Merge pull request 'feat(uploads): choose the language of a recording; Swedish default, Other = auto (#61 interim)' (#62) from sprint/language into main

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>

.gitea/build-image.sh

@@ -0,0 +1,121 @@
+#!/usr/bin/env bash
+# Build an OCI image with buildctl and push it to the local registry.
+#
+# CANONICAL COPY: mathias/infra ci-templates/build-image.sh (infra#309).
+# Every repo carries a byte-identical copy at .gitea/build-image.sh, written by
+# `task ci:sync` and held in place by the drift gate. DO NOT EDIT THE COPIES --
+# edit here, re-sync, and the gate will tell you which repos still disagree.
+#
+# Why this exists: this step was hand-copied into a dozen repos, so migrating it
+# off rootless buildah (infra#241) meant eleven near-identical commits, eleven
+# verifications, and three separate rediscoveries of the same fix in the two
+# repo templates and the gitea-ci skill. The differences between the copies were
+# incidental -- a label here, a build-arg there -- never intentional. Those are
+# inputs now, not forks of the implementation.
+#
+# Why buildctl and not buildah: the gitea-runner pod runs with
+# `capabilities: drop [ALL]` and `allowPrivilegeEscalation: false`, which
+# unconditionally blocks the setuid-root exec that rootless buildah's uid/gid
+# remapping needs, and the pod has no OCI runtime at all. buildctl is a thin
+# client -- it hands the build to the host's buildkitd over a socket, so the
+# privileged layer work happens on the host and this pod's securityContext is
+# not in the picture. buildah is absent from the runner image as of baked-v14
+# and its absence is asserted at pod start (infra#308).
+#
+# Inputs, all via environment:
+#
+#   IMAGE            (required) image name, e.g. cobalt-dingo
+#   REGISTRY         default localhost:5000
+#   SHA_TAG          default `git rev-parse --short HEAD`
+#   VERSION_TAG      optional extra tag, e.g. v1.2.3 on a tag build
+#   PUSH_LATEST      default 1; set 0 to skip the :latest tag
+#   CONTEXT          default . -- build context directory
+#   DOCKERFILE_DIR   default $CONTEXT -- directory holding the Dockerfile
+#   DOCKERFILE_NAME  optional, when the file is not named `Dockerfile`
+#   TARGET           optional, multi-stage build target (--opt target=)
+#   BUILD_ARGS       optional, space-separated K=V pairs
+#   BUILD_SECRETS    optional, space-separated buildkit secret specs, each
+#                    `id=NAME,src=PATH`. Pass the SPEC only -- the caller
+#                    writes the secret to PATH itself. A secret VALUE must
+#                    never be passed here: it would land in this script's
+#                    argv and therefore in the CI log.
+#   EXTRA_LABELS     optional, space-separated K=V pairs
+#   GIT_SHA          optional, stamped as org.opencontainers.image.revision
+#   IMAGE_SOURCE     optional, stamped as org.opencontainers.image.source
+#
+# Note on shell style below: every conditional is a full `if` block, never
+# `[ -n "$X" ] && cmd`. Under `set -e` the && form exits the script whenever the
+# test is false, which is a real bug this estate has already shipped once.
+
+set -euo pipefail
+
+if [ -z "${IMAGE:-}" ]; then
+  echo "build-image: IMAGE is required" >&2
+  exit 2
+fi
+
+REGISTRY="${REGISTRY:-localhost:5000}"
+BUILDKIT_HOST="${BUILDKIT_HOST:-unix:///run/buildkit/buildkitd.sock}"
+CONTEXT="${CONTEXT:-.}"
+DOCKERFILE_DIR="${DOCKERFILE_DIR:-$CONTEXT}"
+PUSH_LATEST="${PUSH_LATEST:-1}"
+VERSION_TAG="${VERSION_TAG:-}"
+
+if [ -z "${SHA_TAG:-}" ]; then
+  SHA_TAG="$(git rev-parse --short HEAD)"
+fi
+
+REF="${REGISTRY}/${IMAGE}:${SHA_TAG}"
+
+TAR_FILE="$(mktemp)"
+trap 'rm -f "$TAR_FILE"' EXIT
+
+args=(
+  --frontend dockerfile.v0
+  --local "context=${CONTEXT}"
+  --local "dockerfile=${DOCKERFILE_DIR}"
+)
+
+if [ -n "${DOCKERFILE_NAME:-}" ]; then
+  args+=(--opt "filename=${DOCKERFILE_NAME}")
+fi
+if [ -n "${TARGET:-}" ]; then
+  args+=(--opt "target=${TARGET}")
+fi
+for spec in ${BUILD_SECRETS:-}; do
+  # Opaque id=NAME,src=PATH -- buildkit reads the file, nothing is echoed.
+  args+=(--secret "${spec}")
+done
+if [ -n "${GIT_SHA:-}" ]; then
+  args+=(--opt "label:org.opencontainers.image.revision=${GIT_SHA}")
+fi
+if [ -n "${IMAGE_SOURCE:-}" ]; then
+  args+=(--opt "label:org.opencontainers.image.source=${IMAGE_SOURCE}")
+fi
+for kv in ${BUILD_ARGS:-}; do
+  args+=(--opt "build-arg:${kv}")
+done
+for kv in ${EXTRA_LABELS:-}; do
+  args+=(--opt "label:${kv}")
+done
+
+echo "build-image: building ${REF}"
+buildctl --addr "$BUILDKIT_HOST" build "${args[@]}" --output "type=oci,dest=${TAR_FILE}"
+
+# buildctl has no registry-push equivalent to `buildah push`, so skopeo copies
+# the OCI archive up -- once per tag, from the same archive, so every tag points
+# at bit-identical content.
+push_tag() {
+  echo "build-image: pushing ${1}"
+  skopeo copy --dest-tls-verify=false "oci-archive:${TAR_FILE}" "docker://${1}"
+}
+
+push_tag "$REF"
+if [ "$PUSH_LATEST" = "1" ]; then
+  push_tag "${REGISTRY}/${IMAGE}:latest"
+fi
+if [ -n "$VERSION_TAG" ]; then
+  push_tag "${REGISTRY}/${IMAGE}:${VERSION_TAG}"
+fi
+
+echo "✓ build-image: ${REF} pushed"

.gitea/workflows/ci.yml

@@ -61,33 +61,163 @@ jobs:
             echo "version-tag=${REF#refs/tags/}" >> "$GITHUB_OUTPUT"
           fi
 
+      # infra#241: rootless buildah inside the gitea-runner pod can't remap the
+      # UID/GID range newuidmap/newgidmap need (the pod drops ALL capabilities
+      # and sets allowPrivilegeEscalation:false — both deliberate, it also holds
+      # CI secrets), so extracting a base-image layer that chowns outside the
+      # single available mapping fails with `ApplyLayer ... requested 0:42 for
+      # /etc/gshadow`. Main has been red on exactly this since 2026-08-13.
+      # buildctl instead talks to koala's host buildkitd over the socket the
+      # runner pod already mounts: the privileged extraction happens inside
+      # buildkitd on the host, never in this pod. Same migration agentsquad
+      # (infra#233) and segment-embedder (infra#243) already made — zero
+      # security tradeoff, unlike loosening the pod's capabilities.
+      #
+      # buildctl has no registry-push flag of its own: it writes an OCI archive
+      # and skopeo pushes that, which is also why every tag below is one copy of
+      # a single build rather than a rebuild.
       - name: Build and push to local registry
-        run: |
-          REGISTRY="localhost:5000"
-          REF="${REGISTRY}/${{ env.IMAGE }}:${{ steps.meta.outputs.sha-tag }}"
-          buildah build \
-            --label "org.opencontainers.image.revision=${{ github.sha }}" \
-            --label "org.opencontainers.image.source=${{ github.repositoryUrl }}" \
-            -t ${REF} \
-            -t ${REGISTRY}/${{ env.IMAGE }}:latest \
-            .
-          buildah push --tls-verify=false ${REF}
-          buildah push --tls-verify=false ${REGISTRY}/${{ env.IMAGE }}:latest
-          [[ -n "${{ steps.meta.outputs.version-tag }}" ]] && \
-            buildah push --tls-verify=false \
-              ${REGISTRY}/${{ env.IMAGE }}:${{ steps.meta.outputs.version-tag }} || true
-          echo "Image pushed to ${REF}"
+      # infra#309: implementation lives in .gitea/build-image.sh, a
+      # byte-identical copy of mathias/infra's ci-templates/build-image.sh.
+      # Do not edit it here -- edit the canonical file, run `task ci:sync`,
+      # and infra's drift gate names any repo still carrying an old copy.
+        env:
+          SHA_TAG: ${{ steps.meta.outputs.sha-tag }}
+          VERSION_TAG: ${{ steps.meta.outputs.version-tag }}
+          GIT_SHA: ${{ github.sha }}
+          IMAGE_SOURCE: ${{ github.repositoryUrl }}
+        run: bash .gitea/build-image.sh
 
-      - name: Smoke test
-        run: |
-          REGISTRY="localhost:5000"
-          REF="${REGISTRY}/${{ env.IMAGE }}:${{ steps.meta.outputs.sha-tag }}"
-          CNAME="smoke-${{ steps.meta.outputs.sha-tag }}"
-          sudo k3s ctr images pull --plain-http ${REF}
-          OUTPUT=$(timeout 5 sudo k3s ctr run --rm ${REF} ${CNAME} /tapir 2>&1 || true)
-          sudo k3s ctr containers delete ${CNAME} 2>/dev/null || true
-          echo "$OUTPUT" | grep -q "tapir" \
-            && echo "Smoke test passed" \
-            || echo "Smoke test inconclusive: $OUTPUT"
+      # The buildah-based smoke test that used to sit here is gone with the
+      # buildctl migration above: `buildah from ${REF}` would pull the image
+      # back into the pod's buildah store and hit the very uid/gid wall the
+      # migration exists to avoid — it would fail the job for a reason that has
+      # nothing to do with the image being good. The `Verify rollout` step in
+      # the deploy job below is the stronger check anyway: it requires the real
+      # pod to reach Ready in-cluster, not just the entrypoint to print its own
+      # name for 5s. Tag builds lose their only image-level check (deploy is
+      # main-push only) — accepted; a tag is always built from a main commit
+      # that already deployed.
 
-  # ── 3. Mirror to GitHub — skipped for now (SSH key rotation pending) ─
+  # ── 3. Deploy via infra repo + Flux ─────────────────────────────────────────
+  # Flux native image-automation can't scan localhost:5000 from inside k3s pods
+  # (mathias/infra k3s/flux/flux-system/image-automation.yaml) — this job
+  # mirrors cobalt-dingo's proven pattern instead: patch the infra repo's
+  # manifest directly on every push to main, then annotate Flux for a fast
+  # reconcile. Fixes infra#111 (image built+pushed but manifest bump was
+  # manual, so merged features silently didn't deploy).
+  deploy:
+    name: Deploy via GitOps
+    needs: build
+    runs-on: self-hosted
+    if: github.ref == 'refs/heads/main' && github.event_name == 'push'
+    environment: staging
+    steps:
+      - name: Update image tag in infra repo
+        env:
+          IMAGE_TAG: ${{ needs.build.outputs.image-tag }}
+          DEPLOY_KEY: ${{ secrets.INFRA_DEPLOY_KEY }}
+        run: |
+          set -euo pipefail
+          # infra#311: fail on an empty tag BEFORE patching anything. The
+          # grep verification further down CANNOT catch it -- with IMAGE_TAG
+          # empty its pattern matches the very corruption it exists to reject.
+          # The result is `image: <repo>:` in the manifest, which is invalid
+          # YAML, so Flux fails the ENTIRE apps Kustomization and every app
+          # stops reconciling, not just this one. That happened on 2026-08-25.
+          [ -n "${IMAGE_TAG:-}" ] || { echo "IMAGE_TAG is empty - refusing to patch"; exit 1; }
+
+          # INFRA_DEPLOY_KEY is a Gitea org secret (mathias org), already
+          # configured per docs/cd-pipeline.md in the infra repo — same key
+          # cobalt-dingo and brain-gardener use, no new secret needed.
+          mkdir -p ~/.ssh
+          echo "$DEPLOY_KEY" > ~/.ssh/id_infra
+          chmod 600 ~/.ssh/id_infra
+          ssh-keyscan -p 30022 10.0.1.20 >> ~/.ssh/known_hosts 2>/dev/null
+
+          export GIT_SSH_COMMAND="ssh -i ~/.ssh/id_infra -o IdentitiesOnly=yes"
+          rm -rf /tmp/infra
+          git clone -b main ssh://git@10.0.1.20:30022/mathias/infra.git /tmp/infra
+          cd /tmp/infra
+
+          DEPLOYMENT="k3s/apps/tapir/deployment.yaml"
+          # In-place update of the image tag. sed (not yq) so we don't
+          # depend on additional tooling on the runner — same as cobalt-dingo.
+          sed -i "s|image: localhost:5000/tapir:.*|image: localhost:5000/tapir:${IMAGE_TAG}|" "$DEPLOYMENT"
+
+          # Verify the patch took effect.
+          grep -q "localhost:5000/tapir:${IMAGE_TAG}" "$DEPLOYMENT" \
+            || { echo "✗ image tag patch failed"; exit 1; }
+
+          if git diff --quiet "$DEPLOYMENT"; then
+            echo "ℹ image tag unchanged — skipping push"
+          else
+            git -c user.name="tapir CI" \
+                -c user.email="ci@tapir.local" \
+                commit -m "chore(deploy): tapir → ${IMAGE_TAG}" "$DEPLOYMENT"
+            git push origin main
+            echo "✓ pushed to infra repo"
+          fi
+
+          shred -u ~/.ssh/id_infra
+
+      - name: Trigger Flux reconcile (immediate)
+        run: |
+          # Without these annotations, Flux would still pick up the change
+          # within 30s (the apps Kustomization interval). The annotations
+          # cut latency to ~1s.
+          kubectl -n flux-system annotate gitrepository flux-system \
+            reconcile.fluxcd.io/requestedAt="$(date +%s)" --overwrite
+          kubectl -n flux-system annotate kustomization apps \
+            reconcile.fluxcd.io/requestedAt="$(date +%s)" --overwrite
+
+      - name: Wait for Flux to apply new image
+        env:
+          IMAGE_TAG: ${{ needs.build.outputs.image-tag }}
+        run: |
+          # Poll the Deployment spec until it reflects the new tag.
+          # Bound to 60s so a stuck Flux doesn't hang CI.
+          EXPECTED="localhost:5000/tapir:${IMAGE_TAG}"
+          for i in $(seq 1 60); do
+            CURRENT=$(kubectl get deploy tapir -n tapir \
+              -o jsonpath='{.spec.template.spec.containers[0].image}' 2>/dev/null || echo "")
+            if [ "$CURRENT" = "$EXPECTED" ]; then
+              echo "✓ Flux applied new image after ${i}s"
+              break
+            fi
+            sleep 1
+          done
+          # Final assertion (in case the loop exited without matching).
+          kubectl get deploy tapir -n tapir \
+            -o jsonpath='{.spec.template.spec.containers[0].image}' \
+            | grep -qx "$EXPECTED" \
+            || { echo "✗ Flux did not apply new image within 60s"; exit 1; }
+
+      - name: Verify rollout
+        run: |
+          kubectl rollout status deployment/tapir \
+            --namespace tapir \
+            --timeout=120s \
+          || {
+            echo "── pod status ──"
+            kubectl get pods -n tapir -o wide
+            echo "── events ──"
+            kubectl get events -n tapir --sort-by='.lastTimestamp' | tail -20
+            echo "── describe ──"
+            kubectl describe pods -n tapir -l app=tapir | tail -40
+            exit 1
+          }
+
+      - name: Confirm pod running new image
+        env:
+          IMAGE_TAG: ${{ needs.build.outputs.image-tag }}
+        run: |
+          kubectl get pods -n tapir \
+            -l app=tapir \
+            --field-selector=status.phase=Running \
+            -o jsonpath='{.items[*].spec.containers[0].image}' \
+          | grep -q "localhost:5000/tapir:${IMAGE_TAG}" \
+          && echo "✓ pod running new image" \
+          || { echo "✗ pod image mismatch"; exit 1; }
+
+  # ── 4. Mirror to GitHub — skipped for now (SSH key rotation pending) ─

.gitignore

@@ -11,3 +11,14 @@
 .env.*
 !.env.example
 *.local
+
+# Spike media (#28): real recordings, their transcripts and derived analyses are
+# private third-party content and this repo is public. Only synthetic fixtures
+# are committed — see scripts/spike-media/README.md.
+/scripts/spike-media/*.mov
+/scripts/spike-media/*.mp4
+/scripts/spike-media/*.wav
+/scripts/spike-media/*.srt
+/scripts/spike-media/*.json
+/scripts/spike-media/*.html
+!/scripts/spike-media/fixtures/

.mcp.json

@@ -0,0 +1,18 @@
+{
+  "mcpServers": {
+    "brain": {
+      "type": "http",
+      "url": "https://brain-mcp.d-ma.be/mcp",
+      "headers": {
+        "Authorization": "Bearer ${BRAIN_MCP_TOKEN}"
+      }
+    },
+    "gitea": {
+      "type": "http",
+      "url": "https://git-mcp.d-ma.be/mcp",
+      "headers": {
+        "Authorization": "Bearer ${GITEA_MCP_TOKEN}"
+      }
+    }
+  }
+}

CLAUDE.md

@@ -12,6 +12,7 @@ docs it indexes.
 3. `DECISIONS.md` — the ADRs. Decisions are settled here; do not re-litigate without a new ADR.
 4. `docs/architecture/architecture.md`, `docs/data-model.md`, `docs/use-cases/*.feature`.
 5. `docs/homelab-integration.md` — the concrete endpoints/conventions you'll need.
+6. `LANGUAGE.md` — the project vocabulary. Apply the caveman rubric before destructive operations.
 
 ## How to work in this repo
 
@@ -49,8 +50,11 @@ These caused real mistakes that were caught and corrected; the corrections are l
 - **No global cross-tenant *video* table** — videos stay per-user (data-model). Transcripts ARE
   shared since ADR-021 (public caption content, keyed by `(provider, provider_video_id)`, non-RLS)
   so re-analysis never re-fetches; the *videos* half of cross-tenant dedup stays a Future C concern.
-- **No audio-download + speech-to-text in the core path** — captions-first (ADR-007). STT is a
-  deferred, bounded optional component.
+- **No YouTube audio-download + speech-to-text in the core path** — captions-first (ADR-007).
+  That rejection stands. ADR-033 superseded the STT deferral **for user-supplied audio only**:
+  an authenticated user may upload a recording and get a *private* transcript (own RLS table,
+  never the shared `transcripts`). It is off by default (`TAPIR_UPLOADS_ENABLED`) and does not
+  open the door to scraping audio.
 - **No public SaaS / sign-up / billing / Google OAuth verification at scale** — Future C,
   deferred behind the Stage 0 gate (ADR-008).
 
@@ -96,7 +100,7 @@ The repo is **green and shipping** — last tag `v0.15.0`. `task check` passes (
 - Adapters present under `internal/adapters/`: `youtube` (captions-first `VideoSource`,
   timedtext/InnerTube acquisition per ADR-010), `summarizer` + `llm` (the copied AI router,
   now a resilient endpoint chain — local primary → local fallback → external worst-case,
-  parse-failure-aware, ADR-004 + ADR-022), `store` (Postgres, golang-migrate migrations 001–015),
+  parse-failure-aware, ADR-004 + ADR-022), `store` (Postgres, golang-migrate migrations 001–019),
   `secrets` (file-backed `SecretStore`). The brain HTTP sink (ADR-005) is the remaining
   optional sink.
 - Stage 1 is open (ADR-012): multi-user with **DB-enforced** isolation — Postgres RLS `FORCE`d
@@ -108,6 +112,28 @@ The repo is **green and shipping** — last tag `v0.15.0`. `task check` passes (
   (`TestTranscriptsTableIsSharedNotRLS`). The engine reads stored transcripts before any caption
   fetch (`usecase.resolveTranscript`), so re-analysis — re-summarize, paste-a-URL, onboarding
   burst — never re-touches YouTube. Per-user summaries/videos stay RLS-scoped.
+- Audio upload (ADR-033, migration 017): `internal/uploads` (validation by **probing**, intake,
+  one-at-a-time in-process worker), `internal/adapters/audio` (ffprobe/ffmpeg), and
+  `internal/adapters/transcribe` (gateway client + iguana→berget chain). Transcripts are private
+  in `uploaded_transcripts`; the audio is never retained; chat reuses `chat.Service`. The
+  isolation test pins the **exact set** of non-RLS tables (`TestRowSecurityOffTablesAreExactlyTheKnownSet`).
+  The runtime image does not yet contain ffmpeg — that is a pending decision recorded in ADR-033.
+  Two-engine reconciliation (ADR-034, migration 019, off unless `TAPIR_RECONCILE_BASE_MODEL`):
+  `internal/reconcile` (word diff, disputes, a local-model adjudicator) and
+  `transcribe.Ensemble`; every model must be a local alias, enforced at start-up.
+  Progress and logging (ADR-035, migration 020): `internal/progress.Tracker` is the ONE writer of a
+  job's phases (saved to the row for the page, logged by the same call); engines and the reconciler
+  report through `progress.From(ctx)`, so `uploads.Transcriber` is unchanged, and a transcriber may
+  implement `uploads.Planner` to declare its own phases. The page tells running / quiet / lost from
+  two timestamps (`alive_at` heartbeat vs `progress_at`) and never shows a percentage that does not
+  come from real counts. **Nothing logged may carry transcript text, words from it, or the spool path.**
+  Errors reach a log, a row or a page as `errkind.Of(err)` (a kind such as `timeout` or `http_502`),
+  never `err.Error()`: error text can hold a provider's response body, the spool path or transcript
+  words (ADR-035 decision 8). Local-only for anything that carries transcript text: `config.IsLocalModel`.
+  `TAPIR_SHOT_DIR=… go test -run TestRenderUploadProgressStatesForReview ./internal/web` renders every
+  state for a screenshot review.
+  **Lint caveat:** `task check` does not fail on lint (#41: the Taskfile `|| echo` swallows
+  findings) — run `golangci-lint run ./...` yourself until that is fixed.
 - `cmd/tapir` subcommands: `list`, `show`, `auth` (interactive host-side OAuth), `run` (batch
   watch→summarize), `serve` (the HTMX+Templ web reader/writer under `internal/web`, a new
   transport over the unchanged engine/ports — ADR-003). `tapir env` prints config.

DECISIONS.md

@@ -123,7 +123,9 @@ secret surface (see VISION Stage 2); it gets the existing, vetted secret path.
 ## ADR-007 — Captions-first; audio-download + speech-to-text deferred
 
 **Status:** Accepted (2026-06-02). Acquisition mechanism superseded by ADR-010 (Data API
-`captions.download` → player/timedtext baseUrl); captions-first stance and STT deferral stand.
+`captions.download` → player/timedtext baseUrl); captions-first stance and STT deferral stand
+**for YouTube**. STT deferral superseded **for user-supplied audio only** by ADR-033; the
+YouTube audio-download rejection is unchanged.
 
 **Context.** YouTube's Data API does not expose transcripts. Options: official captions
 (clean, limited coverage) vs. audio download + local Whisper (broad coverage, ToS-grey,
@@ -1128,6 +1130,700 @@ Spec: `docs/specs/onboarding-wow-burst.md`.
 
