings"
+
+	"git.d-ma.be/mathias/tapir/internal/domain"
+)
+
+// ErrEmptyTranscript is returned when the endpoint answered 200 but with no text.
+// An empty transcript is not a success: chat over it could only answer "nothing".
+var ErrEmptyTranscript = errors.New("transcribe: endpoint returned an empty transcript")
+
+// StatusError is a non-2xx answer from the gateway. Nothing may branch on the
+// body — {"error":"failed to process audio"} has had at least five distinct
+// causes — so only the status code is exposed for logic; Snippet is for logs.
+type StatusError struct {
+	Code    int
+	Snippet string
+}
+
+// HTTPStatus implements errkind.HTTPStatuser: the code, never the snippet.
+func (e *StatusError) HTTPStatus() int { return e.Code }
+
+func (e *StatusError) Error() string {
+	return fmt.Sprintf("transcribe: gateway returned HTTP %d: %s", e.Code, e.Snippet)
+}
+
+// Gateway transcribes through one model alias on the gateway. baseURL is the
+// gateway's /v1 root — the in-cluster NodePort, not the public hostname, whose
+// ingress rejects bodies over ~8-10 MB with 413 (docs/research/audio-length-ceiling).
+type Gateway struct {
+	baseURL string
+	key     string
+	model   string
+	client  *http.Client
+}
+
+// NewGateway builds a Gateway. The client should carry no overall Timeout: a 90
+// minute recording takes minutes, so the deadline belongs to the caller's context.
+func NewGateway(baseURL, key, model string, client *http.Client) *Gateway {
+	return &Gateway{baseURL: strings.TrimRight(baseURL, "/"), key: key, model: model, client: client}
+}
+
+// Name is the model alias; it is what gets recorded as the serving endpoint.
+func (g *Gateway) Name() string { return g.model }
+
+// Transcribe uploads the WAV at path and returns the normalised transcription.
+// The body is streamed from disk with a COMPUTED Content-Length: a 90 minute WAV
+// is ~170 MB, so it is neither buffered in memory nor sent chunked.
+func (g *Gateway) Transcribe(ctx context.Context, path string) (domain.Transcription, error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return domain.Transcription{}, fmt.Errorf("transcribe: open audio: %w", err)
+	}
+	defer func() { _ = f.Close() }()
+	info, err := f.Stat()
+	if err != nil {
+		return domain.Transcription{}, fmt.Errorf("transcribe: stat audio: %w", err)
+	}
+
+	var prefix bytes.Buffer
+	mw := multipart.NewWriter(&prefix)
+	for _, kv := range [][2]string{{"model", g.model}, {"response_format", "verbose_json"}, {"language", "sv"}} {
+		if err := mw.WriteField(kv[0], kv[1]); err != nil {
+			return domain.Transcription{}, fmt.Errorf("transcribe: build request: %w", err)
+		}
+	}
+	// The multipart field is `file`; `audio_file` is rejected as "Invalid request".
+	if _, err := mw.CreateFormFile("file", "audio.wav"); err != nil {
+		return domain.Transcription{}, fmt.Errorf("transcribe: build request: %w", err)
+	}
+	suffix := []byte("\r\n--" + mw.Boundary() + "--\r\n")
+
+	body := io.MultiReader(bytes.NewReader(prefix.Bytes()), f, bytes.NewReader(suffix))
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, g.baseURL+"/audio/transcriptions", body)
+	if err != nil {
+		return domain.Transcription{}, fmt.Errorf("transcribe: build request: %w", err)
+	}
+	req.ContentLength = int64(prefix.Len()) + info.Size() + int64(len(suffix))
+	req.Header.Set("Content-Type", mw.FormDataContentType())
+	req.Header.Set("Authorization", "Bearer "+g.key)
+	// No "Expect: 100-continue": it trips the proxy for any body over 1 KB.
