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>
fix(oidc): retry issuer discovery with bounded back-off at start-up
2026-10-04 22:52:55 +02:00 · 2026-10-04 22:52:26 +02:00 · 2026-10-04 08:47:10 +02:00 · 2026-10-04 00:41:07 +02:00 · 2026-10-04 00:34:59 +02:00 · 2026-10-03 22:31:26 +00:00
47 changed files with 3019 additions and 788 deletions
@@ -1447,6 +1447,25 @@ an hour of Swedish *was* an unknown here; it is now measured, see decision 8.)
   re-uploads. The single-replica assumption gains a second dependent (the semaphore is
   process-local) and is recorded here.

+   *Amended 2026-10-03 (#50, #47).* The process now handles SIGTERM (what Kubernetes sends on
+   every deploy; only SIGINT was handled, so shutdown code never ran). On shutdown the running job
+   is recorded `interrupted` at once and the process waits, bounded to 10 s each, for in-flight
+   requests and for that write; the start-up sweep stays the backstop for a kill. **Known
+   limitation, deliberately not built: a deploy still loses a job in flight** (finishing it needs
+   a drain of up to the job time, a ~330 s termination grace and `strategy: Recreate`, i.e.
+   downtime on every deploy; a 90-minute file takes ~4.5 min and a deploy lands on a running job
+   perhaps one time in five at current volume). Jobs still queued at shutdown stay `queued` until
+   the next start's sweep. The user re-uploads, with an honest message.
+
+   *Receive limits (#47).* Request bodies are received at most 4 at a time, 2 per user, answered
+   as busy beyond that. Together with the 20-job queue and the running job that bounds the spool
+   at about (4 + 20 + 1) x 200 MiB = 5 GiB of the 6 Gi `emptyDir` (infra): raising any of those or
+   `TAPIR_UPLOAD_MAX_BYTES` must revisit that sum (pinned by `TestNewIntake_Defaults...`, not
+   enforced against infra). **A slow-drip body holds its slot without a time bound**: the server
+   sets `ReadHeaderTimeout` only, because a `ReadTimeout` would cut legitimate 200 MiB uploads on
+   slow links; two slow bodies take a user's two slots and two such users make everyone's upload
+   busy. It needs a valid session; recorded as a limitation with a follow-up issue.
+
 8. **Chat reuses `chat.Service`; it is not forked.** The uploaded-transcript chat handlers
   look the row up through the RLS-scoped store, so a non-owner gets 404 by the same mechanism
   as ADR-027's summary gate, **before any model is called**. ADR-027's decisions (ephemeral,
@@ -1531,9 +1550,15 @@ deferred. The scope cut — chat first, no rich analysis, no video — is what k
  #33 specifies, with the serving endpoint shown **(recommended, matches the issue)**; (B)
  disabled by default, a failed iguana job fails honestly and the user retries. *Blocks:*
  the production default only; the adapter ships either way.
- Verify Tapir's ingress `proxy-body-size` against the 200 MB cap (`infra`).
- Per-transcript deletion is not built; account deletion cascades. A private-content store
-  without a delete button is a gap worth a ticket, not a silent omission.
+- ~~Verify Tapir's ingress `proxy-body-size` against the 200 MB cap (`infra`).~~ **Resolved 2026-10-03:** real unauthenticated bodies of 2, 30, 120 and 210 MiB all pass NPM (TLS, piguard) and Traefik to tapir (a 302 to login); there is no body limit in that range. Uploads were enabled in production the same day, local engines only (infra#520).
+- ~~Per-transcript deletion is not built; account deletion cascades.~~ **Built 2026-10-03
+  (#37):** `POST /uploads/{id}/delete`, owner-scoped (a non-owner gets 404 by the same RLS
+  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 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`
@@ -34,7 +34,7 @@ tasks:
  lint:
    desc: golangci-lint (installed by CI; skipped locally if absent).
    cmds:
-      - 'command -v golangci-lint >/dev/null 2>&1 && golangci-lint run ./... || echo "golangci-lint not installed — skipping (CI installs it)"'
+      - 'if command -v golangci-lint >/dev/null 2>&1; then golangci-lint run ./...; else echo "golangci-lint not installed — skipping (CI installs it)"; fi'

  test:
    desc: Run all tests. -p 1 serializes packages so the embedded-postgres tests
@@ -17,9 +17,9 @@ import (
 	"errors"
 	"fmt"
 	"log/slog"
+	"net"
 	"net/http"
 	"os"
-	"os/signal"
 	"sync"
 	"time"

@@ -41,7 +41,7 @@ func main() {
 	}

 	log := slog.New(slog.NewTextHandler(os.Stderr, nil))
-	ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt)
+	ctx, stop := shutdownContext(context.Background())
 	defer stop()

 	var err error
@@ -215,9 +215,11 @@ func cmdServe(ctx context.Context, log *slog.Logger) error {
 	if err != nil {
 		return err
 	}
+	var workerDone chan struct{} // nil without uploads: nothing to wait for at shutdown
 	wireUploads(app, cfg, intake)
 	if intake != nil {
-		go worker.Run(ctx)
+		workerDone = make(chan struct{})
+		go func() { worker.Run(ctx); close(workerDone) }()
 		log.Info("audio upload enabled", "model", cfg.TranscribeModel, "fallback", cfg.TranscribeFallbackModel,
 			"max_mb", cfg.UploadMaxBytes>>20, "max_minutes", int(cfg.UploadMaxDuration.Minutes()))
 		log.Warn("audio upload assumes a SINGLE replica: the transcription queue is process-local (ADR-033, ADR-018)")
@@ -350,20 +352,19 @@ func cmdServe(ctx context.Context, log *slog.Logger) error {
 		}()
 	}

-	// Graceful shutdown on signal: stop accepting, drain in-flight requests.
-	go func() {
-		<-ctx.Done()
-		shutdownCtx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
-		defer cancel()
-		_ = srv.Shutdown(shutdownCtx)
-		if metricsSrv != nil {
-			_ = metricsSrv.Shutdown(shutdownCtx)
-		}
-	}()
-
+	// Graceful shutdown on signal: stop accepting, drain in-flight requests, and let the
+	// worker record the job it was running as interrupted (#50).
+	var extra []*http.Server
+	if metricsSrv != nil {
+		extra = append(extra, metricsSrv)
+	}
+	ln, err := net.Listen("tcp", srv.Addr)
+	if err != nil {
+		return fmt.Errorf("serve: listen: %w", err)
+	}
 	log.Info("serving web ui", "addr", cfg.HTTPAddr, "user", cfg.UserID)
-	if err := srv.ListenAndServe(); err != nil && !errors.Is(err, http.ErrServerClosed) {
-		return fmt.Errorf("serve: %w", err)
+	if err := serveHTTP(ctx, srv, ln, extra, workerDone, shutdownGrace); err != nil {
+		return err
 	}
 	return nil
 }
@@ -0,0 +1,55 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"net"
+	"net/http"
+	"os"
+	"os/signal"
+	"syscall"
+	"time"
+)
+
+// shutdownGrace bounds each wait at shutdown (in-flight requests, then the worker). Two of
+// them stay inside Kubernetes' default 30 s termination grace.
+const shutdownGrace = 10 * time.Second
+
+// shutdownContext is cancelled by SIGINT (a terminal) or SIGTERM (what Kubernetes sends
+// on every deploy). Handling only SIGINT meant a deploy killed the process before any
+// shutdown code ran (#50).
+func shutdownContext(parent context.Context) (context.Context, context.CancelFunc) {
+	return signal.NotifyContext(parent, os.Interrupt, syscall.SIGTERM)
+}
+
+// serveHTTP serves srv on ln until it fails or ctx ends, then shuts every server down and
+// waits for in-flight requests and for workerDone (nil: no worker) to finish, each for at
+// most grace. It must not return earlier: Serve returns http.ErrServerClosed as soon as
+// Shutdown STARTS, and returning from main kills the drain and the worker's last write.
+func serveHTTP(ctx context.Context, srv *http.Server, ln net.Listener, extra []*http.Server, workerDone <-chan struct{}, grace time.Duration) error {
+	errc := make(chan error, 1)
+	go func() { errc <- srv.Serve(ln) }()
+
+	select {
+	case err := <-errc:
+		if !errors.Is(err, http.ErrServerClosed) {
+			return fmt.Errorf("serve: %w", err)
+		}
+		return nil
+	case <-ctx.Done():
+	}
+
+	shutdownCtx, cancel := context.WithTimeout(context.Background(), grace)
+	defer cancel()
+	for _, s := range append([]*http.Server{srv}, extra...) {
+		_ = s.Shutdown(shutdownCtx)
+	}
+	if workerDone != nil {
+		select {
+		case <-workerDone:
+		case <-time.After(grace):
+		}
+	}
+	return nil
+}
@@ -0,0 +1,124 @@
+package main
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"sync"
+	"syscall"
+	"testing"
+	"time"
+)
+
+// Kubernetes stops a pod with SIGTERM; only SIGINT was handled, so a deploy killed the
+// process without running any shutdown code (#50).
+func TestShutdownContext_IsCancelledBySIGTERMAndSIGINT(t *testing.T) {
+	for _, sig := range []syscall.Signal{syscall.SIGTERM, syscall.SIGINT} {
+		t.Run(sig.String(), func(t *testing.T) {
+			ctx, stop := shutdownContext(context.Background())
+			defer stop()
+
+			if err := syscall.Kill(os.Getpid(), sig); err != nil {
+				t.Fatalf("send %v: %v", sig, err)
+			}
+
+			select {
+			case <-ctx.Done():
+			case <-time.After(2 * time.Second):
+				t.Fatalf("context not cancelled by %v", sig)
+			}
+		})
+	}
+}
+
+// serveHTTP must not return while a request is in flight or the worker is still recording
+// its job: Serve returns http.ErrServerClosed as soon as Shutdown STARTS, and returning
+// from it let the process exit before either finished (#50). Each half is tested with the
+// other already satisfied, so neither can hold the call open on the other's behalf.
+
+func pendingFor(t *testing.T, returned <-chan error, why string) {
+	t.Helper()
+	select {
+	case err := <-returned:
+		t.Fatalf("serveHTTP returned (%v) %s", err, why)
+	case <-time.After(150 * time.Millisecond):
+	}
+}
+
+func TestServeHTTP_DrainsAnInFlightRequestBeforeReturning(t *testing.T) {
+	release := make(chan struct{})
+	entered := make(chan struct{})
+	srv := httptest.NewUnstartedServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		close(entered)
+		<-release
+		_, _ = w.Write([]byte("ok"))
+	}))
+	srv.Start()
+	defer srv.Close()
+	var once sync.Once
+	releaseHandler := func() { once.Do(func() { close(release) }) }
+	defer releaseHandler() // runs before srv.Close: a failing test must not hang on the blocked handler
+	ctx, cancel := context.WithCancel(context.Background())
+	workerDone := make(chan struct{})
+	close(workerDone) // the worker is already done: only the request can hold serveHTTP open
+	returned := make(chan error, 1)
+	go func() { returned <- serveHTTP(ctx, srv.Config, srv.Listener, nil, workerDone, 5*time.Second) }()
+
+	go func() { _, _ = http.Get(srv.URL) }()
+	<-entered
+	cancel()
+
+	pendingFor(t, returned, "while a request was still in flight")
+	releaseHandler()
+	select {
+	case err := <-returned:
+		if err != nil {
+			t.Fatalf("serveHTTP: %v", err)
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatal("serveHTTP did not return once the request finished")
+	}
+}
+
+func TestServeHTTP_WaitsForTheWorkerBeforeReturning(t *testing.T) {
+	srv := httptest.NewUnstartedServer(http.NotFoundHandler())
+	srv.Start()
+	defer srv.Close()
+	ctx, cancel := context.WithCancel(context.Background())
+	workerDone := make(chan struct{})
+	returned := make(chan error, 1)
+	go func() { returned <- serveHTTP(ctx, srv.Config, srv.Listener, nil, workerDone, 5*time.Second) }() // no request in flight
+
+	cancel()
+
+	pendingFor(t, returned, "before the worker finished")
+	close(workerDone)
+	select {
+	case err := <-returned:
+		if err != nil {
+			t.Fatalf("serveHTTP: %v", err)
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatal("serveHTTP did not return once the worker finished")
+	}
+}
+
+func TestServeHTTP_GivesUpOnAWorkerThatNeverFinishesAfterTheGrace(t *testing.T) {
+	srv := httptest.NewUnstartedServer(http.NotFoundHandler())
+	srv.Start()
+	defer srv.Close()
+	ctx, cancel := context.WithCancel(context.Background())
+	returned := make(chan error, 1)
+	go func() {
+		returned <- serveHTTP(ctx, srv.Config, srv.Listener, nil, make(chan struct{}), 200*time.Millisecond)
+	}()
+
+	cancel()
+
+	select {
+	case <-returned:
+	case <-time.After(3 * time.Second):
+		t.Fatal("a stuck worker must not hold the process past the grace period")
+	}
+}
@@ -44,7 +44,7 @@ func (f *fakeUploadsStore) SaveProgress(context.Context, string, string, []domai
 	return nil
 }