 ---
 
+## ADR-029 — Stateless session cookie (survives restarts, browser-close, idle)
+
+**Status:** Accepted (2026-06-11). Triggered by pilot feedback: "lots of clicking to log in again
+on iPhone." Supersedes the in-memory session store in the ADR-011 login.
+
+**Context.** Three compounding causes made users re-login constantly:
+1. **In-memory session store** (`sessionStore` map) — wiped on every pod restart, so each deploy
+   logged everyone out. During the active build period that was ~15 logouts.
+2. **1-hour session TTL** — for a "check back tomorrow" reader, idle > 1h forced a re-login on
+   nearly every visit.
+3. **No cookie Max-Age** — a session cookie (deleted on browser/app close); iPhone Safari closing
+   the tab dropped it.
+Each re-login is the full Dex/Authentik redirect dance — many taps on mobile.
+
+**Decision.** Make the session **stateless**: the identity (subject + email) and an absolute
+expiry live INSIDE the existing HMAC-signed (HS256) cookie — no server-side table. Plus:
+- **30-day sliding TTL** (was 1h), re-signed on each request so an active user never lapses.
+- **Persistent cookie** (`Max-Age` set) so it survives browser/app close.
+The cookie is HttpOnly + Secure + SameSite=Lax; the HMAC (keyed by the stable ESO
+`tapir-session-secret`, which does NOT rotate per deploy) makes it tamper-proof. The payload is
+identity, not secrets — the OIDC access/ID tokens are still discarded after callback.
+
+**Consequences.** A deploy/restart no longer logs anyone out (proven by a test: a cookie issued by
+one instance is accepted by a fresh instance with the same secret); works across replicas for
+free. **Trade:** no server-side revocation — `logout` clears the cookie client-side, but a copied
+cookie stays valid until expiry. Accepted for the Stage-0 reader pilot; revisit (server-side
+revocation list, or shorter TTL + refresh) if it ever holds sensitive actions. Rotating
+`tapir-session-secret` invalidates all sessions — the global logout lever.
+
+**Not addressed here:** the tap-count of the IdP login page itself is Authentik's UX; with
+re-login now rare (30-day idle or explicit logout), it matters far less.
+
+---
+
+## ADR-030 — Observability: slog timing + Prometheus metrics (AI-focused)
+
+**Status:** Proposed (2026-06-11). Issue #15. **Draft for review — no code yet.**
+
+**Context / requirements.** Nothing measures the activities that drive Tapir's performance and
+UX, and the Stage-0 eval gate (ADR-016) needs a *performance* dimension to sit beside the
+return-usage one. We need timing for: caption fetches (the scarce op), summarization (which model
+won, how long, fallbacks), Q&A latency, LLM token spend, and basic session/usage (request rate,
+latency by route, logins). Requirements:
+- R1: structured `slog` timing at each AI call site (human-readable, already the logging stack).
+- R2: Prometheus metrics for the same, scrapeable by the cluster's prometheus-operator.
+- R3: **AI metrics are the priority** — summarize latency by `model`/`outcome`/`fallback`,
+  caption-fetch latency by `outcome`, chat latency by `model`, and LLM `tokens` by model+kind.
+- R4: HTTP/session metrics via middleware — request count + latency by route, logins.
+- R5: bounded label cardinality (no per-user, no raw-path labels).
+- R6: `/metrics` must NOT be publicly exposed.
+
+**Decision / architecture.**
+1. **New package `internal/metrics`** owns all Prometheus collectors + a typed API
+   (`ObserveSummarize`, `ObserveCaptionFetch`, `ObserveChat`, `RecordTokens`, `IncLogin`,
+   `HTTPMiddleware`, `Handler`). Adapters call this API; they never import prometheus types.
+2. **New dependency `github.com/prometheus/client_golang`.** Justification: it is *the* standard
+   Go Prometheus client and the cluster already runs prometheus-operator; hand-rolling exposition
+   is not worth it. (Needs the dep-justification note in the commit per repo rules.)
+3. **The copied `llm` package stays stdlib-only (ADR-004).** It must not import `internal/metrics`.
+   Token usage is surfaced via an **optional callback** `llm.WithUsageHook(func(model string, prompt, completion int))`
+   set at wiring time (`buildSummarizer`/`buildChat`) to `metrics.RecordTokens`; `llm.Client` only
+   gains parsing of the response `usage` block. Our own adapters (`summarizer`, `youtube`, `chat`)
+   may import `internal/metrics` directly.
+4. **HTTP middleware** reads `r.Pattern` AFTER routing (Go 1.22 sets it during ServeMux match), so
+   the `route` label is the bounded registered pattern (`GET /v/{videoId}`), satisfying R5;
+   unmatched → `other`.
+5. **Dedicated metrics port** (`TAPIR_METRICS_ADDR`, default `:9090`) served by a second
+   `http.Server` in `cmdServe`; `/metrics` is never on the public app mux (R6). A **PodMonitor**
+   in `mathias/infra` scrapes it; the deployment exposes the port.
+6. **slog** elapsed fields are emitted alongside each metric at the call sites (R1).
+
+**Hook points (where the instrumentation lands).**
+- `summarizer.Summarize` — per-endpoint timing + outcome (`success`/`parse_error`/`error`) + fallback flag.
+- `youtube.FetchTranscript` — fetch timing + outcome from `domain.Transcript.Source`.
+- `chat.Service` answer — timing by model.
+- `llm.Client.Complete` — parse `usage`, fire the usage hook.
+- `oidc.handleCallback` — `IncLogin`.
+- `cmdServe` — wrap `Router()` in `metrics.HTTPMiddleware`; start the metrics server.
+
+**Out of scope / later.** Persisting per-summary latency into Postgres for `tapir report`
+(derive UX latency — publish/discovery → summary — from existing timestamps first; only persist
+op-latency if the scrape proves insufficient). SPA view (#16) and visual refresh (#17).
+
+**Reversibility.** Additive: a new package + middleware + a metrics port. Removing the PodMonitor
+stops scraping; the app is unaffected. No schema change.
+
+**Next steps (gated):** on approval of this ADR → BDD scenarios (`docs/use-cases/observability.feature`
++ scenario-coverage map) → TDD → implement → SemVer + docs + PodMonitor.
+
+---
+
+## ADR-031 — SPA-like reader: inline-expand summary + Q&A in the list (HTMX, no framework)
+
+**Status:** Proposed (2026-06-12). Issue #16. **Draft for review — no code yet.**
+
+**Context / requirements.** The reader is multi-page: a list of compact cards (`/`), then a
+navigation to a separate detail page (`/v/{id}`) for the full summary + the docked chat (ADR-027).
+It feels less fluid than a single integrated view. We want the full summary AND the per-video
+Q&A to open **in place in the list**, no page hop. Requirements:
+- R1: clicking a summarized card expands it in place to the full summary (summary/highlights/
+  takeaways) + the chat dock; a collapse returns it to the compact card.
+- R2: **no SPA framework** — stay HTMX + Templ (ADR-003); reuse existing fragments, not a rewrite.
+- R3: **progressive enhancement** — with JS off, the card link still navigates to `/v/{id}`
+  (the detail page stays as the no-JS + deep-link surface). Nothing becomes JS-only.
+- R4: only **summarized** cards expand; pending/rate-limited/no-caption cards keep their current
+  footer behaviour (Summarize button, waiting/none states).
+- R5: chat inside an expanded card works exactly as on the detail page (reuse `chatReveal`/
+  `chatSection` + the existing `/v/{id}/chat` endpoints, unchanged).
+
+**Decision / architecture.**
+1. **Reuse the existing fragments.** `summaryBody(r)` and `chatReveal(videoID)` already exist and
+   render the detail page; a new `expandedCard(r, chatEnabled)` composes the compact header + a
+   collapse control + `summaryBody` + `chatReveal`. `DetailPage` is refactored to also compose
+   `summaryBody` so the two never drift (DRY).
+2. **Two fragment endpoints** (mirroring the existing list/status HTMX fragment pattern):
+   `GET /v/{videoId}/expand` → `expandedCard`; collapse reuses the existing compact `VideoCard`
+   via `GET /v/{videoId}/card`. Both are list-card `<li>` fragments with the SAME `id`
+   (`video-{id}`), swapped `outerHTML` — same mechanism as `processingCard`/`VideoCard` today.
+3. **The compact card's title/"Read" affordance** becomes `hx-get=/v/{id}/expand`,
+   `hx-target=#video-{id}`, `hx-swap=outerHTML`, with `href=/v/{id}` as the no-JS fallback (R3).
+   The expanded card's collapse control is the inverse (`hx-get=/v/{id}/card`).
+4. **Only when `r.Summarized`** does the expand affordance render (R4); the other states are
+   unchanged.
+5. **v1 does NOT push the URL** (`hx-push-url`) — expand/collapse is ephemeral list UI state; the
+   detail page remains the deep-link/shareable URL. Deep-linking the open state via `hx-push-url`
+   is noted as a later option (needs list-state restore on back).
+
+**Out of scope / later.** URL push / deep-linkable open state; the visual refresh (#17) — though
+the expanded-card markup is where #17's TUI/charm styling will land, so they pair.
+
+**Reversibility.** Additive: two fragment endpoints + one templ + an affordance swap on the
+compact card. Removing the affordance reverts to plain list→detail navigation; the detail page is
+untouched. No schema change.
+
+**Next steps (gated):** on approval → BDD (`docs/use-cases/inline_expand.feature` + coverage
+map) → TDD → implement → SemVer + docs.
+
+---
+
+## ADR-032 — Visual refresh: one charm-reader layout, light + dark themes
+
+**Status:** Proposed (2026-06-12). Issue #17. **Draft for review — no code yet.** Follows the
+sketch-first explore step (3 throwaway mockups in `docs/sketches/`, screenshotted for review).
+
+**Context / decision.** The UI is flat. From the mockups, directions **B (light reader + charm)**
+and **C (dark cozy terminal)** are the SAME layout — readable sans body, monospace meta, charm
+palette accents, lipgloss-style bordered cards — in two palettes. Direction A (full-monospace TUI)
+is dropped as too heavy to read long summaries. Decision: ship that one layout with **both a light
+theme (B) and a dark theme (C)**, user-toggleable, defaulting to the OS preference.
+
+**Requirements.**
+- R1: one set of markup/structure; the two themes are pure palette (CSS variables), no duplicate templates.
+- R2: a **theme toggle** persisted across visits; default to `prefers-color-scheme` when no choice stored.
+- R3: charm language in both — mint/purple/pink accents, mono meta + section labels, lipgloss
+  bordered/gradient cards, the (fixed) ASCII tapir; readable sans body.
+- R4: style the existing pieces — list, compact card, **expanded card incl. the already-present
+  video embed (ADR-031/summaryBody)**, detail page, chat dock, the queue note, the charm spinner.
+- R5: WCAG-AA contrast for body text in BOTH themes; keep `prefers-reduced-motion` (already honoured).
+- R6: stay HTMX+Templ; no CSS framework.
+
+**Architecture.**
+1. **Palette as CSS variables.** `:root` holds the light (B) tokens; `:root[data-theme="dark"]`
+   holds the dark (C) tokens; a `prefers-color-scheme: dark` media block sets the dark tokens when
+   no explicit `data-theme` is set. All component CSS references variables only (R1). The existing
+   `CharmMint/Purple/Pink/Cream/Dim` Go consts remain the source for the spinner's inline colours.
+2. **Theme toggle** = a small inline script (a dozen lines, no framework) in `Layout`: on load,
+   apply stored theme (localStorage) or fall through to the media query; a header toggle button
+   flips `data-theme` on `<html>` and stores it. This is the one new bit of JS; everything else
+   stays server-rendered + HTMX. (Considered: cookie + server-render — rejected, a full round-trip
+   per toggle is clunky for a pure presentation flip.)
+3. **Scope = the `styleTag` CSS in `view.go`** (the single style source) plus tiny class hooks in
+   the templ where needed; content/structure are unchanged, so existing view tests keep passing.
+
+**Verification.** The deployed UI is behind auth (web-shot can't log in), so visual review is via
+the mockups now + a styled full-set mockup screenshot before merge, then a live eyeball on device.
+Automated tests stay structural/behavioural (theme tokens present, toggle persists, expanded card
+embeds the video, no-JS still renders a readable default) — colours are not unit-tested.
+
+**Reversibility.** A CSS theme swap + one small script + a few class hooks; revert `styleTag` to
+roll back. No schema, no structural change.
+
+**Next steps (gated):** on approval → BDD (`docs/use-cases/visual_theme.feature` + coverage map)
+→ TDD → implement → SemVer + docs.
+
+---
+
+## ADR-033 — Upload audio, get a private Swedish transcript, discuss it (user-supplied STT)
+
+**Status:** Proposed (2026-10-01). **Decider: Mathias — pending.** Issue #33 (audio-only subset of
+#28). **Supersedes ADR-007's speech-to-text deferral for user-supplied audio only.** The
+YouTube audio-download rejection in ADR-007 **stands untouched**. Builds on ADR-012 (isolation),
+ADR-018 (single replica), ADR-021 (**the constraint**), ADR-022 (endpoint chain), ADR-025
+(honest status), ADR-027 (chat). ADR-014 item 4 ("Whisper reconsideration is a future ADR,
+gated on caption data") is a different question — whether to transcribe YouTube videos with no
+captions — and this ADR does not answer it. Measurements: `docs/research/audio-length-ceiling-2026-10-01.md`.
+
+**Context.** The maintainer has a concrete need: a Swedish Teams voice-only export, 30–90
+minutes, to upload and then *discuss* (what was decided, what is open, who owes what). Every
+piece exists — a Whisper endpoint on the gateway (infra#274), chat-with-transcript (ADR-027) —
+and nothing connects them.
+
+ADR-007 deferred STT for three reasons: yt-dlp fragility, ToS-grey scraping, and GPU contention
+with the JEPA PoC. Upload-STT shares **none** of them: the user owns the file, nothing is
+scraped, there is no caption rate gate, and the contention moved from koala to iguana. What it
+does share is landing Whisper in the codebase, which ADR-007 deferred — so a new ADR is required,
+and ADR-017's process note records the one time a guardrail-reversing change shipped without one.
+
+**The constraint that shapes everything.** ADR-021 made `transcripts` a shared, **non-RLS**
+table on the explicit grounds that it holds only public caption content and nothing
+user-identifying. A recorded internal meeting is the opposite: private third-party content owned
+by one user. The existing isolation test asserts `transcripts` *is* non-RLS, so a naive reuse
+would leak across tenants **and pass CI green**.
+
+**What was measured, not argued** (research doc above):
+- **No transcription ceiling up to 90 minutes.** One 174 MB request through the LiteLLM NodePort
+  took 440 s (~12× realtime). The first failure is the public ingress `llm-api.d-ma.be`
+  returning `413` above roughly 8–10 MB (~5 min of WAV) — not whisper, and not a Metal OOM.
+  The 8:20 failure recorded in #33 is consistent with that cap.
+- **`response_format=srt` survives LiteLLM, wrapped** in a JSON envelope (`{"text": "<srt>"}`).
+- **The two backends differ.** `iguana/kb-whisper` returns whole-second segments;
+  `berget/kb-whisper` returns sub-second times and a different `language` token.
+- **The existing chat path truncates at 18,000 chars (~22 min of speech), keeping the head.**
+  #33 assumed chat "retrieves rather than generates" and so has no length ceiling. It does
+  not retrieve; it stuffs the transcript into the prompt and cuts the tail. Measured against
+  the offered models: `phi4-mini` 400s at ~13k tokens; `gemma4-26b` is right to 90,000 chars and
+  silently drops the head at 120,000.
+
+**Unknowns, stated.** Real-speech accuracy; compressed-input behaviour; concurrent long jobs on
+iguana; and the Metal OOM, deliberately not provoked. (Whether the offered chat models can hold
+an hour of Swedish *was* an unknown here; it is now measured, see decision 8.)
+
+**Decision.**
+
+1. **Audio is what is used; the transcript is the only thing kept.** An authenticated user
+   uploads a recording; Tapir transcribes its audio and stores the transcript. The audio itself
+   is deleted after transcription (temp file, removed in a `defer`, swept at start-up) — no
+   custody problem to solve, and nothing to hand to #31. **A file that also carries a video
+   track is accepted: its audio is extracted, the video is discarded, and the user is told so**
+   (see 4). *Amended 2026-10-01, the same day this ADR was drafted: the first draft rejected
+   video and named #28. The maintainer's first real recording — a Teams meeting with the camera
+   off — is still an `.mp4` with a video track, so rejecting video rejected the primary use
+   case.* This keeps no video and analyses none: playback, keyframes and what-is-shown analysis
+   remain #28. YouTube audio-download stays rejected. No public/anonymous upload (ADR-008).
+
+2. **Private store: a separate RLS-scoped table, `uploaded_transcripts`** (resolves #28's
+   decision 1). `user_id NOT NULL REFERENCES users ON DELETE CASCADE`, `ENABLE` + `FORCE ROW LEVEL
+   SECURITY`, the same `tapir.current_user_id` policy as every other user-owned table. It carries
+   the transcript text and its timestamped segments (re-transcribing 90 minutes costs ~7
+   minutes; discarding the timestamps is irreversible), `status`, the `endpoint` that served
+   it, `audio_seconds`, and a typed `error_kind`. **Not** a `user_id` column on `transcripts`
+   with a partial policy, even though that would reuse more code: ADR-021's exception is
+   precisely that *that table holds no private row*, the isolation test asserts it has no
+   `user_id`, and a partial policy would turn "single exception" into "single exception, with a
+   private lane inside it" — the shape under which a future query drops the policy clause and
+   leaks. Two tables keep both boundaries independently assertable.
+
+3. **The isolation test asserts both boundaries and the set.** (a) User A cannot read, update or
+   delete user B's uploaded transcript; no GUC set → deny-all. (b) `transcripts` is still
+   non-RLS with no `user_id`. (c) **The set of public-schema tables with row security off is
+   asserted exactly** — so a new table added without RLS fails CI instead of passing it. (c) is
+   new; the existing test only checks the named tables.
+
+4. **Validate by probing, reject clearly, and say what was ignored.** Format, duration and
+   what a file contains cannot be read from a filename or MIME type — a Teams recording is an
+   `.mp4` whether or not the camera was on. The upload is probed (`ffprobe`): **accept any
+   container with an audio stream, whether or not it also has a video stream** (an mp3's cover
+   art counts as neither). Whether a real video track was present is recorded on the row
+   (`had_video`) and the transcript page says so plainly — *this file contained a video track;
+   only its audio was used and the video was discarded* — so the user is never left to wonder
+   what happened to the picture. Reject a file with **no** audio stream (including a screen
+   recording with no microphone) and an unreadable one. Limits, all configurable, all rejected
+   synchronously with a specific message rather than by timeout: **200 MB** upload and
+   **90 minutes** duration. 90 minutes is *the longest length measured*, not a ceiling that was
+   found; raising it is a measurement, not an edit. The size cap now matters more: a video file
+   is much larger than its audio and is spooled whole before the audio is extracted, so a long
+   recording with screen sharing can exceed 200 MB while its audio is a few tens of MB.
+
+5. **Transcode server-side, always, to 16 kHz mono PCM WAV** — even when the input already is,
+   rather than assuming. whisper-server's `--convert` is rejected (its `--tmp-dir` is created at
+   process start and rots under macOS periodic cleanup). The transcoder sits behind a port; the
+   tests use a fake. **This puts `ffmpeg` and `ffprobe` in the runtime image**, which is
+   distroless-static today. **Provisional: copy static ffmpeg/ffprobe binaries into the existing
+   distroless image** (no shell, no package manager retained). This is a heavy new dependency
+   and is **the maintainer's call** — see Open items.
+
+6. **Transcription adapter via the gateway, never the public hostname.** A `Transcriber` port;
+   the implementation calls `POST {TAPIR_GATEWAY_URL}/audio/transcriptions` where the gateway
+   URL is the in-cluster NodePort (`koala:30401/v1`). Primary `iguana/kb-whisper`, fallback
+   `berget/kb-whisper` (immune to iguana contention, which can wedge permanently — infra#273).
+   - **Fallback triggers on any primary failure** (transport error, timeout, non-2xx). It never
+     branches on the body: `{"error":"failed to process audio"}` has had at least five distinct
+     causes.
+   - **`response_format=verbose_json`**, normalised to one segment shape: times clamped to the
+     audio duration (the final segment can end past it), one language token. Not `srt`: the
+     envelope adds a parse step and iguana's segment timing is the same either way.
+   - **No chunking in v1.** Nothing failed to 90 minutes (measured). Chunk-and-stitch is the
+     answer *if* the cap is raised or smaller requests become necessary — see triggers.
+   - **The serving endpoint is recorded per row**, and surfaced, because the two backends
+     differ in timestamp granularity: which one served a transcript is a quality fact.
+   - **Berget is a cloud endpoint.** On fallback the audio of a private meeting leaves the
+     homelab. #33 asks for this fallback; it is configurable (`TAPIR_TRANSCRIBE_FALLBACK_MODEL`,
+     empty disables), recorded per row, and called out in Open items rather than defaulted
+     silently.
+   - Swedish only (`language=sv`). Language detection and routing (#28 decision 2) are not
+     touched. **Amended 2026-10-04 (#61, interim):** the language is now a per-upload choice on
+     the form, `sv` (preselected, and what an upload that names none gets) or `auto` ("Other":
+     `language` is omitted so the engine detects it), a closed set held by the form, the request
+     code and a CHECK constraint on `uploaded_transcripts.language` (migration 021). Automatic
+     detection and a general-purpose engine for other languages are still not built (#61); an
+     `auto` upload goes to the same configured engine, which is not tuned for it.
+   - `-H 'Expect:'`, multipart field `file` — encoded in the adapter and its tests, not rediscovered.
+
+7. **Jobs run in-process, one at a time (ADR-018 stands).** Transcription is minutes (~7.5 min
+   for 90) but a single goroutine behind a one-slot semaphore: two simultaneous 90-minute jobs on
+   iguana were not measured, so concurrency is 1 until they are. No worker, no new deployable,
+   no second replica. Honest status per ADR-025: `queued → transcribing → done | failed`, polled
+   by the same `/status` pattern, every terminal state a rendered card, never a stuck spinner.
+   Audio is not retained, so a restart mid-job **cannot resume**: at start-up any
+   `queued`/`transcribing` row is marked `failed` with `error_kind = interrupted`, and the user
+   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,
+   switchable models, trust-the-model) carry over unchanged. **Two things must differ, and both
+   were measured** (`docs/research/chat-context-over-long-transcripts-2026-10-01.md`):
+   - **The transcript budget.** The 18,000-char caption default cuts a one-hour meeting to its
+     first third. Uploaded transcripts take their own `TAPIR_MAX_UPLOADED_TRANSCRIPT_CHARS`,
+     default **90,000** — the measured ceiling of the best offered model (`iguana/gemma4-26b`),
+     which answers correctly there but **silently loses the head of a 120,000-char transcript**
+     with no error and no truncation flag. (An earlier draft of this ADR proposed 120,000 as a
+     guess; the measurement refuted it.) The mechanism is a per-request `chat.Request.MaxChars`.
+   - **The opening model.** The offered list leads with the summarizer default `koala/phi4-mini`,
+     whose 8,192-token context returns HTTP 400 on an hour of speech (~13,000 tokens). Chat over
+     an upload opens on `TAPIR_UPLOAD_CHAT_MODEL` (default `iguana/gemma4-26b`); the switcher
+     lists **local models only** (amended by ADR-035 after the adversarial review found it listed
+     the cloud model too: the whole private transcript goes to the one chosen), and picking
+     `phi4-mini` on a long recording fails loudly and inline. A per-model budget is not built.
+   Past ~90,000 chars (≈100 minutes) no offered model can answer by stuffing; see triggers.
+
+9. **Not decided here.** Rich `AVSNITT`/`ALTERNATIV` analysis (gated on #30 — `analyze_srt.py`
+   hardcodes `max_tokens: 6000`); video, keyframes, synced playback (#28); language detection;
+   per-transcript deletion UX (account deletion cascades; see Open items); retention of audio.
+
+**On the sequencing question #28 asks of itself** — *is this the thing that best serves the
+Stage-0 gate, or the most interesting thing to build?* Recorded answer: it proceeds **ahead of
+#32** (the demand test) as the maintainer's explicit, knowing call (#28, comment of 2026-10-01),
+on a concrete need rather than speculation. It is not a return-nudge (ADR-020/027): it adds
+nothing that prompts a user to come back, so it does not contaminate the unprompted-return signal
+ADR-016 measures. The honest risk is the one ADR-016 names: it is larger than anything ADR-020
+deferred. The scope cut — chat first, no rich analysis, no video — is what keeps it bounded.
+
+**Consequences.**
+- **Positive.** The user's request is met end to end with existing parts. ADR-021's exception stays
+  singular and now provably so. The measured ceiling removes chunking from v1. No audio is stored.
+- **Negative / watch.**
+  - Tapir's runtime image gains `ffmpeg`: a larger image and a new CVE surface to patch.
+  - iguana becomes load-bearing for a user-facing feature (#28 decision 3; infra#273 is still the
+    dependency that matters) — a Metal OOM now means *every upload fails over to berget*, i.e.
+    private audio leaving the homelab, silently unless the per-row endpoint is surfaced.
+  - A 90-minute job holds the single slot for ~7.5 minutes; a second user's upload queues behind it.
+  - Chat over a long transcript is bounded by model context, not by anything in Tapir, and the
+    failure at the edge is *silent* for gemma (the head is dropped, no error). The 90,000 default
+    sits just inside the measured limit and is only as good as that single measurement.
+  - **The default chat model for uploads is a large ollama model on iguana — the same host as
+    `iguana/kb-whisper`.** infra#273 is exactly that a large resident ollama model can Metal-OOM
+    whisper-server permanently. Chatting with a transcript can therefore load the model that
+    wedges the transcriber. Not solved here; the berget fallback limits the blast radius.
+  - Tapir's own ingress body limit for a 200 MB upload is **unverified** (it lives in `infra`).
+  - Accepting video means large files are spooled before their audio is extracted, and ffmpeg
+    demuxes (but, with `-vn`, does not decode) video it then throws away. A client-side audio
+    extraction would avoid the upload cost; it is not built.
+
+**Validation / resolution triggers** (each observable by someone other than the author):
+- *Chunking*: build it when the duration cap is raised above 90 min, or when a measured
+  request at the current cap fails with whisper-side (not ingress) errors.
+- *Berget-first*: if more than a quarter of uploads in any week are served by the fallback,
+  swap the primary/fallback order and revisit infra#273.
+- *Retrieval chat*: build chunked retrieval when a recording's transcript exceeds the measured
+  90,000-char ceiling (≈100 minutes) — i.e. the moment the duration cap is raised above that, or a
+  real recording turns out denser than the fixture — or when the offered model list changes and
+  the ceiling is re-measured lower.
+- *Worker*: if a transcription job ever degrades the reading UI, or Tapir runs more than one
+  replica, ADR-018 is revisited.
+- *The decision itself*: if the maintainer has not used the feature unprompted within the
+  ADR-016 window, it was the interesting thing, not the useful one — remove it.
+
+**Open items this ADR creates.**
+- **DECIDED (Mathias, 2026-10-02): option A.** Static ffmpeg and ffprobe 7.0.2 are copied into the
+  distroless image, pinned by version and sha256 (Dockerfile in this PR; #34). Measured at that time:
+  **about 160 MB** added, not the 70 MB first estimated (80 MB per binary); the build is **GPLv3**;
+  the checksum proves the bytes reviewed, not the publisher; upstream's `releases/` URL moves on, so
+  vendoring the verified tarball is the durable fix (not done). Uploads stay off until the deployment
+  sets `TAPIR_UPLOADS_ENABLED`. *The question as it was put:* *Options:* (A) static ffmpeg/ffprobe copied
+  into distroless **(recommended — the only option that satisfies "accept m4a/mp3/opus/mp4,
+  validate duration, reject video"; pure-Go decoders cover wav/mp3 only)**; (B) accept WAV
+  only, convert client-side (fails "I just want to upload it"); (C) a transcode sidecar
+  (second deployable, ADR-018's reasoning against). *Blocks:* the Dockerfile slice, not the
+  tests, which use the port.
+- **DECIDED (Mathias, 2026-10-02): option A, enabled by default (#35).** Scope: the audio
+  transcription fallback only. Reconciliation text and chat over a recording never go to a cloud
+  model (ADR-034, ADR-035). *The question as it was put:* *Options:* (A) enabled by default as
+  #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`).~~ **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`
+and ADR-021 are untouched.
+
+---
+
+## ADR-034 — Transcribe with two engines and let a local model reconcile their disagreements
+
+**Status:** Proposed (2026-10-01). **Decider: Mathias — pending.** Issue #40 (decision B) and #33.
+Builds on ADR-033 (private upload transcripts), ADR-022 (endpoint chain, a pattern reused), ADR-018
+(single replica), ADR-025 (honest status). Evidence:
+`docs/research/asr-kb-whisper-vs-pianissimo-2026-10-01.md` and
+`docs/research/local-model-adjudicator-sweep-2026-10-01.md`.
+
+**Context.** On the first real one-hour recording `iguana/kb-whisper` and `iguana/pianissimo` were
+a near tie on accuracy (the speaker's blind verdicts on the 30 meaning-changing disagreements:
+pianissimo 17, kb-whisper 13) and they **err in different places**: pianissimo is better on terms and
+numbers, kb-whisper on ordinary words; kb-whisper's errors garble, pianissimo's read as plausible.
+Pianissimo is ~6x faster (40 s vs 4m08s for 63.7 min) and runs through the same adapter. Two engines
+that disagree in ~380 places per hour, with a ~92% word agreement, is information a single-engine
+pipeline throws away. Separately, eleven local models were scored on adjudicating those disagreements
+against the speaker's answers: **`gemma4-31b` 97% and `gemma4-26b` 92%** passed the pre-set bar
+(>= 80% accuracy, >= 85% order-swap consistency, <= 10% unparseable), nothing else did.
+
+**Unknowns, stated.** The adjudicator evidence is 30 items from one speaker and one domain, chosen by
+Opus as decidable; the other ~350 disagreements (fillers, compounds, small words) have no ground
+truth. Whether a *merged* transcript is better than the better single engine has not been measured.
+Reconcile time per hour is estimated, not measured. `gemma4-31b`'s slowness was Low Power Mode on
+iguana: with it off the 31B decodes at 22.3 tok/s, not 7.9 (2026-10-02, same-machine before/after on
+a different synthetic prompt; see the research note). The reconciler default `gemma4-26b` was chosen
+on the slow numbers and is reopened: 31B scored 97% against 92%.
+
+**Decision.**
+
+1. **Two engines, one WAV, both local, one after the other.** Each upload is transcribed by
+   `iguana/pianissimo` and `iguana/kb-whisper` from the same converted WAV, sequentially (concurrency
+   stays 1, ADR-033). Each engine fails independently; the berget fallback (ADR-033) applies only when
+   **both** fail. Off by default: with no second engine configured nothing changes.
+
+2. **Align by word, with pianissimo as the base text.** Both texts are tokenised (lower case, letters
+   and digits only) and diffed word by word (Myers); neighbouring differences separated by a single
+   equal word are merged into one **dispute**. The base is pianissimo: sentence-level prose with
+   punctuation and millisecond timestamps, where kb-whisper's output is wrapped fragments. **This
+   choice is about readability and timing, not accuracy — accuracy was a tie — so it is the cheaper
+   mistake, not a measured winner.** A replaced span takes the alternative's own words.
+
+3. **Resolve the trivial ones without a model.** A dispute that is identical once spaces are removed
+   (a compound split or joined) keeps the base; so does a dispute that is only a short filler word
+   on one side. Everything else is adjudicated. (Opus's classification of 380 disputes found ~46
+   compound and ~26 punctuation-only residue, ~65 filler.)
+
+4. **A local model adjudicates the rest, bounded.** Each dispute goes to a local chat model (default
+   `iguana/gemma4-26b`; `gemma4-31b` is configurable and scored higher but is ~5x slower) with ~100
+   words of context either side and the two readings, in the wording that was measured. The reading
+   order alternates by dispute so position bias averages out. The verdict is *base*, *alternative* or
+   *unsure*; **unsure keeps the base and is flagged**. An optional order-check mode asks twice with the
+   order swapped and treats disagreement as unsure; it is **off by default** because it doubles the time
+   and the measured consistency of the passing models is 93%.
+   - **The model must be local.** Its alias must begin `koala/` or `iguana/`; a cloud alias is a
+     start-up configuration error. Private meeting content never goes to a cloud model for this step.
+   - **Bounded impact.** The reconciler can only choose between two readings the engines already
+     produced; it cannot write new text. Speech that says "answer 2" can at worst flip one dispute.
+
+5. **Reconciliation never fails a job.** One engine fails → its text is used, recorded as
+   single-engine. The adjudicator is down (three consecutive errors) → the base text is used
+   unreconciled, recorded as skipped with the reason, and the remaining disputes are not retried one
+   timeout at a time. Both engines fail → ADR-033's fallback path.
+
+6. **What is stored and shown.** `content` is the merged text. `segments` stay the base engine's timed
+   segments, untouched (**amended by ADR-035: they now carry the merged words, timing stays the
+   base's**): `content` is authoritative for chat, `segments` only for time. `endpoint`
+   lists every contributor and the reconciler. A new `reconciliation` column (migration 019, same
+   RLS-scoped table) records the status, counts, and the **uncertain and changed spots with a minute
+   mark and both readings**, so the user can check them against the audio. The transcript page says in
+   plain words what was done — "reconciled from two transcripts: N differences, K uncertain" — and
+   lists the uncertain spots. A reconciled transcript is never presented as simply "the" transcript.
+
+7. **Configuration.** `TAPIR_RECONCILE_BASE_MODEL` (empty = off; the engine whose text is the base,
+   e.g. `iguana/pianissimo` — `TAPIR_TRANSCRIBE_MODEL` stays the other engine, e.g. kb-whisper),
+   `TAPIR_RECONCILE_MODEL` (the adjudicator; required when a base engine is set),
+   `TAPIR_RECONCILE_CHECK_ORDER` (off). All three models must be local aliases (`koala/` or `iguana/`):
+   a cloud alias is a start-up error, and the base engine must differ from the other one. The default
+   job timeout rises from 30 to 60 minutes: ~4 min of kb-whisper and ~40 s of pianissimo per hour of
+   audio, plus the adjudication.
+
+   **Amended 2026-10-04 (#61, interim):** reconcile is for Swedish. A recording whose language is
+   `auto` is never sent to the base engine (it is Swedish-only) and is not reconciled: the other
+   engine transcribes alone, then the cloud fallback if that fails and the job is live, and the
+   page does not claim a second opinion.
+8. **Not decided here.** A cloud adjudicator (rejected: private content); tuning the prompt; making
+   `gemma4-31b` the default; a "reconciling" progress state (the page says "transcribing" for the whole
+   job); feeding the uncertain spots to chat; whether pianissimo should also be a middle tier in the
+   fallback chain (#40, option A).
+
+**Consequences.**
+- **Positive.** Uses the finding that the engines err differently; stays local end to end; degrades
+  to today's behaviour at every failure; and is honest on the page about what it did.
+- **Negative / watch.**
+  - Job time grows: roughly 300 adjudicated disputes x ~2 s on `gemma4-26b` ~ 10 minutes per hour of
+    audio — an estimate, not a measurement; double with order-check on, ~5x with `gemma4-31b`.
+  - Two gemma-class calls sit on iguana beside whisper. During the sweep a 17-20 GB model was resident
+    beside whisper-server without wedging it, far below the 52 GB that did (infra#273); the hazard is
+    reduced, not removed.
+  - A merged transcript can be wrong where both engines are wrong, and where the adjudicator picks
+    the wrong one (it was right 92% on a hand-picked set, ~97% for 31b).
+  - Reading the page takes more words: the reconciliation summary is new surface to keep honest.
+
+**First live smoke (2026-10-01, synthetic audio, 3 disputes):** all 3 were changed to the
+alternative's reading; against the known script 2 were right and **1 was wrong** (a name replaced by a
+pronoun, a correct base reading lost). That is one run of three, not a measurement, but it is the failure
+this design must expect: the reconciler can overwrite a correct reading, which is why changed spots, not
+only uncertain ones, are listed on the page, and why the decision below is a blind check on a real
+recording.
+
+**Validation / resolution triggers.**
+- **The test that decides it:** re-upload the same real recording after merge; the speaker judges, blind,
+  ~30 disputes where the reconciled text differs from pianissimo alone. Keep the feature only if the
+  reconciled reading is right at least 85% of the time **and** no worse than the better single engine.
+  If not, switch it off (config) and keep the single-engine path.
+  **Decision (Mathias, 2026-10-02), on the first result below: keep it as an opt-in feature**, reading
+  the bar as 95% of decided items, with the second condition to be answered on the next real run.
+  For deployment the reconciler is `iguana/gemma4-31b` (97% against the speaker's answers, about 22
+  minutes for the 63.6-minute recording with Low Power Mode off). A cloud reconciler was raised and is
+  **not adopted**: it would send private transcript text to a cloud API on every upload; recorded as
+  an evaluation question in #45.
+  **First result, 2026-10-02, before merge** (`docs/research/reconcile-blind-check-2026-10-02.md`): 30
+  random replacement spots on the real 63.6-minute recording, judged blind by Opus from text (not by the
+  speaker, who delegated it): 21 right, 1 worse than pianissimo, 3 neither reading right, 5 undecidable.
+  That is 95% of decided items but 70% of all 30; the bar did not say how ties count, so it is **met on
+  one reading and missed on the other**, and the speaker has not ruled. The second condition, no worse
+  than the better single engine, is **not answered**: kept-base spots are not stored, so kb-whisper alone
+  cannot be compared. Left `Proposed`; the feature stays off by default.
+- Reconcile time above 25 minutes per hour of audio: drop to a smaller adjudicator or skip more
+  disputes automatically.
+- A different recording or speaker in which the engines agree far less (or far more): re-measure before
+  trusting the 92%/97% figures.
+
+**Reversibility.** Off by default and additive: the second engine, the reconciler and the column are
+unused when the settings are empty; the down-migration drops the column; the single-engine path is
+unchanged.
+
+---
+
+## ADR-035 — Show an upload's progress as real phases, log the same transitions, and keep every judged spot
+
+**Status:** Proposed (2026-10-02). Decider: Mathias (pending). Amends ADR-034 decision 6 (segments).
+Closes #42, #43 and #44.
+
+**Context (measured, not assumed).** The first real run (a 63.6-minute recording, 21 min 49 s) exposed
+three gaps. (1) The page said only `queued`/`transcribing`/`done`; during 22 minutes nothing told the
+user whether the job was working, hung or dead, and after a crash the row would say `transcribing`
+until the next start. (2) The server log had no line about the job between start-up and the final
+line, so an operator saw it only from outside (spool files, iguana CPU). (3) Evaluating the result
+showed that spots the model *kept* as the base are counted but not stored (so ADR-034's "no worse than
+the better single engine" cannot be measured after the fact, and the audio is not retained), insertions
+and deletions cannot be placed without context, and the saved `segments` still carry the base engine's
+wording while only `content` is merged.
+
+**Decision.**
+
+1. **Phases are data on the row.** A new `progress` JSONB column (migration 020, same RLS-scoped table)
+   holds an ordered list of steps `{key, label, state, started_at, finished_at, done, total}` with state
+   `pending|running|done|failed|skipped`. The plan is declared up front so the page shows what is coming:
+   received, preparing audio, then the transcriber's own steps (two-engine: engine A, engine B,
+   comparing, deciding n of N; single: transcribing), then saving.
+2. **One source for the page and the log.** An `internal/progress` Tracker is the only writer. Every
+   transition is saved to the row and logged by the same call, so the UI and the log cannot disagree.
+   Engines and the reconciler report through a context-carried `Reporter` (no-op when absent), so the
+   `Transcriber` interface does not change.
+3. **Alive is not the same as progressing, and the page says which it knows.** A heartbeat goroutine,
+   independent of any engine call, stamps `alive_at` every 10 s; `progress_at` moves only on a real
+   transition or advance. The page derives: *running*; *quiet* (the deciding step has not advanced for
+   5 minutes, so the model is probably waiting or slow); *lost* (`alive_at` older than 45 s, so the
+   service stopped reporting and may have restarted or crashed). The two engine calls are single blocking
+   requests, so inside them the page shows elapsed time only: **no percentage is ever shown that does not
+   come from real counts.** Only the deciding step has a bar.
+4. **The page.** A server-rendered step list in tapir's existing tokens and mono font, swapped in place
+   by the existing HTMX poll (every 2 s while unfinished, none once done or failed). Finished steps
+   collapse to one line with their duration; the running step has a CSS-only spinner that respects
+   `prefers-reduced-motion`; a failed step and a lost service are visibly different from running.
+   No new JavaScript, framework or stylesheet.
+5. **Logs.** One INFO line per phase transition and engine call (upload id, endpoint, elapsed), the
+   reconcile counts, `adjudicated n/N` at most once a minute, a WARN when an engine soft-fails or the
+   breaker trips, and a "still in phase X" line every 5 minutes. **Ids, counts, durations, endpoints and
+   error kinds only: never transcript text, disputed words or the filename**, asserted by a test.
+6. **Every judged spot is stored.** `reconciliation.contested` keeps each model-decided spot with its
+   verdict (changed, kept base, uncertain) and a few words of left context, bounded at 2,000. The page
+   still lists only changed and uncertain spots. This makes ADR-034's second condition measurable on the
+   next real run.
+7. **Segments carry the merged words.** The reconciled words replace the base's in the saved `segments`;
+   timing stays the base's. If the segments cannot be aligned to the merged text (a word wrapped across
+   a segment edge), they stay the base's and the result says so in the log. This replaces ADR-034
+   decision 6's "segments stay untouched".
+
+8. **Hardening from the adversarial review** (`skills:pr-gatekeeper`, head `7029dd1`: BLOCK; recorded on
+   PR #39). Each finding was reproduced by a failing test first.
+   - **Chat over a recording offers and accepts local models only** (it listed `berget/mistral-small`);
+     `TAPIR_UPLOAD_CHAT_MODEL` must be local at start-up; with no local model, nothing is sent.
+   - **A model alias counts as local only if it is one plain name after `koala/` or `iguana/`**
+     (`koala/../berget/x`, `koala//x`, `koala/x?host=…` were accepted by a bare prefix check).
+   - **An error reaches a log line, the row or the page as a kind** (`timeout`, `canceled`, `http_502`,
+     `network`, `error`), never its text, which can hold a provider's response body, the spool path or
+     a word from the transcript (`internal/errkind`). The stored reconcile reason and an engine's
+     failure reason are kinds; the page no longer shows provider text.
+   - **Every local engine call has its own deadline** (`TAPIR_TRANSCRIBE_ENGINE_TIMEOUT`, default 20
+     minutes: two wedged engines in sequence then leave 20 of the job's 60 minutes for a fallback),
+     so a wedged engine fails while the job lives and the fallback can run; the job's own
+     deadline or a cancel still never reaches the cloud.
+   - **The duration limit is checked against what is converted**, not the file's header: ffmpeg is
+     capped at the limit plus a minute, a longer result fails the job (`audio_too_long`), and a result
+     under 80% of the claimed length and more than two minutes short is **transcribed and noted** on
+     the audio step ("the audio is 20 min long; the file claimed 1 h 19 min"), not refused. The second
+     adversarial review showed a complete variable-bitrate MP3 whose header over-estimated its length
+     by 4x was wrongly refused by the first version of this check; a note keeps a truncated file
+     visible without turning away real recordings.
+   - **The queue is bounded**, 20 waiting in total and 3 per user, and its room is checked before the
+     upload is spooled (not after 200 MB and an ffprobe run); a full queue is a 429 with a typed reason,
+     never a stuck row or a server error.
+   - **The configured upload chat model is callable.** It is added to the shared chat service's model
+     list when uploads are on and it is local, because the service re-maps a model it does not offer
+     (a valid `iguana/gemma4-31b` was being silently swapped for `koala/phi4-mini`). The video chat's own switcher
+     lists that local model too while uploads are on.
+   - **A client that hangs up mid-upload is not an error**: the handler logs INFO and writes nothing
+     (rendering the error page ran a database read on a cancelled context and logged a false ERROR). The
+     400 "cut off" answer is keyed on a marker for errors from *reading the client's body*
+     (`uploads.ErrClientRead`), not on `io.ErrUnexpectedEOF`, which is also what a database drop looks like.
+   - **A note is not hidden**: a finished upload's record opens by default when any step carries a note.
+   - **Bad input is a 4xx**: a non-UUID id is a 404, a malformed or cut-off multipart body is a 400
+     whose text is not logged, and a title with control bytes, bidirectional controls or invalid UTF-8 is
+     cleaned (emoji joiners, Swedish letters and variation selectors are kept), not a 500.
+   - **The image was built and inspected** (buildah on koala, 2026-10-02): the first build FAILED
+     (`tar: xz: Cannot exec`, the golang image has no xz); fixed, rebuilt at 180 MB, ffmpeg and ffprobe
+     run as `nonroot` in distroless. CI does not build the image on a pull request
+     (`if: github.event_name != 'pull_request'`), so the first CI build of it is the merge to main.
+   Not done, filed as follow-ups: a vendored ffmpeg tarball, CSRF cookie review, ffprobe CPU and
+   stream-count limits, mov external references, `alive_at` clock skew, a hung database blocking the
+   tracker, an unbounded gateway response body, a user-facing notice for a recording that is partial but
+   inside the tolerance.
+
+**Alternatives rejected.** Server-sent events or a websocket: a long-lived connection through the
+ingress, with buffering to get wrong, for a page that changes every few seconds at best; polling is what
+the page already does. A percentage from the engine calls: there is no data behind it. A separate
+progress table: progress belongs to the row and inherits its RLS. Logging only through the UI: an
+operator has no page.
+
+**Consequences.** Positive: running, quiet, lost and failed are distinguishable by the user and the
+operator; the next blind check can answer ADR-034's second condition. Negative / watch: two more writes
+per transition and one per 10 s per running job (one job at a time, so negligible); a heartbeat proves
+the process is alive, not that an engine is making progress inside one request, which only the job
+timeout (60 min) bounds. The stored spot list grows to a few hundred small objects per hour of audio.
+
+**Validation.** A fake engine held open past the thresholds must change the page (quiet, lost); a log
+capture on a synthetic transcript must contain none of its words; a screenshot review of the running,
+quiet, lost, failed and done states against the existing pages (the visual claims tests cannot assert).
+Revisit if the heartbeat write ever shows in the database's load, or if a longer recording cap makes
+five-minute quiet detection too coarse.
+
+---
+
 ## Rejected alternatives
 
 Approaches considered during the 2026-06-02 planning + grill session and **deliberately not
@@ -1152,6 +1848,11 @@ maps to the ADR that settles it.
 | Reverse the Dex-write invite flow (Google OIDC only) | Some intended Future-B users won't use Google; OIDC-only leaves them with no onboarding path — invite flow is load-bearing | ADR-017 |
 | k8s CronJob for scheduled discovery (vs in-process) | At Future-B scale the in-process scheduler is simpler to deploy; CronJob's failure-isolation benefit was weighed and traded away knowingly (revisit if >1 replica or load grows) | ADR-018 |
 | Cached-transcript-first onboarding burst (instant, zero-fetch picks) | Live pilot DB: ~3% cross-user video overlap, 0 cached-and-unsummarized, a new user's newest-20 are 20/20 uncached — newest-first and cached-first are structurally incompatible. Empty lever at pilot scale | ADR-028 |
+| Uploaded transcripts as a `user_id` column + partial RLS policy on `transcripts` | Turns ADR-021's single exception into an exception with a private lane inside it; the isolation test asserts `transcripts` has no `user_id`. A separate RLS table keeps both boundaries independently assertable | ADR-033 |
+| whisper-server `--convert` for server-side transcode | Its `--tmp-dir` is created at process start and rots under macOS periodic cleanup | ADR-033 |
+| Calling the public `llm-api.d-ma.be` from in-cluster Tapir | Its ingress returns 413 above ~8–10 MB (~5 min of WAV); the NodePort has no such cap | ADR-033 |
+| A cloud model (e.g. Opus) as the adjudicator | It scored 30/30 but would send private meeting content to a cloud API; local gemma4-26b/31b passed the same bar | ADR-034 |
+| Pick one engine globally (pianissimo only, or kb-whisper only) | A near tie on accuracy with different error profiles; choosing one discards the information in their disagreement | ADR-034 |
 
 If a future case genuinely reopens one of these, that's a new ADR superseding the relevant one —
 not a silent reversal.

Dockerfile

@@ -16,12 +16,52 @@ COPY . .
 # no libc. Trim symbols/DWARF to shrink the layer.
 RUN CGO_ENABLED=0 GOOS=linux go build -trimpath -ldflags='-s -w' -o /out/tapir ./cmd/tapir
 
+# ── ffmpeg ──────────────────────────────────────────────────────────────────
+# Audio upload probes files with ffprobe and converts them with ffmpeg (ADR-033,
+# #34 option A). Static builds, so they run in the distroless image below.
+#
+# Pinned by version AND sha256: a changed upstream file fails the build instead of
+# shipping. The checksum proves these are the bytes that were reviewed (2026-10-02;
+# the publisher's md5 matched too). It does NOT prove authenticity: the publisher
+# signs nothing we can check. amd64 only (the cluster is one amd64 node).
+#
+# These builds are GPLv3 (--enable-gpl --enable-version3). The service is private and
+# the image is not distributed; the licence text is copied into the image anyway.
+#
+# The release URL moves on when upstream releases a newer version, so the pinned
+# version is tried in releases/ and then old-releases/. If neither has it the build
+# fails; vendoring the verified tarball is the durable fix.
+FROM golang:1.26 AS ffmpeg
+
+ARG FFMPEG_VERSION=7.0.2
+ARG FFMPEG_SHA256=abda8d77ce8309141f83ab8edf0596834087c52467f6badf376a6a2a4c87cf67
+
+# xz is not in the golang image (found by building it: `tar: xz: Cannot exec`).
+RUN set -eu; \
+    apt-get update; \
+    apt-get install -y --no-install-recommends xz-utils; \
+    rm -rf /var/lib/apt/lists/*; \
+    mkdir -p /out; \
+    for dir in releases old-releases; do \
+      curl -fsSL -o /out/ffmpeg.tar.xz "https://johnvansickle.com/ffmpeg/${dir}/ffmpeg-${FFMPEG_VERSION}-amd64-static.tar.xz" && break; \
+    done; \
+    echo "${FFMPEG_SHA256}  /out/ffmpeg.tar.xz" | sha256sum -c -; \
+    tar -xJf /out/ffmpeg.tar.xz -C /out --strip-components=1 --wildcards '*/ffmpeg' '*/ffprobe' '*/GPLv3.txt'; \
+    rm /out/ffmpeg.tar.xz
+
 # ── runtime ─────────────────────────────────────────────────────────────────
 # distroless static + nonroot: no shell, no package manager, runs as uid 65532.
 # ca-certificates are bundled, which the OIDC/HTTPS clients need.
 FROM gcr.io/distroless/static-debian12:nonroot
 
 COPY --from=build /out/tapir /tapir
+COPY --from=ffmpeg /out/ffmpeg /out/ffprobe /usr/local/bin/
+COPY --from=ffmpeg /out/GPLv3.txt /usr/share/doc/ffmpeg/GPLv3.txt
+
+# Where the service finds them. Uploads stay OFF until TAPIR_UPLOADS_ENABLED is set
+# (a deployment decision, not an image one).
+ENV TAPIR_FFMPEG_BIN=/usr/local/bin/ffmpeg \
+    TAPIR_FFPROBE_BIN=/usr/local/bin/ffprobe
 
 EXPOSE 8080
 USER nonroot:nonroot

LANGUAGE.md

@@ -0,0 +1,26 @@
+# Vocabulary (hand-written pilot 2026-06 — will be generated by langgen in Phase 2)
+
+| Term | Means | Never say |
+|---|---|---|
+| transcript | Captions-first text of a video; the source for summarizing. Shared store keyed by `(provider, video_id)`. | "subtitles file", "the audio" |
+| video | Per-user record of a seen video; the unit of work. Not globally deduped. | "the shared video" |
+| subscription | A watched channel on a user's connection that Tapir polls for new videos. | "feed", "follow" |
+| summary | A video's produced output: summary text + highlights + takeaways + AI provenance. | "transcript" |
+| highlight | A notable point pulled from a video (`Summary.Highlights`). | "takeaway" |
+| takeaway | An actionable conclusion from a video (`Summary.Takeaways`). | "highlight" |
+| sink | A delivery destination for a summary (`store`, `brain`). New one = new adapter. | "the database" |
+| AI router | Local-first chain: local Primary → local fallback → external worst-case. | "the API" |
+| BYO-AI fallback | User's own external AI key, used only when local fails; opt-in. Without it, content is never sent externally. | "the default AI" |
+| brain | Persistent homelab knowledge store; in Tapir, one optional HTTP sink (`brain_ingest`), not the filesystem package. | "the database", "the store" |
+| LiteLLM gateway | Local AI gateway (`koala:30401/v1`); the Primary in the AI router. | "piguard:4000", "koala:4000", "the cloud" |
+| connection | A connected video account a user authorizes via OAuth; subscriptions hang off it. | "login", "session" |
+
+## Caveman rubric
+Before any HIGH/CRITICAL operation, output one line:
+`caveman: me <verb> <object>, not <excluded thing>`
+Valid iff: (1) only vocabulary terms + plain verbs, (2) a stranger could identify
+the exact operation, (3) names one thing explicitly NOT being done.
+
+Examples:
+- `caveman: me delete user connection, not the shared transcript`
+- `caveman: me send transcript to BYO-AI fallback, not the LiteLLM gateway`

Taskfile.yml

@@ -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

cmd/tapir/chat_test.go

@@ -4,7 +4,7 @@ import (
 	"strings"
 	"testing"
 
-	"gitea.d-ma.be/mathias/tapir/internal/config"
+	"git.d-ma.be/mathias/tapir/internal/config"
 )
 
 // TestChatModelsReuseTheChainLocalFirst: the switcher offers the ADR-022 chain in

cmd/tapir/discovery.go

@@ -5,7 +5,7 @@ import (
 	"log/slog"
 	"sync"
 
-	"gitea.d-ma.be/mathias/tapir/internal/runner"
+	"git.d-ma.be/mathias/tapir/internal/runner"
 )
 
 // discoveryRunner runs one user's discovery pass.

cmd/tapir/discovery_test.go

@@ -10,7 +10,7 @@ import (
 
 	"github.com/stretchr/testify/require"
 
-	"gitea.d-ma.be/mathias/tapir/internal/runner"
+	"git.d-ma.be/mathias/tapir/internal/runner"
 )
 
