Merge pull request 'FabledCurator can now tell you one of its own parts has stopped' (#250) from dev into main

The ask was a surface AND a path. The path is the part that was missing —
everything that could answer "is it running" lived inside Settings, which you
only open once you already suspect something.

**Re-used the indicator that already existed rather than adding a fourth.**
There were three partial surfaces: TopNav's health dot, PipelineStatusChip's
pulse, and the Settings Activity tab. None answered "is every part alive", and
a fourth would have made the question harder to answer, not easier.

TopNav's dot read /api/health — a no-DB liveness check proving only that the
WEB container is serving. Green there while a worker was dead is exactly what
it looked like, and a green dot beside the product name gets read as
"everything is fine". It now reflects the whole-stack verdict, and it is a
link: the place someone already looks when they suspect something is now also
the way to the detail.

The tooltip names the actual problem. "Scheduler has not checked in for 6 min"
sends someone somewhere; "something is unhealthy" sends them hunting.

/system is deliberately NOT in the nav row — TopNav builds that from routes
with a meta.title, and a sixth top-level tab for a page visited twice a year
costs more attention than it returns. It is reached from the dot.

The page lists every learned part with its state as a sentence rather than a
chip, and prints the staleness thresholds it was judged by, taken from the
endpoint so the UI keeps no second copy of them. PipelineStatusChip still
hand-rolls its own 3-minute scheduler window; that is now a duplicate of a
threshold the server owns, and worth collapsing once this has been watched
working.

The stores stay separate on purpose: system.js is "can I reach the API",
systemActivity.js is "what is the pipeline doing", systemHealth.js is "is
anything broken". Running and alive fail independently — an idle stack with a
dead worker looks identical to a healthy one on every activity surface, which
is the whole reason this milestone exists.

Not yet verified against a real stopped service. Rule 12 keeps a local stack
out of it, and frontend CI has no Vue type-check or visual regression, so this
needs an operator look rather than a green lane.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TTjbZZ6JirCMSaJzQV1RhA

.env.example

@@ -1,24 +1,103 @@
-# Database
+# FabledCurator configuration.
-DB_USER=fabledcurator
+#
-DB_PASSWORD=changeme_use_a_real_password
+# Copy to `.env` and edit before your first production start:
-DB_HOST=postgres
+#
-DB_PORT=5432
+#     cp .env.example .env
-DB_NAME=fabledcurator
+#
+# Only the two values under CHANGE THESE actually need your attention. The
+# rest have working defaults baked into docker-compose.yml and are listed
+# here so you know they exist, not because you have to set them.
+#
+# Almost nothing else lives here on purpose. FabledCurator is configured from
+# its own Settings UI, backed by the database — no restart, no YAML. If you
+# are looking for where to set an import path, a download schedule or an ML
+# threshold, it is in the app, not in this file.
 
-# Redis / Celery
-CELERY_BROKER_URL=redis://redis:6379/0
-CELERY_RESULT_BACKEND=redis://redis:6379/0
 
-# App
+# ---------------------------------------------------------------------------
-# Generate with: openssl rand -hex 32
+# CHANGE THESE
-SECRET_KEY=changeme_32_byte_hex_secret
+# ---------------------------------------------------------------------------
 
-# Extension API key — used in FC-3, lands later but reserved now
+# The Postgres password. docker-compose.yml falls back to a published default
-# Generate with: openssl rand -hex 32
+# (`fabledcurator_dev`) so that `docker compose up` works with no config at
-EXTENSION_API_KEY=
+# all — which is exactly why you must not leave it at that on a real install.
+# It is the credential protecting your stored platform session cookies.
+DB_PASSWORD=
 
-# Logging
+# Sets Quart's app.secret_key. Today it signs nothing: FabledCurator has no
+# login and uses no session cookies, so no value here is protecting anything
+# right now. Set it anyway. It is required at boot rather than defaulted so
+# that the day something session-backed does land, no instance is already
+# running on a value published in this file.
+#
+#     openssl rand -hex 32
+SECRET_KEY=
+
+
+# ---------------------------------------------------------------------------
+# FIRST BOOT ONLY — then delete this line
+# ---------------------------------------------------------------------------
+
+# FabledCurator encrypts your stored platform credentials with a Fernet key it
+# keeps at /images/secrets/credential_key.b64 — inside the ./images bind mount,
+# so it outlives the container. On a brand-new install that file does not exist
+# yet, and the app REFUSES TO START rather than quietly create one:
+#
+#     MissingCredentialKey: Fernet key file not found at
+#     /images/secrets/credential_key.b64
+#
+# That refusal is deliberate. Auto-creating a key is indistinguishable from the
+# disaster case — a restore that brought the database back but lost
+# ./images/secrets — and there it would mint a key that cannot decrypt anything,
+# leaving an instance that looks healthy while every paywalled download fails.
+# So the choice is yours to make explicitly, once.
+#
+# Set this for your first `up`, watch the container come up, then DELETE THE
+# LINE. Leaving it set disarms the protection permanently, on an instance that
+# by then has credentials worth protecting.
+#
+# BACK UP ./images/secrets/ ALONGSIDE YOUR DATABASE. The key is the only thing
+# that can read your stored credentials; a database restored without it needs
+# every credential re-entered by hand.
+CURATOR_BOOTSTRAP_NEW_KEY=1
+
+
+# ---------------------------------------------------------------------------
+# Optional — defaults are fine
+# ---------------------------------------------------------------------------
+
+# Host port the UI is published on. The container always listens on 8080;
+# this is only the left-hand side of the port mapping.
+PORT=8080
+
+# DEBUG | INFO | WARNING | ERROR
 LOG_LEVEL=INFO
 
-# Deployment posture: plain HTTP (no TLS in the app; reverse proxy if needed)
+# Postgres identity. Change these only if you are pointing at a database you
-# See docs/superpowers/specs/2026-05-13-fabledcurator-merge-design.md §2.1
+# manage yourself — the bundled postgres service is created with whatever is
+# set here, so changing them after the first start will not rename anything.
+DB_USER=fabledcurator
+DB_NAME=fabledcurator
+
+# Set by docker-compose.yml to reach the bundled services. Override only when
+# running Postgres or Redis outside this stack.
+# DB_HOST=postgres
+# DB_PORT=5432
+# CELERY_BROKER_URL=redis://redis:6379/0
+# CELERY_RESULT_BACKEND=redis://redis:6379/0
+
+
+# ---------------------------------------------------------------------------
+# There is no authentication variable here, and that is not an omission
+# ---------------------------------------------------------------------------
+#
+# FabledCurator has no login, no accounts and no permission model. Anything
+# that can reach PORT is an administrator and can read the platform session
+# cookies the app stores for Patreon, SubscribeStar and Pixiv.
+#
+# Bind it to a trusted network. See "Before you expose it" in README.md and
+# the deployment posture section of SECURITY.md.
+#
+# The Firefox extension's API key is NOT configured here — it is generated
+# automatically on first use and shown under Settings → Maintenance, where you
+# can also rotate it.

.forgejo/workflows/baseline.yml

@@ -87,10 +87,21 @@ jobs:
           test -n "$PG_IP"
           echo "PG_CONTAINER=$PG" >> "$GITHUB_ENV"
           echo "DB_HOST=$PG_IP" >> "$GITHUB_ENV"
+          # Socket probe in python, not bash's /dev/tcp — these steps run under
+          # `sh -e`, where that path does not exist. Same fix and same reasoning
+          # as ci.yml's integration job; see the comment there.
+          pg_ready=""
           for i in $(seq 1 60); do
-            (echo > "/dev/tcp/$PG_IP/5432") >/dev/null 2>&1 && break
+            if python -c "import socket,sys; s=socket.socket(); s.settimeout(2); sys.exit(0 if s.connect_ex(('$PG_IP', 5432)) == 0 else 1)"; then
+              pg_ready=1
+              break
+            fi
             sleep 2
           done
+          if [ -z "$pg_ready" ]; then
+            echo "postgres at $PG_IP:5432 did not accept a connection within 120s"
+            exit 1
+          fi
           if command -v uv >/dev/null 2>&1; then
             uv pip install --system -r requirements.txt
           else

.forgejo/workflows/build.yml

@@ -44,6 +44,10 @@ on:
         description: 'Rebuild every image even if the published revision matches'
         type: boolean
         default: false
+      refresh:
+        description: 'Behave as the weekly base refresh: build main against fresh bases, publish through the candidate tag'
+        type: boolean
+        default: false
 
   # The base-image refresh (milestone 326 step 4, #3154).
   #
@@ -72,8 +76,43 @@ on:
 # Deriving it per job invites the two halves to disagree: sign-extension would
 # derive dev's extension version while build-web bundled main's, and the
 # release download would 404 on a version that exists perfectly well.
+# IS THIS A BASE REFRESH? Asked in five places and previously spelled five
+# ways — `github.event_name == 'schedule'` in an `if:`, `$GITHUB_EVENT_NAME` in
+# one shell, an `EVENT:` env passed into another, and a bare expression on
+# `pull:`. Five spellings of one fact is how half of them come to disagree
+# after somebody adds a sixth trigger.
+#
+# The `refresh` dispatch input is here so this path can be EXERCISED. A weekly
+# cron is otherwise testable once a week, which is not a cadence anything can
+# be developed against — the same reason `force_build` exists (#3252, added to
+# confirm #3190 was gone rather than wait for it to recur). It is also what
+# makes the milestone-362 gate verifiable at all: a gate has to be watched
+# rejecting something before anyone can believe it is wired up.
+#
+# The input is normalised through `format()` before it is compared, and that
+# is not defensive styling — the direct comparison is WRONG and fails silently.
+#
+# `type: boolean` delivers a real boolean, and GitHub expression semantics cast
+# operands to numbers when their types differ: `true == 'true'` compares 1
+# against NaN and is FALSE. Measured on run 5270, whose own log says it —
+#
+#   expression '(github.event_name == 'schedule'
+#                || github.event.inputs.refresh == 'true') && 'true' || 'false''
+#     evaluated to '%!t(string=false)'
+#   trigger: raw inputs refresh='true'
+#
+# — the input arrived as `true` and the expression still said false. The run
+# then went green with every step skipped, because a refresh that evaluates
+# false behaves exactly like an ordinary push. That is the whole hazard: the
+# failure has no symptom.
+#
+# `force_build` never hit this because it never compares in an expression. It
+# passes the raw value into an env var and tests it in the shell, where
+# everything is already a string. `format('{0}', x)` buys the same thing here,
+# where a step-level `if:` needs the answer before any shell runs.
 env:
-  BUILD_REF: ${{ github.event_name == 'schedule' && 'main' || github.ref }}
+  IS_REFRESH: ${{ (github.event_name == 'schedule' || format('{0}', github.event.inputs.refresh) == 'true') && 'true' || 'false' }}
+  BUILD_REF: ${{ (github.event_name == 'schedule' || format('{0}', github.event.inputs.refresh) == 'true') && 'main' || github.ref }}
 
 # Requires repo secret RELEASE_TOKEN — a Forgejo PAT with scopes:
 #   - write:package, read:package  (for docker push to git.fabledsword.com)
@@ -143,7 +182,7 @@ jobs:
       # evaluate — this file already gates steps on it — so the guard cannot
       # be disabled by the same uncertainty it exists to cover.
       - name: Guard — a scheduled run must have checked out main
-        if: github.event_name == 'schedule'
+        if: env.IS_REFRESH == 'true'
         run: |
           set -eu
           BRANCH=$(git rev-parse --abbrev-ref HEAD)
@@ -413,6 +452,18 @@ jobs:
        # to. Same source of truth; no double-store.
 
   build-web:
+    # Consumed by smoke-web's job-level `if:`. It cannot read `env` — the env
+    # context is available to STEP `if:` and step bodies, never to a job's own
+    # condition, and an unresolvable context there is empty rather than an
+    # error. `smoke-web` skipped silently on run 5290 for exactly that reason.
+    #
+    # Keying off the reuse step's own output is better than re-deriving the
+    # trigger anyway: it is the same single decision the build, the XPI
+    # download and the promote all take (build.yml's "one decision drives
+    # everything downstream"), and it says the thing smoke-web actually needs
+    # to know — a candidate was published — rather than restating why.
+    outputs:
+      candidate: ${{ steps.reuse.outputs.promote }}
     # A plain `needs` — no `always()`. That expression existed to let a
     # SKIPPED sign-extension through on a tag push while still blocking a
     # FAILED one. With no tag trigger, sign-extension always runs, so the
@@ -437,7 +488,7 @@ jobs:
 
       # See sign-extension's copy for why this guard exists.
       - name: Guard — a scheduled run must have checked out main
-        if: github.event_name == 'schedule'
+        if: env.IS_REFRESH == 'true'
         run: |
           set -eu
           BRANCH=$(git rev-parse --abbrev-ref HEAD)
@@ -470,8 +521,18 @@ jobs:
       #     the PREVIOUS XPI while the freshly signed one is orphaned (#3156).
       #   * dev and main derive the same values for the same source.
       - name: Report the derived artifact version
+        env:
+          # Diagnostic for the trigger normalisation. `refresh` is reported RAW
+          # as well as normalised, because the two disagreeing is the whole
+          # failure mode: a dispatch input whose type does not compare the way
+          # the expression assumes evaluates to false silently, and the only
+          # symptom is a refresh that quietly behaves like an ordinary push.
+          RAW_REFRESH: ${{ github.event.inputs.refresh }}
+          RAW_FORCE: ${{ github.event.inputs.force_build }}
         run: |
           set -u
+          echo "trigger: event=$GITHUB_EVENT_NAME IS_REFRESH='${IS_REFRESH:-<unset>}' BUILD_REF='${BUILD_REF:-<unset>}'"
+          echo "trigger: raw inputs refresh='${RAW_REFRESH:-<unset>}' force_build='${RAW_FORCE:-<unset>}'"
           A=web
           V=$(sh scripts/artifacts.sh version "$A" 2>&1 || echo UNAVAILABLE)
           R=$(sh scripts/artifacts.sh revision "$A" 2>&1 || echo UNAVAILABLE)
@@ -528,7 +589,7 @@ jobs:
           # Checked BEFORE the ref test, not after: a scheduled run's
           # GITHUB_REF is the default branch (dev), so the main test would
           # never fire on it.
-          if [ "${GITHUB_EVENT_NAME:-}" = "schedule" ]; then
+          if [ "${IS_REFRESH:-}" = "true" ]; then
             echo "tags=git.fabledsword.com/bvandeusen/fabledcurator:latest" >> "$GITHUB_OUTPUT"
             echo "channel=main" >> "$GITHUB_OUTPUT"
           elif [ "${GITHUB_REF##*/}" = "main" ]; then
@@ -628,7 +689,6 @@ jobs:
           # A scheduled refresh has to bypass reuse by construction: it
           # rebuilds the SAME source, so fc.revision always matches and the
           # check would skip every refresh there has ever been.
-          EVENT: ${{ github.event_name }}
         run: |
           set -eu
           DERIVED=$(sh scripts/artifacts.sh revision web)
@@ -638,11 +698,60 @@ jobs:
           # adds no variability the reuse check would have to account for.
           echo "version=$(sh scripts/artifacts.sh version web)" >> "$GITHUB_OUTPUT"
 
+          # The build clock, pinned to the same commit (#3265). Without it
+          # buildkit stamps the image config with the wall clock of the build,
+          # so identical layers republish under a new config blob and the
+          # channel tag gets a new manifest digest for no reason. Derived from
+          # `newest()` like revision and version, so all three name one commit
+          # and cannot drift apart.
+          echo "epoch=$(sh scripts/artifacts.sh epoch web)" >> "$GITHUB_OUTPUT"
+
           # The moving tag for this channel. Which tag we ask IS the channel —
           # that is why the revision needs no -main/-dev qualifier any more.
           if [ "$CHANNEL" = "main" ]; then T=latest; else T=dev; fi
           echo "channel_ref=$IMAGE:$T" >> "$GITHUB_OUTPUT"
 
+          # WHERE THE BUILD PUBLISHES, which is not always the channel — and
+          # whether the channel then has to be written separately.
+          #
+          # On a push the build writes the channel tag directly: the bytes came
+          # from a commit, and a commit is the thing CI tests. Nothing to hold
+          # it behind.
+          #
+          # On the scheduled refresh it writes a CANDIDATE tag instead. A
+          # refresh rebuilds against freshly resolved base images, and the web
+          # image's runtime is a line of UNPINNED Debian packages (ffmpeg,
+          # libjpeg62-turbo, libpq5, megatools…) re-resolved on every build.
+          # Nothing in ci.yml can see that: its lanes run on ci-python:3.14 and
+          # install requirements.txt, and a base bump changes neither. So
+          # refreshed bytes have to be proven before :latest names them, and
+          # proving needs a moment between "built" and "published" to occupy.
+          # This is that moment; :latest goes on naming the build that works
+          # until something says otherwise.
+          #
+          # `:refresh-candidate` is one moving ref per image, overwritten in
+          # place, holding a build nobody is told to pull — the shape rule 145
+          # already allows for :buildcache, not the per-build tag family that
+          # milestone 318 withdrew.
+          #
+          # Decided HERE, beside `hit`, for the reason the force/schedule
+          # branch below gives: one step decides what this job does. A
+          # condition derived independently could disagree with the tag the
+          # build actually wrote.
+          #
+          # build-web additionally exposes this as `outputs.candidate`, which is
+          # what gates the `promote` job — a job's `if:` cannot read `env`, and
+          # one flag is enough because all three derive it from the same
+          # IS_REFRESH. ml and agent do not re-emit it; a second copy nothing
+          # reads is the kind of thing that later reads as load-bearing.
+          if [ "${IS_REFRESH:-}" = "true" ]; then
+            echo "build_ref=$IMAGE:refresh-candidate" >> "$GITHUB_OUTPUT"
+            echo "promote=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "build_ref=$IMAGE:$T" >> "$GITHUB_OUTPUT"
+            echo "promote=false" >> "$GITHUB_OUTPUT"
+          fi
+
           # Compare VALUES, never exit codes. Measured on buildx v0.36.1
           # (run 4732): a missing key returns an empty string and exits 0, so
           # branching on the exit code would read "no label yet" as success.
@@ -675,7 +784,7 @@ jobs:
           if [ "${FORCE:-false}" = "true" ]; then
             echo "hit=false" >> "$GITHUB_OUTPUT"
             echo "reuse: force_build set — building regardless"
-          elif [ "${EVENT:-}" = "schedule" ]; then
+          elif [ "${IS_REFRESH:-}" = "true" ]; then
             echo "hit=false" >> "$GITHUB_OUTPUT"
             echo "reuse: scheduled base refresh — building regardless"
           elif [ -n "$PUBLISHED" ] && [ "$PUBLISHED" = "$DERIVED" ]; then
@@ -764,6 +873,12 @@ jobs:
 