-func (f *fakeUploadsStore) Heartbeat(context.Context, string, string) error { return nil }
+func (f *fakeUploadsStore) Heartbeat(context.Context, string, string, time.Time) error { return nil }

 func (f *fakeUploadsStore) FailInterruptedUploads(context.Context) (int, error) {
 	f.sweeps++
@@ -14,7 +14,8 @@ it** — endpoints and aliases drift, and this file is a snapshot (2026-06-06),
    not LiteLLM. Use `30401` for off-cluster LiteLLM. `confirm` resolved.
 - **Speech-to-text (ADR-033):** the same gateway serves `POST /v1/audio/transcriptions` for
  `iguana/kb-whisper` (primary, local) and `berget/kb-whisper` (fallback, EU cloud — private audio
-  leaves the homelab if it is used). **Call it via the NodePort / in-cluster address above, NOT
+  leaves the homelab if it is used: when the primary fails **or returns an empty transcript**,
+  e.g. a silent or music-only file; the transcript page then says so, #52). **Call it via the NodePort / in-cluster address above, NOT
  `https://llm-api.d-ma.be`:** that public ingress (openresty) returns `413` for bodies over
  roughly 8–10 MB (~5 minutes of 16 kHz WAV), while the NodePort took 174 MB (90 minutes) in one
  request (measured 2026-10-01, `docs/research/audio-length-ceiling-2026-10-01.md`). Send multipart
@@ -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.
@@ -0,0 +1,115 @@
+# 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: **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).
+
+Context: v0.31.0 shipped audio upload (ADR-033/034/035) and is live in production behind auth at
+`/uploads`, with the Berget cloud fallback on for audio (infra#522). A real 63.7 min upload
+transcribed in 2m53s with no errors. The phase-1 investigation found two pre-existing or new
+weaknesses an authenticated user can reach, so the first release is a hotfix.
+
+## Oath (closing gates, every clause must hold; blocked = descoped and tracked, never "done")
+
+| # | Clause | Evidence |
+|---|--------|----------|
+| SG1 | Every issue in the sprint is closed with a linked commit, or moved out with a stated reason | issue list of both milestones |
+| SG2 | Each fix has a test that was seen red on main first, named after the mechanism it observes | red/green output in the issue |
+| SG3 | `task check` AND `golangci-lint run ./...` exit 0 (checked by `$?`, not by grep); `task check` does not gate lint until #41 | gate output |
+| SG4 | Control code (#48, #49, the upload cap in #47) has a recorded adversarial review verified by execution, verdict in the issue | gatekeeper verdict comment |
+| SG5 | Docs current in the same commit: ADR-033/035 text, feature file, use-case coverage test | diff |
+| SG6 | Each release tagged (v0.31.1, v0.32.0), pushed, deployed, and observed live (pod running the new image, no restart loop, a real upload still works) | `kubectl`/log + one upload |
+| SG7 | Learnings written to the brain; deferred work filed as issues; decisions that are the owner's recorded as DECISION NEEDED | brain ids + issue numbers |
+
+## Work order
+
+### v0.31.1 (security and robustness; hotfix, do first)
+
+1. **#41 lint gate.** `task check` swallows lint findings (`|| echo`). Make the gate fail on findings,
+   fix what it surfaces. First, so every later gate means something. (Decision D1.)
+2. **#48 cross-origin POST protection.** `http.NewCrossOriginProtection()` around `Router()`.
+   Red: foreign-origin POST to `/account/delete`, `/uploads`, `/uploads/{id}/chat` is refused.
+3. **#49 ffprobe/ffmpeg limits.** `-max_streams 32` + narrower probe output; probe stdout cap;
+   probe deadline. Three commits.
+4. **#47 concurrent-upload cap (remaining item).** Bound the spool: reservation per accepted upload
+   against the 6Gi emptyDir so concurrent uploads cannot evict the pod.
+5. **#50 SIGTERM interim.** Handle SIGTERM; record the interrupted job at once; fix the false
+   comment. Full drain recorded as a limitation. (Decision D2.)
+
+Release v0.31.1: one adversarial review round over 2-5, then tag, push, observe.
+
+### v0.32.0 (follow-ups)
+
+6. **#52** label a cloud-served transcript in plain words; document the empty-transcript trigger.
+7. **#53** bound the LLM client's response body read.
+8. **#51** tracker: save timeout, then app-clock heartbeat.
+9. **#37** nav link to uploads and per-recording delete (queued/transcribing: refuse until it
+   ends). (Decision D3.)
+10. **#36** chat default model after measurement. (Decision D4.)
+11. **#38** measurements (Berget length limit, large models on production iguana). (D5, D6.)
+12. **#46** remainder of the hardening list after the items above are split out.
+13. **#45** evaluate Berget as the reconciler: a question, not work. (D8.)
+
+## DECISION NEEDED (owner's call; recommendation first)
+
+| # | Question | Recommendation | Blocks |
+|---|----------|----------------|--------|
+| D1 | #41: should any linter be allowed to opt out of the gate? | No opt-out; fix findings | #41 |
+| D2 | #50: full SIGTERM drain, or record as a limitation and ship the interim fix? | Limitation + interim; a drain costs 4-6 h and downtime on every deploy at current volume | #50 |
+| D3 | #37: delete of a queued/transcribing recording | Refuse until it ends | #37 |
+| D4 | #36: make a 31B model the chat default? | Only after measurement, conditional | #36 |
+| 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 | **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
+
+- **#46 item 1 (`alis`/`dref` external reference):** checked, not a problem. Under the application's
+  flags ffmpeg skips the external track and writes a 78-byte WAV.
+- **#46 item 2 / #47 item 2:** merged into #49.
+- **#47 item 3 (full drain):** reduced to #50 plus a recorded limitation.
+- **#46 `Beat` taking the tracker mutex:** incorrect in #46; only `logStill` does. Corrected in #51.
+
+## Human-present items (do not run unattended)
+
+- Production iguana model loads (#38, D6).
+- Enabling reconcile in production (D10).
+- Any key change or `apply` in infra (infra#523, infra#524 are open and not part of this sprint).
+
+## Release plan
+
+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)).
@@ -240,6 +240,24 @@ Feature: Upload an audio recording, get a private Swedish transcript, discuss it
    Then it is transcribed anyway and its audio step says how long the audio really is
    And a complete recording whose header over-estimated its length is not refused

+  Scenario: A transcript made by the cloud fallback says the audio left the homelab
+    When the transcript was produced by an endpoint that is not local
+    Then its page says in plain words that the audio left the homelab and names the provider
+    And a locally transcribed recording carries no such note
+
+  Scenario: I can delete a finished recording
+    When I confirm deleting a done or failed recording
+    Then the recording and its transcript are gone and I am back on my list
+    And another user's recording is a 404 and is untouched
+
+  Scenario: A recording still being transcribed cannot be deleted
+    When I try to delete a recording that is queued or transcribing
+    Then I am told it is still being transcribed and nothing is removed
+
+  Scenario: The home list links to my recordings when uploads are on
+    Given audio upload is enabled
+    Then the home list links to the recordings page
+
  Scenario: A full queue is a retry, not an error
    When too many recordings are already waiting
    Then the upload is answered as busy, its row is failed honestly and its audio is deleted
@@ -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
 )
@@ -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=
@@ -47,11 +47,15 @@ func (f *FFmpeg) ToWAV(ctx context.Context, src, dst string) error {
 	args = append(args, "-f", "wav", dst)
 	cmd := exec.CommandContext(ctx, bin, args...)
 	cmd.WaitDelay = killGrace
-	if out, err := cmd.CombinedOutput(); err != nil {
+	// Only a bounded head of ffmpeg's output is kept: a hostile file made it print ~0.4
+	// bytes per input byte, all of which CombinedOutput would buffer and copy (#49).
+	out := &headWriter{max: maxStderr}
+	cmd.Stdout, cmd.Stderr = out, out
+	if err := cmd.Run(); err != nil {
 		if ctxErr := ctx.Err(); ctxErr != nil {
 			return fmt.Errorf("audio: ffmpeg: %w", ctxErr)
 		}
-		return fmt.Errorf("audio: ffmpeg: %w: %s", err, strings.TrimSpace(string(out)))
+		return fmt.Errorf("audio: ffmpeg: %w: %s", err, strings.TrimSpace(out.buf.String()))
 	}
 	// A zero exit that wrote nothing is still a failure: never hand whisper a
 	// missing file and let the error surface three layers later.
@@ -109,3 +109,32 @@ printf data > "$last"`)
 	require.NoError(t, err)
 	assert.NotContains(t, " "+strings.ReplaceAll(string(raw), "\n", " ")+" ", " -t ")
 }
+
+func TestFFmpegToWAV_CapsTheStreamCountBeforeTheInput(t *testing.T) {
+	// ffmpeg also allocated ~570 MB opening the 900-stream file (#49).
+	rec := filepath.Join(t.TempDir(), "argv")
+	bin := stub(t, `printf '%s\n' "$@" > `+rec+`
+for a; do last="$a"; done
+printf data > "$last"`)
+
+	err := (&audio.FFmpeg{Bin: bin}).ToWAV(context.Background(), "/spool/in", filepath.Join(t.TempDir(), "o.wav"))
+
+	require.NoError(t, err)
+	args := argvOf(t, rec)
+	ms, in := indexOf(args, "-max_streams"), indexOf(args, "-i")
+	require.GreaterOrEqual(t, ms, 0, "argv: %v", args)
+	assert.Equal(t, "32", args[ms+1])
+	assert.Less(t, ms, in, "an input option placed after -i is applied to the output")
+}
+
+func TestFFmpegToWAV_ABoundedPartOfAFloodingStderrIsKept(t *testing.T) {
+	// A hostile mp3 made ffmpeg print ~0.4 bytes of stderr per input byte (83 MB for a
+	// 209 MB file), all of it buffered, then copied twice into the error (#49 review).
+	bin := stub(t, `head -c 5000000 /dev/zero | tr '\0' 'e' >&2; exit 1`)
+
+	err := (&audio.FFmpeg{Bin: bin}).ToWAV(context.Background(), "/spool/in", filepath.Join(t.TempDir(), "o.wav"))
+
+	require.Error(t, err)
+	assert.Less(t, len(err.Error()), 16*1024, "the error carries a bounded head of stderr, not all of it")
+	assert.Contains(t, err.Error(), "eeee", "and still carries some of it")
+}
@@ -4,6 +4,7 @@
 package audio

 import (
+	"bytes"
 	"context"
 	"encoding/json"
 	"errors"
@@ -19,6 +20,51 @@ import (
 type FFprobe struct {
 	// Bin is the ffprobe executable ("ffprobe" resolves via PATH).
 	Bin string
+	// MaxOutput bounds how much of ffprobe's stdout is read; zero means DefaultMaxOutput.
+	MaxOutput int
+	// Timeout bounds one probe; zero means DefaultProbeTimeout. A probe is milliseconds of
+	// work on a real file, so one that runs this long is the FILE's doing (#49).
+	Timeout time.Duration
+}
+
+// DefaultProbeTimeout is the probe deadline when FFprobe.Timeout is unset.
+const DefaultProbeTimeout = 45 * time.Second
+
+// DefaultMaxOutput is far above the narrowed probe output of a legitimate file (a few
+// hundred bytes) and far below what a hostile one can make ffprobe print.
+const DefaultMaxOutput = 1 << 20
+
+// maxStderr is how much of ffprobe's stderr is kept for the error message.
+const maxStderr = 4096
+
+// errOutputCap is the write error that stops ffprobe once it prints past the cap.
+var errOutputCap = errors.New("audio: ffprobe output exceeds the cap")
+
+// capWriter collects up to max bytes and then refuses further writes, which closes the
+// pipe and ends the process instead of buffering whatever it prints.
+type capWriter struct {
+	buf bytes.Buffer
+	max int
+}
+
+func (w *capWriter) Write(p []byte) (int, error) {
+	if w.buf.Len()+len(p) > w.max {
+		return 0, errOutputCap
+	}
+	return w.buf.Write(p)
+}
+
+// headWriter keeps the first max bytes and silently drops the rest.
+type headWriter struct {
+	buf bytes.Buffer
+	max int
+}
+
+func (w *headWriter) Write(p []byte) (int, error) {
+	if room := w.max - w.buf.Len(); room > 0 {
+		w.buf.Write(p[:min(room, len(p))])
+	}
+	return len(p), nil
 }

 // AllowedDemuxers are the input formats an upload may be. The input is
@@ -30,9 +76,22 @@ type FFprobe struct {
 // formats. "mov" covers .mp4/.m4a/.3gp; "matroska" covers .webm.
 const AllowedDemuxers = "wav,mov,mp4,m4a,3gp,mp3,ogg,matroska,webm,flac,aac,opus,aiff,amr,asf"

+// MaxStreams bounds how many streams a file may declare. A 1.3 MB file declaring 900 of
+// them cost ~570 MB of RSS in each of ffprobe and ffmpeg (#49), against a 512 Mi pod;
+// real recordings have a handful.
+const MaxStreams = "32"
+
 // inputRestrictions are ffmpeg/ffprobe INPUT options and must precede the input.
-// Local files and pipes only: no network, no other protocols.
-var inputRestrictions = []string{"-protocol_whitelist", "file,pipe", "-format_whitelist", AllowedDemuxers}
+// Local files and pipes only: no network, no other protocols, and a bounded stream count.
+var inputRestrictions = []string{
+	"-protocol_whitelist", "file,pipe",
+	"-format_whitelist", AllowedDemuxers,
+	"-max_streams", MaxStreams,
+}
+
+// probeEntries is exactly what ParseProbe reads. -show_format/-show_streams print every
+// tag too, and a 120 MB WAV made 491 MB of JSON that Output() buffers.
+const probeEntries = "format=duration:stream=codec_type:stream_disposition=attached_pic"

