fix(deps): bump x/text past GO-2026-5970

govulncheck reported GO-2026-5970 as called via golang.org/x/text v0.31.0
(reached through the DB driver). Bump to v0.39.0; x/sync moves to v0.21.0
as its required minimum. govulncheck ./...: 0 vulnerabilities.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

DECISIONS.md

@@ -1556,7 +1556,9 @@ deferred. The scope cut — chat first, no rich analysis, no video — is what k
   mechanism), confirm-then-delete on the recording's page. A recording that is **queued or
   transcribing is refused (409) until it ends**, because the worker would otherwise write to a
   row that is gone; a job whose service was lost stays undeletable until the next start's sweep
-  marks it failed. The home list now links to `/uploads` when uploads are mounted.
+  marks it failed. The refusal reflects the row's state when the delete was attempted: a
+  recording that finishes in the same instant still gets a 409 once, and a retry deletes it.
+  The home list now links to `/uploads` when uploads are mounted.
 
 **Reversibility.** Additive: one table (down-migration drops it), one adapter, one upload
 surface, one image change. Removing the upload route removes the feature; `transcripts`, `videos`

docs/research/reconcile-second-check-protocol.md

@@ -0,0 +1,69 @@
+# ADR-034 second blind check: protocol (ready to run; tapir#33, D9, D10)
+
+This repo is public and the recordings are private, so this file holds the method only. The
+sheet, the key and the filled-in verdicts stay outside the repo, and the trial data is deleted
+after the check. Results go back here as aggregate numbers, as in
+`reconcile-blind-check-2026-10-02.md`.
+
+## What the check decides
+
+ADR-034 (two engines, a local model reconciles their disagreements) is `Proposed` and off in
+production. Two conditions were set for keeping it. The first run answered the first one on one
+reading only and did not answer the second:
+
+1. **The reconciled reading is right at least 85% of the time.** Open flaw in the bar: it did not
+   say how ties count. The first run gave 95% (decided items only) or 70% (all 30), so it was met on
+   one reading and missed on the other.
+2. **It is no worse than the better single engine.** Not answered: before ADR-035 the spots the
+   model *kept* as the base were not stored, so kb-whisper alone could not be compared.
+
+ADR-035 now stores every judged spot (`reconciliation.contested`: both readings, which one was
+used, unsure or not, a few words of left context, a time). That makes condition 2 measurable.
+
+## Rules to fix BEFORE looking at any result (the owner confirms or changes them)
+
+| Rule | Proposed |
+|---|---|
+| Who judges | The speaker, hearing the audio where the text cannot settle it (D9). Not Opus: it cannot hear, and its agreement with the speaker was only shown on items picked as decidable. |
+| Sample | 30 spots drawn at random with a fixed seed, from ALL judged spots (changed and kept-base alike), not only the changed ones. Insertions and deletions that cannot be placed are listed separately and not sampled. |
+| Verdict per spot | `x`, `y` (the two readings, shown in random order, key kept apart), `neither`, or `cannot tell`. |
+| Condition 1 | Reconciled is right on at least 85% of the spots where the judge could tell (`cannot tell` excluded, reported, and `neither` counts as not right). Also report all-30. Pass needs the decided-only figure at 85% or more AND the all-30 figure above 70%. |
+| Condition 2 | Over the same spots, the reconciled text is right at least as often as the better single engine (base engine alone is right where the base reading is right; the other engine alone where the alt reading is right). Ties favour the single engine. |
+| If it fails | Switch reconcile off by configuration; the single-engine path is unchanged (ADR-034 reversibility). |
+
+Why a fixed all-30 floor: the literal 70% last time sat well under the bar, and "decided only" can
+hide a model that mostly lands on spelling differences.
+
+## What a run needs (D10: the owner is present)
+
+- The same real recording as before (about 60 minutes). Only the owner has it.
+- A **throwaway** tapir instance with its own Postgres, not the production pod, so production's
+  key pin and deployment are untouched. Settings: `TAPIR_RECONCILE_BASE_MODEL=iguana/pianissimo`,
+  `TAPIR_TRANSCRIBE_MODEL=iguana/kb-whisper`, `TAPIR_RECONCILE_MODEL=iguana/gemma4-31b`, cloud
+  fallbacks empty. Job time was 21 min 49 s per 63.6 min last run.
+- iguana with Low Power Mode off and gemma4-31b able to load (the production-iguana model load
+  that needs a quiet window; D6).
+
+## Steps
+
+1. Start the throwaway instance, upload the recording, wait for `done` (about 22 minutes).
+2. Read `reconciliation.contested` from the throwaway database. From it build, **outside the repo**:
+   `sheet` (spot number, time, left context, reading x, reading y, random order) and `key` (which
+   of x/y is the base, which is the other engine, which one the reconciler used).
+3. The speaker fills in the sheet (about 20 to 30 minutes for 30 spots, audio open at the times).
+   The agent does not see the key until the sheet is returned.
+4. Score with the rules above. Write the aggregate numbers into a dated note next to this one,
+   and amend ADR-034's validation section (kept, or switched off, and the tie rule used).
+5. Delete the throwaway instance, its database and the sheet, as last time.
+6. Only if the check passes and the owner wants it in production: enabling reconcile there is
+   the separate D10 step (it needs `iguana/pianissimo` and the reconciler added to the tapir key
+   pin in the infra repo plus the env vars on the deployment, `risk/medium`, owner present).
+
+## Open and owner-only
+
+- Confirm or change the rules table above before the run.
+- The **tool that builds the sheet and key** is the one piece not in the repo (last time it was a
+  one-off script). It can be built test-first on request; until then the agent builds it ad hoc
+  outside the repo.
+- A Berget adjudicator (D8, tapir#45) is a different question with its own measurement and is
+  not part of this check: this check judges the local setup that ADR-034 describes.