 // serialize must guarantee at most one discovery pass runs at a time, so a

cmd/tapir/e2e_upload_test.go

@@ -0,0 +1,209 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"mime/multipart"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	embeddedpostgres "github.com/fergusstrange/embedded-postgres"
+	"github.com/jackc/pgx/v5/pgxpool"
+	"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/config"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/web"
+)
+
+// TestE2E_UploadTranscribeAndChat is the end-to-end check for tapir#33: upload a
+// real audio file, watch it become a transcript through the REAL ffprobe, ffmpeg
+// and gateway, then ask a question and get a grounded answer from a REAL model.
+// It is opt-in — it needs real binaries and the live gateway — and skipped unless
+// TAPIR_E2E_AUDIO is set:
+//
+//	TAPIR_E2E_AUDIO=/path/meeting.m4a TAPIR_E2E_FFMPEG=/path/ffmpeg \
+//	TAPIR_E2E_FFPROBE=/path/ffprobe TAPIR_E2E_QUESTION='Vem ansvarar för virket?' \
+//	TAPIR_E2E_EXPECT=anders op run --env-file ~/.op-env -- \
+//	  go test ./cmd/tapir -run E2E_Upload -v -count=1
+//
+// Use SYNTHETIC audio only: this repo is public, and a fallback to berget sends
+// the audio to a cloud endpoint (#33).
+func TestE2E_UploadTranscribeAndChat(t *testing.T) {
+	audioPath := os.Getenv("TAPIR_E2E_AUDIO")
+	if audioPath == "" {
+		t.Skip("set TAPIR_E2E_AUDIO (and friends) to run the live end-to-end check")
+	}
+	key := os.Getenv("LITELLM_MASTER_KEY")
+	require.NotEmpty(t, key, "LITELLM_MASTER_KEY must be set (via op run)")
+	ctx, cancel := context.WithCancel(context.Background())
+	t.Cleanup(cancel)
+
+	// A private embedded Postgres: nothing here may touch a real database.
+	port := uint32(55000 + os.Getpid()%1000)
+	rt := filepath.Join(os.TempDir(), fmt.Sprintf("tapir-epg-e2e-%d", os.Getpid()))
+	pg := embeddedpostgres.NewDatabase(embeddedpostgres.DefaultConfig().Port(port).
+		RuntimePath(rt).DataPath(filepath.Join(rt, "data")).BinariesPath(filepath.Join(rt, "bin")).
+		CachePath(filepath.Join(os.TempDir(), "tapir-epg-cache")))
+	require.NoError(t, pg.Start())
+	t.Cleanup(func() { _ = pg.Stop(); _ = os.RemoveAll(rt) })
+	dsn := fmt.Sprintf("postgres://postgres:postgres@localhost:%d/postgres?sslmode=disable", port)
+
+	st, err := store.New(ctx, dsn)
+	require.NoError(t, err)
+	t.Cleanup(st.Close)
+	raw, err := pgxpool.New(ctx, dsn)
+	require.NoError(t, err)
+	t.Cleanup(raw.Close)
+	const uid = "11111111-1111-1111-1111-111111111111"
+	_, err = raw.Exec(ctx, `INSERT INTO users (id) VALUES ($1)`, uid)
+	require.NoError(t, err)
+	_, err = raw.Exec(ctx, `INSERT INTO user_identities (dex_subject, user_id) VALUES ('e2e', $1)`, uid)
+	require.NoError(t, err)
+
+	cfg, err := config.Load()
+	require.NoError(t, err)
+	cfg.GatewayKey = key
+	cfg.UploadsEnabled = true
+	cfg.UploadTmpDir = filepath.Join(t.TempDir(), "spool")
+	if os.Getenv("TAPIR_E2E_RECONCILE") == "1" {
+		// Two engines and a local model, as ADR-034 describes.
+		cfg.ReconcileBaseModel = envOrDefault("TAPIR_E2E_BASE_MODEL", "iguana/pianissimo")
+		cfg.ReconcileModel = envOrDefault("TAPIR_E2E_RECONCILE_MODEL", "iguana/gemma4-26b")
+	}
+	cfg.FFmpegBin = envOrDefault("TAPIR_E2E_FFMPEG", "ffmpeg")
+	cfg.FFprobeBin = envOrDefault("TAPIR_E2E_FFPROBE", "ffprobe")
+
+	intake, worker, err := buildUploads(ctx, cfg, st, quietLog())
+	require.NoError(t, err)
+	go worker.Run(ctx)
+	app := &web.App{
+		Store: st, Identity: st, Auth: web.StubAuth{U: web.User{Subject: "e2e"}},
+		Chat: buildChat(cfg),
+	}
+	wireUploads(app, cfg, intake) // the same wiring cmdServe uses
+	srv := httptest.NewServer(app.Router())
+	t.Cleanup(srv.Close)
+	noRedirect := &http.Client{CheckRedirect: func(*http.Request, []*http.Request) error { return http.ErrUseLastResponse }}
+
+	// 1. Upload.
+	f, err := os.Open(audioPath)
+	require.NoError(t, err)
+	defer func() { _ = f.Close() }()
+	pr, pw := io.Pipe()
+	mw := multipart.NewWriter(pw)
+	go func() {
+		_ = mw.WriteField("title", "E2E recording")
+		fw, _ := mw.CreateFormFile("file", filepath.Base(audioPath))
+		_, _ = io.Copy(fw, f)
+		_ = mw.Close()
+		_ = pw.Close()
+	}()
+	req, err := http.NewRequest(http.MethodPost, srv.URL+"/uploads", pr)
+	require.NoError(t, err)
+	req.Header.Set("Content-Type", mw.FormDataContentType())
+	resp, err := noRedirect.Do(req)
+	require.NoError(t, err)
+	_ = resp.Body.Close()
+	require.Equal(t, http.StatusSeeOther, resp.StatusCode, "an accepted upload redirects to its page")
+	loc := resp.Header.Get("Location")
+	require.True(t, strings.HasPrefix(loc, "/uploads/"), loc)
+	t.Logf("accepted -> %s", loc)
+
+	// 2. Wait for the transcript: unfinished states poll, terminal states do not.
+	var page string
+	require.Eventually(t, func() bool {
+		r, err := http.Get(srv.URL + loc + "/status")
+		if err != nil {
+			return false
+		}
+		b, _ := io.ReadAll(r.Body)
+		_ = r.Body.Close()
+		page = string(b)
+		return !strings.Contains(page, "hx-trigger")
+	}, 12*time.Minute, 3*time.Second, "the upload never reached a terminal state")
+	require.Contains(t, page, "Transcribed", "terminal state was not 'done': %.300s", page)
+
+	var status, endpoint, content string
+	var secs float64
+	var hadVideo bool
+	require.NoError(t, raw.QueryRow(ctx,
+		`SELECT status, endpoint, content, audio_seconds, had_video FROM uploaded_transcripts WHERE user_id = $1`, uid).
+		Scan(&status, &endpoint, &content, &secs, &hadVideo))
+	t.Logf("done: endpoint=%s audio_s=%.1f transcript_chars=%d had_video=%v", endpoint, secs, len(content), hadVideo)
+	if os.Getenv("TAPIR_E2E_RECONCILE") == "1" {
+		var recJSON string
+		require.NoError(t, raw.QueryRow(ctx,
+			`SELECT COALESCE(reconciliation::text, 'NULL') FROM uploaded_transcripts WHERE user_id = $1`, uid).Scan(&recJSON))
+		require.NotEqual(t, "NULL", recJSON, "a two-engine run must record its reconciliation")
+		t.Logf("reconciliation: %.600s", recJSON)
+		assert.Contains(t, endpoint, "reconciled by", "every contributor is named in the endpoint")
+		assert.Contains(t, page, "Reconciled from two transcripts")
+	}
+	if os.Getenv("TAPIR_E2E_HAD_VIDEO") == "1" {
+		// A Teams meeting with the camera off: a video file whose audio is what matters.
+		assert.True(t, hadVideo, "the file had a video track and the row must say so")
+		assert.Contains(t, page, "Only its audio was used", "and the user must be told")
+	}
+	assert.Equal(t, "done", status)
+	assert.NotEmpty(t, endpoint)
+	assert.NotEmpty(t, content)
+
+	// ADR-035: a finished job's phases are all closed, and it was reporting while it ran.
+	var progJSON string
+	var progressAt, aliveAt *time.Time
+	require.NoError(t, raw.QueryRow(ctx,
+		`SELECT COALESCE(progress::text, '[]'), progress_at, alive_at FROM uploaded_transcripts WHERE user_id = $1`, uid).
+		Scan(&progJSON, &progressAt, &aliveAt))
+	var steps []domain.ProgressStep
+	require.NoError(t, json.Unmarshal([]byte(progJSON), &steps))
+	require.NotEmpty(t, steps, "the job recorded no phases")
+	for _, st := range steps {
+		assert.Contains(t, []domain.StepState{domain.StepDone, domain.StepSkipped}, st.State,
+			"phase %q was left %q after the job finished", st.Key, st.State)
+	}
+	assert.Equal(t, "save", steps[len(steps)-1].Key)
+	assert.NotNil(t, progressAt, "progress_at must have been stamped")
+	t.Logf("phases: %d, heartbeat stamped: %v", len(steps), aliveAt != nil)
+	entries, err := os.ReadDir(cfg.UploadTmpDir)
+	require.NoError(t, err)
+	assert.Empty(t, entries, "the audio must be deleted once transcribed")
+
+	// 3. Ask a question; the answer must come from the transcript.
+	q := envOrDefault("TAPIR_E2E_QUESTION", "Who is responsible for the timber order?")
+	chatResp, err := http.PostForm(srv.URL+loc+"/chat", url.Values{"question": {q}})
+	require.NoError(t, err)
+	answerPage, _ := io.ReadAll(chatResp.Body)
+	_ = chatResp.Body.Close()
+	require.Equal(t, http.StatusOK, chatResp.StatusCode)
+	t.Logf("question: %s", q)
+	if want := os.Getenv("TAPIR_E2E_EXPECT"); want != "" {
+		assert.Contains(t, strings.ToLower(string(answerPage)), strings.ToLower(want),
+			"the answer should be grounded in the transcript")
+	}
+	t.Logf("answer page excerpt: %.600s", afterMarker(string(answerPage), `class="chat-turn chat-a"`))
+}
+
+func envOrDefault(k, d string) string {
+	if v := os.Getenv(k); v != "" {
+		return v
+	}
+	return d
+}
+
+func afterMarker(s, m string) string {
+	if i := strings.Index(s, m); i >= 0 {
+		return s[i:]
+	}
+	return "(no answer turn found)"
+}

cmd/tapir/env.go

@@ -5,7 +5,7 @@ import (
 	"fmt"
 	"os"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
 )
 
 // Env var names for the read-only CLI. DSN and user id are never hardcoded — the

cmd/tapir/format_test.go

@@ -7,7 +7,7 @@ import (
 
 	"github.com/stretchr/testify/require"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
 )
 
 func TestFormatListColumnsAndOrdering(t *testing.T) {

cmd/tapir/list.go

@@ -9,7 +9,7 @@ import (
 	"strings"
 	"text/tabwriter"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
 )
 
 // runList prints the user's stored summaries as a table, most recent first.

cmd/tapir/main.go

@@ -17,20 +17,21 @@ import (
 	"errors"
 	"fmt"
 	"log/slog"
+	"net"
 	"net/http"
 	"os"
-	"os/signal"
 	"sync"
 	"time"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/secrets"
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/youtube"
-	"gitea.d-ma.be/mathias/tapir/internal/auth"
-	"gitea.d-ma.be/mathias/tapir/internal/config"
-	"gitea.d-ma.be/mathias/tapir/internal/runner"
-	"gitea.d-ma.be/mathias/tapir/internal/web"
-	"gitea.d-ma.be/mathias/tapir/internal/web/oidc"
+	"git.d-ma.be/mathias/tapir/internal/adapters/secrets"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/adapters/youtube"
+	"git.d-ma.be/mathias/tapir/internal/auth"
+	"git.d-ma.be/mathias/tapir/internal/config"
+	"git.d-ma.be/mathias/tapir/internal/metrics"
+	"git.d-ma.be/mathias/tapir/internal/runner"
+	"git.d-ma.be/mathias/tapir/internal/web"
+	"git.d-ma.be/mathias/tapir/internal/web/oidc"
 )
 
 func main() {
@@ -40,7 +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
@@ -206,6 +207,28 @@ func cmdServe(ctx context.Context, log *slog.Logger) error {
 		log.Info("web chat enabled (stored-transcript only)", "models", chatModels(cfg))
 	}
 
+	// Audio upload -> private transcript (ADR-033). Off unless TAPIR_UPLOADS_ENABLED:
+	// the image must carry ffmpeg first. The worker is IN-PROCESS and one job at a
+	// time, so it inherits the single-replica assumption below (its queue is
+	// process-local); it stops with the same ctx as the server.
+	intake, worker, err := buildUploads(ctx, cfg, st, log)
+	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 {
+		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)")
+		if cfg.TranscribeFallbackModel != "" {
+			log.Warn("audio upload fallback is a CLOUD endpoint: if the primary fails, private audio leaves the homelab",
+				"fallback", cfg.TranscribeFallbackModel)
+		}
+	}
+
 	// User onboarding is handled by the IdP (Authentik invite flow), not Tapir —
 	// the Dex local-password provisioning path was removed (ADR-019). An
 	// authenticated subject with no Tapir user is routed to /register.
@@ -310,21 +333,38 @@ func cmdServe(ctx context.Context, log *slog.Logger) error {
 
 	srv := &http.Server{
 		Addr:              cfg.HTTPAddr,
-		Handler:           app.Router(),
+		Handler:           metrics.HTTPMiddleware(app.Router()),
 		ReadHeaderTimeout: 10 * time.Second,
 	}
 
-	// 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)
-	}()
+	// Prometheus /metrics on a SEPARATE port (ADR-030) — never on the public app
+	// mux, so a scrape is in-cluster only. Empty TAPIR_METRICS_ADDR disables it.
+	var metricsSrv *http.Server
+	if cfg.MetricsAddr != "" {
+		mmux := http.NewServeMux()
+		mmux.Handle("GET /metrics", metrics.Handler())
+		metricsSrv = &http.Server{Addr: cfg.MetricsAddr, Handler: mmux, ReadHeaderTimeout: 10 * time.Second}
+		go func() {
+			log.Info("serving metrics", "addr", cfg.MetricsAddr)
+			if err := metricsSrv.ListenAndServe(); err != nil && !errors.Is(err, http.ErrServerClosed) {
+				log.Error("metrics server", "err", err)
+			}
+		}()
+	}
 
+	// 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
 }

cmd/tapir/processor.go

@@ -5,17 +5,18 @@ import (
 	"fmt"
 	"strings"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/chat"
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/llm"
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/secrets"
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/summarizer"
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/youtube"
-	"gitea.d-ma.be/mathias/tapir/internal/config"
-	"gitea.d-ma.be/mathias/tapir/internal/domain"
-	"gitea.d-ma.be/mathias/tapir/internal/ports"
-	"gitea.d-ma.be/mathias/tapir/internal/usecase"
-	"gitea.d-ma.be/mathias/tapir/internal/web"
+	"git.d-ma.be/mathias/tapir/internal/adapters/chat"
+	"git.d-ma.be/mathias/tapir/internal/adapters/llm"
+	"git.d-ma.be/mathias/tapir/internal/adapters/secrets"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/adapters/summarizer"
+	"git.d-ma.be/mathias/tapir/internal/adapters/youtube"
+	"git.d-ma.be/mathias/tapir/internal/config"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/metrics"
+	"git.d-ma.be/mathias/tapir/internal/ports"
+	"git.d-ma.be/mathias/tapir/internal/usecase"
+	"git.d-ma.be/mathias/tapir/internal/web"
 )
 
 // videoFetcher adapts the YouTube adapter to web.VideoFetcher for the paste flow
@@ -69,7 +70,7 @@ func buildSummarizer(cfg config.Config) *summarizer.Summarizer {
 func summarizerEndpoint(cfg config.Config) func(model string) summarizer.Endpoint {
 	return func(model string) summarizer.Endpoint {
 		return summarizer.Endpoint{
-			Client:   llm.New(cfg.GatewayURL, cfg.GatewayKey, model, cfg.SummarizerTimeout, llm.WithMaxTokens(cfg.SummaryMaxTokens)),
+			Client:   llm.New(cfg.GatewayURL, cfg.GatewayKey, model, cfg.SummarizerTimeout, llm.WithMaxTokens(cfg.SummaryMaxTokens), llm.WithUsageHook(metrics.RecordTokens)),
 			Provider: providerOf(model),
 			Model:    model,
 		}
@@ -133,9 +134,30 @@ func chatModels(cfg config.Config) []string {
 	add(cfg.SummarizerModel)
 	add(cfg.FallbackModel)
 	add(cfg.CloudFallbackModel)
+	// The model chat over an uploaded recording opens on must be one the shared service can
+	// call: it re-maps a model it does not offer, which silently swapped a configured 31B for
+	// phi4-mini, whose context cannot hold an hour of speech. Local only, and only when
+	// uploads are on, so the video chat is unchanged otherwise (second adversarial review).
+	if cfg.UploadsEnabled && config.IsLocalModel(cfg.UploadChatModel) {
+		add(cfg.UploadChatModel)
+	}
 	return models
 }
 
+// uploadChatModels is the chat switcher for a PRIVATE uploaded recording: the same
+// ordered set as chatModels, local models only. The whole transcript (up to 90,000
+// characters) is sent to the one chosen, so a cloud model the video chat offers must not
+// be offered here; with none local, none is offered (ADR-035 review).
+func uploadChatModels(cfg config.Config) []string {
+	var local []string
+	for _, m := range chatModels(cfg) {
+		if config.IsLocalModel(m) {
+			local = append(local, m)
+		}
+	}
+	return local
+}
+
 // buildChat wires the per-video chat service (ADR-027): a Completer factory over
 // the SAME LiteLLM gateway the summarizer uses (a different alias per model, not a
 // second client config) and the same transcript-truncation budget. It returns nil
@@ -150,7 +172,7 @@ func buildChat(cfg config.Config) *chat.Service {
 		return nil
 	}
 	newClient := func(model string) chat.Completer {
-		return llm.New(cfg.GatewayURL, cfg.GatewayKey, model, cfg.SummarizerTimeout, llm.WithMaxTokens(cfg.SummaryMaxTokens))
+		return llm.New(cfg.GatewayURL, cfg.GatewayKey, model, cfg.SummarizerTimeout, llm.WithMaxTokens(cfg.SummaryMaxTokens), llm.WithUsageHook(metrics.RecordTokens))
 	}
 	return chat.New(newClient, models, cfg.MaxTranscriptChars)
 }

cmd/tapir/processor_test.go

@@ -3,7 +3,7 @@ package main
 import (
 	"testing"
 
-	"gitea.d-ma.be/mathias/tapir/internal/config"
+	"git.d-ma.be/mathias/tapir/internal/config"
 )
 
 // TestBuildProcessorNilOnIncompleteConfig asserts the queue-only fallback: when a

cmd/tapir/report.go

@@ -8,7 +8,7 @@ import (
 	"text/tabwriter"
 	"time"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
 )
 
 // gateThreshold is the Stage-0 gate (VISION/ADR-016): usage in >= 2 distinct

cmd/tapir/report_test.go

@@ -7,7 +7,7 @@ import (
 
 	"github.com/stretchr/testify/require"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
 )
 
 var testSince = time.Date(2026, 6, 11, 0, 0, 0, 0, time.UTC)

cmd/tapir/scheduler.go

@@ -6,13 +6,13 @@ import (
 	"log/slog"
 	"time"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/youtube"
-	"gitea.d-ma.be/mathias/tapir/internal/config"
-	"gitea.d-ma.be/mathias/tapir/internal/ports"
-	"gitea.d-ma.be/mathias/tapir/internal/runner"
-	"gitea.d-ma.be/mathias/tapir/internal/usecase"
-	"gitea.d-ma.be/mathias/tapir/internal/web"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/adapters/youtube"
+	"git.d-ma.be/mathias/tapir/internal/config"
+	"git.d-ma.be/mathias/tapir/internal/ports"
+	"git.d-ma.be/mathias/tapir/internal/runner"
+	"git.d-ma.be/mathias/tapir/internal/usecase"
+	"git.d-ma.be/mathias/tapir/internal/web"
 )
 
 // buildUserRunner constructs a runner.Runner for one user, reusing the same

cmd/tapir/scheduler_test.go

@@ -11,8 +11,8 @@ import (
 
 	"github.com/stretchr/testify/require"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
-	"gitea.d-ma.be/mathias/tapir/internal/runner"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/runner"
 )
 
 func quietLog() *slog.Logger {

cmd/tapir/show.go

@@ -8,7 +8,7 @@ import (
 	"os"
 	"strings"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
 )
 
 // runShow prints the full summary for one video: text, highlights, takeaways,

cmd/tapir/shutdown.go

@@ -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
+}

cmd/tapir/shutdown_test.go

@@ -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")
+	}
+}

cmd/tapir/uploads.go

@@ -0,0 +1,142 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"log/slog"
+	"net/http"
+	"os"
+	"path/filepath"
+	"time"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/audio"
+	"git.d-ma.be/mathias/tapir/internal/adapters/llm"
+	"git.d-ma.be/mathias/tapir/internal/adapters/transcribe"
+	"git.d-ma.be/mathias/tapir/internal/config"
+	"git.d-ma.be/mathias/tapir/internal/reconcile"
+	"git.d-ma.be/mathias/tapir/internal/uploads"
+	"git.d-ma.be/mathias/tapir/internal/web"
+)
+
+// uploadsStore is the store surface the upload wiring needs. *store.Store
+// satisfies it.
+type uploadsStore interface {
+	uploads.Store
+	uploads.WorkerStore
+	FailInterruptedUploads(ctx context.Context) (int, error)
+}
+
+// buildUploads wires audio upload and private transcription (ADR-033). It returns
+// nils when uploads are disabled — never a typed-nil Intake, which would mount the
+// /uploads routes over nothing. When enabled it, in order: prepares a private spool
+// directory, deletes any audio a crashed run left in it, fails the rows such a run
+// left queued or transcribing (the audio is not retained, so they can never
+// resume), and only then builds the intake and the worker. The caller runs the
+// worker. The fallback endpoint is a cloud service; an empty
+// TAPIR_TRANSCRIBE_FALLBACK_MODEL keeps private audio on the homelab.
+func buildUploads(ctx context.Context, cfg config.Config, st uploadsStore, log *slog.Logger) (*uploads.Intake, *uploads.Worker, error) {
+	if !cfg.UploadsEnabled {
+		return nil, nil, nil
+	}
+	if cfg.GatewayURL == "" {
+		return nil, nil, fmt.Errorf("uploads: TAPIR_UPLOADS_ENABLED is set but there is no TAPIR_GATEWAY_URL to transcribe through")
+	}
+
+	if err := os.MkdirAll(cfg.UploadTmpDir, 0o700); err != nil {
+		return nil, nil, fmt.Errorf("uploads: spool dir: %w", err)
+	}
+	removeSpooledAudio(cfg.UploadTmpDir, log)
+
+	n, err := st.FailInterruptedUploads(ctx)
+	if err != nil {
+		return nil, nil, fmt.Errorf("uploads: fail interrupted rows: %w", err)
+	}
+	if n > 0 {
+		log.Warn("uploads: marked interrupted by a restart (audio is not retained; they must be re-uploaded)", "count", n)
+	}
+
+	worker := uploads.NewWorker(st, newConverter(cfg), buildTranscriber(cfg),
+		cfg.UploadTmpDir, cfg.TranscribeTimeout, uploads.WithMaxDuration(cfg.UploadMaxDuration))
+	intake := uploads.NewIntake(st, &audio.FFprobe{Bin: cfg.FFprobeBin}, worker,
+		uploadLimits(cfg), cfg.UploadTmpDir)
+	return intake, worker, nil
+}
+
+func uploadLimits(cfg config.Config) uploads.Limits {
+	return uploads.Limits{MaxBytes: cfg.UploadMaxBytes, MaxDuration: cfg.UploadMaxDuration}
+}
+
+// removeSpooledAudio deletes the spool files this process family creates
+// (upload-*, wav-*) and nothing else.
+func removeSpooledAudio(dir string, log *slog.Logger) {
+	for _, pat := range []string{"upload-*", "wav-*"} {
+		matches, _ := filepath.Glob(filepath.Join(dir, pat))
+		for _, m := range matches {
+			if err := os.Remove(m); err != nil {
+				log.Warn("uploads: remove leftover audio", "path", m, "err", err)
+			}
+		}
+	}
+}
+
+// wireUploads connects a built Intake to the web App: the routes, the limits the
+// form states, and the chat settings for an uploaded recording. A nil intake
+// leaves the App untouched — assigning a nil *Intake to the Uploader interface
+// would produce a non-nil interface and mount the routes over nothing.
+func wireUploads(app *web.App, cfg config.Config, in *uploads.Intake) {
+	if in == nil {
+		return
+	}
+	app.Uploads = in
+	app.UploadLimits = uploadLimits(cfg)
+	app.UploadChatMaxChars = cfg.MaxUploadedTranscriptChars
+	app.UploadChatModel = cfg.UploadChatModel
+	app.UploadChatModels = uploadChatModels(cfg)
+}
+
+// newConverter is the ffmpeg adapter capped at the configured maximum duration, so a file
+// whose header lies about its length cannot become hours of WAV (ADR-035 review).
+func newConverter(cfg config.Config) *audio.FFmpeg {
+	return &audio.FFmpeg{Bin: cfg.FFmpegBin, MaxSeconds: int(cfg.UploadMaxDuration.Seconds())}
+}
+
+// adjudicatorTimeout bounds one adjudication call. A call takes ~2 s on gemma4-26b and ~10 s
+// on 31b; this is the wait before an unresponsive model is counted as an error.
+const adjudicatorTimeout = 2 * time.Minute
+
+// buildTranscriber builds what the worker transcribes with. By default that is the
+// single-engine chain (the primary, then the optional cloud fallback). When
+// TAPIR_RECONCILE_BASE_MODEL is set it is the two-engine ensemble (ADR-034): the base
+// engine and the primary run on the same WAV and a LOCAL model reconciles them. Config
+// has already refused any non-local model, so nothing here can send private audio or
+// text to a cloud model for reconciliation; the cloud fallback is unchanged.
+func buildTranscriber(cfg config.Config) uploads.Transcriber {
+	// No overall client timeout: a 90 minute recording takes minutes. The deadline is
+	// the worker's per-job context.
+	client := &http.Client{}
+	// Each LOCAL engine call gets its own deadline: a wedged engine does not answer, and
+	// without one it used the whole job before the fallback could run.
+	local := func(model string) transcribe.Backend {
+		return transcribe.WithTimeout(transcribe.NewGateway(cfg.GatewayURL, cfg.GatewayKey, model, client), cfg.TranscribeEngineTimeout)
+	}
+	primary := local(cfg.TranscribeModel)
+	var fallback transcribe.Backend
+	if cfg.TranscribeFallbackModel != "" && cfg.TranscribeFallbackModel != cfg.TranscribeModel {
+		fallback = transcribe.NewGateway(cfg.GatewayURL, cfg.GatewayKey, cfg.TranscribeFallbackModel, client)
+	}
+	if cfg.ReconcileBaseModel == "" {
+		return transcribe.NewChain(primary, fallback)
+	}
+
+	base := local(cfg.ReconcileBaseModel)
+	// Temperature 0 and a short reply: the adjudicator's accuracy was measured that way.
+	chat := llm.New(cfg.GatewayURL, cfg.GatewayKey, cfg.ReconcileModel, adjudicatorTimeout,
+		llm.WithTemperature(0), llm.WithMaxTokens(200))
+	var adj reconcile.Adjudicator = reconcile.NewLLMAdjudicator(chat)
+	if cfg.ReconcileCheckOrder {
+		adj = reconcile.NewOrderChecked(adj)
+	}
+	return transcribe.NewEnsemble(base, primary, fallback, adj, reconcile.Options{
+		BaseEngine: cfg.ReconcileBaseModel, AltEngine: cfg.TranscribeModel, Reconciler: cfg.ReconcileModel,
+	})
+}

cmd/tapir/uploads_chat_models_test.go

@@ -0,0 +1,85 @@
+package main
+
+import (
+	"reflect"
+	"testing"
+
+	"git.d-ma.be/mathias/tapir/internal/config"
+)
+
+// The chat over an uploaded recording sends the whole private transcript to the model
+// the user picks, so the switcher for it may offer local models only (ADR-035 review).
+
+func TestUploadChatModels_NeverOffersACloudModelButKeepsTheLocalOrder(t *testing.T) {
+	var cfg config.Config
+	cfg.SummarizerModel = "koala/phi4-mini"
+	cfg.FallbackModel = "iguana/gemma4-26b"
+	cfg.CloudFallbackModel = "berget/mistral-small"
+
+	got := uploadChatModels(cfg)
+
+	want := []string{"koala/phi4-mini", "iguana/gemma4-26b"}
+	if !reflect.DeepEqual(got, want) {
+		t.Errorf("uploadChatModels = %v, want %v: the cloud model that the video chat offers must not be offered for a private recording", got, want)
+	}
+}
+
+func TestUploadChatModels_ACloudPrimaryIsDroppedToo(t *testing.T) {
+	var cfg config.Config
+	cfg.SummarizerModel = "berget/mistral-small"
+	cfg.FallbackModel = "iguana/gemma4-26b"
+
+	got := uploadChatModels(cfg)
+
+	if !reflect.DeepEqual(got, []string{"iguana/gemma4-26b"}) {
+		t.Errorf("uploadChatModels = %v", got)
+	}
+}
+
+func TestUploadChatModels_NoLocalModelMeansNoneOfferedNotACloudDefault(t *testing.T) {
+	var cfg config.Config
+	cfg.SummarizerModel = "berget/mistral-small"
+
+	if got := uploadChatModels(cfg); len(got) != 0 {
+		t.Errorf("uploadChatModels = %v, want none", got)
+	}
+}
+
+// A valid local TAPIR_UPLOAD_CHAT_MODEL that is not in the video chat's chain used to be
+// accepted by config and then silently ignored: the shared chat service re-maps a model it
+// does not offer, so the user got koala/phi4-mini, whose 8,192-token context cannot hold an
+// hour of speech (second adversarial review).
+
+func TestChatModels_AnUploadChatModelOutsideTheChainIsOfferedWhenUploadsAreOn(t *testing.T) {
+	var cfg config.Config
+	cfg.SummarizerModel = "koala/phi4-mini"
+	cfg.FallbackModel = "iguana/gemma4-26b"
+	cfg.UploadsEnabled = true
+	cfg.UploadChatModel = "iguana/gemma4-31b"
+
+	service := chatModels(cfg)
+	upload := uploadChatModels(cfg)
+
+	if !reflect.DeepEqual(service, []string{"koala/phi4-mini", "iguana/gemma4-26b", "iguana/gemma4-31b"}) {
+		t.Errorf("chatModels = %v: the shared service must be able to call the configured upload model", service)
+	}
+	if !reflect.DeepEqual(upload, service) {
+		t.Errorf("uploadChatModels = %v, want %v", upload, service)
+	}
+}
+
+func TestChatModels_NothingIsAddedWhenUploadsAreOffOrTheModelIsNotLocal(t *testing.T) {
+	var cfg config.Config
+	cfg.SummarizerModel = "koala/phi4-mini"
+	cfg.UploadChatModel = "iguana/gemma4-31b"
+
+	if got := chatModels(cfg); !reflect.DeepEqual(got, []string{"koala/phi4-mini"}) {
+		t.Errorf("uploads off: chatModels = %v, want the video chain unchanged", got)
+	}
+
+	cfg.UploadsEnabled = true
+	cfg.UploadChatModel = "berget/mistral-small"
+	if got := chatModels(cfg); !reflect.DeepEqual(got, []string{"koala/phi4-mini"}) {
+		t.Errorf("a cloud upload chat model was added to the service: %v", got)
+	}
+}

cmd/tapir/uploads_converter_test.go

@@ -0,0 +1,26 @@
+package main
+
+import (
+	"testing"
+	"time"
+
+	"git.d-ma.be/mathias/tapir/internal/config"
+)
+
+// The converter is told the configured maximum, so a file whose header lies about its
+// length cannot become hours of WAV (adversarial review F2).
+
+func TestNewConverter_IsCappedAtTheConfiguredMaximumDuration(t *testing.T) {
+	var cfg config.Config
+	cfg.FFmpegBin = "/usr/local/bin/ffmpeg"
+	cfg.UploadMaxDuration = 90 * time.Minute
+
+	c := newConverter(cfg)
+
+	if c.Bin != "/usr/local/bin/ffmpeg" {
+		t.Errorf("Bin = %q", c.Bin)
+	}
+	if c.MaxSeconds != 5400 {
+		t.Errorf("MaxSeconds = %v, want 5400", c.MaxSeconds)
+	}
+}

cmd/tapir/uploads_test.go

@@ -0,0 +1,196 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/transcribe"
+	"git.d-ma.be/mathias/tapir/internal/config"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/web"
+)
+
+// fakeUploadsStore satisfies the store surface buildUploads needs, and counts the
+// start-up sweep.
+type fakeUploadsStore struct {
+	sweeps      int
+	sweepErr    error
+	failedKinds []domain.UploadErrorKind
+}
+
+func (f *fakeUploadsStore) CreateUpload(context.Context, string, string, domain.Language, bool) (string, error) {
+	return "id", nil
+}
+
+func (f *fakeUploadsStore) FailUpload(_ context.Context, _, _ string, k domain.UploadErrorKind) error {
+	f.failedKinds = append(f.failedKinds, k)
+	return nil
+}
+
+func (f *fakeUploadsStore) StartTranscribing(context.Context, string, string) error { return nil }
+
+func (f *fakeUploadsStore) CompleteUpload(context.Context, string, string, domain.Transcription) error {
+	return nil
+}
+
+func (f *fakeUploadsStore) SaveProgress(context.Context, string, string, []domain.ProgressStep, time.Time) 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++
+	return 2, f.sweepErr
+}
+
+func uploadCfg(t *testing.T, enabled bool) config.Config {
+	t.Helper()
+	var cfg config.Config
+	cfg.GatewayURL = "http://gateway.invalid/v1"
+	cfg.UploadsEnabled = enabled
+	cfg.UploadMaxBytes = 200 << 20
+	cfg.UploadTmpDir = filepath.Join(t.TempDir(), "spool")
+	cfg.TranscribeModel = "iguana/kb-whisper"
+	cfg.TranscribeFallbackModel = "berget/kb-whisper"
+	return cfg
+}
+
+func TestBuildUploads_DisabledMountsNothingAndSweepsNothing(t *testing.T) {
+	st := &fakeUploadsStore{}
+
+	in, w, err := buildUploads(context.Background(), uploadCfg(t, false), st, quietLog())
+
+	require.NoError(t, err)
+	assert.Nil(t, in, "a typed-nil Intake must never reach web.App.Uploads: it would mount the routes over nothing")
+	assert.Nil(t, w)
+	assert.Zero(t, st.sweeps)
+}
+
+func TestBuildUploads_EnabledFailsInterruptedRowsBeforeAnythingElseStarts(t *testing.T) {
+	st := &fakeUploadsStore{}
+
+	in, w, err := buildUploads(context.Background(), uploadCfg(t, true), st, quietLog())
+
+	require.NoError(t, err)
+	require.NotNil(t, in)
+	require.NotNil(t, w)
+	assert.Equal(t, 1, st.sweeps,
+		"audio is not retained, so rows left queued/transcribing by a restart can never resume: fail them at start-up, not never")
+}
+
+func TestBuildUploads_SweepFailureStopsStartup(t *testing.T) {
+	st := &fakeUploadsStore{sweepErr: errors.New("db down")}
+
+	_, _, err := buildUploads(context.Background(), uploadCfg(t, true), st, quietLog())
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "db down")
+}
+
+func TestBuildUploads_CreatesTheSpoolDirectoryPrivately(t *testing.T) {
+	cfg := uploadCfg(t, true)
+
+	_, _, err := buildUploads(context.Background(), cfg, &fakeUploadsStore{}, quietLog())
+
+	require.NoError(t, err)
+	info, err := os.Stat(cfg.UploadTmpDir)
+	require.NoError(t, err)
+	assert.True(t, info.IsDir())
+	assert.Equal(t, os.FileMode(0o700), info.Mode().Perm(), "spooled audio is private third-party content")
+}
+
+func TestBuildUploads_EnabledWithoutAGatewayIsAnError(t *testing.T) {
+	cfg := uploadCfg(t, true)
+	cfg.GatewayURL = ""
+
+	_, _, err := buildUploads(context.Background(), cfg, &fakeUploadsStore{}, quietLog())
+
+	assert.Error(t, err, "uploads on with nowhere to transcribe must fail loudly at start-up, not at the first upload")
+}
+
+func TestBuildUploads_RemovesAudioLeftBehindByACrash(t *testing.T) {
+	cfg := uploadCfg(t, true)
+	require.NoError(t, os.MkdirAll(cfg.UploadTmpDir, 0o700))
+	leftovers := []string{"upload-123", "wav-456"}
+	for _, n := range leftovers {
+		require.NoError(t, os.WriteFile(filepath.Join(cfg.UploadTmpDir, n), []byte("audio"), 0o600))
+	}
+	unrelated := filepath.Join(cfg.UploadTmpDir, "someone-elses-file")
+	require.NoError(t, os.WriteFile(unrelated, []byte("keep"), 0o600))
+
+	_, _, err := buildUploads(context.Background(), cfg, &fakeUploadsStore{}, quietLog())
+
+	require.NoError(t, err)
+	for _, n := range leftovers {
+		_, err := os.Stat(filepath.Join(cfg.UploadTmpDir, n))
+		assert.True(t, os.IsNotExist(err), "%s: audio is never retained, so a spooled file from a crashed run must not survive start-up", n)
+	}
+	_, err = os.Stat(unrelated)
+	assert.NoError(t, err, "only our own spool files are removed")
+}
+
+func TestWireUploads_SetsEverythingTheWebAppNeedsForUploadsAndChat(t *testing.T) {
+	cfg := uploadCfg(t, true)
+	cfg.UploadMaxDuration = 90 * 60 * 1e9
+	cfg.MaxUploadedTranscriptChars = 90000
+	cfg.UploadChatModel = "iguana/gemma4-26b"
+	in, _, err := buildUploads(context.Background(), cfg, &fakeUploadsStore{}, quietLog())
+	require.NoError(t, err)
+	app := &web.App{}
+
+	wireUploads(app, cfg, in)
+
+	assert.NotNil(t, app.Uploads)
+	assert.Equal(t, cfg.UploadMaxBytes, app.UploadLimits.MaxBytes, "the form states the limits Intake enforces")
+	assert.Equal(t, cfg.UploadMaxDuration, app.UploadLimits.MaxDuration)
+	assert.Equal(t, 90000, app.UploadChatMaxChars,
+		"without this the chat would use the 18,000-char caption default and cut an hour to its first third")
+	assert.Equal(t, "iguana/gemma4-26b", app.UploadChatModel)
+}
+
+func TestWireUploads_NilIntakeLeavesTheAppUntouched(t *testing.T) {
+	app := &web.App{}
+
+	wireUploads(app, uploadCfg(t, false), nil)
+
+	assert.Nil(t, app.Uploads, "a nil *Intake must not become a non-nil interface that mounts the routes")
+	assert.Zero(t, app.UploadChatMaxChars)
+}
+
+func TestBuildTranscriber_DefaultIsTheSingleEngineChainAsBefore(t *testing.T) {
+	cfg := uploadCfg(t, true)
+
+	tr := buildTranscriber(cfg)
+
+	assert.IsType(t, &transcribe.Chain{}, tr, "with no base engine configured nothing changes")
+}
+
+func TestBuildTranscriber_ABaseEngineTurnsOnTheEnsemble(t *testing.T) {
+	cfg := uploadCfg(t, true)
+	cfg.ReconcileBaseModel = "iguana/pianissimo"
+	cfg.ReconcileModel = "iguana/gemma4-26b"
+
+	tr := buildTranscriber(cfg)
+
+	assert.IsType(t, &transcribe.Ensemble{}, tr)
+}
+
+func TestBuildUploads_WiresTheEnsembleWhenConfigured(t *testing.T) {
+	cfg := uploadCfg(t, true)
+	cfg.ReconcileBaseModel = "iguana/pianissimo"
+	cfg.ReconcileModel = "iguana/gemma4-26b"
+
+	in, w, err := buildUploads(context.Background(), cfg, &fakeUploadsStore{}, quietLog())
+
+	require.NoError(t, err)
+	assert.NotNil(t, in)
+	assert.NotNil(t, w)
+}

cmd/tapir/uploads_timeout_test.go

@@ -0,0 +1,78 @@
+package main
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// The wiring of the per-engine deadline: a gateway whose primary model never answers
+// must still reach the fallback model inside the job's own deadline (adversarial review
+// F4). The fake gateway answers per model from the multipart `model` field.
+
+func TestBuildTranscriber_AWedgedPrimaryReachesTheFallbackBeforeTheJobDeadline(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_ = r.ParseMultipartForm(1 << 20)
+		if r.FormValue("model") == "iguana/kb-whisper" {
+			<-r.Context().Done() // wedged: never answers
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"text":"hej","duration":1,"segments":[{"start":0,"end":1,"text":"hej"}]}`))
+	}))
+	defer srv.Close()
+	cfg := uploadCfg(t, true)
+	cfg.GatewayURL = srv.URL
+	cfg.TranscribeModel = "iguana/kb-whisper"
+	cfg.TranscribeFallbackModel = "berget/kb-whisper"
+	cfg.TranscribeEngineTimeout = 150 * time.Millisecond
+	wav := filepath.Join(t.TempDir(), "a.wav")
+	require.NoError(t, os.WriteFile(wav, []byte("RIFF....WAVEfmt "), 0o600))
+	job, cancel := context.WithTimeout(context.Background(), 20*time.Second)
+	defer cancel()
+
+	got, err := buildTranscriber(cfg).Transcribe(job, wav)
+
+	require.NoError(t, err, "the wedged primary used the whole job and the fallback never ran")
+	assert.Equal(t, "berget/kb-whisper", got.Endpoint)
+}
+
+func TestBuildTranscriber_AWedgedBaseEngineIsASoftFailureInTheEnsemble(t *testing.T) {
+	// The same deadline must guard the ensemble's BASE engine, not only the chain's primary
+	// (second adversarial review: the wiring was tested for the chain alone).
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_ = r.ParseMultipartForm(1 << 20)
+		if r.FormValue("model") == "iguana/pianissimo" {
+			<-r.Context().Done()
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"text":"hej","duration":1,"segments":[{"start":0,"end":1,"text":"hej"}]}`))
+	}))
+	defer srv.Close()
+	cfg := uploadCfg(t, true)
+	cfg.GatewayURL = srv.URL
+	cfg.TranscribeModel = "iguana/kb-whisper"
+	cfg.ReconcileBaseModel = "iguana/pianissimo"
+	cfg.ReconcileModel = "iguana/gemma4-31b"
+	cfg.TranscribeEngineTimeout = 150 * time.Millisecond
+	wav := filepath.Join(t.TempDir(), "a.wav")
+	require.NoError(t, os.WriteFile(wav, []byte("RIFF....WAVEfmt "), 0o600))
+	job, cancel := context.WithTimeout(context.Background(), 20*time.Second)
+	defer cancel()
+
+	got, err := buildTranscriber(cfg).Transcribe(job, wav)
+
+	require.NoError(t, err, "a wedged base engine must not use up the job")
+	assert.Equal(t, "hej", got.Text)
+	require.NotNil(t, got.Reconciliation)
+	assert.Equal(t, "single_engine", string(got.Reconciliation.Status))
+	assert.Contains(t, got.Reconciliation.Reason, "timeout")
+}

cmd/tapir/uploads_worker_limit_test.go

@@ -0,0 +1,39 @@
+package main
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/uploads"
+)
+
+// The worker buildUploads returns enforces the configured maximum on what was really
+// converted (adversarial review F2). A stub ffmpeg writes a sparse 100-minute WAV however
+// short the file claimed to be; the gateway is never reached because the job fails first.
+
+func TestBuildUploads_TheWorkerRefusesAudioLongerThanTheConfiguredMaximum(t *testing.T) {
+	dir := t.TempDir()
+	stub := filepath.Join(dir, "ffmpeg")
+	require.NoError(t, os.WriteFile(stub, []byte("#!/bin/sh\nfor a; do last=\"$a\"; done\ntruncate -s 192000044 \"$last\"\n"), 0o700))
+	cfg := uploadCfg(t, true)
+	cfg.FFmpegBin = stub
+	cfg.UploadMaxDuration = 90 * time.Minute
+	cfg.TranscribeTimeout = time.Minute
+	st := &fakeUploadsStore{}
+	_, w, err := buildUploads(context.Background(), cfg, st, quietLog())
+	require.NoError(t, err)
+	src := filepath.Join(cfg.UploadTmpDir, "upload-x")
+	require.NoError(t, os.WriteFile(src, []byte("audio"), 0o600))
+
+	w.Process(context.Background(), uploads.Job{UserID: "u", UploadID: "id", Path: src, Duration: 10 * time.Second})
+
+	assert.Equal(t, []domain.UploadErrorKind{domain.UploadErrTooLong}, st.failedKinds,
+		"a file that claimed 10 s but converted to 100 minutes must fail, not be sent to an engine")
+}

docs/data-model.md

@@ -18,7 +18,12 @@ only opaque references to them; the secret material lives in ESO/1Password (ADR-
   not LLM re-summarization but a rate-gated, reputation-risky caption fetch (ADR-010/014), which
   is paid per re-fetch regardless of user count — so persisting public caption content once and
   sharing it strictly beats the coupling it removes. Everything else each user owns is
-  self-contained; `rls_test.go` proves transcripts is the single exception.
+  self-contained; `rls_test.go` proves transcripts is the single exception for *content*.
+  **Uploaded-audio transcripts are the opposite of public caption content** (private, owned by
+  one user) and therefore live in their own RLS-scoped `uploaded_transcripts` table, never in
+  `transcripts` (ADR-033). The isolation test now also pins the **exact set** of tables with row
+  security off (`transcripts`, `user_identities`, `invitations`, `schema_migrations`), so a new
+  table added without RLS fails CI.
 - **Secrets by reference only.** Tables hold a `secret_ref` (opaque string/UUID resolved via
   the `SecretStore` port), never tokens or keys.
 - **The brain sink is just a delivery target.** No brain-specific tables. Whether a summary
@@ -41,6 +46,7 @@ erDiagram
     VIDEO ||--o| SUMMARY : "has at most one"
     SUMMARY ||--o{ SINK_DELIVERY : "delivered via"
     USER ||--o{ CHANNEL_ERROR : "reports unavailable channels"
+    USER ||--o{ UPLOADED_TRANSCRIPT : "uploads (private, ADR-033)"
 