+
+	resp, err := g.client.Do(req)
+	if err != nil {
+		return domain.Transcription{}, fmt.Errorf("transcribe: %s: %w", g.model, err)
+	}
+	defer func() { _ = resp.Body.Close() }()
+	if resp.StatusCode < 200 || resp.StatusCode > 299 {
+		snip, _ := io.ReadAll(io.LimitReader(resp.Body, 200))
+		return domain.Transcription{}, &StatusError{Code: resp.StatusCode, Snippet: strings.TrimSpace(string(snip))}
+	}
+
+	t, err := parseVerbose(resp.Body)
+	if err != nil {
+		return domain.Transcription{}, fmt.Errorf("transcribe: %s: %w", g.model, err)
+	}
+	t.Endpoint = g.model
+	return t, nil
+}
+
+// verboseJSON is the subset of the verbose_json response used. Both backends
+// return it; they differ in segment granularity (iguana: whole seconds and
+// sub-sentence pieces, berget: sub-second sentences), in the language token, and in
+// whether a segment's text starts with a space.
+type verboseJSON struct {
+	Text     string  `json:"text"`
+	Duration float64 `json:"duration"`
+	Segments []struct {
+		Start float64 `json:"start"`
+		End   float64 `json:"end"`
+		Text  string  `json:"text"`
+	} `json:"segments"`
+}
+
+// parseVerbose normalises a verbose_json body. Whisper's final window can end past
+// the audio (9.0 s for 4.25 s of audio was observed), so segment ends are clamped
+// to the reported duration. Text is rebuilt from the segments, not taken from the
+// response's `text` field: see joinSegments.
+func parseVerbose(r io.Reader) (domain.Transcription, error) {
+	var v verboseJSON
+	if err := json.NewDecoder(r).Decode(&v); err != nil {
+		return domain.Transcription{}, fmt.Errorf("decode response: %w", err)
+	}
+	raw := make([]string, 0, len(v.Segments))
+	for _, s := range v.Segments {
+		raw = append(raw, s.Text)
+	}
+	text := joinSegments(raw)
+	if text == "" {
+		text = collapseSpace(v.Text)
+	}
+	if text == "" {
+		return domain.Transcription{}, ErrEmptyTranscript
+	}
+	t := domain.Transcription{Text: text, AudioSeconds: v.Duration}
+	for _, s := range v.Segments {
+		seg := domain.Segment{Start: s.Start, End: s.End, Text: strings.TrimSpace(s.Text)}
+		if v.Duration > 0 {
+			seg.End = min(seg.End, v.Duration)
+			seg.Start = min(seg.Start, seg.End)
+		}
+		t.Segments = append(t.Segments, seg)
+	}
+	return t, nil
+}
+
+// joinSegments rebuilds the running text from raw (untrimmed) segment texts.
+//
+// whisper.cpp wraps its output at ~60 characters and a wrap can land INSIDE a word
+// ("nyt" | "t godkännande"): a segment that begins with a space starts a new word,
+// one that does not continues the previous one, and the response's own `text`
+// field joins them with newlines, which breaks the word. So when any segment
+// carries a leading space the raw texts are concatenated as they are. Backends
+// that return whole sentences with no leading space (berget) are joined with a
+// space instead — concatenating those would glue sentences together.
+func joinSegments(raw []string) string {
+	leadingSpace := false
+	for _, s := range raw {
+		if strings.HasPrefix(s, " ") {
+			leadingSpace = true
+			break
+		}
+	}
+	sep := " "
+	if leadingSpace {
+		sep = ""
+	}
+	return collapseSpace(strings.Join(raw, sep))
+}
+
+// collapseSpace turns every run of whitespace (including newlines) into one space.
+func collapseSpace(s string) string {
+	return strings.Join(strings.Fields(s), " ")
+}

internal/adapters/transcribe/gateway_test.go

@@ -0,0 +1,257 @@
+package transcribe_test
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/transcribe"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/errkind"
+)
+
+// The httptest server stands in for the LiteLLM gateway: it is the process
+// boundary, so a fake here is legitimate. The JSON it returns mirrors what the
+// real gateway returned for iguana/kb-whisper and berget/kb-whisper on 2026-10-01
+// (docs/research/audio-length-ceiling-2026-10-01.md). All text is synthetic.