docs/sprints/audio-upload-followups-2026-10.md

@@ -1,7 +1,8 @@
 # Sprint: audio upload follow-ups (Sweep Sprint, 2026-10)
 
 Method: Sweep Sprint (phase 1 parallel read-only research per issue: **done**; phase 2 sequential
-execution, one issue per commit: **not started, waits for the owner's go**). Closing oath below.
+execution, one issue per commit: **v0.31.1 and v0.32.0 shipped 2026-10-03/04, the owner-held items
+below remain open**). Closing oath below.
 
 Milestones: [v0.31.1 - security and robustness](https://git.d-ma.be/mathias/tapir/milestone/10),
 [v0.32.0 - audio upload follow-ups](https://git.d-ma.be/mathias/tapir/milestone/11).
@@ -63,8 +64,8 @@ Release v0.31.1: one adversarial review round over 2-5, then tag, push, observe.
 | D5 | #38: Berget length measurement needs a throwaway key | Yes, throwaway key | #38 |
 | D6 | #38: load large models on production iguana | Yes, in a quiet window with the owner present | #38 |
 | D7 | #46: where the ffmpeg tarball lives | Do nothing now; file an issue | none |
-| D8 | #45: evaluate Berget as the reconciler | No; stay local-only | #45 |
-| D9 | Who judges ADR-034's second blind check | The owner; do not reuse the one-off Opus authorisation | ADR-034 |
+| D8 | #45: evaluate Berget as the reconciler | **Decided 2026-10-04 by the owner: yes, Berget may be used, and each user must be able to keep a recording on local models only** (reverses the recommendation). Opt-out first: #60; #45 stays as the model measurement | #60, #45 |
+| D9 | Who judges ADR-034's second blind check | The owner; do not reuse the one-off Opus authorisation. Protocol: `docs/research/reconcile-second-check-protocol.md` | ADR-034 |
 | D10 | When to enable reconcile in production | Needs `iguana/pianissimo` in the key pin and env vars; owner present | ADR-034 |
 
 ## Dropped or reduced, with reasons
@@ -86,3 +87,29 @@ Release v0.31.1: one adversarial review round over 2-5, then tag, push, observe.
 SemVer: v0.31.1 patch (fixes, no new behaviour); v0.32.0 minor (nav link, delete, cloud labelling).
 Every push to tapir main redeploys the pod and fails any job in flight: push when no upload is
 running (check the pod log first).
+
+## Status (2026-10-04)
+
+**Shipped and deployed:** v0.31.1 (PR #55, merge e037c6b) and v0.32.0 (PR #58, merge 85f4cc3; pod
+tapir-597dbfccc-ng78q 1/1, 0 restarts, image 85f4cc3). Each went through an adversarial review
+round by agent with the verdicts recorded on the PR (v0.31.1: BLOCK then APPROVE; v0.32.0: BLOCK,
+BLOCK, APPROVE, APPROVE; every block was a real defect, fixed test-first).
+
+**Live evidence:** a foreign-origin `POST` to `/account/delete` and to `/uploads/{id}/delete` is
+refused (403) on production, a same-origin one reaches the login redirect (302).
+
+**Not observed live (needs an authenticated session, so it is the owner's look; closed on word or screenshot):**
+
+- #52, cloud side: a transcript served by the Berget fallback names `berget`; a reconciled all-local
+  one shows no note. The local side was observed on 2026-10-04 (a new clip served by iguana: chip
+  `iguana/…`, no cloud note). The cloud side is carried as an acceptance item in #60.
+
+**Observed live since:** #37 (owner, 2026-10-04: link, confirm-then-delete, no control while transcribing; closed). #50: at the 2026-10-04 rollout the old pod went `Terminating` then `Completed`
+(not `Error`), 0 restarts; #50 closed. No job was in flight, so a drained job is still unobserved
+(a recorded limitation, D2).
+
+**Still open (owner's call, milestone 11):** #36, #38, #45 (now the Berget adjudicator measurement), #60 (per-recording local-only choice, 4 sub-decisions), #46 remainder, #56 (slow-drip bodies hold
+a receive slot), #57 (shutdown logging, second signal). Decisions D4-D6, D9, D10 above are unchanged; D8 is decided (above).
+
+**Deferred and filed from the reviews:** #56, #57, #59 (start-up exits when the identity provider's OIDC
+discovery returns 503 (seen once at 19:40 on 2026-10-03; the pod restarted itself and recovered)).

