fix: extension updates install the new build — no 12h-cached "latest" XPI, and the popup's Update opens FC's install page

Operator, 2026-09-25: "the extension update trigger from inside the
extension doesn't work and the manual update seems to not move it to the
most recent version or at least mark it the most recent."

- fabledcurator-latest.xpi was served with Quart's default
  `public, max-age=43200`: one URL whose bytes change every release, so a
  browser that had fetched it reinstalled the previous build for 12 hours
  (measured on the instance). It is now `no-cache` (the ETag keeps an
  unchanged file a 304); versioned XPIs are `immutable`.
- The web Settings card installs/downloads the VERSIONED xpi_url, which can
  only ever be that build's bytes.
- The popup's Update button did tabs.create() on the .xpi, which Firefox
  refuses (NS_ERROR_FAILURE on a 200: it only installs from a user click on
  a web page). It now opens FC's install card (/subscriptions?tab=settings).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LVjrnpQjRgHdvq95rASoiR

.forgejo/workflows/baseline.yml

@@ -79,7 +79,7 @@ jobs:
       - name: Resolve the Postgres service and install deps
         run: |
           set -eux
-          # Same service-IP dance as ci.yml's integration job; see the long
+          # Same service-IP dance as build.yml's integration job; see the long
           # comment there for why the job name must stay separator-free.
           PG=$(docker ps --filter "name=compare" --filter "ancestor=pgvector/pgvector:pg16" -q | head -n1)
           test -n "$PG"
@@ -89,7 +89,7 @@ jobs:
           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.
+          # as build.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

.forgejo/workflows/ci.yml

@@ -1,294 +0,0 @@
-name: CI
-
-# CI lanes per FabledRulebook/forgejo.md "CI philosophy":
-#   - lint: ruff only, no dep install — fast-fail for the common lint bounce.
-#   - extension-version: the derived version resolves and is a shape AMO takes.
-#   - backend-lint-and-test: `pytest -m "not integration"`, no service containers.
-#   - frontend-build: vitest unit + vite build.
-#   - integration: pgvector + redis service containers; alembic + `pytest -m integration`.
-
-on:
-  push:
-    branches: [dev, main]
-  # Renovate opens PRs from `renovate/*` branches into `dev`. Those branches
-  # never push to dev/main, so the push trigger above gives them NO pre-merge
-  # CI — a bump could only be validated after it was already merged. This
-  # pull_request trigger (base `dev` only) validates Renovate PRs before merge.
-  # It deliberately does NOT fire on dev→main PRs (base `main`), which still
-  # rely on the dev push run — so no duplicate runs. FC has no fork PRs
-  # (single-operator Forgejo repo), so secrets-on-PR is not a concern.
-  pull_request:
-    branches: [dev]
-
-jobs:
-  # Fast-fail lint lane. ruff is pre-installed in the ci-python image, so
-  # this runs with NO dependency install and surfaces the most common bounce
-  # class (lint: I001 / UP037 / ASYNC109 / W293 …) in seconds — instead of
-  # after the backend job's ~30-60s wheel install. ruff is static analysis,
-  # so no DB/secret env is needed.
-  lint:
-    runs-on: python-ci
-    container:
-      image: git.fabledsword.com/bvandeusen/ci-python:3.14
-    steps:
-      - uses: actions/checkout@v4
-      - name: Ruff lint
-        # agent/ included so the GPU-agent is linted before its image is built
-        # (build.yml only `docker build`s it — this is where it gets checked).
-        # scripts/ likewise: release_notes.py runs only on a tag push, so a
-        # syntax or import error there would otherwise surface at the one
-        # moment nobody wants to debug a workflow.
-        run: ruff check backend/ tests/ alembic/ agent/ scripts/
-      - name: Agent syntax check
-        # The agent's runtime deps (torch/transformers/ultralytics) aren't in the
-        # CI image, so we can't import it — but compileall parses every module,
-        # catching syntax errors before the image build.
-        run: python -m compileall -q agent/fc_agent
-
-  # The extension version is DERIVED, not hand-maintained (milestone 271 step
-  # 4): build.yml computes it from the commit TIME of the newest packaged
-  # extension change and stamps it into manifest.json / package.json at build
-  # time. The guard that used to live here — "packaged files changed but nobody
-  # bumped the version" — was therefore checking a fact that had stopped
-  # existing. Worse than useless: it would have failed this lane on every real
-  # extension change, demanding a bump that decides nothing. Retired 2026-08-27
-  # rather than left running beside the new mechanism (rule 22).
-  #
-  # Two things are still worth asserting, and this is the only lane that can:
-  # the extension.yml suite runs on node:24-slim, which is exactly why
-  # version.spec.js sticks to packaging.sh's git-free subcommands.
-  #   1. the derivation actually resolves on this commit
-  #   2. the derived string is one AMO will accept, checked against Mozilla's
-  #      own published grammar rather than a loose "digits and dots"
-  #
-  # The MAJOR.MINOR-agreement check that used to be (2) is gone with milestone
-  # 318 step 8: the committed version no longer seeds anything, so there is no
-  # hand-set part left for the two files to disagree about.
-  #
-  # Deliberately NOT checked here: that the derived value beats what has already
-  # been signed. That guard belongs in build.yml, where it compares against the
-  # real ext-* releases. Comparing against origin/main here would be wrong —
-  # dev legitimately derives a LOWER value whenever main is ahead on the
-  # extension, and a lane that fails for being behind is a lane people learn to
-  # ignore.
-  extension-version:
-    runs-on: python-ci
-    container:
-      image: git.fabledsword.com/bvandeusen/ci-python:3.14
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          # The derivation needs real history: a depth-1 clone sees one commit
-          # and produces a wrong, too-low value RATHER THAN FAILING. Checking
-          # that here is half the point of the lane.
-          fetch-depth: 0
-      - name: Extension version derives cleanly
-        run: |
-          set -eu
-          # busybox sh on the act_runner — no bashisms (family rule).
-          VERSION=$(sh extension/scripts/packaging.sh version)
-          echo "derived: $VERSION"
-
-          # Mozilla's published grammar for AMO, transcribed verbatim from
-          # MDN's manifest.json/version page:
-          #
-          #     ^(0|[1-9][0-9]{0,8})([.](0|[1-9][0-9]{0,8})){0,3}$
-          #
-          # Not the looser `^[0-9]+(\.[0-9]+)*$` this lane used to carry. That
-          # one passes `2026.08.29.0201`, which AMO REJECTS — a segment must be
-          # the single digit 0 or start 1-9 — and it also passes five segments,
-          # where AMO allows four. Both would surface as a failed sign with the
-          # version already burned: AMO 409s on re-signing, so a rejected value
-          # cannot be reclaimed and cannot be reused. This lane is the cheap
-          # place to find out. (#3138, milestone 318 step 8.)
-          if ! echo "$VERSION" | grep -qE '^(0|[1-9][0-9]{0,8})(\.(0|[1-9][0-9]{0,8})){0,3}$'; then
-            echo "ERROR: derived version '$VERSION' is not a version AMO accepts."
-            echo "AMO's grammar: ^(0|[1-9][0-9]{0,8})([.](0|[1-9][0-9]{0,8})){0,3}$"
-            echo "Most likely cause: a zero-padded segment (08, 0201). The rest"
-            echo "of the family pads; the extension must not — see packaging.sh."
-            exit 1
-          fi
-
-          # ...and the shape this project actually derives. AMO would happily
-          # take `1.0.3500147` too, so the grammar check alone would not notice
-          # a regression to the pre-318 shape — which orders BELOW everything
-          # signed since, and is unrecoverable once Firefox has the higher one.
-          if ! echo "$VERSION" | grep -qE '^20[0-9][0-9]\.[0-9]{1,2}\.[0-9]{1,2}\.[0-9]{1,4}$'; then
-            echo "ERROR: derived version '$VERSION' is not YYYY.M.D.HHMM."
-            echo "Rule 148's CalVer is what build.yml signs; the old"
-            echo "1.0.<minutes> shape would order below every ext-2026.* release."
-            exit 1
-          fi
-          echo "OK: derived version $VERSION"
-
-  backend-lint-and-test:
-    runs-on: python-ci
-    container:
-      image: git.fabledsword.com/bvandeusen/ci-python:3.14
-    env:
-      # DB_PASSWORD and SECRET_KEY are required by config.py at import time
-      # even though unit tests don't actually touch the DB or use the secret.
-      DB_PASSWORD: ci_unit_test_placeholder
-      SECRET_KEY: ci_unit_test_placeholder
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          # Full history for tests/test_artifact_identity.py, which derives
-          # each artifact's revision to check the identity scheme. On a
-          # depth-1 clone that derivation either fails or returns the tip sha
-          # — so the lane would go green while asserting nothing, which is
-          # the one outcome worse than a red one.
-          fetch-depth: 0
-
-      # Cache step removed 2026-05-26: act_runner's cache backend has been
-      # broken on this homelab runner since 2026-05-15 (first as request-
-      # timeout warnings, then as hard "Cannot find module .../dist/restore/
-      # index.js" failures that tank the whole job). The cache step targeted
-      # ~/.cache/pip but the install below uses `uv pip install` primarily,
-      # whose own cache lives at ~/.cache/uv — so the cache step's real
-      # benefit was marginal even when working. Cost of removal: ~30s of
-      # wheel downloads per job. Future re-enable: mount ~/.cache/uv as a
-      # docker volume at the runner level (skips actions/cache entirely),
-      # or fix the runner-side cache backend (clear /var/run/act/actions/*,
-      # pin act_runner version, etc.).
-
-      - name: Install Python deps
-        # ruff is pre-installed in the ci-python image (see CI-Runner/CI-python/
-        # Dockerfile's RUFF_VERSION). Per FabledRulebook ci-runners.md, toolchain
-        # versions live on the runner image, not here.
-        # uv: 5-10x faster wheel resolve than pip for cold caches.
-        # Falls back to pip install on uv-missing runners (older images).
-        run: |
-          if command -v uv >/dev/null 2>&1; then
-            uv pip install --system -r requirements.txt pytest pytest-asyncio
-          else
-            pip install -r requirements.txt pytest pytest-asyncio
-          fi
-
-      # Ruff moved to the dedicated fast `lint` job above (fails in seconds,
-      # no dep install). This job is now unit tests only.
-      - name: Pytest (unit only — integration runs in the integration job)
-        run: pytest tests/ -v -m "not integration"
-
-  frontend-build:
-    runs-on: python-ci
-    container:
-      image: git.fabledsword.com/bvandeusen/ci-python:3.14
-    defaults:
-      run:
-        working-directory: frontend
-    steps:
-      - uses: actions/checkout@v4
-      # No package-lock.json is tracked yet (we don't run npm locally per
-      # feedback-no-local-runs). Using `npm install` instead of `npm ci`.
-      # If we want strict lockfile-based reproducibility later, commit a
-      # package-lock.json and flip this back to `npm ci`.
-      - run: npm install --no-audit --no-fund
-      # No type-check step: the frontend is pure JS (no .ts files, no JSDoc),
-      # so a type-checker has nothing to do. The vue-tsc devDep + its `check`
-      # script were dropped 2026-07-11 rather than bumped to v3. If we add
-      # TS/JSDoc later, re-add a tsconfig.json + vue-tsc + a type-check step.
-      - run: npm run test:unit
-      - run: npm run build
-
-  # Single integration job — collapsed from a 3-way shard split on 2026-06-04.
-  # The shards existed to parallelize ~8.5min of integration tests; once the
-  # throwaway Postgres runs with fsync OFF (the durability step below) the whole
-  # suite runs in ~45s, so the split only triplicated the ~2min fixed overhead
-  # (container + `uv pip install` + `alembic upgrade head`) and burned 3 of 6
-  # runner slots for no wall-clock gain. One job now: spin up once, install
-  # once, migrate once, run every integration test.
-  #
-  # The docker-ps filter scopes to THIS job's own Postgres/Redis service
-  # containers by job name. act_runner strips underscores from job names when
-  # labelling containers (`int_api` matched nothing on 2026-05-25), so the name
-  # stays separator-free (`integration`). The step prints `docker ps -a` first
-  # so a future naming-convention shift surfaces in the log without a
-  # guess-and-push cycle.
-  #
-  # Pre-baking requirements.txt into ci-python:3.14 is intentionally NOT done —
-  # per ci-requirements.md, FC is the only Python consumer of that image and the
-  # CI-Runner "add deps to image when used by >1 project" rule keeps it per-job.
-  integration:
-    runs-on: python-ci
-    container:
-      image: git.fabledsword.com/bvandeusen/ci-python:3.14
-    env:
-      DB_USER: fabledcurator
-      DB_PASSWORD: ci_integration
-      DB_PORT: "5432"
-      DB_NAME: fabledcurator_test
-      SECRET_KEY: ci_integration_placeholder
-    services:
-      postgres:
-        image: pgvector/pgvector:pg16
-        env:
-          POSTGRES_USER: fabledcurator
-          POSTGRES_PASSWORD: ci_integration
-          POSTGRES_DB: fabledcurator_test
-        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
-      - name: Integration suite (resolve service IPs, migrate, test)
-        run: |
-          set -eux
-          echo "=== container landscape (diagnostic for filter scoping) ==="
-          docker ps -a --format '{{.ID}} {{.Image}} -> {{.Names}}'
-          echo "=== end landscape ==="
-          PG=$(docker ps --filter "name=integration" --filter "ancestor=pgvector/pgvector:pg16" -q | head -n1)
-          RD=$(docker ps --filter "name=integration" --filter "ancestor=redis:7-alpine" -q | head -n1)
-          test -n "$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 "$RD_IP"
-          export DB_HOST="$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 $(seq 1 60); do
-            if python -c "import socket,sys; s=socket.socket(); s.settimeout(2); sys.exit(0 if s.connect_ex(('$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
-            pip install -r requirements.txt pytest pytest-asyncio
-          fi
-          # Relax durability on the throwaway CI Postgres so the per-test
-          # TRUNCATE's commit-fsync — the integration teardown's dominant cost
-          # (~1.5-2s/test, which collapsed the suite from ~13min to ~45s) — is
-          # skipped. fsync/full_page_writes are sighup GUCs and synchronous_commit
-          # is user-context, so ALTER SYSTEM + pg_reload_conf() applies them with
-          # NO restart. Ephemeral DB ⇒ fsync-off is safe. Non-fatal so a perms
-          # surprise can't red the job; fabledcurator is the postgres image's
-          # bootstrap superuser.
-          python -c "import os,psycopg; c=psycopg.connect(host=os.environ['DB_HOST'],port=5432,user=os.environ['DB_USER'],password=os.environ['DB_PASSWORD'],dbname=os.environ['DB_NAME'],autocommit=True); [c.execute(q) for q in ('ALTER SYSTEM SET fsync=off','ALTER SYSTEM SET synchronous_commit=off','ALTER SYSTEM SET full_page_writes=off','SELECT pg_reload_conf()')]; c.close()" || echo 'WARN: durability GUC relax failed (continuing)'
-          alembic upgrade head
-          pytest tests/ -v -m integration --durations=15