       - name: Build and push web image
         if: steps.reuse.outputs.hit != 'true'
+        # Read by buildx out of the ENVIRONMENT, not passed as a build-arg —
+        # it normalises the image config's `created` field and the history
+        # timestamps rather than being consumed by the Dockerfile. See #3265
+        # and the reuse step's `epoch` output.
+        env:
+          SOURCE_DATE_EPOCH: ${{ steps.reuse.outputs.epoch }}
         uses: docker/build-push-action@v5
         with:
           context: .
@@ -776,20 +891,17 @@ jobs:
           # invalidates, and the image genuinely rebuilds.
           #
           # MEASURED on the first real fire, run 4934 (#3265): when the base
-          # did NOT move, the build is ~13s and every content step reports
+          # did NOT move, the build was ~13s with every content step CACHED —
-          # CACHED — but the channel tag STILL gets a new manifest digest.
+          # and the channel tag STILL got a new manifest digest, because
-          # buildkit mints a fresh image config each run, so identical layers
+          # buildkit stamps a fresh image config per run and republishes the
-          # are republished under a new config blob. All three images moved
+          # identical layers under it. All three images moved that way on
-          # that way on 2026-08-30 with nothing whatsoever changed in them.
+          # 2026-08-30 with nothing whatsoever changed in them.
           #
-          # So a refresh currently rewrites :latest every Sunday whether or
+          # SOURCE_DATE_EPOCH (below) is the fix: pinned to the commit the
-          # not there is anything new in it, and :c-<sha> is handed a new
+          # content came from, the config is byte-identical across runs, so
-          # manifest to diverge from on the same cadence. Layers are shared,
+          # the manifest digest is too and the push is a registry no-op. A
-          # so the storage cost is a config blob; the cost that matters is
+          # digest change means the content changed again, which is the only
-          # that a digest change no longer MEANS anything. Tracked in #3265 —
+          # thing a digest is any use for.
-          # the likely fix is a deterministic SOURCE_DATE_EPOCH, which would
-          # make "same source, same bytes" true and turn the no-op case into
-          # a genuine no-op.
           #
           # What `pull` does NOT catch either: a Debian package update inside
           # the `apt-get install` layer while the base tag itself stands
@@ -799,14 +911,14 @@ jobs:
           # churn #3265 is about.
           #
           # Only on the schedule. An ordinary push wants the cached base.
-          pull: ${{ github.event_name == 'schedule' }}
+          pull: ${{ env.IS_REFRESH == 'true' }}
           # ONE tag, the channel's. Every other tag is written by the step
           # below, registry-side. buildx here pushes the first tag to the
           # registry and then re-pushes the rest through the DOCKER driver,
           # out of a local image store a registry-direct build never filled —
           # #3190, which cost `main` its :c-<sha> on 2026-08-29 while :latest
           # published perfectly well.
-          tags: ${{ steps.reuse.outputs.channel_ref }}
+          tags: ${{ steps.reuse.outputs.build_ref }}
           # The reuse key. Read back off the channel tag on the next push to
           # decide whether that push needs to build at all, so this is not
           # decoration — an unstamped image is one that will always rebuild.
@@ -937,6 +1049,282 @@ jobs:
           docker buildx imagetools create $ARGS "$SOURCE"
           echo "repointed from $SOURCE:$ARGS"
 
+  # Does the image a refresh just built still work?
+  #
+  # This is the gate the base refresh never had. `ci.yml` cannot be it: its
+  # lanes run on ci-python:3.14 and install requirements.txt, and a base bump
+  # changes neither — all five stay green through a refresh that breaks the
+  # product. What a refresh re-resolves is the Dockerfile's apt layer (ffmpeg,
+  # unar, libpq5, postgresql-client, zstd, megatools, libjpeg62-turbo,
+  # libwebp7, libpng16-16), unpinned, every build.
+  #
+  # So this runs the CANDIDATE IMAGE, against real Postgres and Redis. Not the
+  # source tree, and not a static inspection: `ffmpeg -version` exiting 0 would
+  # pass while a codec removal broke every thumbnail in the library.
+  #
+  # Refresh-only. On a push the bytes came from a commit, and a commit is what
+  # ci.yml already tests.
+  #
+  # Reports a verdict; it does not yet gate the promote (milestone 362 step 4).
+  # Landing the gate and the thing it gates in one change would mean the first
+  # time anyone saw this job run would also be the first time it could stop a
+  # publish.
+  smoke-web:
+    needs: [build-web]
+    if: needs.build-web.outputs.candidate == 'true'
+    runs-on: python-ci
+    container:
+      image: git.fabledsword.com/bvandeusen/ci-python:3.14
+    env:
+      DB_USER: fabledcurator
+      DB_PASSWORD: ci_smoke
+      DB_PORT: "5432"
+      DB_NAME: fabledcurator_smoke
+      SECRET_KEY: ci_smoke_placeholder
+      IMAGE: git.fabledsword.com/bvandeusen/fabledcurator
+    services:
+      postgres:
+        image: pgvector/pgvector:pg16
+        env:
+          POSTGRES_USER: fabledcurator
+          POSTGRES_PASSWORD: ci_smoke
+          POSTGRES_DB: fabledcurator_smoke
+        options: >-
+          --health-cmd "pg_isready -U fabledcurator"
+          --health-interval 10s
+          --health-timeout 5s
+          --health-retries 10
+      redis:
+        image: redis:7-alpine
+        options: >-
+          --health-cmd "redis-cli ping"
+          --health-interval 10s
+          --health-timeout 5s
+          --health-retries 10
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          # The same ref the image was built from, so the smoke script matches
+          # the code inside the candidate.
+          ref: ${{ env.BUILD_REF }}
+
+      - name: Smoke the candidate image
+        env:
+          TOKEN: ${{ secrets.RELEASE_TOKEN }}
+          ACTOR: ${{ github.actor }}
+        run: |
+          set -eux
+          # Service discovery mirrors ci.yml's integration lane: these jobs run
+          # in a container against a mounted docker socket, so the services are
+          # SIBLINGS reachable by IP, not by hostname.
+          PG=$(docker ps --filter "name=smoke" --filter "ancestor=pgvector/pgvector:pg16" -q | head -n1)
+          RD=$(docker ps --filter "name=smoke" --filter "ancestor=redis:7-alpine" -q | head -n1)
+          test -n "$PG" && test -n "$RD"
+          PG_IP=$(docker inspect -f '{{range .NetworkSettings.Networks}}{{.IPAddress}}{{end}}' "$PG")
+          RD_IP=$(docker inspect -f '{{range .NetworkSettings.Networks}}{{.IPAddress}}{{end}}' "$RD")
+          test -n "$PG_IP" && test -n "$RD_IP"
+
+          # Socket probe in python, not bash's /dev/tcp — these steps run under
+          # `sh -e`, where that path does not exist. Same fix and reasoning as
+          # ci.yml's integration job; see the comment there.
+          pg_ready=""
+          for i in $(seq 1 60); do
+            if python -c "import socket,sys; s=socket.socket(); s.settimeout(2); sys.exit(0 if s.connect_ex(('$PG_IP', 5432)) == 0 else 1)"; then
+              pg_ready=1
+              break
+            fi
+            sleep 2
+          done
+          if [ -z "$pg_ready" ]; then
+            echo "postgres at $PG_IP:5432 did not accept a connection within 120s"
+            exit 1
+          fi
+
+          echo "$TOKEN" | docker login git.fabledsword.com -u "$ACTOR" --password-stdin
+          CANDIDATE="$IMAGE:refresh-candidate"
+          docker pull "$CANDIDATE"
+
+          ENVOPTS="-e DB_USER=$DB_USER -e DB_PASSWORD=$DB_PASSWORD -e DB_HOST=$PG_IP"
+          ENVOPTS="$ENVOPTS -e DB_PORT=5432 -e DB_NAME=$DB_NAME -e SECRET_KEY=$SECRET_KEY"
+          ENVOPTS="$ENVOPTS -e CELERY_BROKER_URL=redis://$RD_IP:6379/0"
+          ENVOPTS="$ENVOPTS -e CELERY_RESULT_BACKEND=redis://$RD_IP:6379/0"
+          # A throwaway CI instance IS first-time setup, which is the one case
+          # credential_crypto allows a key to be minted in. Without it the web
+          # role refuses to boot — deliberately, since silently generating a
+          # key on a restored-DB-but-lost-secrets deployment would leave every
+          # Credential row undecryptable (the 2026-06-02 audit). Discovered by
+          # this job on its first real run; see #3422 for the fact that no
+          # user-facing file mentions this variable at all.
+          ENVOPTS="$ENVOPTS -e CURATOR_BOOTSTRAP_NEW_KEY=1"
+
+          # 1. The schema builds from empty, using the image's OWN libpq and
+          #    psycopg. This is the same call entrypoint.sh makes before it
+          #    serves anything, so a failure here is a failure to boot.
+          echo "smoke: alembic upgrade head"
+          docker run --rm $ENVOPTS "$CANDIDATE" alembic upgrade head
+
+          # 2. The apt layer's binaries and the app's own thumbnail path, run
+          #    inside the image. Piped over stdin rather than bind-mounted: the
+          #    workspace is a docker VOLUME belonging to this job's container,
+          #    so a host bind of $PWD would not resolve for a sibling.
+          echo "smoke: image-internal checks"
+          docker run --rm -i $ENVOPTS "$CANDIDATE" shell -c 'python3 -' < scripts/smoke_image.py
+
+          # 3. It actually serves. `docker run -d` then poll the container's own
+          #    IP — no port publishing, because the job container reaches
+          #    siblings directly and a published port would collide with
+          #    whatever else the runner is hosting.
+          echo "smoke: web boots and answers /api/health"
+          CID=$(docker run -d $ENVOPTS "$CANDIDATE" web)
+          # Clean up the container however this ends, and dump its log ONLY
+          # on failure — a boot that never answers must fail with the reason
+          # visible rather than as a bare timeout (rule 156), while a green run
+          # has nothing to say. `exit $rc` preserves the real status, which a
+          # trap that ends on a successful `docker rm` would otherwise mask.
+          trap 'rc=$?; [ $rc -eq 0 ] || docker logs "$CID" 2>&1 | tail -40; docker rm -f "$CID" >/dev/null 2>&1 || true; exit $rc' EXIT
+          WEB_IP=$(docker inspect -f '{{range .NetworkSettings.Networks}}{{.IPAddress}}{{end}}' "$CID")
+          test -n "$WEB_IP"
+          healthy=""
+          for i in $(seq 1 60); do
+            if curl -fsS --max-time 5 "http://$WEB_IP:8080/api/health" >/dev/null 2>&1; then
+              healthy=1
+              break
+            fi
+            # A container that has EXITED will never answer, so stop asking.
+            # Without this the loop spent 3m35s polling a dead container on
+            # this job's first run, and — because docker recycles the IP — got
+            # a confusing mix of connection-refused and 5s timeouts from
+            # whatever took the address next. The trap's log dump had the real
+            # answer the whole time; this just stops burying it.
+            if [ "$(docker inspect -f '{{.State.Running}}' "$CID" 2>/dev/null)" != "true" ]; then
+              echo "smoke: FAILED — the web container exited during boot." >&2
+              echo "smoke: its log follows; entrypoint runs alembic BEFORE" >&2
+              echo "smoke: serving, so a startup exception lands here." >&2
+              exit 1
+            fi
+            sleep 2
+          done
+          if [ -z "$healthy" ]; then
+            # 60 iterations of (up to 5s connect + 2s sleep) — up to ~7min, not
+            # the 120s an earlier version of this message claimed.
+            echo "smoke: FAILED — web is running but never answered" >&2
+            echo "smoke: /api/health. It is up, so look at hypercorn and the" >&2
+            echo "smoke: python base rather than at startup." >&2
+            exit 1
+          fi
+          curl -fsS --max-time 5 "http://$WEB_IP:8080/api/health"
+          echo
+
+          echo "smoke: all checks passed against $CANDIDATE"
+
+  # Move the channel tags — the whole point of the gate.
+  #
+  # Lives in its own job because the verdict it depends on cannot exist until
+  # after build-web has finished, and the promote used to run INSIDE build-web.
+  #
+  # `needs` on smoke-web is the gate. A failed smoke skips this job, so a
+  # refresh that broke something leaves :latest naming the build that works —
+  # "the refresh failed" and "production is broken" must not be the same event.
+  # A SKIPPED smoke also skips this job, which is the behaviour that matters
+  # most: on run 5290 the gate silently skipped itself, and a design where only
+  # a FAILED gate blocks would have published unverified images while reporting
+  # success. Not running is not the same as passing.
+  #
+  # All three images promote TOGETHER, or none do. They are one stack: build.yml
+  # already refuses to publish a :dev web image beside a stale :dev ml, because
+  # the mismatch only shows up as a runtime failure. A refresh that published ml
+  # and withheld web would be that same trap, arrived at through the gate.
+  #
+  # The gate covers the web image only (milestone 362 step 3 scoped it there),
+  # so ml and agent are being held to web's verdict rather than their own. That
+  # is deliberate and it is the conservative direction — they ship together, so
+  # the weakest evidence should govern all three — but it is not the same as
+  # having smoked them, and it should not be read as if it were.
+  promote:
+    needs: [build-web, build-ml, build-agent, smoke-web]
+    # Only a refresh publishes through a candidate; a push writes its channel
+    # tag directly from the build. Reads the same reuse-step decision the build
+    # took, via a job output — a job's `if:` cannot see the `env` context.
+    if: needs.build-web.outputs.candidate == 'true'
+    runs-on: python-ci
+    container:
+      image: git.fabledsword.com/bvandeusen/ci-python:3.14
+    steps:
+      - name: Point the channel tags at the smoked candidates
+        env:
+          TOKEN: ${{ secrets.RELEASE_TOKEN }}
+          ACTOR: ${{ github.actor }}
+        run: |
+          set -eu
+          # `latest` is not a guess: a refresh always builds `main` (BUILD_REF),
+          # and the "must have checked out main" guard in every build job fails
+          # the run if that did not hold. So the channel is main's.
+          TAG=latest
+          FAILED=""
+
+          for NAME in fabledcurator fabledcurator-ml fabledcurator-agent; do
+            REPO="bvandeusen/$NAME"
+            echo "promote: $REPO"
+
+            # Registry auth is its own token exchange — `docker login`
+            # authenticates the docker client, not curl. Deadline on every call
+            # (rule 156): a registry that stops answering must fail this step,
+            # not hang the weekly refresh until the job times out.
+            BEARER=$(curl -fsS --max-time 30 -u "$ACTOR:$TOKEN" \
+              "https://git.fabledsword.com/v2/token?scope=repository:$REPO:pull,push&service=git.fabledsword.com" \
+              | python3 -c 'import sys,json; print(json.load(sys.stdin)["token"])')
+
+            # Ask for the IMAGE manifest media types only. Offering the index
+            # types too would let the registry hand back an index if one ever
+            # existed at this tag, and we would faithfully copy the thing this
+            # whole approach exists to avoid creating.
+            ACCEPT='application/vnd.oci.image.manifest.v1+json, application/vnd.docker.distribution.manifest.v2+json'
+            CT=$(curl -fsS --max-time 60 -o manifest.json -D headers.txt \
+                   -H "Authorization: Bearer $BEARER" -H "Accept: $ACCEPT" \
+                   "https://git.fabledsword.com/v2/$REPO/manifests/refresh-candidate" \
+                 && tr -d '\r' < headers.txt | awk -F': ' '/^[Cc]ontent-[Tt]ype:/{print $2}')
+            test -n "$CT"
+            SRC=$(tr -d '\r' < headers.txt | awk -F': ' '/^[Dd]ocker-[Cc]ontent-[Dd]igest:/{print $2}')
+            echo "promote:   candidate $SRC ($CT)"
+
+            # NOT `imagetools create`. That wraps its source in an INDEX, and
+            # `.Image.Config.Labels` does not resolve through one — the
+            # fc.revision the reuse check reads off the channel tag would come
+            # back empty, every later push would miss and rebuild, and nothing
+            # would go red (#3183, run 4751). A manifest PUT is what "make this
+            # tag name that image" means at the registry: same bytes, same media
+            # type, same digest, no layer transfer.
+            curl -fsS --max-time 120 -X PUT \
+              -H "Authorization: Bearer $BEARER" -H "Content-Type: $CT" \
+              --data-binary @manifest.json \
+              "https://git.fabledsword.com/v2/$REPO/manifests/$TAG"
+
+            # Read it back. A PUT that returned 2xx but landed something else is
+            # exactly the silent-and-plausible failure this pipeline keeps
+            # producing, and the check costs one request.
+            NOW=$(curl -fsS --max-time 30 -o /dev/null -D - \
+                    -H "Authorization: Bearer $BEARER" -H "Accept: $ACCEPT" \
+                    "https://git.fabledsword.com/v2/$REPO/manifests/$TAG" \
+                  | tr -d '\r' | awk -F': ' '/^[Dd]ocker-[Cc]ontent-[Dd]igest:/{print $2}')
+            if [ "$NOW" != "$SRC" ]; then
+              echo "promote:   FAILED — $NAME:$TAG is $NOW, expected $SRC" >&2
+              FAILED="$FAILED $NAME"
+              continue
+            fi
+            echo "promote:   $NAME:$TAG now names $NOW"
+          done
+
+          if [ -n "$FAILED" ]; then
+            echo "" >&2
+            echo "promote: FAILED for:$FAILED" >&2
+            echo "promote: the channel tags are now INCONSISTENT — some images" >&2
+            echo "promote: moved and some did not. Re-run this refresh; the" >&2
+            echo "promote: candidates are still published and the promote is" >&2
+            echo "promote: idempotent." >&2
+            exit 1
+          fi
+          echo "promote: all three channel tags moved"
+
   build-ml:
     runs-on: python-ci
     container:
@@ -957,7 +1345,7 @@ jobs:
 