go.mod

@@ -33,9 +33,9 @@ require (
 	github.com/rogpeppe/go-internal v1.15.0 // indirect
 	github.com/xi2/xz v0.0.0-20171230120015-48954b6210f8 // indirect
 	go.yaml.in/yaml/v2 v2.4.2 // indirect
-	golang.org/x/sync v0.18.0 // indirect
+	golang.org/x/sync v0.21.0 // indirect
 	golang.org/x/sys v0.41.0 // indirect
-	golang.org/x/text v0.31.0 // indirect
+	golang.org/x/text v0.39.0 // indirect
 	google.golang.org/protobuf v1.36.8 // indirect
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -113,12 +113,12 @@ go.yaml.in/yaml/v2 v2.4.2 h1:DzmwEr2rDGHl7lsFgAHxmNz/1NlQ7xLIrlN2h5d1eGI=
 go.yaml.in/yaml/v2 v2.4.2/go.mod h1:081UH+NErpNdqlCXm3TtEran0rJZGxAYx9hb/ELlsPU=
 golang.org/x/oauth2 v0.36.0 h1:peZ/1z27fi9hUOFCAZaHyrpWG5lwe0RJEEEeH0ThlIs=
 golang.org/x/oauth2 v0.36.0/go.mod h1:YDBUJMTkDnJS+A4BP4eZBjCqtokkg1hODuPjwiGPO7Q=
-golang.org/x/sync v0.18.0 h1:kr88TuHDroi+UVf+0hZnirlk8o8T+4MrK6mr60WkH/I=
-golang.org/x/sync v0.18.0/go.mod h1:9KTHXmSnoGruLpwFjVSX0lNNA75CykiMECbovNTZqGI=
+golang.org/x/sync v0.21.0 h1:HLII4xRRTtCRkxYp4HNFF0Js/Og6q2i++KXbg0gHCwM=
+golang.org/x/sync v0.21.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
 golang.org/x/sys v0.41.0 h1:Ivj+2Cp/ylzLiEU89QhWblYnOE9zerudt9Ftecq2C6k=
 golang.org/x/sys v0.41.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
-golang.org/x/text v0.31.0 h1:aC8ghyu4JhP8VojJ2lEHBnochRno1sgL6nEi9WGFGMM=
-golang.org/x/text v0.31.0/go.mod h1:tKRAlv61yKIjGGHX/4tP1LTbc13YSec1pxVEWXzfoeM=
+golang.org/x/text v0.39.0 h1:UbZz4pLOvn600D6Oh6GGEI6VAmndrEBLv8/6BEXzyus=
+golang.org/x/text v0.39.0/go.mod h1:3UwRclnC2g0TU9x8PZiyfOajCd1zaUNHF9cvqcQZ+ZM=
 golang.org/x/time v0.15.0 h1:bbrp8t3bGUeFOx08pvsMYRTCVSMk89u4tKbNOZbp88U=
 golang.org/x/time v0.15.0/go.mod h1:Y4YMaQmXwGQZoFaVFk4YpCt4FLQMYKZe9oeV/f4MSno=
 google.golang.org/protobuf v1.36.8 h1:xHScyCOEuuwZEc6UtSOvPbAT4zRh0xcNRYekJwfqyMc=