 // killGrace bounds how long Probe waits for a cancelled ffprobe's pipes to close.
 const killGrace = 2 * time.Second
@@ -48,21 +107,38 @@ func (f *FFprobe) Probe(ctx context.Context, path string) (uploads.Probe, error)
 		bin = "ffprobe"
 	}
 	args := append([]string{"-v", "error"}, inputRestrictions...)
-	args = append(args, "-print_format", "json", "-show_format", "-show_streams", path)
-	cmd := exec.CommandContext(ctx, bin, args...)
+	args = append(args, "-print_format", "json", "-show_entries", probeEntries, path)
+	timeout := f.Timeout
+	if timeout <= 0 {
+		timeout = DefaultProbeTimeout
+	}
+	probeCtx, cancel := context.WithTimeout(ctx, timeout)
+	defer cancel()
+	cmd := exec.CommandContext(probeCtx, bin, args...)
 	cmd.WaitDelay = killGrace
-	out, err := cmd.Output()
-	if err != nil {
-		if ctxErr := ctx.Err(); ctxErr != nil {
+	maxOut := f.MaxOutput
+	if maxOut <= 0 {
+		maxOut = DefaultMaxOutput
+	}
+	stdout, stderr := &capWriter{max: maxOut}, &headWriter{max: maxStderr}
+	cmd.Stdout, cmd.Stderr = stdout, stderr
+	if err := cmd.Run(); err != nil {
+		if ctxErr := ctx.Err(); ctxErr != nil { // the caller gave up: not a verdict on the file
 			return uploads.Probe{}, fmt.Errorf("audio: ffprobe: %w", ctxErr)
 		}
+		if probeCtx.Err() != nil { // our own deadline passed while the caller was still waiting
+			return uploads.Probe{}, fmt.Errorf("%w: ffprobe did not finish within %s", uploads.ErrUnreadable, timeout)
+		}
+		if errors.Is(err, errOutputCap) {
+			return uploads.Probe{}, fmt.Errorf("%w: ffprobe printed more than %d bytes", uploads.ErrUnreadable, maxOut)
+		}
 		var exit *exec.ExitError
 		if errors.As(err, &exit) {
-			return uploads.Probe{}, fmt.Errorf("%w: ffprobe: %s", uploads.ErrUnreadable, exit.Stderr)
+			return uploads.Probe{}, fmt.Errorf("%w: ffprobe: %s", uploads.ErrUnreadable, stderr.buf.String())
 		}
 		return uploads.Probe{}, fmt.Errorf("audio: run ffprobe: %w", err)
 	}
-	p, err := ParseProbe(out)
+	p, err := ParseProbe(stdout.buf.Bytes())
 	if err != nil {
 		return uploads.Probe{}, fmt.Errorf("audio: ffprobe output: %w", err)
 	}
@@ -184,3 +184,78 @@ func TestAllowedDemuxers_ExcludeThePlaylistFormats(t *testing.T) {
 		assert.Contains(t, strings.Split(audio.AllowedDemuxers, ","), good)
 	}
 }
+
+func TestFFprobeProbe_CapsTheStreamCountBeforeTheInput(t *testing.T) {
+	// A 1.3 MB file declaring 900 streams cost ~570 MB of RSS in ffprobe (#49) against
+	// a 512 Mi pod. -max_streams is an input option: it must precede the path.
+	bin, argvFile := argvStub(t)
+
+	_, err := (&audio.FFprobe{Bin: bin}).Probe(context.Background(), "/spool/upload-1")
+
+	require.NoError(t, err)
+	args := argvOf(t, argvFile)
+	ms := indexOf(args, "-max_streams")
+	require.GreaterOrEqual(t, ms, 0, "argv: %v", args)
+	assert.Equal(t, "32", args[ms+1])
+	assert.Less(t, ms, len(args)-1, "must come before the path, which stays last")
+}
+
+func TestFFprobeProbe_AsksOnlyForTheFieldsParseProbeReads(t *testing.T) {
+	// -show_format/-show_streams print every tag: a 120 MB WAV produced 491 MB of JSON
+	// that cmd.Output() buffers in Go (#49).
+	bin, argvFile := argvStub(t)
+
+	_, err := (&audio.FFprobe{Bin: bin}).Probe(context.Background(), "/spool/upload-1")
+
+	require.NoError(t, err)
+	args := argvOf(t, argvFile)
+	assert.Equal(t, -1, indexOf(args, "-show_format"), "argv: %v", args)
+	assert.Equal(t, -1, indexOf(args, "-show_streams"), "argv: %v", args)
+	se := indexOf(args, "-show_entries")
+	require.GreaterOrEqual(t, se, 0, "argv: %v", args)
+	assert.Equal(t, "format=duration:stream=codec_type:stream_disposition=attached_pic", args[se+1])
+}
+
+func TestFFprobeProbe_OutputPastTheCapIsUnreadableNotBuffered(t *testing.T) {
+	// ffprobe's stdout is attacker-influenced (tags, 491 MB observed). Past the cap the
+	// probe is stopped and the file is reported as unreadable (#49).
+	bin := stub(t, `head -c 200000 /dev/zero | tr '\0' 'a'`)
+
+	_, err := (&audio.FFprobe{Bin: bin, MaxOutput: 4096}).Probe(context.Background(), "/x")
+
+	assert.ErrorIs(t, err, uploads.ErrUnreadable)
+}
+
+func TestFFprobeProbe_OutputWithinTheCapIsParsed(t *testing.T) {
+	bin := stub(t, `echo '{"streams":[{"codec_type":"audio"}],"format":{"duration":"90"}}'`)
+
+	got, err := (&audio.FFprobe{Bin: bin, MaxOutput: 4096}).Probe(context.Background(), "/x")
+
+	require.NoError(t, err)
+	assert.True(t, got.HasAudio)
+}
+
+func TestFFprobeProbe_AHungProbeTimesOutOnItsOwnDeadlineAsUnreadableNotAsOurFault(t *testing.T) {
+	// Accept runs the probe on the request context, which has no deadline of its own;
+	// a probe that never returns would pin a request (#49). `exec` so the sleep is the
+	// process itself and not a child holding the pipe open.
+	bin := stub(t, `exec sleep 30`)
+
+	start := time.Now()
+	_, err := (&audio.FFprobe{Bin: bin, Timeout: 200 * time.Millisecond}).Probe(context.Background(), "/x")
+
+	assert.ErrorIs(t, err, uploads.ErrUnreadable)
+	assert.Less(t, time.Since(start), 3*time.Second)
+}
+
+func TestFFprobeProbe_ParentCancellationIsStillNotUnreadableWhenATimeoutIsSet(t *testing.T) {
+	bin := stub(t, `exec sleep 30`)
+	ctx, cancel := context.WithTimeout(context.Background(), 100*time.Millisecond)
+	defer cancel()
+
+	_, err := (&audio.FFprobe{Bin: bin, Timeout: time.Minute}).Probe(ctx, "/x")
+
+	require.Error(t, err)
+	assert.False(t, errors.Is(err, uploads.ErrUnreadable), "the caller giving up is not a verdict on the file")
+	assert.ErrorIs(t, err, context.DeadlineExceeded)
+}
@@ -0,0 +1,156 @@
+package llm
+
+// Observes how much of a gateway response the client will read. The body is read before
+// the status is looked at, so a misbehaving gateway or proxy that streams forever was
+// bounded only by the client's overall timeout (#53).
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// endlessGateway answers with the given status and streams filler until the client hangs
+// up; wrote reports how many bytes it managed to send.
+func endlessGateway(t *testing.T, status int, wrote *atomic.Int64) *httptest.Server {
+	t.Helper()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(status)
+		chunk := []byte(strings.Repeat("a", 4096))
+		for {
+			n, err := w.Write(chunk)
+			wrote.Add(int64(n))
+			if err != nil {
+				return
+			}
+			if f, ok := w.(http.Flusher); ok {
+				f.Flush()
+			}
+		}
+	}))
+	t.Cleanup(srv.Close)
+	return srv
+}
+
+func TestClient_ASuccessBodyLargerThanTheCapIsAnErrorAndTheClientStopsReading(t *testing.T) {
+	var wrote atomic.Int64
+	srv := endlessGateway(t, http.StatusOK, &wrote)
+	c := New(srv.URL, "", "m", 30*time.Second, WithMaxResponseBytes(8192))
+
+	start := time.Now()
+	_, err := c.Complete(context.Background(), "s", "u")
+
+	if err == nil || !strings.Contains(err.Error(), "exceeds") {
+		t.Fatalf("want a response-too-large error, got %v", err)
+	}
+	if time.Since(start) > 5*time.Second {
+		t.Fatalf("the client kept reading: took %s", time.Since(start))
+	}
+	// Hanging up lets the server's writes fail; allow what the transport had already buffered.
+	time.Sleep(100 * time.Millisecond)
+	if got := wrote.Load(); got > 8192+1<<20 {
+		t.Fatalf("server managed to send %d bytes after a cap of 8192", got)
+	}
+}
+
+func TestClient_AFailingBodyIsReadOnlyUpToTheErrorBound(t *testing.T) {
+	var wrote atomic.Int64
+	srv := endlessGateway(t, http.StatusBadGateway, &wrote)
+	c := New(srv.URL, "", "m", 30*time.Second)
+
+	start := time.Now()
+	_, err := c.Complete(context.Background(), "s", "u")
+
+	he, ok := err.(*HTTPError)
+	if !ok {
+		t.Fatalf("want *HTTPError, got %T: %v", err, err)
+	}
+	if he.Code != http.StatusBadGateway {
+		t.Fatalf("code %d", he.Code)
+	}
+	if len(he.Body) > maxErrorBody+1 {
+		t.Fatalf("kept %d bytes of an error body, bound is %d", len(he.Body), maxErrorBody)
+	}
+	if time.Since(start) > 5*time.Second {
+		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)
+		}
+	}
+}
@@ -37,6 +37,21 @@ type Client struct {
 	temp       float64
 	httpClient *http.Client
 	usageHook  func(model string, prompt, completion int)
+	maxBody    int64
+}
+
+// defaultMaxResponseBytes bounds a success body. A completion is a few KB of text; this
+// leaves room for a long one and none for a gateway that streams forever (#53).
+const defaultMaxResponseBytes = 8 << 20
+
+// WithMaxResponseBytes overrides the bound on a success response body. A non-positive n is
+// ignored (keeps the default).
+func WithMaxResponseBytes(n int64) Option {
+	return func(c *Client) {
+		if n > 0 {
+			c.maxBody = n
+		}
+	}
 }

 // Option configures a Client at construction. Variadic so the existing 4-arg
@@ -101,6 +116,7 @@ func New(baseURL, apiKey, model string, timeout time.Duration, opts ...Option) *
 		maxTokens:  defaultMaxTokens,
 		temp:       defaultTemperature,
 		httpClient: &http.Client{Timeout: timeout},
+		maxBody:    defaultMaxResponseBytes,
 	}
 	for _, opt := range opts {
 		opt(c)
@@ -184,12 +200,18 @@ func (c *Client) Complete(ctx context.Context, system, user string) (string, err
 	}
 	defer resp.Body.Close() //nolint:errcheck

-	out, err := io.ReadAll(resp.Body)
+	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)}
+	}
+	out, err := io.ReadAll(io.LimitReader(resp.Body, c.maxBody+1))
 	if err != nil {
 		return "", fmt.Errorf("read response: %w", err)
 	}
-	if resp.StatusCode != http.StatusOK {
-		return "", &HTTPError{Code: resp.StatusCode, Body: string(out)}
+	if int64(len(out)) > c.maxBody {
+		return "", fmt.Errorf("LLM response exceeds %d bytes", c.maxBody)
 	}

 	var cr chatResponse
@@ -17,6 +17,10 @@ import (
 // nothing about whether someone else's upload exists (ADR-033).
 var ErrUploadNotFound = errors.New("store: uploaded transcript not found")

+// ErrUploadInProgress is returned by DeleteUpload for an upload that is still queued or
+// transcribing: the worker would write its result to a row that no longer exists.
+var ErrUploadInProgress = errors.New("store: uploaded transcript is still being processed")
+
 // Uploaded transcripts are PRIVATE per-user content, stored in their own
 // RLS-scoped table (migration 017) — never in the shared `transcripts` (ADR-021).
 // Every method goes through withUser, so row-level security confines it to the
@@ -80,23 +84,56 @@ func (s *Store) SaveProgress(ctx context.Context, userID, id string, steps []dom
 		  WHERE id = $1 AND user_id = $2`, b, at)
 }

-// Heartbeat stamps the upload as alive now. It does not touch progress_at: alive says
-// the service is still reporting, not that work is moving (ADR-035).
-func (s *Store) Heartbeat(ctx context.Context, userID, id string) error {
+// Heartbeat stamps the upload as alive at the given time. It does not touch progress_at:
+// alive says the service is still reporting, not that work is moving (ADR-035). The time
+// comes from the application clock, the one progress_at and the page's health check use,
+// so a database clock that disagrees cannot make a healthy job look lost (#51).
+func (s *Store) Heartbeat(ctx context.Context, userID, id string, at time.Time) error {
 	return s.updateUpload(ctx, userID, id, "heartbeat",
-		`UPDATE uploaded_transcripts SET alive_at = NOW()
-		  WHERE id = $1 AND user_id = $2`)
+		`UPDATE uploaded_transcripts SET alive_at = $3
+		  WHERE id = $1 AND user_id = $2`, at)
 }

-// FailUpload moves an upload to failed with a typed reason.
+// FailUpload moves a queued or transcribing upload to failed with a typed reason. A
+// finished upload is left as it was: a shutdown can record a job as failed after its
+// transcript was already committed (the commit lands, the caller sees a cancelled
+// context), and that must not hide the text. Such a call touches no row and returns
+// ErrUploadNotFound.
 func (s *Store) FailUpload(ctx context.Context, userID, id string, kind domain.UploadErrorKind) error {
 	return s.updateUpload(ctx, userID, id, "fail upload",
 		`UPDATE uploaded_transcripts
 		    SET status = 'failed', error_kind = $3, finished_at = NOW()
-		  WHERE id = $1 AND user_id = $2`,
+		  WHERE id = $1 AND user_id = $2 AND status IN ('queued', 'transcribing')`,
 		string(kind))
 }