+
+const verboseJSON = `{
+  "text": " God morgon allihopa.\n Vi går igenom läget.\n",
+  "language": "swedish",
+  "duration": 4.25,
+  "segments": [
+    {"id": 0, "start": 0.0, "end": 1.0, "text": " God morgon allihopa."},
+    {"id": 1, "start": 1.0, "end": 9.0, "text": " Vi går igenom läget."}
+  ]
+}`
+
+func wavFile(t *testing.T, content string) string {
+	t.Helper()
+	p := filepath.Join(t.TempDir(), "a.wav")
+	require.NoError(t, os.WriteFile(p, []byte(content), 0o600))
+	return p
+}
+
+type seen struct {
+	method, path, auth, expect, contentType string
+	contentLength                           int64
+	transferEncoding                        []string
+	fields                                  map[string]string
+	file                                    string
+	fileField                               string
+	bodyLen                                 int
+}
+
+// serve starts a gateway fake that records the request it received and replies
+// with status and body.
+func serve(t *testing.T, status int, reply string, got *seen) *httptest.Server {
+	t.Helper()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if got != nil {
+			got.method, got.path = r.Method, r.URL.Path
+			got.auth = r.Header.Get("Authorization")
+			got.expect = r.Header.Get("Expect")
+			got.contentType = r.Header.Get("Content-Type")
+			got.contentLength = r.ContentLength
+			got.transferEncoding = r.TransferEncoding
+			raw, _ := io.ReadAll(r.Body)
+			got.bodyLen = len(raw)
+			r.Body = io.NopCloser(bytes.NewReader(raw))
+			require.NoError(t, r.ParseMultipartForm(32<<20))
+			got.fields = map[string]string{}
+			for k, v := range r.MultipartForm.Value {
+				got.fields[k] = v[0]
+			}
+			for field, fhs := range r.MultipartForm.File {
+				got.fileField = field
+				f, _ := fhs[0].Open()
+				b, _ := io.ReadAll(f)
+				got.file = string(b)
+			}
+		}
+		w.WriteHeader(status)
+		_, _ = io.WriteString(w, reply)
+	}))
+	t.Cleanup(srv.Close)
+	return srv
+}
+
+func TestGateway_SendsTheRequestTheGatewayExpects(t *testing.T) {
+	var got seen
+	srv := serve(t, 200, verboseJSON, &got)
+	g := transcribe.NewGateway(srv.URL+"/v1", "secret-key", "iguana/kb-whisper", srv.Client())
+
+	_, err := g.Transcribe(context.Background(), wavFile(t, "RIFF-synthetic-pcm"))
+
+	require.NoError(t, err)
+	assert.Equal(t, http.MethodPost, got.method)
+	assert.Equal(t, "/v1/audio/transcriptions", got.path)
+	assert.Equal(t, "Bearer secret-key", got.auth)
+	assert.Equal(t, "file", got.fileField, "the multipart field is `file`; `audio_file` is rejected as 'Invalid request'")
+	assert.Equal(t, "RIFF-synthetic-pcm", got.file)
+	assert.Equal(t, "iguana/kb-whisper", got.fields["model"])
+	assert.Equal(t, "verbose_json", got.fields["response_format"])
+	assert.Equal(t, "sv", got.fields["language"])
+	assert.Empty(t, got.expect, "no Expect: 100-continue: it trips the proxy for any body over 1 KB")
+}
+
+func TestGateway_SendsAKnownContentLengthNotChunked(t *testing.T) {
+	var got seen
+	srv := serve(t, 200, verboseJSON, &got)
+	g := transcribe.NewGateway(srv.URL+"/v1", "k", "m", srv.Client())
+
+	_, err := g.Transcribe(context.Background(), wavFile(t, "0123456789"))
+
+	require.NoError(t, err)
+	assert.Empty(t, got.transferEncoding, "a 90-minute WAV is ~170 MB: stream it with a computed length, not chunked and not buffered")
+	assert.Equal(t, int64(got.bodyLen), got.contentLength)
+}
+
+func TestGateway_NormalisesTheVerboseJSONResponse(t *testing.T) {
+	srv := serve(t, 200, verboseJSON, nil)
+	g := transcribe.NewGateway(srv.URL+"/v1", "k", "iguana/kb-whisper", srv.Client())
+
+	got, err := g.Transcribe(context.Background(), wavFile(t, "x"))
+
+	require.NoError(t, err)
+	assert.Equal(t, "God morgon allihopa. Vi går igenom läget.", got.Text)
+	assert.Equal(t, "iguana/kb-whisper", got.Endpoint, "the serving endpoint is part of the result")
+	assert.InDelta(t, 4.25, got.AudioSeconds, 1e-9)
+	assert.Equal(t, []domain.Segment{
+		{Start: 0, End: 1, Text: "God morgon allihopa."},
+		// The final segment ended at 9.0 s for 4.25 s of audio: whisper's last
+		// window can run past the end. Clamped to the audio duration.