internal/adapters/llm/bodycap_test.go

@@ -79,3 +79,78 @@ func TestClient_AFailingBodyIsReadOnlyUpToTheErrorBound(t *testing.T) {
 		t.Fatalf("the client kept reading an endless error body: took %s", time.Since(start))
 	}
 }
+
+// sizedGateway answers 200 with a valid completion padded with trailing whitespace to
+// exactly size bytes, so the boundary of the cap can be probed to the byte.
+func sizedGateway(t *testing.T, size int) *httptest.Server {
+	t.Helper()
+	const base = `{"choices":[{"message":{"content":"ok"}}]}`
+	payload := base + strings.Repeat(" ", size-len(base))
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(payload))
+	}))
+	t.Cleanup(srv.Close)
+	return srv
+}
+
+func TestClient_ABodyOfExactlyTheCapIsReadAndOneByteMoreIsRefused(t *testing.T) {
+	const capBytes = 4096
+	for _, tt := range []struct {
+		name    string
+		size    int
+		wantErr bool
+	}{
+		{"one under", capBytes - 1, false},
+		{"exactly the cap", capBytes, false},
+		{"one over", capBytes + 1, true},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			c := New(sizedGateway(t, tt.size).URL, "", "m", 5*time.Second, WithMaxResponseBytes(capBytes))
+
+			got, err := c.Complete(context.Background(), "s", "u")
+
+			if tt.wantErr {
+				if err == nil || !strings.Contains(err.Error(), "exceeds") {
+					t.Fatalf("want a response-too-large error, got %q, %v", got, err)
+				}
+				return
+			}
+			if err != nil || got != "ok" {
+				t.Fatalf("want ok, got %q, %v", got, err)
+			}
+		})
+	}
+}
+
+// New is what production calls: no option. The default is therefore the live bound.
+func TestClient_TheDefaultCapIsEightMiBWhenNoOptionIsGiven(t *testing.T) {
+	for _, tt := range []struct {
+		name    string
+		size    int
+		wantErr bool
+	}{
+		{"exactly 8 MiB", 8 << 20, false},
+		{"8 MiB plus one", 8<<20 + 1, true},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			c := New(sizedGateway(t, tt.size).URL, "", "m", 10*time.Second)
+
+			_, err := c.Complete(context.Background(), "s", "u")
+
+			if (err != nil) != tt.wantErr {
+				t.Fatalf("size %d: err = %v, wantErr %v", tt.size, err, tt.wantErr)
+			}
+		})
+	}
+}
+
+func TestClient_ANonPositiveCapOptionKeepsTheDefault(t *testing.T) {
+	for _, n := range []int64{0, -1} {
+		c := New(sizedGateway(t, 1024).URL, "", "m", 5*time.Second, WithMaxResponseBytes(n))
+
+		if got, err := c.Complete(context.Background(), "s", "u"); err != nil || got != "ok" {
+			t.Fatalf("WithMaxResponseBytes(%d): want ok from the default cap, got %q, %v", n, got, err)
+		}
+	}
+}

internal/adapters/llm/client.go

@@ -202,6 +202,7 @@ func (c *Client) Complete(ctx context.Context, system, user string) (string, err
 
 	if resp.StatusCode != http.StatusOK {
 		// Only the bounded head is ever used (HTTPError.Error truncates to it), so read no more.
+		// A read error here only shortens a diagnostic; the status code below is what callers act on.
 		head, _ := io.ReadAll(io.LimitReader(resp.Body, maxErrorBody+1))
 		return "", &HTTPError{Code: resp.StatusCode, Body: string(head)}
 	}

internal/progress/tracker.go

@@ -57,8 +57,11 @@ func WithClock(now func() time.Time) Option { return func(t *Tracker) { t.now =
 func WithMinSave(d time.Duration) Option { return func(t *Tracker) { t.minSave = d } }
 
 // WithSaveTimeout bounds one save or heartbeat write. Progress is cosmetic, so a database
-// that stops answering costs a dropped write, never a stalled job: saves run under the
-// tracker's lock, which Begin and Advance (called by the engines) need (#51).
+// that stops answering costs a dropped write rather than an unbounded stall. It is still a
+// delay: saves run under the tracker's lock, which Begin and Advance (called by the engines)
+// need, so each transition can wait up to this long, and a hung database slows a job by
+// that much per call, out of the job's own time budget (#51). A dropped final save is not
+// retried; the row keeps its last saved step.
 func WithSaveTimeout(d time.Duration) Option { return func(t *Tracker) { t.saveTimeout = d } }
 
 // Tracker implements Reporter for one job.

internal/progress/tracker_stall_test.go

@@ -96,3 +96,59 @@ func TestTracker_TheStillRunningLineIsNotHeldUpBehindAHungSave(t *testing.T) {
 
 	require.Eventually(t, func() bool { return strings.Contains(buf.String(), "still running") }, 3*time.Second, 5*time.Millisecond)
 }
+
+// deadlineSink records how long each write was given, and answers at once.
+type deadlineSink struct {
+	mu     sync.Mutex
+	saves  []time.Duration
+	beats  []time.Duration
+	noDLOK bool
+}
+
+func (d *deadlineSink) record(ctx context.Context, into *[]time.Duration) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	dl, ok := ctx.Deadline()
+	if !ok {
+		d.noDLOK = true
+		return
+	}
+	*into = append(*into, time.Until(dl))
+}
+
+func (d *deadlineSink) SaveProgress(ctx context.Context, _ []domain.ProgressStep, _ time.Time) error {
+	d.record(ctx, &d.saves)
+	return nil
+}
+
+func (d *deadlineSink) Beat(ctx context.Context, _ time.Time) error {
+	d.record(ctx, &d.beats)
+	return nil
+}
+
+// The production value is the one the worker gets: it builds a Tracker without options.
+// A zero or tiny default would drop every write, a huge one would bound nothing.
+func TestTracker_TheDefaultSaveAndBeatDeadlineIsAFewSecondsAway(t *testing.T) {
+	sink := &deadlineSink{}
+	tr := progress.NewTracker(context.Background(), sink, slog.New(slog.NewTextHandler(&lockedBuf{}, nil)),
+		[]progress.Planned{{Key: "a", Label: "A"}})
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	go tr.Heartbeat(ctx, 5*time.Millisecond, time.Hour)
+
+	tr.Begin("a", 0)
+	require.Eventually(t, func() bool {
+		sink.mu.Lock()
+		defer sink.mu.Unlock()
+		return len(sink.beats) > 0
+	}, 3*time.Second, 5*time.Millisecond)
+
+	sink.mu.Lock()
+	defer sink.mu.Unlock()
+	require.False(t, sink.noDLOK, "a write was given no deadline at all")
+	require.NotEmpty(t, sink.saves)
+	for _, d := range append(append([]time.Duration{}, sink.saves...), sink.beats...) {
+		require.Greater(t, d, time.Second, "the deadline is already (nearly) spent when the write starts")
+		require.LessOrEqual(t, d, 5*time.Second, "the default bound is 5 s")
+	}
+}