       # See sign-extension's copy for why this guard exists.
       - name: Guard — a scheduled run must have checked out main
-        if: github.event_name == 'schedule'
+        if: env.IS_REFRESH == 'true'
         run: |
           set -eu
           BRANCH=$(git rev-parse --abbrev-ref HEAD)
@@ -990,8 +1378,18 @@ jobs:
       #     the PREVIOUS XPI while the freshly signed one is orphaned (#3156).
       #   * dev and main derive the same values for the same source.
       - name: Report the derived artifact version
+        env:
+          # Diagnostic for the trigger normalisation. `refresh` is reported RAW
+          # as well as normalised, because the two disagreeing is the whole
+          # failure mode: a dispatch input whose type does not compare the way
+          # the expression assumes evaluates to false silently, and the only
+          # symptom is a refresh that quietly behaves like an ordinary push.
+          RAW_REFRESH: ${{ github.event.inputs.refresh }}
+          RAW_FORCE: ${{ github.event.inputs.force_build }}
         run: |
           set -u
+          echo "trigger: event=$GITHUB_EVENT_NAME IS_REFRESH='${IS_REFRESH:-<unset>}' BUILD_REF='${BUILD_REF:-<unset>}'"
+          echo "trigger: raw inputs refresh='${RAW_REFRESH:-<unset>}' force_build='${RAW_FORCE:-<unset>}'"
           A=ml
           V=$(sh scripts/artifacts.sh version "$A" 2>&1 || echo UNAVAILABLE)
           R=$(sh scripts/artifacts.sh revision "$A" 2>&1 || echo UNAVAILABLE)
@@ -1008,7 +1406,7 @@ jobs:
           SHORT_SHA=$(printf '%s' "$GITHUB_SHA" | cut -c1-7)
           # Mirrors build-web's tag list and its schedule handling; see
           # the comments there.
-          if [ "${GITHUB_EVENT_NAME:-}" = "schedule" ]; then
+          if [ "${IS_REFRESH:-}" = "true" ]; then
             echo "tags=git.fabledsword.com/bvandeusen/fabledcurator-ml:latest" >> "$GITHUB_OUTPUT"
             echo "channel=main" >> "$GITHUB_OUTPUT"
           elif [ "${GITHUB_REF##*/}" = "main" ]; then
@@ -1091,17 +1489,63 @@ jobs:
           # A scheduled refresh has to bypass reuse by construction: it
           # rebuilds the SAME source, so fc.revision always matches and the
           # check would skip every refresh there has ever been.
-          EVENT: ${{ github.event_name }}
         run: |
           set -eu
           DERIVED=$(sh scripts/artifacts.sh revision ml)
           echo "revision=$DERIVED" >> "$GITHUB_OUTPUT"
 
+          # The build clock, pinned to the same commit (#3265). Without it
+          # buildkit stamps the image config with the wall clock of the build,
+          # so identical layers republish under a new config blob and the
+          # channel tag gets a new manifest digest for no reason. Derived from
+          # `newest()` like revision and version, so all three name one commit
+          # and cannot drift apart.
+          echo "epoch=$(sh scripts/artifacts.sh epoch ml)" >> "$GITHUB_OUTPUT"
+
           # The moving tag for this channel. Which tag we ask IS the channel —
           # that is why the revision needs no -main/-dev qualifier any more.
           if [ "$CHANNEL" = "main" ]; then T=latest; else T=dev; fi
           echo "channel_ref=$IMAGE:$T" >> "$GITHUB_OUTPUT"
 
+          # WHERE THE BUILD PUBLISHES, which is not always the channel — and
+          # whether the channel then has to be written separately.
+          #
+          # On a push the build writes the channel tag directly: the bytes came
+          # from a commit, and a commit is the thing CI tests. Nothing to hold
+          # it behind.
+          #
+          # On the scheduled refresh it writes a CANDIDATE tag instead. A
+          # refresh rebuilds against freshly resolved base images, and the web
+          # image's runtime is a line of UNPINNED Debian packages (ffmpeg,
+          # libjpeg62-turbo, libpq5, megatools…) re-resolved on every build.
+          # Nothing in ci.yml can see that: its lanes run on ci-python:3.14 and
+          # install requirements.txt, and a base bump changes neither. So
+          # refreshed bytes have to be proven before :latest names them, and
+          # proving needs a moment between "built" and "published" to occupy.
+          # This is that moment; :latest goes on naming the build that works
+          # until something says otherwise.
+          #
+          # `:refresh-candidate` is one moving ref per image, overwritten in
+          # place, holding a build nobody is told to pull — the shape rule 145
+          # already allows for :buildcache, not the per-build tag family that
+          # milestone 318 withdrew.
+          #
+          # Decided HERE, beside `hit`, for the reason the force/schedule
+          # branch below gives: one step decides what this job does. A
+          # condition derived independently could disagree with the tag the
+          # build actually wrote.
+          #
+          # build-web additionally exposes this as `outputs.candidate`, which is
+          # what gates the `promote` job — a job's `if:` cannot read `env`, and
+          # one flag is enough because all three derive it from the same
+          # IS_REFRESH. ml and agent do not re-emit it; a second copy nothing
+          # reads is the kind of thing that later reads as load-bearing.
+          if [ "${IS_REFRESH:-}" = "true" ]; then
+            echo "build_ref=$IMAGE:refresh-candidate" >> "$GITHUB_OUTPUT"
+          else
+            echo "build_ref=$IMAGE:$T" >> "$GITHUB_OUTPUT"
+          fi
+
           # Compare VALUES, never exit codes. Measured on buildx v0.36.1
           # (run 4732): a missing key returns an empty string and exits 0, so
           # branching on the exit code would read "no label yet" as success.
@@ -1134,7 +1578,7 @@ jobs:
           if [ "${FORCE:-false}" = "true" ]; then
             echo "hit=false" >> "$GITHUB_OUTPUT"
             echo "reuse: force_build set — building regardless"
-          elif [ "${EVENT:-}" = "schedule" ]; then
+          elif [ "${IS_REFRESH:-}" = "true" ]; then
             echo "hit=false" >> "$GITHUB_OUTPUT"
             echo "reuse: scheduled base refresh — building regardless"
           elif [ -n "$PUBLISHED" ] && [ "$PUBLISHED" = "$DERIVED" ]; then
@@ -1147,6 +1591,12 @@ jobs:
 
       - name: Build and push ml image
         if: steps.reuse.outputs.hit != 'true'
+        # Read by buildx out of the ENVIRONMENT, not passed as a build-arg —
+        # it normalises the image config's `created` field and the history
+        # timestamps rather than being consumed by the Dockerfile. See #3265
+        # and the reuse step's `epoch` output.
+        env:
+          SOURCE_DATE_EPOCH: ${{ steps.reuse.outputs.epoch }}
         uses: docker/build-push-action@v5
         with:
           context: .
@@ -1159,20 +1609,17 @@ jobs:
           # invalidates, and the image genuinely rebuilds.
           #
           # MEASURED on the first real fire, run 4934 (#3265): when the base
-          # did NOT move, the build is ~13s and every content step reports
+          # did NOT move, the build was ~13s with every content step CACHED —
-          # CACHED — but the channel tag STILL gets a new manifest digest.
+          # and the channel tag STILL got a new manifest digest, because
-          # buildkit mints a fresh image config each run, so identical layers
+          # buildkit stamps a fresh image config per run and republishes the
-          # are republished under a new config blob. All three images moved
+          # identical layers under it. All three images moved that way on
-          # that way on 2026-08-30 with nothing whatsoever changed in them.
+          # 2026-08-30 with nothing whatsoever changed in them.
           #
-          # So a refresh currently rewrites :latest every Sunday whether or
+          # SOURCE_DATE_EPOCH (below) is the fix: pinned to the commit the
-          # not there is anything new in it, and :c-<sha> is handed a new
+          # content came from, the config is byte-identical across runs, so
-          # manifest to diverge from on the same cadence. Layers are shared,
+          # the manifest digest is too and the push is a registry no-op. A
-          # so the storage cost is a config blob; the cost that matters is
+          # digest change means the content changed again, which is the only
-          # that a digest change no longer MEANS anything. Tracked in #3265 —
+          # thing a digest is any use for.
-          # the likely fix is a deterministic SOURCE_DATE_EPOCH, which would
-          # make "same source, same bytes" true and turn the no-op case into
-          # a genuine no-op.
           #
           # What `pull` does NOT catch either: a Debian package update inside
           # the `apt-get install` layer while the base tag itself stands
@@ -1182,14 +1629,14 @@ jobs:
           # churn #3265 is about.
           #
           # Only on the schedule. An ordinary push wants the cached base.
-          pull: ${{ github.event_name == 'schedule' }}
+          pull: ${{ env.IS_REFRESH == 'true' }}
           # ONE tag, the channel's. Every other tag is written by the step
           # below, registry-side. buildx here pushes the first tag to the
           # registry and then re-pushes the rest through the DOCKER driver,
           # out of a local image store a registry-direct build never filled —
           # #3190, which cost `main` its :c-<sha> on 2026-08-29 while :latest
           # published perfectly well.
-          tags: ${{ steps.reuse.outputs.channel_ref }}
+          tags: ${{ steps.reuse.outputs.build_ref }}
           # The reuse key. Read back off the channel tag on the next push to
           # decide whether that push needs to build at all, so this is not
           # decoration — an unstamped image is one that will always rebuild.
@@ -1331,7 +1778,7 @@ jobs:
 
       # See sign-extension's copy for why this guard exists.
       - name: Guard — a scheduled run must have checked out main
-        if: github.event_name == 'schedule'
+        if: env.IS_REFRESH == 'true'
         run: |
           set -eu
           BRANCH=$(git rev-parse --abbrev-ref HEAD)
@@ -1364,8 +1811,18 @@ jobs:
       #     the PREVIOUS XPI while the freshly signed one is orphaned (#3156).
       #   * dev and main derive the same values for the same source.
       - name: Report the derived artifact version
+        env:
+          # Diagnostic for the trigger normalisation. `refresh` is reported RAW
+          # as well as normalised, because the two disagreeing is the whole
+          # failure mode: a dispatch input whose type does not compare the way
+          # the expression assumes evaluates to false silently, and the only
+          # symptom is a refresh that quietly behaves like an ordinary push.
+          RAW_REFRESH: ${{ github.event.inputs.refresh }}
+          RAW_FORCE: ${{ github.event.inputs.force_build }}
         run: |
           set -u
+          echo "trigger: event=$GITHUB_EVENT_NAME IS_REFRESH='${IS_REFRESH:-<unset>}' BUILD_REF='${BUILD_REF:-<unset>}'"
+          echo "trigger: raw inputs refresh='${RAW_REFRESH:-<unset>}' force_build='${RAW_FORCE:-<unset>}'"
           A=agent
           V=$(sh scripts/artifacts.sh version "$A" 2>&1 || echo UNAVAILABLE)
           R=$(sh scripts/artifacts.sh revision "$A" 2>&1 || echo UNAVAILABLE)
@@ -1377,7 +1834,7 @@ jobs:
           SHORT_SHA=$(printf '%s' "$GITHUB_SHA" | cut -c1-7)
           # Mirrors build-web's tag list and its schedule handling; see
           # the comments there.
-          if [ "${GITHUB_EVENT_NAME:-}" = "schedule" ]; then
+          if [ "${IS_REFRESH:-}" = "true" ]; then
             echo "tags=git.fabledsword.com/bvandeusen/fabledcurator-agent:latest" >> "$GITHUB_OUTPUT"
             echo "channel=main" >> "$GITHUB_OUTPUT"
           elif [ "${GITHUB_REF##*/}" = "main" ]; then
@@ -1460,17 +1917,63 @@ jobs:
           # A scheduled refresh has to bypass reuse by construction: it
           # rebuilds the SAME source, so fc.revision always matches and the
           # check would skip every refresh there has ever been.
-          EVENT: ${{ github.event_name }}
         run: |
           set -eu
           DERIVED=$(sh scripts/artifacts.sh revision agent)
           echo "revision=$DERIVED" >> "$GITHUB_OUTPUT"
 
+          # The build clock, pinned to the same commit (#3265). Without it
+          # buildkit stamps the image config with the wall clock of the build,
+          # so identical layers republish under a new config blob and the
+          # channel tag gets a new manifest digest for no reason. Derived from
+          # `newest()` like revision and version, so all three name one commit
+          # and cannot drift apart.
+          echo "epoch=$(sh scripts/artifacts.sh epoch agent)" >> "$GITHUB_OUTPUT"
+
           # The moving tag for this channel. Which tag we ask IS the channel —
           # that is why the revision needs no -main/-dev qualifier any more.
           if [ "$CHANNEL" = "main" ]; then T=latest; else T=dev; fi
           echo "channel_ref=$IMAGE:$T" >> "$GITHUB_OUTPUT"
 
+          # WHERE THE BUILD PUBLISHES, which is not always the channel — and
+          # whether the channel then has to be written separately.
+          #
+          # On a push the build writes the channel tag directly: the bytes came
+          # from a commit, and a commit is the thing CI tests. Nothing to hold
+          # it behind.
+          #
+          # On the scheduled refresh it writes a CANDIDATE tag instead. A
+          # refresh rebuilds against freshly resolved base images, and the web
+          # image's runtime is a line of UNPINNED Debian packages (ffmpeg,
+          # libjpeg62-turbo, libpq5, megatools…) re-resolved on every build.
+          # Nothing in ci.yml can see that: its lanes run on ci-python:3.14 and
+          # install requirements.txt, and a base bump changes neither. So
+          # refreshed bytes have to be proven before :latest names them, and
+          # proving needs a moment between "built" and "published" to occupy.
+          # This is that moment; :latest goes on naming the build that works
+          # until something says otherwise.
+          #
+          # `:refresh-candidate` is one moving ref per image, overwritten in
+          # place, holding a build nobody is told to pull — the shape rule 145
+          # already allows for :buildcache, not the per-build tag family that
+          # milestone 318 withdrew.
+          #
+          # Decided HERE, beside `hit`, for the reason the force/schedule
+          # branch below gives: one step decides what this job does. A
+          # condition derived independently could disagree with the tag the
+          # build actually wrote.
+          #
+          # build-web additionally exposes this as `outputs.candidate`, which is
+          # what gates the `promote` job — a job's `if:` cannot read `env`, and
+          # one flag is enough because all three derive it from the same
+          # IS_REFRESH. ml and agent do not re-emit it; a second copy nothing
+          # reads is the kind of thing that later reads as load-bearing.
+          if [ "${IS_REFRESH:-}" = "true" ]; then
+            echo "build_ref=$IMAGE:refresh-candidate" >> "$GITHUB_OUTPUT"
+          else
+            echo "build_ref=$IMAGE:$T" >> "$GITHUB_OUTPUT"
+          fi
+
           # Compare VALUES, never exit codes. Measured on buildx v0.36.1
           # (run 4732): a missing key returns an empty string and exits 0, so
           # branching on the exit code would read "no label yet" as success.
@@ -1503,7 +2006,7 @@ jobs:
           if [ "${FORCE:-false}" = "true" ]; then
             echo "hit=false" >> "$GITHUB_OUTPUT"
             echo "reuse: force_build set — building regardless"
-          elif [ "${EVENT:-}" = "schedule" ]; then
+          elif [ "${IS_REFRESH:-}" = "true" ]; then
             echo "hit=false" >> "$GITHUB_OUTPUT"
             echo "reuse: scheduled base refresh — building regardless"
           elif [ -n "$PUBLISHED" ] && [ "$PUBLISHED" = "$DERIVED" ]; then
@@ -1516,6 +2019,12 @@ jobs:
 
       - name: Build and push agent image
         if: steps.reuse.outputs.hit != 'true'
+        # Read by buildx out of the ENVIRONMENT, not passed as a build-arg —
+        # it normalises the image config's `created` field and the history
+        # timestamps rather than being consumed by the Dockerfile. See #3265
+        # and the reuse step's `epoch` output.
+        env:
+          SOURCE_DATE_EPOCH: ${{ steps.reuse.outputs.epoch }}
         uses: docker/build-push-action@v5
         with:
           context: agent
@@ -1528,20 +2037,17 @@ jobs:
           # invalidates, and the image genuinely rebuilds.
           #
           # MEASURED on the first real fire, run 4934 (#3265): when the base
-          # did NOT move, the build is ~13s and every content step reports
+          # did NOT move, the build was ~13s with every content step CACHED —
-          # CACHED — but the channel tag STILL gets a new manifest digest.
+          # and the channel tag STILL got a new manifest digest, because
-          # buildkit mints a fresh image config each run, so identical layers
+          # buildkit stamps a fresh image config per run and republishes the
-          # are republished under a new config blob. All three images moved
+          # identical layers under it. All three images moved that way on
-          # that way on 2026-08-30 with nothing whatsoever changed in them.
+          # 2026-08-30 with nothing whatsoever changed in them.
           #
-          # So a refresh currently rewrites :latest every Sunday whether or
+          # SOURCE_DATE_EPOCH (below) is the fix: pinned to the commit the
-          # not there is anything new in it, and :c-<sha> is handed a new
+          # content came from, the config is byte-identical across runs, so
-          # manifest to diverge from on the same cadence. Layers are shared,
+          # the manifest digest is too and the push is a registry no-op. A
-          # so the storage cost is a config blob; the cost that matters is
+          # digest change means the content changed again, which is the only
-          # that a digest change no longer MEANS anything. Tracked in #3265 —
+          # thing a digest is any use for.
-          # the likely fix is a deterministic SOURCE_DATE_EPOCH, which would
-          # make "same source, same bytes" true and turn the no-op case into
-          # a genuine no-op.
           #
           # What `pull` does NOT catch either: a Debian package update inside
           # the `apt-get install` layer while the base tag itself stands
@@ -1551,14 +2057,14 @@ jobs:
           # churn #3265 is about.
           #
           # Only on the schedule. An ordinary push wants the cached base.
-          pull: ${{ github.event_name == 'schedule' }}
+          pull: ${{ env.IS_REFRESH == 'true' }}
           # ONE tag, the channel's. Every other tag is written by the step
           # below, registry-side. buildx here pushes the first tag to the
           # registry and then re-pushes the rest through the DOCKER driver,
           # out of a local image store a registry-direct build never filled —
           # #3190, which cost `main` its :c-<sha> on 2026-08-29 while :latest
           # published perfectly well.
-          tags: ${{ steps.reuse.outputs.channel_ref }}
+          tags: ${{ steps.reuse.outputs.build_ref }}
           # The reuse key. Read back off the channel tag on the next push to
           # decide whether that push needs to build at all, so this is not
           # decoration — an unstamped image is one that will always rebuild.