+		{Start: 1, End: 4.25, Text: "Vi går igenom läget."},
+	}, got.Segments)
+}
+
+func TestGateway_ErrorsDoNotDependOnTheBody(t *testing.T) {
+	// {"error":"failed to process audio"} has had at least five distinct causes, so
+	// no code may branch on it. Two different bodies must be treated identically.
+	for _, reply := range []string{`{"error":"failed to process audio"}`, `{"error":"something else entirely"}`} {
+		srv := serve(t, 500, reply, nil)
+		g := transcribe.NewGateway(srv.URL+"/v1", "k", "m", srv.Client())
+
+		_, err := g.Transcribe(context.Background(), wavFile(t, "x"))
+
+		var se *transcribe.StatusError
+		require.ErrorAs(t, err, &se)
+		assert.Equal(t, 500, se.Code)
+	}
+}
+
+func TestGateway_AnEmptyTranscriptIsAFailureNotASuccess(t *testing.T) {
+	srv := serve(t, 200, `{"text":"  ","language":"swedish","duration":3,"segments":[]}`, nil)
+	g := transcribe.NewGateway(srv.URL+"/v1", "k", "m", srv.Client())
+
+	_, err := g.Transcribe(context.Background(), wavFile(t, "x"))
+
+	assert.ErrorIs(t, err, transcribe.ErrEmptyTranscript)
+}
+
+func TestGateway_MalformedJSONIsAnError(t *testing.T) {
+	srv := serve(t, 200, `not json`, nil)
+	g := transcribe.NewGateway(srv.URL+"/v1", "k", "m", srv.Client())
+
+	_, err := g.Transcribe(context.Background(), wavFile(t, "x"))
+
+	assert.Error(t, err)
+}
+
+func TestGateway_MissingAudioFileIsAnError(t *testing.T) {
+	g := transcribe.NewGateway("http://127.0.0.1:1/v1", "k", "m", http.DefaultClient)
+
+	_, err := g.Transcribe(context.Background(), filepath.Join(t.TempDir(), "nope.wav"))
+
+	assert.Error(t, err)
+}
+
+// whisper.cpp wraps its output at ~60 characters, and a wrap can land INSIDE a word:
+// the continuation segment has no leading space, the `text` field joins segments
+// with newlines, and a segment can be a lone ".". Measured on real gateway output
+// 2026-10-01: 7% of a one-hour recording's segments were such fragments. Joining by
+// newline hands the chat model broken words ("nyt\nt"), so Text is rebuilt from the
+// segments: a segment that begins with a space starts a new word, one that does not
+// continues the previous one.