+// DeleteUpload removes one of the caller's finished (done or failed) uploads. An upload
+// still queued or transcribing is refused with ErrUploadInProgress and left untouched; one
+// that is not the caller's is ErrUploadNotFound, as for every other write (#37).
+func (s *Store) DeleteUpload(ctx context.Context, userID, id string) error {
+	return s.withUser(ctx, userID, func(tx pgx.Tx) error {
+		ct, err := tx.Exec(ctx,
+			`DELETE FROM uploaded_transcripts
+			  WHERE id = $1 AND user_id = $2 AND status IN ('done', 'failed')`, id, userID)
+		if err != nil {
+			return fmt.Errorf("store: delete upload: %w", err)
+		}
+		if ct.RowsAffected() > 0 {
+			return nil
+		}
+		var exists bool
+		if err := tx.QueryRow(ctx,
+			`SELECT EXISTS (SELECT 1 FROM uploaded_transcripts WHERE id = $1 AND user_id = $2)`,
+			id, userID).Scan(&exists); err != nil {
+			return fmt.Errorf("store: delete upload: %w", err)
+		}
+		if exists {
+			return ErrUploadInProgress
+		}
+		return ErrUploadNotFound
+	})
+}
+
 // updateUpload runs one UPDATE keyed on ($1 = id, $2 = userID) plus extra args,
 // and maps "no row touched" to ErrUploadNotFound — which is what a write to another
 // user's upload looks like under RLS (the row is invisible, not forbidden).
@@ -0,0 +1,68 @@
+package store_test
+
+// Deleting one recording (#37). A finished or failed upload can be removed by its owner;
+// one that is queued or transcribing is refused until it ends, because the worker would
+// otherwise write its result to a row that is gone (D3 in the sprint plan).
+
+import (
+	"context"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+)
+
+func TestDeleteUpload_RemovesAFinishedOrFailedUploadOfTheOwner(t *testing.T) {
+	for name, finish := range map[string]func(*store.Store, string) error{
+		"done": func(s *store.Store, id string) error {
+			return s.CompleteUpload(context.Background(), userA, id, domain.Transcription{Text: "synthetic", Endpoint: "iguana/kb-whisper"})
+		},
+		"failed": func(s *store.Store, id string) error {
+			return s.FailUpload(context.Background(), userA, id, domain.UploadErrTranscriptionFailed)
+		},
+	} {
+		t.Run(name, func(t *testing.T) {
+			s := newStore(t)
+			id := startedUpload(t, s)
+			require.NoError(t, s.StartTranscribing(context.Background(), userA, id))
+			require.NoError(t, finish(s, id))
+
+			require.NoError(t, s.DeleteUpload(context.Background(), userA, id))
+
+			_, err := s.GetUpload(context.Background(), userA, id)
+			assert.ErrorIs(t, err, store.ErrUploadNotFound)
+		})
+	}
+}
+
+func TestDeleteUpload_RefusesAnUploadThatIsStillQueuedOrTranscribingAndKeepsTheRow(t *testing.T) {
+	s := newStore(t)
+	ctx := context.Background()
+	queued := startedUpload(t, s)
+	running, err := s.CreateUpload(ctx, userA, "running", false)
+	require.NoError(t, err)
+	require.NoError(t, s.StartTranscribing(ctx, userA, running))
+
+	for _, id := range []string{queued, running} {
+		assert.ErrorIs(t, s.DeleteUpload(ctx, userA, id), store.ErrUploadInProgress)
+		_, err := s.GetUpload(ctx, userA, id)
+		assert.NoError(t, err, "a refused delete must leave the row")
+	}
+}
+
+func TestDeleteUpload_AnotherUsersUploadIsNotFoundAndSurvives(t *testing.T) {
+	s := newStore(t)
+	ctx := context.Background()
+	id := startedUpload(t, s)
+	require.NoError(t, s.StartTranscribing(ctx, userA, id))
+	require.NoError(t, s.FailUpload(ctx, userA, id, domain.UploadErrTranscriptionFailed))
+
+	assert.ErrorIs(t, s.DeleteUpload(ctx, userB, id), store.ErrUploadNotFound,
+		"existence is not disclosed to a non-owner")
+
+	_, err := s.GetUpload(ctx, userA, id)
+	assert.NoError(t, err, "B's delete must not have removed A's row")
+}
@@ -0,0 +1,47 @@
+package store_test
+
+// A shutdown can record a job as failed after its transcript was already committed (the
+// commit lands, the caller sees a cancelled context). FailUpload must not be able to turn a
+// finished row into a failed one and hide the text (#50 review).
+
+import (
+	"context"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/domain"
+)
+
+func TestFailUpload_LeavesAFinishedUploadAsItWas(t *testing.T) {
+	s := newStore(t)
+	ctx := context.Background()
+	id := startedUpload(t, s)
+	require.NoError(t, s.StartTranscribing(ctx, userA, id))
+	require.NoError(t, s.CompleteUpload(ctx, userA, id, domain.Transcription{Text: "synthetic text", Endpoint: "iguana/kb-whisper"}))
+
+	_ = s.FailUpload(ctx, userA, id, domain.UploadErrInterrupted)
+
+	got, err := s.GetUpload(ctx, userA, id)
+	require.NoError(t, err)
+	assert.Equal(t, domain.UploadDone, got.Status)
+	assert.Equal(t, "synthetic text", got.Content)
+	assert.Empty(t, got.ErrorKind)
+}
+
+func TestFailUpload_StillFailsAQueuedOrTranscribingUpload(t *testing.T) {
+	s := newStore(t)
+	ctx := context.Background()
+	queued := startedUpload(t, s)
+	running, err := s.CreateUpload(ctx, userA, "r", false)
+	require.NoError(t, err)
+	require.NoError(t, s.StartTranscribing(ctx, userA, running))
+
+	for _, id := range []string{queued, running} {
+		require.NoError(t, s.FailUpload(ctx, userA, id, domain.UploadErrBusy))
+		got, err := s.GetUpload(ctx, userA, id)
+		require.NoError(t, err)
+		assert.Equal(t, domain.UploadFailed, got.Status)
+	}
+}
@@ -68,13 +68,13 @@ func TestUploadProgress_AHeartbeatMovesAliveAtAndNotProgressAt(t *testing.T) {
 	ctx := context.Background()
 	at := time.Date(2026, 10, 2, 8, 30, 0, 0, time.UTC)
 	require.NoError(t, s.SaveProgress(ctx, userA, id, []domain.ProgressStep{{Key: "audio", Label: "a", State: domain.StepRunning}}, at))
-	before := time.Now().Add(-time.Second)
+	beatAt := at.Add(90 * time.Second) // deliberately not the database's NOW()

-	require.NoError(t, s.Heartbeat(ctx, userA, id))
+	require.NoError(t, s.Heartbeat(ctx, userA, id, beatAt))

 	got, err := s.GetUpload(ctx, userA, id)
 	require.NoError(t, err)
-	assert.True(t, got.AliveAt.After(before), "the heartbeat must stamp the row with now")
+	assert.True(t, got.AliveAt.Equal(beatAt), "the heartbeat is stamped with the application's clock, the same one progress_at uses, not the database's NOW() (#51)")
 	assert.True(t, got.ProgressAt.Equal(at), "alive is not progress: a heartbeat must not move progress_at")
 }

@@ -99,7 +99,7 @@ func TestUploadProgress_ByAnotherUserChangesNothing(t *testing.T) {
 	ctx := context.Background()

 	errSave := s.SaveProgress(ctx, userB, id, []domain.ProgressStep{{Key: "x", Label: "X", State: domain.StepDone}}, time.Now())
-	errBeat := s.Heartbeat(ctx, userB, id)
+	errBeat := s.Heartbeat(ctx, userB, id, time.Now())

 	assert.ErrorIs(t, errSave, store.ErrUploadNotFound)
 	assert.ErrorIs(t, errBeat, store.ErrUploadNotFound)
@@ -37,12 +37,13 @@ type Reporter interface {
 // is for people, so a failed write is logged and never stops the job.
 type Sink interface {
 	SaveProgress(ctx context.Context, steps []domain.ProgressStep, at time.Time) error
-	Beat(ctx context.Context) error
+	Beat(ctx context.Context, at time.Time) error
 }

 const (
-	defaultMinSave = 2 * time.Second
-	logAdvanceEach = time.Minute
+	defaultMinSave     = 2 * time.Second
+	defaultSaveTimeout = 5 * time.Second
+	logAdvanceEach     = time.Minute
 )

 // Option tunes a Tracker.
@@ -55,6 +56,14 @@ func WithClock(now func() time.Time) Option { return func(t *Tracker) { t.now =
 // Done, Fail and Finish always save.
 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 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.
 type Tracker struct {
 	ctx     context.Context
@@ -63,6 +72,8 @@ type Tracker struct {
 	now     func() time.Time
 	minSave time.Duration

+	saveTimeout time.Duration
+
 	mu              sync.Mutex
 	steps           []domain.ProgressStep
 	lastAdvanceSave time.Time
@@ -72,7 +83,7 @@ type Tracker struct {
 // NewTracker plans a job's phases. ctx is used for the saves, so it should outlive the
 // job's own timeout: a job that timed out still gets its failure recorded.
 func NewTracker(ctx context.Context, sink Sink, log *slog.Logger, plan []Planned, opts ...Option) *Tracker {
-	t := &Tracker{ctx: ctx, sink: sink, log: log, now: time.Now, minSave: defaultMinSave}
+	t := &Tracker{ctx: ctx, sink: sink, log: log, now: time.Now, minSave: defaultMinSave, saveTimeout: defaultSaveTimeout}
 	for _, o := range opts {
 		o(t)
 	}
@@ -212,7 +223,10 @@ func (t *Tracker) Heartbeat(ctx context.Context, beatEvery, stillEvery time.Dura
 		case <-ctx.Done():
 			return
 		case <-beat.C:
-			if err := t.sink.Beat(ctx); err != nil && ctx.Err() == nil {
+			beatCtx, cancel := context.WithTimeout(ctx, t.saveTimeout)
+			err := t.sink.Beat(beatCtx, t.now())
+			cancel()
+			if err != nil && ctx.Err() == nil {
 				t.log.Warn("progress: heartbeat not saved", "err", err)
 			}
 		case <-still.C:
@@ -254,7 +268,9 @@ func (t *Tracker) find(key string) *domain.ProgressStep {
 // dropped: progress is cosmetic and must never fail the job it describes.
 func (t *Tracker) save() {
 	snapshot := append([]domain.ProgressStep(nil), t.steps...)
-	if err := t.sink.SaveProgress(t.ctx, snapshot, t.now()); err != nil && t.ctx.Err() == nil {
+	ctx, cancel := context.WithTimeout(t.ctx, t.saveTimeout)
+	defer cancel()
+	if err := t.sink.SaveProgress(ctx, snapshot, t.now()); err != nil && t.ctx.Err() == nil {
 		t.log.Warn("progress: not saved", "err", err)
 	}
 }
@@ -0,0 +1,154 @@
+package progress_test
+
+// Observes what a database that stops answering does to the job it reports on (#51):
+// progress is cosmetic, so a save that never returns must not stall the engine calls that
+// report through the tracker, nor the heartbeat that says the process is alive.
+
+import (
+	"bytes"
+	"context"
+	"log/slog"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/progress"
+)
+
+// stallSink never answers until its context ends, like a database connection that hung.
+type stallSink struct{ beats atomic.Int32 }
+
+func (s *stallSink) SaveProgress(ctx context.Context, _ []domain.ProgressStep, _ time.Time) error {
+	<-ctx.Done()
+	return ctx.Err()
+}
+
+func (s *stallSink) Beat(ctx context.Context, _ time.Time) error {
+	s.beats.Add(1)
+	<-ctx.Done()
+	return ctx.Err()
+}
+
+type lockedBuf struct {
+	mu sync.Mutex
+	b  bytes.Buffer
+}
+
+func (l *lockedBuf) Write(p []byte) (int, error) {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	return l.b.Write(p)
+}
+
+func (l *lockedBuf) String() string {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	return l.b.String()
+}
+
+func stallTracker(sink progress.Sink, buf *lockedBuf) *progress.Tracker {
+	log := slog.New(slog.NewTextHandler(buf, nil))
+	return progress.NewTracker(context.Background(), sink, log,
+		[]progress.Planned{{Key: "a", Label: "A"}}, progress.WithSaveTimeout(50*time.Millisecond))
+}
+
+func TestTracker_AHungSaveDoesNotBlockTheReporterPastItsTimeout(t *testing.T) {
+	buf := &lockedBuf{}
+	tr := stallTracker(&stallSink{}, buf)
+
+	done := make(chan struct{})
+	go func() { tr.Begin("a", 0); tr.Advance("a", 1); tr.Done("a"); close(done) }()
+
+	select {
+	case <-done:
+	case <-time.After(3 * time.Second):
+		t.Fatal("the reporter was blocked by a save that never returned")
+	}
+	require.Contains(t, buf.String(), "progress: not saved", "a dropped save is logged")
+}
+
+func TestTracker_AHungBeatDoesNotStopTheHeartbeat(t *testing.T) {
+	sink := &stallSink{}
+	tr := stallTracker(sink, &lockedBuf{})
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	go tr.Heartbeat(ctx, 5*time.Millisecond, time.Hour)
+
+	require.Eventually(t, func() bool { return sink.beats.Load() >= 3 }, 3*time.Second, 5*time.Millisecond,
+		"each beat is abandoned at its timeout and the next one is tried")
+}
+
+func TestTracker_TheStillRunningLineIsNotHeldUpBehindAHungSave(t *testing.T) {
+	buf := &lockedBuf{}
+	tr := stallTracker(&stallSink{}, buf)
+	go tr.Begin("a", 0) // holds the tracker's lock inside a save that hangs until its timeout
+	time.Sleep(10 * time.Millisecond)
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	go tr.Heartbeat(ctx, time.Hour, 5*time.Millisecond)
+
+	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")
+	}
+}
@@ -17,11 +17,12 @@ import (

 // fakeSink records every saved snapshot and every heartbeat.
 type fakeSink struct {
-	mu    sync.Mutex
-	saves [][]domain.ProgressStep
-	at    []time.Time
-	beats int
-	err   error
+	mu     sync.Mutex
+	saves  [][]domain.ProgressStep
+	at     []time.Time
+	beats  int
+	beatAt []time.Time
+	err    error
 }

 func (f *fakeSink) SaveProgress(_ context.Context, steps []domain.ProgressStep, at time.Time) error {
@@ -32,10 +33,11 @@ func (f *fakeSink) SaveProgress(_ context.Context, steps []domain.ProgressStep,
 	return f.err
 }

-func (f *fakeSink) Beat(context.Context) error {
+func (f *fakeSink) Beat(_ context.Context, at time.Time) error {
 	f.mu.Lock()
 	defer f.mu.Unlock()
 	f.beats++
+	f.beatAt = append(f.beatAt, at)
 	return f.err
 }

@@ -297,7 +299,7 @@ func (r *retainingSink) SaveProgress(_ context.Context, steps []domain.ProgressS
 	r.got = append(r.got, steps)
 	return nil
 }
-func (r *retainingSink) Beat(context.Context) error { return nil }
+func (r *retainingSink) Beat(context.Context, time.Time) error { return nil }

 func TestTracker_ASnapshotHandedToTheSinkIsNotChangedByLaterTransitions(t *testing.T) {
 	sink := &retainingSink{}
@@ -400,3 +402,20 @@ func TestTracker_AnnotateAttachesANoteToAStepAndSavesItAtOnce(t *testing.T) {
 	require.Equal(t, "20 min of the 79 the file claimed", sink.last(t)[0].Note)
 	require.NotPanics(t, func() { tr.Annotate("no-such-step", "x") })
 }
+
+func TestTracker_BeatIsStampedWithTheTrackersClock(t *testing.T) {
+	// alive_at used the database's NOW() while the page compared it with the application
+	// clock; a skewed database clock showed a healthy job as lost (#51).
+	tr, sink, clk, _ := newTracker(t)
+	clk.tick(42 * time.Second)
+	want := clk.now()
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	go tr.Heartbeat(ctx, 5*time.Millisecond, time.Hour)
+
+	require.Eventually(t, func() bool { return sink.beatCount() >= 1 }, 2*time.Second, 5*time.Millisecond)
+	sink.mu.Lock()
+	defer sink.mu.Unlock()
+	require.True(t, sink.beatAt[0].Equal(want), "beat stamped %v, the tracker's clock says %v", sink.beatAt[0], want)
+}
@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"io"
 	"os"
+	"sync"
 	"time"