.forgejo/workflows/ci.yml

@@ -255,10 +255,27 @@ jobs:
           export DB_HOST="$PG_IP"
           export CELERY_BROKER_URL="redis://$RD_IP:6379/0"
           export CELERY_RESULT_BACKEND="redis://$RD_IP:6379/0"
+          # These steps run under `sh -e`, not bash, so bash's /dev/tcp magic
+          # path does not exist here — the probe this loop used to run could
+          # never succeed and simply burned the full 120s on every run, green
+          # or red, then continued without having established anything. Python
+          # is in the image and needs no installed package for a socket
+          # connect, so it is the probe. Exhausting the budget is now a named
+          # failure rather than a silent fall-through (rule 156): if Postgres
+          # is genuinely not up, that is what the log should say, instead of
+          # whatever the first query happens to raise two minutes later.
+          pg_ready=""
           for i in $(seq 1 60); do
-            (echo > "/dev/tcp/$PG_IP/5432") >/dev/null 2>&1 && break
+            if python -c "import socket,sys; s=socket.socket(); s.settimeout(2); sys.exit(0 if s.connect_ex(('$PG_IP', 5432)) == 0 else 1)"; then
+              pg_ready=1
+              break
+            fi
             sleep 2
           done
+          if [ -z "$pg_ready" ]; then
+            echo "postgres at $PG_IP:5432 did not accept a connection within 120s"
+            exit 1
+          fi
           if command -v uv >/dev/null 2>&1; then
             uv pip install --system -r requirements.txt pytest pytest-asyncio
           else

README.md

@@ -1,14 +1,230 @@
 # FabledCurator
 
-Self-hosted media curation — gallery, ML tagging, and subscription-driven downloading in one app. Part of the FabledSword family.
+<!-- overview:start -->
+Self-hosted media curation — a gallery, ML auto-tagging, and subscription-driven
+downloading in one application. Part of the FabledSword family.
 
-Combines what was [ImageRepo](https://git.fabledsword.com/bvandeusen/ImageRepo) (gallery, ML, importer) and [GallerySubscriber](https://git.fabledsword.com/bvandeusen/GallerySubscriber) (gallery-dl wrapper, subscriptions, credential capture) into a single product.
+## What it does
 
-## Status
+You point it at creators you follow. It downloads what they post, files it,
+tags it, and gives you something better than a folder full of images to look
+through afterwards.
 
-In production. `main` is continuously deployed — every merge to `main` builds
+- **Gallery and browsing.** Images, videos and multi-page works, organised by
-and publishes `:latest` images, so whatever is on `main` is what is running.
+  artist, tag, post and series. A Showcase front page, a filterable gallery, a
-Day-to-day work happens on `dev`, which publishes `:dev` images.
+  similarity-driven Explore view, and a page-turning reader for series.
+- **Subscriptions.** Follows creators on Patreon, SubscribeStar, Pixiv and
+  anything `gallery-dl` supports, on a schedule. Handles paywalled posts using
+  your own logged-in session.
+- **ML tagging.** Runs image models in-container to suggest tags, group
+  characters, find near-duplicates and power similarity search. Suggestions are
+  reviewable — it proposes, you confirm, and it learns which proposals you keep
+  rejecting.
+- **Deduplication and provenance.** Everything that arrives is hashed and
+  deduplicated by content, metadata sidecars are read wherever the source
+  writes them, and every file keeps a record of where it came from.
+- **Maintenance.** Backups, library audits, thumbnail and embedding backfills,
+  orphan cleanup — all from the UI, all as background jobs you can watch.
+
+Everything is configured from the Settings UI and stored in the database. There
+is no config file to edit beyond a handful of bootstrap environment variables.
+<!-- overview:end -->
+
+## Before you expose it
+
+**FabledCurator has no login.** There are no user accounts, no passwords and no
+permission model. Anything that can reach the port is an administrator.
+
+That matters more here than it would in most self-hosted apps, because of what
+this one stores: **live platform session cookies for Patreon, SubscribeStar and
+Pixiv** — accounts that usually have a payment method attached. Whoever reaches
+the port can read them, alongside your entire library.
+
+So:
+
+- Bind it to a LAN, a VPN, or a tunnel you control.
+- Do not port-forward it. Do not put it on a public hostname.
+- A reverse proxy that adds TLS but no authentication **does not help**. If you
+  want it reachable from outside, put an authenticating proxy in front of it —
+  a forward-auth provider, HTTP basic auth, an identity-aware tunnel — and treat
+  that layer as the only thing standing between the internet and your accounts.
+
+This is a deliberate design decision for a single-operator tool on a trusted
+network, not a bug and not an oversight. It is stated here because it decides
+how you are allowed to deploy it. [SECURITY.md](SECURITY.md) covers the rest of
+the threat model.
+
+## Requirements
+
+- **Docker** with Compose v2.
+- **~4 GB RAM** for the app, plus whatever Postgres needs for your library size.
+- **Disk** for your media, plus several GB for ML model weights.
+- **No GPU required.** The ML worker runs on CPU — tagging and embedding are
+  slower, and that is the whole difference. A GPU is only involved if you
+  separately run the optional agent (below), which is a different machine's job.
+
+## Install
+
+```bash
+git clone https://git.fabledsword.com/bvandeusen/FabledCurator.git
+cd FabledCurator
+
+cp .env.example .env
+$EDITOR .env          # set DB_PASSWORD and SECRET_KEY
+
+docker compose -f docker-compose.yml up -d
+```
+
+Then open <http://localhost:8080>.
+
+**The `-f docker-compose.yml` is required, not decoration.** Compose
+auto-merges `docker-compose.override.yml` when you leave it off, and that
+override builds the images locally from source — the contributor path, not
+yours. Naming the file explicitly skips the override and pulls the published
+`:latest` images, which is the stable channel built from `main`.
+
+If you forget it, the symptom is a long build instead of a quick pull.
+
+## First run
+
+The database schema is created automatically on first start — the web container
+runs its migrations before serving. Nothing to initialise by hand.
+
+**One thing does need a deliberate act, and the app will not start without it.**
+FabledCurator encrypts your stored platform credentials with a key it keeps at
+`./images/secrets/credential_key.b64`. On a brand-new install that file does not
+exist, and rather than quietly creating one the app stops:
+
+```
+MissingCredentialKey: Fernet key file not found at /images/secrets/credential_key.b64
+```
+
+Set `CURATOR_BOOTSTRAP_NEW_KEY=1` in your `.env` for the first `up`, then delete
+the line once the container is running. `.env.example` ships it with that
+instruction attached.
+
+The refusal is deliberate, and worth understanding rather than working around:
+auto-creating a key is indistinguishable from the disaster case — a restore that
+brought the database back but lost `./images/secrets` — where it would mint a key
+that cannot decrypt anything, leaving an instance that looks healthy while every
+paywalled download fails. Making you say so once, on an empty install, is the
+price of that not happening silently later.
+
+**Which means: back up `./images/secrets/` alongside your database.** It is the
+only thing that can read your stored credentials. A database restored without it
+needs every credential entered again by hand.
+
+A few other things are worth knowing about the first few minutes:
+
+- **The ML worker downloads its model weights on first boot**, several GB from
+  HuggingFace into `./models`. Until that finishes, tagging is queued rather
+  than broken. It is idempotent — a restart resumes rather than refetches.
+- **The gallery starts empty**, and that is the expected state. Add a creator
+  under **Subscriptions** and it fills as posts come down.
+- **If you already have a library on disk**, there is no screen that imports
+  it, and there is not going to be one. Folder ingestion had a UI until July
+  2026; it was retired once posts began arriving entirely through
+  subscriptions and the browser extension, and the decision to leave it
+  retired is deliberate — the folder path carries complexity the product does
+  not need in order to do its job. The supported way to fill a new install is
+  to add the creators you follow under **Subscriptions** and let it pull.
+
+  The `/api/import/trigger` endpoint is still wired up for anyone who wants to
+  script a one-off against a folder mounted at `./import`, and its progress
+  shows under **Settings → Activity**. Treat it as an unsupported escape
+  hatch rather than a feature: nothing in the UI drives it and nothing else
+  in this README depends on it.
+- **To download from a paywalled account**, FabledCurator needs that account's
+  session — see the browser extension below. Without one it can still fetch
+  public posts.
+- **Check Settings → Overview** to confirm the workers are alive. Every long
+  operation in FabledCurator is a background job, so if the queues are not
+  running, the UI will look like it is ignoring you rather than like it is
+  broken.
+
+## The browser extension
+
+A Firefox extension does two jobs: it hands your logged-in platform sessions to
+FabledCurator so it can download on your behalf, and it adds a creator as a
+subscription in one click from their page.
+
+It ships **inside the web image** — there is no add-on store listing to find.
+Go to **Subscriptions → Settings**, find the *Browser extension* card, and click
+**Install Firefox extension**. The XPI is Mozilla-signed, so Firefox installs it
+like any other add-on; the button serves it directly rather than making you
+download and side-load a file.
+
+It pairs with your instance using an API key generated automatically on first
+use. The bar directly under that card shows the key and can rotate it.
+
+See [extension/README.md](extension/README.md) for what it does in detail.
+
+## The GPU agent
+
+Optional, and separate. If you have a desktop with a graphics card, you can run
+an agent on it that leases ML jobs from FabledCurator over HTTP, does them on
+the GPU, and hands the results back. It never touches the database or Redis, so
+it is safe to run somewhere the rest of the stack is not.
+
+Run it for a burst of tagging, stop it to get your card back. It deploys from
+`agent/docker-compose.yml`, not the main stack — see
+[agent/README.md](agent/README.md).
+
+## Upgrading
+
+```bash
+docker compose -f docker-compose.yml pull
+docker compose -f docker-compose.yml up -d
+```
+
+Migrations run automatically on start. Take a database backup first — Settings →
+Maintenance has one — because the schema moves forward and does not move back.
+
+## Deployment posture
+
+FabledCurator is built to run inside a homelab over plain HTTP. It does not
+generate certificates, redirect to HTTPS, or set HSTS. If you want TLS,
+terminate it at your reverse proxy. See [Before you expose it](#before-you-expose-it)
+for why TLS alone is not enough.
+
+## Troubleshooting
+
+**The UI loads but nothing ever finishes.** The web container is up and the
+workers are not. `docker compose -f docker-compose.yml ps` — check `worker`,
+`scheduler` and `ml-worker` are healthy, not restarting.
+
+**`docker compose up` started building instead of pulling.** You left off
+`-f docker-compose.yml`, so the dev override took over. See [Install](#install).
+
+**Downloads fail with an auth error.** The stored session for that platform has
+expired. Re-capture it with the extension; sessions do not last forever.
+
+**Which build am I running?** The foot of Settings shows a version and a
+channel, and `/api/health` returns the same two fields. There are no version
+tags on the images, so this is the authoritative answer.
+
+---
+
+# Developing FabledCurator
+
+Everything below is about working on FabledCurator rather than running it. If
+you are installing it, you are done — see [CONTRIBUTING.md](CONTRIBUTING.md) if
+you want to send a patch.
+
+## Status and channels
+
+In production. `main` is continuously deployed — every merge builds and
+publishes `:latest`, so whatever is on `main` is what is running. Day-to-day
+work happens on `dev`, which publishes `:dev`.
+
+For local development, the dev override handles everything:
+
+```bash
+docker compose up -d          # note: no -f, so the override applies
+```
+
+That builds the images from source, turns on DEBUG logging, and exposes
+Postgres and Redis on the host. No `.env` required.
 
 ## Versions and tags
 
@@ -47,48 +263,10 @@ Five deployable pieces, built by `.forgejo/workflows/build.yml`:
 | --- | --- | --- | --- |
 | **Web / workers** | `Dockerfile` | `fabledcurator` | Quart API + the built Vue SPA in one image. `entrypoint.sh` picks the role: `web`, `worker`, `scheduler`. The `maintenance-long` service is a second `worker` pinned to the long-running maintenance queue. |
 | **ML worker** | `Dockerfile.ml` | `fabledcurator-ml` | Same app, plus `requirements-ml.txt` — tagging and embedding models that run in-container. |
-| **GPU agent** | `agent/Dockerfile` | `fabledcurator-agent` | Optional desktop-GPU worker (`agent/`). Leases jobs over **HTTP only** — never touches the database or Redis. Run it for a burst, stop it to reclaim the card. See `agent/README.md`. |
+| **GPU agent** | `agent/Dockerfile` | `fabledcurator-agent` | Optional desktop-GPU worker (`agent/`). Leases jobs over **HTTP only** — never touches the database or Redis. See `agent/README.md`. |
 | **Firefox extension** | `extension/` | signed XPI | MV3 extension: pushes platform session cookies into FC and adds a creator as a Source in one click. AMO-signed on both `dev` and `main` (one signature per extension change, shared by the two channels), bundled into that channel's web image and served from Settings → Maintenance. See `extension/README.md`. |
 | **Data** | — | `pgvector/pgvector:pg16`, `redis:7-alpine` | Postgres with pgvector for embeddings; Redis as the Celery broker. |
 
-## Quick start
-
-For local development and testing, just:
-
-```bash
-docker compose up -d
-# UI: http://localhost:8080
-```
-
-That uses sane dev defaults baked into `docker-compose.yml` and the dev
-override (`docker-compose.override.yml`, auto-merged) — local builds, DEBUG
-logging, exposed Postgres + Redis ports on the host. No `.env` required.
-
-For a production-like deployment, override the dev defaults via shell env
-or a `.env` file (see `.env.example` for the variable names) and use:
-
-```bash
-docker compose -f docker-compose.yml up -d
-# (skips the dev override, so containers pull published :latest images)
-```
-
-`-f` is doing real work there: it tells Compose to use *only* that file, which
-skips `docker-compose.override.yml` and its local builds. What you get is the
-`:latest` images — the stable channel, built from `main`. This is the install
-path, and it is the one to use if you are running FabledCurator rather than
-working on it.
-
-`:dev` is the other channel: rebuilt from the `dev` branch several times a day,
-bleeding edge, no stability promise. Nothing in this repo points an installer at
-it, and nothing should.
-
-The GPU agent is deployed separately, on the machine with the card —
-`agent/docker-compose.yml`, not this stack.
-
-## Deployment posture
-
-FabledCurator is designed to run inside a self-hosted homelab environment over plain HTTP. If you want TLS, terminate it at your reverse proxy. The app does not generate certificates, redirect to HTTPS, or set HSTS.
-
 ## CI / Forgejo setup
 
 Four workflows: `ci.yml` (lint, extension-version check, backend unit tests,
@@ -118,6 +296,15 @@ source, so `main` finds `dev`'s signature already cached and makes no second AMO
 call. That cache is why signing must be one-shot — AMO rejects a re-signed
 version.
 
+## History
+
+FabledCurator combines what was
+[ImageRepo](https://git.fabledsword.com/bvandeusen/ImageRepo) (gallery, ML,
+importer) and
+[GallerySubscriber](https://git.fabledsword.com/bvandeusen/GallerySubscriber)
+(gallery-dl wrapper, subscriptions, credential capture) into a single product.
+Both are superseded; neither is maintained.
+
 ## License
 
 **GNU Affero General Public License v3.0** — see [LICENSE](LICENSE).

SECURITY.md

@@ -32,29 +32,43 @@ they shape what counts as a serious bug here:
   them, or lets one user of a shared instance read another's is high severity.
 - **An extension API key.** The Firefox extension authenticates to the backend
   with a shared key. Anything that leaks it or lets it be bypassed is a way in.
-- **A multi-user sharing ACL.** Instances can be shared. A bug that lets one
+- **No authentication of its own.** This is the most important thing on this
-  account see content another has not shared is an access-control failure, not
+  page. FabledCurator has no login, no user accounts and no permission model —
-  a cosmetic one.
+  there is no `User` table and no session auth anywhere in the backend. Every
+  HTTP client that can reach the port is the administrator, with full read and
+  write access to everything above, including the stored platform credentials.
+  Access control is entirely the operator's job, done at the network layer.
+  Reports that an unauthenticated caller can reach an endpoint are therefore
+  describing the design; reports that something *crosses the network boundary
+  the operator drew* — an SSRF, a request forgery that rides a browser the
+  operator already has open, a path that leaks state to an origin the operator
+  did not authorise — are in scope and are serious.
 - **Arbitrary media from the internet.** Downloaded files are decoded, hashed,
   thumbnailed and fed to ML models. Anything that turns a hostile file into
   code execution is in scope.
 
 ## Deployment posture — read this before reporting
 
-FabledCurator is designed to run **inside a private network, over plain HTTP**.
+FabledCurator is designed to run **inside a private network, over plain HTTP,
-It does not terminate TLS, redirect to HTTPS, or set HSTS; if you want
+reachable only by its operator**. It does not terminate TLS, redirect to
-transport security, terminate it at your reverse proxy. This is a documented
+HTTPS, or set HSTS; if you want transport security, terminate it at your
-design decision, not an oversight.
+reverse proxy. It also does not authenticate anyone — see above. These are
+documented design decisions, not oversights.
 