internal/uploads/jobsink_test.go

@@ -0,0 +1,48 @@
+package uploads
+
+// Observes what jobSink hands the store: the heartbeat's timestamp is the one the tracker
+// stamped (its application clock), not a second reading taken here (#51).
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"git.d-ma.be/mathias/tapir/internal/domain"
+)
+
+type atStore struct{ beatAt, saveAt time.Time }
+
+func (*atStore) StartTranscribing(context.Context, string, string) error { return nil }
+func (*atStore) CompleteUpload(context.Context, string, string, domain.Transcription) error {
+	return nil
+}
+func (*atStore) FailUpload(context.Context, string, string, domain.UploadErrorKind) error { return nil }
+func (s *atStore) SaveProgress(_ context.Context, _, _ string, _ []domain.ProgressStep, at time.Time) error {
+	s.saveAt = at
+	return nil
+}
+func (s *atStore) Heartbeat(_ context.Context, _, _ string, at time.Time) error {
+	s.beatAt = at
+	return nil
+}
+
+func TestJobSink_ForwardsTheTrackersTimestampToTheStoreUnchanged(t *testing.T) {
+	st := &atStore{}
+	sink := jobSink{st: st, userID: "u", id: "i"}
+	stamp := time.Date(2020, 1, 2, 3, 4, 5, 0, time.UTC)
+
+	if err := sink.Beat(context.Background(), stamp); err != nil {
+		t.Fatal(err)
+	}
+	if err := sink.SaveProgress(context.Background(), nil, stamp); err != nil {
+		t.Fatal(err)
+	}
+
+	if !st.beatAt.Equal(stamp) {
+		t.Errorf("heartbeat reached the store stamped %v, want the tracker's %v", st.beatAt, stamp)
+	}
+	if !st.saveAt.Equal(stamp) {
+		t.Errorf("progress save reached the store stamped %v, want the tracker's %v", st.saveAt, stamp)
+	}
+}

internal/web/crossorigin_test.go

@@ -113,7 +113,6 @@ func TestCrossOrigin_AForeignPostIsRefusedOnEachListedStateChangingPath(t *testi
 		"/v/abc/summarize",
 		"/v/abc/retry-now",
 		"/v/abc/chat",
-		"/uploads/00000000-0000-0000-0000-000000000000/delete",
 	}
 	for _, path := range routes {
 		t.Run(path, func(t *testing.T) {
@@ -126,6 +125,20 @@ func TestCrossOrigin_AForeignPostIsRefusedOnEachListedStateChangingPath(t *testi
 	}
 }
 
+func TestCrossOrigin_ARefusedRecordingDeletePostLeavesTheRecordingInPlace(t *testing.T) {
+	for name, h := range foreign {
+		t.Run(name, func(t *testing.T) {
+			app, _ := newUploadApp(t, audioOK, 200<<20)
+			id := seedUpload(t, userID, "m", "done", "Synthetic transcript text.")
+
+			rec := do(t, app, withHeaders(httptest.NewRequest(http.MethodPost, "/uploads/"+id+"/delete", nil), h))
+
+			assert.Equal(t, http.StatusForbidden, rec.Code)
+			assert.True(t, uploadExists(t, id), "a refused post must not delete the recording")
+		})
+	}
+}
+
 // A login redirect comes back to /auth/callback as a cross-site GET navigation, and
 // health probes carry no headers: safe methods must never be refused on origin grounds.
 func TestCrossOrigin_ACrossSiteGETIsNotRefused(t *testing.T) {