 	"git.d-ma.be/mathias/tapir/internal/domain"
@@ -51,6 +52,53 @@ type Intake struct {
 	sub    Submitter
 	limits Limits
 	tmpDir string
+
+	// MaxReceives and MaxReceivesPerUser bound how many request bodies Accept reads at
+	// once; zero means the defaults. Without them N parallel POSTs each spool up to
+	// MaxBytes while the queue still looks empty, and the spool volume fills (#47).
+	MaxReceives        int
+	MaxReceivesPerUser int
+
+	mu       sync.Mutex
+	receives int
+	byUser   map[string]int
+}
+
+const (
+	defaultMaxReceives        = 4
+	defaultMaxReceivesPerUser = 2
+)
+
+// acquireReceive reserves a receive slot for userID, or reports there is none. Every
+// successful call must be paired with releaseReceive.
+func (in *Intake) acquireReceive(userID string) bool {
+	maxAll, maxUser := in.MaxReceives, in.MaxReceivesPerUser
+	if maxAll <= 0 {
+		maxAll = defaultMaxReceives
+	}
+	if maxUser <= 0 {
+		maxUser = defaultMaxReceivesPerUser
+	}
+	in.mu.Lock()
+	defer in.mu.Unlock()
+	if in.receives >= maxAll || in.byUser[userID] >= maxUser {
+		return false
+	}
+	if in.byUser == nil {
+		in.byUser = map[string]int{}
+	}
+	in.receives++
+	in.byUser[userID]++
+	return true
+}
+
+func (in *Intake) releaseReceive(userID string) {
+	in.mu.Lock()
+	defer in.mu.Unlock()
+	in.receives--
+	if in.byUser[userID]--; in.byUser[userID] <= 0 {
+		delete(in.byUser, userID)
+	}
 }

 // NewIntake wires an Intake. tmpDir is where uploads are spooled; it must exist.
@@ -94,6 +142,11 @@ func (in *Intake) Accept(ctx context.Context, userID, title string, src io.Reade
 	if rc, ok := in.sub.(interface{ HasRoom(userID string) bool }); ok && !rc.HasRoom(userID) {
 		return "", &RejectError{Reason: ReasonBusy, Message: busyMessage}
 	}
+	// HasRoom cannot see bodies still being received; bound those separately.
+	if !in.acquireReceive(userID) {
+		return "", &RejectError{Reason: ReasonBusy, Message: busyMessage}
+	}
+	defer in.releaseReceive(userID)
 	f, err := os.CreateTemp(in.tmpDir, "upload-*")
 	if err != nil {
 		return "", fmt.Errorf("uploads: spool: %w", err)
@@ -0,0 +1,131 @@
+package uploads_test
+
+// Observes how many request bodies Accept will read at once (#47). HasRoom and the
+// per-user queue cap act before and after the receive; N parallel POSTs all pass
+// HasRoom while the queue is still empty, so the receive itself is bounded.
+
+import (
+	"context"
+	"errors"
+	"io"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/uploads"
+)
+
+// slowReceive starts an Accept whose body stays open until the returned writer is closed,
+// and waits until it holds a spool file, which proves it is inside the receive.
+type slowReceive struct {
+	w        *io.PipeWriter
+	done     chan error
+	finished bool
+}
+
+func startSlowReceive(t *testing.T, r *rig, in *uploads.Intake, user string, wantFiles int) *slowReceive {
+	t.Helper()
+	pr, pw := io.Pipe()
+	s := &slowReceive{w: pw, done: make(chan error, 1)}
+	go func() {
+		_, err := in.Accept(context.Background(), user, "t", pr)
+		s.done <- err
+	}()
+	require.Eventually(t, func() bool { return len(r.tmpFiles(t)) == wantFiles }, 2*time.Second, 5*time.Millisecond,
+		"the receive for %s should have started", user)
+	// An abandoned receive ends in a read error, so it never reaches the rig's fakes, which
+	// are not safe for concurrent use. A receive that finish() already ended is a no-op.
+	t.Cleanup(func() {
+		if s.finished {
+			return
+		}
+		_ = pw.CloseWithError(errors.New("test over"))
+		select {
+		case <-s.done:
+		case <-time.After(2 * time.Second):
+		}
+	})
+	return s
+}
+
+func (s *slowReceive) finish(t *testing.T) error {
+	t.Helper()
+	_, _ = s.w.Write([]byte("data"))
+	require.NoError(t, s.w.Close())
+	s.finished = true
+	select {
+	case err := <-s.done:
+		return err
+	case <-time.After(2 * time.Second):
+		t.Fatal("Accept did not return after its body ended")
+		return nil
+	}
+}
+
+func requireBusy(t *testing.T, err error) {
+	t.Helper()
+	var rej *uploads.RejectError
+	require.True(t, errors.As(err, &rej), "want a busy rejection, got %v", err)
+	assert.Equal(t, uploads.ReasonBusy, rej.Reason)
+}
+
+func TestAccept_APerUserLimitOnConcurrentReceivesRefusesTheNextBodyBeforeSpoolingIt(t *testing.T) {
+	r := newRig(t)
+	in := uploads.NewIntake(r.st, r.pr, r.sub, uploads.Limits{MaxBytes: 1024, MaxDuration: time.Hour}, r.tmp)
+	in.MaxReceivesPerUser = 2
+	in.MaxReceives = 10
+	startSlowReceive(t, r, in, "user-1", 1)
+	startSlowReceive(t, r, in, "user-1", 2)
+
+	_, err := in.Accept(context.Background(), "user-1", "t", strings.NewReader("data"))
+
+	requireBusy(t, err)
+	assert.Len(t, r.tmpFiles(t), 2, "the refused body was never spooled")
+	// Another user is not blocked by user-1's two receives.
+	other := startSlowReceive(t, r, in, "user-2", 3)
+	require.NoError(t, other.finish(t))
+}
+
+func TestAccept_AnOverallLimitOnConcurrentReceivesRefusesTheNextUser(t *testing.T) {
+	r := newRig(t)
+	in := uploads.NewIntake(r.st, r.pr, r.sub, uploads.Limits{MaxBytes: 1024, MaxDuration: time.Hour}, r.tmp)
+	in.MaxReceivesPerUser = 5
+	in.MaxReceives = 3
+	for i, u := range []string{"a", "b", "c"} {
+		startSlowReceive(t, r, in, u, i+1)
+	}
+
+	_, err := in.Accept(context.Background(), "d", "t", strings.NewReader("data"))
+
+	requireBusy(t, err)
+	assert.Len(t, r.tmpFiles(t), 3)
+}
+
+func TestAccept_AFinishedReceiveFreesItsSlot(t *testing.T) {
+	r := newRig(t)
+	in := uploads.NewIntake(r.st, r.pr, r.sub, uploads.Limits{MaxBytes: 1024, MaxDuration: time.Hour}, r.tmp)
+	in.MaxReceivesPerUser = 1
+	in.MaxReceives = 5
+	first := startSlowReceive(t, r, in, "user-1", 1)
+	require.NoError(t, first.finish(t))
+
+	_, err := in.Accept(context.Background(), "user-1", "t", strings.NewReader("data"))
+
+	require.NoError(t, err, "the slot is released when Accept returns")
+}
+
+func TestAccept_AFailedReceiveFreesItsSlot(t *testing.T) {
+	r := newRig(t)
+	in := uploads.NewIntake(r.st, r.pr, r.sub, uploads.Limits{MaxBytes: 1024, MaxDuration: time.Hour}, r.tmp)
+	in.MaxReceivesPerUser = 1
+	in.MaxReceives = 5
+	_, err := in.Accept(context.Background(), "user-1", "t", &brokenBody{n: 1})
+	require.ErrorIs(t, err, uploads.ErrClientRead)
+
+	_, err = in.Accept(context.Background(), "user-1", "t", strings.NewReader("data"))
+
+	require.NoError(t, err, "a dropped client must not leak a slot")
+}
@@ -0,0 +1,31 @@
+package uploads_test
+
+// The limits production runs on are the DEFAULTS: nothing outside tests sets
+// MaxReceives or MaxReceivesPerUser. Pinned here with the literal numbers, so changing a
+// default is a visible decision (#47 review). With 4 receives, 20 queued and 1 running at
+// the 200 MiB cap, the spool needs ~5 GiB of the 6 Gi emptyDir (infra); raising any of
+// them or TAPIR_UPLOAD_MAX_BYTES must revisit that arithmetic.
+
+import (
+	"context"
+	"strings"
+	"testing"
+	"time"
+
+	"git.d-ma.be/mathias/tapir/internal/uploads"
+)
+
+func TestNewIntake_DefaultsAllowTwoReceivesPerUserAndFourOverall(t *testing.T) {
+	r := newRig(t)
+	in := uploads.NewIntake(r.st, r.pr, r.sub, uploads.Limits{MaxBytes: 1024, MaxDuration: time.Hour}, r.tmp)
+	startSlowReceive(t, r, in, "a", 1)
+	startSlowReceive(t, r, in, "a", 2)
+
+	_, err := in.Accept(context.Background(), "a", "t", strings.NewReader("x"))
+	requireBusy(t, err) // a third body from one user
+
+	startSlowReceive(t, r, in, "b", 3)
+	startSlowReceive(t, r, in, "b", 4)
+	_, err = in.Accept(context.Background(), "c", "t", strings.NewReader("x"))
+	requireBusy(t, err) // a fifth body overall
+}
@@ -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)
+	}
+}
@@ -47,7 +47,7 @@ type WorkerStore interface {
 	CompleteUpload(ctx context.Context, userID, id string, t domain.Transcription) error
 	FailUpload(ctx context.Context, userID, id string, kind domain.UploadErrorKind) error
 	SaveProgress(ctx context.Context, userID, id string, steps []domain.ProgressStep, at time.Time) error
-	Heartbeat(ctx context.Context, userID, id string) error
+	Heartbeat(ctx context.Context, userID, id string, at time.Time) error
 }

 // WorkerOption tunes a Worker.
@@ -91,7 +91,9 @@ func (s jobSink) SaveProgress(ctx context.Context, steps []domain.ProgressStep,
 	return s.st.SaveProgress(ctx, s.userID, s.id, steps, at)
 }

-func (s jobSink) Beat(ctx context.Context) error { return s.st.Heartbeat(ctx, s.userID, s.id) }
+func (s jobSink) Beat(ctx context.Context, at time.Time) error {
+	return s.st.Heartbeat(ctx, s.userID, s.id, at)
+}

 // Worker transcribes accepted uploads in-process, ONE AT A TIME (ADR-033). It is
 // not a second deployable: it shares the web process's lifetime, which is why it
@@ -176,9 +178,10 @@ func (w *Worker) Submit(j Job) error {
 }

 // Run processes queued jobs until ctx ends, then deletes the audio of anything
-// still queued and refuses new jobs. The rows of those jobs stay queued or
-// transcribing and are failed as interrupted by the start-up sweep
-// (store.FailInterruptedUploads): the audio is not retained, so they cannot resume.
+// still queued and refuses new jobs. The job running at shutdown is recorded interrupted
+// by Process; the rows of jobs still queued stay queued and are failed as interrupted by
+// the start-up sweep (store.FailInterruptedUploads): the audio is not retained, so they
+// cannot resume.
 func (w *Worker) Run(ctx context.Context) {
 	defer w.stop()
 	for {
@@ -227,7 +230,7 @@ func (w *Worker) stop() {
 // Process runs one job: row -> transcribing, convert, transcribe, store. The
 // uploaded file and the converted WAV are deleted on every path — the audio is
 // never retained (ADR-033). A failure is recorded with a TYPED kind; a shutdown
-// mid-job is not a failure and leaves the row for the start-up sweep.
+// mid-job is recorded as interrupted at once (see fail).
 //
 // Every phase is reported through one progress.Tracker, which saves it to the row for
 // the page and logs it, and a heartbeat stamps the row on its own while the job runs
@@ -269,7 +272,7 @@ func (w *Worker) Process(ctx context.Context, job Job) {
 	defer func() { stopBeat(); <-hbDone }() // no heartbeat outlives its job

 	// abort records where the job stopped. A shutdown (the parent context ended) is not
-	// a failure: the row is left for the start-up sweep, so the step stays as it was.
+	// a failure of the step: it stays as it was, and fail records the row as interrupted.
 	abort := func(kind domain.UploadErrorKind) {
 		if ctx.Err() == nil {
 			tr.FailRunning()
@@ -332,11 +335,19 @@ func (w *Worker) Process(ctx context.Context, job Job) {
 		"segments", len(t.Segments), "elapsed", time.Since(start).Round(time.Second))
 }

-// fail records a transcription failure — unless the PARENT context ended, which
-// is a shutdown, not a failure: the row is left for the start-up sweep.
+// interruptedWriteTimeout bounds the one write a dying process makes to record its job.
+const interruptedWriteTimeout = 3 * time.Second
+
+// fail records a transcription failure. When the PARENT context ended the process is
+// shutting down: the job is recorded interrupted at once, on a short context of its own,
+// because the process may exit before any other pod's start-up sweep would do it (#50).
+// The sweep remains the backstop for a process that is killed outright.
 func (w *Worker) fail(ctx context.Context, job Job, log *slog.Logger, kind domain.UploadErrorKind) {
 	if ctx.Err() != nil {
-		return
+		kind = domain.UploadErrInterrupted
+		var cancel context.CancelFunc
+		ctx, cancel = context.WithTimeout(context.WithoutCancel(ctx), interruptedWriteTimeout)
+		defer cancel()
 	}
 	if err := w.store.FailUpload(ctx, job.UserID, job.UploadID, kind); err != nil {
 		log.Error("upload: record failure", "kind", errkind.Of(err))
@@ -53,7 +53,10 @@ func (s *wStore) CompleteUpload(_ context.Context, _, id string, t domain.Transc
 	return nil
 }