-Reports that reduce to "the application is served over HTTP" or "there is no
+Putting this on the public internet, with or without TLS, hands whoever finds
-HSTS header" describe that decision rather than a vulnerability. Reports that
+it your Patreon, SubscribeStar and Pixiv sessions. A reverse proxy that adds
-an authenticated operator can cause the software to do something destructive
+TLS but not an authentication layer does not change that.
-are usually also by design — the operator is the administrator of their own
+
-instance.
+Reports that reduce to "the application is served over HTTP", "there is no
+HSTS header", or "the API needs no credentials" describe those decisions
+rather than vulnerabilities. Reports that the operator can cause the software
+to do something destructive are usually also by design — the operator is the
+administrator of their own instance.
 
 What remains in scope is everything that crosses a boundary the software is
-supposed to hold: between one user and another, between an unauthenticated
+actually supposed to hold: between untrusted downloaded content and the host,
-visitor and any of it, and between untrusted downloaded content and the host.
+between a third-party origin and an operator's open browser session, and
+between the credentials at rest and anything that is not the operator.
 
 ## Supported versions
 

alembic/versions/0090_service_seen.py

@@ -0,0 +1,64 @@
+"""service_seen — the learned roster that makes a stopped part observable.
+
+Milestone 365. Nothing in FabledCurator knew what was SUPPOSED to be running:
+`celery inspect` reports the workers that answer, so a dead worker was a
+shorter list rather than a red light, and the only surface that could tell an
+operator otherwise was Portainer. This table is the memory that turns an
+absence into something the app can see.
+
+Keyed on the queue set for a celery role and on agent_id for the GPU agent —
+NOT on the celery worker name, which here is `celery@<container id>` and is
+minted fresh on every deploy. See the model docstring for why that choice is
+the whole design.
+
+## First migration on the collapsed baseline
+
+0089 is the single generated baseline that replaced revisions 0001..0089
+(milestone 328). This is the first revision written on top of it, so it is
+also the first evidence that the chain steps forward from the collapse rather
+than merely reproducing the schema — which nothing had demonstrated yet.
+
+An existing install is at 0089 because it ran the real 0089; a fresh one is at
+0089 because it ran the baseline. Both arrive here identically, which was the
+property the collapse was designed around.
+
+Revision ID: 0090
+Revises: 0089
+Create Date: 2026-09-02
+
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0090"
+down_revision: Union[str, None] = "0089"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "service_seen",
+        sa.Column("key", sa.String(length=128), nullable=False),
+        sa.Column("kind", sa.String(length=16), nullable=False),
+        sa.Column("display_name", sa.String(length=64), nullable=False),
+        sa.Column(
+            "first_seen_at", sa.DateTime(timezone=True),
+            server_default=sa.text("now()"), nullable=False,
+        ),
+        sa.Column(
+            "last_seen_at", sa.DateTime(timezone=True),
+            server_default=sa.text("now()"), nullable=False,
+        ),
+        sa.Column("details", sa.JSON(), nullable=False),
+        sa.PrimaryKeyConstraint("key", name=op.f("pk_service_seen")),
+    )
+    # No secondary indexes, deliberately: one row per moving part means every
+    # read is a handful of rows and an index would be write cost buying
+    # nothing (#3301 removed seven of exactly that shape).
+
+
+def downgrade() -> None:
+    op.drop_table("service_seen")

backend/app/api/__init__.py

@@ -38,6 +38,7 @@ def all_blueprints() -> list[Blueprint]:
     from .suggestions import suggestions_bp
     from .system_activity import system_activity_bp
     from .system_backup import system_backup_bp
+    from .system_health import system_health_bp
     from .tags import tags_bp
     from .thumbnails import thumbnails_bp
     return [
@@ -51,6 +52,7 @@ def all_blueprints() -> list[Blueprint]:
         showcase_bp,
         settings_bp,
         system_activity_bp,
+        system_health_bp,
         system_backup_bp,
         admin_bp,
         cleanup_bp,

backend/app/api/gpu.py

@@ -21,6 +21,7 @@ from ..services.gallery_service import image_url
 from ..services.ml.gpu_jobs import GpuJobService, error_dedupe_statements
 from ..services.ml.gpu_triage import classify_reason, recover_defective_image
 from ..services.ml.regions import RegionService
+from ..services.service_roster import touch_service
 
 gpu_bp = Blueprint("gpu", __name__, url_prefix="/api/gpu")
 
@@ -256,6 +257,18 @@ async def lease():
         if not await _agent_authed(session):
             return jsonify({"error": "unauthorized"}), 401
         jobs = await GpuJobService(session).lease(agent_id, batch_size=batch)
+        # The agent cannot be polled — it is HTTP-only and pulls from here, so
+        # web never dials it. A lease IS the check-in, and until milestone 365
+        # it was thrown away: an agent sitting idle with nothing to lease left
+        # no trace at all and was indistinguishable from one switched off a
+        # week ago. Recorded on the call that was already happening.
+        await touch_service(
+            session,
+            key=f"agent:{agent_id}",
+            kind="agent",
+            display_name="GPU agent" if agent_id == "agent" else f"GPU agent ({agent_id})",
+            details={"agent_id": agent_id, "last_call": "lease", "leased": len(jobs)},
+        )
         ml = await MLSettings.load(session)
         # image rows for url/mime in one shot
         ids = [j.image_record_id for j in jobs]
@@ -329,6 +342,13 @@ async def heartbeat():
         if not await _agent_authed(session):
             return jsonify({"error": "unauthorized"}), 401
         n = await GpuJobService(session).heartbeat(agent_id, job_ids)
+        await touch_service(
+            session,
+            key=f"agent:{agent_id}",
+            kind="agent",
+            display_name="GPU agent" if agent_id == "agent" else f"GPU agent ({agent_id})",
+            details={"agent_id": agent_id, "last_call": "heartbeat", "extended": n},
+        )
         await session.commit()
     return jsonify({"extended": n})
 

backend/app/api/system_health.py

@@ -0,0 +1,192 @@
+"""Is every part of FabledCurator running? One verdict, one endpoint.
+
+Milestone 365. The nav indicator and the System page both read this and
+nothing else — composing a verdict is this module's job, not the UI's.
+
+## Two kinds of part, answered two different ways
+
+**Learned** — celery roles and the GPU agent, from `service_seen`. The
+question is "how long since it checked in", and these are the parts that can
+be ABSENT, which is the whole point: `celery inspect` alone reports presence,
+so a dead worker is a shorter list rather than a red light.
+
+**Probed live** — Postgres and Redis. Always expected, never learned, and a
+last-seen for them would be actively misleading: that Redis answered thirty
+seconds ago says nothing about now.
+
+## This endpoint must never fail because something it checks has failed
+
+The inversion is easy to write by accident and it destroys the feature exactly
+when it is needed — a 500 when Redis is down, instead of `redis: down`. Every
+probe is wrapped, every wait has a deadline (rule 156), and the roster refresh
+swallows its own errors. The worst case is a part reported `unknown`, which is
+a true statement.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import time
+from datetime import UTC, datetime
+
+from quart import Blueprint, jsonify
+from sqlalchemy import select, text
+
+from ..config import get_config
+from ..extensions import get_session
+from ..models import ServiceSeen
+from ..services.service_roster import refresh_if_stale
+
+log = logging.getLogger(__name__)
+
+system_health_bp = Blueprint("system_health", __name__, url_prefix="/api/system")
+
+# How long a learned part may go quiet before it is doubted, then disbelieved.
+#
+# These are deliberately generous, and the reason is a deploy rather than a
+# worker: `docker compose up -d` rolls start-first, so a role is briefly served
+# by two containers and then by neither while the old one drains. Thresholds
+# tight enough to catch a crash in seconds would paint the page red every time
+# the stack is updated, and an alarm that cries wolf on every deploy is one
+# nobody reads. Tune down only after watching a real deploy pass through.
+STALE_AFTER_SECONDS = 90
+DOWN_AFTER_SECONDS = 300
+
+# Probes cross a process boundary, so they carry deadlines. A hung Postgres
+# must make this endpoint say "postgres: down", not hang alongside it.
+PROBE_TIMEOUT_SECONDS = 2.0
+
+_OK, _STALE, _DOWN, _UNKNOWN = "ok", "stale", "down", "unknown"
+
+# Worst-first, so an overall verdict is just the max.
+_SEVERITY = {_OK: 0, _UNKNOWN: 1, _STALE: 2, _DOWN: 3}
+
+
+def _age_state(age_seconds: float) -> str:
+    if age_seconds >= DOWN_AFTER_SECONDS:
+        return _DOWN
+    if age_seconds >= STALE_AFTER_SECONDS:
+        return _STALE
+    return _OK
+
+
+def _describe_learned(name: str, state: str, age: float, details: dict) -> str:
+    """Say what the state MEANS. A red chip tells an operator less than a
+    sentence does at the moment they are deciding whether to go and look."""
+    if state == _OK:
+        replicas = details.get("replicas")
+        if replicas and replicas > 1:
+            return f"{name} is running ({replicas} replicas)"
+        return f"{name} is running"
+    mins = int(age // 60)
+    ago = f"{mins} min" if mins else f"{int(age)}s"
+    if state == _STALE:
+        return f"{name} has not checked in for {ago}"
+    return f"{name} has not checked in for {ago} — treat it as stopped"
+
+
+async def _probe_postgres(session) -> dict:
+    started = time.monotonic()
+    try:
+        await asyncio.wait_for(
+            session.execute(text("SELECT 1")), timeout=PROBE_TIMEOUT_SECONDS
+        )
+    except Exception as exc:  # noqa: BLE001 — a probe reports, it never raises
+        return {
+            "key": "postgres", "kind": "datastore", "name": "PostgreSQL",
+            "state": _DOWN, "detail": f"not answering: {type(exc).__name__}",
+        }
+    return {
+        "key": "postgres", "kind": "datastore", "name": "PostgreSQL", "state": _OK,
+        "detail": "answering", "latency_ms": round((time.monotonic() - started) * 1000, 1),
+    }
+
+
+def _ping_redis_sync() -> None:
+    import redis  # local import; mirrors system_activity's pattern
+
+    client = redis.Redis.from_url(
+        get_config().celery_broker_url,
+        socket_connect_timeout=PROBE_TIMEOUT_SECONDS,
+        socket_timeout=PROBE_TIMEOUT_SECONDS,
+    )
+    client.ping()
+
+
+async def _probe_redis() -> dict:
+    started = time.monotonic()
+    try:
+        await asyncio.wait_for(
+            asyncio.to_thread(_ping_redis_sync), timeout=PROBE_TIMEOUT_SECONDS * 2
+        )
+    except Exception as exc:  # noqa: BLE001
+        return {
+            "key": "redis", "kind": "datastore", "name": "Redis",
+            "state": _DOWN,
+            "detail": f"not answering: {type(exc).__name__} — queues and workers "
+                      f"cannot be reached either",
+        }
+    return {
+        "key": "redis", "kind": "datastore", "name": "Redis", "state": _OK,
+        "detail": "answering", "latency_ms": round((time.monotonic() - started) * 1000, 1),
+    }
+
+
+@system_health_bp.route("/health", methods=["GET"])
+async def system_health():
+    """Every part, its state, and one overall verdict.
+
+    Response: {overall, parts: [{key, kind, name, state, detail, last_seen_at,
+    …}], checked_at}
+    """
+    parts: list[dict] = []
+    now = datetime.now(UTC)
+
+    async with get_session() as session:
+        # Postgres first, and if it is unreachable nothing else can be read —
+        # say so rather than failing, because "the database is down" is the
+        # single most useful thing this endpoint can ever report.
+        pg = await _probe_postgres(session)
+        parts.append(pg)
+
+        if pg["state"] == _OK:
+            # Rate-limited inside; see service_roster on why the web process
+            # is the right observer.
+            try:
+                await refresh_if_stale(session)
+                await session.commit()
+            except Exception:  # noqa: BLE001
+                log.warning("system health: roster refresh failed", exc_info=True)
+
+            rows = (
+                await session.execute(select(ServiceSeen).order_by(ServiceSeen.display_name))
+            ).scalars().all()
+            for row in rows:
+                age = (now - row.last_seen_at).total_seconds()
+                state = _age_state(age)
+                parts.append({
+                    "key": row.key,
+                    "kind": row.kind,
+                    "name": row.display_name,
+                    "state": state,
+                    "detail": _describe_learned(row.display_name, state, age, row.details or {}),
+                    "last_seen_at": row.last_seen_at.isoformat(),
+                    "first_seen_at": row.first_seen_at.isoformat(),
+                    **{k: v for k, v in (row.details or {}).items() if k != "agent_id"},
+                })
+
+    parts.append(await _probe_redis())
+
+    overall = max((p["state"] for p in parts), key=lambda s: _SEVERITY[s], default=_UNKNOWN)
+    return jsonify({
+        "overall": overall,
+        "parts": sorted(parts, key=lambda p: (-_SEVERITY[p["state"]], p["name"])),
+        "checked_at": now.isoformat(),
+        # So the UI can explain a `stale` without hard-coding the same numbers
+        # in a second place.
+        "thresholds": {
+            "stale_after_seconds": STALE_AFTER_SECONDS,
+            "down_after_seconds": DOWN_AFTER_SECONDS,
+        },
+    })

backend/app/config.py

@@ -16,8 +16,11 @@ class Config:
     celery_broker_url: str
     celery_result_backend: str
 
+    # Sets Quart's app.secret_key. Nothing signs a cookie today (FC has no
+    # login and no session use), so this currently protects nothing — it is
+    # required rather than defaulted so that the day something session-backed
+    # does land, no instance is already running on a value we published.
     secret_key: str
-    extension_api_key: str  # used by the Firefox extension; lands in FC-3 but read here
     log_level: str
 
     @property
@@ -47,6 +50,5 @@ def get_config() -> Config:
         celery_broker_url=os.environ.get("CELERY_BROKER_URL", "redis://redis:6379/0"),
         celery_result_backend=os.environ.get("CELERY_RESULT_BACKEND", "redis://redis:6379/0"),
         secret_key=os.environ["SECRET_KEY"],
-        extension_api_key=os.environ.get("EXTENSION_API_KEY", ""),
         log_level=os.environ.get("LOG_LEVEL", "INFO"),
     )

backend/app/models/__init__.py

@@ -32,6 +32,7 @@ from .presentation_review import PresentationReview
 from .series_chapter import SeriesChapter
 from .series_page import SeriesPage
 from .series_suggestion import SeriesSuggestion
+from .service_seen import ServiceSeen
 from .source import Source
 from .subscribestar_failed_media import SubscribeStarFailedMedia
 from .subscribestar_seen_media import SubscribeStarSeenMedia
@@ -63,6 +64,7 @@ __all__ = [
     "SeriesChapter",
     "SeriesPage",
     "SeriesSuggestion",
+    "ServiceSeen",
     "ImageRecord",
     "ImageProvenance",
     "ImageRegion",

backend/app/models/service_seen.py

@@ -0,0 +1,88 @@
+"""service_seen — the learned roster of FabledCurator's own moving parts.
+
+Nothing else in this application knows what is SUPPOSED to be running.
+`celery inspect` reports the workers that answer, so a stopped worker is a
+shorter list rather than a red light, and Postgres and Redis have no
+representation at all. That is why the only place an operator could see a
+dead service was Portainer, which knows the intended set (milestone 365).
+
+This table is the memory that makes an absence observable: every part that
+has ever checked in, and when it last did. A row that stops advancing is a
+part that stopped.
+
+## Why the key is not the hostname
+
+`_read_workers_sync()` returns celery's worker names, which here are
+`celery@<container id>`. Those are minted fresh on every deploy. Keyed on
+them, this table would record a death and a birth every time the stack is
+updated — and a status page that goes red on every deploy is a status page
+nobody reads, which is worse than not having one.
+
+So a celery role is keyed on its **queue set**, which is assigned per role in
+docker-compose.yml (`CELERY_QUEUES`) and survives container replacement:
+
+    default,import,thumbnail,download  ->  worker
+    maintenance,scan                   ->  scheduler   (celery worker --beat)
+    ml                                 ->  ml-worker
+
+Two replicas of one role share a queue set and are therefore ONE row — which
+is right, because the question being answered is "is that role being served",
+not "how many containers exist". The replica count and their hostnames go in
+`details`, where they can change without the identity changing.
+
+The GPU agent is keyed on its `agent_id`, the identity its lease protocol
+already uses (`api/gpu.py`).
+
+## What is NOT in here
+
+Postgres and Redis. They are always expected and never learned, and a
+last-seen for them would be actively misleading — that one answered thirty
+seconds ago says nothing about now. They are probed live at request time.
+
+## kind
+
+Plain `String`, not a Postgres ENUM and not CHECK-gated, matching
+`gpu_job.status` and `backup_run.status`. The value set here is expected to
+grow as parts are added, and a constraint swap per new kind (rule 36) would
+be cost with no invariant behind it.
+
+    celery  — a worker role, keyed on its queue set
+    agent   — a GPU agent, keyed on its agent_id
+"""
+
+from datetime import datetime
+
+from sqlalchemy import JSON, DateTime, String, func
+from sqlalchemy.orm import Mapped, mapped_column
+
+from .base import Base
+
+
+class ServiceSeen(Base):
+    __tablename__ = "service_seen"
+
+    # No indexes beyond the primary key, deliberately. This table holds one row
+    # per moving part — a handful, forever — so every query against it is a
+    # full read of a few rows and an index would be write cost buying nothing
+    # (the lesson of #3301, which removed seven redundant ones).
+    key: Mapped[str] = mapped_column(String(128), primary_key=True)
+    kind: Mapped[str] = mapped_column(String(16), nullable=False)
+
+    # What to call it in the UI. Derived from the queue set where it is
+    # recognised, and falling back to the raw queue list where it is not — a
+    # deployment that slices its queues differently should still show something
+    # true rather than a name this code invented for it.
+    display_name: Mapped[str] = mapped_column(String(64), nullable=False)
+
+    first_seen_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now(),
+    )
+    last_seen_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now(),
+    )
+
+    # The parts that change without changing identity: replica hostnames,
+    # active task counts, the queues actually being served. Kept as a blob
+    # because it is displayed and never queried — giving it columns would
+    # invite filtering on it, which is what the activity endpoints are for.
+    details: Mapped[dict] = mapped_column(JSON, nullable=False, default=dict)