+func TestGateway_RejoinsWordsWhisperCppSplitAcrossSegments(t *testing.T) {
+	reply := `{
+	  "text": " Vi ligger just nu 12 procent över den ursprungliga kalkylen\n.\n Det kräver ett nyt\nt godkännande från beställaren.\n Framförallt på arbetspl\natsen.\n",
+	  "language": "swedish", "duration": 20,
+	  "segments": [
+	    {"start": 0, "end": 4, "text": " Vi ligger just nu 12 procent över den ursprungliga kalkylen"},
+	    {"start": 4, "end": 5, "text": "."},
+	    {"start": 5, "end": 8, "text": " Det kräver ett nyt"},
+	    {"start": 8, "end": 11, "text": "t godkännande från beställaren."},
+	    {"start": 11, "end": 14, "text": " Framförallt på arbetspl"},
+	    {"start": 14, "end": 16, "text": "atsen."}
+	  ]}`
+	srv := serve(t, 200, reply, nil)
+	g := transcribe.NewGateway(srv.URL+"/v1", "k", "iguana/kb-whisper", srv.Client())
+
+	got, err := g.Transcribe(context.Background(), wavFile(t, "x"))
+
+	require.NoError(t, err)
+	assert.Equal(t, "Vi ligger just nu 12 procent över den ursprungliga kalkylen. "+
+		"Det kräver ett nytt godkännande från beställaren. Framförallt på arbetsplatsen.", got.Text)
+	assert.Len(t, got.Segments, 6, "the timed segments are kept as the backend gave them (trimmed)")
+	assert.Equal(t, ".", got.Segments[1].Text)
+}
+
+func TestGateway_SentenceLevelSegmentsWithNoLeadingSpaceAreJoinedWithASpace(t *testing.T) {
+	// berget/kb-whisper returns whole sentences with no leading space; rebuilding by
+	// concatenation would glue them ("allihopa!Vi").
+	reply := `{"text": "God morgon allihopa!\nVi går igenom läget.", "language": "sv", "duration": 5,
+	  "segments": [
+	    {"start": 0.05, "end": 1.2, "text": "God morgon allihopa!"},
+	    {"start": 1.6, "end": 4.2, "text": "Vi går igenom läget."}
+	  ]}`
+	srv := serve(t, 200, reply, nil)
+	g := transcribe.NewGateway(srv.URL+"/v1", "k", "berget/kb-whisper", srv.Client())
+
+	got, err := g.Transcribe(context.Background(), wavFile(t, "x"))
+
+	require.NoError(t, err)
+	assert.Equal(t, "God morgon allihopa! Vi går igenom läget.", got.Text)
+}
+
+func TestGateway_WithNoSegmentsTheTextFieldIsUsedWithWhitespaceCollapsed(t *testing.T) {
+	srv := serve(t, 200, `{"text": "  Hej  allihopa.\n Vi börjar.  ", "duration": 3, "segments": []}`, nil)
+	g := transcribe.NewGateway(srv.URL+"/v1", "k", "m", srv.Client())
+
+	got, err := g.Transcribe(context.Background(), wavFile(t, "x"))
+
+	require.NoError(t, err)
+	assert.Equal(t, "Hej allihopa. Vi börjar.", got.Text)
+}
+
+func TestStatusError_ExposesItsCodeAsAKindAndNotItsSnippet(t *testing.T) {
+	err := fmt.Errorf("transcribe: %w", &transcribe.StatusError{Code: 502, Snippet: "echoed: Anders ansvarar"})
+
+	assert.Equal(t, "http_502", errkind.Of(err))
+}
+
+func TestGateway_AnErrorBodyIsBoundedToTwoHundredBytesInTheStatusError(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusInternalServerError)
+		_, _ = w.Write([]byte(strings.Repeat("x", 100_000)))
+	}))
+	defer srv.Close()
+	wav := filepath.Join(t.TempDir(), "a.wav")
+	require.NoError(t, os.WriteFile(wav, []byte("RIFF....WAVEfmt "), 0o600))
+
+	_, err := transcribe.NewGateway(srv.URL, "", "iguana/kb-whisper", srv.Client()).Transcribe(context.Background(), wav)
+
+	var se *transcribe.StatusError
+	require.ErrorAs(t, err, &se)
+	assert.LessOrEqual(t, len(se.Snippet), 200, "a gateway's whole body must not become part of an error")
+}

internal/adapters/transcribe/live_test.go

@@ -0,0 +1,49 @@
+package transcribe_test
+
+import (
+	"context"
+	"net/http"
+	"os"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/transcribe"
+)
+
+// TestGateway_LiveGateway is an opt-in check against the REAL gateway, because
+// the httptest fake above can only prove what we think the gateway accepts. It
+// is skipped unless TAPIR_LIVE_WAV names a 16 kHz mono WAV and the gateway
+// coordinates are set:
+//
+//	TAPIR_LIVE_WAV=/path/a.wav TAPIR_LIVE_GATEWAY_URL=http://koala:30401/v1 \
+//	TAPIR_LIVE_MODEL=iguana/kb-whisper op run --env-file ~/.op-env -- \
+//	  go test ./internal/adapters/transcribe -run LiveGateway -v
+//
+// Use SYNTHETIC audio only: this repo is public and the gateway may be a cloud
+// endpoint (#33).