     USER {
         uuid id PK
@@ -136,12 +142,32 @@ erDiagram
         timestamptz first_seen
         timestamptz last_seen
     }
+    UPLOADED_TRANSCRIPT {
+        uuid id PK
+        uuid user_id FK "-> USER, ON DELETE CASCADE, NOT NULL; RLS ENABLE+FORCE (migration 017)"
+        text title
+        text status "queued | transcribing | done | failed"
+        text content "nullable until done; the transcript text"
+        jsonb segments "nullable until done; [{start,end,text}] in seconds"
+        text endpoint "which backend served it: iguana/... | berget/... (berget = audio left the homelab)"
+        float audio_seconds
+        text error_kind "typed: interrupted | transcription_failed — never a provider string"
+        bool had_video "the file carried a video track; only its audio was used (migration 018)"
+        jsonb reconciliation "NULL for a single engine; else status, exact counts and a bounded list of every judged spot (migration 019, ADR-034; all spots stored, ADR-035)"
+        jsonb progress "the job's ordered phases: key, label, state, times, done/total; NULL until the job starts (migration 020, ADR-035)"
+        timestamptz progress_at "when a phase last really changed or advanced (migration 020)"
+        timestamptz alive_at "heartbeat from the running job, every ~10 s; not progress (migration 020)"
+        timestamptz created_at
+        timestamptz finished_at
+    }
 ```
 
 `SUMMARY_ACTION` has `UNIQUE (user_id, video_id, action)`; `VIDEO_CONNECTION` has
 `UNIQUE (user_id, provider)` (one connection per provider — reconnect upserts in place).
 RLS (`ENABLE` + `FORCE`) is on **every solid user-owned table above** — `users`, `videos`,
-`transcripts`, `summaries`, `summary_actions`, `video_connections`, `channel_errors`.
+`summaries`, `summary_actions`, `video_connections`, `channel_errors`, `uploaded_transcripts`.
+(`transcripts` is deliberately **not** in this list: since ADR-021 it is shared public caption
+content with no `user_id`.)
 `sink_deliveries` is RLS'd via an `EXISTS` on its parent summary; `user_identities` is
 intentionally **not** RLS'd (auth plumbing). See the *Isolation invariant* section for the
 mechanism.
@@ -197,6 +223,14 @@ mechanism.
   the most recent discovery pass. Upserted per scheduler pass (`last_seen` refreshed each run);
   surfaced on the account page as a warning. Cascades on user deletion. Primary key is
   `(user_id, channel_id)`. FORCE RLS'd.
+- **UPLOADED_TRANSCRIPTS** (migration 017, ADR-033) — the **private** transcript of audio a user
+  uploaded. Deliberately a separate table from the shared `transcripts`: that one is non-RLS on the
+  grounds that it holds only public caption content, and a recorded meeting is the opposite.
+  `user_id NOT NULL ... ON DELETE CASCADE`, FORCE RLS'd. `status` runs `queued → transcribing →
+  done | failed`; `done` requires `content`, `failed` requires a typed `error_kind` (CHECK
+  constraints). `reconciliation` (migration 019, ADR-034) is NULL when one engine produced the text, so the page can never claim a second opinion that was not there; otherwise it holds the status (`reconciled` / `single_engine` / `skipped`), exact counts (differences, settled automatically, decided by the model, changed, uncertain, model errors) and up to 2,000 judged spots (ADR-035: every one the model decided, kept or changed or unsure, each a position in seconds, both readings, which was used, and the few words before it; the page lists only the changed and uncertain ones). `progress`, `progress_at` and `alive_at` (migration 020, ADR-035) let the page show the job's phases and tell a running job from a quiet one from a lost service: `alive_at` is a heartbeat from a goroutine independent of any engine call, `progress_at` moves only when work really moved; the saved `segments` carry the reconciled words with the base engine's timing. The spots hold fragments of the private transcript, which is why they live in this RLS-scoped row. `had_video` (migration 018) records that the file carried a video track — a Teams meeting with the camera off is still a video file — so the page can say only its audio was used and the video was discarded. `endpoint` records which backend served it — `berget/...` means the audio left the
+  homelab. The audio itself is never stored. `segments` keeps the timestamps (re-transcribing 90
+  minutes costs ~7). Unlike `summary_actions`/`login_events` it **does** cascade on user deletion.
 - **LOGIN_EVENTS** (migration 010) — throttled one-row-per-(user, date) login stamp. Used by the
   Stage-0 gate query (VISION §Stage 0: "returned and used in ≥2 distinct weeks").
 
@@ -209,7 +243,10 @@ enforcement dormant); **ADR-012 opened Stage 1 and turned enforcement on in the
 Enforcement is **Postgres Row-Level Security** (migration `003_rls.up.sql`):
 
 - RLS is `ENABLE`d **and** `FORCE`d on every user-owned table — `users`, `videos`,
-  `transcripts`, `summaries`, `summary_actions`, `video_connections`, `channel_errors`.
+  `summaries`, `summary_actions`, `video_connections`, `channel_errors`, and (migration 017)
+  `uploaded_transcripts`. `transcripts` is **not** one of them: ADR-021 made it shared public
+  content, and migration 003's original per-user `transcripts` policy was dropped with that
+  table in migration 015.
   `FORCE` is load-bearing: the app connects as the table **owner** (`tapir` role), and owners
   bypass RLS unless forced.
 - Each policy keys off the per-request GUC `tapir.current_user_id`, set transaction-locally by

docs/homelab-integration.md

@@ -12,6 +12,29 @@ it** — endpoints and aliases drift, and this file is a snapshot (2026-06-06),
   - **Off-cluster (host / NodePort):** `http://koala:30401/v1/` — service `litellm-nodeport` (`4000:30401`).
   - NOTE: the earlier `31234` was **wrong** — `31234` is the `llama-swap` NodePort (`8080:31234`),
     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: 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
+  field `file`, `response_format=verbose_json`, no `Expect: 100-continue`. iguana's whisper-server
+  can wedge permanently after a Metal OOM (infra#273); never branch on its
+  `{"error":"failed to process audio"}` string — it has had at least five causes. Tapir's knobs:
+  `TAPIR_UPLOADS_ENABLED` (off by default), `TAPIR_UPLOAD_MAX_MB` (200), `TAPIR_UPLOAD_MAX_MINUTES`
+  (90), `TAPIR_TRANSCRIBE_MODEL`, `TAPIR_TRANSCRIBE_FALLBACK_MODEL` (empty disables the cloud
+  fallback), `TAPIR_TRANSCRIBE_TIMEOUT` (60m, the whole job), `TAPIR_TRANSCRIBE_ENGINE_TIMEOUT` (20m, one engine call: a wedged engine then fails while the job lives and the fallback can run), `TAPIR_UPLOAD_TMP_DIR`, `TAPIR_FFMPEG_BIN`,
+  `TAPIR_FFPROBE_BIN`, `TAPIR_MAX_UPLOADED_TRANSCRIPT_CHARS` (90000 — the measured ceiling of
+  `iguana/gemma4-26b`, which silently drops the head above it), `TAPIR_UPLOAD_CHAT_MODEL`
+  (`iguana/gemma4-26b`; must be a local `koala/` or `iguana/` alias, since the whole private transcript is sent to it; `koala/phi4-mini`'s 8,192-token context 400s on an hour of speech). `ffmpeg` + `ffprobe` are in the runtime image (static, pinned by sha256; #34 decided A on 2026-10-02); Tapir's own ingress body limit for a 200 MB upload is unverified (infra).
+  **Two-engine mode (ADR-034), off unless `TAPIR_RECONCILE_BASE_MODEL` is set:** it runs
+  `iguana/pianissimo` (the base text) and `TAPIR_TRANSCRIBE_MODEL` (kb-whisper) on the same audio, then a
+  LOCAL chat model (`TAPIR_RECONCILE_MODEL`) chooses between their disagreements. **Deployment choice (2026-10-02): `iguana/gemma4-31b`** (97% against a speaker's blind answers; 92% for `iguana/gemma4-26b`). With iguana's macOS **Low Power Mode off** the 31B decodes at 22 tok/s and reconciled a 63.6-minute recording in 21 min 49 s end to end; with it on it ran at 8 tok/s. Check `pmset -g | grep lowpowermode` before blaming a model or an engine for slowness.
+  `TAPIR_RECONCILE_CHECK_ORDER=true` asks twice per dispute with the order swapped (doubles the time).
+  Every model must start `koala/` or `iguana/`: a cloud alias is a start-up error. The default job timeout is
+  60 minutes. `iguana/pianissimo` is KlangAI's Swedish ASR (a Parakeet/FastConformer on MLX, served by a small
+  FastAPI shell on iguana:8102) and speaks the same `/v1/audio/transcriptions` API.
 - **Auth key:** treat as a secret via the SecretStore port; do not hardcode in committed code.
   NOTE: the documented `sk-local-123` returned `401` on `30401/v1/models` (2026-06-02) — it is
   stale or the gateway now enforces the `LITELLM_MASTER_KEY` (1Password `op://HomeLab/LITELLM_MASTER_KEY`).
@@ -227,6 +250,12 @@ knobs plus one load-bearing deployment constraint:
 - `TAPIR_USAGE_GATE_START` — `YYYY-MM-DD`, default **`2026-06-11`** (the morning the pilot was
   unblocked and summaries started flowing). `tapir report` counts return-usage (distinct active
   weeks, ADR-016) only from this date, so pre-launch testing and the blocked period are excluded.
+- `TAPIR_METRICS_ADDR` — listen address for the Prometheus `/metrics` endpoint (ADR-030).
+  **Default `:9090`** — a SEPARATE port from `TAPIR_HTTP_ADDR` so metrics are never on the public
+  app; scraped in-cluster only (PodMonitor). Empty disables the metrics server. Key series:
+  `tapir_summarize_duration_seconds{model,outcome,fallback}`, `tapir_caption_fetch_duration_seconds{outcome}`,
+  `tapir_chat_duration_seconds{model}`, `tapir_llm_tokens_total{model,kind}`,
+  `tapir_http_request_duration_seconds{method,route}`, `tapir_logins_total`.
 - `TAPIR_FETCH_RATE` — Go duration, default `2s`. The **process-wide per-egress-IP caption-fetch
   rate gate** (ADR-014 item 2). Every caption fetch — scheduler runners *and* the web "Summarize"
   click-path — serialises through this one limiter so the pod cannot collectively trip 429s. `0`

docs/research/asr-kb-whisper-vs-pianissimo-2026-10-01.md

@@ -0,0 +1,81 @@
+# kb-whisper vs pianissimo on one real hour of Swedish speech (tapir#33, 2026-10-01)
+
+One recording, one speaker: a 63.7-minute Teams session of the maintainer dictating project
+notes (domain-heavy, English loanwords and product names mixed into Swedish). **No transcript
+text and no domain terms are recorded here**: this repo is public and the content is private.
+Only aggregate numbers.
+
+Backends, both through the LiteLLM gateway and Tapir's own adapter, unchanged:
+`iguana/kb-whisper` (whisper.cpp, KB's Swedish fine-tune) and `iguana/pianissimo` (KlangAI
+Swedish Parakeet/FastConformer-TDT, MLX 8-bit, served by a small FastAPI shell on iguana:8102).
+
+## Result: a near tie on accuracy, a large difference in speed
+
+| | kb-whisper | pianissimo |
+|---|---|---|
+| Wall time for 63.7 min | 4m08s (~15x realtime) | **40 s (~95x realtime)** |
+| Segments | 1,302 (timed pieces, wrapped at ~60 chars) | 434 (sentences) |
+| Words (normalised) | 6,606 | 6,510 |
+| Punctuation | light, every segment ends in a full stop | heavier (138 vs 82 commas) |
+| Repetition loops | none | none |
+
+Word-level agreement between them: **92.4%** as first computed, **93.1%** after the segment
+rejoin fix below. 380 aligned disagreement spans, median one word each.
+
+**Blind A/B by the speaker** (the only ground truth available), on the 30 spans a model
+(Claude Opus 5.5) had picked as changing meaning, shuffled and unlabelled: **pianissimo 17,
+kb-whisper 13**, no "neither", no "can't tell".
+
+| Category of the 30 | pianissimo | kb-whisper |
+|---|---|---|
+| Technical term / English loanword | 8 | 4 |
+| Number / date / unit | 4 | 1 |
+| Name / proper noun | 3 | 3 |
+| Function word / small word | 1 | 2 |
+| Other (everyday words) | 1 | 3 |
+
+The reading: pianissimo is better on terminology and numbers, kb-whisper on ordinary Swedish
+words; names are even. Their errors differ in kind: kb-whisper's tend to be garbled words, a
+couple were wrong numbers; pianissimo's tend to be fluent substitutions that read as plausible
+(a wrong institution, an invented person, a dropped sign), which a reader or an LLM is less
+likely to catch.
+
+Opus's own lean agreed with the speaker's blind answer on **30 of 30**. That is not a measure
+of Opus over all 380 spans: it chose spans it could decide, and every one of the 30 carried a
+lean. Its full classification of the 380 (counts only): function/small word 81, filler or
+false start 65, technical term 58, compound split/join 46, other 39, name 28, casing or
+punctuation residue 26, inflection 23, number/date/unit 10, content omission or insertion
+(>= 3 words) 4.
+
+## A bug this comparison found in Tapir (fixed)
+
+kb-whisper's output is wrapped at ~60 characters and **a wrap can fall inside a word**; the
+continuation segment has no leading space, the response's `text` field joins segments with
+newlines, and a segment can be a lone `.`. Tapir stored that joined text, so chat received
+broken words. Measured on real gateway output: 7% of segments (96/1,332) in a synthetic hour,
+42/1,302 lines in the real one. berget's segments are whole sentences with no leading space, so
+the two need different joins. `transcribe.joinSegments` now rebuilds the text from the segments.
+Checked against the real transcript with pianissimo's vocabulary as an independent yardstick:
+36 spans changed, 31 now entirely known words, words unknown to pianissimo 224 -> 177.
+Five rejoined spans still contain an unknown word (a name or an unlucky glue): small residual
+risk. Earlier tests used short sentences and could not see this.
+
+## What this does not show
+
+- **One recording, one speaker, one domain.** No claim about other voices, accents, crosstalk,
+  or room noise. 17 vs 13 is not a significant difference.
+- **The other ~350 spans** were not judged by the speaker. They are mostly spelling, compounds,
+  fillers and small words; Opus's counts there are a model's judgement, not ground truth.
+- Errors **both** models share in text they agree on are invisible to this method.
+- Speed was measured once each, on an otherwise idle iguana; pianissimo's 40 s includes the
+  gateway hop and a 122 MB WAV upload.
+- Whether pianissimo survives the infra#273 Metal-OOM contention the way whisper does not was
+  not tested; both run on iguana.
+
+## What it suggests for Tapir (not decided, filed as an issue)
+
+Pianissimo is a **second local backend on the same host, through the same adapter**. That
+makes it a natural middle tier ahead of the cloud fallback (kb-whisper -> pianissimo ->
+berget), so a wedged whisper no longer sends private audio off the homelab. At 40 s an hour it
+is also cheap enough to run **both** on every upload and give chat both texts, or flag where
+they disagree; the maintainer's blind answers show they err in different places.

docs/research/audio-length-ceiling-2026-10-01.md

@@ -0,0 +1,110 @@
+# Audio length ceiling — measured before designing chunking (tapir#33, 2026-10-01)
+
+Question (#33, "the main technical risk"): at what recording length does Swedish
+transcription through the LiteLLM gateway first fail, and with what failure mode?
+Measured **before** any chunking design, with iguana not holding a large ollama model,
+so the result is the ceiling and not the Metal OOM of infra#273.
+
+## Answer
+
+**No transcription ceiling was found at or below 90 minutes.** A single request carrying
+90 minutes of 16 kHz mono PCM (174 MB) transcribed in one pass in 440 s. The first thing
+that actually fails is **not whisper**: it is the public gateway ingress, which rejects the
+upload with `413` at roughly 5.5 minutes of WAV.
+
+| Route | First failure | Failure mode |
+|---|---|---|
+| `https://llm-api.d-ma.be` (openresty ingress) | between 8.0 MB and 10.5 MB body (~4–5.5 min of WAV) | `413 Request Entity Too Large`, HTML body from `openresty`, returned instantly |
+| `http://koala:30401` (LiteLLM NodePort) → `iguana/kb-whisper` | **none up to 90 min / 174 MB** | — |
+| `http://koala:30401` → `berget/kb-whisper` | **none up to 22 min / 42 MB** (largest tried) | — |
+
+The ingress limit is the reason the 8:20 / 15 MB recording in #33 "failed repeatedly" and a
+30-second slice succeeded. That failure was attributed to a Metal OOM; at least in the
+size-dependent form described, a body-size cap explains it without one. (The OOM
+mechanism in infra#273 is separately real; this run did not provoke it.)
+
+The exact ingress cap was bracketed (8.0 MB passed, 10.5 MB failed), not read from the
+ingress configuration, which lives in `infra`. "About 10 MB" is an inference.
+
+## Method
+
+- **Audio.** No known-good Swedish recording is available that is safe to use: the only real
+  one is a private recording (#31 custody still open) and was
+  deliberately not touched. Instead: macOS `say -v Alva` (Swedish) on iguana, a 1,171-char
+  five-paragraph meeting script → 82.5 s of 16 kHz mono PCM WAV (`ffmpeg -ar 16000 -ac 1`).
+  Concatenated ×4, ×8, ×16, ×32, ×44, ×66 with Python `wave` (the issue's own recipe).
+  Nothing here is a real transcript; none of it is committed.
+- **Request.** `POST /v1/audio/transcriptions`, multipart `file=`, `model=iguana/kb-whisper`,
+  `response_format=verbose_json`, `-H 'Expect:'` (per #33's traps), master key supplied only
+  through `op run` and never on argv.
+- **Confound control.** `ollama /api/ps` on iguana was read before every step. Resident:
+  nothing, or `nomic-embed-text` (0.37 GB). Never a large model.
+
+## Results (`iguana/kb-whisper`, NodePort)
+
+| Audio | WAV size | HTTP | Wall | Realtime factor | Segments | Chars |
+|---|---|---|---|---|---|---|
+| 0:30 | 0.96 MB | 200 | 2.8 s | 11× | — | — |
+| 2:30 / 3:20 / 4:10 | 4.8 / 6.4 / 8.0 MB | 200 (ingress) | 12.9 / 16.0 / 19.9 s | 12× | — | — |
+| 5:30 (×4) | 10.5 MB | **413 ingress**, 200 NodePort in 27 s | 27 s | 12× | 121 | 4,557 |
+| 11:00 (×8) | 21.1 MB | 200 | 53 s | 12× | 241 | 9,094 |
+| 22:00 (×16) | 42.3 MB | 200 | 105 s | 13× | 482 | 18,201 |
+| 44:01 (×32) | 84.5 MB | 200 | 209 s | 13× | 966 | 36,438 |
+| 60:32 (×44) | 116.2 MB | 200 | 291 s | 12× | 1,332 | 50,145 |
+| 90:48 (×66) | 174.3 MB | 200 | 440 s | 12× | 2,005 | 75,165 |
+
+- Throughput is flat at ~12× realtime. 90 minutes costs ~7.5 minutes of wall time. This is
+  a minutes-long job, not a request/response one (see ADR-033 on ADR-018).
+- Segments per source repetition stay at 30.1–30.4 from ×4 to ×66, and segment times are
+  monotonic with `last_end` within 0.2 s of `duration`. No drift and no repetition-loop
+  hallucination appeared across 90 minutes. This is a weak check — the input is a repeated
+  loop and synthetic voice — not evidence of accuracy.
+- **Only `iguana/kb-whisper`'s `x4` run reported `last_end` 360.0 s against `duration`
+  330.2 s.** Whisper's final segment can end past the audio; clamp to `duration`.
+
+## Findings that change the design
+
+1. **Use the NodePort route, not the public hostname.** Tapir runs in-cluster beside the
+   gateway; `llm-api.d-ma.be` is the external door and has a body cap. #33 said "use the
+   gateway, never the host" and named the public URL; the gateway is right, that URL is not.
+   `TAPIR_GATEWAY_URL` already points at `koala:30401/v1`.
+2. **`response_format=srt` survives the LiteLLM hop, but wrapped.** The response is JSON
+   `{"text": "<the SRT>", "usage": …}`, not a bare SRT body — a client that parses the body
+   as SRT gets a parse error on `{`. #33 question 2 is answered: it passes through, inside
+   an envelope.
+3. **The two backends return different timestamp granularity.** `iguana/kb-whisper`
+   segments are **whole seconds** (`0→1`, `1→4`, `4→7`); `berget/kb-whisper` returns real
+   sub-second times (`0.051→1.192`) and a different `language` token (`swedish` vs `sv`).
+   `verbose_json` and the default JSON were byte-identical on iguana. An adapter that
+   falls back between the two must normalise both, and "which endpoint served this" is
+   a quality fact, not just a log line.
+4. **The existing chat path cannot answer about most of a long recording.** At
+   828 chars/min (this fixture; real speech is plausibly a little denser),
+   `TAPIR_MAX_TRANSCRIPT_CHARS`'s default of **18,000** is ~21.7 minutes. A 60-minute
+   transcript is 50 k chars, a 90-minute one 75 k. `chat.Service` truncates by keeping
+   the head (`internal/adapters/chat/chat.go`), so questions about the last two-thirds of
+   a one-hour meeting get an answer from a transcript that does not contain it. #33 says
+   chat "retrieves rather than generates" and therefore has no ceiling; it **stuffs and
+   truncates**, and its ceiling is the model context. See ADR-033.
+5. **Chunking is not needed for v1 on length.** Nothing failed up to 90 minutes. The
+   chunking-and-stitching recipe in #33 (offset timestamps, renumber cues, strip `\r`) is
+   not built. It stays the answer *if* a longer cap is wanted or the OOM mechanism forces
+   smaller requests — see ADR-033's resolution triggers.
+
+## What this did not measure
+
+- **Metal OOM under contention.** Deliberately avoided. A big resident ollama model
+  during a long transcription is still the known way to wedge iguana permanently.
+- **Real speech.** TTS is clean, single-voice, no crosstalk or room noise. WER and
+  speaker overlap in a Teams export are untested. Speed is probably representative;
+  accuracy is not claimed.
+- **Compressed input.** Only WAV was sent. Whether `m4a`/`mp3`/`opus` can reach whisper
+  without a server-side transcode is not established here (ADR-033 treats it as needing one).
+- **Concurrency.** One request at a time. Two simultaneous 90-minute jobs on iguana were
+  not tried.
+- **`berget/kb-whisper` beyond 22 min / 42 MB.** Enough to know the fallback is not
+  size-capped at meeting length for 22 minutes; not proven at 90.
+- **Time-of-day / other tenants.** One afternoon, one run per length.
+
+Scripts were throwaway shell/Python in a session scratchpad and are not committed; the
+method above reproduces them.

docs/research/chat-context-over-long-transcripts-2026-10-01.md

@@ -0,0 +1,66 @@
+# Chat over a long transcript — what the offered models really hold (tapir#33, 2026-10-01)
+
+Question: ADR-033 reuses `chat.Service`, which stuffs the transcript into the prompt and
+cuts the tail. #33 assumed chat "retrieves rather than generates" and so has no length
+ceiling. It does not retrieve, so its ceiling is the model's context. What is it, for the
+models chat actually offers?
+
+## Answer
+
+| Model (via the NodePort gateway) | Result | Limit found |
+|---|---|---|
+| `koala/phi4-mini` | correct to 30,000 chars; **HTTP 400 at 50,000** (13,332 tokens vs an 8,192-token context) | ~22,000 chars in production, where the client adds `max_tokens: 1500` |
+| `iguana/gemma4-26b` | correct at 18k, 50k and **90,000 chars** (26.5 k tokens); **at 120,000 chars it silently loses the head** | **90,000 chars** |
+| `berget/mistral-small` | answered "the transcript does not contain it" for a fact **at the end** of an 18,000-char and a 50,000-char transcript | unreliable here, one run each — see caveats |
+
+The dangerous one is gemma at 120,000 chars: **no error, no truncation flag** — the reported
+prompt tokens fell from 26,548 (90 k chars) to 16,387 (120 k chars), and the model answered
+"it does not say who the chair is" about a sentence at the very start of the transcript. The
+serving layer truncated the prompt from the front. `chat.Service` reports `Truncated: false`
+because *it* did not cut anything.
+
+## What this changes
+
+1. **The budget for an uploaded transcript is 90,000 chars, not a round 120,000.** That is the
+   best offered model's measured limit, ~100 minutes of speech at the fixture's density.
+   `TAPIR_MAX_UPLOADED_TRANSCRIPT_CHARS` defaults to it.
+2. **Chat over an upload must not open on the summarizer default.** The offered list is
+   `[SummarizerModel, FallbackModel, CloudFallbackModel]`, so `koala/phi4-mini` is first — and
+   400s on an hour of speech. `TAPIR_UPLOAD_CHAT_MODEL` (default `iguana/gemma4-26b`) is what
+   the upload chat opens with; the switcher still lists every model.
+3. **A per-model budget does not exist yet.** Someone who switches the dropdown to
+   `phi4-mini` on a long recording gets a loud HTTP 400, shown inline as "couldn't answer".
+   That fails honestly but unhelpfully. Filed as a follow-up.
+4. **The "retrieval chat" trigger in ADR-033 is now concrete:** any recording past 90,000 chars
+   (≈100 min) cannot be answered by stuffing, for any offered model.
+
+## Method
+
+`POST /v1/chat/completions` through `http://koala:30401/v1`, `temperature: 0`,
+`max_tokens: 200` (production sends 1500 for chat via the summary token cap), the same system
+framing as `chat.Service`. The transcript is the 60-minute synthetic fixture from
+`audio-length-ceiling-2026-10-01.md` truncated/repeated to N chars, with one unique sentence
+planted either at the **end** ("lagret på Kvarngatan ska tömmas … av Pelle", question: who/when)
+or at the **head** ("kvällens ordförande är Ingrid", question: who chairs). Needle at the
+end alone would have missed the gemma failure — a serving layer that truncates from the front
+keeps the tail, so the end-needle passed at 120,000 chars; only the head-needle exposed it.
+Grounded = the answer contains the planted fact.
+
+## What this did not measure
+
+- **Real speech**: the fixture is a repeated 82-second script, so retrieval "difficulty" is
+  artificially easy except for the planted needle. One question per run, one run per cell.
+- **`berget/mistral-small`'s failure**: it missed an end-needle at 18,000 chars where the
+  others succeeded. One run; could be the model, the prompt framing, or position bias. Not
+  investigated; do not read it as a ranking.
+- **Latency under load**, **concurrent** chat while whisper is transcribing, and in particular
+  the hazard below.
+- **`koala/qwen36-35b-a3b` or any frontier model** — not in the offered list.
+
+## A hazard this surfaces and does not solve
+
+`iguana/gemma4-26b` is a large ollama model on the same host as `iguana/kb-whisper`. infra#273
+is precisely that a large resident ollama model can Metal-OOM whisper-server permanently. Making
+gemma the default *chat* model for uploads means chatting with a transcript can load the model
+that wedges the transcriber. The primary-fails-over-to-berget design limits the blast radius but
+does not remove it. Recorded in ADR-033 and filed.

docs/research/local-model-adjudicator-sweep-2026-10-01.md

@@ -0,0 +1,85 @@
+# Can a local model adjudicate two ASR transcripts? (tapir#33, 2026-10-01)
+
+Question: when `iguana/kb-whisper` and `iguana/pianissimo` disagree about a word or phrase,
+can a **local** model say which reading is right, well enough to merge the two transcripts?
+Ground truth: the speaker's own blind answers on 30 disagreements of one real recording
+(see `asr-kb-whisper-vs-pianissimo-2026-10-01.md`; no transcript text is recorded here,
+this repo is public). Reference: Claude Opus 5.5 agreed with the speaker on all 30.
+
+## Method
+
+- **Items:** 30 disagreement spans that Opus had picked as changing meaning, the speaker's
+  verdict as truth (pianissimo right 17, kb-whisper right 13).
+- **Prompt:** one fixed English instruction (two transcriptions of Swedish speech differ at
+  ONE marked spot; say which reading the speaker most likely said; reply `ANSWER: 1|2|UNSURE`
+  plus one sentence). Each item was asked in **both reading orders** and in **two contexts**:
+  narrow (+-8 words) and wide (+-100 words). 120 calls per model, temperature 0.
+- **Models:** 11, all local (`koala/*`, `iguana/*`): nothing left the homelab. Run one iguana
+  model at a time, each unloaded afterwards, to avoid the memory pressure behind infra#273;
+  whisper was confirmed healthy afterwards.
+- **Pass criteria, fixed before the run:** wide context, strict accuracy >= 80% (an
+  unparseable or "unsure" answer counts as wrong), order-swap consistency >= 85%,
+  unparseable <= 10%.
+
+## Results (wide context, 60 calls per model)
+
+| Model | Correct | 95% CI | Same pick when order swapped | Unparseable | s/call | Pass |
+|---|---|---|---|---|---|---|
+| `iguana/gemma4-31b` | 58 = 97% | 89-99% | 93% | 0 | 10.4 | **yes** |
+| `iguana/gemma4-26b` | 55 = 92% | 82-96% | 93% | 0 | 2.0 | **yes** |
+| `iguana/ornith-35b` | 51 = 85% | 74-92% | 77% | 0 | 2.5 | no (consistency) |
+| `koala/phi4-14b` | 50 = 83% | 72-91% | 80% | 0 | 1.0 | no (consistency) |
+| `koala/qwen36-35b-a3b` | 49 = 82% | 70-89% | 83% | 0 | 3.3 | no (consistency) |
+| `iguana/granite4` | 47 = 78% | 66-87% | 83% | 0 | 4.6 | no |
+| `iguana/qwen3-14b` | 44 = 73% | 61-83% | 87% | 0 | 4.4 | no (accuracy) |
+| `iguana/gpt-oss-20b` | 40 = 67% | 54-77% | 100% | 17 | 11.2 | no (unparseable) |
+| `koala/phi4-mini` | 37 = 62% | 49-73% | 24% | 1 | 0.5 | no |
+| `iguana/gemma4-26b-think` | 23 = 38% | 27-51% | 100% | 37 | 15.3 | no (unparseable) |
+| `koala/qwen35-9b-fast` | 0 | n/a | n/a | 60 | 7.2 | no (no output) |
+
+Narrow context only (+-8 words) costs 6-12 points for the models that pass: gemma4-31b 92%,
+gemma4-26b 83%, ornith-35b 80%. Context is not decoration: Opus also used evidence from
+elsewhere in the recording (an amount stated earlier, a term that recurs).
+
+Position bias: the passing models pick "Reading 1" 47-51% of the time. `phi4-mini` picks it 15%
+and flips its answer when the order is swapped 76% of the time: it carries almost no signal.
+
+## Reading it
+
+- **Two models pass:** the gemma4 pair. Both clear the bar and neither shows position bias.
+  26b is ~5x faster; their intervals overlap, so 30 items cannot separate 92% from 97%.
+- **The thinking models were not tested fairly.** `gpt-oss-20b` and `gemma4-26b-think` spend
+  the 600-token budget reasoning and often never reach the answer line. When gemma4-26b-think
+  did answer, it was right 23 of 23; gpt-oss-20b 93% of the answered. A larger budget would
+  make them eligible; at 15 s a call it would not make them practical.
+
+## What this does not show
+
+- **Thirty items, one recording, one speaker, one domain.** The 30 were chosen by Opus as
+  decidable, so they overstate how easy the typical disagreement is; ~350 others are mostly
+  spelling, compounds and fillers, which no ground truth covers.
+- Opus agreeing 30/30 is not a measure of Opus over all 380 disagreements.
+- Accuracy was measured at temperature 0 on one prompt; no prompt tuning was tried.
+- The speaker's answers are the only truth; they are one person's recollection of what they said.
+
+## Why gemma4-31b is slow, and what was ruled out
+
+On iguana (M2 Ultra, 64 GB) the dense 31B runs at **103 tok/s prefill, 7.9 tok/s decode**; the
+26B MoE at 641 and 39.6. Ruled out as the cause: memory pressure (no swap, 45% free, model 100%
+on GPU), context length (32k vs 4k made no difference), and the **inference engine** — MLX 4-bit
+of the same model measured 76 tok/s prefill and 8.6 decode, the same within noise. (Terse
+output saves only ~3.6 s of ~9.3 s because prefill dominates.) Two unrelated engines at the same
+speed points at the machine: ~140 GB/s effective memory bandwidth on a part rated for 800, and
+**iguana had macOS Low Power Mode on** (`pmset` `lowpowermode 1`). A first MLX benchmark was
+wrong because `mlx_lm.server` caches prompts; the figures above are from the corrected run.
+
+**Confirmed 2026-10-02.** The owner turned Low Power Mode off (`lowpowermode 0`, read back over
+ssh) and `gemma4:31b` was re-measured with nothing else resident: **217 tok/s prefill, 22.3
+tok/s decode** (4 runs, spread 22.2-22.6), against 103 and 7.9 with it on, so 2.1x and 2.8x. The
+slowness was the power mode, not the engine or the model. Caveats on the comparison: the new run
+used a different, synthetic Swedish prompt (~700 words, the original items held real transcript
+text and were deleted) and the model answered in 26-43 tokens, not the ~200 of the first run, so
+it is a same-machine before/after, not a controlled A/B. `gemma4:26b` and the MLX figures were
+not re-measured. The 31B's 97% judge accuracy is therefore usable at roughly a third of the
+latency the sweep implied; the 26B-vs-31B choice in ADR-034 was made on the slow numbers and is
+reopened by this.

docs/research/reconcile-blind-check-2026-10-02.md

@@ -0,0 +1,68 @@
+# Two-engine reconcile, first check on a real recording (tapir#33, ADR-034, 2026-10-02)
+
+Aggregate numbers only. This repo is public and the recording is private (#31 custody is open), so no
+transcript text, no disputed words and no per-item reading appears here.
+
+## What was run
+
+One real 63.6-minute Teams recording of one speaker dictating project notes (the video track, camera
+off, was discarded). Build: PR #39 head `7f8e367`. Everything local: `iguana/pianissimo` as base,
+`iguana/kb-whisper` as the second engine, `iguana/gemma4-31b` as reconciler (iguana Low Power Mode
+off), cloud fallbacks off, a throwaway Postgres.
+
+| Measure | Value |
+|---|---|
+| Job time, upload to saved | 21 min 49 s for 63.6 min of audio (20.6 min per hour of audio) |
+| Differences between the engines | 362 |
+| Settled automatically (compound splits, up to two filler words) | 80 |
+| Decided by the model | 282 |
+| Of those: alternative reading used / base kept / uncertain | 130 / 151 / 1 |
+| Model call errors | 0 |
+
+20.6 min per hour is inside the 25-minute resolution trigger in ADR-034, so no smaller adjudicator is
+needed on speed grounds. The earlier estimate was about 10 min per hour on gemma4-26b; it was not
+measured and this run used the 31B.
+
+## Blind check
+
+30 spots drawn at random (fixed seed) from the 89 replacement spots that could be placed in the base
+text, across the whole recording (0:44 to 60:57). The two readings were shown as x and y in random
+order, with the key kept apart from the judge. Judge: Claude Opus, text only, no audio, at the
+speaker's request ("I trust Opus"). The speaker did not judge these items themselves.
+
+| Verdict | Items |
+|---|---|
+| The reconciler's reading (the alternative) is right | 21 |
+| The base reading (pianissimo) is right, so the change made it worse | 1 |
+| Neither reading is right | 3 |
+| Cannot tell from text (compound, hyphen and number-format differences) | 5 |
+
+- Right of the items with a verdict: 21 of 22, 95% (95% interval 78% to 99%).
+- Right of all 30, counting "neither" and "cannot tell" as not right: 21 of 30, 70% (52% to 83%).
+- Right of the 25 that were not "cannot tell": 21 of 25, 84% (65% to 94%).
+- Net effect against the base on these 30: 21 better, 1 worse, 8 unchanged in quality.
+
+## Reading the result
+
+ADR-034's bar was "right at least 85% of the time". It did not say how ties count, and that was a
+flaw in the bar, not in the data. The literal reading (all 30) gives 70% and **does not meet** the bar;
+counting only decided items gives 95% and meets it. Neither is a hidden pass or a hidden failure; both
+are reported. What is not in doubt: on a random sample the reconciler improved the text far more often
+than it harmed it (21 to 1), and the 8 unchanged cases are mostly spellings the text cannot settle.
+
+## What this did not measure
+
+- **Better single engine.** The second ADR-034 condition, "no worse than the better single engine", is
+  **not answered**. Only changed spots are stored, not the 151 where the model kept pianissimo, so the
+  reconciler cannot be compared with kb-whisper alone. The 21-to-1 result is against pianissimo.
+- **The speaker as judge.** Opus judged from text. It cannot hear the audio, so it can be wrong where
+  the right word is audible but not inferable (items such as a mis-heard number or a name). Earlier,
+  Opus and the speaker agreed 30 of 30, but on items chosen as decidable; that does not transfer to a
+  random sample.
+- **Insertions and deletions.** 42 of the 131 stored spots were not sampled: 39 are an added or
+  removed word, which cannot be placed in the base text, and 3 are the uncertain spot or could not be
+  located. The 39 are mostly small words ("och", "det"), so this sample says little about them.
+- **One recording, one speaker, 30 items.** Wide intervals, and neighbouring spots (three pairs within
+  seconds of each other) are not independent.
+- **Segments keep the base text.** The saved timed segments hold pianissimo's wording, not the merged
+  text; only the stored full text is merged. Found while building this sample; not yet an issue.