backend/app/services/service_roster.py

@@ -0,0 +1,175 @@
+"""The learned roster: which of FabledCurator's parts have checked in, and when.
+
+Milestone 365. `celery inspect` answers "who is here"; this answers "who is
+missing", which nothing in the application could do before — see
+`models/service_seen.py` for why the identity is a queue set and not a
+worker hostname.
+
+## Who does the observing, and why it is the web process
+
+Three candidates, and the choice matters more than the code:
+
+* **A celery beat sweep.** Rejected. If the scheduler dies, the sweep stops,
+  every row goes stale, and the page reports that everything is down when one
+  thing is. An alarm that cannot distinguish "one part died" from "the
+  observer died" is worse than no alarm.
+* **A background task in web.** Rejected on a detail of how this deploys:
+  hypercorn runs `--workers 4`, so a `before_serving` loop would be FOUR
+  concurrent inspect loops hammering the broker, forever, per container.
+* **Refresh on demand, rate-limited by the data itself.** Taken. Whichever web
+  process happens to serve a health request refreshes the roster if it is
+  older than REFRESH_TTL, and otherwise reads what is already there.
+
+The third has the property the other two lack: **the observer is the thing
+serving the page.** If web is down you get a browser error rather than a
+confidently green page, which is the honest failure. It also self-limits
+without coordination — the TTL lives in the row everybody can see.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+
+from sqlalchemy import func, select
+from sqlalchemy.dialects.postgresql import insert as pg_insert
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import ServiceSeen
+
+log = logging.getLogger(__name__)
+
+# How stale the roster may be before a health request refreshes it. Comfortably
+# under the staleness thresholds that decide a service is missing, so the
+# verdict is never limited by how often anyone looked.
+REFRESH_TTL_SECONDS = 20.0
+
+# celery inspect is a broker round trip and this sits on a request path, so it
+# gets a deadline (rule 156). A broker that has stopped answering must make the
+# roster stale — which is a true statement about the system — not hang the one
+# page that exists to explain it.
+INSPECT_TIMEOUT_SECONDS = 2.0
+
+# Queue set -> the name an operator recognises. Sorted-tuple keys, because the
+# order celery reports them in is not guaranteed.
+#
+# A deployment that slices CELERY_QUEUES differently falls through to the raw
+# queue list rather than being given a name this table invented for it: a
+# wrong-but-confident label on a status page is worse than an ugly true one.
+ROLE_NAMES: dict[tuple[str, ...], str] = {
+    ("default", "download", "import", "thumbnail"): "Worker",
+    ("maintenance", "scan"): "Scheduler",
+    ("ml",): "ML worker",
+}
+
+
+def role_display_name(queues: tuple[str, ...]) -> str:
+    known = ROLE_NAMES.get(queues)
+    if known:
+        return known
+    return "Worker (" + ", ".join(queues) + ")"
+
+
+def _inspect_celery_sync() -> dict[tuple[str, ...], dict]:
+    """celery inspect, grouped by queue set rather than by worker.
+
+    Returns {queue_set: {"hostnames": [...], "active": int}}. Two replicas of
+    one role collapse into one entry on purpose — the question is whether the
+    role is being served, not how many containers exist.
+    """
+    from ..celery_app import celery as celery_app
+
+    insp = celery_app.control.inspect(timeout=INSPECT_TIMEOUT_SECONDS)
+    active_queues = insp.active_queues() or {}
+    active_tasks = insp.active() or {}
+
+    grouped: dict[tuple[str, ...], dict] = {}
+    for hostname, queues in active_queues.items():
+        key = tuple(sorted({q["name"] for q in queues}))
+        entry = grouped.setdefault(key, {"hostnames": [], "active": 0})
+        entry["hostnames"].append(hostname)
+        entry["active"] += len(active_tasks.get(hostname, []))
+    for entry in grouped.values():
+        entry["hostnames"].sort()
+    return grouped
+
+
+async def touch_service(
+    session: AsyncSession, *, key: str, kind: str, display_name: str, details: dict
+) -> None:
+    """Record that a part checked in just now.
+
+    Upsert rather than read-modify-write: several web processes and several
+    agents can be doing this at once, and the last writer is simply the most
+    recent sighting. `first_seen_at` is deliberately NOT updated — it is the
+    one field that answers "has this ever run", which the learned-roster design
+    depends on.
+    """
+    stmt = pg_insert(ServiceSeen).values(
+        key=key, kind=kind, display_name=display_name, details=details,
+    )
+    stmt = stmt.on_conflict_do_update(
+        index_elements=[ServiceSeen.key],
+        set_={
+            "kind": stmt.excluded.kind,
+            "display_name": stmt.excluded.display_name,
+            "details": stmt.excluded.details,
+            "last_seen_at": func.now(),
+        },
+    )
+    await session.execute(stmt)
+
+
+async def refresh_celery_roster(session: AsyncSession) -> None:
+    """Inspect the broker and record what answered. Never raises.
+
+    A failure here means the roster does not advance, and the rows going stale
+    is then a TRUE report about a broker nobody can reach. Letting the
+    exception out would instead break the health endpoint, which is the one
+    thing that must keep answering when the stack is unwell.
+    """
+    try:
+        grouped = await asyncio.wait_for(
+            asyncio.to_thread(_inspect_celery_sync),
+            timeout=INSPECT_TIMEOUT_SECONDS * 2,
+        )
+    except Exception:
+        log.warning("service roster: celery inspect failed; roster not refreshed", exc_info=True)
+        return
+
+    for queues, entry in grouped.items():
+        await touch_service(
+            session,
+            key="celery:" + ",".join(queues),
+            kind="celery",
+            display_name=role_display_name(queues),
+            details={
+                "queues": list(queues),
+                "hostnames": entry["hostnames"],
+                "replicas": len(entry["hostnames"]),
+                "active": entry["active"],
+            },
+        )
+
+
+async def refresh_if_stale(session: AsyncSession) -> None:
+    """Refresh the celery roster if nobody has for REFRESH_TTL_SECONDS.
+
+    Rate-limited by the data rather than by a lock: the gate is the newest
+    last_seen_at across the celery rows, which every web process can see. Two
+    processes racing through the gate costs one redundant inspect and writes
+    the same values twice, so the benign outcome needs no coordination to
+    prevent.
+    """
+    newest = (
+        await session.execute(
+            select(func.max(ServiceSeen.last_seen_at)).where(ServiceSeen.kind == "celery")
+        )
+    ).scalar_one_or_none()
+
+    if newest is not None:
+        age = (await session.execute(select(func.now()))).scalar_one() - newest
+        if age.total_seconds() < REFRESH_TTL_SECONDS:
+            return
+
+    await refresh_celery_roster(session)

ci-requirements.md

@@ -198,14 +198,21 @@ per `docs/process.md`'s "add deps to the image when used by >1 project".
     refresh from being undone.
   - **`pull: true` on the scheduled path only** is the mechanism: a moved base
     tag changes the `FROM` layer's cache key and everything above it rebuilds.
-    **It does not currently make the unmoved case free.** Measured on the first
+    It did not always make the unmoved case free. Measured on the first real
-    real fire (run 4934, 2026-08-30): every content step reported `CACHED` and
+    fire (run 4934, 2026-08-30): every content step reported `CACHED` and the
-    the bases resolved to unchanged digests, yet all three `:latest` tags got a
+    bases resolved to unchanged digests, yet all three `:latest` tags got a NEW
-    NEW manifest digest, because buildkit mints a fresh image config per run and
+    manifest digest, because buildkit stamps a fresh image config per run and
-    republishes identical layers under it. So `:latest` is rewritten weekly
+    republishes identical layers under it — so `:latest` was rewritten weekly
-    whether or not anything changed, and `:c-<sha>` is handed a new manifest to
+    whether or not anything changed, and a digest change stopped meaning
-    diverge from on the same cadence — a digest change stops meaning anything.
+    anything (#3265).
-    Tracked as #3265; the likely fix is a deterministic `SOURCE_DATE_EPOCH`.
+  - **`SOURCE_DATE_EPOCH` is what makes it free.** Set on each build step from
+    `artifacts.sh epoch <artifact>` — the unix timestamp of the same commit
+    `revision` and `version` name, so all three are views of one `newest()`
+    lookup and cannot drift into disagreeing. With the config's `created` field
+    and history timestamps pinned to the content rather than to the wall clock,
+    identical source produces an identical manifest digest and the push is a
+    registry no-op. That restores the property the whole scheme rests on: a
+    channel tag's digest changes when, and only when, its content does.
     Separately not caught: a Debian package update inside the `apt-get install`
     layer while the base tag stands still — a lag rather than a hole, since the
     official python/cuda images rebuild with those updates baked in.

docker-compose.yml

@@ -1,12 +1,21 @@
-# Base compose stack. Uses ${VAR:-default} interpolation throughout so the
+# Base compose stack, and the install path. Uses ${VAR:-default} throughout so
-# stack boots with zero config — sane dev defaults baked in. For production
+# the stack boots with zero config — but those defaults are DEV defaults, and
-# deployments, override the defaults via shell env vars or a .env file:
+# two of them (DB_PASSWORD, SECRET_KEY) are published in this file. Copy
+# .env.example to .env and set them before running this anywhere real.
 #
-#     DB_PASSWORD=...real... SECRET_KEY=...real... docker compose up
+# To run FabledCurator:
 #
-# The dev override (docker-compose.override.yml) is auto-merged when you
+#     docker compose -f docker-compose.yml up -d
-# run `docker compose up` from this directory and switches images to
+#
-# local builds + DEBUG logging.
+# The -f is load-bearing. Without it Compose auto-merges
+# docker-compose.override.yml, which replaces every image: with a local
+# build: and turns on DEBUG logging — the contributor path. Naming this file
+# explicitly skips the override and pulls the published :latest images.
+#
+# FabledCurator has no authentication. Whatever can reach ${PORT} is an
+# administrator, including over the stored Patreon/SubscribeStar/Pixiv session
+# cookies. Do not publish this port beyond a network you trust — see
+# "Before you expose it" in README.md.
 
 # Rolling-deploy safety (Swarm / `docker stack deploy`): update one task at a
 # time, START the new task before stopping the old (zero-downtime via the ingress
@@ -123,18 +132,40 @@ services:
       CELERY_BROKER_URL: redis://redis:6379/0
       CELERY_RESULT_BACKEND: redis://redis:6379/0
       SECRET_KEY: ${SECRET_KEY:-dev_secret_key_not_for_production_change_me}
-      EXTENSION_API_KEY: ${EXTENSION_API_KEY:-}
       LOG_LEVEL: ${LOG_LEVEL:-INFO}
+      # First boot only. FabledCurator refuses to start until the credential
+      # encryption key at /images/secrets/credential_key.b64 exists, and
+      # refuses to create one unless told to — auto-creating is
+      # indistinguishable from a restore that lost ./images/secrets, where it
+      # would mint a key that decrypts nothing and leave an instance that looks
+      # healthy while every paywalled download fails.
+      #
+      # Passed through EXPLICITLY because a variable in `.env` is only used for
+      # ${...} interpolation; it does not reach the container unless it is
+      # named here. Defaulted to empty so the refusal stands for everyone who
+      # has not opted in — the app tests for exactly "1".
+      #
+      # Set it in .env for one `up`, then remove it. See .env.example.
+      CURATOR_BOOTSTRAP_NEW_KEY: ${CURATOR_BOOTSTRAP_NEW_KEY:-}
     volumes:
       - ./images:/images
       - ./import:/import
-      # FC-5 legacy migration: bind-mount the host's ImageRepo images dir
+      # /import is a staging area for scripting a one-off ingest of a library
-      # under /import (FC's existing filesystem scan picks them up). Read-only
+      # you already have on disk. Drop files in ./import, or bind-mount an
-      # is sufficient — FC copies into /images during the scan. The worker +
+      # existing directory under it as below, then trigger the scan:
-      # scheduler services see the same /import via their own mounts below
+      #
-      # because of /import volume reuse. Edit the host path to match your
+      #     curl -X POST http://localhost:8080/api/import/trigger
-      # install before running Settings → Maintenance → Legacy migration.
+      #
-      # - /var/lib/imagerepo/images:/import/imagerepo:ro
+      # Read-only is sufficient — FC copies into /images during the scan. The
+      # worker + scheduler services mount the same /import so the scan can run
+      # on whichever lane picks it up.
+      #
+      # Deliberately has no UI. The manual-scan surface was retired 2026-07-02
+      # once imports arrived via subscriptions + the extension, and the call
+      # not to restore it stands (operator, 2026-09-02): folder ingestion
+      # brings complexity the product does not need. The endpoint stays as an
+      # unsupported escape hatch; the supported way in is Subscriptions.
+      # - /srv/media/my-library:/import/my-library:ro
     depends_on:
       postgres: { condition: service_healthy }
       redis:    { condition: service_healthy }

frontend/src/components/TopNav.vue

@@ -5,9 +5,12 @@
         <img src="/favicon.svg" alt="" class="fc-brand__glyph" width="22" height="22" />
         <span class="fc-brand__text">FabledCurator</span>
       </RouterLink>
-      <span class="fc-health" :title="health.label">
+      <RouterLink
+        :to="{ name: 'system' }" class="fc-health" :title="health.label"
+        :aria-label="`System health: ${health.label}`"
+      >
         <v-icon size="x-small" :color="health.color">{{ health.icon }}</v-icon>
-      </span>
+      </RouterLink>
       <PipelineStatusChip />
     </div>
 
@@ -64,13 +67,15 @@
 </template>
 
 <script setup>
-import { computed, onBeforeUnmount, onMounted, ref } from 'vue'
+import { computed, onBeforeUnmount, onMounted, onUnmounted, ref } from 'vue'
 import { useRoute } from 'vue-router'
 import router, { FRONT_DOOR } from '../router.js'
 import { useSystemStore } from '../stores/system.js'
+import { useSystemHealthStore } from '../stores/systemHealth.js'
 import PipelineStatusChip from './PipelineStatusChip.vue'
 
 const system = useSystemStore()
+const healthStore = useSystemHealthStore()
 
 // Publish the nav's REAL height as --fc-nav-h so full-height workspaces
 // (Explore/Subscriptions) and sticky sub-headers pin to it exactly instead of a
@@ -116,15 +121,55 @@ const settingsRoute = computed(() =>
   navRoutes.value.find(r => r.name === 'settings') || null
 )
 
+// The dot beside the brand, and the only ambient signal that something in the
+// stack has stopped (milestone 365).
+//
+// It used to read /api/health — a no-DB liveness check that proves the WEB
+// container is serving and nothing else. Green there while the worker was dead
+// is exactly what it looked like, and a green dot next to the product name is
+// read as "everything is fine". It now reflects the whole-stack verdict.
+//
+// Deliberately re-using this element rather than adding a second indicator:
+// there were already three partial surfaces (this, the pipeline chip, the
+// Settings Activity tab) and a fourth would have made the question harder to
+// answer, not easier. This is the one that already occupied the slot.
 const health = computed(() => {
-  if (system.healthy === null) {
+  const overall = healthStore.overall
+  if (overall === null) {
     return { icon: 'mdi-circle-outline', color: 'on-surface', label: 'checking…' }
   }
-  if (system.healthy === true) {
+  if (overall === 'ok') {
-    return { icon: 'mdi-circle', color: 'success', label: 'healthy' }
+    return { icon: 'mdi-circle', color: 'success', label: 'All parts running' }
   }
-  return { icon: 'mdi-alert-circle', color: 'error', label: 'unreachable' }
+  // Name what is wrong in the tooltip. "Something is unhealthy" sends someone
+  // hunting; "Scheduler has not checked in for 6 min" does not.
+  const worst = healthStore.problems[0]
+  const others = healthStore.problems.length - 1
+  const suffix = others > 0 ? ` (+${others} more)` : ''
+  if (overall === 'down') {
+    return {
+      icon: 'mdi-alert-circle', color: 'error',
+      label: (worst?.detail || 'A part has stopped') + suffix,
+    }
+  }
+  if (overall === 'stale') {
+    return {
+      icon: 'mdi-alert', color: 'warning',
+      label: (worst?.detail || 'A part is quiet') + suffix,
+    }
+  }
+  return { icon: 'mdi-help-circle-outline', color: 'on-surface', label: 'Health unknown' }
 })
+
+const HEALTH_POLL_MS = 15_000
+let healthTimer = null
+onMounted(() => {
+  healthStore.refresh()
+  healthTimer = setInterval(() => {
+    if (!document.hidden) healthStore.refresh()
+  }, HEALTH_POLL_MS)
+})
+onUnmounted(() => { if (healthTimer) clearInterval(healthTimer) })
 </script>
 
 <style scoped>
@@ -237,7 +282,14 @@ const health = computed(() => {
   display: flex;
   align-items: center;
   flex-shrink: 0;
+  /* A RouterLink since milestone 365 — it is the path to /system, not just an
+     indicator. Reset the anchor so turning a span into a link changed nothing
+     about how the nav reads. */
+  text-decoration: none;
+  color: inherit;
+  border-radius: 50%;
 }
+.fc-health:hover { background: rgb(var(--v-theme-on-surface) / 0.12); }
 .fc-nav-right {
   flex: 1 1 0;
   min-width: 0;

frontend/src/router.js

@@ -1,5 +1,6 @@
 import { createRouter, createWebHistory, createMemoryHistory } from 'vue-router'
 import SettingsView from './views/SettingsView.vue'
+import SystemView from './views/SystemView.vue'
 import GalleryView from './views/GalleryView.vue'
 import ShowcaseView from './views/ShowcaseView.vue'
 import ExploreView from './views/ExploreView.vue'
@@ -45,6 +46,12 @@ const routes = [
 