.forgejo/workflows/extension.yml

@@ -1,87 +0,0 @@
-name: extension
-# Lint + unit tests. The sign-and-publish dance moved into build.yml's
-# `sign-extension` job (2026-05-25) — `:latest` now always bundles the XPI
-# because sign-extension runs as a build-web dependency in the SAME workflow,
-# eliminating the prior race between build.yml and a separate extension.yml.
-# Signed XPIs are cached in Forgejo Release Assets named `ext-<version>`.
-on:
-  push:
-    branches: [dev, main]
-    paths:
-      - 'extension/**'
-      - '.forgejo/workflows/extension.yml'
-      # test/version.spec.js asserts things ABOUT the other two workflows —
-      # that neither inlines the packaged-file set, and that build.yml derives
-      # the shipped version rather than reading it out of the repo. A
-      # workflow-only edit can therefore break this suite, so it has to trigger
-      # it. build.yml joined the list at milestone 271 step 5, when the spec
-      # started asserting against it.
-      - '.forgejo/workflows/ci.yml'
-      - '.forgejo/workflows/build.yml'
-  pull_request:
-    branches: [main]
-    paths:
-      - 'extension/**'
-      - '.forgejo/workflows/ci.yml'
-      - '.forgejo/workflows/build.yml'
-  workflow_dispatch:
-
-jobs:
-  lint:
-    runs-on: python-ci
-    container:
-      image: node:24-bookworm-slim
-    steps:
-      - uses: actions/checkout@v4
-      # Not --no-save: vitest and web-ext are both real devDependencies now,
-      # and the suite needs vitest resolvable from node_modules.
-      - name: Install dev dependencies
-        run: cd extension && npm install --no-audit --no-fund
-      - name: Lint
-        run: cd extension && npm run lint
-      # Pure-logic specs over lib/url.js and lib/platforms.js plus manifest /
-      # package version-consistency checks. No browser, no network.
-      - name: Unit tests
-        run: cd extension && npm run test:unit
-
-      # Everything else about packaging is asserted against our own declaration
-      # of what ships. This is the only check that asks web-ext what it ACTUALLY
-      # put in the archive. Until now that was an unverified assumption about
-      # glob semantics — and a fragile one: `test/**` reaches web-ext intact
-      # only because callers `set -f` first, so losing that quoting would
-      # silently start shipping dev files with no other signal.
-      - name: Verify XPI contents
-        run: |
-          set -eu
-          command -v unzip >/dev/null 2>&1 || { apt-get update -qq && apt-get install -y -qq unzip; }
-          cd extension
-          npm run build
-          ZIP=$(ls web-ext-artifacts/*.zip | head -1)
-          echo "=== packaged entries in $ZIP ==="
-          unzip -Z1 "$ZIP" | sort
-          echo "=== end ==="
-          ENTRIES=$(unzip -Z1 "$ZIP")
-          fail=0
-          # Must NOT ship: repo infrastructure with no business in a user's browser.
-          for pat in 'test/' 'scripts/' 'vitest.config.js' 'package.json' 'package-lock.json' 'README.md' 'node_modules/' 'web-ext-artifacts/'; do
-            if echo "$ENTRIES" | grep -q "^$pat"; then
-              echo "ERROR: '$pat' was packaged into the XPI but must not be"
-              fail=1
-            fi
-          done
-          # Must ship: if an exclusion pattern ever over-matches, the extension
-          # breaks at runtime rather than at build time, so assert presence too.
-          for req in 'manifest.json' 'lib/url.js' 'lib/api.js' 'lib/platforms.js' 'lib/cookies.js'; do
-            if ! echo "$ENTRIES" | grep -q "^$req$"; then
-              echo "ERROR: '$req' is missing from the XPI"
-              fail=1
-            fi
-          done
-          for dir in 'background/' 'popup/' 'options/' 'content/' 'icons/'; do
-            if ! echo "$ENTRIES" | grep -q "^$dir"; then
-              echo "ERROR: nothing from '$dir' was packaged"
-              fail=1
-            fi
-          done
-          [ "$fail" -eq 0 ] || exit 1
-          echo "XPI contents verified."