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

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

docs/sketches/a-tui-panels.html

@@ -0,0 +1,63 @@
+<!DOCTYPE html><html lang="en"><head><meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>tapir — A · TUI panels</title>
+<style>
+:root{--bg:#0d0d12;--panel:#14141c;--line:#2a2a3a;--mint:#0EF9B6;--pink:#F740A0;--purple:#7653fc;--cream:#F5E9D6;--dim:#6b6b86;--text:#d9d9e6}
+*{box-sizing:border-box}
+body{margin:0;background:var(--bg);color:var(--text);font:14px/1.5 ui-monospace,SFMono-Regular,Menlo,"Cascadia Code",monospace}
+.wrap{max-width:760px;margin:0 auto;padding:18px 14px 60px}
+header{display:flex;align-items:center;gap:10px;border-bottom:1px solid var(--line);padding-bottom:12px;margin-bottom:18px}
+.brand{color:var(--mint);font-weight:700;letter-spacing:.5px}
+.brand b{color:var(--cream)}
+.tip{margin-left:auto;color:var(--dim);font-size:12px}
+.note{color:var(--dim);font-size:12.5px;border-left:2px solid var(--purple);padding:6px 10px;margin:0 0 16px;background:#11111a}
+/* lipgloss-style bordered panel */
+.card{border:1px solid var(--line);border-radius:8px;background:var(--panel);padding:12px 14px;margin:0 0 12px;position:relative}
+.card::before{content:"";position:absolute;left:0;top:10px;bottom:10px;width:3px;border-radius:3px;background:var(--purple)}
+.card.ready::before{background:var(--mint)}
+.title{color:var(--cream);font-size:15px;font-weight:600;margin:0 0 4px}
+.meta{color:var(--dim);font-size:12px}
+.chip{display:inline-block;border:1px solid var(--mint);color:var(--mint);border-radius:4px;padding:0 6px;font-size:11px;margin-left:6px}
+.chip.q{border-color:var(--pink);color:var(--pink)}
+.preview{color:var(--dim);margin:6px 0 0;font-size:13px}
+/* expanded */
+.exp{border-color:var(--purple)}
+.exp .head{display:flex;justify-content:space-between;align-items:baseline}
+.collapse{color:var(--mint);font-size:12px;text-decoration:none;border:1px solid var(--line);border-radius:4px;padding:1px 7px}
+.sec h2{color:var(--mint);font-size:12px;text-transform:uppercase;letter-spacing:1px;margin:16px 0 6px;border-bottom:1px dashed var(--line);padding-bottom:3px}
+.sec ul{margin:0;padding-left:18px}.sec li{margin:3px 0}
+.body{color:var(--text)}
+.dock{margin-top:16px;border-top:1px solid var(--line);padding-top:12px}
+.ask{display:inline-block;background:linear-gradient(90deg,var(--purple),var(--pink));color:#fff;border:0;border-radius:6px;padding:7px 12px;font:inherit;font-size:13px;cursor:pointer}
+pre.tapir{margin:0;color:var(--mint);font-size:10px;line-height:1.05}
+</style></head><body><div class="wrap">
+<header>
+  <pre class="tapir">  ▄█▓▓█▄  ∩
+ █▓( ◕ ◕ )▓█──┘</pre>
+  <span class="brand">◆ <b>tapir</b> · watch less, know more</span>
+  <span class="tip">261 in queue</span>
+</header>
+<p class="note">Tapir fetches captions slowly on purpose, to respect YouTube's limits — new summaries land gradually.</p>
+
+<div class="card ready"><div class="title">How the Attention Economy Rewires Your Brain</div>
+<div class="meta">youtube · 2026-06-11 · 18 min <span class="chip">ready</span></div>
+<div class="preview">A tour of the incentive loops behind infinite feeds and three concrete ways to claw back focus…</div></div>
+
+<article class="card exp ready">
+  <div class="head"><div class="title">Postgres 18 — What's Actually New</div><a class="collapse" href="#">collapse ↑</a></div>
+  <div class="meta">youtube · 2026-06-10 · 42 min · phi4-mini</div>
+  <div class="sec"><h2>Takeaways</h2><ul>
+    <li>Async I/O cuts cold-cache read latency materially on NVMe.</li>
+    <li>Skip-scan makes more multicolumn indexes usable without rewrites.</li>
+    <li>Upgrade path is smooth; test the new planner stats first.</li></ul></div>
+  <div class="sec"><h2>Highlights</h2><ul>
+    <li>Async I/O subsystem (effective_io_concurrency now matters more).</li>
+    <li>B-tree skip scan for leading-column gaps.</li>
+    <li>Better partition-wise joins.</li></ul></div>
+  <div class="sec"><h2>Summary</h2><p class="body">Postgres 18 is an incremental but meaningful release: the headline is the new asynchronous I/O path, with skip-scan and planner improvements close behind. For most homelabs the upgrade is low-risk and worth it for the read-latency wins.</p></div>
+  <div class="dock"><button class="ask">Dig deeper — ask about this video →</button></div>
+</article>
+
+<div class="card"><div class="title">RAG is dead, right?? — Conference Talk</div>
+<div class="meta">youtube · 2026-06-09 · 31 min <span class="chip q">queued</span></div></div>
+</div></body></html>

docs/sketches/b-reader-charm.html

@@ -0,0 +1,62 @@
+<!DOCTYPE html><html lang="en"><head><meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>tapir — B · reader + charm accents</title>
+<style>
+:root{--bg:#faf7f2;--card:#fff;--ink:#1c1b22;--soft:#6a6878;--line:#e7e2d8;--mint:#0bbf8c;--purple:#6a4cf0;--pink:#e0379a}
+*{box-sizing:border-box}
+body{margin:0;background:var(--bg);color:var(--ink);font:16px/1.6 -apple-system,BlinkMacSystemFont,"Segoe UI",Inter,sans-serif}
+.mono{font-family:ui-monospace,SFMono-Regular,Menlo,monospace}
+.wrap{max-width:680px;margin:0 auto;padding:20px 16px 60px}
+header{display:flex;align-items:center;gap:10px;margin-bottom:18px}
+.brand{font-weight:800;font-size:18px;letter-spacing:-.3px}
+.brand .dot{color:var(--mint)}
+.tag{color:var(--soft);font-size:13px}
+.tip{margin-left:auto;color:var(--soft);font-size:12px}
+.note{color:var(--soft);font-size:13px;background:#fff;border:1px solid var(--line);border-left:3px solid var(--mint);border-radius:8px;padding:8px 12px;margin:0 0 18px}
+.card{background:var(--card);border:1px solid var(--line);border-radius:12px;padding:16px 18px;margin:0 0 14px;box-shadow:0 1px 2px rgba(20,18,40,.04)}
+.title{font-size:18px;font-weight:700;letter-spacing:-.2px;margin:0 0 4px;line-height:1.3}
+.meta{color:var(--soft);font-size:13px}
+.meta .mono{font-size:12.5px}
+.chip{display:inline-block;background:rgba(11,191,140,.12);color:var(--mint);border-radius:999px;padding:1px 9px;font-size:12px;font-weight:600;margin-left:6px}
+.chip.q{background:rgba(224,55,154,.12);color:var(--pink)}
+.preview{color:var(--soft);margin:8px 0 0}
+.exp{border-color:#d9d0ee;box-shadow:0 6px 24px rgba(106,76,240,.10)}
+.head{display:flex;justify-content:space-between;align-items:baseline;gap:10px}
+.collapse{color:var(--purple);font-size:13px;font-weight:600;text-decoration:none;white-space:nowrap}
+.sec h2{font-size:12px;text-transform:uppercase;letter-spacing:1.2px;color:var(--purple);margin:18px 0 6px}
+.sec ul{margin:0;padding-left:20px}.sec li{margin:5px 0}
+.body{line-height:1.7}
+.dock{margin-top:18px;border-top:1px solid var(--line);padding-top:14px}
+.ask{display:inline-flex;align-items:center;gap:6px;background:var(--ink);color:#fff;border:0;border-radius:999px;padding:9px 16px;font:inherit;font-size:14px;font-weight:600;cursor:pointer}
+pre.tapir{margin:0;color:var(--mint);font-size:10px;line-height:1.05}
+</style></head><body><div class="wrap">
+<header>
+  <pre class="tapir">  ▄█▓▓█▄  ∩
+ █▓( ◕ ◕ )▓█──┘</pre>
+  <span class="brand">tapir<span class="dot">.</span></span><span class="tag">watch less, know more</span>
+  <span class="tip mono">261 in queue</span>
+</header>
+<p class="note">Tapir fetches captions slowly on purpose, to respect YouTube's limits — new summaries land gradually.</p>
+
+<div class="card"><div class="title">How the Attention Economy Rewires Your Brain</div>
+<div class="meta mono">youtube · 2026-06-11 · 18 min <span class="chip">ready</span></div>
+<div class="preview">A tour of the incentive loops behind infinite feeds and three concrete ways to claw back focus…</div></div>
+
+<article class="card exp">
+  <div class="head"><div class="title">Postgres 18 — What's Actually New</div><a class="collapse" href="#">collapse ↑</a></div>
+  <div class="meta mono">youtube · 2026-06-10 · 42 min · phi4-mini</div>
+  <div class="sec"><h2>Takeaways</h2><ul>
+    <li>Async I/O cuts cold-cache read latency materially on NVMe.</li>
+    <li>Skip-scan makes more multicolumn indexes usable without rewrites.</li>
+    <li>Upgrade path is smooth; test the new planner stats first.</li></ul></div>
+  <div class="sec"><h2>Highlights</h2><ul>
+    <li>Async I/O subsystem (effective_io_concurrency now matters more).</li>
+    <li>B-tree skip scan for leading-column gaps.</li>
+    <li>Better partition-wise joins.</li></ul></div>
+  <div class="sec"><h2>Summary</h2><p class="body">Postgres 18 is an incremental but meaningful release: the headline is the new asynchronous I/O path, with skip-scan and planner improvements close behind. For most homelabs the upgrade is low-risk and worth it for the read-latency wins.</p></div>
+  <div class="dock"><button class="ask">✦ Ask about this video</button></div>
+</article>
+
+<div class="card"><div class="title">RAG is dead, right?? — Conference Talk</div>
+<div class="meta mono">youtube · 2026-06-09 · 31 min <span class="chip q">queued</span></div></div>
+</div></body></html>

docs/sketches/c-cozy-terminal.html

@@ -0,0 +1,65 @@
+<!DOCTYPE html><html lang="en"><head><meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>tapir — C · cozy terminal</title>
+<style>
+:root{--bg:#16131d;--card:#1e1a28;--ink:#ece7f5;--soft:#9a92b4;--line:#322b44;--mint:#2ee6b6;--purple:#9d7bff;--pink:#ff6bdb;--cream:#f3ead8}
+*{box-sizing:border-box}
+body{margin:0;background:radial-gradient(1200px 600px at 70% -10%,#231b33,transparent),var(--bg);color:var(--ink);font:15px/1.6 -apple-system,BlinkMacSystemFont,"Segoe UI",Inter,sans-serif}
+.mono{font-family:ui-monospace,SFMono-Regular,Menlo,monospace}
+.wrap{max-width:700px;margin:0 auto;padding:20px 16px 60px}
+header{display:flex;align-items:center;gap:12px;margin-bottom:18px}
+.brand{font-weight:800;font-size:18px}.brand .c{color:var(--mint)}
+.tag{color:var(--soft);font-size:13px}
+.tip{margin-left:auto;color:var(--soft);font-size:12px}
+.note{color:var(--soft);font-size:13px;background:#1b1726;border:1px solid var(--line);border-radius:10px;padding:9px 12px;margin:0 0 18px}
+.note b{color:var(--mint);font-weight:600}
+.card{position:relative;background:var(--card);border:1px solid var(--line);border-radius:12px;padding:14px 16px 14px 18px;margin:0 0 13px;overflow:hidden}
+.card::before{content:"";position:absolute;left:0;top:0;bottom:0;width:4px;background:var(--purple)}
+.card.ready::before{background:linear-gradient(var(--mint),var(--purple))}
+.title{font-size:16px;font-weight:700;margin:0 0 4px;color:var(--cream)}
+.meta{color:var(--soft);font-size:12.5px}
+.chip{display:inline-block;border-radius:999px;padding:1px 9px;font-size:11px;font-weight:600;margin-left:6px;background:rgba(46,230,182,.14);color:var(--mint)}
+.chip.q{background:rgba(157,123,255,.18);color:var(--purple)}
+.preview{color:var(--soft);margin:7px 0 0;font-size:14px}
+.exp{border-color:#473a63;box-shadow:0 10px 40px rgba(0,0,0,.35)}
+.head{display:flex;justify-content:space-between;align-items:baseline;gap:10px}
+.collapse{color:var(--mint);font-size:12px;text-decoration:none;border:1px solid var(--line);border-radius:6px;padding:2px 8px;white-space:nowrap}
+.sec h2{font-size:11px;text-transform:uppercase;letter-spacing:1.3px;color:var(--mint);margin:18px 0 7px;display:flex;align-items:center;gap:8px}
+.sec h2::after{content:"";flex:1;height:1px;background:var(--line)}
+.sec ul{margin:0;padding-left:18px}.sec li{margin:5px 0}
+.body{line-height:1.7;color:var(--ink)}
+.dock{margin-top:18px;border-top:1px dashed var(--line);padding-top:14px;display:flex;align-items:center;gap:10px}
+.ask{display:inline-flex;align-items:center;gap:7px;background:linear-gradient(90deg,var(--purple),var(--mint));color:#10101a;border:0;border-radius:999px;padding:9px 16px;font:inherit;font-weight:700;font-size:14px;cursor:pointer}
+.dockhint{color:var(--soft);font-size:12px}
+pre.tapir{margin:0;color:var(--mint);font-size:11px;line-height:1.05}
+</style></head><body><div class="wrap">
+<header>
+  <pre class="tapir">  ▄█▓▓█▄  ∩
+ █▓( ◕ ◕ )▓█──┘</pre>
+  <span class="brand">tapir<span class="c">_</span></span><span class="tag">watch less, know more</span>
+  <span class="tip mono">261 in queue</span>
+</header>
+<p class="note">Tapir fetches captions slowly on purpose, to respect YouTube's limits — <b>new summaries land gradually</b>.</p>
+
+<div class="card ready"><div class="title">How the Attention Economy Rewires Your Brain</div>
+<div class="meta mono">youtube · 2026-06-11 · 18 min <span class="chip">ready</span></div>
+<div class="preview">A tour of the incentive loops behind infinite feeds and three concrete ways to claw back focus…</div></div>
+
+<article class="card exp ready">
+  <div class="head"><div class="title">Postgres 18 — What's Actually New</div><a class="collapse" href="#">collapse ↑</a></div>
+  <div class="meta mono">youtube · 2026-06-10 · 42 min · phi4-mini</div>
+  <div class="sec"><h2>Takeaways</h2><ul>
+    <li>Async I/O cuts cold-cache read latency materially on NVMe.</li>
+    <li>Skip-scan makes more multicolumn indexes usable without rewrites.</li>
+    <li>Upgrade path is smooth; test the new planner stats first.</li></ul></div>
+  <div class="sec"><h2>Highlights</h2><ul>
+    <li>Async I/O subsystem (effective_io_concurrency now matters more).</li>
+    <li>B-tree skip scan for leading-column gaps.</li>
+    <li>Better partition-wise joins.</li></ul></div>
+  <div class="sec"><h2>Summary</h2><p class="body">Postgres 18 is an incremental but meaningful release: the headline is the new asynchronous I/O path, with skip-scan and planner improvements close behind. For most homelabs the upgrade is low-risk and worth it for the read-latency wins.</p></div>
+  <div class="dock"><button class="ask">◆ Ask about this video</button><span class="dockhint mono">answers only from this video's transcript</span></div>
+</article>
+
+<div class="card"><div class="title">RAG is dead, right?? — Conference Talk</div>
+<div class="meta mono">youtube · 2026-06-09 · 31 min <span class="chip q">queued</span></div></div>
+</div></body></html>

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

@@ -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)).

docs/ui-spec.md

@@ -179,4 +179,5 @@ distinguishable.
 | **"Summarize now" foreground path** | Unified quiet nudge button on actionable non-summarized cards. Five explicit card states — (1) summarized: chip + no button; (2) no captions (`transcript_status = 'none'`): "No transcript available", no button; (3) queued: "Queued" chip, no button; (4) rate-limited: "Fetching soon…" + "Summarize now" → `POST /v/{id}/retry-now` (clears `rate_limited_at`, triggers engine); (5) pending: "Not summarized" + "Summarize now" → `POST /v/{id}/summarize` (queues + triggers engine). One verb, one style (`.btn-quiet`); backend difference invisible to user. Both handlers call `ProcessVideo` through `globalFetchGate`. Rate gate respected, not bypassed — this is onboarding prioritisation. | Fast onboarding value; honest dead-end for no-captions videos (no button that fails). | `internal/web/handlers.go` (`handleRetryNow`, `handleRequestSummarize`); `internal/web/views.templ` (`VideoCard`) |
 | **Pipeline stats bar** | A one-line status bar above the video list: `N summarized · M fetching soon · K no captions`. Computed from the unfiltered row set; hidden when all videos are summarized. Gives the user a clear read on pipeline state without any interaction. | Replaces the "why is nothing happening?" confusion when most videos are pending or rate-limited. | `internal/web/view.go` (`PipelineStats`, `pipelineStats`) |
 | **Unavailable channels (account page)** | The `/account` page shows a "Unavailable channels" section when any channels returned HTTP 404 on the last discovery pass. Lists channel name, an "unavailable" badge, and the first-seen date. Data sourced from the `channel_errors` table (migration 013). | Surfaces silent failures so users know why some subscribed channels produce no new videos. | migration 013; `internal/web/account.go`; `internal/adapters/youtube/youtube.go` (`domain.ErrChannelUnavailable`) |
+| **Visual refresh — charm-reader theme + light/dark toggle (ADR-032)** | One layout in two palettes expressed as CSS custom properties: a warm "reader" light theme (sketch B) and a "cozy terminal" dark theme (sketch C). Palette is chosen in cascade order — `:root` light default, an OS-preference dark block scoped to `:root:not([data-theme])` so it only applies absent an explicit choice, and `:root[data-theme="dark"\|"light"]` set by a header toggle that outranks the media query and persists in `localStorage` (guarded; degrades to OS default). An init script in `<head>` applies the stored choice before paint (no flash). Charm touches: monospace meta lines, accent uppercase section dividers with a trailing rule, pill buttons, a lifted/accent-edged expanded card. Error/danger shades became `--err-*` tokens so they follow the theme without per-block dark overrides. Sketches kept as the design record under `docs/sketches/`. | The UI read "flat and boring"; the charm/TUI aesthetic makes it distinctive and gives a real light/dark choice rather than OS-only. | ADR-032; `docs/use-cases/visual_theme.feature`; `internal/web/visual_theme_test.go`; `internal/web/view.go` (`stylesheet`, `themeScript`), `internal/web/views.templ` (Layout/PublicLayout) |
 | **Recency window + sparse-state honesty (ADR-020)** | Supersedes the copy/sort in the rows above. Auto-summarize is bounded to videos published within `TAPIR_AUTO_SUMMARIZE_WINDOW` (~7d); older un-summarized videos collapse behind a single "Show N older videos — summarize on demand" disclosure, and caption-less videos collapse to a one-line count (not N cards). List order is now `summarized-first, published_at DESC NULLS LAST`. Copy reframed for honest scarcity: pipeline bar reads "N ready · M in queue · K no captions" (no "fetching soon"); a gradual-fill note explains the rate limit; the nudge verb is "Summarize" (not "Summarize now"); the queued card says "summarizing shortly"; the empty-connected state drops the impossible `tapir run` instruction. Detail leads with Takeaways. Filters slimmed (no date pickers; hidden when empty); watched/skipped segmented; back link on detail; empty terms checkbox removed. | Make the sparse reality legible and honest instead of implying abundance/imminence; bound auto load so the back-catalogue doesn't re-drive the caption gate. Never fetch harder — scarcity is surfaced, not engineered around. | ADR-020; `2384c47`, `3df0459`, `40b703e`, `a1a5217`, `4a0a56e`, `9bf1c31`, `980638d`, `12fb031`, `f775441`, `51aa5d9` |

docs/use-cases/audio_upload.feature

@@ -0,0 +1,293 @@
+Feature: Upload an audio recording, get a private Swedish transcript, discuss it
+  As someone with a recorded meeting on my computer
+  I want to upload it and then ask questions about what was said
+  So that I can find out what was decided, what is open and who owes what
+
+  # ADR-033, tapir#33. The load-bearing constraint: an uploaded transcript is
+  # PRIVATE content, so it lives in its own RLS-scoped store and never in the
+  # shared public-caption `transcripts` table (ADR-021). Rejection is synchronous,
+  # the audio is never retained, and a failed job ends in a terminal state with a
+  # message — never a spinner (ADR-025). A file with a video track (a Teams meeting
+  # with the camera off is one) is accepted: only its audio is used and the user is
+  # told. Video playback, keyframes and analysis of what is shown remain #28.
+
+  Background:
+    Given I am a registered user with audio upload enabled
+
+  Scenario: The upload form states the limits up front
+    When I open the upload page
+    Then I see the size limit and the length limit before I choose a file
+
+  Scenario: An accepted recording is queued and I am taken to its page
+    When I upload an audio file within the limits
+    Then it is recorded as queued for me
+    And I am redirected to its page
+
+  Scenario: A recording with no title is named after its file
+    When I upload an audio file and give no title
+    Then its title is the filename without its extension
+
+  Scenario: A file with a video track is accepted and I am told the video was ignored
+    When I upload a recording that also has a video track
+    Then it is accepted and queued like any other
+    And its page says only its audio was used and the video was discarded
+    And it says so while queued, while transcribing, when done and when failed
+
+  Scenario: A video file with no audio is rejected for having no audio
+    When I upload a video that has no audio track
+    Then it is rejected with a message that it has no audio
+    And nothing is recorded
+
+  Scenario: A recording over the length limit is rejected with both lengths
+    When I upload a recording longer than the limit
+    Then it is rejected stating the recording's length and the limit
+    And nothing is recorded
+
+  Scenario: A file over the size limit is rejected as too large
+    When I upload a file larger than the size limit
+    Then it is rejected with 413 and a message about the limit
+    And nothing is recorded
+
+  Scenario: Cover art on an audio file is not a video track
+    When I upload an mp3 whose only video stream is attached cover art
+    Then it is treated as audio and no video notice is shown
+
+  Scenario: A file that is not media is rejected as unreadable
+    When I upload a file ffprobe cannot read
+    Then it is rejected as the file's fault
+
+  Scenario: A broken ffprobe is reported as our fault, not as a bad file
+    When the prober itself fails
+    Then the upload is an error rather than a rejection of my file
+
+  Scenario: A recording becomes a transcript and the audio is deleted
+    Given I have uploaded a recording
+    When it has been transcribed
+    Then the transcript is stored privately for me
+    And the audio is deleted
+
+  Scenario: The endpoint that served a transcript is recorded and shown
+    Given a recording was transcribed by the fallback endpoint
+    When I view its page
+    Then it names the fallback endpoint, because that means the audio left the homelab
+
+  Scenario: The local endpoint is tried first and the fallback only on failure
+    When the primary transcription endpoint fails for any reason
+    Then the fallback endpoint transcribes it
+    But when the primary succeeds the fallback is never contacted
+
+  Scenario: A failed transcription ends in a message, never a spinner
+    When every transcription endpoint fails
+    Then the upload is failed with a typed reason
+    And its page shows that reason and stops polling
+
+  Scenario: A restart fails the jobs it interrupted
+    Given the service restarted while a recording was queued or transcribing
+    When it starts again
+    Then those uploads are failed as interrupted
+    And any audio left in the spool is deleted
+
+  Scenario: Only one recording is transcribed at a time
+    When several recordings are queued
+    Then they are transcribed one after another
+
+  Scenario: My uploads are invisible to other users
+    Given another user has uploaded a recording
+    When I ask for it by its id
+    Then I get 404 and no part of its title or text
+    And the chat for it calls no model
+
+  Scenario: I can ask about my recording and get a grounded answer
+    Given my recording has been transcribed
+    When I ask a question about it
+    Then the answer comes from my transcript through the same chat as a video's
+
+  Scenario: A long recording is not cut to its first third
+    Given my recording's transcript is longer than the caption default
+    When I ask a question about it
+    Then the transcript is sent with the upload budget, not the 18,000-character caption default
+
+  # ADR-034, two engines and a local reconciler (off unless TAPIR_RECONCILE_BASE_MODEL is set)
+
+  Scenario: Two engines transcribe the audio and their disagreements are reconciled
+    Given two-engine transcription is enabled
+    When a recording is transcribed
+    Then both engines run on the same audio, the fast one first
+    And a local model chooses between their disagreements
+    And the endpoint names every engine and the model that took part
+
+  Scenario: Trivial disagreements never reach the model
+    When the engines differ only by a compound written as one word or two, or by a filler word
+    Then the base reading is kept without a model call
+
+  Scenario: Differences close together are one decision
+    When two differences are separated by a single shared word
+    Then they are asked about as one dispute
+
+  Scenario: An unsure or changed reading is flagged for me
+    When the model changes a reading or cannot decide
+    Then the page lists the spot with a minute mark, both readings and which was used
+    And an unsure spot keeps the base reading and is marked uncertain
+
+  Scenario: The model cannot be a cloud model
+    When the reconciliation model or base engine is not a koala/ or iguana/ alias
+    Then the service refuses to start
+    And a cloud alias that merely contains a local host name is refused too
+
+  Scenario: One engine failing does not fail the upload
+    When one of the two engines fails
+    Then the other engine's transcript is used
+    And the page says only one engine produced it and why
+    And the cloud fallback is not used
+
+  Scenario: Only when both local engines fail does the cloud fallback run
+    When both local engines fail and a fallback is configured
+    Then the fallback transcribes it and the row records that it did
+
+  Scenario: A model that is down gives an unreconciled transcript, not a failed job
+    When the reconciliation model fails three times in a row
+    Then the base transcript is kept unreconciled with the reason
+    And the page says it could not be reconciled
+
+  Scenario: A cancelled job never reaches the cloud
+    When the job is cancelled during either engine
+    Then neither the second engine nor the cloud fallback is called
+
+  Scenario: A single-engine transcript claims no second opinion
+    When only one engine was configured
+    Then the page does not mention reconciliation
+
+  # ADR-035: progress, the log, and the judged spots
+
+  Scenario: I can see which phase my upload is in
+    Given a recording is being transcribed
+    When I open its page
+    Then it lists every phase in order, the finished ones with how long each took
+    And the running phase is marked, with how long it has run
+    And the page updates in place every two seconds until the job ends
+
+  Scenario: A bar appears only where the count is real
+    When the deciding phase has decided 142 of 282 differences
+    Then the page shows a bar with "142 of 282"
+    And an engine call, which is one request with nothing to count, never gets a bar or a percentage
+
+  Scenario: A slow engine call is not mistaken for a dead service
+    When an engine call has run for 25 minutes and the service is still reporting
+    Then the page shows the elapsed time and says nothing is wrong
+
+  Scenario: A service that stopped reporting is called lost
+    When nothing has been heard from the service for longer than the heartbeat allows
+    Then the page says the service stopped reporting and may have restarted or crashed
+    And the running phase carries a warning mark and no spinner
+    And the page keeps checking
+
+  Scenario: A model that stops advancing is called quiet
+    When the deciding phase has not advanced for five minutes and the service is still reporting
+    Then the page says nothing new has happened for that long and that the service is still running
+    And it does not say the service is lost
+
+  Scenario: A failed job shows where it stopped
+    When a job fails
+    Then the failed phase is marked and what never started stays pending
+    And the page stops checking
+
+  Scenario: A finished job keeps its record of what was done
+    When a job is done
+    Then the page keeps a collapsed list of the phases with how long each took
+
+  Scenario: The heartbeat runs on its own while an engine call is stuck
+    When an engine call does not return
+    Then the row is still stamped alive every few seconds
+    And no heartbeat outlives its job
+
+  Scenario: The log names every phase and never the transcript
+    When a job runs
+    Then the log has a line per phase in the order it ran, each naming the upload
+    And no line contains transcript text or the spool path
+
+  Scenario: Every judged spot is stored but only the ones to check are listed
+    When the model has changed, kept and been unsure about spots
+    Then every one is stored with its verdict and the words before it
+    And the page lists only the changed and uncertain ones, with the words before each
+
+  Scenario: The saved segments carry the merged words
+    When the model changes a word
+    Then the segment containing it carries the reconciled word
+    And the segment's timing is unchanged
+
+  # ADR-035 decision 8: hardening from the adversarial review
+
+  Scenario: Chat about a recording is offered with local models only
+    When I open the chat for my recording
+    Then the model list holds only models that run on the homelab
+    And a cloud model the shared chat service knows is not in it
+
+  Scenario: A forged cloud model is never sent the transcript
+    When a chat request names a cloud model
+    Then a local model answers instead, and nothing was sent to the cloud one
+
+  Scenario: A wedged engine does not use up the whole job
+    When the primary engine never answers
+    Then it fails at its own deadline while the job is still alive
+    And the fallback runs, unless the job itself was cancelled or timed out
+
+  Scenario: A file that lies about its length is refused
+    When a file claims to be short but converts to more than the limit
+    Then the job fails with a plain explanation and no engine receives the audio
+
+  Scenario: A recording much shorter than its file claimed is transcribed and noted
+    When a file converts to under 80 percent of the length it claimed and more than two minutes short
+    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
+
+  Scenario: One user cannot fill the queue for everyone
+    When one user has several recordings waiting
+    Then further uploads from that user are refused and other users are not affected
+
+  Scenario: Errors are shown as kinds, never their text
+    When a model or an engine fails
+    Then the log, the stored reason and the page carry a kind such as timeout or http_502
+    And never the provider's response, a file path or a word from the transcript
+
+  Scenario: Input that cannot be right is a client error
+    When the id is not an id, or a header or title holds control bytes
+    Then the answer is a 4xx or the title is cleaned, never a server error
+
+  Scenario: I choose the language of a recording
+    When I upload a recording and choose Swedish or Other on the form
+    Then the recording is transcribed as that language and its page says which
+
+  Scenario: Swedish stays the default
+    When an upload names no language
+    Then it is transcribed as Swedish, as before the choice existed
+
+  Scenario: A language outside the list is refused
+    When an upload names a language the form does not offer
+    Then it is answered 400 before anything is stored or sent to an engine
+
+  Scenario: A recording in another language never reaches the Swedish-only engine
+    When I choose Other for a recording
+    Then the language is left for the engine to detect and the Swedish-only base engine and the reconcile are not used
+    And the engine that does run is the configured one, which is not tuned for other languages

docs/use-cases/inline_expand.feature

@@ -0,0 +1,37 @@
+Feature: Inline-expand summary + Q&A in the list (ADR-031, #16)
+  As a reader skimming my summaries
+  I want to open a summary and its Q&A in place in the list
+  So that I get the full read and follow-up without leaving the list (SPA-like, no page hop)
+
+  # HTMX inline-expand, no SPA framework (ADR-031). Each scenario maps to a Go test
+  # in scenario_coverage_test.go (the BDD name-coverage gate).
+
+  Scenario: A summarized card expands to the full summary in place
+    Given a summarized video in my list
+    When I expand its card
+    Then the full summary, highlights, and takeaways are returned as an in-place card fragment, not a full page
+
+  Scenario: An expanded card collapses back to the compact card
+    Given an expanded card
+    When I collapse it
+    Then the compact card fragment is returned in its place
+
+  Scenario: The expanded card offers the Q&A dock
+    Given chat is enabled
+    When a summarized card is expanded
+    Then the expanded card includes the deeper-dive chat affordance for that video
+
+  Scenario: Only a summarized card offers expand
+    Given a discovered but not-yet-summarized card
+    When the card is rendered
+    Then it shows its summarize/queue footer and no expand affordance
+
+  Scenario: With JS off the card still reaches the full summary
+    Given a summarized card
+    When it is rendered
+    Then its expand affordance carries an href to the detail page as a no-JS fallback
+
+  Scenario: The detail page and the expanded card show the same summary
+    Given a summarized video
+    When I view it on the detail page and as an expanded card
+    Then both render the same summary body (one shared fragment, no drift)

docs/use-cases/observability.feature

@@ -0,0 +1,46 @@
+Feature: Observability — timing and metrics for performance and UX (ADR-030, #15)
+  As the maintainer running Tapir for pilot users
+  I want timing and Prometheus metrics for the activities that drive performance and UX
+  So that I can see latency, model behaviour, and usage — and feed the Stage-0 eval gate
+
+  # AI metrics are the priority (ADR-030 R3). Each scenario maps to a Go test in
+  # test/acceptance/scenario_coverage_test.go (the BDD name-coverage gate).
+
+  Scenario: Summarization latency is recorded per endpoint
+    Given the summarizer runs a transcript through its endpoint chain
+    When an endpoint returns a parseable summary
+    Then the summarize latency is recorded with the model, outcome "success", and whether it was a fallback
+
+  Scenario: A failing summarizer endpoint records its failure outcome
+    Given the summarizer runs a transcript through its endpoint chain
+    When an endpoint errors or returns unparseable output
+    Then the summarize latency is recorded with outcome "error" or "parse_error" before the chain advances
+
+  Scenario: Caption fetch latency is recorded by outcome
+    Given a caption fetch is attempted for a video
+    When it resolves to captions, no captions, or a rate limit
+    Then the caption-fetch latency is recorded labelled by that outcome
+
+  Scenario: LLM token usage is recorded from the completion
+    Given an LLM completion returns a usage block with prompt and completion tokens
+    When the client finishes the call
+    Then the prompt and completion tokens are recorded for that model
+
+  Scenario: Q&A answer latency is recorded
+    Given a user asks a question about a video
+    When the answer is produced from the stored transcript
+    Then the chat answer latency is recorded for the answering model
+
+  Scenario: HTTP requests are counted by route, method, and status
+    Given the metrics HTTP middleware wraps the app
+    When a request is served against a registered route
+    Then it is counted and timed under the bounded route pattern, not the raw path
+
+  Scenario: A successful login is counted
+    Given a user completes the OIDC callback and a session is established
+    Then the login counter is incremented
+
+  Scenario: The metrics endpoint is not on the public app port
+    Given the service is running
+    When the public app mux is inspected
+    Then it exposes no /metrics route — metrics are served on the dedicated metrics port only

docs/use-cases/visual_theme.feature

@@ -0,0 +1,29 @@
+Feature: Visual refresh — one charm-reader layout, light + dark themes (ADR-032, #17)
+  As a reader who likes a TUI/charm aesthetic
+  I want a fresh look with a light and a dark theme
+  So that the app feels distinctive and reads well in either mode
+