   // Settings — config, pinned to the right of the nav (TopNav special-cases it).
   { path: '/settings', name: 'settings', component: SettingsView, meta: { title: 'Settings', stickyChrome: true } },
+  // Deliberately NO meta.title: TopNav builds its nav row from routes that
+  // have one, and this is reached from the health indicator beside the
+  // brand — the place someone already looks when they suspect something is
+  // wrong. A sixth top-level tab for a page you visit twice a year would
+  // cost more attention than it returns.
+  { path: '/system', name: 'system', component: SystemView },
 
   // The old standalone paths now redirect into the Browse hub, preserving any
   // deep-link query (e.g. /posts?post_id=N → /browse?tab=posts&post_id=N). The

frontend/src/stores/system.js

@@ -4,7 +4,12 @@ import { useApi } from '../composables/useApi.js'
 
 export const useSystemStore = defineStore('system', () => {
   const api = useApi()
-  const healthy = ref(null)  // null=unknown, true=ok, false=down
+  // NOT what the nav dot reads any more (milestone 365): that is the
+  // whole-stack verdict in systemHealth.js. /api/health only proves the web
+  // container is serving, which is why a green dot here sat happily beside a
+  // dead worker. refreshHealth() is still called — it is also how build/version
+  // info arrives — so this stays as its by-product rather than its purpose.
+  const healthy = ref(null)
   // What the instance says it is. Since milestone 318 stopped publishing
   // version image tags, this is the only answer to "which build is this?" —
   // there is no registry name left to check it against.

frontend/src/stores/systemHealth.js

@@ -0,0 +1,49 @@
+import { defineStore } from 'pinia'
+import { computed, ref } from 'vue'
+
+import { useApi } from '../composables/useApi.js'
+
+// Whole-stack health: is every part of FabledCurator running (milestone 365)?
+//
+// Distinct from `system.js`, which polls /api/health — a no-DB liveness check
+// that only proves the web container is serving. That endpoint answers "can I
+// reach the API"; this one answers "is anything broken", which is the question
+// a green dot beside the brand was already being read as answering.
+//
+// Also distinct from `systemActivity.js`, which is about what the pipeline is
+// DOING — queue depths, running tasks, failures. Running and alive are
+// different questions and they fail independently: a perfectly idle stack with
+// a dead worker looks identical to a healthy one on the activity surfaces.
+export const useSystemHealthStore = defineStore('systemHealth', () => {
+  const api = useApi()
+
+  const overall = ref(null)      // null until the first answer: unknown ≠ ok
+  const parts = ref([])
+  const checkedAt = ref(null)
+  const thresholds = ref(null)   // server-owned, so the UI keeps no second copy
+  const lastError = ref(null)
+
+  async function refresh() {
+    try {
+      const body = await api.get('/api/system/health')
+      overall.value = body.overall
+      parts.value = body.parts || []
+      checkedAt.value = body.checked_at
+      thresholds.value = body.thresholds || null
+      lastError.value = null
+    } catch (e) {
+      // The endpoint is built never to fail because a dependency failed, so a
+      // throw here means the API itself is unreachable — which is its own kind
+      // of unhealthy and must not be shown as "ok".
+      lastError.value = e.message
+      overall.value = 'unknown'
+    }
+    return overall.value
+  }
+
+  // The parts worth naming in a tooltip — everything that is not ok, worst
+  // first. The endpoint already sorts that way.
+  const problems = computed(() => parts.value.filter(p => p.state !== 'ok'))
+
+  return { overall, parts, checkedAt, thresholds, lastError, problems, refresh }
+})

frontend/src/views/SystemView.vue

@@ -0,0 +1,129 @@
+<template>
+  <v-container class="py-6" style="max-width: 900px">
+    <div class="d-flex align-center mb-1">
+      <h1 class="text-h5">System</h1>
+      <v-spacer />
+      <span class="fc-sys__checked">
+        {{ store.checkedAt ? `checked ${formatRelative(store.checkedAt)}` : 'checking…' }}
+      </span>
+    </div>
+
+    <p class="fc-sys__lede text-body-2 mb-5">
+      Every moving part of FabledCurator and whether it is still checking in.
+      Parts are learned as they appear, so anything that has run at least once
+      stays listed — that is what lets a stopped one be noticed rather than
+      simply vanishing.
+    </p>
+
+    <v-alert
+      v-if="store.lastError" type="error" variant="tonal" density="compact" class="mb-4"
+    >
+      Could not reach FabledCurator: {{ store.lastError }}
+    </v-alert>
+
+    <v-card v-else variant="flat" class="fc-sys__card">
+      <div v-if="!store.parts.length" class="pa-6 text-center fc-sys__muted">
+        Still gathering — this fills in on the first check.
+      </div>
+
+      <div
+        v-for="part in store.parts" :key="part.key"
+        class="fc-sys__row" :class="`fc-sys__row--${part.state}`"
+      >
+        <span class="fc-sys__dot" :class="`fc-sys__dot--${part.state}`" />
+
+        <div class="fc-sys__body">
+          <div class="fc-sys__name">
+            {{ part.name }}
+            <span class="fc-sys__kind">{{ kindLabel(part.kind) }}</span>
+          </div>
+          <!-- The sentence, not just a chip. At the moment someone is deciding
+               whether to go and open Portainer, "has not checked in for 6 min"
+               is the thing that answers them. -->
+          <div class="fc-sys__detail">{{ part.detail }}</div>
+        </div>
+
+        <div class="fc-sys__meta">
+          <div v-if="part.last_seen_at" :title="part.last_seen_at">
+            seen {{ formatRelative(part.last_seen_at) }}
+          </div>
+          <div v-if="part.latency_ms != null">{{ part.latency_ms }} ms</div>
+          <div v-if="part.queues?.length" class="fc-sys__queues">{{ part.queues.join(', ') }}</div>
+        </div>
+      </div>
+    </v-card>
+
+    <p v-if="store.thresholds" class="fc-sys__foot text-caption mt-4">
+      A part is called stale after
+      {{ Math.round(store.thresholds.stale_after_seconds / 60) }} min without a
+      check-in and treated as stopped after
+      {{ Math.round(store.thresholds.down_after_seconds / 60) }} min. The window
+      is deliberately wide: a rolling deploy briefly runs two of a service and
+      then neither, and an indicator that reddened on every update would stop
+      being read.
+    </p>
+  </v-container>
+</template>
+
+<script setup>
+import { onMounted, onUnmounted } from 'vue'
+
+import { useSystemHealthStore } from '../stores/systemHealth.js'
+import { formatRelative } from '../utils/date.js'
+
+const store = useSystemHealthStore()
+
+// Slower than the pipeline chip's 8s: liveness changes on the scale of
+// container restarts, not task starts, and this page is open while someone
+// watches it.
+const POLL_MS = 10_000
+let timer = null
+
+function kindLabel(kind) {
+  if (kind === 'celery') return 'background worker'
+  if (kind === 'agent') return 'GPU agent'
+  if (kind === 'datastore') return 'datastore'
+  return kind
+}
+
+onMounted(() => {
+  store.refresh()
+  timer = setInterval(() => { if (!document.hidden) store.refresh() }, POLL_MS)
+})
+onUnmounted(() => { if (timer) clearInterval(timer) })
+</script>
+
+<style scoped>
+.fc-sys__lede, .fc-sys__muted, .fc-sys__checked, .fc-sys__foot {
+  color: rgb(var(--v-theme-on-surface) / 0.66);
+}
+.fc-sys__checked { font-size: 0.78rem; }
+.fc-sys__card { background: rgb(var(--v-theme-on-surface) / 0.04); }
+
+.fc-sys__row {
+  display: flex; align-items: center; gap: 12px;
+  padding: 12px 16px;
+  border-bottom: 1px solid rgb(var(--v-theme-on-surface) / 0.08);
+}
+.fc-sys__row:last-child { border-bottom: 0; }
+
+.fc-sys__dot { width: 9px; height: 9px; border-radius: 50%; flex: 0 0 auto; }
+.fc-sys__dot--ok      { background: rgb(var(--v-theme-success)); }
+.fc-sys__dot--stale   { background: rgb(var(--v-theme-warning)); }
+.fc-sys__dot--down    { background: rgb(var(--v-theme-error)); }
+.fc-sys__dot--unknown { background: rgb(var(--v-theme-on-surface) / 0.35); }
+
+.fc-sys__body { min-width: 0; flex: 1 1 auto; }
+.fc-sys__name { font-weight: 600; }
+.fc-sys__kind {
+  margin-left: 8px; font-weight: 400; font-size: 0.72rem; text-transform: uppercase;
+  letter-spacing: 0.04em; color: rgb(var(--v-theme-on-surface) / 0.5);
+}
+.fc-sys__detail { font-size: 0.82rem; color: rgb(var(--v-theme-on-surface) / 0.72); }
+
+.fc-sys__meta {
+  text-align: right; font-size: 0.75rem; flex: 0 0 auto;
+  font-variant-numeric: tabular-nums; color: rgb(var(--v-theme-on-surface) / 0.6);
+}
+.fc-sys__queues { opacity: 0.75; }
+</style>

scripts/artifacts.sh

@@ -90,7 +90,7 @@ DERIVER='scripts/artifacts.sh'
 
 
 usage() {
-  echo "usage: artifacts.sh {paths|revision|version} {web|ml|agent|extension}" >&2
+  echo "usage: artifacts.sh {paths|revision|version|epoch} {web|ml|agent|extension}" >&2
   exit 2
 }
 
@@ -149,6 +149,26 @@ cmd_revision() {
   echo "$(newest "$1")" | cut -d' ' -f2 | cut -c1-12
 }
 
+# The BUILD CLOCK: the same commit's unix timestamp, for SOURCE_DATE_EPOCH.
+#
+# buildkit stamps the image config's `created` field and every history entry
+# with the wall clock of the build unless this is set, so two builds of
+# identical source produce different config blobs and therefore different
+# manifest digests. That is #3265: the weekly refresh republished all three
+# `:latest` tags on 2026-08-30 with every content step CACHED and the bases
+# resolved to unchanged digests — nothing was different, and the digest moved
+# anyway. A digest that changes on a calendar cannot also mean "the content
+# changed", which is the only thing anyone wants it for.
+#
+# It is the same commit `revision` and `version` name — deliberately, and this
+# is the point of routing it through `newest()` rather than taking git's word
+# separately. Three values derived from three lookups can disagree; three
+# views of one lookup cannot. Note #3127 §2 is the record of what a second
+# clock costs.
+cmd_epoch() {
+  echo "$(newest "$1")" | cut -d' ' -f1
+}
+
 # The VERSION: `YYYY.MM.DD.HHMM`, zero-padded, UTC. One shape across the whole
 # family (note #3127 §1, rule 148) — the number an instance reports about
 # itself, and, with a `v` in front, the release tag naming the same build.
@@ -197,5 +217,6 @@ case "$1" in
   paths)    cmd_paths "$2" ;;
   revision) cmd_revision "$2" ;;
   version)  cmd_version "$2" ;;
+  epoch)    cmd_epoch "$2" ;;
   *)        usage ;;
 esac

scripts/release_notes.py

@@ -33,6 +33,32 @@ history. Ancestry is immune to the shape change, and it is also the more honest
 question: "what is in this that was not in the last one" IS a reachability
 question.
 
+Ancestry alone is not enough, though, and milestone 328 is where that showed.
+The 28 `v26.*` tags are still in the repo — the operator kept them as history
+when their releases were deleted — so `--match v*` walks straight back to
+`v26.06.04.0` and reports 533 commits. That span is not a changelog: nobody has
+run `v26.06.04.0`, its release page no longer exists to compare against, and
+the 200 lines that survive truncation are precisely the internal build-out that
+milestone 328 exists to stop shipping. So the match is `v[0-9][0-9][0-9][0-9].*`
+— rule 148's four-digit-year shape — which is exactly the set of tags that name
+a release a reader could have been running. A pre-convention tag is history,
+not a predecessor.
+
+## The first release has no changelog, and should not pretend to
+
+Once the match is narrowed, the first rule-148 tag reaches no predecessor at
+all, and the old fallback — diff against the whole history — is worse than the
+problem it replaced. The honest content for a release nobody has a previous
+version of is what the thing IS.
+
+So a release with no reachable predecessor renders the product overview instead
+of a commit list. It is read out of README.md between `<!-- overview:start -->`
+and `<!-- overview:end -->` rather than written here, for the same reason the
+changelog is derived: two hand-maintained descriptions of one product drift,
+and nothing ever catches it. The release page and the repo front page are one
+source. Every later release goes back to being a changelog, which is what §5 of
+note #3127 says a release is for.
+
 ## Re-runs update, they do not fall through
 
 Note #3127 §6.7: a publisher that POSTs and recovers the id from a `409` never
@@ -91,18 +117,45 @@ def git_ok(*args: str) -> str | None:
 
 
 def previous_tag(ref: str, tag: str | None) -> str | None:
-    """The most recent `v*` tag reachable from `ref`, excluding `tag` itself.
+    """The most recent rule-148 tag reachable from `ref`, excluding `tag` itself.
 
     `--exclude` rather than `<ref>^` so this is the same call whether or not
     `ref` is the tag being released — and so it does not blow up on a root
     commit that has no parent to walk to.
+
+    The glob deliberately does NOT match the old `v26.*` tags. They are kept as
+    history and their releases are gone, so naming one as the predecessor emits
+    a span nobody can look up. See the module docstring.
     """
-    args = ["describe", "--tags", "--abbrev=0", "--match", "v*"]
+    args = ["describe", "--tags", "--abbrev=0", "--match", "v[0-9][0-9][0-9][0-9].*"]
     if tag:
         args += ["--exclude", tag]
     return git_ok(*args, ref)
 
 
+def product_overview() -> str | None:
+    """The product description, lifted verbatim from README.md.
+
+    Returns None if the markers are absent or empty — a missing overview is
+    reported as a note and the release still publishes, on the same reasoning
+    as cross_checks(): the release is the useful object even when one part of
+    the derivation could not run.
+    """
+    root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+    try:
+        with open(os.path.join(root, "README.md"), encoding="utf-8") as fh:
+            readme = fh.read()
+    except OSError:
+        return None
+    match = re.search(
+        r"<!--\s*overview:start\s*-->(.*?)<!--\s*overview:end\s*-->",
+        readme, re.S,
+    )
+    if not match:
+        return None
+    return match.group(1).strip() or None
+
+
 def commits(previous: str | None, ref: str) -> list[str]:
     """The subjects between the previous release and this one.
 
@@ -125,7 +178,10 @@ def truncate(log: list[str]) -> tuple[list[str], str | None]:
     )
 
 
-def render(tag: str, sha: str, previous: str | None, log: list[str], notes: list[str]) -> str:
+def render(
+    tag: str, sha: str, previous: str | None, log: list[str], notes: list[str],
+    overview: str | None,
+) -> str:
     short = sha[:7]
     parts = []
 
@@ -135,6 +191,25 @@ def render(tag: str, sha: str, previous: str | None, log: list[str], notes: list
         # is the failure this whole milestone is about.
         parts.append("\n".join(f"> **Note:** {n}" for n in notes))
 
+    # No predecessor means nobody reading this has run an earlier one, so the
+    # release describes the product rather than a diff. The overview is
+    # README.md's own words — see the module docstring on why it is not
+    # written here.
+    if previous is None and overview:
+        parts.append(overview)
+        parts.append(
+            "## Installing\n\n"
+            "```\ncurl -O https://git.fabledsword.com/bvandeusen/FabledCurator/raw/"
+            f"tag/{tag}/docker-compose.yml\ncurl -O https://git.fabledsword.com/"
+            f"bvandeusen/FabledCurator/raw/tag/{tag}/.env.example\n"
+            "mv .env.example .env   # then set SECRET_KEY, DB_PASSWORD\n"
+            "docker compose -f docker-compose.yml up -d\n```\n\n"
+            "**Read \"Before you expose it\" in the README first.** FabledCurator "
+            "has no login, and it stores live platform session cookies for "
+            "accounts that usually have a payment method attached. Bind it to a "
+            "network you trust."
+        )
+
     parts.append(
         f"Built from `{short}`. The rollback unit is the immutable `:c-` tag "
         f"(rule 145) — these three move together:\n\n```\n"
@@ -142,7 +217,19 @@ def render(tag: str, sha: str, previous: str | None, log: list[str], notes: list
         + "\n```"
     )
 
-    heading = f"## Changes since {previous}" if previous else "## Changes"
+    if previous is None:
+        # Deliberately NOT a commit list. The alternative is the whole history
+        # truncated to MAX_COMMITS, which is 200 lines of internal build-out
+        # presented to someone who has never seen this project.
+        parts.append(
+            "---\n\n_First release under rule 148's `vYYYY.MM.DD.HHMM` shape, so "
+            "there is no predecessor to diff against and no changelog to derive. "
+            "The description above is README.md's, quoted at publish time. Later "
+            "releases carry the commits since the previous one._"
+        )
+        return "\n\n".join(parts)
+
+    heading = f"## Changes since {previous}"
     if log:
         parts.append(heading + "\n\n" + "\n".join(f"- {line}" for line in log))
     else:
@@ -152,10 +239,9 @@ def render(tag: str, sha: str, previous: str | None, log: list[str], notes: list
             "names the same source under a new name._"
         )
 
-    span = f"{previous}..{tag}" if previous else tag
     parts.append(
-        f"---\n\n_Derived at publish time from `git log --no-merges {span}`. "
+        f"---\n\n_Derived at publish time from "
-        f"Nothing here is hand-maintained._"
+        f"`git log --no-merges {previous}..{tag}`. Nothing here is hand-maintained._"
     )
     return "\n\n".join(parts)
 
@@ -289,13 +375,31 @@ def main() -> None:
     for note in notes:
         print(f"release: NOTE {note}")
 