Dockerfile

@@ -28,6 +28,10 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
         postgresql-client \
         zstd \
         megatools \
+    # PID 1 for every role. See the ENTRYPOINT note at the foot of this file:
+    # without it the image needs `init: true` in whatever runs it, which is a
+    # deployment remembering a flag for the image to behave correctly.
+        tini \
         libjpeg62-turbo \
         libwebp7 \
         libpng16-16 \
@@ -36,9 +40,59 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 
 WORKDIR /app
 
-COPY requirements.txt ./
+COPY requirements.txt requirements-ml.txt ./
 RUN pip install -r requirements.txt
 
+# --- ML, merged from Dockerfile.ml (milestone 422 step 6) --------------------
+#
+# ONE image now serves every lane. It was two because the ML lane ran in its
+# own container; with the single-container layout (step 5) running every lane
+# in one process tree, a second image would mean the `ml` lane could never be
+# enabled from the UI — there would be no worker in this container to enable.
+#
+# THE COST, MEASURED from run 7273 rather than guessed — and it is far
+# smaller than the estimate this comment first carried, which said "everyone
+# pulls ~4GB":
+#
+#   torch 2.12.1+cpu wheel     192.3 MB
+#   torchvision 0.27.1+cpu       1.8 MB
+#   transformers / onnxruntime / opencv / sklearn and friends (opencv and
+#   onnxruntime since dropped, #1451 — nothing here imported them)
+#                               62.0, 35.3, 23.6, 16.7, 12.3, 9.2, 6.9 MB
+#   largest newly-pushed layer 222.07 MB
+#
+# So the ML code adds a few hundred MB to the pull, not gigabytes. The CPU
+# index is what makes that true: the default PyPI torch wheel bundles the
+# NVIDIA CUDA runtime and is ~2GB on its own.
+#
+# The GIGABYTES are in the MODEL — ~3.5GB of SigLIP weights — and those are
+# NOT in this image. They arrive only when the operator enables the lane,
+# which is what lets rule 164 permit a runtime fetch at all ("optional and
+# clearly off"). That also settles the trade this step was asked to weigh:
+# baking the weights in would add ~3.5GB to every pull for a feature many
+# adopters never enable, against ~350MB for the code that makes the switch
+# available. Off-by-default wins by an order of magnitude, which was NOT
+# obvious before measuring — the estimate had the two costs within 15% of
+# each other.
+#
+# `--index-url`, not `--extra-index-url`: the latter would let pip resolve a
+# +cu wheel anyway, and the whole saving above depends on it not doing that.
+#
+# CPU-only torch from the PyTorch CPU index. Nothing here uses a GPU — the
+# GPU agent is a separate service with its own image.
+RUN pip install --index-url https://download.pytorch.org/whl/cpu \
+        "torch>=2.14" "torchvision>=0.29"
+RUN pip install -r requirements-ml.txt
+
+# Where the model lands. Deliberately NOT a VOLUME instruction: that mints an
+# anonymous volume when nobody mounts one, which survives `docker rm` and
+# accumulates 3.5GB copies nobody can find. The compose files mount it
+# explicitly instead, so an unmounted run simply re-downloads — visible, and
+# recoverable.
+ENV HF_HOME=/models/.huggingface \
+    TRANSFORMERS_CACHE=/models/.huggingface \
+    ML_MODEL_DIR=/models
+
 COPY backend/ ./backend/
 COPY alembic/ ./alembic/
 COPY alembic.ini ./
@@ -72,5 +126,51 @@ ENV FC_VERSION=${FC_VERSION}
 
 EXPOSE 8080
 
-ENTRYPOINT ["./entrypoint.sh"]
-CMD ["web"]
+# ONE healthcheck for every role, because the image knows which role it is
+# running and a deployment should not have to repeat it. `healthcheck` reads
+# the role entrypoint.sh recorded and asks the right question: HTTP for web,
+# a self-addressed celery ping for a worker lane, both-for-every-lane for the
+# consolidated `all`.
+#
+# start-period covers the SLOWEST role, which is `all`: alembic, then
+# hypercorn, then four celery workers registering with the broker. A web-only
+# container is ready long before this; the cost of the shared number is that
+# a broken one takes a little longer to be called broken.
+#
+# A service may still declare its own healthcheck and docker will prefer it —
+# the escape hatch for a deployment that wants something different.
+HEALTHCHECK --interval=30s --timeout=15s --start-period=90s --retries=3 \
+  CMD ["python", "-m", "backend.app.scripts.healthcheck"]
+
+# tini is PID 1, and the image brings its own rather than asking the
+# deployment for one.
+#
+# PID 1 carries a duty no other process has: every orphaned process in the
+# container reparents to it and must be reaped, or it stays a zombie holding
+# a PID slot. This app makes orphans in normal operation — six service
+# modules shell out (gallery-dl, ffmpeg, pg_dump, the external fetchers) and
+# celery's prefork pool forks children that spawn them.
+#
+# Whatever the role, something that is not an init ends up as PID 1:
+# supervisord for `all`, hypercorn for `web`, celery for a worker. The fix
+# was `init: true` in the compose/stack file, which is out of the norm and
+# put correct process handling in the hands of whoever deploys the image —
+# the same mistake as declaring the healthcheck per service. A flag that is
+# silently dropped (an older Swarm, a `docker run` without it) costs reaping
+# with no signal at all.
+#
+# So the image owns it. `docker run <image>` is correct on its own, and
+# nothing downstream has to know. The smoke asserts /proc/1/comm is tini.
+ENTRYPOINT ["/usr/bin/tini", "--", "./entrypoint.sh"]
+# The DEFAULT is the whole application, not one lane of it.
+#
+# `docker run fabledcurator` with no command starts hypercorn plus every
+# worker lane under supervisord — the shape an adopter wants and the shape the
+# consolidated stack runs. It was `web`, which meant the single-container
+# layout only worked if you knew to ask for it by name, and a compose file
+# that forgot `command:` got a web server with nothing processing its queues:
+# a gallery that loads, accepts an import, and never finishes one.
+#
+# The multi-service stack is unaffected — every service there names its role
+# explicitly, which is exactly what makes it the multi-service stack.
+CMD ["all"]

Dockerfile.ml

@@ -1,43 +0,0 @@
-# syntax=docker/dockerfile:1.25
-
-FROM python:3.14-slim
-ENV PYTHONUNBUFFERED=1 \
-    PYTHONDONTWRITEBYTECODE=1 \
-    PIP_NO_CACHE_DIR=1 \
-    PIP_DISABLE_PIP_VERSION_CHECK=1 \
-    HF_HOME=/models/.huggingface \
-    TRANSFORMERS_CACHE=/models/.huggingface \
-    ML_MODEL_DIR=/models
-
-RUN apt-get update && apt-get install -y --no-install-recommends \
-        ffmpeg \
-        libpq5 \
-        libjpeg62-turbo \
-        libwebp7 \
-        libpng16-16 \
-        libgl1 \
-        libglib2.0-0 \
-        ca-certificates \
-    && rm -rf /var/lib/apt/lists/*
-
-WORKDIR /app
-
-COPY requirements-ml.txt requirements.txt ./
-# CPU-only torch: the default PyPI wheel bundles the CUDA runtime (~5.6GB
-# layer); this pipeline never uses a GPU. --index-url (not --extra-index-url)
-# guarantees only +cpu wheels are considered, so no nvidia-*-cu12 deps.
-RUN pip install --index-url https://download.pytorch.org/whl/cpu \
-        "torch>=2.12,<3.0" "torchvision>=0.27,<0.28"
-RUN pip install -r requirements-ml.txt
-
-COPY backend/ ./backend/
-COPY alembic/ ./alembic/
-COPY alembic.ini ./
-COPY entrypoint.sh ./
-RUN chmod +x entrypoint.sh
-
-# Models self-heal into /models on first start (FC-2 implements this)
-VOLUME ["/models"]
-
-ENTRYPOINT ["./entrypoint.sh"]
-CMD ["ml-worker"]

README.md

@@ -245,9 +245,9 @@ reasoning is note #3127 §5). Rolling back is `docker pull …:c-<sha>`.
 
 Each artifact still has a version, derived rather than chosen: the commit time
 of the newest change to that artifact's *own* shipped files, as
-`YYYY.MM.DD.HHMM` UTC (rule 148). Four artifacts, four independent versions —
-a push touching only `agent/` re-versions the agent and leaves web and ml
-alone, and CI skips the builds whose content did not move.
+`YYYY.MM.DD.HHMM` UTC (rule 148). Three artifacts, three independent versions
+— a push touching only `agent/` re-versions the agent and leaves web and the
+extension alone, and CI skips the builds whose content did not move.
 
 Because no registry name carries it, the running instance's own report is the
 only answer to "which build is this?". The foot of Settings shows
@@ -260,22 +260,32 @@ commits since the previous tag; it builds no image.
 
 ## What's in here
 
-Five deployable pieces, built by `.forgejo/workflows/build.yml`:
+Four deployable pieces, built by `.forgejo/workflows/build.yml`:
 
 | Piece | Built from | Image | Role |
 | --- | --- | --- | --- |
-| **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. |
+| **Web / workers** | `Dockerfile` | `fabledcurator` | Quart API + the built Vue SPA in one image. `entrypoint.sh` picks the role: `web`, `worker`, `scheduler`, `ml-worker`, or `all` (every lane under supervisord, the single-container layout). The `maintenance-long` service is a second `worker` pinned to the long-running maintenance queue. |
 | **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. |
 
 ## CI / Forgejo setup
 
-Four workflows: `ci.yml` (lint, extension-version check, backend unit tests,
-frontend build, integration), `extension.yml` (extension lint, vitest, XPI
-content verification), `build.yml` (sign + publish), and `release.yml`, which
-runs only on a `v*` tag and publishes a changelog without building anything.
+Two workflows that matter here: `build.yml` (the six verification lanes — lint,
+extension-version check, backend unit tests, frontend build, extension lint +
+vitest + XPI content check, integration — and then sign + publish), and
+`release.yml`, which runs only on a `v*` tag and publishes a changelog without
+building anything. The extension lane was its own `extension.yml` until
+milestone 429, which let a red extension suite sign and ship the XPI anyway.
+
+**The lanes and the publish are one workflow on purpose.** They were two
+(`ci.yml` and `build.yml`) until 2026-09-23, on the same push trigger, which
+meant the build could not see the tests' verdict and published whatever it
+built — a red unit lane and a fresh `:dev` image, in the same minute. A
+`needs:` edge only exists inside one workflow graph, so the two are one graph
+and the gate is that edge: a lane that fails, **or that merely skips**, leaves
+the publishing jobs unrun. Pull-request runs (Renovate bumps into `dev`) are
+the lanes and nothing else.
 