internal/web/oidc/discovery_retry_test.go

@@ -0,0 +1,61 @@
+package oidc_test
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/web/oidc"
+)
+
+func testConfig(issuer string) oidc.Config {
+	return oidc.Config{
+		Issuer:        issuer,
+		ClientID:      testClientID,
+		ClientSecret:  "test-client-secret",
+		RedirectURL:   "http://tapir.test/auth/callback",
+		SessionSecret: "test-session-secret-please-change",
+	}
+}
+
+// #59: a discovery 503 at start-up is retried instead of failing New.
+func TestNewRetriesDiscoveryAfter503UntilIssuerAnswers(t *testing.T) {
+	f := newFakeIssuer(t)
+	f.unavailableFor = 2
+
+	auth, err := oidc.New(context.Background(), testConfig(f.server.URL),
+		oidc.WithDiscoveryRetry(5, time.Millisecond))
+
+	require.NoError(t, err)
+	require.NotNil(t, auth)
+	require.Equal(t, int32(3), f.discoveryHits.Load())
+}
+
+func TestNewReturnsDiscoveryErrorAfterAttemptsExhausted(t *testing.T) {
+	f := newFakeIssuer(t)
+	f.unavailableFor = 1000
+
+	_, err := oidc.New(context.Background(), testConfig(f.server.URL),
+		oidc.WithDiscoveryRetry(3, time.Millisecond))
+
+	require.ErrorContains(t, err, "503")
+	require.Equal(t, int32(3), f.discoveryHits.Load())
+}
+
+func TestNewStopsDiscoveryBackoffWhenContextCancelled(t *testing.T) {
+	f := newFakeIssuer(t)
+	f.unavailableFor = 1000
+	ctx, cancel := context.WithTimeout(context.Background(), 50*time.Millisecond)
+	defer cancel()
+
+	start := time.Now()
+	_, err := oidc.New(ctx, testConfig(f.server.URL),
+		oidc.WithDiscoveryRetry(5, time.Hour))
+
+	require.True(t, errors.Is(err, context.DeadlineExceeded), "err = %v", err)
+	require.Less(t, time.Since(start), 5*time.Second)
+	require.Equal(t, int32(1), f.discoveryHits.Load())
+}

internal/web/oidc/oidc.go

@@ -57,6 +57,10 @@ const (
 	pendingTTL        = 10 * time.Minute
 	sessionCookie     = "tapir_session"
 	loginPath         = "/auth/login"
+
+	// Discovery retry (#59): 5 tries, 2+4+8+16 s back-off ≈ 30 s, then fail.
+	defaultDiscoveryAttempts = 5
+	defaultDiscoveryDelay    = 2 * time.Second
 )
 
 // DexAuth implements web.Auth against Dex.
@@ -70,6 +74,9 @@ type DexAuth struct {
 	sessionTTL time.Duration
 	now        func() time.Time
 	insecure   bool // tests only: drop the Secure attribute so cookies flow over http
+
+	discoveryAttempts int
+	discoveryDelay    time.Duration
 }
 
 var _ web.Auth = (*DexAuth)(nil)
@@ -96,14 +103,55 @@ func WithSessionTTL(ttl time.Duration) Option {
 	}
 }
 
+// WithDiscoveryRetry overrides the discovery attempt count and the initial
+// back-off delay (doubled after each failed attempt).
+func WithDiscoveryRetry(attempts int, delay time.Duration) Option {
+	return func(d *DexAuth) {
+		if attempts > 0 {
+			d.discoveryAttempts = attempts
+		}
+		if delay > 0 {
+			d.discoveryDelay = delay
+		}
+	}
+}
+
 // WithInsecureCookies drops the Secure cookie attribute. TEST ONLY — it lets
 // the session cookie travel over plain http; never use it in production.
 func WithInsecureCookies() Option {
 	return func(d *DexAuth) { d.insecure = true }
 }
 
-// New discovers the issuer and returns a configured DexAuth. ctx bounds the
-// discovery request only.
+// discover runs OIDC discovery up to attempts times, doubling delay between
+// tries (#59: an issuer briefly unavailable behind its proxy answered 503 and
+// serve exited). A cancelled ctx ends the back-off immediately with ctx's error.
+func discover(ctx context.Context, issuer string, attempts int, delay time.Duration) (*oidc.Provider, error) {
+	var lastErr error
+	for i := range attempts {
+		if i > 0 {
+			t := time.NewTimer(delay)
+			select {
+			case <-ctx.Done():
+				t.Stop()
+				return nil, fmt.Errorf("after %d attempts (last: %v): %w", i, lastErr, ctx.Err())
+			case <-t.C:
+			}
+			delay *= 2
+		}
+		p, err := oidc.NewProvider(ctx, issuer)
+		if err == nil {
+			return p, nil
+		}
+		lastErr = err
+		if ctx.Err() != nil {
+			return nil, fmt.Errorf("after %d attempts: %w", i+1, ctx.Err())
+		}
+	}
+	return nil, fmt.Errorf("after %d attempts: %w", attempts, lastErr)
+}
+
+// New discovers the issuer and returns a configured DexAuth. ctx bounds discovery,
+// including its retry back-off.
 func New(ctx context.Context, cfg Config, opts ...Option) (*DexAuth, error) {
 	for name, val := range map[string]string{
 		"issuer":         cfg.Issuer,
@@ -117,29 +165,31 @@ func New(ctx context.Context, cfg Config, opts ...Option) (*DexAuth, error) {
 		}
 	}
 