-    log = commits(previous, ref)
+    # A first release renders the overview instead of a changelog, so the
-    print(f"release: {len(log)} non-merge commits in the span")
+    # commit walk is skipped entirely rather than computed and discarded —
-    log, overflow = truncate(log)
+    # `commits(None, ref)` is the whole history and there is no reason to ask
-    if overflow:
+    # for it.
-        print(f"release: NOTE {overflow}")
+    overview = None
-        notes.append(overflow)
+    log: list[str] = []
-    body = render(tag or ref, sha, previous, log, notes)
+    if previous is None:
+        overview = product_overview()
+        if overview is None:
+            note = (
+                "No `<!-- overview:start -->` block found in README.md, so this "
+                "first release has no product description. Published anyway; add "
+                "the markers and re-run the workflow to fill it in."
+            )
+            print(f"release: NOTE {note}")
+            notes.append(note)
+        print("release: no rule-148 predecessor — rendering the product overview")
+    else:
+        log = commits(previous, ref)
+        print(f"release: {len(log)} non-merge commits in the span")
+        log, overflow = truncate(log)
+        if overflow:
+            print(f"release: NOTE {overflow}")
+            notes.append(overflow)
+    body = render(tag or ref, sha, previous, log, notes, overview)
 
     if args.dry_run or not tag:
         print("--- body ---")

scripts/smoke_image.py

@@ -0,0 +1,155 @@
+"""Prove a freshly built image can still do the things its OS packages provide.
+
+Run INSIDE the image, not against the source tree. That distinction is the
+entire reason this file exists.
+
+`ci.yml`'s lanes run on `ci-python:3.14` and install `requirements.txt`. A base
+refresh changes neither, so all five lanes stay green through a base bump that
+breaks the product. What a refresh actually re-resolves is this, from the
+Dockerfile:
+
+    RUN apt-get update && apt-get install -y --no-install-recommends \
+            ffmpeg unar libpq5 postgresql-client zstd megatools \
+            libjpeg62-turbo libwebp7 libpng16-16 ca-certificates
+
+Unpinned, every build. Nothing else in this repo looks at it.
+
+So the checks below run the APPLICATION'S OWN code — `Thumbnailer`, which needs
+no database and no app context — against whatever Pillow and ffmpeg have
+become. `ffmpeg -version` exiting 0 would pass while a codec removal or an
+soname bump broke every thumbnail in the library; producing a thumbnail would
+not.
+
+Every failure names the package it implicates. This fires on a Sunday,
+unattended, about a change nobody made deliberately — "assertion failed" a week
+later teaches nobody anything.
+
+Usage:  docker run --rm -i <image> shell -c 'python3 -' < scripts/smoke_image.py
+"""
+
+from __future__ import annotations
+
+import shutil
+import subprocess
+import tempfile
+from pathlib import Path
+
+try:
+    from PIL import Image
+
+    from backend.app.services.thumbnailer import Thumbnailer
+except Exception as exc:  # noqa: BLE001 — a smoke test reports, it never raises
+    print(f"smoke: FAILED — could not import the thumbnail path at all: {exc}")
+    print("  Implicates Pillow or its shared libraries (libjpeg62-turbo,")
+    print("  libpng16-16, libwebp7), or the python base image itself.")
+    raise SystemExit(1) from exc
+
+
+# Binary → what stops working without it. Listed individually because
+# `--no-install-recommends` means any one of them can vanish on its own when a
+# dependency chain higher up changes.
+REQUIRED_BINARIES = {
+    "ffmpeg": "video thumbnails and transcoding (Dockerfile: ffmpeg)",
+    "unar": "archive import — cbz/zip/rar members (Dockerfile: unar)",
+    "pg_dump": "database backup (Dockerfile: postgresql-client)",
+    "zstd": "backup compression, pg_dump | tar --zstd (Dockerfile: zstd)",
+    "megatools": "mega.nz public-link downloads, #830 (Dockerfile: megatools)",
+}
+
+
+def check_jpeg(thumbs: Thumbnailer, src: Path) -> None:
+    path = src / "flat.jpg"
+    Image.new("RGB", (900, 400), (30, 90, 160)).save(path, "JPEG")
+    result = thumbs.generate_image_thumbnail(path, "a" * 64)
+    assert result.mime == "image/jpeg", f"mime was {result.mime}"
+    assert result.path.stat().st_size > 0, "no bytes written"
+    # Re-open it. A file that writes but cannot be read back is the shape a
+    # half-broken codec produces, and size alone would not catch it.
+    with Image.open(result.path) as im:
+        im.load()
+
+
+def check_png_alpha(thumbs: Thumbnailer, src: Path) -> None:
+    path = src / "alpha.png"
+    Image.new("RGBA", (400, 900), (200, 40, 40, 128)).save(path, "PNG")
+    result = thumbs.generate_image_thumbnail(path, "b" * 64)
+    assert result.mime == "image/png", f"mime was {result.mime}"
+    with Image.open(result.path) as im:
+        im.load()
+        assert im.mode in ("RGBA", "LA", "P"), f"alpha lost, mode={im.mode}"
+
+
+def check_webp(thumbs: Thumbnailer, src: Path) -> None:
+    path = src / "sample.webp"
+    Image.new("RGB", (500, 500), (10, 140, 70)).save(path, "WEBP")
+    result = thumbs.generate_image_thumbnail(path, "c" * 64)
+    assert result.path.stat().st_size > 0, "no bytes written"
+
+
+def check_video(thumbs: Thumbnailer, src: Path) -> None:
+    # Synthesised rather than committed as a fixture: a checked-in video is a
+    # binary blob nobody can review, and lavfi ships with every ffmpeg build.
+    #
+    # 3 seconds, not 2. The seek lands at max(1.0, duration * 0.05) = 1.0s, and
+    # a clip barely longer than its own seek is how #1231 produced zero frames.
+    # This check exists to exercise ffmpeg, not to re-litigate that edge.
+    clip = src / "clip.mp4"
+    subprocess.run(
+        ["ffmpeg", "-nostdin", "-f", "lavfi", "-i", "testsrc=size=640x360:rate=10",
+         "-t", "3", "-pix_fmt", "yuv420p", "-y", str(clip)],
+        check=True, capture_output=True, timeout=120,
+    )
+    result = thumbs.generate_video_thumbnail(clip, "d" * 64, duration_seconds=3.0)
+    assert result.path.stat().st_size > 0, "no bytes written"
+    with Image.open(result.path) as im:
+        im.load()
+
+
+CHECKS = (
+    ("JPEG thumbnail", "libjpeg62-turbo / Pillow", check_jpeg),
+    ("PNG thumbnail (alpha)", "libpng16-16 / Pillow", check_png_alpha),
+    ("WebP decode", "libwebp7 / Pillow", check_webp),
+    ("video thumbnail", "ffmpeg", check_video),
+)
+
+
+def main() -> int:
+    failures: list[str] = []
+
+    print("smoke: binaries the apt layer provides")
+    for binary, purpose in REQUIRED_BINARIES.items():
+        if shutil.which(binary) is None:
+            print(f"  FAIL  {binary}: not on PATH")
+            failures.append(f"{binary} — {purpose}")
+        else:
+            print(f"  ok    {binary}")
+
+    print("smoke: the application's own thumbnail path, against this image's libraries")
+    with tempfile.TemporaryDirectory() as tmp:
+        root = Path(tmp)
+        src = root / "src"
+        src.mkdir()
+        thumbs = Thumbnailer(root)
+        for name, implicates, fn in CHECKS:
+            try:
+                fn(thumbs, src)
+                print(f"  ok    {name}")
+            except Exception as exc:  # noqa: BLE001 — report every check, then fail once
+                print(f"  FAIL  {name}: {exc}")
+                failures.append(f"{name} — {implicates}")
+
+    if failures:
+        print(f"\nsmoke: FAILED — {len(failures)} check(s)")
+        for failure in failures:
+            print(f"  - {failure}")
+        print("\nThis image was built against freshly resolved base layers. The")
+        print("named packages are where to look: compare this build's apt versions")
+        print("against the previous :latest before assuming the app changed.")
+        return 1
+
+    print("\nsmoke: all checks passed")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

tests/test_artifact_identity.py

@@ -234,3 +234,56 @@ def test_version_and_revision_describe_the_same_commit(artifact):
             f"in AMO_UNPADDED may differ here."
         )
     assert sha.startswith(revision(artifact))
+
+
+@pytest.mark.parametrize("artifact", ARTIFACTS)
+def test_epoch_is_the_same_commit_the_version_names(artifact):
+    """The build clock and the version must be one lookup, not two.
+
+    `epoch` feeds SOURCE_DATE_EPOCH, which decides the image config's bytes and
+    therefore the manifest digest; `version` is what the instance reports about
+    itself. If they could name different commits, an image would be stamped
+    reproducibly against one commit while claiming to be another — and both
+    values would look perfectly well-formed, exactly like the divergence the
+    test above guards.
+
+    They cannot, because `cmd_epoch` and `cmd_version` are two fields of one
+    `newest()` result. This pins that they stay that way: a future refactor
+    that gave epoch its own `git log` would pass every other test here.
+    """
+    epoch = artifacts("epoch", artifact).strip()
+    assert epoch.isdigit(), f"{artifact} epoch is {epoch!r}, not a unix timestamp"
+
+    sha = newest_by_commit_time(artifact)
+    committed = subprocess.run(
+        ["git", "show", "-s", "--format=%ct", sha],
+        capture_output=True, text=True, check=True, cwd=ROOT,
+    ).stdout.strip()
+    assert epoch == committed, (
+        f"{artifact} derives epoch {epoch}, but its newest shipped commit "
+        f"{sha[:12]} was committed at {committed}. SOURCE_DATE_EPOCH would "
+        f"pin the image config to a commit the version does not name."
+    )
+
+    # And the two renderings must agree, which is the property that actually
+    # matters at build time: same commit in, same digest and same reported
+    # version out.
+    rendered = subprocess.run(
+        ["git", "show", "-s", "--format=%cd", "--date=format-local:%Y.%m.%d.%H%M", sha],
+        capture_output=True, text=True, check=True, cwd=ROOT,
+        env={"TZ": "UTC", "PATH": os.environ.get("PATH", "")},
+    ).stdout.strip()
+    assert segments(artifacts("version", artifact).strip()) == segments(rendered)
+
+
+def test_epoch_is_stable_across_calls():
+    """SOURCE_DATE_EPOCH's entire job is to be the same on the next build.
+
+    A value that moved between two invocations on one unchanged checkout would
+    reintroduce #3265 through the very mechanism meant to close it, and the
+    symptom would be indistinguishable: a digest that changes for no reason.
+    """
+    for artifact in ARTIFACTS:
+        first = artifacts("epoch", artifact).strip()
+        second = artifacts("epoch", artifact).strip()
+        assert first == second, f"{artifact} epoch moved: {first} then {second}"

tests/test_release_notes.py

@@ -24,12 +24,36 @@ SCRIPT = ROOT / "scripts" / "release_notes.py"
 
 
 def notes(*args: str, cwd: Path | None = None) -> str:
+    """Run the script the way release.yml does.
+
+    A synthetic repo runs its OWN copy of the script, because the overview is
+    read relative to `__file__` rather than to the cwd — which is right in
+    production (release.yml checks out the tag, so the script IS the tagged
+    tree's copy) and would otherwise make every synthetic repo silently quote
+    FabledCurator's real README.
+    """
+    root = cwd or ROOT
+    script = root / "scripts" / "release_notes.py"
     return subprocess.run(
-        ["python3", str(SCRIPT), "--dry-run", *args],
+        ["python3", str(script if script.exists() else SCRIPT), "--dry-run", *args],
-        capture_output=True, text=True, check=True, cwd=cwd or ROOT,
+        capture_output=True, text=True, check=True, cwd=root,
     ).stdout
 
 
+OVERVIEW_TEXT = "A synthetic product, described once."
+
+
+def install_script(repo: Path, *, overview: bool = True) -> None:
+    """Give a synthetic repo the script and a README to quote."""
+    (repo / "scripts").mkdir(exist_ok=True)
+    (repo / "scripts" / "release_notes.py").write_text(SCRIPT.read_text())
+    (repo / "scripts" / "artifacts.sh").write_text("#!/bin/sh\nexit 1\n")
+    readme = "# Synthetic\n\n"
+    if overview:
+        readme += f"<!-- overview:start -->\n{OVERVIEW_TEXT}\n<!-- overview:end -->\n"
+    (repo / "README.md").write_text(readme)
+
+
 def body_of(out: str) -> str:
     assert "--- body ---" in out, f"no body was rendered:\n{out}"
     return out.split("--- body ---", 1)[1]
@@ -56,9 +80,10 @@ def shaped_history(tmp_path: Path) -> Path:
     repo = tmp_path / "shaped"
     repo.mkdir()
     git(repo, "init", "-q", "-b", "main")
+    install_script(repo)
     for i, tag in enumerate(("v26.06.04.0", "v2026.08.28.2208", "v2026.08.29.1000")):
         (repo / "f.txt").write_text(f"{i}\n")
-        git(repo, "add", "f.txt")
+        git(repo, "add", "-A")
         git(repo, "commit", "-q", "-m", f"work landing in {tag}")
         git(repo, "tag", tag)
     # One more commit and a merge, so the merge-exclusion test has something to
@@ -109,12 +134,57 @@ def test_merges_are_excluded_so_the_list_is_the_work(shaped_history):
     assert "Merge pull request #999" not in body
 
 
-def test_the_first_release_still_renders_with_nothing_behind_it(shaped_history):
+def test_a_pre_convention_tag_is_history_not_a_predecessor(shaped_history):
-    """No previous tag is reachable from the oldest one. That is a real state,
+    """The defect milestone 328 hit, and the reason the match glob narrowed.
-    not an error, and it must not take the release down with it."""
+
-    out = notes("v26.06.04.0", cwd=shaped_history)
+    The 28 `v26.*` tags are kept as history while their releases were deleted.
+    Ancestry alone happily names `v26.06.04.0` as the predecessor of the first
+    rule-148 tag — and then the body offers "changes since" a release that no
+    longer exists, over a span (533 commits in the real repo) that is the
+    internal build-out this milestone exists to stop publishing.
+
+    Reachable is not the same as comparable. Only a `vYYYY.` tag names a
+    release a reader could have been running.
+    """
+    out = notes("v2026.08.28.2208", cwd=shaped_history)
     assert "previous=<none>" in out
-    assert "## Changes" in body_of(out)
+    assert "v26.06.04.0" not in body_of(out)
+
+
+def test_the_first_release_describes_the_product_instead_of_diffing(shaped_history):
+    """No predecessor means nobody reading has run an earlier version, so a
+    changelog has no referent. The alternative the script used to take — diff
+    against the whole history, truncated — puts 200 lines of internal build-out
+    in front of someone meeting the project for the first time."""
+    body = body_of(notes("v2026.08.28.2208", cwd=shaped_history))
+    assert OVERVIEW_TEXT in body
+    assert "## Changes" not in body
+    assert not [ln for ln in body.split("\n") if ln.startswith("- work landing")]
+
+
+def test_the_overview_is_readmes_words_not_a_second_copy(shaped_history):
+    """Two hand-maintained descriptions of one product drift and nothing
+    catches it. The release page quotes README.md so there is one source."""
+    readme = (shaped_history / "README.md").read_text()
+    assert OVERVIEW_TEXT in readme
+    assert OVERVIEW_TEXT in body_of(notes("v2026.08.28.2208", cwd=shaped_history))
+
+
+def test_a_missing_overview_block_is_reported_and_still_publishes(tmp_path):
+    """Same reasoning as cross_checks(): the release is the useful object even
+    when part of the derivation could not run. Say what is missing, publish
+    anyway — do not leave the operator with a tag and no release."""
+    repo = tmp_path / "no-markers"
+    repo.mkdir()
+    git(repo, "init", "-q", "-b", "main")
+    install_script(repo, overview=False)
+    git(repo, "add", "-A")
+    git(repo, "commit", "-q", "-m", "first")
+    git(repo, "tag", "v2026.09.01.1200")
+
+    out = notes("v2026.09.01.1200", cwd=repo)
+    assert "No `<!-- overview:start -->` block found" in out
+    assert "No `<!-- overview:start -->` block found" in body_of(out)
 
 
 def test_a_non_tag_ref_renders_but_refuses_to_claim_it_published():
@@ -135,15 +205,29 @@ def test_the_rollback_refs_name_all_three_images():
         assert f"bvandeusen/{image}:c-" in body, f"{image} missing from the rollback refs"
 
 
-def test_an_unbounded_span_is_truncated_and_says_so():
+def test_a_long_span_between_two_releases_is_truncated_and_says_so(tmp_path):
-    """With no reachable previous tag the span is the whole history. Emitting
+    """The cap is still reachable, just not by the route it used to be.
-    eleven hundred lines would bury the one line explaining why there are
+
-    eleven hundred of them, so the cap is part of the message, not a silent
+    It no longer fires on "no predecessor" — that renders the overview now.
-    slice."""
+    What it still guards is two real releases far enough apart that the list
-    out = notes("HEAD")
+    stops being something anyone reads, which is the ordinary case for a
-    if "previous=<none>" not in out:
+    project that cuts a bookmark twice a year. The cap is part of the message,
-        pytest.skip("a previous tag is reachable from HEAD in this checkout")
+    not a silent slice.
+    """
+    repo = tmp_path / "long"
+    repo.mkdir()
+    git(repo, "init", "-q", "-b", "main")
+    install_script(repo)
+    git(repo, "add", "-A")
+    git(repo, "commit", "-q", "-m", "scaffold")
+    git(repo, "tag", "v2026.01.01.0000")
+    for i in range(205):
+        git(repo, "commit", "-q", "--allow-empty", "-m", f"fix: change {i}")
+    git(repo, "tag", "v2026.07.01.0000")
+
+    out = notes("v2026.07.01.0000", cwd=repo)
+    assert "previous=v2026.01.01.0000" in out
     body = body_of(out)
     listed = [ln for ln in body.split("\n") if ln.startswith("- ")]
-    assert len(listed) <= 200
+    assert len(listed) == 200
     assert "more than a changelog is for" in body