+  # Direction B (light) + C (dark) are one layout in two palettes (CSS variables),
+  # plus a persisted toggle and an OS-preference default. Colours are reviewed via
+  # the mockups in docs/sketches/, not unit-tested; these scenarios cover the
+  # testable structure. Each maps to a Go test in scenario_coverage_test.go.
+
+  Scenario: Light and dark themes share one layout via CSS variables
+    Given the app stylesheet
+    When it is rendered
+    Then it defines a light palette on the root and a dark palette under data-theme="dark", with no duplicate markup
+
+  Scenario: Without a stored choice the theme follows the OS preference
+    Given a visitor with no saved theme
+    When the page loads
+    Then a prefers-color-scheme dark block applies the dark palette automatically
+
+  Scenario: A persisted toggle switches light and dark
+    Given any page
+    When it is rendered
+    Then it includes a theme-toggle control and a small script that flips data-theme and persists the choice
+
+  Scenario: The expanded card embeds the video player
+    Given a summarized video with a valid provider id
+    When its card is expanded
+    Then the expanded card includes the embedded video player

go.mod

@@ -1,4 +1,4 @@
-module gitea.d-ma.be/mathias/tapir
+module git.d-ma.be/mathias/tapir
 
 go 1.26.1
 
@@ -9,23 +9,33 @@ require (
 	github.com/go-jose/go-jose/v4 v4.1.4
 	github.com/golang-migrate/migrate/v4 v4.19.1
 	github.com/jackc/pgx/v5 v5.9.2
+	github.com/prometheus/client_golang v1.23.2
+	github.com/prometheus/client_model v0.6.2
 	github.com/stretchr/testify v1.11.1
-	golang.org/x/crypto v0.45.0
 	golang.org/x/oauth2 v0.36.0
 	golang.org/x/time v0.15.0
 )
 
 require (
+	github.com/beorn7/perks v1.0.1 // indirect
+	github.com/cespare/xxhash/v2 v2.3.0 // indirect
 	github.com/davecgh/go-spew v1.1.2-0.20180830191138-d8f796af33cc // indirect
 	github.com/jackc/pgerrcode v0.0.0-20220416144525-469b46aa5efa // indirect
 	github.com/jackc/pgpassfile v1.0.0 // indirect
 	github.com/jackc/pgservicefile v0.0.0-20240606120523-5a60cdf6a761 // indirect
 	github.com/jackc/puddle/v2 v2.2.2 // indirect
+	github.com/kylelemons/godebug v1.1.0 // indirect
 	github.com/lib/pq v1.10.9 // indirect
+	github.com/munnerz/goautoneg v0.0.0-20191010083416-a7dc8b61c822 // indirect
 	github.com/pmezard/go-difflib v1.0.1-0.20181226105442-5d4384ee4fb2 // indirect
+	github.com/prometheus/common v0.66.1 // indirect
+	github.com/prometheus/procfs v0.16.1 // indirect
 	github.com/rogpeppe/go-internal v1.15.0 // indirect
 	github.com/xi2/xz v0.0.0-20171230120015-48954b6210f8 // indirect
-	golang.org/x/sync v0.18.0 // indirect
-	golang.org/x/text v0.31.0 // indirect
+	go.yaml.in/yaml/v2 v2.4.2 // indirect
+	golang.org/x/sync v0.21.0 // indirect
+	golang.org/x/sys v0.41.0 // indirect
+	golang.org/x/text v0.39.0 // indirect
+	google.golang.org/protobuf v1.36.8 // indirect
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -4,6 +4,10 @@ github.com/Microsoft/go-winio v0.6.2 h1:F2VQgta7ecxGYO8k3ZZz3RS8fVIXVxONVUPlNERo
 github.com/Microsoft/go-winio v0.6.2/go.mod h1:yd8OoFMLzJbo9gZq8j5qaps8bJ9aShtEA8Ipt1oGCvU=
 github.com/a-h/templ v0.3.1020 h1:ypAT/L5ySWEnZ6Zft/5yfoWXYYkhFNvEFOeeqecg4tw=
 github.com/a-h/templ v0.3.1020/go.mod h1:A2DlK61v+K+NRoGnhmYbNYVmtYHcFO5/AisMvBdDxTM=
+github.com/beorn7/perks v1.0.1 h1:VlbKKnNfV8bJzeqoa4cOKqO6bYr3WgKZxO8Z16+hsOM=
+github.com/beorn7/perks v1.0.1/go.mod h1:G2ZrVWU2WbWT9wwq4/hrbKbnv/1ERSJQ0ibhJ6rlkpw=
+github.com/cespare/xxhash/v2 v2.3.0 h1:UL815xU9SqsFlibzuggzjXhog7bL6oX9BbNZnL2UFvs=
+github.com/cespare/xxhash/v2 v2.3.0/go.mod h1:VGX0DQ3Q6kWi7AoAeZDth3/j3BFtOZR5XLFGgcrjCOs=
 github.com/containerd/errdefs v1.0.0 h1:tg5yIfIlQIrxYtu9ajqY42W3lpS19XqdxRQeEwYG8PI=
 github.com/containerd/errdefs v1.0.0/go.mod h1:+YBYIdtsnF4Iw6nWZhJcqGSg/dwvV7tyJ/kCkyJ2k+M=
 github.com/containerd/errdefs/pkg v0.3.0 h1:9IKJ06FvyNlexW690DXuQNx2KA2cUJXx151Xdx3ZPPE=
@@ -37,8 +41,8 @@ github.com/gogo/protobuf v1.3.2 h1:Ov1cvc58UF3b5XjBnZv7+opcTcQFZebYjWzi34vdm4Q=
 github.com/gogo/protobuf v1.3.2/go.mod h1:P1XiOD3dCwIKUDQYPy72D8LYyHL2YPYrpS2s69NZV8Q=
 github.com/golang-migrate/migrate/v4 v4.19.1 h1:OCyb44lFuQfYXYLx1SCxPZQGU7mcaZ7gH9yH4jSFbBA=
 github.com/golang-migrate/migrate/v4 v4.19.1/go.mod h1:CTcgfjxhaUtsLipnLoQRWCrjYXycRz/g5+RWDuYgPrE=
-github.com/google/go-cmp v0.6.0 h1:ofyhxvXcZhMsU5ulbFiLKl/XBFqE1GSq7atu8tAmTRI=
-github.com/google/go-cmp v0.6.0/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
+github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=
+github.com/google/go-cmp v0.7.0/go.mod h1:pXiqmnSA92OHEEa9HXL2W4E7lf9JzCmGVUdgjX3N/iU=
 github.com/jackc/pgerrcode v0.0.0-20220416144525-469b46aa5efa h1:s+4MhCQ6YrzisK6hFJUX53drDT4UsSW3DEhKn0ifuHw=
 github.com/jackc/pgerrcode v0.0.0-20220416144525-469b46aa5efa/go.mod h1:a/s9Lp5W7n/DD0VrVoyJ00FbP2ytTPDVOivvn2bMlds=
 github.com/jackc/pgpassfile v1.0.0 h1:/6Hmqy13Ss2zCq62VdNG8tM1wchn8zjSGOBJ6icpsIM=
@@ -49,10 +53,14 @@ github.com/jackc/pgx/v5 v5.9.2 h1:3ZhOzMWnR4yJ+RW1XImIPsD1aNSz4T4fyP7zlQb56hw=
 github.com/jackc/pgx/v5 v5.9.2/go.mod h1:mal1tBGAFfLHvZzaYh77YS/eC6IX9OWbRV1QIIM0Jn4=
 github.com/jackc/puddle/v2 v2.2.2 h1:PR8nw+E/1w0GLuRFSmiioY6UooMp6KJv0/61nB7icHo=
 github.com/jackc/puddle/v2 v2.2.2/go.mod h1:vriiEXHvEE654aYKXXjOvZM39qJ0q+azkZFrfEOc3H4=
-github.com/kr/pretty v0.3.0 h1:WgNl7dwNpEZ6jJ9k1snq4pZsg7DOEN8hP9Xw0Tsjwk0=
-github.com/kr/pretty v0.3.0/go.mod h1:640gp4NfQd8pI5XOwp5fnNeVWj67G7CFk/SaSQn7NBk=
+github.com/klauspost/compress v1.18.0 h1:c/Cqfb0r+Yi+JtIEq73FWXVkRonBlf0CRNYc8Zttxdo=
+github.com/klauspost/compress v1.18.0/go.mod h1:2Pp+KzxcywXVXMr50+X0Q/Lsb43OQHYWRCY2AiWywWQ=
+github.com/kr/pretty v0.3.1 h1:flRD4NNwYAUpkphVc1HcthR4KEIFJ65n8Mw5qdRn3LE=
+github.com/kr/pretty v0.3.1/go.mod h1:hoEshYVHaxMs3cyo3Yncou5ZscifuDolrwPKZanG3xk=
 github.com/kr/text v0.2.0 h1:5Nx0Ya0ZqY2ygV366QzturHI13Jq95ApcVaJBhpS+AY=
 github.com/kr/text v0.2.0/go.mod h1:eLer722TekiGuMkidMxC/pM04lWEeraHUUmBw8l2grE=
+github.com/kylelemons/godebug v1.1.0 h1:RPNrshWIDI6G2gRW9EHilWtl7Z6Sb1BR0xunSBf0SNc=
+github.com/kylelemons/godebug v1.1.0/go.mod h1:9/0rRGxNHcop5bhtWyNeEfOS8JIWk580+fNqagV/RAw=
 github.com/lib/pq v1.10.9 h1:YXG7RB+JIjhP29X+OtkiDnYaXQwpS4JEWq7dtCCRUEw=
 github.com/lib/pq v1.10.9/go.mod h1:AlVN5x4E4T544tWzH6hKfbfQvm3HdbOxrmggDNAPY9o=
 github.com/moby/docker-image-spec v1.3.1 h1:jMKff3w6PgbfSa69GfNg+zN/XLhfXJGnEx3Nl2EsFP0=
@@ -61,6 +69,8 @@ github.com/moby/term v0.5.0 h1:xt8Q1nalod/v7BqbG21f8mQPqH+xAaC9C3N3wfWbVP0=
 github.com/moby/term v0.5.0/go.mod h1:8FzsFHVUBGZdbDsJw/ot+X+d5HLUbvklYLJ9uGfcI3Y=
 github.com/morikuni/aec v1.0.0 h1:nP9CBfwrvYnBRgY6qfDQkygYDmYwOilePFkwzv4dU8A=
 github.com/morikuni/aec v1.0.0/go.mod h1:BbKIizmSmc5MMPqRYbxO4ZU0S0+P200+tUnFx7PXmsc=
+github.com/munnerz/goautoneg v0.0.0-20191010083416-a7dc8b61c822 h1:C3w9PqII01/Oq1c1nUAm88MOHcQC9l5mIlSMApZMrHA=
+github.com/munnerz/goautoneg v0.0.0-20191010083416-a7dc8b61c822/go.mod h1:+n7T8mK8HuQTcFwEeznm/DIxMOiR9yIdICNftLE1DvQ=
 github.com/opencontainers/go-digest v1.0.0 h1:apOUWs51W5PlhuyGyz9FCeeBIOUDA/6nW8Oi/yOhh5U=
 github.com/opencontainers/go-digest v1.0.0/go.mod h1:0JzlMkj0TRzQZfJkVvzbP0HBR3IKzErnv2BNG4W4MAM=
 github.com/opencontainers/image-spec v1.1.0 h1:8SG7/vwALn54lVB/0yZ/MMwhFrPYtpEHQb2IpWsCzug=
@@ -70,6 +80,14 @@ github.com/pkg/errors v0.9.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINE
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
 github.com/pmezard/go-difflib v1.0.1-0.20181226105442-5d4384ee4fb2 h1:Jamvg5psRIccs7FGNTlIRMkT8wgtp5eCXdBlqhYGL6U=
 github.com/pmezard/go-difflib v1.0.1-0.20181226105442-5d4384ee4fb2/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
+github.com/prometheus/client_golang v1.23.2 h1:Je96obch5RDVy3FDMndoUsjAhG5Edi49h0RJWRi/o0o=
+github.com/prometheus/client_golang v1.23.2/go.mod h1:Tb1a6LWHB3/SPIzCoaDXI4I8UHKeFTEQ1YCr+0Gyqmg=
+github.com/prometheus/client_model v0.6.2 h1:oBsgwpGs7iVziMvrGhE53c/GrLUsZdHnqNwqPLxwZyk=
+github.com/prometheus/client_model v0.6.2/go.mod h1:y3m2F6Gdpfy6Ut/GBsUqTWZqCUvMVzSfMLjcu6wAwpE=
+github.com/prometheus/common v0.66.1 h1:h5E0h5/Y8niHc5DlaLlWLArTQI7tMrsfQjHV+d9ZoGs=
+github.com/prometheus/common v0.66.1/go.mod h1:gcaUsgf3KfRSwHY4dIMXLPV0K/Wg1oZ8+SbZk/HH/dA=
+github.com/prometheus/procfs v0.16.1 h1:hZ15bTNuirocR6u0JZ6BAHHmwS1p8B4P6MRqxtzMyRg=
+github.com/prometheus/procfs v0.16.1/go.mod h1:teAbpZRB1iIAJYREa1LsoWUXykVXA1KlTmWl8x/U+Is=
 github.com/rogpeppe/go-internal v1.15.0 h1:D0RCU5rMAp+SpgkiNdrjfJ+LX4J1M32V2NeCY7EJ6hc=
 github.com/rogpeppe/go-internal v1.15.0/go.mod h1:DrUVZyrJU+txYW5/1kwtXQSMFio52ZOxX7yM1VHvnxs=
 github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
@@ -91,18 +109,20 @@ go.opentelemetry.io/otel/trace v1.37.0 h1:HLdcFNbRQBE2imdSEgm/kwqmQj1Or1l/7bW6mx
 go.opentelemetry.io/otel/trace v1.37.0/go.mod h1:TlgrlQ+PtQO5XFerSPUYG0JSgGyryXewPGyayAWSBS0=
 go.uber.org/goleak v1.3.0 h1:2K3zAYmnTNqV73imy9J1T3WC+gmCePx2hEGkimedGto=
 go.uber.org/goleak v1.3.0/go.mod h1:CoHD4mav9JJNrW/WLlf7HGZPjdw8EucARQHekz1X6bE=
-golang.org/x/crypto v0.45.0 h1:jMBrvKuj23MTlT0bQEOBcAE0mjg8mK9RXFhRH6nyF3Q=
-golang.org/x/crypto v0.45.0/go.mod h1:XTGrrkGJve7CYK7J8PEww4aY7gM3qMCElcJQ8n8JdX4=
+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=
+google.golang.org/protobuf v1.36.8/go.mod h1:fuxRtAxBytpl4zzqUh6/eyUujkJdNiuEkXntxiD/uRU=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c h1:Hei/4ADfdWqJk1ZMxUNpqntNwaWcugrBjAiHlqqRiVk=
 gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c/go.mod h1:JHkPIbrfpd72SG/EVd6muEfDQjcINNoR0C8j2r3qZ4Q=

internal/adapters/audio/ffmpeg.go

@@ -0,0 +1,66 @@
+package audio
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"os/exec"
+	"strconv"
+	"strings"
+)
+
+// FFmpeg converts any audio container to the strict format whisper wants: 16 kHz
+// mono signed 16-bit PCM WAV (ADR-033). It ALWAYS transcodes, even when the input
+// already is that format — a Teams export often is, and the transcode is then
+// nearly free; assuming it would be the bug. This replaces whisper-server's
+// --convert, whose --tmp-dir is created at process start and rots under macOS
+// periodic cleanup.
+type FFmpeg struct {
+	// Bin is the ffmpeg executable ("ffmpeg" resolves via PATH).
+	Bin string
+	// MaxSeconds caps the converted length when positive. The duration limit is a claim
+	// in the file's header and can be lied about (a 2 h file claiming 10 s converted to a
+	// 120-minute WAV); this bounds what a lie can cost. A minute of slack is added so a
+	// file at the limit is never cut, and the worker measures the result.
+	MaxSeconds int
+}
+
+// lengthSlackSeconds is added to MaxSeconds so ffmpeg's rounding never clips a file that
+// is really at the limit.
+const lengthSlackSeconds = 60
+
+// ToWAV writes src, converted, to dst. Video streams are dropped (-vn): by now the
+// file has already been probed as audio-only, but a cover-art stream must not end
+// up in the WAV.
+func (f *FFmpeg) ToWAV(ctx context.Context, src, dst string) error {
+	bin := f.Bin
+	if bin == "" {
+		bin = "ffmpeg"
+	}
+	// Input restrictions go BEFORE -i: placed after it they would apply to the output.
+	args := append([]string{"-y", "-v", "error"}, inputRestrictions...)
+	args = append(args, "-i", src, "-vn", "-ac", "1", "-ar", "16000", "-c:a", "pcm_s16le")
+	if f.MaxSeconds > 0 {
+		// After -i: an OUTPUT cap. Before it, it would limit what is read, not written.
+		args = append(args, "-t", strconv.Itoa(f.MaxSeconds+lengthSlackSeconds))
+	}
+	args = append(args, "-f", "wav", dst)
+	cmd := exec.CommandContext(ctx, bin, args...)
+	cmd.WaitDelay = killGrace
+	// 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(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.
+	if info, err := os.Stat(dst); err != nil || info.Size() == 0 {
+		return fmt.Errorf("audio: ffmpeg wrote no output at %s", dst)
+	}
+	return nil
+}

internal/adapters/audio/ffmpeg_test.go

@@ -0,0 +1,140 @@
+package audio_test
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/audio"
+)
+
+func TestFFmpegToWAV_AsksForSixteenKilohertzMonoPCMAlways(t *testing.T) {
+	// Always transcode, even when the input already is 16 kHz mono: a Teams export
+	// often is, and the point is to not ASSUME (ADR-033). Records the argv.
+	rec := filepath.Join(t.TempDir(), "argv")
+	bin := stub(t, `printf '%s\n' "$@" > `+rec+`
+for a; do last="$a"; done
+printf data > "$last"`)
+
+	err := (&audio.FFmpeg{Bin: bin}).ToWAV(context.Background(), "/spool/in", filepath.Join(t.TempDir(), "out.wav"))
+
+	require.NoError(t, err)
+	raw, err := os.ReadFile(rec)
+	require.NoError(t, err)
+	argv := strings.Split(strings.TrimSpace(string(raw)), "\n")
+	joined := " " + strings.Join(argv, " ") + " "
+	for _, want := range []string{" -ac 1 ", " -ar 16000 ", " -c:a pcm_s16le ", " -vn ", " -i /spool/in "} {
+		assert.Contains(t, joined, want)
+	}
+	assert.True(t, strings.HasSuffix(strings.TrimSpace(joined), "out.wav"), "the output path is the last argument")
+}
+
+func TestFFmpegToWAV_NonzeroExitIsAnErrorCarryingStderr(t *testing.T) {
+	bin := stub(t, `echo "moov atom not found" >&2; exit 1`)
+
+	err := (&audio.FFmpeg{Bin: bin}).ToWAV(context.Background(), "/in", filepath.Join(t.TempDir(), "o.wav"))
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "moov atom not found")
+}
+
+func TestFFmpegToWAV_ZeroExitWithNoOutputIsAnError(t *testing.T) {
+	bin := stub(t, `exit 0`)
+
+	err := (&audio.FFmpeg{Bin: bin}).ToWAV(context.Background(), "/in", filepath.Join(t.TempDir(), "never-written.wav"))
+
+	require.Error(t, err)
+}
+
+func TestFFmpegToWAV_MissingBinaryIsAnError(t *testing.T) {
+	err := (&audio.FFmpeg{Bin: filepath.Join(t.TempDir(), "no-ffmpeg")}).ToWAV(context.Background(), "/in", "/out.wav")
+
+	require.Error(t, err)
+	assert.False(t, errors.Is(err, context.Canceled))
+}
+
+func TestFFmpegToWAV_RestrictsWhatTheFileMayMakeFFmpegOpen(t *testing.T) {
+	rec := filepath.Join(t.TempDir(), "argv")
+	bin := stub(t, `printf '%s\n' "$@" > `+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)
+	pw, fw, in := indexOf(args, "-protocol_whitelist"), indexOf(args, "-format_whitelist"), indexOf(args, "-i")
+	require.GreaterOrEqual(t, pw, 0, "argv: %v", args)
+	require.GreaterOrEqual(t, fw, 0, "argv: %v", args)
+	assert.Equal(t, "file,pipe", args[pw+1])
+	assert.Equal(t, audio.AllowedDemuxers, args[fw+1])
+	assert.Less(t, pw, in, "input options must come BEFORE -i or ffmpeg applies them to the output")
+	assert.Less(t, fw, in)
+}
+
+func TestFFmpegToWAV_PassesAnOutputLengthCapAfterTheInputWhenAMaximumIsSet(t *testing.T) {
+	// The duration limit is a CLAIM in the file's header; a file that lies about it can
+	// convert to hours of audio (adversarial review F2: a 2 h file claiming 10 s became a
+	// 120-minute WAV). The cap is an OUTPUT option, after -i, with a minute of slack.
+	rec := filepath.Join(t.TempDir(), "argv")
+	bin := stub(t, `printf '%s\n' "$@" > `+rec+`
+for a; do last="$a"; done
+printf data > "$last"`)
+
+	err := (&audio.FFmpeg{Bin: bin, MaxSeconds: 5400}).ToWAV(context.Background(), "/spool/in", filepath.Join(t.TempDir(), "out.wav"))
+
+	require.NoError(t, err)
+	raw, err := os.ReadFile(rec)
+	require.NoError(t, err)
+	joined := " " + strings.Join(strings.Split(strings.TrimSpace(string(raw)), "\n"), " ") + " "
+	assert.Contains(t, joined, " -t 5460 ", "limit plus a minute of slack")
+	assert.Less(t, strings.Index(joined, " -i /spool/in "), strings.Index(joined, " -t 5460 "), "-t after -i is an output cap; before it, an input one")
+}
+
+func TestFFmpegToWAV_PassesNoLengthCapWhenNoMaximumIsSet(t *testing.T) {
+	rec := filepath.Join(t.TempDir(), "argv")
+	bin := stub(t, `printf '%s\n' "$@" > `+rec+`
+for a; do last="$a"; done
+printf data > "$last"`)
+
+	require.NoError(t, (&audio.FFmpeg{Bin: bin}).ToWAV(context.Background(), "/spool/in", filepath.Join(t.TempDir(), "out.wav")))
+
+	raw, err := os.ReadFile(rec)
+	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")
+}

internal/adapters/audio/ffprobe.go

@@ -0,0 +1,185 @@
+// Package audio holds the adapters that inspect and convert user-supplied audio
+// with ffmpeg's tools (ADR-033). They sit behind small ports so the use case is
+// tested with fakes and only this package knows a subprocess is involved.
+package audio
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os/exec"
+	"strconv"
+	"time"
+
+	"git.d-ma.be/mathias/tapir/internal/uploads"
+)
+
+// FFprobe implements uploads.Prober by running ffprobe on the spooled file.
+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
+// attacker-supplied, and a file's CONTENT can tell ffmpeg to open other files or
+// URLs (HLS and concat playlists, image2 sequences, network demuxers). Recent
+// ffmpeg refuses the ones tried (7.0.2: HLS without a .m3u8 name is not sniffed,
+// concat rejects unsafe paths), but that is version luck; this list is not. It
+// holds the containers real recordings use and none of the playlist or network
+// 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, 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
+
+// Probe runs ffprobe and classifies the outcome. A non-zero exit means ffprobe
+// looked at the file and could not read it: that is the FILE's fault and returns
+// uploads.ErrUnreadable. A missing binary, a cancelled context or unparsable
+// output is OURS and is returned as an ordinary error, so it is never reported to
+// the user as their file being bad.
+func (f *FFprobe) Probe(ctx context.Context, path string) (uploads.Probe, error) {
+	bin := f.Bin
+	if bin == "" {
+		bin = "ffprobe"
+	}
+	args := append([]string{"-v", "error"}, inputRestrictions...)
+	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
+	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, stderr.buf.String())
+		}
+		return uploads.Probe{}, fmt.Errorf("audio: run ffprobe: %w", err)
+	}
+	p, err := ParseProbe(stdout.buf.Bytes())
+	if err != nil {
+		return uploads.Probe{}, fmt.Errorf("audio: ffprobe output: %w", err)
+	}
+	return p, nil
+}
+
+// probeJSON is the subset of `ffprobe -print_format json` output that matters.
+type probeJSON struct {
+	Streams []struct {
+		CodecType   string `json:"codec_type"`
+		Disposition struct {
+			AttachedPic int `json:"attached_pic"`
+		} `json:"disposition"`
+	} `json:"streams"`
+	Format struct {
+		Duration string `json:"duration"`
+	} `json:"format"`
+}
+
+// ParseProbe turns ffprobe's JSON into a uploads.Probe. A video stream that is
+// only attached cover art (an mp3/m4a with an album picture) is not video. A
+// missing or "N/A" duration yields zero, which uploads.Validate rejects as
+// unreadable rather than accepting an empty recording.
+func ParseProbe(data []byte) (uploads.Probe, error) {
+	var j probeJSON
+	if err := json.Unmarshal(data, &j); err != nil {
+		return uploads.Probe{}, fmt.Errorf("decode: %w", err)
+	}
+	var p uploads.Probe
+	for _, s := range j.Streams {
+		switch s.CodecType {
+		case "audio":
+			p.HasAudio = true
+		case "video":
+			if s.Disposition.AttachedPic == 0 {
+				p.HasVideo = true
+			}
+		}
+	}
+	if secs, err := strconv.ParseFloat(j.Format.Duration, 64); err == nil && secs > 0 {
+		p.Duration = time.Duration(secs * float64(time.Second))
+	}
+	return p, nil
+}

internal/adapters/audio/ffprobe_test.go

@@ -0,0 +1,261 @@
+package audio_test
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/audio"
+	"git.d-ma.be/mathias/tapir/internal/uploads"
+)
+
+func TestParseProbe(t *testing.T) {
+	tests := []struct {
+		name string
+		json string
+		want uploads.Probe
+	}{
+		{
+			name: "audio-only m4a",
+			json: `{"streams":[{"codec_type":"audio","disposition":{"attached_pic":0}}],"format":{"duration":"2701.5"}}`,
+			want: uploads.Probe{Duration: 2701*time.Second + 500*time.Millisecond, HasAudio: true},
+		},
+		{
+			name: "mp4 with a real video stream",
+			json: `{"streams":[{"codec_type":"video","disposition":{"attached_pic":0}},{"codec_type":"audio"}],"format":{"duration":"240.0"}}`,
+			want: uploads.Probe{Duration: 240 * time.Second, HasAudio: true, HasVideo: true},
+		},
+		{
+			name: "mp3 whose only video stream is attached cover art is NOT video",
+			json: `{"streams":[{"codec_type":"audio"},{"codec_type":"video","disposition":{"attached_pic":1}}],"format":{"duration":"60"}}`,
+			want: uploads.Probe{Duration: 60 * time.Second, HasAudio: true},
+		},
+		{
+			name: "no streams at all",
+			json: `{"streams":[],"format":{"duration":"10"}}`,
+			want: uploads.Probe{Duration: 10 * time.Second},
+		},
+		{
+			name: "duration absent means zero, which Validate rejects as unreadable",
+			json: `{"streams":[{"codec_type":"audio"}],"format":{}}`,
+			want: uploads.Probe{HasAudio: true},
+		},
+		{
+			name: "duration reported as N/A means zero",
+			json: `{"streams":[{"codec_type":"audio"}],"format":{"duration":"N/A"}}`,
+			want: uploads.Probe{HasAudio: true},
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got, err := audio.ParseProbe([]byte(tt.json))
+
+			require.NoError(t, err)
+			assert.Equal(t, tt.want, got)
+		})
+	}
+}
+
+func TestParseProbe_GarbageIsAnError(t *testing.T) {
+	_, err := audio.ParseProbe([]byte("not json"))
+	assert.Error(t, err)
+}
+
+// stub writes an executable shell script standing in for ffprobe. This is a fake
+// at the process boundary: the real binary is not on every host (it is not on
+// koala), and what is under test is how Probe drives and interprets a process.
+func stub(t *testing.T, body string) string {
+	t.Helper()
+	p := filepath.Join(t.TempDir(), "ffprobe")
+	require.NoError(t, os.WriteFile(p, []byte("#!/bin/sh\n"+body+"\n"), 0o755))
+	return p
+}
+
+func TestFFprobeProbe_ParsesWhatTheProcessPrints(t *testing.T) {
+	bin := stub(t, `echo '{"streams":[{"codec_type":"audio"}],"format":{"duration":"90"}}'`)
+
+	got, err := (&audio.FFprobe{Bin: bin}).Probe(context.Background(), "/ignored")
+
+	require.NoError(t, err)
+	assert.Equal(t, uploads.Probe{Duration: 90 * time.Second, HasAudio: true}, got)
+}
+
+func TestFFprobeProbe_PassesTheFilePathAsTheLastArgument(t *testing.T) {
+	out := filepath.Join(t.TempDir(), "args")
+	bin := stub(t, `for a; do last="$a"; done; printf '%s' "$last" > `+out+`
+echo '{"streams":[{"codec_type":"audio"}],"format":{"duration":"1"}}'`)
+
+	_, err := (&audio.FFprobe{Bin: bin}).Probe(context.Background(), "/spool/upload-123")
+
+	require.NoError(t, err)
+	got, err := os.ReadFile(out)
+	require.NoError(t, err)
+	assert.Equal(t, "/spool/upload-123", string(got))
+}
+
+func TestFFprobeProbe_NonzeroExitMeansTheFileIsUnreadable(t *testing.T) {
+	bin := stub(t, `echo "Invalid data found when processing input" >&2; exit 1`)
+
+	_, err := (&audio.FFprobe{Bin: bin}).Probe(context.Background(), "/x")
+
+	assert.ErrorIs(t, err, uploads.ErrUnreadable)
+}
+
+func TestFFprobeProbe_MissingBinaryIsOurFaultNotAnUnreadableFile(t *testing.T) {
+	_, err := (&audio.FFprobe{Bin: filepath.Join(t.TempDir(), "no-such-ffprobe")}).Probe(context.Background(), "/x")
+
+	require.Error(t, err)
+	assert.False(t, errors.Is(err, uploads.ErrUnreadable),
+		"a missing ffprobe must not be reported to the user as their file being bad")
+}
+
+func TestFFprobeProbe_StopsWhenTheContextIsCancelled(t *testing.T) {
+	bin := stub(t, `sleep 30`)
+	ctx, cancel := context.WithTimeout(context.Background(), 100*time.Millisecond)
+	defer cancel()
+
+	start := time.Now()
+	_, err := (&audio.FFprobe{Bin: bin}).Probe(ctx, "/x")
+
+	require.Error(t, err)
+	assert.Less(t, time.Since(start), 5*time.Second)
+	assert.False(t, errors.Is(err, uploads.ErrUnreadable), "a timeout is not a verdict on the file")
+}
+
+// argvStub records the arguments it was called with into a file and then behaves as
+// a successful ffprobe, so a test can assert exactly what the process was asked.
+func argvStub(t *testing.T) (bin, argvFile string) {
+	t.Helper()
+	argvFile = filepath.Join(t.TempDir(), "argv")
+	bin = stub(t, `printf '%s\n' "$@" > `+argvFile+`
+echo '{"streams":[{"codec_type":"audio"}],"format":{"duration":"1"}}'`)
+	return bin, argvFile
+}
+
+func argvOf(t *testing.T, file string) []string {
+	t.Helper()
+	raw, err := os.ReadFile(file)
+	require.NoError(t, err)
+	return strings.Split(strings.TrimSpace(string(raw)), "\n")
+}
+
+func indexOf(args []string, want string) int {
+	for i, a := range args {
+		if a == want {
+			return i
+		}
+	}
+	return -1
+}
+
+func TestFFprobeProbe_RestrictsWhatTheFileMayMakeFFmpegOpen(t *testing.T) {
+	// The input is attacker-supplied. ffmpeg can be told by a file's CONTENT to open
+	// other files or URLs (HLS and concat playlists); recent versions refuse by
+	// default, but that is version luck. Pin it: local files and pipes only, and
+	// only the demuxers real audio recordings use.
+	bin, argvFile := argvStub(t)
+
+	_, err := (&audio.FFprobe{Bin: bin}).Probe(context.Background(), "/spool/upload-1")
+
+	require.NoError(t, err)
+	args := argvOf(t, argvFile)
+	pw := indexOf(args, "-protocol_whitelist")
+	require.GreaterOrEqual(t, pw, 0, "argv: %v", args)
+	assert.Equal(t, "file,pipe", args[pw+1])
+	fw := indexOf(args, "-format_whitelist")
+	require.GreaterOrEqual(t, fw, 0, "argv: %v", args)
+	assert.Equal(t, audio.AllowedDemuxers, args[fw+1])
+	assert.Equal(t, "/spool/upload-1", args[len(args)-1], "the path stays the last argument")
+}
+
+func TestAllowedDemuxers_ExcludeThePlaylistFormats(t *testing.T) {
+	for _, bad := range []string{"concat", "hls", "applehttp", "dash", "rtsp", "sdp", "image2"} {
+		assert.NotContains(t, strings.Split(audio.AllowedDemuxers, ","), bad,
+			"%s can make ffmpeg open other files or URLs named inside the upload", bad)
+	}
+	for _, good := range []string{"wav", "mov", "mp4", "m4a", "mp3", "ogg", "matroska", "flac"} {
+		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)
+}

internal/adapters/audio/hostile_live_test.go

@@ -0,0 +1,32 @@
+package audio_test
+
+import (
+	"context"
+	"os"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/audio"
+	"git.d-ma.be/mathias/tapir/internal/uploads"
+)
+
+// TestFFprobeLive_HostilePlaylistsAreRefusedAndRealAudioIsNot runs the REAL ffprobe
+// (opt-in: TAPIR_LIVE_FFPROBE) on crafted concat/HLS files and on a real m4a.
+func TestFFprobeLive_HostilePlaylistsAreRefusedAndRealAudioIsNot(t *testing.T) {
+	bin := os.Getenv("TAPIR_LIVE_FFPROBE")
+	if bin == "" {
+		t.Skip("set TAPIR_LIVE_FFPROBE (and TAPIR_LIVE_HOSTILE_DIR, TAPIR_LIVE_M4A) to run against the real ffprobe")
+	}
+	p := &audio.FFprobe{Bin: bin}
+	for _, name := range []string{"up-concat", "up-hls"} {
+		_, err := p.Probe(context.Background(), os.Getenv("TAPIR_LIVE_HOSTILE_DIR")+"/"+name)
+		assert.ErrorIs(t, err, uploads.ErrUnreadable, name)
+	}
+	got, err := p.Probe(context.Background(), os.Getenv("TAPIR_LIVE_M4A"))
+	require.NoError(t, err)
+	assert.True(t, got.HasAudio)
+	assert.False(t, got.HasVideo)
+	assert.InDelta(t, 82.5, got.Duration.Seconds(), 1)
+}

internal/adapters/chat/chat.go

@@ -13,8 +13,12 @@ package chat
 import (
 	"context"
 	"fmt"
+	"log/slog"
 	"strings"
+	"time"
 	"unicode/utf8"
+
+	"git.d-ma.be/mathias/tapir/internal/metrics"
 )
 
 // Completer is the minimal LLM chat surface the Service needs. *llm.Client
@@ -37,6 +41,11 @@ type Request struct {
 	Transcript string
 	History    []Turn
 	Question   string
+	// MaxChars, when positive, overrides the Service's transcript budget for this
+	// request. The Service default is sized for YouTube captions (~22 minutes of
+	// speech); an uploaded recording is much longer, and truncation keeps only the
+	// head (ADR-033). Zero or negative means "use the Service default".
+	MaxChars int
 }
 
 // Reply is the model's answer plus whether the transcript was bounded to fit the
@@ -109,14 +118,22 @@ func (s *Service) Answer(ctx context.Context, req Request) (Reply, error) {
 		model = s.DefaultModel("")
 	}
 
-	transcript, truncated := truncate(req.Transcript, s.maxChars)
+	budget := s.maxChars
+	if req.MaxChars > 0 {
+		budget = req.MaxChars
+	}
+	transcript, truncated := truncate(req.Transcript, budget)
 	system := buildSystem(transcript, truncated)
 	user := buildUser(req.History, req.Question)
 
+	start := time.Now()
 	out, err := s.newClient(model).Complete(ctx, system, user)
 	if err != nil {
 		return Reply{}, fmt.Errorf("chat: %s: %w", model, err)
 	}
+	dur := time.Since(start)
+	metrics.ObserveChat(model, dur)
+	slog.Default().Info("chat answer", "model", model, "elapsed_ms", dur.Milliseconds())
 	answer := strings.TrimSpace(out)
 	if answer == "" {
 		return Reply{}, fmt.Errorf("chat: %s returned an empty answer", model)

internal/adapters/chat/maxchars_test.go

@@ -0,0 +1,52 @@
+package chat
+
+import (
+	"context"
+	"strings"
+	"testing"
+)
+
+// A Request may carry its own transcript budget. The Service's cap (18,000 chars by
+// default) is sized for YouTube captions; an uploaded recording is much longer
+// (~50,000 chars an hour), and keeping only the head of it would answer questions
+// about the last two-thirds of a meeting from text that does not contain them
+// (ADR-033). The per-request value overrides; zero means "the service default".
+
+func TestAnswerMaxCharsOverridesTheServiceCap(t *testing.T) {
+	tests := []struct {
+		name          string
+		serviceCap    int
+		requestCap    int
+		transcriptLen int
+		wantKept      int
+		wantTruncated bool
+	}{
+		{"request cap above the service cap keeps the whole transcript", 10, 100, 50, 50, false},
+		{"request cap below the service cap bounds tighter", 100, 10, 50, 10, true},
+		{"zero request cap falls back to the service cap", 20, 0, 50, 20, true},
+		{"negative request cap is treated as unset", 20, -1, 50, 20, true},
+		{"request cap exceeded is reported as truncated", 10, 30, 50, 30, true},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			f := &factory{answer: "answer"}
+			s := New(f.make, []string{"koala/phi4-mini"}, tt.serviceCap)
+
+			reply, err := s.Answer(context.Background(), Request{
+				Model:      "koala/phi4-mini",
+				Transcript: strings.Repeat("x", tt.transcriptLen),
+				Question:   "what was decided?",
+				MaxChars:   tt.requestCap,
+			})
+			if err != nil {
+				t.Fatalf("Answer: %v", err)
+			}
+			if got := strings.Count(f.used.lastSystem, "x"); got != tt.wantKept {
+				t.Errorf("transcript chars sent = %d, want %d", got, tt.wantKept)
+			}
+			if reply.Truncated != tt.wantTruncated {
+				t.Errorf("Truncated = %v, want %v", reply.Truncated, tt.wantTruncated)
+			}
+		})
+	}
+}

internal/adapters/llm/bodycap_test.go

@@ -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)
+		}
+	}
+}

internal/adapters/llm/client.go

@@ -25,13 +25,33 @@ import (
 // docs/homelab-integration.md.
 const defaultMaxTokens = 8192
 
+// defaultTemperature is what every caller got before WithTemperature existed.
+const defaultTemperature = 0.2
+
 // Client calls an OpenAI-compatible chat completions endpoint.
 type Client struct {
 	baseURL    string
 	apiKey     string
 	model      string
 	maxTokens  int
+	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
@@ -50,6 +70,43 @@ func WithMaxTokens(n int) Option {
 	}
 }
 
+// maxErrorBody bounds how much of a failing response ends up in an error message.
+const maxErrorBody = 512
+
+// HTTPError is a non-200 answer from the gateway. Its text keeps the form callers that
+// log it already rely on, with the body bounded; its Code is what code may branch on and
+// what errkind turns into a kind. The body is provider text and can echo the prompt, so
+// nothing private may persist or display Error() (ADR-035 review F3).
+type HTTPError struct {
+	Code int
+	Body string
+}
+
+func (e *HTTPError) Error() string {
+	body := e.Body
+	if len(body) > maxErrorBody {
+		body = body[:maxErrorBody] + "…"
+	}
+	return fmt.Sprintf("LLM returned %d: %s", e.Code, body)
+}
+
+// HTTPStatus implements errkind.HTTPStatuser.
+func (e *HTTPError) HTTPStatus() int { return e.Code }
+
+// WithTemperature overrides the sampling temperature. Zero IS a valid value and is sent
+// as zero: the reconciler's accuracy was measured at temperature 0 and must run there.
+func WithTemperature(t float64) Option {
+	return func(c *Client) { c.temp = t }
+}
+
+// WithUsageHook registers a callback fired after a successful completion with the
+// model and the prompt/completion token counts from the response usage block. It
+// keeps this copied, stdlib-only package (ADR-004) decoupled from metrics: the
+// caller wires it to internal/metrics, the client imports nothing. nil is ignored.
+func WithUsageHook(fn func(model string, prompt, completion int)) Option {
+	return func(c *Client) { c.usageHook = fn }
+}
+
 // New constructs a Client.
 func New(baseURL, apiKey, model string, timeout time.Duration, opts ...Option) *Client {
 	c := &Client{
@@ -57,7 +114,9 @@ func New(baseURL, apiKey, model string, timeout time.Duration, opts ...Option) *
 		apiKey:     apiKey,
 		model:      model,
 		maxTokens:  defaultMaxTokens,
+		temp:       defaultTemperature,
 		httpClient: &http.Client{Timeout: timeout},
+		maxBody:    defaultMaxResponseBytes,
 	}
 	for _, opt := range opts {
 		opt(c)
@@ -81,6 +140,10 @@ type chatResponse struct {
 	Choices []struct {
 		Message message `json:"message"`
 	} `json:"choices"`
+	Usage struct {
+		PromptTokens     int `json:"prompt_tokens"`
+		CompletionTokens int `json:"completion_tokens"`
+	} `json:"usage"`
 }
 
 // Complete sends a system + user message and returns the assistant's reply.
@@ -92,7 +155,7 @@ func (c *Client) Complete(ctx context.Context, system, user string) (string, err
 			{Role: "system", Content: system},
 			{Role: "user", Content: user},
 		},
-		Temperature: 0.2,
+		Temperature: c.temp,
 		MaxTokens:   c.maxTokens,
 	}
 	b, err := json.Marshal(body)
@@ -137,12 +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 "", fmt.Errorf("LLM returned %d: %s", resp.StatusCode, out)
+	if int64(len(out)) > c.maxBody {
+		return "", fmt.Errorf("LLM response exceeds %d bytes", c.maxBody)
 	}
 
 	var cr chatResponse
@@ -152,5 +221,8 @@ func (c *Client) Complete(ctx context.Context, system, user string) (string, err
 	if len(cr.Choices) == 0 {
 		return "", fmt.Errorf("LLM returned no choices")
 	}
+	if c.usageHook != nil {
+		c.usageHook(c.model, cr.Usage.PromptTokens, cr.Usage.CompletionTokens)
+	}
 	return cr.Choices[0].Message.Content, nil
 }

internal/adapters/llm/client_test.go

@@ -85,6 +85,30 @@ func TestClient_WithMaxTokens(t *testing.T) {
 	}
 }
 
+// TestClient_UsageHookRecordsTokens: the usage hook fires with the model and the
+// prompt/completion token counts parsed from the response usage block.
+func TestClient_UsageHookRecordsTokens(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"choices": []map[string]any{{"message": map[string]any{"content": "ok"}}},
+			"usage":   map[string]any{"prompt_tokens": 123, "completion_tokens": 45},
+		})
+	}))
+	defer srv.Close()
+
+	var gotModel string
+	var gotPrompt, gotCompletion int
+	c := New(srv.URL, "", "test-model", 10*time.Second, WithUsageHook(func(model string, p, comp int) {
+		gotModel, gotPrompt, gotCompletion = model, p, comp
+	}))
+	if _, err := c.Complete(context.Background(), "sys", "user"); err != nil {
+		t.Fatalf("Complete: %v", err)
+	}
+	if gotModel != "test-model" || gotPrompt != 123 || gotCompletion != 45 {
+		t.Errorf("usage hook got (%q, %d, %d), want (test-model, 123, 45)", gotModel, gotPrompt, gotCompletion)
+	}
+}
+
 func TestClient_ReturnsErrorOnNon200(t *testing.T) {
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		http.Error(w, "overloaded", http.StatusServiceUnavailable)

internal/adapters/llm/status_error_test.go

@@ -0,0 +1,63 @@
+package llm
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	"git.d-ma.be/mathias/tapir/internal/errkind"
+)
+
+// A non-200 from the gateway is an HTTP status error: its kind can be logged and stored
+// without its body, which can echo the prompt (adversarial review F3).
+
+func failingGateway(t *testing.T, status int, body string) *httptest.Server {
+	t.Helper()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(status)
+		_, _ = w.Write([]byte(body))
+	}))
+	t.Cleanup(srv.Close)
+	return srv
+}
+
+func TestClient_ANon200IsAnHTTPStatusErrorWhoseKindCarriesNoBody(t *testing.T) {
+	srv := failingGateway(t, http.StatusBadRequest, `{"error":"prompt was: Anders ansvarar för virket"}`)
+
+	_, err := New(srv.URL, "", "m", 5*time.Second).Complete(context.Background(), "s", "u")
+
+	if err == nil {
+		t.Fatal("expected an error")
+	}
+	if got := errkind.Of(err); got != "http_400" {
+		t.Errorf("errkind.Of = %q, want http_400", got)
+	}
+	var he *HTTPError
+	if !errors.As(err, &he) || he.Code != http.StatusBadRequest {