 **The toolchain each job runs in is its `container.image`, not its `runs-on`
 label.** `runs-on: python-ci` only schedules the job onto a runner; every job

agent/Dockerfile

@@ -1,10 +1,21 @@
 # FabledCurator GPU agent — runs on the desktop with the GPU.
-# CUDA 12.9 + cuDNN 9 runtime so onnxruntime-gpu can use the card (it needs
-# cuDNN 9 — the plain -runtime image lacks it: "libcudnn.so.9: cannot open
-# shared object file"); ffmpeg for video frames. Ubuntu 24.04 → Python 3.12.
-# Stays on the CUDA-12 / cuDNN-9 line the default onnxruntime-gpu + torch are
-# built against (CUDA 13 has only nascent ONNX Runtime support).
-FROM nvidia/cuda:12.9.2-cudnn-runtime-ubuntu24.04
+#
+# The `base` flavour, not `cudnn-runtime`: CUDA and cuDNN arrive as the
+# `nvidia-*` pip packages torch and onnxruntime-gpu depend on, so the base only
+# has to hand the container the driver (it sets NVIDIA_VISIBLE_DEVICES /
+# NVIDIA_DRIVER_CAPABILITIES for the Container Toolkit). Until #1451 this was
+# `12.9.2-cudnn-runtime` under a `torch==2.6.0+cu124` — and requirements.txt then
+# REPLACED that torch with PyPI's CUDA-13 build (ultralytics pulls torchvision,
+# which pulls its matching torch), beside a CUDA-13 onnxruntime-gpu. The image
+# ran CUDA 13 on a CUDA-12 base, carrying ~3 GB of base libraries and a ~3 GB
+# torch nothing loaded: 10 GB compressed.
+#
+# 13.0 because that is the line both wheels are built for (torch's cu130 index,
+# onnxruntime-gpu's `nvidia-cuda-runtime~=13.0`). Needs an NVIDIA driver that
+# supports CUDA 13 (580+); fc_agent/accel.py logs at startup whether torch and
+# onnxruntime actually got the GPU, since both fall back to the CPU silently.
+# ffmpeg for video frames. Ubuntu 24.04 → Python 3.12.
+FROM nvidia/cuda:13.0.3-base-ubuntu24.04
 
 # PIP_BREAK_SYSTEM_PACKAGES: Ubuntu 24.04 marks its system Python as externally
 # managed (PEP 668), so a global `pip install` errors without this. It's a
@@ -16,10 +27,12 @@ RUN apt-get update \
     && rm -rf /var/lib/apt/lists/*
 
 WORKDIR /app
-# torch from the CUDA-12.4 wheel index; its wheels bundle their own CUDA + cuDNN
-# so they run on the 12.9 base and coexist with onnxruntime-gpu. Installed first
-# + separately so the GPU build of torch is deterministic and layer-cached.
-RUN pip3 install --no-cache-dir torch==2.6.0 --index-url https://download.pytorch.org/whl/cu124
+# torch AND torchvision from the cu130 index, together and first. Installing
+# torch alone is what let the next step swap it out: ultralytics needs
+# torchvision, PyPI's torchvision pins its own torch, and pip replaced ours to
+# match. With both present, requirements.txt finds them satisfied.
+RUN pip3 install --no-cache-dir --index-url https://download.pytorch.org/whl/cu130 \
+        torch torchvision
 COPY requirements.txt .
 RUN pip3 install --no-cache-dir -r requirements.txt
 COPY fc_agent ./fc_agent
@@ -27,6 +40,23 @@ COPY fc_agent ./fc_agent
 # imgutils ONNX models + the transformers SigLIP weights both cache here; mount
 # a volume to persist them across restarts (the SigLIP download is ~3.5 GB once).
 ENV HF_HOME=/models
+
+# Declared LAST on purpose, exactly as the web Dockerfile does: an ARG/ENV
+# invalidates every layer below it, and these are the only values that differ
+# between builds of otherwise identical source. Any earlier and the ~6.3 GB
+# CUDA + torch layers could never be shared between the dev and main builds of
+# one commit — which is the cost #3114 measured at 9m26s cold.
+#
+# Three values, never folded together (rule 149) — the NAME a person reads, the
+# CHANNEL it came from, and the REVISION that identifies the content. See
+# fc_agent/build_info.py; CI derives all three from scripts/artifacts.sh.
+ARG FC_CHANNEL=""
+ENV FC_CHANNEL=${FC_CHANNEL}
+ARG FC_VERSION=""
+ENV FC_VERSION=${FC_VERSION}
+ARG FC_REVISION=""
+ENV FC_REVISION=${FC_REVISION}
+
 EXPOSE 8770
 