-func (s *wStore) FailUpload(_ context.Context, _, id string, k domain.UploadErrorKind) error {
+func (s *wStore) FailUpload(ctx context.Context, _, id string, k domain.UploadErrorKind) error {
+	if err := ctx.Err(); err != nil { // like the real store: a cancelled context cannot write
+		return err
+	}
 	s.mu.Lock()
 	defer s.mu.Unlock()
 	s.failed[id] = k
@@ -67,7 +70,7 @@ func (s *wStore) SaveProgress(_ context.Context, _, _ string, steps []domain.Pro
 	return nil
 }

-func (s *wStore) Heartbeat(context.Context, string, string) error {
+func (s *wStore) Heartbeat(context.Context, string, string, time.Time) error {
 	time.Sleep(s.beatDelay)
 	s.mu.Lock()
 	defer s.mu.Unlock()
@@ -242,7 +245,9 @@ func TestWorkerProcess_FailuresAreTypedAndStillDeleteTheAudio(t *testing.T) {
 	})
 }

-func TestWorkerProcess_ShutdownMidJobLeavesTheRowForTheStartupSweep(t *testing.T) {
+func TestWorkerProcess_ShutdownMidJobRecordsTheRowInterruptedAtOnce(t *testing.T) {
+	// SIGTERM on a deploy used to leave the row "transcribing" until the NEXT pod's
+	// start-up sweep; the process may exit before any sweep runs (#50).
 	r := newWRig(t, time.Minute)
 	r.tr.block = make(chan struct{})
 	ctx, cancel := context.WithCancel(context.Background())
@@ -255,8 +260,8 @@ func TestWorkerProcess_ShutdownMidJobLeavesTheRowForTheStartupSweep(t *testing.T
 	<-done

 	_, f := r.st.snapshot()
-	assert.Empty(t, f, "a shutdown is not a transcription failure: the row stays for FailInterruptedUploads to mark interrupted")
-	assert.Empty(t, r.leftovers(t), "but the audio is still deleted")
+	assert.Equal(t, domain.UploadErrInterrupted, f["a"], "the row is recorded interrupted by the dying process, not left for a sweep")
+	assert.Empty(t, r.leftovers(t), "and the audio is still deleted")
 }

 func TestWorkerProcess_AJobThatOutlivesItsTimeoutIsFailedNotStuck(t *testing.T) {
@@ -0,0 +1,151 @@
+package web_test
+
+// Observes the cross-origin gate around Router (#48): SameSite=Lax does not stop a POST
+// from a sibling *.d-ma.be origin, so the router itself must refuse a non-safe request
+// whose Origin or Sec-Fetch-Site says it did not come from this site's own pages.
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// foreign is a header set a browser sends for a form post made by another page.
+var foreign = map[string]map[string]string{
+	"sibling subdomain, Sec-Fetch-Site": {"Origin": "https://evil.d-ma.be", "Sec-Fetch-Site": "same-site"},
+	"other site, Sec-Fetch-Site":        {"Origin": "https://evil.example", "Sec-Fetch-Site": "cross-site"},
+	"sibling subdomain, Origin only":    {"Origin": "https://evil.d-ma.be"},
+	"opaque origin":                     {"Origin": "null"},
+}
+
+// own is a header set for the app's own pages (or a non-browser client): must pass.
+var own = map[string]map[string]string{
+	"same-origin Sec-Fetch-Site": {"Sec-Fetch-Site": "same-origin", "Origin": "http://example.com"},
+	"matching Origin only":       {"Origin": "http://example.com"},
+	"no browser headers":         {},
+}
+
+func withHeaders(req *http.Request, h map[string]string) *http.Request {
+	for k, v := range h {
+		req.Header.Set(k, v)
+	}
+	return req
+}
+
+func userCount(t *testing.T) int {
+	t.Helper()
+	var n int
+	require.NoError(t, rawPool(t).QueryRow(context.Background(), `SELECT count(*) FROM users WHERE id = $1`, userID).Scan(&n))
+	return n
+}
+
+func TestCrossOrigin_ARefusedDeleteAccountPostLeavesTheAccountInPlace(t *testing.T) {
+	for name, h := range foreign {
+		t.Run(name, func(t *testing.T) {
+			app, _ := newAccountApp(t)
+			resetDB(t, rawPool(t))
+			seedConnection(t, app, "youtube", "ada@channel", "ref")
+
+			rec := do(t, app, withHeaders(httptest.NewRequest(http.MethodPost, "/account/delete", nil), h))
+
+			assert.Equal(t, http.StatusForbidden, rec.Code)
+			assert.Equal(t, 1, userCount(t), "the account must survive a refused post")
+		})
+	}
+}
+
+func TestCrossOrigin_ASameOriginOrHeaderlessDeleteAccountPostIsServed(t *testing.T) {
+	for name, h := range own {
+		t.Run(name, func(t *testing.T) {
+			app, _ := newAccountApp(t)
+			resetDB(t, rawPool(t))
+			seedConnection(t, app, "youtube", "ada@channel", "ref")
+
+			rec := do(t, app, withHeaders(httptest.NewRequest(http.MethodPost, "/account/delete", nil), h))
+
+			assert.Equal(t, http.StatusSeeOther, rec.Code)
+			assert.Equal(t, "/auth/logout", rec.Header().Get("Location"))
+		})
+	}
+}
+
+func TestCrossOrigin_ARefusedUploadPostSubmitsNoJob(t *testing.T) {
+	for name, h := range foreign {
+		t.Run(name, func(t *testing.T) {
+			app, sub := newUploadApp(t, audioOK, 200<<20)
+
+			rec := do(t, app, withHeaders(uploadReq(t, "m", "m.wav", []byte("RIFF....WAVE")), h))
+
+			assert.Equal(t, http.StatusForbidden, rec.Code)
+			assert.Empty(t, sub.jobs, "a refused post must not enqueue transcription")
+		})
+	}
+}
+
+func TestCrossOrigin_ARefusedUploadChatPostCallsNoModel(t *testing.T) {
+	for name, h := range foreign {
+		t.Run(name, func(t *testing.T) {
+			app, fc := newUploadChatApp(t)
+			id := seedUpload(t, userID, "Roadmap sync", "done", meetingText)
+
+			req := chatPost("/uploads/"+id+"/chat", url.Values{"message": {"hej"}})
+			rec := do(t, app, withHeaders(req, h))
+
+			assert.Equal(t, http.StatusForbidden, rec.Code)
+			assert.Zero(t, fc.callCount)
+		})
+	}
+}
+
+func TestCrossOrigin_AForeignPostIsRefusedOnEachListedStateChangingPath(t *testing.T) {
+	routes := []string{
+		"/account/disconnect/youtube",
+		"/account/summarize-mode",
+		"/register",
+		"/paste",
+		"/v/abc/action",
+		"/v/abc/summarize",
+		"/v/abc/retry-now",
+		"/v/abc/chat",
+	}
+	for _, path := range routes {
+		t.Run(path, func(t *testing.T) {
+			app := newApp(t)
+			req := httptest.NewRequest(http.MethodPost, path, strings.NewReader(""))
+			req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+			rec := do(t, app, withHeaders(req, foreign["sibling subdomain, Sec-Fetch-Site"]))
+			assert.Equal(t, http.StatusForbidden, rec.Code)
+		})
+	}
+}
+
+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) {
+	app := newApp(t)
+	for _, path := range []string{"/healthz", "/welcome"} {
+		req := httptest.NewRequest(http.MethodGet, path, nil)
+		req.Header.Set("Sec-Fetch-Site", "cross-site")
+		assert.Equal(t, http.StatusOK, do(t, app, req).Code, path)
+	}
+}
@@ -59,6 +59,9 @@ type Store interface {
 	// non-owner's id is ErrUploadNotFound.
 	GetUpload(ctx context.Context, userID, id string) (*domain.UploadedTranscript, error)
 	ListUploads(ctx context.Context, userID string) ([]domain.UploadedTranscript, error)
+	// DeleteUpload removes a finished upload; ErrUploadInProgress while it is still being
+	// processed, ErrUploadNotFound for a non-owner.
+	DeleteUpload(ctx context.Context, userID, id string) error

 	// SetTranscriptStatus clears or updates a video's transcript backoff state.
 	// Used by handleRetryNow to clear rate_limited_at before immediate processing.
@@ -200,6 +203,7 @@ func (a *App) Router() http.Handler {
 		app.HandleFunc("POST /uploads", a.handleUploadPost)
 		app.HandleFunc("GET /uploads/{id}", a.handleUploadDetail)
 		app.HandleFunc("GET /uploads/{id}/status", a.handleUploadStatus)
+		app.HandleFunc("POST /uploads/{id}/delete", a.handleUploadDelete)
 		// Chat over the uploaded transcript reuses the ADR-027 Chatter; mounted only
 		// when one is wired.
 		if a.Chat != nil {
@@ -229,7 +233,12 @@ func (a *App) Router() http.Handler {
 	// resolved user_id. /register lives inside the auth guard but is exempt from
 	// the registration gate (you must be able to reach it before you have a user).
 	root.Handle("/", a.Auth.Middleware(a.registrationGate(app)))
-	return root
+
+	// The session cookie is SameSite=Lax, which still travels on a POST made from a
+	// sibling *.d-ma.be page (same site, different origin). Refuse any non-safe request
+	// whose Origin or Sec-Fetch-Site says it did not come from this origin (#48). It wraps
+	// everything, so a route added later is covered without remembering to ask.
+	return http.NewCrossOriginProtection().Handler(root)
 }

 // handleWelcome renders the public landing page (/welcome). It is mounted outside
@@ -304,7 +313,7 @@ func (a *App) handleList(w http.ResponseWriter, r *http.Request) {
 		a.render(w, r, summaryList(buckets, hasConnected, autoSummarize))
 		return
 	}
-	a.render(w, r, ListPage(buckets, f, stats, takeFlash(w, r), hasConnected, channels, autoSummarize))
+	a.render(w, r, ListPage(buckets, f, stats, takeFlash(w, r), hasConnected, channels, autoSummarize, a.Uploads != nil))
 }

 // handleDetail renders one summary in full (highlights, takeaways, action group).
@@ -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())
+}
@@ -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
 }

@@ -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",
@@ -66,7 +66,7 @@ func TestParseYouTubeVideoID(t *testing.T) {
 func TestListPageShowsPasteFormOnlyWhenConnected(t *testing.T) {
 	render := func(connected bool) string {
 		var buf bytes.Buffer
-		if err := ListPage(listBuckets{}, Filter{}, PipelineStats{}, "", connected, nil, true).Render(context.Background(), &buf); err != nil {
+		if err := ListPage(listBuckets{}, Filter{}, PipelineStats{}, "", connected, nil, true, false).Render(context.Background(), &buf); err != nil {
 			t.Fatalf("render: %v", err)
 		}
 		return buf.String()
@@ -18,6 +18,7 @@ import (

 	"git.d-ma.be/mathias/tapir/internal/adapters/chat"
 	"git.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/config"
 	"git.d-ma.be/mathias/tapir/internal/domain"
 	"git.d-ma.be/mathias/tapir/internal/errkind"
 	"git.d-ma.be/mathias/tapir/internal/uploads"
@@ -209,6 +210,32 @@ func (a *App) handleUploadStatus(w http.ResponseWriter, r *http.Request) {
 	a.render(w, r, uploadStatus(*u, a.Chat != nil, a.now()))
 }

+// handleUploadDelete removes one of the caller's finished recordings and returns to the
+// list. Another user's id is a 404 like every other upload route; a recording still being
+// transcribed is refused (409) rather than deleted under the worker (#37).
+func (a *App) handleUploadDelete(w http.ResponseWriter, r *http.Request) {
+	userID, ok := a.currentUserID(w, r)
+	if !ok {
+		return
+	}
+	id := r.PathValue("id")
+	if !uuidShape.MatchString(id) {
+		http.NotFound(w, r)
+		return
+	}
+	switch err := a.Store.DeleteUpload(r.Context(), userID, id); {
+	case err == nil:
+		http.Redirect(w, r, "/uploads", http.StatusSeeOther)
+	case errors.Is(err, store.ErrUploadNotFound):
+		http.NotFound(w, r)
+	case errors.Is(err, store.ErrUploadInProgress):
+		a.renderUploadsPage(w, r, userID, http.StatusConflict,
+			"This recording is still being transcribed. You can delete it once it has finished.")
+	default:
+		a.serverError(w, r, "delete upload", err)
+	}
+}
+
 func (a *App) loadOwnedUpload(w http.ResponseWriter, r *http.Request) (*domain.UploadedTranscript, bool) {
 	userID, ok := a.currentUserID(w, r)
 	if !ok {
@@ -243,6 +270,36 @@ func titleFromFilename(name string) string {
 	return t
 }

+// 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) {
+	endpoint = strings.TrimSpace(endpoint)
+	if endpoint == "" {
+		return "", false
+	}
+	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
 // recording that is shorter than its file claimed, so a record that has one is shown open.
 func hasNote(steps []domain.ProgressStep) bool {
@@ -66,6 +66,12 @@ templ uploadStatus(u domain.UploadedTranscript, chatEnabled bool, now time.Time)
 					Transcribed { uploads.FormatDuration(time.Duration(u.AudioSeconds * float64(time.Second))) }
 					of audio by <span class="chip">{ u.Endpoint }</span>.
 				</p>
+				if provider, cloud := cloudProvider(u.Endpoint); cloud {
+					<p class="cloud-note">
+						This recording's audio left the homelab: it was transcribed by a cloud service ({ provider })
+						because the local engine could not.
+					</p>
+				}
 				@reconciliationNote(u.Reconciliation)
 				if len(u.Progress) > 0 {
 					<details class="what-was-done" if hasNote(u.Progress) { open }>
@@ -80,6 +86,7 @@ templ uploadStatus(u domain.UploadedTranscript, chatEnabled bool, now time.Time)
 					<summary>Transcript</summary>
 					<pre class="transcript">{ u.Content }</pre>
 				</details>
+				@deleteRecording(u.ID)
 			</div>
 		case domain.UploadFailed:
 			<div class="upload-status upload-failed flash flash-error" role="alert">
@@ -91,6 +98,7 @@ templ uploadStatus(u domain.UploadedTranscript, chatEnabled bool, now time.Time)
 						@progressSteps(u.Progress, now, false)
 					</details>
 				}
+				@deleteRecording(u.ID)
 			</div>
 		default:
 			<div
@@ -116,6 +124,20 @@ templ uploadStatus(u domain.UploadedTranscript, chatEnabled bool, now time.Time)
 	}
 }