+		t.Errorf("error is not an *HTTPError with the status code: %v", err)
+	}
+}
+
+func TestClient_TheErrorTextIsUnchangedForCallersThatLogIt(t *testing.T) {
+	srv := failingGateway(t, http.StatusBadGateway, "upstream down")
+
+	_, err := New(srv.URL, "", "m", 5*time.Second).Complete(context.Background(), "s", "u")
+
+	if err == nil || err.Error() != "LLM returned 502: upstream down" {
+		t.Errorf("error text = %v: callers that log it already rely on this form", err)
+	}
+}
+
+func TestClient_AnEnormousErrorBodyIsBoundedInTheErrorText(t *testing.T) {
+	srv := failingGateway(t, http.StatusInternalServerError, strings.Repeat("x", 100_000))
+
+	_, err := New(srv.URL, "", "m", 5*time.Second).Complete(context.Background(), "s", "u")
+
+	if err == nil || len(err.Error()) > 1024 {
+		t.Errorf("error text is %d bytes: a gateway's whole body must not become an error message", len(err.Error()))
+	}
+}

internal/adapters/llm/temperature_test.go

@@ -0,0 +1,57 @@
+package llm
+
+import (
+	"context"
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+	"time"
+)
+
+// capturedTemperature runs one Complete against a fake gateway and returns the
+// "temperature" field the request carried, and whether it was present at all.
+func capturedTemperature(t *testing.T, opts ...Option) (float64, bool) {
+	t.Helper()
+	var body map[string]any
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		raw, _ := io.ReadAll(r.Body)
+		_ = json.Unmarshal(raw, &body)
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(map[string]any{"choices": []map[string]any{{"message": map[string]any{"role": "assistant", "content": "ok"}}}})
+	}))
+	defer srv.Close()
+
+	if _, err := New(srv.URL, "", "m", 5*time.Second, opts...).Complete(context.Background(), "s", "u"); err != nil {
+		t.Fatalf("Complete: %v", err)
+	}
+	v, ok := body["temperature"].(float64)
+	return v, ok
+}
+
+func TestClient_DefaultTemperatureIsUnchanged(t *testing.T) {
+	got, ok := capturedTemperature(t)
+
+	if !ok || got != 0.2 {
+		t.Fatalf("default temperature = %v (present %v), want 0.2: existing callers must not change behaviour", got, ok)
+	}
+}
+
+func TestClient_WithTemperatureZeroSendsExactlyZero(t *testing.T) {
+	// The reconciler's accuracy (92%/97%) was measured at temperature 0. Zero must be
+	// SENT, not treated as "unset" and replaced by the default.
+	got, ok := capturedTemperature(t, WithTemperature(0))
+
+	if !ok || got != 0 {
+		t.Fatalf("temperature = %v (present %v), want an explicit 0", got, ok)
+	}
+}
+
+func TestClient_WithTemperatureSendsTheGivenValue(t *testing.T) {
+	got, ok := capturedTemperature(t, WithTemperature(0.7))
+
+	if !ok || got != 0.7 {
+		t.Fatalf("temperature = %v (present %v), want 0.7", got, ok)
+	}
+}

internal/adapters/secrets/file.go

@@ -18,7 +18,7 @@ import (
 	"path/filepath"
 	"sync"
 
-	"gitea.d-ma.be/mathias/tapir/internal/ports"
+	"git.d-ma.be/mathias/tapir/internal/ports"
 )
 
 // FileStore is a SecretStore backed by a single 0600 JSON file mapping opaque

internal/adapters/secrets/file_test.go

@@ -7,7 +7,7 @@ import (
 	"path/filepath"
 	"testing"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/secrets"
+	"git.d-ma.be/mathias/tapir/internal/adapters/secrets"
 )
 
 func TestPutThenGet(t *testing.T) {

internal/adapters/store/connections_test.go

@@ -6,7 +6,7 @@ import (
 
 	"github.com/stretchr/testify/require"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
 )
 
 // seedUserRow inserts a bare users row (FK target for a connection) as the

internal/adapters/store/identity_test.go

@@ -7,7 +7,7 @@ import (
 
 	"github.com/stretchr/testify/require"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
 )
 
 const (

internal/adapters/store/migrate_test.go

@@ -3,6 +3,7 @@ package store_test
 import (
 	"context"
 	"database/sql"
+	"errors"
 	"os"
 	"testing"
 
@@ -34,6 +35,30 @@ func fileMigrator(t *testing.T) *migrate.Migrate {
 	return m
 }
 
+// headVersion reports the current (HEAD) schema version so a test can restore
+// to it after stepping down, without hard-coding what HEAD is. Adding a
+// migration on top changes HEAD but no test that uses this needs editing.
+func headVersion(t *testing.T, m *migrate.Migrate) uint {
+	t.Helper()
+	v, dirty, err := m.Version()
+	require.NoError(t, err)
+	require.False(t, dirty, "schema must not be dirty")
+	return v
+}
+
+// migrateTo drives the schema to an exact version *by version number*, not by
+// step count. This is the whole point of the migrate-test design: a migration
+// added above the target does not shift any count here, so unrelated tests stay
+// green (see issue #8). ErrNoChange (already at that version) is not a failure.
+func migrateTo(t *testing.T, m *migrate.Migrate, version uint) {
+	t.Helper()
+	err := m.Migrate(version)
+	if errors.Is(err, migrate.ErrNoChange) {
+		return
+	}
+	require.NoError(t, err)
+}
+
 // loginEventsExists reports whether the login_events relation is present.
 func loginEventsExists(t *testing.T) bool {
 	t.Helper()
@@ -53,25 +78,15 @@ func TestMigration010LoginEventsUpDown(t *testing.T) {
 	require.True(t, loginEventsExists(t), "login_events must exist at latest migration")
 
 	m := fileMigrator(t)
-	// 011..016 sit above 010; step them down first so 010 is exercised in isolation.
-	require.NoError(t, m.Steps(-1), "down 016 drops channel_caption_state, login_events intact")
-	require.True(t, loginEventsExists(t), "016 down leaves login_events intact")
-	require.NoError(t, m.Steps(-1), "down 015 reshapes transcripts, login_events intact")
-	require.True(t, loginEventsExists(t), "015 down leaves login_events intact")
-	require.NoError(t, m.Steps(-1), "down 014 drops channel_title, login_events intact")
-	require.True(t, loginEventsExists(t), "014 down leaves login_events intact")
-	require.NoError(t, m.Steps(-1), "down 013 drops channel_errors, login_events intact")
-	require.True(t, loginEventsExists(t), "013 down leaves login_events intact")
-	require.NoError(t, m.Steps(-1), "down 012 is a no-op, login_events intact")
-	require.True(t, loginEventsExists(t), "012 down leaves login_events intact")
-	require.NoError(t, m.Steps(-1), "down 011 must not touch login_events")
-	require.True(t, loginEventsExists(t), "011 down leaves login_events intact")
+	head := headVersion(t, m)
 
-	require.NoError(t, m.Steps(-1), "down 010 must drop login_events")
+	migrateTo(t, m, 9) // just below 010 — everything above steps down
 	require.False(t, loginEventsExists(t), "login_events must be gone after the down migration")
 
-	require.NoError(t, m.Steps(7), "up must recreate 010 then re-apply 011..016")
+	migrateTo(t, m, 10) // up 010
 	require.True(t, loginEventsExists(t), "login_events must be restored after the up migration")
+
+	migrateTo(t, m, head) // restore to HEAD for sibling tests
 }
 
 // autoSummarizeDefault reads the users.auto_summarize column default as text
@@ -93,21 +108,15 @@ func TestMigration011AutoSummarizeDefaultUpDown(t *testing.T) {
 	require.Equal(t, "true", autoSummarizeDefault(t), "011 sets the default to TRUE")
 
 	m := fileMigrator(t)
-	require.NoError(t, m.Steps(-1), "down 016 drops channel_caption_state")
-	require.NoError(t, m.Steps(-1), "down 015 reshapes transcripts")
-	require.NoError(t, m.Steps(-1), "down 014 drops channel_title")
-	require.NoError(t, m.Steps(-1), "down 013 drops channel_errors")
-	require.NoError(t, m.Steps(-1), "down 012 is a no-op")
-	require.NoError(t, m.Steps(-1), "down 011 reverts the column default")
+	head := headVersion(t, m)
+
+	migrateTo(t, m, 10) // just below 011 — reverts the column default
 	require.Equal(t, "false", autoSummarizeDefault(t), "default is FALSE after the down migration")
 
-	require.NoError(t, m.Steps(1), "up 011 re-applies the TRUE default")
+	migrateTo(t, m, 11) // up 011 re-applies the TRUE default
 	require.Equal(t, "true", autoSummarizeDefault(t))
-	require.NoError(t, m.Steps(1), "up 012 runs clean (no FORCE RLS on fresh schema)")
-	require.NoError(t, m.Steps(1), "up 013 creates channel_errors")
-	require.NoError(t, m.Steps(1), "up 014 recreates channel_title")
-	require.NoError(t, m.Steps(1), "up 015 reshapes transcripts to shared")
-	require.NoError(t, m.Steps(1), "up 016 recreates channel_caption_state (HEAD)")
+
+	migrateTo(t, m, head) // restore to HEAD for sibling tests
 }
 
 // channelTitleExists reports whether videos.channel_title is present.
@@ -127,17 +136,15 @@ func TestMigration014VideoChannelTitleUpDown(t *testing.T) {
 	require.True(t, channelTitleExists(t), "channel_title exists at latest migration")
 
 	m := fileMigrator(t)
-	require.NoError(t, m.Steps(-1), "down 016 drops channel_caption_state, channel_title intact")
-	require.True(t, channelTitleExists(t), "016 down leaves channel_title intact")
-	require.NoError(t, m.Steps(-1), "down 015 reshapes transcripts, channel_title intact")
-	require.True(t, channelTitleExists(t), "015 down leaves channel_title intact")
-	require.NoError(t, m.Steps(-1), "down 014 must drop channel_title")
+	head := headVersion(t, m)
+
+	migrateTo(t, m, 13) // just below 014 — drops channel_title
 	require.False(t, channelTitleExists(t), "channel_title must be gone after the down migration")
 
-	require.NoError(t, m.Steps(1), "up 014 must recreate channel_title")
+	migrateTo(t, m, 14) // up 014 recreates channel_title
 	require.True(t, channelTitleExists(t), "channel_title must be restored after the up migration")
-	require.NoError(t, m.Steps(1), "up 015 restores the shared transcripts shape")
-	require.NoError(t, m.Steps(1), "up 016 recreates channel_caption_state (HEAD)")
+
+	migrateTo(t, m, head) // restore to HEAD for sibling tests
 }
 
 // TestMigration012FixAutoSummarizeRLS proves 012 runs cleanly and flips any
@@ -148,7 +155,11 @@ func TestMigration012FixAutoSummarizeRLS(t *testing.T) {
 
 	// Round-trip: down 012, then up 012 — must be idempotent.
 	m := fileMigrator(t)
-	require.NoError(t, m.Steps(-1), "down 012 must not error")
-	require.NoError(t, m.Steps(1), "up 012 must re-apply cleanly")
+	head := headVersion(t, m)
+
+	migrateTo(t, m, 11) // down 012 must not error
+	migrateTo(t, m, 12) // up 012 must re-apply cleanly
 	require.Equal(t, "true", autoSummarizeDefault(t), "default still TRUE after 012 re-applied")
+
+	migrateTo(t, m, head) // restore to HEAD for sibling tests
 }

internal/adapters/store/migrations/017_uploaded_transcripts.down.sql

@@ -0,0 +1,3 @@
+-- Drops the private uploaded-transcript store and every transcript in it. There
+-- is no revert of the data: the original audio was never kept (ADR-033).
+DROP TABLE uploaded_transcripts;

internal/adapters/store/migrations/017_uploaded_transcripts.up.sql

@@ -0,0 +1,48 @@
+-- Migration 017: uploaded_transcripts — PRIVATE transcripts of audio a user
+-- uploaded themselves (ADR-033, tapir#33).
+--
+-- This is deliberately NOT the shared `transcripts` table (migration 015,
+-- ADR-021). That table is non-RLS on the explicit grounds that it holds only
+-- public caption content and nothing user-identifying. A recorded meeting is the
+-- opposite: private third-party content owned by exactly one user. Reusing
+-- `transcripts` — or adding a user_id to it with a partial policy — would put a
+-- private row in the one table the isolation test asserts is shared, and would
+-- pass CI green. A separate table keeps ADR-021's exception singular.
+--
+-- Row-level security follows migration 003 exactly: ENABLE + FORCE (the app
+-- connects as the table owner, which bypasses RLS unless forced), and a policy
+-- keyed off the per-request GUC tapir.current_user_id. With no GUC set the
+-- predicate is NULL and the table is deny-all.
+--
+-- The original audio is never stored (ADR-033): this row is all that is kept.
+-- segments holds the timestamped cues; re-transcribing 90 minutes costs ~7
+-- minutes, so discarding the timestamps would be irreversible.
+CREATE TABLE uploaded_transcripts (
+    id            UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    user_id       UUID NOT NULL REFERENCES users(id) ON DELETE CASCADE,
+    title         TEXT NOT NULL,
+    status        TEXT NOT NULL DEFAULT 'queued'
+                  CHECK (status IN ('queued', 'transcribing', 'done', 'failed')),
+    content       TEXT,
+    segments      JSONB,
+    endpoint      TEXT,             -- which backend served it (iguana | berget); a quality fact
+    audio_seconds DOUBLE PRECISION,
+    error_kind    TEXT,             -- typed, never a provider's error string (ADR-033)
+    created_at    TIMESTAMPTZ NOT NULL DEFAULT NOW(),
+    finished_at   TIMESTAMPTZ,
+    CONSTRAINT uploaded_transcripts_done_has_content
+        CHECK (status <> 'done' OR content IS NOT NULL),
+    CONSTRAINT uploaded_transcripts_failed_has_error_kind
+        CHECK (status <> 'failed' OR error_kind IS NOT NULL)
+);
+
+CREATE INDEX idx_uploaded_transcripts_user ON uploaded_transcripts (user_id, created_at DESC);
+
+ALTER TABLE uploaded_transcripts ENABLE ROW LEVEL SECURITY;
+ALTER TABLE uploaded_transcripts FORCE ROW LEVEL SECURITY;
+CREATE POLICY uploaded_transcripts_isolation ON uploaded_transcripts
+    FOR ALL
+    USING (user_id = current_setting('tapir.current_user_id', true)::uuid);
+
+COMMENT ON TABLE uploaded_transcripts IS
+    'PRIVATE per-user transcripts of user-supplied audio. RLS-scoped (ADR-012). Never merge into the shared transcripts table (ADR-021, ADR-033).';

internal/adapters/store/migrations/018_uploaded_transcripts_had_video.down.sql

@@ -0,0 +1 @@
+ALTER TABLE uploaded_transcripts DROP COLUMN had_video;

internal/adapters/store/migrations/018_uploaded_transcripts_had_video.up.sql

@@ -0,0 +1,9 @@
+-- Migration 018: record whether an uploaded file carried a video track (ADR-033,
+-- amended 2026-10-01).
+--
+-- A Teams meeting with the camera off is still an .mp4 with a video track, so video
+-- files are ACCEPTED: only the audio is extracted and the video is discarded. This
+-- flag lets the transcript page say so — "this file contained a video track; only its
+-- audio was used" — so the user is never left to wonder what happened to the picture.
+-- It outlives the job: completing or failing an upload does not clear it.
+ALTER TABLE uploaded_transcripts ADD COLUMN had_video BOOLEAN NOT NULL DEFAULT FALSE;

internal/adapters/store/migrations/019_uploaded_transcripts_reconciliation.down.sql

@@ -0,0 +1 @@
+ALTER TABLE uploaded_transcripts DROP COLUMN reconciliation;

internal/adapters/store/migrations/019_uploaded_transcripts_reconciliation.up.sql

@@ -0,0 +1,12 @@
+-- Migration 019: how a merged transcript came about (ADR-034).
+--
+-- When two engines transcribed the same audio and a local model chose between their
+-- disagreements, this records the status, the exact counts, and a bounded list of the
+-- contested spots (changed or unsure, each with a position and both readings) so the
+-- page can say what was done and the user can check the uncertain spots against the
+-- audio. NULL means a single engine: no second opinion, and the page must not claim one.
+--
+-- It lives on the already-FORCE'd, RLS-scoped uploaded_transcripts, so the exact-set
+-- isolation test is unaffected. The contested spots hold fragments of the private
+-- transcript, which is why they are here and nowhere else.
+ALTER TABLE uploaded_transcripts ADD COLUMN reconciliation JSONB;

internal/adapters/store/migrations/020_uploaded_transcripts_progress.down.sql

@@ -0,0 +1,4 @@
+ALTER TABLE uploaded_transcripts
+    DROP COLUMN alive_at,
+    DROP COLUMN progress_at,
+    DROP COLUMN progress;

internal/adapters/store/migrations/020_uploaded_transcripts_progress.up.sql

@@ -0,0 +1,15 @@
+-- Migration 020: where an upload job is, and whether the service is still reporting (ADR-035).
+--
+-- progress    the job's ordered phases as JSON (key, label, state, times, done/total).
+-- progress_at when a phase last REALLY changed or advanced.
+-- alive_at    a heartbeat stamped every few seconds by a goroutine independent of any
+--             engine call. The two differ on purpose: alive says the service is still
+--             reporting, progress says work is moving, and a page that cannot tell a
+--             crashed service from a slow model is the failure this exists to end.
+--
+-- All three are NULL until the job starts. They sit on the already-FORCE'd, RLS-scoped
+-- uploaded_transcripts, so the exact-set isolation test is unaffected.
+ALTER TABLE uploaded_transcripts
+    ADD COLUMN progress    JSONB,
+    ADD COLUMN progress_at TIMESTAMPTZ,
+    ADD COLUMN alive_at    TIMESTAMPTZ;

internal/adapters/store/migrations/021_uploaded_transcripts_language.down.sql

@@ -0,0 +1,2 @@
+ALTER TABLE uploaded_transcripts
+    DROP COLUMN language;

internal/adapters/store/migrations/021_uploaded_transcripts_language.up.sql

@@ -0,0 +1,13 @@
+-- Migration 021: the language an upload is transcribed as (tapir#61).
+--
+-- 'sv'   Swedish: the engines are sent language=sv. Every row before this migration was
+--        transcribed that way (the request hard-coded it), so the default is true history.
+-- 'auto' Other: the engine detects the language and the Swedish-only engine is not used.
+--
+-- A closed set held by the table as well as by the form and the request code, so a second
+-- writer or a bug cannot store a value the engines are never meant to receive. The column
+-- sits on the already-FORCE'd, RLS-scoped uploaded_transcripts, so the exact-set isolation
+-- test is unaffected.
+ALTER TABLE uploaded_transcripts
+    ADD COLUMN language TEXT NOT NULL DEFAULT 'sv'
+        CONSTRAINT uploaded_transcripts_language_check CHECK (language IN ('sv', 'auto'));

internal/adapters/store/reads_test.go

@@ -8,7 +8,7 @@ import (
 	"github.com/jackc/pgx/v5/pgxpool"
 	"github.com/stretchr/testify/require"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
 )
 
 // seedVideo inserts a videos row whose id matches a summary's video_id, so the

internal/adapters/store/store.go

@@ -24,7 +24,7 @@ import (
 
 	_ "github.com/jackc/pgx/v5/stdlib" // register the "pgx" database/sql driver for migrate
 
-	"gitea.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/domain"
 )
 
 //go:embed migrations/*.sql

internal/adapters/store/store_test.go

@@ -11,9 +11,9 @@ import (
 	"github.com/jackc/pgx/v5/pgxpool"
 	"github.com/stretchr/testify/require"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
-	"gitea.d-ma.be/mathias/tapir/internal/domain"
-	"gitea.d-ma.be/mathias/tapir/internal/ports"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/ports"
 )
 
 // Static check: Store satisfies the Sink port.

internal/adapters/store/summarize_mode_test.go

@@ -8,7 +8,7 @@ import (
 	"github.com/jackc/pgx/v5/pgxpool"
 	"github.com/stretchr/testify/require"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
 )
 
 // seedBareVideo inserts a videos row with no summary, so the all-videos read and

internal/adapters/store/transcript.go

@@ -7,7 +7,7 @@ import (
 
 	"github.com/jackc/pgx/v5"
 
-	"gitea.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/domain"
 )
 
 // GetTranscript returns the shared, stored transcript for a video keyed by the

internal/adapters/store/transcript_status_test.go

@@ -6,7 +6,7 @@ import (
 
 	"github.com/stretchr/testify/require"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
 )
 
 func TestSetTranscriptStatus_RoundTrip(t *testing.T) {

internal/adapters/store/transcript_test.go

@@ -6,9 +6,9 @@ import (
 
 	"github.com/stretchr/testify/require"
 
-	"gitea.d-ma.be/mathias/tapir/internal/adapters/store"
-	"gitea.d-ma.be/mathias/tapir/internal/domain"
-	"gitea.d-ma.be/mathias/tapir/internal/ports"
+	"git.d-ma.be/mathias/tapir/internal/adapters/store"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/ports"
 )
 
 // Static check: Store satisfies the shared TranscriptStore port (ADR-021).

internal/adapters/store/uploaded_transcripts.go

@@ -0,0 +1,288 @@
+package store
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"time"
+
+	"github.com/jackc/pgx/v5"
+
+	"git.d-ma.be/mathias/tapir/internal/domain"
+)
+
+// ErrUploadNotFound is returned when an uploaded transcript does not exist OR is
+// not the caller's. The two are deliberately indistinguishable: a non-owner learns
+// 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
+// caller's rows; each statement ALSO names user_id in its predicate, so the
+// isolation does not rest on a single mechanism and stays testable under the
+// superuser connection the unit tests use.
+
+// CreateUpload records a new upload in the queued state and returns its id.
+// hadVideo records that the file carried a video track (only its audio is used);
+// it is kept for the life of the row so the page can say so.
+func (s *Store) CreateUpload(ctx context.Context, userID, title string, lang domain.Language, hadVideo bool) (string, error) {
+	var id string
+	if err := s.withUser(ctx, userID, func(tx pgx.Tx) error {
+		return tx.QueryRow(ctx,
+			`INSERT INTO uploaded_transcripts (user_id, title, language, had_video) VALUES ($1, $2, $3, $4) RETURNING id`,
+			userID, title, string(lang), hadVideo).Scan(&id)
+	}); err != nil {
+		return "", fmt.Errorf("store: create upload: %w", err)
+	}
+	return id, nil
+}
+
+// StartTranscribing moves an upload to transcribing.
+func (s *Store) StartTranscribing(ctx context.Context, userID, id string) error {
+	return s.updateUpload(ctx, userID, id, "start transcribing",
+		`UPDATE uploaded_transcripts SET status = 'transcribing'
+		  WHERE id = $1 AND user_id = $2`)
+}
+
+// CompleteUpload stores the finished transcription and moves the upload to done.
+func (s *Store) CompleteUpload(ctx context.Context, userID, id string, t domain.Transcription) error {
+	segs, err := json.Marshal(t.Segments)
+	if err != nil {
+		return fmt.Errorf("store: complete upload: marshal segments: %w", err)
+	}
+	// A nil Reconciliation must be SQL NULL, not the JSON text "null": the page decides
+	// whether to claim a second opinion by whether this column is NULL.
+	var rec []byte
+	if t.Reconciliation != nil {
+		if rec, err = json.Marshal(t.Reconciliation); err != nil {
+			return fmt.Errorf("store: complete upload: marshal reconciliation: %w", err)
+		}
+	}
+	return s.updateUpload(ctx, userID, id, "complete upload",
+		`UPDATE uploaded_transcripts
+		    SET status = 'done', content = $3, segments = $4, endpoint = $5,
+		        audio_seconds = $6, reconciliation = $7, error_kind = NULL, finished_at = NOW()
+		  WHERE id = $1 AND user_id = $2`,
+		t.Text, segs, t.Endpoint, t.AudioSeconds, rec)
+}
+
+// SaveProgress records the job's phases and when something last really changed
+// (ADR-035). It replaces the previous snapshot: the Tracker holds the whole plan.
+func (s *Store) SaveProgress(ctx context.Context, userID, id string, steps []domain.ProgressStep, at time.Time) error {
+	b, err := json.Marshal(steps)
+	if err != nil {
+		return fmt.Errorf("store: save progress: marshal: %w", err)
+	}
+	return s.updateUpload(ctx, userID, id, "save progress",
+		`UPDATE uploaded_transcripts SET progress = $3, progress_at = $4
+		  WHERE id = $1 AND user_id = $2`, b, at)
+}
+
+// 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 = $3
+		  WHERE id = $1 AND user_id = $2`, at)
+}
+
+// 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 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).
+func (s *Store) updateUpload(ctx context.Context, userID, id, op, sql string, args ...any) error {
+	all := append([]any{id, userID}, args...)
+	return s.withUser(ctx, userID, func(tx pgx.Tx) error {
+		ct, err := tx.Exec(ctx, sql, all...)
+		if err != nil {
+			return fmt.Errorf("store: %s: %w", op, err)
+		}
+		if ct.RowsAffected() == 0 {
+			return ErrUploadNotFound
+		}
+		return nil
+	})
+}
+
+// GetUpload returns one of the caller's uploads, including the transcript text.
+func (s *Store) GetUpload(ctx context.Context, userID, id string) (*domain.UploadedTranscript, error) {
+	var u domain.UploadedTranscript
+	err := s.withUser(ctx, userID, func(tx pgx.Tx) error {
+		row := tx.QueryRow(ctx, uploadSelect+` WHERE id = $1 AND user_id = $2`, id, userID)
+		return scanUpload(row, &u, true)
+	})
+	if errors.Is(err, pgx.ErrNoRows) {
+		return nil, ErrUploadNotFound
+	}
+	if err != nil {
+		return nil, fmt.Errorf("store: get upload: %w", err)
+	}
+	return &u, nil
+}
+
+// ListUploads returns the caller's uploads, newest first, WITHOUT transcript text
+// or segments: a list view must not drag every transcript along.
+func (s *Store) ListUploads(ctx context.Context, userID string) ([]domain.UploadedTranscript, error) {
+	var out []domain.UploadedTranscript
+	err := s.withUser(ctx, userID, func(tx pgx.Tx) error {
+		rows, err := tx.Query(ctx, uploadSelect+` WHERE user_id = $1 ORDER BY created_at DESC, id`, userID)
+		if err != nil {
+			return err
+		}
+		defer rows.Close()
+		for rows.Next() {
+			var u domain.UploadedTranscript
+			if err := scanUpload(rows, &u, false); err != nil {
+				return err
+			}
+			out = append(out, u)
+		}
+		return rows.Err()
+	})
+	if err != nil {
+		return nil, fmt.Errorf("store: list uploads: %w", err)
+	}
+	return out, nil
+}
+
+// FailInterruptedUploads is the start-up sweep. The audio of an upload is never
+// retained (ADR-033), so a job that was queued or transcribing when the process
+// stopped can never resume; leaving it would be a stuck spinner. Every such row,
+// for every user, becomes failed/interrupted and the user re-uploads. It
+// enumerates users from the un-RLS'd identity map and scopes each pass with
+// withUser, exactly as scheduled discovery does (ADR-018). Returns rows failed.
+func (s *Store) FailInterruptedUploads(ctx context.Context) (int, error) {
+	users, err := s.ListAllUsers(ctx)
+	if err != nil {
+		return 0, fmt.Errorf("store: fail interrupted uploads: %w", err)
+	}
+	total := 0
+	for _, u := range users {
+		err := s.withUser(ctx, u.UserID, func(tx pgx.Tx) error {
+			ct, err := tx.Exec(ctx,
+				`UPDATE uploaded_transcripts
+				    SET status = 'failed', error_kind = $2, finished_at = NOW()
+				  WHERE user_id = $1 AND status IN ('queued', 'transcribing')`,
+				u.UserID, string(domain.UploadErrInterrupted))
+			total += int(ct.RowsAffected())
+			return err
+		})
+		if err != nil {
+			return total, fmt.Errorf("store: fail interrupted uploads: user %s: %w", u.UserID, err)
+		}
+	}
+	return total, nil
+}
+
+const uploadSelect = `SELECT id, user_id, title, language, status, content, segments, endpoint,
+       audio_seconds, error_kind, had_video, reconciliation, progress, progress_at, alive_at,
+       created_at, finished_at FROM uploaded_transcripts`
+
+// scanUpload scans one uploadSelect row. withText controls whether the transcript
+// body and segments are kept (GetUpload) or dropped (ListUploads).
+func scanUpload(row pgx.Row, u *domain.UploadedTranscript, withText bool) error {
+	var (
+		status, errKind, endpoint, content *string
+		segments, reconciliation, progress []byte
+		seconds                            *float64
+		finished, progressAt, aliveAt      *time.Time
+	)
+	var language string
+	if err := row.Scan(&u.ID, &u.UserID, &u.Title, &language, &status, &content, &segments, &endpoint,
+		&seconds, &errKind, &u.HadVideo, &reconciliation, &progress, &progressAt, &aliveAt,
+		&u.CreatedAt, &finished); err != nil {
+		return err
+	}
+	u.Language = domain.Language(language)
+	u.Status = domain.UploadStatus(deref(status))
+	u.Endpoint = deref(endpoint)
+	u.ErrorKind = domain.UploadErrorKind(deref(errKind))
+	if seconds != nil {
+		u.AudioSeconds = *seconds
+	}
+	if finished != nil {
+		u.FinishedAt = *finished
+	}
+	if !withText {
+		return nil
+	}
+	if len(progress) > 0 {
+		if err := json.Unmarshal(progress, &u.Progress); err != nil {
+			return fmt.Errorf("unmarshal progress: %w", err)
+		}
+	}
+	if progressAt != nil {
+		u.ProgressAt = *progressAt
+	}
+	if aliveAt != nil {
+		u.AliveAt = *aliveAt
+	}
+	if len(reconciliation) > 0 {
+		var r domain.Reconciliation
+		if err := json.Unmarshal(reconciliation, &r); err != nil {
+			return fmt.Errorf("unmarshal reconciliation: %w", err)
+		}
+		u.Reconciliation = &r
+	}
+	u.Content = deref(content)
+	if len(segments) > 0 {
+		if err := json.Unmarshal(segments, &u.Segments); err != nil {
+			return fmt.Errorf("unmarshal segments: %w", err)
+		}
+	}
+	return nil
+}
+
+func deref(s *string) string {
+	if s == nil {
+		return ""
+	}
+	return *s
+}

internal/adapters/store/uploaded_transcripts_delete_test.go

@@ -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", domain.LangSwedish, 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")
+}

internal/adapters/store/uploaded_transcripts_failguard_test.go

@@ -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", domain.LangSwedish, 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)
+	}
+}

internal/adapters/store/uploaded_transcripts_isolation_test.go

@@ -0,0 +1,196 @@
+package store_test
+
+import (
+	"context"
+	"sort"
+	"testing"
+
+	"github.com/jackc/pgx/v5/pgxpool"
+	"github.com/stretchr/testify/require"
+)
+
+// Isolation proof for ADR-033: an uploaded transcript is PRIVATE content owned by
+// one user, so it lives in its own RLS-scoped table — never in the shared,
+// non-RLS `transcripts` (ADR-021). These tests drive the database directly (raw
+// SQL as the non-superuser `app` role) so they observe row-level security itself,
+// not any application-layer WHERE clause. The superuser pool is used only to seed
+// and to read pg_catalog; it bypasses RLS and so must never be the pool whose
+// answers the assertions rest on.
+
+// seedUploadedTranscript inserts one finished upload for userID as the superuser
+// (RLS bypassed so both users' rows land regardless of any GUC) and returns its id.
+// The text is synthetic — this repo is public, and a real transcript must never
+// appear in it (#33).
+func seedUploadedTranscript(t *testing.T, super *pgxpool.Pool, userID, text string) string {
+	t.Helper()
+	ctx := context.Background()
+	_, err := super.Exec(ctx, `INSERT INTO users (id) VALUES ($1) ON CONFLICT DO NOTHING`, userID)
+	require.NoError(t, err)
+	var id string
+	require.NoError(t, super.QueryRow(ctx,
+		`INSERT INTO uploaded_transcripts (user_id, title, status, content)
+		 VALUES ($1, 'synthetic fixture', 'done', $2) RETURNING id`,
+		userID, text).Scan(&id))
+	return id
+}
+
+func TestUploadedTranscriptsRLS_ScopedReadsSeeOnlyOwnRows(t *testing.T) {
+	newStore(t)
+	super := rawPool(t)
+	resetDB(t, super)
+	seedUploadedTranscript(t, super, userA, "text owned by A")
+	seedUploadedTranscript(t, super, userB, "text owned by B")
+	app := appPool(t, super)
+
+	require.Equal(t, 0, scopedCount(t, app, "", "uploaded_transcripts"),
+		"no tapir.current_user_id must be deny-all: this proves RLS is live, not bypassed")
+	require.Equal(t, 1, scopedCount(t, app, userA, "uploaded_transcripts"),
+		"user A must see exactly its own upload")
+	require.Equal(t, 1, scopedCount(t, app, userB, "uploaded_transcripts"),
+		"user B must see exactly its own upload")
+
+	// Count alone cannot distinguish "A's row" from "B's row" — read the content.
+	var got string
+	ctx := context.Background()
+	tx, err := app.Begin(ctx)
+	require.NoError(t, err)
+	defer tx.Rollback(ctx) //nolint:errcheck
+	_, err = tx.Exec(ctx, `SELECT set_config('tapir.current_user_id', $1, true)`, userA)
+	require.NoError(t, err)
+	require.NoError(t, tx.QueryRow(ctx, `SELECT content FROM uploaded_transcripts`).Scan(&got))
+	require.Equal(t, "text owned by A", got, "A's scoped read must return A's text, never B's")
+}
+
+func TestUploadedTranscriptsRLS_CrossUserWritesAffectZeroRows(t *testing.T) {
+	newStore(t)
+	super := rawPool(t)
+	resetDB(t, super)
+	seedUploadedTranscript(t, super, userA, "text owned by A")
+	bID := seedUploadedTranscript(t, super, userB, "text owned by B")
+	app := appPool(t, super)
+
+	require.Equal(t, int64(0), scopedRowsAffected(t, app, userA,
+		`UPDATE uploaded_transcripts SET content = 'hacked' WHERE id = $1`, bID),
+		"user A scoped UPDATE must touch zero of B's uploads")
+	require.Equal(t, int64(0), scopedRowsAffected(t, app, userA,
+		`DELETE FROM uploaded_transcripts WHERE id = $1`, bID),
+		"user A scoped DELETE must touch zero of B's uploads")
+
+	var content string
+	require.NoError(t, super.QueryRow(context.Background(),
+		`SELECT content FROM uploaded_transcripts WHERE id = $1`, bID).Scan(&content))
+	require.Equal(t, "text owned by B", content, "B's upload must survive A's writes unchanged")
+}
+
+func TestUploadedTranscriptsRLS_InsertForAnotherUserIsRejected(t *testing.T) {
+	newStore(t)
+	super := rawPool(t)
+	resetDB(t, super)
+	seedUploadedTranscript(t, super, userA, "text owned by A")
+	_, err := super.Exec(context.Background(), `INSERT INTO users (id) VALUES ($1)`, userB)
+	require.NoError(t, err)
+	app := appPool(t, super)
+
+	ctx := context.Background()
+	tx, err := app.Begin(ctx)
+	require.NoError(t, err)
+	defer tx.Rollback(ctx) //nolint:errcheck
+	_, err = tx.Exec(ctx, `SELECT set_config('tapir.current_user_id', $1, true)`, userA)
+	require.NoError(t, err)
+
+	// Scoped as A, plant a row that claims to belong to B. The policy's WITH CHECK
+	// (defaulting to USING under FOR ALL) must refuse it.
+	_, err = tx.Exec(ctx,
+		`INSERT INTO uploaded_transcripts (user_id, title, status, content)
+		 VALUES ($1, 'planted', 'done', 'planted')`, userB)
+	require.Error(t, err, "a row scoped as A must not be insertable on behalf of B")
+	require.Contains(t, err.Error(), "row-level security")
+}
+
+func TestUploadedTranscripts_RowSecurityIsEnabledForcedAndKeyedByUser(t *testing.T) {
+	newStore(t)
+	super := rawPool(t)
+	ctx := context.Background()
+
+	var enabled, forced bool
+	require.NoError(t, super.QueryRow(ctx,
+		`SELECT relrowsecurity, relforcerowsecurity FROM pg_class
+		  WHERE oid = 'public.uploaded_transcripts'::regclass`).Scan(&enabled, &forced))
+	require.True(t, enabled, "uploaded_transcripts must have row-level security enabled")
+	require.True(t, forced,
+		"FORCE is what binds the table-owner role tapir; without it the policy is dead in production")
+
+	var userIDNotNull bool
+	require.NoError(t, super.QueryRow(ctx,
+		`SELECT is_nullable = 'NO' FROM information_schema.columns
+		  WHERE table_name = 'uploaded_transcripts' AND column_name = 'user_id'`).Scan(&userIDNotNull))
+	require.True(t, userIDNotNull, "an upload with no owner would match no policy and be unreachable")
+}
+
+func TestUploadedTranscripts_DeleteUserRemovesTheirUploads(t *testing.T) {
+	s := newStore(t)
+	super := rawPool(t)
+	resetDB(t, super)
+	seedUploadedTranscript(t, super, userA, "text owned by A")
+	bID := seedUploadedTranscript(t, super, userB, "text owned by B")
+
+	require.NoError(t, s.DeleteUser(context.Background(), userA))
+
+	var aRows, bRows int
+	ctx := context.Background()
+	require.NoError(t, super.QueryRow(ctx,
+		`SELECT count(*) FROM uploaded_transcripts WHERE user_id = $1`, userA).Scan(&aRows))
+	require.NoError(t, super.QueryRow(ctx,
+		`SELECT count(*) FROM uploaded_transcripts WHERE id = $1`, bID).Scan(&bRows))
+	require.Equal(t, 0, aRows, "deleting a user must remove their private transcripts (ON DELETE CASCADE)")
+	require.Equal(t, 1, bRows, "deleting A must not touch B's transcripts")
+}
+
+// rowSecurityOffTables are the public-schema tables that deliberately have row
+// security OFF. The list is exact on purpose: a NEW table added without RLS must
+// fail this test and force a decision, instead of passing CI by omission — the
+// failure mode ADR-033 names, where a naive reuse "leaks private content across
+// tenants and passes CI green".
+//
+//   - transcripts: shared public caption content, the single exception for CONTENT
+//     (ADR-021). It is the one that must never gain a private row.
+//   - user_identities: the Dex-subject -> user_id map, read at login before any
+//     user is known (migration 004). Holds no user content.
+//   - invitations: exists before the user does, so there is no user to scope by;
+//     the single-use token is the capability (migration 009).
+//   - schema_migrations: golang-migrate bookkeeping.
+//
+// #33 describes `transcripts` as "the only non-RLS table". That is true of content
+// and false of tables; this test asserts the real set.
+var rowSecurityOffTables = []string{
+	"invitations",
+	"schema_migrations",
+	"transcripts",
+	"user_identities",
+}
+
+func TestRowSecurityOffTablesAreExactlyTheKnownSet(t *testing.T) {
+	newStore(t)
+	super := rawPool(t)
+
+	rows, err := super.Query(context.Background(),