+func TestGateway_LiveGateway(t *testing.T) {
+	wav := os.Getenv("TAPIR_LIVE_WAV")
+	url := os.Getenv("TAPIR_LIVE_GATEWAY_URL")
+	if wav == "" || url == "" {
+		t.Skip("set TAPIR_LIVE_WAV and TAPIR_LIVE_GATEWAY_URL to run against the real gateway")
+	}
+	model := os.Getenv("TAPIR_LIVE_MODEL")
+	if model == "" {
+		model = "iguana/kb-whisper"
+	}
+	key := os.Getenv("LITELLM_MASTER_KEY")
+	require.NotEmpty(t, key, "LITELLM_MASTER_KEY must be set (via op run)")
+
+	ctx, cancel := context.WithTimeout(context.Background(), 25*time.Minute)
+	defer cancel()
+	got, err := transcribe.NewGateway(url, key, model, &http.Client{}).Transcribe(ctx, wav)
+
+	require.NoError(t, err)
+	assert.NotEmpty(t, got.Text)
+	assert.NotEmpty(t, got.Segments)
+	assert.Equal(t, model, got.Endpoint)
+	t.Logf("endpoint=%s audio_s=%.1f segments=%d chars=%d", got.Endpoint, got.AudioSeconds, len(got.Segments), len(got.Text))
+}

internal/adapters/transcribe/timeout.go

@@ -0,0 +1,33 @@
+package transcribe
+
+import (
+	"context"
+	"time"
+
+	"git.d-ma.be/mathias/tapir/internal/domain"
+)
+
+// WithTimeout gives every call to b its own deadline. A wedged engine does not answer at
+// all (iguana after a Metal OOM, infra#273), and with only the job's deadline it used up
+// the whole job before a fallback could run -- and Chain then refuses the fallback because
+// the job context has expired. With its own deadline the call fails while the job is still
+// alive, which is exactly when the fallback, or the other engine, can take over.
+//
+// A zero or negative d returns b unchanged. The wrapper is invisible: Name is b's.
+func WithTimeout(b Backend, d time.Duration) Backend {
+	if d <= 0 {
+		return b
+	}
+	return timeoutBackend{Backend: b, d: d}
+}
+
+type timeoutBackend struct {
+	Backend
+	d time.Duration
+}
+
+func (t timeoutBackend) Transcribe(ctx context.Context, wavPath string) (domain.Transcription, error) {
+	ctx, cancel := context.WithTimeout(ctx, t.d)
+	defer cancel()
+	return t.Backend.Transcribe(ctx, wavPath)
+}

internal/adapters/transcribe/timeout_test.go

@@ -0,0 +1,104 @@
+package transcribe_test
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/transcribe"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/errkind"
+	"git.d-ma.be/mathias/tapir/internal/reconcile"
+)
+
+// A wedged engine (iguana after a Metal OOM, infra#273) does not answer at all. With only
+// the job's deadline, the primary used up the whole job before any fallback could run, and
+// the chain then refused the fallback because the job context had expired. Each engine call
+// gets its own deadline (adversarial review F4).