-	provider, err := oidc.NewProvider(ctx, cfg.Issuer)
-	if err != nil {
-		return nil, fmt.Errorf("oidc: discover issuer: %w", err)
-	}
-
 	d := &DexAuth{
-		cfg:      cfg,
-		verifier: provider.Verifier(&oidc.Config{ClientID: cfg.ClientID}),
-		oauth: &oauth2.Config{
-			ClientID:     cfg.ClientID,
-			ClientSecret: cfg.ClientSecret,
-			Endpoint:     provider.Endpoint(),
-			RedirectURL:  cfg.RedirectURL,
-			Scopes:       []string{oidc.ScopeOpenID, "profile", "email"},
-		},
-		pending:    newPendingStore(),
-		secret:     []byte(cfg.SessionSecret),
-		sessionTTL: defaultSessionTTL,
-		now:        time.Now,
+		cfg:               cfg,
+		pending:           newPendingStore(),
+		secret:            []byte(cfg.SessionSecret),
+		sessionTTL:        defaultSessionTTL,
+		now:               time.Now,
+		discoveryAttempts: defaultDiscoveryAttempts,
+		discoveryDelay:    defaultDiscoveryDelay,
 	}
 	for _, o := range opts {
 		o(d)
 	}
+
+	provider, err := discover(ctx, cfg.Issuer, d.discoveryAttempts, d.discoveryDelay)
+	if err != nil {
+		return nil, fmt.Errorf("oidc: discover issuer: %w", err)
+	}
+	d.verifier = provider.Verifier(&oidc.Config{ClientID: cfg.ClientID})
+	d.oauth = &oauth2.Config{
+		ClientID:     cfg.ClientID,
+		ClientSecret: cfg.ClientSecret,
+		Endpoint:     provider.Endpoint(),
+		RedirectURL:  cfg.RedirectURL,
+		Scopes:       []string{oidc.ScopeOpenID, "profile", "email"},
+	}
 	return d, nil
 }
 

internal/web/oidc/oidc_test.go

@@ -8,6 +8,7 @@ import (
 	"net/http"
 	"net/http/httptest"
 	"net/url"
+	"sync/atomic"
 	"testing"
 	"time"
 
@@ -34,6 +35,11 @@ type fakeIssuer struct {
 	sub   string
 	email string
 	nonce string
+
+	// discoveryHits counts requests to the discovery document; the first
+	// unavailableFor of them answer 503 (issuer briefly down behind its proxy).
+	discoveryHits  atomic.Int32
+	unavailableFor int32
 }
 
 const testKID = "test-key"
@@ -47,6 +53,10 @@ func newFakeIssuer(t *testing.T) *fakeIssuer {
 
 	mux := http.NewServeMux()
 	mux.HandleFunc("/.well-known/openid-configuration", func(w http.ResponseWriter, _ *http.Request) {
+		if f.discoveryHits.Add(1) <= f.unavailableFor {
+			http.Error(w, "no available server", http.StatusServiceUnavailable)
+			return
+		}
 		writeJSON(t, w, map[string]any{
 			"issuer":                                f.server.URL,
 			"authorization_endpoint":                f.server.URL + "/authorize",

internal/web/uploads.go

@@ -273,12 +273,31 @@ func titleFromFilename(name string) string {
 // cloudProvider names the cloud service that produced a transcript, or false when it was
 // served locally (or by an unknown endpoint). "Local" is config.IsLocalModel: the same
 // definition the start-up validation uses, not a second string match here (#52).
+//
+// The endpoint is one alias, except for the two-engine path (ADR-034), which writes
+// "<base> + <alt>, reconciled by <model>" (ensemble.go). That shape is taken apart exactly
+// and each alias is judged whole, so a composite of local models is local, and anything
+// else (a cloud alias, one with no provider prefix, one with a space or a connector word in
+// it) is not. A string that is not that shape is one alias.
 func cloudProvider(endpoint string) (string, bool) {
-	if endpoint == "" || config.IsLocalModel(endpoint) {
+	endpoint = strings.TrimSpace(endpoint)
+	if endpoint == "" {
 		return "", false
 	}
-	provider, _, _ := strings.Cut(endpoint, "/")
-	return provider, true
+	aliases := []string{endpoint}
+	if engines, reconciler, ok := strings.Cut(endpoint, ", reconciled by "); ok {
+		aliases = append(strings.Split(engines, " + "), reconciler)
+	}
+	for _, a := range aliases {
+		if a = strings.TrimSpace(a); !config.IsLocalModel(a) {
+			provider, _, _ := strings.Cut(a, "/")
+			if provider = strings.TrimSpace(provider); provider == "" {
+				provider = a
+			}
+			return provider, true
+		}
+	}
+	return "", false
 }
 