+		`SELECT c.relname FROM pg_class c
+		   JOIN pg_namespace n ON n.oid = c.relnamespace
+		  WHERE n.nspname = 'public' AND c.relkind IN ('r', 'p') AND NOT c.relrowsecurity
+		  ORDER BY c.relname`)
+	require.NoError(t, err)
+	defer rows.Close()
+	var got []string
+	for rows.Next() {
+		var name string
+		require.NoError(t, rows.Scan(&name))
+		got = append(got, name)
+	}
+	require.NoError(t, rows.Err())
+
+	want := append([]string{}, rowSecurityOffTables...)
+	sort.Strings(want)
+	require.Equal(t, want, got,
+		"every public table must have row security ON unless it is in rowSecurityOffTables; "+
+			"adding a table here is a decision that needs an ADR")
+}

internal/adapters/store/uploaded_transcripts_progress_test.go

@@ -0,0 +1,137 @@
+package store_test
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"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"
+)
+
+// Progress and heartbeat of an upload job (ADR-035, migration 020). Like the other
+// upload store tests these run as the superuser, so the cross-user cases prove the
+// statement's own user_id predicate; the database-enforced isolation is covered by the
+// RLS tests, and migration 020 adds columns to the already-FORCE'd table, not a table.
+
+func startedUpload(t *testing.T, s *store.Store) string {
+	t.Helper()
+	seedUsers(t, userA, userB)
+	id, err := s.CreateUpload(context.Background(), userA, "m", domain.LangSwedish, false)
+	require.NoError(t, err)
+	return id
+}
+
+func TestUploadProgress_AFreshUploadHasNoProgressAndNoHeartbeat(t *testing.T) {
+	s := newStore(t)
+	id := startedUpload(t, s)
+
+	got, err := s.GetUpload(context.Background(), userA, id)
+
+	require.NoError(t, err)
+	assert.Empty(t, got.Progress)
+	assert.True(t, got.ProgressAt.IsZero())
+	assert.True(t, got.AliveAt.IsZero())
+}
+
+func TestUploadProgress_StepsAndTheirTimesRoundTripInOrder(t *testing.T) {
+	s := newStore(t)
+	id := startedUpload(t, s)
+	ctx := context.Background()
+	at := time.Date(2026, 10, 2, 8, 30, 0, 0, time.UTC)
+	steps := []domain.ProgressStep{
+		{Key: "audio", Label: "Preparing audio", State: domain.StepDone, StartedAt: at, FinishedAt: at.Add(40 * time.Second)},
+		{Key: "decide", Label: "Deciding", State: domain.StepRunning, StartedAt: at.Add(time.Minute), Done: 12, Total: 282},
+		{Key: "save", Label: "Saving", State: domain.StepPending},
+	}
+
+	require.NoError(t, s.SaveProgress(ctx, userA, id, steps, at.Add(2*time.Minute)))
+
+	got, err := s.GetUpload(ctx, userA, id)
+	require.NoError(t, err)
+	require.Len(t, got.Progress, 3)
+	assert.Equal(t, []string{"audio", "decide", "save"}, []string{got.Progress[0].Key, got.Progress[1].Key, got.Progress[2].Key})
+	assert.Equal(t, domain.StepRunning, got.Progress[1].State)
+	assert.Equal(t, 12, got.Progress[1].Done)
+	assert.Equal(t, 282, got.Progress[1].Total)
+	assert.True(t, got.Progress[0].FinishedAt.Equal(at.Add(40*time.Second)), "a step's finish time must survive the round trip")
+	assert.True(t, got.Progress[2].StartedAt.IsZero(), "a pending step has no start time, not a zero date")
+	assert.True(t, got.ProgressAt.Equal(at.Add(2*time.Minute)))
+}
+
+func TestUploadProgress_AHeartbeatMovesAliveAtAndNotProgressAt(t *testing.T) {
+	s := newStore(t)
+	id := startedUpload(t, s)
+	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))
+	beatAt := at.Add(90 * time.Second) // deliberately not the database's NOW()
+
+	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.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")
+}
+
+func TestUploadProgress_ASecondSaveReplacesTheFirst(t *testing.T) {
+	s := newStore(t)
+	id := startedUpload(t, s)
+	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: "a", Label: "A", State: domain.StepRunning}}, at))
+
+	require.NoError(t, s.SaveProgress(ctx, userA, id, []domain.ProgressStep{{Key: "a", Label: "A", State: domain.StepDone}, {Key: "b", Label: "B", State: domain.StepRunning}}, at.Add(time.Minute)))
+
+	got, err := s.GetUpload(ctx, userA, id)
+	require.NoError(t, err)
+	require.Len(t, got.Progress, 2)
+	assert.Equal(t, domain.StepDone, got.Progress[0].State)
+}
+
+func TestUploadProgress_ByAnotherUserChangesNothing(t *testing.T) {
+	s := newStore(t)
+	id := startedUpload(t, s)
+	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, time.Now())
+
+	assert.ErrorIs(t, errSave, store.ErrUploadNotFound)
+	assert.ErrorIs(t, errBeat, store.ErrUploadNotFound)
+	got, err := s.GetUpload(ctx, userA, id)
+	require.NoError(t, err)
+	assert.Empty(t, got.Progress, "B's write must not have reached A's upload")
+	assert.True(t, got.AliveAt.IsZero())
+}
+
+func TestUploadProgress_ListDoesNotDragItAlong(t *testing.T) {
+	s := newStore(t)
+	id := startedUpload(t, s)
+	ctx := context.Background()
+	require.NoError(t, s.SaveProgress(ctx, userA, id, []domain.ProgressStep{{Key: "a", Label: "A", State: domain.StepRunning}}, time.Now()))
+
+	rows, err := s.ListUploads(ctx, userA)
+
+	require.NoError(t, err)
+	require.Len(t, rows, 1)
+	assert.Empty(t, rows[0].Progress, "the list view needs a status, not every row's phases")
+}
+
+func TestUploadProgress_SurvivesCompletion(t *testing.T) {
+	s := newStore(t)
+	id := startedUpload(t, s)
+	ctx := context.Background()
+	require.NoError(t, s.StartTranscribing(ctx, userA, id))
+	require.NoError(t, s.SaveProgress(ctx, userA, id, []domain.ProgressStep{{Key: "a", Label: "A", State: domain.StepDone}}, time.Now()))
+
+	require.NoError(t, s.CompleteUpload(ctx, userA, id, domain.Transcription{Text: "x", Endpoint: "e"}))
+
+	got, err := s.GetUpload(ctx, userA, id)
+	require.NoError(t, err)
+	require.Len(t, got.Progress, 1, "a finished upload keeps its record of what was done, and how long each step took")
+}

internal/adapters/store/uploaded_transcripts_test.go

@@ -0,0 +1,338 @@
+package store_test
+
+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"
+)
+
+// These exercise the Store API over uploaded_transcripts. The test Store connects
+// as the superuser, which BYPASSES row-level security — so the cross-user cases
+// here prove the query's own user_id predicate (defence in depth), while the
+// database-enforced guarantee is proven separately, as the non-superuser app
+// role, in uploaded_transcripts_isolation_test.go. Text is synthetic; this repo is
+// public and no real transcript may appear in it (#33).
+
+func seedUsers(t *testing.T, ids ...string) {
+	t.Helper()
+	super := rawPool(t)
+	resetDB(t, super)
+	for _, id := range ids {
+		_, err := super.Exec(context.Background(), `INSERT INTO users (id) VALUES ($1)`, id)
+		require.NoError(t, err)
+	}
+}
+
+func TestCreateUpload_StartsQueuedAndIsOwnedByTheCreator(t *testing.T) {
+	s := newStore(t)
+	seedUsers(t, userA)
+	ctx := context.Background()
+
+	id, err := s.CreateUpload(ctx, userA, "Teams export", domain.LangSwedish, false)
+	require.NoError(t, err)
+
+	got, err := s.GetUpload(ctx, userA, id)
+	require.NoError(t, err)
+	assert.Equal(t, id, got.ID)
+	assert.Equal(t, userA, got.UserID)
+	assert.Equal(t, "Teams export", got.Title)
+	assert.Equal(t, domain.UploadQueued, got.Status)
+	assert.Empty(t, got.Content)
+	assert.True(t, got.FinishedAt.IsZero(), "an unfinished upload has no finish time")
+}
+
+func TestUploadLifecycle_CompletedKeepsTextSegmentsEndpointAndDuration(t *testing.T) {
+	s := newStore(t)
+	seedUsers(t, userA)
+	ctx := context.Background()
+	id, err := s.CreateUpload(ctx, userA, "m", domain.LangSwedish, false)
+	require.NoError(t, err)
+
+	require.NoError(t, s.StartTranscribing(ctx, userA, id))
+	mid, err := s.GetUpload(ctx, userA, id)
+	require.NoError(t, err)
+	assert.Equal(t, domain.UploadTranscribing, mid.Status)
+
+	want := domain.Transcription{
+		Text: "God morgon allihopa. Vi går igenom läget.",
+		Segments: []domain.Segment{
+			{Start: 0.0, End: 1.5, Text: "God morgon allihopa."},
+			{Start: 1.5, End: 4.25, Text: "Vi går igenom läget."},
+		},
+		Endpoint:     "iguana/kb-whisper",
+		AudioSeconds: 4.25,
+	}
+	require.NoError(t, s.CompleteUpload(ctx, userA, id, want))
+
+	got, err := s.GetUpload(ctx, userA, id)
+	require.NoError(t, err)
+	assert.Equal(t, domain.UploadDone, got.Status)
+	assert.Equal(t, want.Text, got.Content)
+	assert.Equal(t, want.Segments, got.Segments, "segment times must survive the JSONB round trip")
+	assert.Equal(t, "iguana/kb-whisper", got.Endpoint, "the serving endpoint is a quality fact and is recorded per row")
+	assert.InDelta(t, 4.25, got.AudioSeconds, 1e-9)
+	assert.False(t, got.FinishedAt.IsZero())
+	assert.Empty(t, got.ErrorKind)
+}
+
+func TestUploadLifecycle_FailedRecordsTypedKindAndNoContent(t *testing.T) {
+	s := newStore(t)
+	seedUsers(t, userA)
+	ctx := context.Background()
+	id, err := s.CreateUpload(ctx, userA, "m", domain.LangSwedish, false)
+	require.NoError(t, err)
+	require.NoError(t, s.StartTranscribing(ctx, userA, id))
+
+	require.NoError(t, s.FailUpload(ctx, userA, id, domain.UploadErrTranscriptionFailed))
+
+	got, err := s.GetUpload(ctx, userA, id)
+	require.NoError(t, err)
+	assert.Equal(t, domain.UploadFailed, got.Status)
+	assert.Equal(t, domain.UploadErrTranscriptionFailed, got.ErrorKind)
+	assert.Empty(t, got.Content)
+	assert.False(t, got.FinishedAt.IsZero())
+}
+
+func TestGetUpload_AnotherUsersIDIsNotFound(t *testing.T) {
+	s := newStore(t)
+	seedUsers(t, userA, userB)
+	ctx := context.Background()
+	id, err := s.CreateUpload(ctx, userA, "A's meeting", domain.LangSwedish, false)
+	require.NoError(t, err)
+
+	_, err = s.GetUpload(ctx, userB, id)
+
+	assert.ErrorIs(t, err, store.ErrUploadNotFound,
+		"a non-owner must get not-found, not forbidden: existence is not disclosed")
+}
+
+func TestUploadWrites_ByAnotherUserChangeNothing(t *testing.T) {
+	s := newStore(t)
+	seedUsers(t, userA, userB)
+	ctx := context.Background()
+	id, err := s.CreateUpload(ctx, userA, "A's meeting", domain.LangSwedish, false)
+	require.NoError(t, err)
+
+	assert.ErrorIs(t, s.StartTranscribing(ctx, userB, id), store.ErrUploadNotFound)
+	assert.ErrorIs(t, s.CompleteUpload(ctx, userB, id, domain.Transcription{Text: "planted"}), store.ErrUploadNotFound)
+	assert.ErrorIs(t, s.FailUpload(ctx, userB, id, domain.UploadErrTranscriptionFailed), store.ErrUploadNotFound)
+
+	got, err := s.GetUpload(ctx, userA, id)
+	require.NoError(t, err)
+	assert.Equal(t, domain.UploadQueued, got.Status, "B's writes must leave A's upload exactly as it was")
+	assert.Empty(t, got.Content)
+}
+
+func TestListUploads_ReturnsOnlyTheCallersNewestFirstWithoutTranscriptText(t *testing.T) {
+	s := newStore(t)
+	seedUsers(t, userA, userB)
+	ctx := context.Background()
+	first, err := s.CreateUpload(ctx, userA, "first", domain.LangSwedish, false)
+	require.NoError(t, err)
+	second, err := s.CreateUpload(ctx, userA, "second", domain.LangSwedish, false)
+	require.NoError(t, err)
+	_, err = s.CreateUpload(ctx, userB, "B's", domain.LangSwedish, false)
+	require.NoError(t, err)
+	require.NoError(t, s.StartTranscribing(ctx, userA, first))
+	require.NoError(t, s.CompleteUpload(ctx, userA, first, domain.Transcription{Text: "long text body", Endpoint: "e"}))
+
+	got, err := s.ListUploads(ctx, userA)
+
+	require.NoError(t, err)
+	require.Len(t, got, 2, "A must see only A's uploads")
+	assert.Equal(t, second, got[0].ID, "newest first")
+	assert.Equal(t, first, got[1].ID)
+	assert.Empty(t, got[1].Content, "a list row must not drag the whole transcript along")
+	assert.Equal(t, domain.UploadDone, got[1].Status)
+}
+
+func TestFailInterruptedUploads_FailsOnlyUnfinishedRowsForEveryUser(t *testing.T) {
+	s := newStore(t)
+	seedUsers(t, userA, userB)
+	super := rawPool(t)
+	ctx := context.Background()
+	for subject, uid := range map[string]string{"sub-a": userA, "sub-b": userB} {
+		_, err := super.Exec(ctx, `INSERT INTO user_identities (dex_subject, user_id) VALUES ($1, $2)`, subject, uid)
+		require.NoError(t, err)
+	}
+	queued, err := s.CreateUpload(ctx, userA, "queued", domain.LangSwedish, false)
+	require.NoError(t, err)
+	running, err := s.CreateUpload(ctx, userB, "running", domain.LangSwedish, false)
+	require.NoError(t, err)
+	require.NoError(t, s.StartTranscribing(ctx, userB, running))
+	done, err := s.CreateUpload(ctx, userA, "done", domain.LangSwedish, false)
+	require.NoError(t, err)
+	require.NoError(t, s.StartTranscribing(ctx, userA, done))
+	require.NoError(t, s.CompleteUpload(ctx, userA, done, domain.Transcription{Text: "kept", Endpoint: "e"}))
+
+	n, err := s.FailInterruptedUploads(ctx)
+
+	require.NoError(t, err)
+	assert.Equal(t, 2, n, "the queued and the transcribing row, across both users")
+	for _, c := range []struct{ user, id string }{{userA, queued}, {userB, running}} {
+		got, err := s.GetUpload(ctx, c.user, c.id)
+		require.NoError(t, err)
+		assert.Equal(t, domain.UploadFailed, got.Status)
+		assert.Equal(t, domain.UploadErrInterrupted, got.ErrorKind,
+			"audio is not retained, so a restart cannot resume: the honest state is failed/interrupted")
+	}
+	kept, err := s.GetUpload(ctx, userA, done)
+	require.NoError(t, err)
+	assert.Equal(t, domain.UploadDone, kept.Status, "a finished transcript must survive the sweep")
+	assert.Equal(t, "kept", kept.Content)
+}
+
+func TestCreateUpload_RecordsWhetherTheFileHadAVideoTrack(t *testing.T) {
+	s := newStore(t)
+	seedUsers(t, userA)
+	ctx := context.Background()
+	withVideo, err := s.CreateUpload(ctx, userA, "camera off", domain.LangSwedish, true)
+	require.NoError(t, err)
+	audioOnly, err := s.CreateUpload(ctx, userA, "voice only", domain.LangSwedish, false)
+	require.NoError(t, err)
+
+	gotVideo, err := s.GetUpload(ctx, userA, withVideo)
+	require.NoError(t, err)
+	gotAudio, err := s.GetUpload(ctx, userA, audioOnly)
+	require.NoError(t, err)
+	listed, err := s.ListUploads(ctx, userA)
+	require.NoError(t, err)
+
+	assert.True(t, gotVideo.HadVideo, "the page must be able to say the video was ignored")
+	assert.False(t, gotAudio.HadVideo)
+	byID := map[string]bool{}
+	for _, u := range listed {
+		byID[u.ID] = u.HadVideo
+	}
+	assert.Equal(t, map[string]bool{withVideo: true, audioOnly: false}, byID, "the list carries it too")
+}
+
+func TestUploadHadVideo_SurvivesCompletionAndFailure(t *testing.T) {
+	s := newStore(t)
+	seedUsers(t, userA)
+	ctx := context.Background()
+	done, err := s.CreateUpload(ctx, userA, "d", domain.LangSwedish, true)
+	require.NoError(t, err)
+	failed, err := s.CreateUpload(ctx, userA, "f", domain.LangSwedish, true)
+	require.NoError(t, err)
+	require.NoError(t, s.StartTranscribing(ctx, userA, done))
+	require.NoError(t, s.CompleteUpload(ctx, userA, done, domain.Transcription{Text: "x", Endpoint: "e"}))
+	require.NoError(t, s.FailUpload(ctx, userA, failed, domain.UploadErrTranscriptionFailed))
+
+	for _, id := range []string{done, failed} {
+		got, err := s.GetUpload(ctx, userA, id)
+		require.NoError(t, err)
+		assert.True(t, got.HadVideo, "completing or failing must not clear it: the notice outlives the job")
+	}
+}
+
+func TestUploadLifecycle_ReconciliationRoundTripsWithItsContestedSpots(t *testing.T) {
+	s := newStore(t)
+	seedUsers(t, userA)
+	ctx := context.Background()
+	id, err := s.CreateUpload(ctx, userA, "m", domain.LangSwedish, false)
+	require.NoError(t, err)
+	require.NoError(t, s.StartTranscribing(ctx, userA, id))
+	want := &domain.Reconciliation{
+		Status: domain.ReconcileDone, Base: "iguana/pianissimo", Alt: "iguana/kb-whisper", Reconciler: "iguana/gemma4-26b",
+		Disputes: 380, AutoResolved: 90, Adjudicated: 285, Changed: 120, Unsure: 5, Errors: 0,
+		Contested: []domain.ContestedSpot{
+			{Seconds: 2210.5, Base: "Axel", Alt: "Alex", Used: "alt"},
+			{Seconds: 3400, Base: "A+2", Alt: "en A", Used: "base", Unsure: true},
+		},
+	}
+
+	require.NoError(t, s.CompleteUpload(ctx, userA, id, domain.Transcription{Text: "x", Endpoint: "e", Reconciliation: want}))
+
+	got, err := s.GetUpload(ctx, userA, id)
+	require.NoError(t, err)
+	assert.Equal(t, want, got.Reconciliation)
+}
+
+func TestUploadLifecycle_ASingleEngineTranscriptHasNoReconciliation(t *testing.T) {
+	s := newStore(t)
+	seedUsers(t, userA)
+	ctx := context.Background()
+	id, err := s.CreateUpload(ctx, userA, "m", domain.LangSwedish, false)
+	require.NoError(t, err)
+	require.NoError(t, s.StartTranscribing(ctx, userA, id))
+	require.NoError(t, s.CompleteUpload(ctx, userA, id, domain.Transcription{Text: "x", Endpoint: "e"}))
+
+	got, err := s.GetUpload(ctx, userA, id)
+
+	require.NoError(t, err)
+	assert.Nil(t, got.Reconciliation, "no reconciliation is not an empty one: the page must not claim a second opinion")
+}
+
+func TestUploadLifecycle_ASkippedReconciliationKeepsItsReason(t *testing.T) {
+	s := newStore(t)
+	seedUsers(t, userA)
+	ctx := context.Background()
+	id, err := s.CreateUpload(ctx, userA, "m", domain.LangSwedish, false)
+	require.NoError(t, err)
+	require.NoError(t, s.StartTranscribing(ctx, userA, id))
+	rec := &domain.Reconciliation{Status: domain.ReconcileSkipped, Reason: "the model failed 3 times in a row: gateway down", Errors: 3}
+
+	require.NoError(t, s.CompleteUpload(ctx, userA, id, domain.Transcription{Text: "x", Endpoint: "e", Reconciliation: rec}))
+
+	got, err := s.GetUpload(ctx, userA, id)
+	require.NoError(t, err)
+	assert.Equal(t, rec, got.Reconciliation)
+}
+
+func TestCreateUpload_RecordsTheLanguageTheUserChoseAndTheListCarriesIt(t *testing.T) {
+	s := newStore(t)
+	seedUsers(t, userA)
+	ctx := context.Background()
+	swedish, err := s.CreateUpload(ctx, userA, "sv", domain.LangSwedish, false)
+	require.NoError(t, err)
+	auto, err := s.CreateUpload(ctx, userA, "auto", domain.LangAuto, false)
+	require.NoError(t, err)
+
+	gotSv, err := s.GetUpload(ctx, userA, swedish)
+	require.NoError(t, err)
+	gotAuto, err := s.GetUpload(ctx, userA, auto)
+	require.NoError(t, err)
+	listed, err := s.ListUploads(ctx, userA)
+	require.NoError(t, err)
+
+	assert.Equal(t, domain.LangSwedish, gotSv.Language)
+	assert.Equal(t, domain.LangAuto, gotAuto.Language)
+	byID := map[string]domain.Language{}
+	for _, u := range listed {
+		byID[u.ID] = u.Language
+	}
+	assert.Equal(t, map[string]domain.Language{swedish: domain.LangSwedish, auto: domain.LangAuto}, byID)
+}
+
+// The closed set is also held by the table, so a bug or a second writer cannot store a
+// value the engines are never meant to receive (tapir#61).
+func TestUploadedTranscripts_TheTableRefusesALanguageOutsideTheClosedSet(t *testing.T) {
+	newStore(t)
+	seedUsers(t, userA)
+
+	_, err := rawPool(t).Exec(context.Background(),
+		`INSERT INTO uploaded_transcripts (user_id, title, language) VALUES ($1, 'x', 'en')`, userA)
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "language", "the CHECK constraint names the column")
+}
+
+// A row from before the column existed was transcribed as Swedish; the default says so.
+func TestUploadedTranscripts_ARowInsertedWithoutALanguageIsSwedish(t *testing.T) {
+	s := newStore(t)
+	seedUsers(t, userA)
+	var id string
+	require.NoError(t, rawPool(t).QueryRow(context.Background(),
+		`INSERT INTO uploaded_transcripts (user_id, title) VALUES ($1, 'old') RETURNING id`, userA).Scan(&id))
+
+	got, err := s.GetUpload(context.Background(), userA, id)
+
+	require.NoError(t, err)
+	assert.Equal(t, domain.LangSwedish, got.Language)
+}

internal/adapters/store/videos.go

@@ -7,7 +7,7 @@ import (
 
 	"github.com/jackc/pgx/v5"
 
-	"gitea.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/domain"
 )
 
 // UpsertVideo persists a video's metadata and returns its durable store id (the

internal/adapters/store/videos_test.go

@@ -7,7 +7,7 @@ import (
 
 	"github.com/stretchr/testify/require"
 
-	"gitea.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/domain"
 )
 
 func ytVideo(userID, provVideoID, title string) domain.Video {

internal/adapters/summarizer/summarizer.go

@@ -13,11 +13,13 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"log/slog"
 	"strings"
 	"time"
 	"unicode/utf8"
 
-	"gitea.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/metrics"
 )
 
 // Completer is the minimal LLM chat surface the Summarizer needs.
@@ -95,16 +97,23 @@ func (s *Summarizer) Summarize(ctx context.Context, v domain.Video, t domain.Tra
 
 	var errs []error
 	for i, ep := range s.endpoints {
+		fallback := i > 0
+		start := time.Now()
 		out, err := ep.Client.Complete(ctx, systemPrompt, user)
+		dur := time.Since(start)
 		if err != nil {
+			metrics.ObserveSummarize(ep.Model, "error", fallback, dur)
 			errs = append(errs, fmt.Errorf("%s/%s call: %w", ep.Provider, ep.Model, err))
 			continue
 		}
-		sum, perr := s.build(v, ep, i > 0, out)
+		sum, perr := s.build(v, ep, fallback, out)
 		if perr != nil {
+			metrics.ObserveSummarize(ep.Model, "parse_error", fallback, dur)
 			errs = append(errs, fmt.Errorf("%s/%s output: %w", ep.Provider, ep.Model, perr))
 			continue
 		}
+		metrics.ObserveSummarize(ep.Model, "success", fallback, dur)
+		slog.Default().Info("summarized", "model", ep.Model, "fallback", fallback, "elapsed_ms", dur.Milliseconds())
 		return sum, nil
 	}
 	return domain.Summary{}, fmt.Errorf("summarize: all %d endpoint(s) failed: %w", len(s.endpoints), errors.Join(errs...))

internal/adapters/summarizer/summarizer_test.go

@@ -7,13 +7,36 @@ package summarizer
 import (
 	"context"
 	"errors"
+	"net/http"
+	"net/http/httptest"
 	"strings"
 	"testing"
 
-	"gitea.d-ma.be/mathias/tapir/internal/domain"
-	"gitea.d-ma.be/mathias/tapir/internal/ports"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/metrics"
+	"git.d-ma.be/mathias/tapir/internal/ports"
 )
 
+// TestSummarizerRecordsMetric verifies the summarizer→metrics wiring (ADR-030)
+// black-box: after a successful summarize, the public /metrics scrape shows a
+// success observation for that endpoint's model.
+func TestSummarizerRecordsMetric(t *testing.T) {
+	const model = "metrics-test-model"
+	s := New(Endpoint{Client: &fakeClient{reply: goodReply}, Provider: "local", Model: model}, nil)
+	if _, err := s.Summarize(context.Background(), testVideo(), testTranscript()); err != nil {
+		t.Fatalf("Summarize: %v", err)
+	}
+
+	rec := httptest.NewRecorder()
+	metrics.Handler().ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/metrics", nil))
+	body := rec.Body.String()
+	if !strings.Contains(body, `tapir_summarize_duration_seconds`) ||
+		!strings.Contains(body, `model="`+model+`"`) ||
+		!strings.Contains(body, `outcome="success"`) {
+		t.Errorf("metrics scrape missing summarize success for %s", model)
+	}
+}
+
 // compile-time check: Summarizer satisfies the port.
 var _ ports.Summarizer = (*Summarizer)(nil)
 

internal/adapters/transcribe/chain.go

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

internal/adapters/transcribe/chain_test.go

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

internal/adapters/transcribe/ensemble.go

@@ -0,0 +1,149 @@
+package transcribe
+
+import (
+	"context"
+	"errors"
+	"fmt"
+
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/errkind"
+	"git.d-ma.be/mathias/tapir/internal/progress"
+	"git.d-ma.be/mathias/tapir/internal/reconcile"
+)
+
+// Ensemble transcribes the same WAV with two engines, one after the other, and has a
+// local model reconcile their disagreements (ADR-034). It satisfies uploads.Transcriber,
+// so the worker neither knows nor cares that two engines were involved.
+//
+// It degrades in steps and never turns a recoverable problem into a failed job:
+// one engine down → the other's text, said so on the result; the model down → the
+// base text unreconciled; both engines down → the cloud fallback, if configured, and
+// recorded as such. A cancelled job stops everything and never reaches the cloud.
+type Ensemble struct {
+	base, alt Backend
+	fallback  Backend // nil = no cloud fallback
+	adj       reconcile.Adjudicator
+	opts      reconcile.Options
+}
+
+// NewEnsemble builds an Ensemble. base is the engine whose text stands where the two
+// disagree and is the faster one, so it runs first; alt supplies the other reading.
+func NewEnsemble(base, alt, fallback Backend, adj reconcile.Adjudicator, opts reconcile.Options) *Ensemble {
+	return &Ensemble{base: base, alt: alt, fallback: fallback, adj: adj, opts: opts}
+}
+
+// Plan implements uploads.Planner: the phases of a Swedish job, the two-engine one. The
+// fallback is a phase only when one is configured.
+func (e *Ensemble) Plan() []progress.Planned { return e.PlanFor(domain.LangSwedish) }
+
+// PlanFor implements uploads.LanguagePlanner. A recording that is not Swedish never runs the
+// base engine, the comparison or the decisions (transcribeOther), so the page lists only
+// what can run: the other engine, then the fallback if one is configured.
+func (e *Ensemble) PlanFor(lang domain.Language) []progress.Planned {
+	var plan []progress.Planned
+	if lang == domain.LangSwedish || lang == "" {
+		plan = []progress.Planned{
+			{Key: "engine_a", Label: "Transcribing with " + e.base.Name()},
+			{Key: "engine_b", Label: "Transcribing with " + e.alt.Name()},
+			{Key: "compare", Label: "Comparing the two transcripts"},
+			{Key: "decide", Label: "Deciding the differences"},
+		}
+	} else {
+		plan = []progress.Planned{{Key: "engine_b", Label: "Transcribing with " + e.alt.Name()}}
+	}
+	if e.fallback != nil {
+		plan = append(plan, progress.Planned{Key: "fallback", Label: "Cloud fallback (" + e.fallback.Name() + ")"})
+	}
+	return plan
+}
+
+// Transcribe implements uploads.Transcriber. Each phase is reported through the
+// progress.Reporter in ctx (a no-op when nobody listens).
+func (e *Ensemble) Transcribe(ctx context.Context, wavPath string) (domain.Transcription, error) {
+	rep := progress.From(ctx)
+	if domain.LanguageFrom(ctx) != domain.LangSwedish {
+		return e.transcribeOther(ctx, rep, wavPath)
+	}
+	b, berr := e.run(ctx, rep, "engine_a", e.base, wavPath)
+	if err := ctx.Err(); err != nil {
+		return domain.Transcription{}, err
+	}
+	a, aerr := e.run(ctx, rep, "engine_b", e.alt, wavPath)
+	if err := ctx.Err(); err != nil {
+		return domain.Transcription{}, err
+	}
+
+	switch {
+	case berr == nil && aerr == nil:
+		out, err := reconcile.Reconcile(ctx, b, a, e.adj, e.opts)
+		if err != nil {
+			return domain.Transcription{}, err
+		}
+		out.Endpoint = fmt.Sprintf("%s + %s, reconciled by %s", e.base.Name(), e.alt.Name(), e.opts.Reconciler)
+		if rc := out.Reconciliation; rc != nil {
+			rep.Note("reconciled", "status", string(rc.Status), "disputes", rc.Disputes, "auto_resolved", rc.AutoResolved,
+				"adjudicated", rc.Adjudicated, "changed", rc.Changed, "unsure", rc.Unsure, "errors", rc.Errors)
+		}
+		return out, nil
+	case berr == nil:
+		return e.single(b, e.base.Name(), fmt.Sprintf("%s failed (%s)", e.alt.Name(), errkind.Of(aerr))), nil
+	case aerr == nil:
+		return e.single(a, e.alt.Name(), fmt.Sprintf("%s failed (%s)", e.base.Name(), errkind.Of(berr))), nil
+	}
+
+	both := errors.Join(fmt.Errorf("%s: %w", e.base.Name(), berr), fmt.Errorf("%s: %w", e.alt.Name(), aerr))
+	if e.fallback == nil {
+		return domain.Transcription{}, both
+	}
+	f, ferr := e.run(ctx, rep, "fallback", e.fallback, wavPath)
+	if ferr != nil {
+		return domain.Transcription{}, errors.Join(both, fmt.Errorf("%s: %w", e.fallback.Name(), ferr))
+	}
+	return e.single(f, e.fallback.Name(), "both local engines failed"), nil
+}
+
+// transcribeOther handles a recording that is not Swedish (tapir#61). The base engine is
+// Swedish-only, so it never receives the audio, and with one reading there is nothing to
+// reconcile: the other engine transcribes alone, then the cloud fallback if it fails and the
+// job is still live. The phases that did not run are shown as skipped by the tracker.
+func (e *Ensemble) transcribeOther(ctx context.Context, rep progress.Reporter, wavPath string) (domain.Transcription, error) {
+	t, err := e.run(ctx, rep, "engine_b", e.alt, wavPath)
+	if err == nil {
+		t.Endpoint = e.alt.Name()
+		return t, nil
+	}
+	if e.fallback == nil || ctx.Err() != nil {
+		return domain.Transcription{}, fmt.Errorf("%s: %w", e.alt.Name(), err)
+	}
+	f, ferr := e.run(ctx, rep, "fallback", e.fallback, wavPath)
+	if ferr != nil {
+		return domain.Transcription{}, errors.Join(
+			fmt.Errorf("%s: %w", e.alt.Name(), err), fmt.Errorf("%s: %w", e.fallback.Name(), ferr))
+	}
+	f.Endpoint = e.fallback.Name()
+	return f, nil
+}
+
+// run is one engine call, reported as a phase. A job that was stopped is not a broken
+// engine, so a cancelled context is never reported as a failed phase.
+func (e *Ensemble) run(ctx context.Context, rep progress.Reporter, key string, be Backend, wav string) (domain.Transcription, error) {
+	rep.Begin(key, 0)
+	t, err := be.Transcribe(ctx, wav)
+	switch {
+	case err == nil:
+		rep.Done(key)
+	case ctx.Err() == nil:
+		rep.Fail(key, err)
+	}
+	return t, err
+}
+
+// single is a transcript from one engine only: nothing to reconcile, and the result
+// says why, because a reader should know the second opinion was missing.
+func (e *Ensemble) single(t domain.Transcription, engine, reason string) domain.Transcription {
+	t.Endpoint = engine
+	t.Reconciliation = &domain.Reconciliation{
+		Status: domain.ReconcileSingle, Reason: reason, Base: e.opts.BaseEngine, Alt: e.opts.AltEngine, Reconciler: e.opts.Reconciler,
+	}
+	return t
+}

internal/adapters/transcribe/ensemble_progress_test.go

@@ -0,0 +1,184 @@
+package transcribe_test
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"sync"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.d-ma.be/mathias/tapir/internal/adapters/transcribe"
+	"git.d-ma.be/mathias/tapir/internal/domain"
+	"git.d-ma.be/mathias/tapir/internal/progress"
+	"git.d-ma.be/mathias/tapir/internal/reconcile"
+)
+
+// What the Ensemble reports to the page and the log (ADR-035). The reporter here is a
+// recording double carried in the context, exactly as the worker's Tracker is.
+
+type recReporter struct {
+	mu     sync.Mutex
+	events []string
+	notes  []map[string]any
+}
+
+func (r *recReporter) add(e string) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.events = append(r.events, e)
+}
+
+func (r *recReporter) Begin(key string, _ int)   { r.add("begin:" + key) }
+func (r *recReporter) Advance(key string, _ int) { r.add("advance:" + key) }
+func (r *recReporter) Done(key string)           { r.add("done:" + key) }
+func (r *recReporter) Fail(key string, _ error)  { r.add("fail:" + key) }
+func (r *recReporter) Note(msg string, kv ...any) {
+	m := map[string]any{"msg": msg}
+	for i := 0; i+1 < len(kv); i += 2 {
+		m[kv[i].(string)] = kv[i+1]
+	}
+	r.mu.Lock()
+	r.notes = append(r.notes, m)
+	r.mu.Unlock()
+	r.add("note:" + msg)
+}
+
+func (r *recReporter) joined() string {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	return strings.Join(r.events, ",")
+}
+
+func TestEnsemblePlan_DeclaresBothEnginesAndTheReconcileStepsByName(t *testing.T) {
+	r := newEnsemble("a", "a", nil, nil, false, &verdictAdj{})
+
+	plan := r.e.Plan()
+
+	require.Len(t, plan, 4)
+	assert.Equal(t, []string{"engine_a", "engine_b", "compare", "decide"}, []string{plan[0].Key, plan[1].Key, plan[2].Key, plan[3].Key})
+	assert.Contains(t, plan[0].Label, "iguana/pianissimo", "the page says which engine is running")
+	assert.Contains(t, plan[1].Label, "iguana/kb-whisper")
+}
+
+func TestEnsemblePlan_TheCloudFallbackIsAStepOnlyWhenOneIsConfigured(t *testing.T) {
+	with := newEnsemble("a", "a", nil, nil, true, &verdictAdj{}).e.Plan()
+	without := newEnsemble("a", "a", nil, nil, false, &verdictAdj{}).e.Plan()
+
+	assert.Equal(t, "fallback", with[len(with)-1].Key)
+	assert.Len(t, without, 4, "no fallback configured: the page must not promise one")
+}
+
+func TestEnsembleProgress_BothEnginesAreReportedInTheOrderTheyRun(t *testing.T) {
+	rep := &recReporter{}
+	r := newEnsemble("Totalt 1015 kronor idag.", "Totalt 1050 kronor idag.", nil, nil, false, &verdictAdj{v: reconcile.VerdictAlt})
+
+	_, err := r.e.Transcribe(progress.With(context.Background(), rep), "/a.wav")
+
+	require.NoError(t, err)
+	got := rep.joined()
+	a, aDone := strings.Index(got, "begin:engine_a"), strings.Index(got, "done:engine_a")
+	b, bDone := strings.Index(got, "begin:engine_b"), strings.Index(got, "done:engine_b")
+	require.True(t, a >= 0 && aDone > a && b > aDone && bDone > b, "events out of order: %s", got)
+	assert.NotContains(t, got, "fail:")
+}
+
+func TestEnsembleProgress_AnEngineThatFailsIsShownFailedAndTheOtherStillRuns(t *testing.T) {
+	rep := &recReporter{}
+	r := newEnsemble("text", "text", errors.New("pianissimo down"), nil, false, &verdictAdj{})
+
+	got, err := r.e.Transcribe(progress.With(context.Background(), rep), "/a.wav")
+
+	require.NoError(t, err, "one engine down is a degraded result, not a failed job")
+	assert.Equal(t, domain.ReconcileSingle, got.Reconciliation.Status)
+	ev := rep.joined()
+	assert.Contains(t, ev, "fail:engine_a")
+	assert.Contains(t, ev, "done:engine_b")
+	assert.NotContains(t, ev, "begin:compare", "nothing to compare with one engine")
+}
+
+func TestEnsembleProgress_TheFallbackIsReportedOnlyWhenBothEnginesFailed(t *testing.T) {
+	rep := &recReporter{}
+	r := newEnsemble("t", "t", errors.New("a down"), errors.New("b down"), true, &verdictAdj{})
+
+	_, err := r.e.Transcribe(progress.With(context.Background(), rep), "/a.wav")
+
+	require.NoError(t, err)
+	ev := rep.joined()
+	assert.Contains(t, ev, "fail:engine_a")
+	assert.Contains(t, ev, "fail:engine_b")
+	assert.Contains(t, ev, "begin:fallback")
+	assert.Contains(t, ev, "done:fallback")
+
+	rep2 := &recReporter{}
+	ok := newEnsemble("x", "x", nil, nil, true, &verdictAdj{})
+	_, err = ok.e.Transcribe(progress.With(context.Background(), rep2), "/a.wav")
+	require.NoError(t, err)
+	assert.NotContains(t, rep2.joined(), "fallback", "a healthy run never touches the cloud step")
+}
+
+// cancellingBackend ends the job in the middle of its own call, like a shutdown.
+type cancellingBackend struct {
+	name   string
+	cancel context.CancelFunc
+}
+
+func (c cancellingBackend) Name() string { return c.name }
+func (c cancellingBackend) Transcribe(ctx context.Context, _ string) (domain.Transcription, error) {
+	c.cancel()
+	return domain.Transcription{}, ctx.Err()
+}
+
+func TestEnsembleProgress_AShutdownMidEngineIsNotReportedAsAFailedEngine(t *testing.T) {
+	rep := &recReporter{}
+	ctx, cancel := context.WithCancel(progress.With(context.Background(), rep))
+	calls := &[]string{}
+	alt := &orderedBackend{name: "iguana/kb-whisper", text: "x", calls: calls}
+	e := transcribe.NewEnsemble(cancellingBackend{name: "iguana/pianissimo", cancel: cancel}, alt, nil, &verdictAdj{}, ensembleOpts)
+
+	_, err := e.Transcribe(ctx, "/a.wav")
+
+	require.ErrorIs(t, err, context.Canceled)
+	assert.NotContains(t, rep.joined(), "fail:", "a stopped job is not a broken engine: the page must not blame one")
+	assert.Empty(t, *calls, "and the second engine never starts")
+}
+
+func TestEnsembleProgress_TheReconcileCountsAreNotedForTheOperator(t *testing.T) {
+	rep := &recReporter{}
+	r := newEnsemble("Totalt 1015 kronor idag.", "Totalt 1050 kronor idag.", nil, nil, false, &verdictAdj{v: reconcile.VerdictAlt})
+
+	_, err := r.e.Transcribe(progress.With(context.Background(), rep), "/a.wav")
+
+	require.NoError(t, err)
+	require.Len(t, rep.notes, 1)
+	n := rep.notes[0]
+	assert.Equal(t, "reconciled", n["msg"])
+	assert.Equal(t, 1, n["disputes"])
+	assert.Equal(t, 1, n["changed"])
+	assert.Equal(t, 0, n["unsure"])
+	assert.Equal(t, 0, n["errors"])
+}
+
+// An auto-language upload never runs the base engine, the comparison or the decisions, so the
+// page must not list them as coming (tapir#61, PR #62 review).
+func TestEnsemblePlanFor_AnAutoLanguageListsOnlyThePhasesThatCanRun(t *testing.T) {
+	with := newEnsemble("a", "a", nil, nil, true, &verdictAdj{}).e
+	without := newEnsemble("a", "a", nil, nil, false, &verdictAdj{}).e
+	keys := func(p []progress.Planned) []string {
+		var k []string
+		for _, s := range p {
+			k = append(k, s.Key)
+		}
+		return k
+	}
+
+	assert.Equal(t, []string{"engine_b", "fallback"}, keys(with.PlanFor(domain.LangAuto)))
+	assert.Equal(t, []string{"engine_b"}, keys(without.PlanFor(domain.LangAuto)), "no fallback configured: none promised")
+	assert.Equal(t, with.Plan(), with.PlanFor(domain.LangSwedish), "Swedish is the plan it always was")
+	assert.Contains(t, with.PlanFor(domain.LangAuto)[0].Label, "iguana/kb-whisper")
+	assert.Equal(t, with.Plan(), with.PlanFor(""), "a job queued before the choice existed is Swedish")
+	assert.Equal(t, with.PlanFor(domain.LangAuto), with.PlanFor(domain.Language("xx")),
+		"an unknown language is planned as it is run: not the Swedish phases")
+}

internal/adapters/transcribe/ensemble_reason_test.go

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