+// deleteRecording is the confirm-then-delete control, shown only once a recording has
+// finished (done or failed): deleting one that is still being processed is refused.
+templ deleteRecording(id string) {
+	<details class="confirm-delete">
+		<summary class="btn-danger">Delete this recording…</summary>
+		<div class="confirm-body">
+			<p>This permanently deletes the recording and its transcript. Are you sure?</p>
+			<form method="post" action={ templ.SafeURL("/uploads/" + id + "/delete") }>
+				<button type="submit" class="btn-danger">Yes, permanently delete it</button>
+			</form>
+		</div>
+	</details>
+}
+
 // UploadChatPage is the no-JS full page for chatting with an uploaded transcript:
 // the same chat panel as a video's (ADR-027), scoped to this recording.
 templ UploadChatPage(u domain.UploadedTranscript, v chatView) {
@@ -0,0 +1,140 @@
+package web_test
+
+// A transcript produced by the cloud fallback means this recording's audio left the
+// homelab, which the reader is told in plain words (#52). "Local" is config.IsLocalModel:
+// one definition, not a second string match in the view.
+
+import (
+	"context"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func setEndpoint(t *testing.T, id, endpoint string) {
+	t.Helper()
+	_, err := rawPool(t).Exec(context.Background(), `UPDATE uploaded_transcripts SET endpoint = $2 WHERE id = $1`, id, endpoint)
+	require.NoError(t, err)
+}
+
+func TestUploadStatus_ACloudServedTranscriptSaysTheAudioLeftTheHomelab(t *testing.T) {
+	app, _ := newUploadApp(t, audioOK, 200<<20)
+	id := seedUpload(t, userID, "m", "done", "Synthetic transcript text.")
+	setEndpoint(t, id, "berget/kb-whisper")
+
+	b := body(t, do(t, app, get("/uploads/"+id+"/status")))
+
+	assert.Contains(t, b, "left the homelab")
+	assert.Contains(t, b, "cloud service (berget)")
+}
+
+func TestUploadStatus_ALocallyServedTranscriptCarriesNoCloudNote(t *testing.T) {
+	for _, endpoint := range []string{"iguana/kb-whisper", "koala/whisper", ""} {
+		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_AnEndpointThatOnlyLooksLocalIsStillNamedAsCloud(t *testing.T) {
+	// The prefix alone must not decide it: IsLocalModel refuses path tricks.
+	app, _ := newUploadApp(t, audioOK, 200<<20)
+	id := seedUpload(t, userID, "m", "done", "Synthetic transcript text.")
+	setEndpoint(t, id, "iguana/../berget/kb-whisper")
+
+	b := body(t, do(t, app, get("/uploads/"+id+"/status")))
+
+	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)")
+}
@@ -0,0 +1,85 @@
+package web_test
+
+// Deleting one recording (#37). Owner-scoped like every other upload route: another user's
+// id is a 404 and the row survives; an upload still being processed is refused.
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func deleteReq(id string) *http.Request {
+	return httptest.NewRequest(http.MethodPost, "/uploads/"+id+"/delete", nil)
+}
+
+func uploadExists(t *testing.T, id string) bool {
+	t.Helper()
+	var n int
+	require.NoError(t, rawPool(t).QueryRow(context.Background(), `SELECT count(*) FROM uploaded_transcripts WHERE id = $1`, id).Scan(&n))
+	return n == 1
+}
+
+func TestUploadDelete_AnOwnerRemovesAFinishedOrFailedRecordingAndLandsOnTheList(t *testing.T) {
+	for _, status := range []string{"done", "failed"} {
+		t.Run(status, func(t *testing.T) {
+			app, _ := newUploadApp(t, audioOK, 200<<20)
+			id := seedUpload(t, userID, "Roadmap sync", status, "transcription_failed")
+
+			rec := do(t, app, deleteReq(id))
+
+			assert.Equal(t, http.StatusSeeOther, rec.Code)
+			assert.Equal(t, "/uploads", rec.Header().Get("Location"))
+			assert.False(t, uploadExists(t, id))
+		})
+	}
+}
+
+func TestUploadDelete_AnotherUsersRecordingIs404AndSurvives(t *testing.T) {
+	app, _ := newUploadApp(t, audioOK, 200<<20)
+	theirs := seedUpload(t, otherUserID, "Their private meeting", "done", meetingText)
+
+	rec := do(t, app, deleteReq(theirs))
+
+	assert.Equal(t, http.StatusNotFound, rec.Code)
+	assert.True(t, uploadExists(t, theirs), "a non-owner's delete must not remove the row")
+}
+
+func TestUploadDelete_ARecordingStillBeingTranscribedIsRefusedWith409AndKept(t *testing.T) {
+	for _, status := range []string{"queued", "transcribing"} {
+		t.Run(status, func(t *testing.T) {
+			app, _ := newUploadApp(t, audioOK, 200<<20)
+			id := seedUpload(t, userID, "m", status, "")
+
+			rec := do(t, app, deleteReq(id))
+
+			assert.Equal(t, http.StatusConflict, rec.Code)
+			assert.Contains(t, body(t, rec), "still being transcribed")
+			assert.True(t, uploadExists(t, id))
+		})
+	}
+}
+
+func TestUploadDelete_ANonIDIsA404NotAServerError(t *testing.T) {
+	app, _ := newUploadApp(t, audioOK, 200<<20)
+
+	assert.Equal(t, http.StatusNotFound, do(t, app, deleteReq("not-a-uuid")).Code)
+}
+
+func TestUploadPage_OffersDeleteOnlyForAFinishedOrFailedRecording(t *testing.T) {
+	for status, want := range map[string]bool{"done": true, "failed": true, "queued": false, "transcribing": false} {
+		t.Run(status, func(t *testing.T) {
+			app, _ := newUploadApp(t, audioOK, 200<<20)
+			id := seedUpload(t, userID, "m", status, "transcription_failed")
+
+			b := body(t, do(t, app, get("/uploads/"+id)))
+
+			assert.Equal(t, want, strings.Contains(b, `action="/uploads/`+id+`/delete"`))
+		})
+	}
+}
@@ -0,0 +1,19 @@
+package web_test
+
+// /uploads was mounted but nothing linked to it (#37). The shared Layout has no access to
+// the App, so the entry point is the home list, shown only when uploads are mounted.
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestHomeList_LinksToRecordingsOnlyWhenUploadsAreMounted(t *testing.T) {
+	with, _ := newUploadApp(t, audioOK, 200<<20)
+	without := newApp(t)
+
+	assert.Contains(t, body(t, do(t, with, get("/"))), `href="/uploads"`)
+	assert.NotContains(t, body(t, do(t, without, get("/"))), `href="/uploads"`,
+		"a link to a route that is not mounted would be a dead end")
+}
@@ -107,9 +107,12 @@ templ flashBanner(code string) {
 // #summary-list region; a non-HTMX request renders the whole page. flash carries
 // a one-shot notification (e.g. "connected", "registered") surfaced on arrival
 // after a POST→redirect.
-templ ListPage(b listBuckets, f Filter, stats PipelineStats, flash string, hasConnected bool, channels []string, autoSummarize bool) {
+templ ListPage(b listBuckets, f Filter, stats PipelineStats, flash string, hasConnected bool, channels []string, autoSummarize bool, uploadsEnabled bool) {
 	@Layout("Tapir — Summaries") {
 		@flashBanner(flash)
+		if uploadsEnabled {
+			<p class="muted"><a href="/uploads">Recordings</a>: upload audio and get a private transcript.</p>
+		}
 		if hasConnected {
 			@pasteForm()
 		}
@@ -122,6 +122,12 @@ var scenarioCoverage = map[string]string{
 	"I can ask about my recording and get a grounded answer":                    "TestUploadChatPost_SendsTheOwnersTranscriptWithTheUploadBudget",
 	"A long recording is not cut to its first third":                            "TestAnswerMaxCharsOverridesTheServiceCap",