 # The control UI; the worker is started from it (or POST /start).

agent/README.md

@@ -15,13 +15,28 @@ sudo pacman -S nvidia-container-toolkit
 sudo nvidia-ctk runtime configure --runtime=docker
 sudo systemctl restart docker
 # verify:
-docker run --rm --gpus all nvidia/cuda:12.4.1-base-ubuntu22.04 nvidia-smi
+docker run --rm --gpus all nvidia/cuda:13.0.3-base-ubuntu24.04 nvidia-smi
+# the header's CUDA version must be 13.0 or later (driver 580+)
+```
+
+### After a driver update: regenerate the CDI spec
+If the agent's first log lines say `accel: torch is NOT on the GPU` or report
+`cudaGetDeviceCount: unknown error (999)` while `nvidia-smi` still works, the
+toolkit's saved device list (`/etc/cdi/nvidia.yaml`) is out of date. The
+`nvidia-uvm` device number changes between driver versions, and a spec
+generated before the update hands the container a device node that no longer
+exists (2026-09-24: host `511,0`, container `235,0`). Compare
+`ls -l /dev/nvidia-uvm` on the host with the same inside the container, then:
+```sh
+sudo nvidia-ctk cdi generate --output=/etc/cdi/nvidia.yaml
+# if your toolkit ships it, this keeps it current on every driver update:
+sudo systemctl enable --now nvidia-cdi-refresh.path
 ```
 
 ## 1. Get a token
 In FC: **Settings → Tagging → GPU agent → Generate token** (or Rotate). Copy it.
 
-## 2. Pull (CI publishes it alongside the web/ml images)
+## 2. Pull (CI publishes it alongside the web image)
 ```sh
 docker pull git.fabledsword.com/bvandeusen/fabledcurator-agent:latest
 ```

agent/fc_agent/accel.py

@@ -0,0 +1,124 @@
+"""Which accelerator each runtime actually got — reported once, at startup.
+
+The agent has two GPU runtimes and both fall back to the CPU without raising:
+torch when the driver is too old for its CUDA build, and onnxruntime (the imgutils
+detector + CCIP models) when its CUDA provider cannot load its libraries. A
+fallback shows up only as slower work, and nothing reported it. On 2026-09-24 the
+image turned out to be running a CUDA-13 torch and onnxruntime on a CUDA-12 base
+(#1451), and whether the ONNX half was on the GPU could not be answered from
+anything the agent had ever logged.
+
+Also the fix for the likeliest way the ONNX half misses: onnxruntime-gpu's CUDA
+provider finds libcudart/cuBLAS/cuDNN only on the loader path, and in this image
+they live in the `nvidia-*` pip packages torch installs. `preload_dlls()` (ORT
+1.21+) loads them from there, so the provider resolves them by soname.
+
+Stdlib-only at import, so the unit suite can import it — torch and onnxruntime
+are imported inside the functions.
+"""
+
+from __future__ import annotations
+
+import ctypes
+import importlib
+import logging
+from pathlib import Path
+
+log = logging.getLogger("fc_agent.accel")
+
+# Filled by report(); /status carries it so the page can show it too.
+LAST: dict = {}
+
+
+def torch_status(imp=importlib.import_module) -> dict:
+    try:
+        torch = imp("torch")
+    except Exception as e:
+        return {"device": "unavailable", "error": str(e)}
+    out = {"version": torch.__version__, "cuda_build": torch.version.cuda}
+    if torch.cuda.is_available():
+        out["device"] = "cuda"
+        out["gpu"] = torch.cuda.get_device_name(0)
+    else:
+        out["device"] = "cpu"
+    return out
+
+
+def onnx_status(imp=importlib.import_module, load=ctypes.CDLL) -> dict:
+    try:
+        ort = imp("onnxruntime")
+    except Exception as e:
+        return {"device": "unavailable", "error": str(e)}
+    out = {"version": ort.__version__, "providers": list(ort.get_available_providers())}
+    if "CUDAExecutionProvider" not in out["providers"]:
+        out["device"] = "cpu"
+        return out
+    preload = getattr(ort, "preload_dlls", None)
+    if preload is not None:
+        try:
+            preload()
+        except Exception as e:
+            out["preload_error"] = str(e)
+    # "Available" only means the build HAS the provider. Loading its library is
+    # what resolves libcudart/cuBLAS/cuDNN — the step that fails when they are
+    # missing, and the one a session would otherwise fail silently on.
+    capi = Path(ort.__file__).parent / "capi"
+    try:
+        load(str(capi / "libonnxruntime_providers_shared.so"), mode=ctypes.RTLD_GLOBAL)
+        load(str(capi / "libonnxruntime_providers_cuda.so"))
+    except OSError as e:
+        out["device"] = "cpu"
+        out["error"] = str(e)
+        return out
+    # Loading proves the libraries resolve, NOT that a GPU can be used: on
+    # 2026-09-24 this reported "onnx on GPU" beside torch failing cuInit with
+    # "CUDA unknown error" (a driver update awaiting a reboot). Asking the CUDA
+    # runtime for a device initialises the driver the provider would use.
+    error = _cuda_device_error(load)
+    out["device"] = "cpu" if error else "cuda"
+    if error:
+        out["error"] = error
+    return out
+
+
+def _cuda_device_error(load=ctypes.CDLL) -> str | None:
+    """None when the CUDA runtime can reach a device, else why it cannot."""
+    try:
+        cudart = load("libcudart.so.13")
+    except OSError as e:
+        return str(e)
+    count = ctypes.c_int(0)
+    rc = cudart.cudaGetDeviceCount(ctypes.byref(count))
+    if rc != 0:
+        cudart.cudaGetErrorString.restype = ctypes.c_char_p
+        return f"cudaGetDeviceCount: {cudart.cudaGetErrorString(rc).decode()} ({rc})"
+    return None if count.value > 0 else "no CUDA device visible"
+
+
+def summary() -> dict | None:
+    """The report as FabledCurator stores it: each runtime's device, and why
+    when it is not the GPU. Sent on every lease and heartbeat, so the System
+    view can call a running agent that fell back to the CPU "degraded" rather
+    than "running" — the 2026-09-24 fallback went unseen for weeks because
+    only this agent's own log said so. None before report() has run."""
+    if not LAST:
+        return None
+    out = {}
+    for name, s in LAST.items():
+        entry = {"device": s.get("device")}
+        if s.get("error"):
+            entry["error"] = str(s["error"])[:200]
+        out[name] = entry
+    return out
+
+
+def report() -> dict:
+    """Check both runtimes, log the result, and keep it for /status."""
+    LAST.clear()
+    LAST.update(torch=torch_status(), onnx=onnx_status())
+    for name, s in LAST.items():
+        if s.get("device") == "cuda":
+            log.info("accel: %s on GPU (%s)", name, s)
+        else:
+            log.warning("accel: %s is NOT on the GPU — work runs on the CPU (%s)", name, s)
+    return dict(LAST)

agent/fc_agent/app.py

@@ -11,17 +11,22 @@ import logging
 from fastapi import FastAPI, Request
 from fastapi.responses import HTMLResponse, JSONResponse
 
-from . import logbuf
+from . import accel, logbuf
+from .build_info import FC_CHANNEL, FC_REVISION, FC_VERSION, build_id, display_version
 from .config import Config
 from .gpu import read_gpu
 from .worker import Worker
 
 log = logging.getLogger("fc_agent.app")
 
-# Bump on every agent change. The page embeds this and /status reports it; the UI
-# warns to reload when they differ — so a stale browser-cached page can't be
-# mistaken for "the new image didn't deploy". (Belt-and-braces with no-store.)
-VERSION = "2026-07-17.1 · idle model-unload: after ~5 min idle the GPU models release their VRAM and reload on the next job (env IDLE_UNLOAD_SECONDS, 0=off) · sleep mode sheds to one downloader"
+# DERIVED at image build time, not hand-maintained — see build_info. This was a
+# literal an author was asked to bump on every agent change, and the September
+# image printed the same "2026-07-17.1" as the July one, so the surface meant to
+# answer "did my pull work?" answered the same either way.
+#
+# Two values with two jobs, kept apart (rule 149): the page SHOWS the version
+# and COMPARES the build id. /status reports both, plus the raw fields, so a
+# reader never has to take a formatted string apart to get at one of them.
 
 logbuf.install()
 cfg = Config.from_env()
@@ -42,6 +47,9 @@ async def _no_store(request, call_next):
 
 @app.on_event("startup")
 def _maybe_autostart() -> None:
+    # Before the worker: the report also preloads the CUDA libraries the ONNX
+    # models need, and it says in the log which runtimes landed on the GPU.
+    accel.report()
     # With AUTO_START set, a container restart (host reboot, or `restart:
     # unless-stopped` after a crash) resumes the worker on its own — the slots
     # then ride out a still-down curator via lease backoff. Lets the agent
@@ -52,7 +60,14 @@ def _maybe_autostart() -> None:
 
 @app.get("/", response_class=HTMLResponse)
 def index() -> str:
-    return _PAGE.replace("__BUILD__", VERSION)
+    # Two substitutions, not one: `__VERSION__` is what a person reads in the
+    # meta line, `__BUILD_ID__` is what the script compares against /status to
+    # notice the page is a cached copy from a previous build.
+    return (
+        _PAGE
+        .replace("__VERSION__", display_version())
+        .replace("__BUILD_ID__", build_id())
+    )
 
 
 @app.post("/start")
@@ -117,7 +132,15 @@ def status():
     s["fc_url"] = cfg.fc_url
     s["configured"] = bool(cfg.token)
     s["queue"] = worker.latest_queue()
-    s["build"] = VERSION
+    # `build` is the comparison token the page checks — see build_info.
+    # `version`/`channel`/`revision` ride BESIDE it rather than inside it, so a
+    # reader wanting the version never has to parse it back out of something
+    # else. Absent rather than empty when the image carries no stamp.
+    s["build"] = build_id()
+    s["version"] = FC_VERSION or None
+    s["channel"] = FC_CHANNEL or None
+    s["revision"] = FC_REVISION or None
+    s["accel"] = accel.LAST or None
     return JSONResponse(s)
 
 
@@ -169,7 +192,11 @@ _PAGE = """<!doctype html><html><head><meta charset=utf-8>
   width:30px;height:32px;font:700 16px system-ui;cursor:pointer}
  .step:hover{border-color:var(--acc)}
  #conc,#bw{width:3.4rem;height:32px;text-align:center;font:700 16px system-ui;background:#11151a;
-  color:var(--fg);border:1px solid var(--bd);border-radius:8px}
+  color:var(--fg);border:1px solid var(--bd);border-radius:8px;appearance:textfield;-moz-appearance:textfield}
+ /* The browser's own spin arrows, hidden: the − / + beside each field are the
+    control, styled like the rest of the page (operator, 2026-09-24). */
+ #conc::-webkit-inner-spin-button,#conc::-webkit-outer-spin-button,
+ #bw::-webkit-inner-spin-button,#bw::-webkit-outer-spin-button{-webkit-appearance:none;margin:0}
  .unit{color:var(--mut);font-size:12px;font-weight:600}
  .hint{color:var(--mut);font-size:12px;margin-top:12px}
  .tiles{display:grid;grid-template-columns:repeat(6,1fr);gap:8px;margin-bottom:16px}
@@ -203,11 +230,12 @@ _PAGE = """<!doctype html><html><head><meta charset=utf-8>
   <div class=brand><span class=logo>◆</span> FabledCurator <span class=sub>GPU agent</span></div>
   <div class=conn><span class="dot" id=dot></span><span id=connlbl>—</span></div>
  </header>
- <p class=meta>Server <code id=fc>—</code> · token <code id=cfg>—</code> · build <code id=build>__BUILD__</code></p>
+ <p class=meta>Server <code id=fc>—</code> · token <code id=cfg>—</code> · build <code id=build>__VERSION__</code></p>
 