+
+type hangingBackend struct{ name string }
+
+func (h hangingBackend) Name() string { return h.name }
+func (hangingBackend) Transcribe(ctx context.Context, _ string) (domain.Transcription, error) {
+	<-ctx.Done()
+	return domain.Transcription{}, ctx.Err()
+}
+
+func TestWithTimeout_AHungBackendFailsAtItsOwnDeadlineWhileTheJobIsStillAlive(t *testing.T) {
+	job, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+	b := transcribe.WithTimeout(hangingBackend{name: "iguana/kb-whisper"}, 40*time.Millisecond)
+
+	start := time.Now()
+	_, err := b.Transcribe(job, "/a.wav")
+
+	require.Error(t, err)
+	assert.Equal(t, "timeout", errkind.Of(err))
+	assert.Less(t, time.Since(start), 2*time.Second, "it must give up at its own deadline, not wait for the job's")
+	assert.NoError(t, job.Err(), "the job context is untouched, which is what lets a fallback run")
+	assert.Equal(t, "iguana/kb-whisper", b.Name(), "the wrapper is invisible in the recorded endpoint")
+}
+
+func TestWithTimeout_ASuccessPassesStraightThrough(t *testing.T) {
+	calls := &[]string{}
+	inner := &orderedBackend{name: "iguana/kb-whisper", text: "hej", calls: calls}
+
+	got, err := transcribe.WithTimeout(inner, time.Minute).Transcribe(context.Background(), "/a.wav")
+
+	require.NoError(t, err)
+	assert.Equal(t, "hej", got.Text)
+}
+
+func TestWithTimeout_ZeroMeansNoDeadlineOfItsOwn(t *testing.T) {
+	inner := &orderedBackend{name: "x", text: "hej", calls: &[]string{}}
+
+	assert.Same(t, inner, transcribe.WithTimeout(inner, 0), "a zero timeout must not wrap or change anything")
+}
+
+func TestChain_AHungPrimaryReachesTheFallbackWhenItHasItsOwnDeadline(t *testing.T) {
+	calls := &[]string{}
+	fb := &orderedBackend{name: "berget/kb-whisper", text: "fallback text", calls: calls}
+	c := transcribe.NewChain(transcribe.WithTimeout(hangingBackend{name: "iguana/kb-whisper"}, 40*time.Millisecond), fb)
+	job, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+
+	got, err := c.Transcribe(job, "/a.wav")
+
+	require.NoError(t, err, "a wedged primary is exactly what the fallback exists for")
+	assert.Equal(t, "berget/kb-whisper", got.Endpoint)
+}
+
+func TestChain_AJobThatEndedWhileThePrimaryHungStillNeverReachesTheFallback(t *testing.T) {
+	calls := &[]string{}
+	fb := &orderedBackend{name: "berget/kb-whisper", text: "fallback text", calls: calls}
+	c := transcribe.NewChain(transcribe.WithTimeout(hangingBackend{name: "iguana/kb-whisper"}, time.Minute), fb)
+	job, cancel := context.WithTimeout(context.Background(), 40*time.Millisecond)
+	defer cancel()
+
+	_, err := c.Transcribe(job, "/a.wav")
+
+	require.Error(t, err)
+	assert.Empty(t, *calls, "the job's own deadline or a cancel must still stop the audio from reaching a cloud endpoint")
+}
+
+func TestEnsemble_AHungEngineIsASoftFailureAndTheOtherEnginesTextIsUsed(t *testing.T) {
+	calls := &[]string{}
+	alt := &orderedBackend{name: "iguana/kb-whisper", text: "text from the other engine", calls: calls}
+	e := transcribe.NewEnsemble(
+		transcribe.WithTimeout(hangingBackend{name: "iguana/pianissimo"}, 40*time.Millisecond), alt, nil,
+		&verdictAdj{v: reconcile.VerdictAlt}, ensembleOpts)
+	job, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+
+	got, err := e.Transcribe(job, "/a.wav")
+
+	require.NoError(t, err)
+	assert.Equal(t, "text from the other engine", got.Text)
+	assert.Equal(t, domain.ReconcileSingle, got.Reconciliation.Status)
+	assert.Contains(t, got.Reconciliation.Reason, "timeout")
+	assert.False(t, errors.Is(err, context.DeadlineExceeded))
+}