+	"A transcript made by the cloud fallback says the audio left the homelab": "TestUploadStatus_ACloudServedTranscriptSaysTheAudioLeftTheHomelab",
+
+	"I can delete a finished recording":                        "TestUploadDelete_AnOwnerRemovesAFinishedOrFailedRecordingAndLandsOnTheList",
+	"A recording still being transcribed cannot be deleted":    "TestUploadDelete_ARecordingStillBeingTranscribedIsRefusedWith409AndKept",
+	"The home list links to my recordings when uploads are on": "TestHomeList_LinksToRecordingsOnlyWhenUploadsAreMounted",
+
 	// audio_upload.feature, ADR-034
 	"Two engines transcribe the audio and their disagreements are reconciled": "TestEnsemble_BothEnginesRunInOrderAndTheirDisagreementIsReconciled",
 	"Trivial disagreements never reach the model":                             "TestReconcile_TrivialDisputesAreResolvedWithoutAskingTheModel",
@@ -0,0 +1,143 @@
+package lintgate
+
+// Observes the exit status of `task lint` (the lint step of `task check`), not the lint
+// rules themselves. It runs the repo's real Taskfile in a scratch module so a change to the
+// lint command that stops it failing on findings (#41) turns this red.
+
+import (
+	"bytes"
+	"errors"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"regexp"
+	"strings"
+	"testing"
+)
+
+const scratchWithFinding = `package main
+
+import "os"
+
+func main() { os.Remove("x") }
+`
+
+func scratchModule(t *testing.T) string {
+	t.Helper()
+	taskfile, err := os.ReadFile(filepath.Join("..", "..", "Taskfile.yml"))
+	if err != nil {
+		t.Fatalf("read Taskfile.yml: %v", err)
+	}
+	dir := t.TempDir()
+	for name, body := range map[string]string{
+		"Taskfile.yml": string(taskfile),
+		"go.mod":       "module scratch\n\ngo 1.21\n",
+		"main.go":      scratchWithFinding,
+	} {
+		if err := os.WriteFile(filepath.Join(dir, name), []byte(body), 0o644); err != nil {
+			t.Fatalf("write %s: %v", name, err)
+		}
+	}
+	return dir
+}
+
+// ansi matches the colour escape sequences task prints when it thinks it has a terminal.
+var ansi = regexp.MustCompile("\x1b\\[[0-9;]*m")
+
+// hasSkipNotice reports whether the skip notice was printed as output. Task echoes the command
+// line first, and that line contains the same words, so a substring match would be vacuous.
+func hasSkipNotice(out string) bool {
+	for _, line := range strings.Split(ansi.ReplaceAllString(out, ""), "\n") {
+		if strings.HasPrefix(line, "golangci-lint not installed") {
+			return true
+		}
+	}
+	return false
+}
+
+// runLint runs `task lint` in dir with the given PATH and returns its combined output and exit code.
+func runLint(t *testing.T, dir, path string) (string, int) {
+	t.Helper()
+	return runTask(t, dir, path, "lint")
+}
+
+func runTask(t *testing.T, dir, path string, args ...string) (string, int) {
+	t.Helper()
+	taskBin, err := exec.LookPath("task")
+	if err != nil {
+		t.Skip("task not installed")
+	}
+	cmd := exec.Command(taskBin, args...)
+	cmd.Dir = dir
+	cmd.Env = append(os.Environ(), "PATH="+path)
+	var out bytes.Buffer
+	cmd.Stdout, cmd.Stderr = &out, &out
+	err = cmd.Run()
+	var exitErr *exec.ExitError
+	switch {
+	case err == nil:
+		return out.String(), 0
+	case errors.As(err, &exitErr):
+		return out.String(), exitErr.ExitCode()
+	default:
+		t.Fatalf("run task lint: %v", err)
+		return "", -1
+	}
+}
+
+func TestTaskLint_ExitsNonZeroWhenGolangciLintReportsAFinding(t *testing.T) {
+	if _, err := exec.LookPath("golangci-lint"); err != nil {
+		t.Skip("golangci-lint not installed")
+	}
+	out, code := runLint(t, scratchModule(t), os.Getenv("PATH"))
+	if code == 0 {
+		t.Fatalf("task lint exited 0 on a module with an unchecked error; output:\n%s", out)
+	}
+	if hasSkipNotice(out) {
+		t.Fatalf("task lint blamed a missing binary for a lint failure; output:\n%s", out)
+	}
+	if !strings.Contains(out, "errcheck") {
+		t.Fatalf("a non-zero exit without the finding means the linter crashed, not that it found something; output:\n%s", out)
+	}
+}
+
+// The lint step only gates anything if `task check` runs it.
+func TestTaskCheck_RunsTheLintStep(t *testing.T) {
+	out, code := runTask(t, scratchModule(t), os.Getenv("PATH"), "--dry", "check")
+	if code != 0 {
+		t.Fatalf("task --dry check exited %d; output:\n%s", code, out)
+	}
+	if !strings.Contains(out, "task: [lint]") {
+		t.Fatalf("task check no longer runs the lint step; output:\n%s", out)
+	}
+}
+
+func TestTaskLint_SkipsWithANoticeWhenGolangciLintIsAbsent(t *testing.T) {
+	taskBin, err := exec.LookPath("task")
+	if err != nil {
+		t.Skip("task not installed")
+	}
+	bin := t.TempDir() // a PATH holding task and nothing else
+	if err := os.Symlink(taskBin, filepath.Join(bin, "task")); err != nil {
+		t.Fatalf("symlink task: %v", err)
+	}
+	out, code := runLint(t, scratchModule(t), bin)
+	if code != 0 {
+		t.Fatalf("task lint exited %d without golangci-lint on PATH; output:\n%s", code, out)
+	}
+	if !hasSkipNotice(out) {
+		t.Fatalf("expected the skip notice; output:\n%s", out)
+	}
+}
+
+// CI runs task with colour on, so the notice line starts with an escape sequence; a bare
+// prefix match missed it there while passing on a terminal-less local run.
+func TestHasSkipNotice_SeesTheNoticeBehindColourCodes(t *testing.T) {
+	out := "\x1b[32mtask: [lint] ...\n\x1b[0mgolangci-lint not installed — skipping (CI installs it)\n"
+	if !hasSkipNotice(out) {
+		t.Fatalf("notice not found in coloured output: %q", out)
+	}
+	if hasSkipNotice("\x1b[32mtask: [lint] if command -v golangci-lint ...; echo \"golangci-lint not installed\"\n") {
+		t.Fatal("the echoed command line must not count as the notice")
+	}
+}
Author	SHA1	Message	Date
mathiasandClaude Opus 5.5	cace7d179d	fix(deps): bump x/text past GO-2026-5970 CI / Lint / Test / Vet (push) Successful in 30s Details CI / Build & Import (push) Successful in 11s Details CI / Deploy via GitOps (push) Skipped Details 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>	2026-10-04 22:52:55 +02:00
mathiasandClaude Opus 5.5	550764d43c	fix(oidc): retry issuer discovery with bounded back-off at start-up A 503 from the issuer's discovery endpoint made `serve` exit and the pod restart. New now retries discovery 5 times (2+4+8+16 s back-off, ~30 s) before failing as before; a cancelled ctx ends the back-off at once. WithDiscoveryRetry lets tests shrink the schedule. Closes #59 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-04 22:52:26 +02:00
mathiasandClaude Sonnet 5.5	4affeb12e8	docs: ADR-034 second blind check protocol; D8 decided (Berget allowed, per-recording local-only opt-out first, #60 ); #37 verified CI / Lint / Test / Vet (push) Successful in 33s Details CI / Build & Import (push) Successful in 10s Details CI / Deploy via GitOps (push) Successful in 5s Details Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-04 08:47:10 +02:00
mathiasandClaude Sonnet 5.5	460e07719d	docs(sprint): #50 observed live (old pod Completed); #37 and #52 wait for an authenticated look CI / Build & Import (push) Successful in 10s Details CI / Lint / Test / Vet (push) Successful in 32s Details CI / Deploy via GitOps (push) Successful in 4s Details Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-04 00:41:07 +02:00
mathiasandClaude Sonnet 5.5	7b7da650d4	docs(sprint): v0.31.1 and v0.32.0 shipped; what is observed live and what waits for an authenticated look CI / Deploy via GitOps (push) Successful in 5s Details CI / Lint / Test / Vet (push) Successful in 31s Details CI / Build & Import (push) Successful in 10s Details Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-04 00:34:59 +02:00
mathias	85f4cc3236	Merge pull request 'v0.32.0: per-recording delete and nav link, cloud-served note, tracker hardening, bounded LLM reads (#37 #51 #52 #53 )' (#58 ) from sprint/v032 into main CI / Deploy via GitOps (push) Successful in 5s Details CI / Lint / Test / Vet (push) Successful in 34s Details CI / Build & Import (push) Successful in 11s Details	2026-10-03 22:31:26 +00:00
mathiasandClaude Sonnet 5.5	41ab1f5d50	fix(web): take the reconciled endpoint apart exactly, so a connector word used as a model name is still judged (#52 ) CI / Lint / Test / Vet (pull_request) Successful in 35s Details CI / Build & Import (pull_request) Skipped Details CI / Deploy via GitOps (pull_request) Skipped Details Skipping 'by', 'reconciled' and '+' as words everywhere let an alias with one of those names pass as local. The two-engine shape is now split on its own separators and each alias judged whole; a reconciler that is not local is pinned too. Hardening suggested by the third PR #58 review. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-04 00:28:09 +02:00
mathiasandClaude Sonnet 5.5	67bd0350a3	fix(web): judge every word of a reconciled endpoint that is not a connector, so an unprefixed alias is not called local (#52 ) CI / Lint / Test / Vet (pull_request) Successful in 33s Details CI / Build & Import (pull_request) Skipped Details CI / Deploy via GitOps (pull_request) Skipped Details The previous version kept only words containing '/' and dropped the rest, so a bare TAPIR_TRANSCRIBE_MODEL alias in the composite produced a false 'all local' result. Found by the PR #58 re-review; red tests for a bare alias, an alias with a space and a split 'berget / x' first. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-04 00:22:41 +02:00
mathiasandClaude Sonnet 5.5	c24573bd9f	docs: state what the save timeout and a racing delete actually guarantee (#51 , #37 ) CI / Lint / Test / Vet (pull_request) Successful in 29s Details CI / Build & Import (pull_request) Skipped Details CI / Deploy via GitOps (pull_request) Skipped Details The tracker comment said 'never a stalled job'; each transition can wait up to the timeout. The delete refusal is judged at the moment of the attempt. Found by the PR #58 review. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:32:14 +02:00
mathiasandClaude Sonnet 5.5	fc1e8e7ba8	test(web): a foreign post to the recording delete route leaves the recording, on an app that mounts uploads (#37 , #48 ) The earlier entry posted to an app without uploads, where the route does not exist, and stayed green with the route removed. This one seeds a finished row and goes red without the gate. Found by the PR #58 review. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:32:14 +02:00
mathiasandClaude Sonnet 5.5	4a0226d8bd	test(llm): the response cap is pinned at its boundary and as the production default (#53 ) Only the 8-byte override was exercised; >= vs >, a huge default and a non-positive option all stayed green. Also says why the error-body read ignores its error. Found by the PR #58 review. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:31:30 +02:00
mathiasandClaude Sonnet 5.5	2e54d80a78	test(uploads): jobSink forwards the tracker's timestamp to the store (#51 ) The worker fake ignored the heartbeat time, so the store-boundary half of the app-clock fix had no assertion. Found by the PR #58 review. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:31:14 +02:00
mathiasandClaude Sonnet 5.5	01c1606caa	test(progress): the default save and heartbeat deadline is pinned, not only the override (#51 ) Every earlier test passed WithSaveTimeout; the worker passes nothing, so the 5 s default was the only production value and a flip to 0, 1ms or 1h stayed green. Found by the PR #58 review; each flip is now red. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:30:58 +02:00
mathiasandClaude Sonnet 5.5	9aa01468ce	fix(web): judge each alias of a reconciled endpoint, so an all-local transcript is not called cloud-served (#52 ) The two-engine path writes '<base> + <alt>, reconciled by <model>'. IsLocalModel is a single-alias match, so the composite was 'not local' and the note claimed the audio left the homelab, naming a local host as the provider. Found by the PR #58 review. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:30:44 +02:00
mathiasandClaude Sonnet 5.5	25628a9994	test(web): the per-recording delete route is covered by the cross-origin gate (#37 , #48 ) CI / Lint / Test / Vet (pull_request) Successful in 30s Details CI / Build & Import (pull_request) Skipped Details CI / Deploy via GitOps (pull_request) Skipped Details Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:24 +02:00
mathiasandClaude Sonnet 5.5	a6691ac46b	docs: per-recording delete and the recordings link are built; feature scenarios and coverage map (#37 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	c909b4e775	feat(web): confirm-then-delete control on a finished or failed recording (#37 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	f309a39027	feat(web): POST /uploads/{id}/delete, owner-scoped and refused while the recording is active (#37 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	89611164e3	test(web): per-recording delete is owner-scoped, refused while active, offered only when finished (#37 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	93860210ef	feat(store): DeleteUpload removes a finished upload of the owner (#37 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	8509004eb7	test(store): an owner can delete a finished upload; an active or foreign one is refused (#37 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	453464a22b	feat(web): the home list links to recordings when uploads are mounted (#37 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	4ef3e85d16	test(web): the home list links to recordings only when uploads are mounted (#37 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	b8af7f32c7	fix(progress): stamp the heartbeat with the application clock the page compares it with (#51 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	813192e5ba	test: heartbeats are stamped with the application clock, not the database's NOW() (#51 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	2ec1557a8c	fix(progress): bound every save and heartbeat write so a hung database cannot stall the job (#51 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	68815755e8	test(progress): a hung save or beat must not stall the reporter or the heartbeat (#51 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	ba18668796	docs(feature): cloud-served transcripts say the audio left the homelab; empty transcript also triggers the fallback (#52 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	a65f2f2492	feat(web): say in plain words when a recording's audio left the homelab (#52 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	a43ee19505	feat(web): helper naming the cloud provider of a transcript, one definition of local (#52 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	4301e691f5	test(web): a cloud-served transcript says the audio left the homelab (#52 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	61d1ff2255	fix(llm): read a bounded response body, 8 MiB on success and the error bound on failure (#53 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathiasandClaude Sonnet 5.5	9cc954df40	test(llm): the client reads a bounded response body, success and error (#53 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:22:17 +02:00
mathias	e037c6b683	Merge pull request 'v0.31.1: cross-origin POST protection, ffprobe/ffmpeg limits, receive cap, SIGTERM, lint gate (#41 #47 #48 #49 #50 )' (#55 ) from sprint/audio-followups into main CI / Deploy via GitOps (push) Successful in 6s Details CI / Lint / Test / Vet (push) Successful in 30s Details CI / Build & Import (push) Successful in 10s Details	2026-10-03 19:19:55 +00:00
mathiasandClaude Sonnet 5.5	5baffd0189	fix(test): strip colour codes before looking for the lint skip notice (CI run 8937) CI / Lint / Test / Vet (pull_request) Successful in 30s Details CI / Build & Import (pull_request) Skipped Details CI / Deploy via GitOps (pull_request) Skipped Details Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:18:05 +02:00
mathiasandClaude Sonnet 5.5	228bcc77da	test(lint): the skip notice is found behind CI's colour codes, red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:18:01 +02:00
mathiasandClaude Sonnet 5.5	4ef446548d	docs(adr): spool arithmetic and the unbounded slow-drip limitation for the receive cap (#47 ) CI / Lint / Test / Vet (pull_request) Failing after 30s Details CI / Build & Import (pull_request) Skipped Details CI / Deploy via GitOps (pull_request) Skipped Details Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:12:48 +02:00
mathiasandClaude Sonnet 5.5	145949875e	fix(audio): keep a bounded head of ffmpeg's output instead of buffering all of it (#49 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:12:48 +02:00
mathiasandClaude Sonnet 5.5	82aeac7c19	fix(store): FailUpload only fails a queued or transcribing upload (review of #55 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:12:48 +02:00
mathiasandClaude Sonnet 5.5	1b4747ed03	test: FailUpload must not overwrite a finished upload; ffmpeg stderr is bounded, red (review of #55 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:12:34 +02:00
mathiasandClaude Sonnet 5.5	9939afb90c	test: pin the production receive limits, split the vacuous drain test, harden the lint-gate test (review of #55 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:12:34 +02:00
mathiasandClaude Sonnet 5.5	3e69c00a64	docs(adr): ADR-033 amended for SIGTERM handling and the recorded deploy limitation (#50 , #47 ) CI / Lint / Test / Vet (pull_request) Failing after 31s Details CI / Build & Import (pull_request) Skipped Details CI / Deploy via GitOps (pull_request) Skipped Details Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:00:55 +02:00
mathiasandClaude Sonnet 5.5	253aa567aa	fix(serve): handle SIGTERM and wait for in-flight requests and the worker before exiting (#50 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:00:55 +02:00
mathiasandClaude Sonnet 5.5	5644dc3fc8	fix(uploads): record a job interrupted by shutdown at once, on its own short context (#50 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 21:00:55 +02:00
mathiasandClaude Sonnet 5.5	7df1de6628	test(serve): shutdown waits for in-flight requests and the worker, bounded (#50 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:59:57 +02:00
mathiasandClaude Sonnet 5.5	7e920102f6	test: SIGTERM cancels the process context; a shutdown mid-job records the row interrupted at once (#50 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:59:40 +02:00
mathiasandClaude Sonnet 5.5	480c7227b5	fix(uploads): bound concurrent request-body receives, 4 total and 2 per user (#47 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:59:09 +02:00
mathiasandClaude Sonnet 5.5	0eb3be9631	test(uploads): concurrent receives are bounded in total and per user (#47 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:57:31 +02:00
mathiasandClaude Sonnet 5.5	e8c5ac50bf	fix(audio): give ffprobe its own 45s deadline, a timeout is the file's fault (#49 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:56:58 +02:00
mathiasandClaude Sonnet 5.5	6101dd4e2f	test(audio): a hung ffprobe times out on its own deadline as unreadable (#49 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:56:47 +02:00
mathiasandClaude Sonnet 5.5	089444c19f	fix(audio): stop reading ffprobe output past a cap and call the file unreadable (#49 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:56:47 +02:00
mathiasandClaude Sonnet 5.5	d44db88787	test(audio): ffprobe output past a cap is unreadable, not buffered (#49 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:56:37 +02:00
mathiasandClaude Sonnet 5.5	1f8bd849fe	fix(audio): cap ffprobe/ffmpeg streams at 32 and narrow the probe output (#49 ) A 900-stream file cost ~570 MB RSS in each tool; with the cap the same file fails in ~18 MB and is reported as unreadable. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:56:28 +02:00
mathiasandClaude Sonnet 5.5	7b19801bfa	test(audio): ffprobe/ffmpeg must cap streams, ffprobe must narrow its output (#49 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:55:54 +02:00
mathiasandClaude Sonnet 5.5	1260500d32	fix(web): refuse cross-origin non-safe requests at the router (#48 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:55:19 +02:00
mathiasandClaude Sonnet 5.5	10b4c43bd2	test(web): foreign-origin POSTs must be refused by the router (#48 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:55:00 +02:00
mathiasandClaude Sonnet 5.5	f1f8b6d02e	fix(task): lint step fails on findings instead of printing 'not installed' (#41 ) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:54:20 +02:00
mathiasandClaude Sonnet 5.5	0d4c5c9d1d	test(lint): task lint must exit non-zero on a finding (#41 ), red Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:53:50 +02:00
mathiasandClaude Sonnet 5.5	24d65b2085	docs(sprint): audio upload follow-ups sweep sprint plan, oath and decision list CI / Lint / Test / Vet (push) Successful in 27s Details CI / Build & Import (push) Successful in 22s Details CI / Deploy via GitOps (push) Successful in 5s Details Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 20:50:34 +02:00
mathiasandClaude Sonnet 5.5	df7bd459ed	docs(adr): the ingress body-size item is resolved, and uploads are on in production (infra#520) CI / Build & Import (push) Successful in 13s Details CI / Deploy via GitOps (push) Successful in 5s Details CI / Lint / Test / Vet (push) Successful in 36s Details Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>	2026-10-03 16:29:32 +02:00