  <div id=verbanner class=banner style="display:none;background:#3a1212;border-color:#5a1717;color:#ffb3b3">
   a newer agent version is running — reload this page (Ctrl+Shift+R) to update the controls
  </div>
+ <div id=accelbanner class=banner style="display:none;background:#3a1212;border-color:#5a1717;color:#ffb3b3"></div>
  <div id=banner class=banner style=display:none>
   curator unreachable — holding work + retrying, resumes on its own (no restart needed)
  </div>
@@ -225,7 +253,9 @@ _PAGE = """<!doctype html><html><head><meta charset=utf-8>
     <button class=step onclick=setc(1)>+</button>
    </div>
    <div class=stepper title="aggregate download cap, downloads + video streams combined — 0 = unlimited">
+    <button class=step onclick=stepbw(-1)>−</button>
     <input id=bw type=number min=0 step=1 value=8 onchange="setbw(this.value)">
+    <button class=step onclick=stepbw(1)>+</button>
     <span class=unit>MB/s</span>
    </div>
   </div>
@@ -262,7 +292,7 @@ _PAGE = """<!doctype html><html><head><meta charset=utf-8>
  </section>
 </div>
 <script>
- const PAGE_BUILD="__BUILD__"
+ const PAGE_BUILD="__BUILD_ID__"
  let CAP=8
  // Optimistic transitional state on click, then apply the POST's own status
  // response (it returns worker.status()) for instant feedback — don't wait on the
@@ -293,6 +323,14 @@ _PAGE = """<!doctype html><html><head><meta charset=utf-8>
   await fetch('/auto',{method:'POST',headers:{'Content-Type':'application/json'},
    body:JSON.stringify({value:on})});refresh()
  }
+ function stepbw(d){ setbw((parseFloat(bw.value)||0)+d) }
+ // Runtimes that did NOT get the GPU, from the startup report. Both fall back
+ // to the CPU without raising, so this banner and the pill are the only place
+ // on this page a slow, CPU-bound agent announces itself.
+ function cpuOnly(s){
+  const a=s.accel||{}
+  return Object.keys(a).filter(k=>a[k] && a[k].device!=='cuda')
+ }
  async function setbw(v){
   v=Math.max(0,parseFloat(v)||0); bw.value=v
   await fetch('/bandwidth',{method:'POST',headers:{'Content-Type':'application/json'},
@@ -363,11 +401,17 @@ _PAGE = """<!doctype html><html><head><meta charset=utf-8>
    // unreachable curator; grey when stopped; red with no token.
    let dc='dot', lbl='stopped'
    if(!ok){ dc='dot red'; lbl='no token' }
-   else if(st==='running'){ dc='dot '+(s.queue?'green':'amber'); lbl=s.queue?'running':'running · curator unreachable' }
+   else if(st==='running'){ dc='dot '+(s.queue?'green':'amber'); lbl=s.queue?'running':'running · curator unreachable'
+    if(s.queue && cpuOnly(s).length){ dc='dot amber'; lbl='running · CPU only (degraded)' } }
    else if(st==='starting'){ dc='dot amber'; lbl='starting…' }
    else if(st==='stopping'){ dc='dot amber'; lbl='stopping…' }
    dot.className=dc; connlbl.textContent=lbl
    banner.style.display=(st==='running' && !s.queue)?'block':'none'
+   const slow=cpuOnly(s)
+   accelbanner.style.display=slow.length?'block':'none'
+   accelbanner.textContent=slow.length?('degraded — '+slow.join(' + ')+' not on the GPU, so that work runs on the CPU: '
+    +slow.map(k=>k+': '+(s.accel[k].error||s.accel[k].device)).join(' · ')
+    +'. After a driver update, regenerate the CDI spec (agent README).'):''
    queue.textContent=s.queue?('queue · pending '+s.queue.pending+' · in flight '+s.queue.leased+' · done '+s.queue.done+' · errored '+s.queue.error):'queue · unreachable'
   }
  }

agent/fc_agent/build_info.py

@@ -0,0 +1,79 @@
+"""What this agent build IS — stamped at image build time, not configurable.
+
+The mirror of `backend/app/build_info.py`, for the same reasons and with the
+same posture. Kept as its own module rather than as constants in `app.py`
+because it is stdlib-only and therefore importable by the test suite, which
+cannot import `app` (torch, transformers and ultralytics are not in the CI
+image — see build.yml's "Agent syntax check").
+
+## Why this replaced a hand-written string
+
+`app.VERSION` used to be a literal an author was asked to bump, carrying a
+version AND a changelog in one string:
+
+    VERSION = "2026-07-17.1 · idle model-unload: after ~5 min idle ..."
+
+Nobody bumped it. The September image printed the identical string to the July
+one, so the one surface that was supposed to answer *"did my pull work?"*
+answered *"2026-07-17.1"* either way. An artifact that cannot identify itself
+is worse than one that says nothing, because the stale value reads as an
+answer.
+
+The values are now derived by `scripts/artifacts.sh` from the commit its
+shipped files last changed in — the same derivation the web image has used
+since milestone 313, and the same one the reuse check already ran for the
+agent and discarded.
+
+## Three values, never folded together (rule 149)
+
+* `FC_VERSION` — the NAME, `YYYY.MM.DD.HHMM` UTC, derived from COMMIT time.
+  For people to read and quote. Identical on `dev` and `main` for the same
+  source, which is the property that makes "am I running the same code as
+  production?" answerable at a glance.
+* `FC_CHANNEL` — a SIBLING field, never a suffix inside the name.
+* `FC_REVISION` — the 12-char commit sha, the artifact's IDENTITY. This is
+  what the reuse check already keys on as the `fc.revision` image label.
+
+**Absent rather than empty when unknown.** A locally built image has no
+stamp, and neither does any image predating this module. One spelling of
+"cannot say" instead of two.
+"""
+
+import os
+
+FC_VERSION = os.environ.get("FC_VERSION", "").strip()
+FC_CHANNEL = os.environ.get("FC_CHANNEL", "").strip()
+FC_REVISION = os.environ.get("FC_REVISION", "").strip()
+
+
+def display_version() -> str:
+    """The build, as a line for a human: `2026.09.24.1052 (dev)`.
+
+    `unknown` rather than a blank when unstamped — an empty slot in the meta
+    line reads as "no version", which is a different and false claim from "this
+    build does not carry one".
+    """
+    if not FC_VERSION:
+        return "unknown"
+    return f"{FC_VERSION} ({FC_CHANNEL})" if FC_CHANNEL else FC_VERSION
+
+
+def build_id() -> str:
+    """The token the control page compares against `/status` to notice it is
+    showing a CACHED page from a previous build.
+
+    Deliberately NOT `display_version()`. That is the value for reading; this
+    is the value for deciding, and folding the two is what rule 149 is about.
+    The revision is the better discriminator of the two — two builds of the
+    same commit ARE the same agent and should not prompt a reload, and two
+    different commits always differ here even when they land in the same
+    minute and derive one version name.
+
+    A locally built image falls through to a constant, so the reload banner
+    cannot fire for it. That is honest rather than a gap: nothing in an
+    unstamped image knows what source it was built from, and a per-process
+    nonce would make every ordinary container RESTART claim a new version had
+    arrived — a false positive on the exact surface the banner exists to keep
+    trustworthy.
+    """
+    return FC_REVISION or FC_VERSION or "local"

agent/fc_agent/client.py

@@ -7,6 +7,8 @@ import requests
 from requests.adapters import HTTPAdapter
 from urllib3.util.retry import Retry
 
+from . import accel
+
 
 class FcClient:
     def __init__(self, base_url: str, token: str, agent_id: str):
@@ -72,7 +74,10 @@ class FcClient:
     def lease(self, batch_size: int) -> list[dict]:
         r = self.s.post(
             f"{self.base}/api/gpu/jobs/lease",
-            json={"agent_id": self.agent_id, "batch_size": batch_size},
+            json={
+                "agent_id": self.agent_id, "batch_size": batch_size,
+                "accel": accel.summary(),
+            },
             timeout=30,
         )
         r.raise_for_status()
@@ -90,7 +95,9 @@ class FcClient:
         })
 
     def heartbeat(self, job_ids: list[int]) -> None:
-        self._post_quiet("/api/gpu/jobs/heartbeat", {"job_ids": job_ids})
+        self._post_quiet(
+            "/api/gpu/jobs/heartbeat", {"job_ids": job_ids, "accel": accel.summary()},
+        )
 
     def fail(self, job_id: int, error: str) -> None:
         self._post_quiet("/api/gpu/jobs/fail", {"job_id": job_id, "error": error})

agent/fc_agent/worker.py

@@ -342,15 +342,44 @@ class Worker:
 
     # --- background loops ---------------------------------------------------
     def _heartbeat_loop(self) -> None:
-        """Keep every held lease alive so buffered jobs waiting on the GPU aren't
-        reclaimed by curator's 180s TTL. Errors are swallowed by client.heartbeat;
-        a reclaimed lease just re-leases elsewhere — never fatal."""
+        """Keep every held lease alive, and say we are here even when holding none.
+
+        Leases: buffered jobs waiting on the GPU would otherwise be reclaimed by
+        curator's 180s TTL. Errors are swallowed by client.heartbeat; a reclaimed
+        lease just re-leases elsewhere — never fatal.
+
+        ## Why this sends with an EMPTY list rather than skipping
+
+        Curator's roster records a check-in on this call (and on `lease`), and
+        calls an agent stopped after 300s of silence. This loop used to be
+        gated on `if ids:` — so an agent holding no leases sent nothing at all,
+        and the only check-in left was the lease poll, which sleep mode backs
+        off exponentially to a 900s ceiling (see IDLE_POLL_MAX_SECONDS).
+
+        900 against 300: an IDLE agent was structurally guaranteed to read as
+        stopped. Operator, 2026-09-23: *"I'm running the gpu agent on my device
+        and it currently reads as 'offline' but it's running and has checked in
+        recently."* It had — twelve minutes ago, partway up the backoff ladder.
+
+        The two halves were written ten weeks apart and never reconciled: sleep
+        mode landed 2026-07-02, and the roster adopted the lease as its
+        check-in on 2026-09-02 without noticing the call it was piggybacking on
+        had been deliberately slowed.
+
+        An empty heartbeat extends nothing (`id.in_([])` matches no rows) and
+        costs one small POST every 45s — against the 6/min lease poll sleep
+        mode exists to avoid, that is not a cadence worth protecting, and it is
+        what makes "is the agent alive" answerable at all.
+
+        Still gated on `self._running`: a worker that has been stopped is not
+        checking in for work, and reporting it as present would be a different
+        lie.
+        """
         while True:
             if self._running:
                 with self._held_lock:
                     ids = list(self._held)
-                if ids:
-                    self.client.heartbeat(ids)
+                self.client.heartbeat(ids)
             time.sleep(HEARTBEAT_INTERVAL)
 
     def _queue_poll_loop(self):

agent/requirements.txt

@@ -1,10 +1,12 @@
 # CCIP + figure detection (ONNX models, auto-downloaded from HuggingFace).
 dghs-imgutils>=0.4
 # GPU inference for the ONNX models. Swap to onnxruntime (CPU) for a slow
-# server-side fallback run.
-onnxruntime-gpu
-# The crop EMBEDDER (concept bag). torch is installed separately in the
-# Dockerfile from the CUDA-12.4 wheel index so the GPU build is deterministic;
+# server-side fallback run. The extras declare the CUDA/cuDNN pip packages its