 // hasNote reports whether any phase carries a note. A note is the only warning for a

internal/web/uploads_cloud_note_test.go

@@ -53,3 +53,88 @@ func TestUploadStatus_AnEndpointThatOnlyLooksLocalIsStillNamedAsCloud(t *testing
 
 	assert.Contains(t, b, "left the homelab")
 }
+
+// The two-engine path (ADR-034) writes one composite endpoint string. Every name in it is
+// a local model, so the audio never left the homelab and the note must not say it did.
+func TestUploadStatus_AReconciledAllLocalEndpointCarriesNoCloudNote(t *testing.T) {
+	for _, endpoint := range []string{
+		"iguana/kb-whisper + koala/whisper, reconciled by iguana/gemma4-26b",
+		"iguana/kb-whisper + koala/whisper, reconciled by iguana/gemma4-26b\n",
+	} {
+		t.Run(endpoint, func(t *testing.T) {
+			app, _ := newUploadApp(t, audioOK, 200<<20)
+			id := seedUpload(t, userID, "m", "done", "Synthetic transcript text.")
+			setEndpoint(t, id, endpoint)
+
+			b := body(t, do(t, app, get("/uploads/"+id+"/status")))
+
+			assert.NotContains(t, b, "left the homelab")
+		})
+	}
+}
+
+func TestUploadStatus_AReconciledEndpointNamingACloudEngineStillNamesTheCloudProvider(t *testing.T) {
+	app, _ := newUploadApp(t, audioOK, 200<<20)
+	id := seedUpload(t, userID, "m", "done", "Synthetic transcript text.")
+	setEndpoint(t, id, "iguana/kb-whisper + berget/kb-whisper, reconciled by iguana/gemma4-26b")
+
+	b := body(t, do(t, app, get("/uploads/"+id+"/status")))
+
+	assert.Contains(t, b, "cloud service (berget)")
+}
+
+// An alias without a provider prefix is not a local model, wherever it sits in the
+// composite: it must never be dropped from the judgement (PR #58 re-review).
+func TestUploadStatus_AReconciledEndpointWithAnUnprefixedAliasIsNotCalledLocal(t *testing.T) {
+	for _, endpoint := range []string{
+		"iguana/kb-whisper + whisper-1, reconciled by iguana/gemma4-26b",
+		"iguana/kb-whisper + whisper 1, reconciled by iguana/gemma4-26b",
+		"iguana/kb-whisper + berget / x, reconciled by iguana/gemma4-26b",
+	} {
+		t.Run(endpoint, func(t *testing.T) {
+			app, _ := newUploadApp(t, audioOK, 200<<20)
+			id := seedUpload(t, userID, "m", "done", "Synthetic transcript text.")
+			setEndpoint(t, id, endpoint)
+
+			b := body(t, do(t, app, get("/uploads/"+id+"/status")))
+
+			assert.Contains(t, b, "left the homelab")
+			assert.NotContains(t, b, "cloud service ()", "the note names a provider")
+		})
+	}
+}
+
+// "by", "reconciled" and "+" are connectors only inside the composite format. As a whole
+// endpoint, or as an alias within the composite, they are model names and are judged.
+func TestUploadStatus_AConnectorWordUsedAsAModelNameIsStillJudged(t *testing.T) {
+	for _, endpoint := range []string{
+		"by",
+		"reconciled",
+		"+",
+		"iguana/kb-whisper + by, reconciled by iguana/gemma4-26b",
+		"iguana/kb-whisper + reconciled, reconciled by iguana/gemma4-26b",
+		"iguana/kb-whisper + +, reconciled by iguana/gemma4-26b",
+		"/x", // no provider at all: the note must still name something
+	} {
+		t.Run(endpoint, func(t *testing.T) {
+			app, _ := newUploadApp(t, audioOK, 200<<20)
+			id := seedUpload(t, userID, "m", "done", "Synthetic transcript text.")
+			setEndpoint(t, id, endpoint)
+
+			b := body(t, do(t, app, get("/uploads/"+id+"/status")))
+
+			assert.Contains(t, b, "left the homelab")
+			assert.NotContains(t, b, "cloud service ()")
+		})
+	}
+}
+
+func TestUploadStatus_AReconcilerThatIsNotLocalNamesTheCloudProvider(t *testing.T) {
+	app, _ := newUploadApp(t, audioOK, 200<<20)
+	id := seedUpload(t, userID, "m", "done", "Synthetic transcript text.")
+	setEndpoint(t, id, "iguana/kb-whisper + koala/whisper, reconciled by berget/mistral-large")
+
+	b := body(t, do(t, app, get("/uploads/"+id+"/status")))
+
+	assert.Contains(t, b, "cloud service (berget)")
+}