+# CUDA provider loads (fc_agent/accel.py preloads them) rather than relying on
+# torch happening to install the same ones.
+onnxruntime-gpu[cuda,cudnn]
+# The crop EMBEDDER (concept bag). torch + torchvision are installed separately
+# in the Dockerfile from the cu130 wheel index, so pip never swaps them out;
 # transformers loads whatever SigLIP-family model the server announces.
 transformers>=4.45
 # Crop PROPOSERS — small YOLO detectors (booru_yolo anatomy, COCO person, comic

alembic/versions/0098_phash_256_bit_rehash.py

@@ -0,0 +1,90 @@
+"""Widen image_record.phash to 256-bit and re-hash the library (issue #4223).
+
+The operator reported a 15-image variant pack landing as 3 records, and then
+that variants were STILL being dropped with `phash_threshold` at 0. Zero was
+already the floor of the dial, so no setting could have fixed it: at
+`hash_size=8` a pHash is 64 bits of coarse light/dark layout, and variant
+artwork sharing a composition produces the SAME 64 bits. Distance 0 meant
+"identical hash", never "identical image".
+
+`utils/phash.py` moves to `hash_size=16` (256 bits, what ImageRepo always
+used) and adds an aspect-ratio gate plus a pixel-level confirm, so a merge is
+accepted on the files rather than on the hash.
+
+## Why this NULLs every phash
+
+Widening the column does not correct the values already in it. Every stored
+hash is a 64-bit hash of an image the app will now hash at 256 bits, and the
+two cannot be compared — `find_similar` skips a mismatched-length candidate
+rather than guessing, so leaving them would silently mean "no dedup, forever,
+for everything imported before today". NULL is the state `backfill_phash`
+already knows how to repair: it is NULL-only, keyset-paginated and
+restart-safe, and the beat schedule runs it daily.
+
+Until that backfill finishes, image dedup degrades to sha256 only —
+duplicates may be kept. That is the safe direction, and the only one
+available: the alternative is comparing hashes of different widths, which
+would drop artwork. NOTHING here deletes or supersedes a file.
+
+## Why the threshold is reset rather than carried over
+
+`phash_threshold` counts bits, and the denominator went from 64 to 256. The
+stored number would keep its value while meaning something four times
+tighter. There is no honest carry-over, so every row goes to the new default
+of 24 — including the operator's 0, which was a workaround for the bug this
+revision fixes.
+
+Revision ID: 0098
+Revises: 0097
+Create Date: 2026-09-21
+
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0098"
+down_revision: Union[str, None] = "0097"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # varchar(32) -> varchar(64): widening a length limit is a catalog-only
+    # change in Postgres, so this does not rewrite the table or its index.
+    op.alter_column(
+        "image_record", "phash",
+        existing_type=sa.String(32),
+        type_=sa.String(64),
+        existing_nullable=True,
+    )
+    op.execute("UPDATE image_record SET phash = NULL WHERE phash IS NOT NULL")
+    op.alter_column(
+        "import_settings", "phash_threshold",
+        existing_type=sa.Integer(),
+        server_default="24",
+        existing_nullable=False,
+    )
+    op.execute("UPDATE import_settings SET phash_threshold = 24")
+
+
+def downgrade() -> None:
+    # The 64-bit hashes this replaced are gone, and a 64-char value does not
+    # fit back into varchar(32) — so the column is cleared again on the way
+    # down and left for backfill_phash to refill at whatever HASH_SIZE the
+    # code is running. Rule #22: no legacy to preserve.
+    op.execute("UPDATE image_record SET phash = NULL WHERE phash IS NOT NULL")
+    op.alter_column(
+        "image_record", "phash",
+        existing_type=sa.String(64),
+        type_=sa.String(32),
+        existing_nullable=True,
+    )
+    op.alter_column(
+        "import_settings", "phash_threshold",
+        existing_type=sa.Integer(),
+        server_default="10",
+        existing_nullable=False,
+    )
+    op.execute("UPDATE import_settings SET phash_threshold = 10")

alembic/versions/0099_library_placement_run.py

@@ -0,0 +1,99 @@
+"""library_placement_run — the placement reconciler's plan/apply/undo ledger.
+
+Milestone #421 step 3. The survey (#4245) measured 33,789 ImageRecord rows
+sitting outside their artist's canonical directory, across 56 artists. This
+table holds one run of the sweep that trues them up: the plan, what it did,
+and where every file came from.
+
+## Why the moves live in a table rather than a log line
+
+`ImageRecord.path` is the only pointer at the bytes, so a move rewrites the
+row. Once that write lands, the previous location exists nowhere — unless it
+was recorded first. `moves` is that record, which is what makes a 33,789-file
+operation something the operator can undo per artist after looking at the
+result, rather than a one-way door.
+
+An `applied` row is therefore HISTORY, not state (lesson #4226). Any future
+retention on this table may prune `ready`, `cancelled` and `error` runs; an
+`applied` one is only disposable once someone decides undo is no longer
+wanted. That is deliberately not a timer's decision, and no pruning is added
+here.
+
+Revision ID: 0099
+Revises: 0098
+Create Date: 2026-09-21
+
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects import postgresql
+
+revision: str = "0099"
+down_revision: Union[str, None] = "0098"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "library_placement_run",
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column(
+            "status", sa.String(length=16), server_default="running",
+            nullable=False,
+        ),
+        # SET NULL, not CASCADE: deleting an artist must not destroy the
+        # record of where their files were moved.
+        sa.Column("artist_id", sa.Integer(), nullable=True),
+        sa.Column(
+            "started_at", sa.DateTime(timezone=True),
+            server_default=sa.text("now()"), nullable=False,
+        ),
+        sa.Column("finished_at", sa.DateTime(timezone=True), nullable=True),
+        sa.Column(
+            "planned_count", sa.Integer(), server_default="0", nullable=False,
+        ),
+        sa.Column(
+            "moved_count", sa.Integer(), server_default="0", nullable=False,
+        ),
+        sa.Column(
+            "refused_count", sa.Integer(), server_default="0", nullable=False,
+        ),
+        sa.Column(
+            "moves", postgresql.JSONB(astext_type=sa.Text()),
+            server_default=sa.text("'[]'::jsonb"), nullable=False,
+        ),
+        sa.Column(
+            "refusals", postgresql.JSONB(astext_type=sa.Text()),
+            server_default=sa.text("'[]'::jsonb"), nullable=False,
+        ),
+        sa.Column("error", sa.Text(), nullable=True),
+        sa.ForeignKeyConstraint(
+            ["artist_id"], ["artist.id"],
+            name="fk_library_placement_run_artist_id", ondelete="SET NULL",
+        ),
+        sa.PrimaryKeyConstraint("id"),
+    )
+    op.create_index(
+        "ix_library_placement_run_status", "library_placement_run", ["status"],
+    )
+    op.create_index(
+        "ix_library_placement_run_artist_id", "library_placement_run",
+        ["artist_id"],
+    )
+
+
+def downgrade() -> None:
+    # Dropping this table destroys the only record of where moved files came
+    # from. That is correct for a downgrade — the code that reads it is going
+    # away too — but it is worth saying out loud rather than discovering.
+    op.drop_index(
+        "ix_library_placement_run_artist_id",
+        table_name="library_placement_run",
+    )
+    op.drop_index(
+        "ix_library_placement_run_status", table_name="library_placement_run",
+    )
+    op.drop_table("library_placement_run")

alembic/versions/0100_drop_library_placement_run.py

@@ -0,0 +1,109 @@
+"""Drop library_placement_run — the placement reconciler is removed.
+
+Milestone #421 built a sweep that compared each image's `artist_id` to the
+name of the directory its file sat in, and called every mismatch a misplaced
+file. On the operator's library that reported 33,789 of 63,605 images as
+wrongly filed.
+
+That number was an artefact of the comparison, not a fact about the library:
+
+- **97.1%** of it was one artist's own folder spelled differently —
+  `Telepurte/` versus `telepurte/`. Same artist, same art, nothing wrong.
+- Of the 1% that sat in a differently-named folder, querying `ImageProvenance`
+  — which records the post and source each file was actually downloaded from —
+  showed 87 where provenance agreed with the FOLDER and not the record, and 40
+  genuinely posted by several creators. The sweep would have misfiled or
+  arbitrarily picked for roughly 41% of that set.
+
+The system already knows where every file came from. The reconciler inferred
+it from a column and a directory name instead, and manufactured work out of a
+naming convention. Operator's call, 2026-09-21: *"the current system
+consistently records where items are and where they came from this is just
+complicating something works and doesn't need fixing."*
+
+Rule #22 — no legacy to preserve. The table goes with the code.
+
+## What is deliberately kept
+
+`utils.paths.canonical_subdir` stays: new filesystem imports derive their
+directory from the artist's slug, matching what the downloader has always
+done. It is not part of this tool and removing it would be churn for no fix.
+Run 1's 327 moved files (`InsoUwu/` -> `insouwu/`) also stay where they are —
+same artist either way, and the gallery renders them correctly.
+
+Revision ID: 0100
+Revises: 0099
+Create Date: 2026-09-21
+
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects import postgresql
+
+revision: str = "0100"
+down_revision: Union[str, None] = "0099"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.drop_index(
+        "ix_library_placement_run_artist_id",
+        table_name="library_placement_run",
+    )
+    op.drop_index(
+        "ix_library_placement_run_status", table_name="library_placement_run",
+    )
+    op.drop_table("library_placement_run")
+
+
+def downgrade() -> None:
+    # Recreates the table only. The three runs it held (one applied, two
+    # planned-and-never-run) are not restored and are not worth restoring —
+    # the code that reads them is gone.
+    op.create_table(
+        "library_placement_run",
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column(
+            "status", sa.String(length=16), server_default="running",
+            nullable=False,
+        ),
+        sa.Column("artist_id", sa.Integer(), nullable=True),
+        sa.Column(
+            "started_at", sa.DateTime(timezone=True),
+            server_default=sa.text("now()"), nullable=False,
+        ),
+        sa.Column("finished_at", sa.DateTime(timezone=True), nullable=True),
+        sa.Column(
+            "planned_count", sa.Integer(), server_default="0", nullable=False,
+        ),
+        sa.Column(
+            "moved_count", sa.Integer(), server_default="0", nullable=False,
+        ),
+        sa.Column(
+            "refused_count", sa.Integer(), server_default="0", nullable=False,
+        ),
+        sa.Column(
+            "moves", postgresql.JSONB(astext_type=sa.Text()),
+            server_default=sa.text("'[]'::jsonb"), nullable=False,
+        ),
+        sa.Column(
+            "refusals", postgresql.JSONB(astext_type=sa.Text()),
+            server_default=sa.text("'[]'::jsonb"), nullable=False,
+        ),
+        sa.Column("error", sa.Text(), nullable=True),
+        sa.ForeignKeyConstraint(
+            ["artist_id"], ["artist.id"],
+            name="fk_library_placement_run_artist_id", ondelete="SET NULL",
+        ),
+        sa.PrimaryKeyConstraint("id"),
+    )
+    op.create_index(
+        "ix_library_placement_run_status", "library_placement_run", ["status"],
+    )
+    op.create_index(
+        "ix_library_placement_run_artist_id", "library_placement_run",
+        ["artist_id"],
+    )

alembic/versions/0101_clear_failure_state_on_disabled_sources.py

@@ -0,0 +1,66 @@
+"""Clear failure state on sources that are disabled (#4279).
+
+`failing_sources_clause()` now means "enabled AND erroring", so a disabled
+source no longer counts as failing. That fixes what the surfaces REPORT; it
+does not touch what the rows already CARRY, and the rows are the reason the
+operator saw a banner for six days with no way to act on it (lesson #4202 —
+a guard does not undo the value already stored).
+
+## The row this exists for
+
+Ebi77 (source 19): the membership sweep stopped it as `former_patron` at
+02:50 on 2026-09-15 and correctly cleared its failure state. A deep scan was
+armed twenty minutes later — `/backfill` had no `enabled` guard, which this
+release also fixes — and could not complete without access, so the recovery
+sweep stranded it:
+
+    consecutive_failures = 1
+    last_error = "stranded by recovery sweep (no terminal status after time_limit)"
+
+Nothing could clear that. A disabled source is never scheduled, so no
+successful run resets the counter; `SourceService.update` clears failure
+state only on an explicit disable, and the source was already disabled; and
+the card's Retry routes to `/check`, which refuses a disabled source.
+
+## Why every disabled source, not just that one
+
+The clear matches what `SourceService.update` already does when a source is
+disabled through the app — "disable the subs you're not paying for without
+them lingering as failing" — so this brings rows disabled by any OTHER path
+(the membership sweep, a retired platform in 0097) into line with the rows
+disabled by hand. Same shape as 0097: a repair migration reaches the live
+instance on deploy rather than waiting for someone to find the row.
+
+Enabled sources are untouched — a real failure on a live source must keep
+showing.
+
+Revision ID: 0101
+Revises: 0100
+Create Date: 2026-09-21
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+
+revision: str = "0101"
+down_revision: Union[str, None] = "0100"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.execute(
+        "UPDATE source SET last_error = NULL, error_type = NULL, "
+        "consecutive_failures = 0 "
+        "WHERE NOT enabled "
+        "AND (last_error IS NOT NULL OR error_type IS NOT NULL "
+        "     OR consecutive_failures <> 0)"
+    )
+
+
+def downgrade() -> None:
+    # Irreversible by design: the cleared strings and counts are not recorded
+    # anywhere, and restoring a failure state nobody can act on would only
+    # re-create the banner this removes. Rule #22 owes no story backwards.
+    pass

alembic/versions/0102_drop_pixiv_data_and_orphan_credentials.py

@@ -0,0 +1,138 @@
+"""Drop pixiv's ledgers, and delete credentials for platforms FC no longer has.
+
+Milestone #406 step 6 (with issue #3980 folded in). Phase 1 unregistered pixiv
+and the commit alongside this one deleted its client, downloader, ingester and
+models. This removes the data those models described, and the stored secrets of
+every platform that has been retired.
+
+## The two ledger tables
+
+`pixiv_seen_media` and `pixiv_failed_media` are the per-source seen / dead-letter
+ledgers for a downloader that no longer exists. They were created in
+`0089_baseline.py`, so dropping them needs a new revision rather than an edit
+there.
+
+## The credentials
+
+Written as *delete every credential whose platform is not registered* rather
+than as `platform = 'pixiv'`, at the explicit ask in this step's plan. That is
+what makes one migration cover two retirements:
+
+- **pixiv** — a live OAuth refresh token for a service FC no longer talks to.
+- **deviantart** — issue #3980. #3069 retired DeviantArt in code on 2026-08-27
+  and left its stored session behind; seven weeks later it was still there.
+
+And it is the only way either row can go. The credentials UI
+(`subscriptions/SettingsTab.vue`) renders one card per platform returned by
+`/api/platforms`, then looks the credential up by key — so a row whose platform
+is unregistered has no card, no Remove button, and no way for the operator to
+reach it. `CredentialService.list()` would return it; nothing asks.
+
+The registered set is written out literally instead of importing
+`known_platform_keys()`. A migration is a statement about one moment in the
+schema's history: if it imported the live registry, retiring a fifth platform
+in 2027 would silently change what this 2026 revision did on a fresh database.
+The list below is the registry as of 2026-09-21.
+
+## What is deliberately kept
+
+**Every pixiv `Source` row.** The original plan deleted them; the operator's
+call on 2026-09-21 was to keep them, and the reason is that `platform` is
+stored ONLY on `Source` — neither `Post` nor `ImageRecord` carries it. Both
+FKs are `ON DELETE SET NULL`, so a delete would not lose the art, but it would
+drop every pixiv image into the gallery's `__unsourced__` bucket and strip the
+platform chip off every pixiv post. The rows stay disabled (0097) and their
+platform is unregistered, so nothing schedules them, nothing downloads through
+them, and `POST /api/sources` will not make another. Keeping them costs
+nothing and keeps the attribution the milestone's goal — *"the art already
+downloaded from pixiv stays"* — is actually about.
+
+Every pixiv `Post` and `ImageRecord` is likewise untouched.
+
+Revision ID: 0102
+Revises: 0101
+Create Date: 2026-09-21
+
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0102"
+down_revision: Union[str, None] = "0101"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+# services/platforms/__init__.py's PLATFORMS as of this revision. See the
+# docstring for why this is a literal and not an import.
+_REGISTERED_PLATFORMS = ("patreon", "subscribestar", "hentaifoundry", "discord")
+
+
+def upgrade() -> None:
+    op.execute(
+        sa.text(
+            "DELETE FROM credential WHERE platform NOT IN :registered"
+        ).bindparams(
+            sa.bindparam("registered", value=_REGISTERED_PLATFORMS, expanding=True)
+        )
+    )
+    op.drop_index("ix_pixiv_failed_media_source_id", table_name="pixiv_failed_media")
+    op.drop_table("pixiv_failed_media")
+    op.drop_index("ix_pixiv_seen_media_source_id", table_name="pixiv_seen_media")
+    op.drop_table("pixiv_seen_media")
+
+
+def downgrade() -> None:
+    # The tables come back empty, and the credentials do not come back at all:
+    # they were encrypted blobs, this migration does not copy them anywhere,
+    # and restoring a live token for a platform FC cannot talk to would only
+    # re-create the liability. Rule #22 owes no story backwards.
+    op.create_table(
+        "pixiv_seen_media",
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("source_id", sa.Integer(), nullable=False),
+        sa.Column("filehash", sa.String(length=128), nullable=False),
+        sa.Column("url", sa.Text(), nullable=True),
+        sa.Column(
+            "created_at", sa.DateTime(timezone=True),
+            server_default=sa.text("now()"), nullable=False,
+        ),
+        sa.ForeignKeyConstraint(
+            ["source_id"], ["source.id"],
+            name=op.f("fk_pixiv_seen_media_source_id_source"), ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_pixiv_seen_media")),
+        sa.UniqueConstraint(
+            "source_id", "filehash", name="uq_pixiv_seen_media_source_id",
+        ),
+    )
+    op.create_index(
+        op.f("ix_pixiv_seen_media_source_id"), "pixiv_seen_media", ["source_id"],
+        unique=False,
+    )
+    op.create_table(
+        "pixiv_failed_media",
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("source_id", sa.Integer(), nullable=False),
+        sa.Column("filehash", sa.String(length=128), nullable=False),
+        sa.Column("url", sa.Text(), nullable=True),
+        sa.Column("error", sa.Text(), nullable=True),
+        sa.Column("attempts", sa.Integer(), server_default="1", nullable=False),
+        sa.Column(
+            "created_at", sa.DateTime(timezone=True),
+            server_default=sa.text("now()"), nullable=False,
+        ),
+        sa.ForeignKeyConstraint(
+            ["source_id"], ["source.id"],
+            name=op.f("fk_pixiv_failed_media_source_id_source"), ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_pixiv_failed_media")),
+        sa.UniqueConstraint(
+            "source_id", "filehash", name="uq_pixiv_failed_media_source_id",
+        ),
+    )
+    op.create_index(
+        op.f("ix_pixiv_failed_media_source_id"), "pixiv_failed_media", ["source_id"],
+        unique=False,
+    )

alembic/versions/0103_worker_lane_settings.py

@@ -0,0 +1,121 @@
+"""worker_lane — settings-backed slots for each celery lane.
+
+Milestone 422 step 1. One row per lane, holding only what an operator can
+change: how many slots it runs, the ceiling they have set for themselves, and
+whether it consumes its queues at all.
+
+## What is deliberately not a column
+
+**The queues.** They are decided by `celery_app.py`'s `task_routes`, not by
+preference, so a stored copy could contradict the routing table with nothing
+to notice until a queue had no consumer. They live in
+`services/worker_lanes.py`.
+
+**The derived ceiling.** Computed from the container's cgroup limits on every
+read. A row written on a 32GB host and later run in a 4GB container must be
+bounded by the 4GB; a stored ceiling would quietly authorise what the box can
+no longer hold.
+
+## The seeded values
+
+Written out literally rather than imported from `worker_lanes.LANES`. A
+migration is a statement about one moment in the schema's history — if it
+imported the live defaults, changing them in 2027 would silently change what
+this 2026 revision does on a fresh database. The two are allowed to diverge
+afterwards, and that is correct: `LANES` supplies defaults for a lane added
+later, this file records what was seeded today.
+
+    lane              slots  cap  enabled
+    worker                1    4  yes
+    scheduler             1    2  yes
+    maintenance_long      1    2  yes
+    ml                    0    1  NO
+
+One of each, per the operator (2026-09-22: *"that starting value should be one
+of each"*), and far below their own production numbers — worker 8 and ml 2 are
+tuned for their hardware and are not a sane first boot for a stranger.
+
+**ml ships at zero and disabled**, which is milestone 422 step 6's requirement
+arriving early: enabling the lane is what triggers the SigLIP download, and
+rule 164 permits a runtime fetch only for a feature that is "optional and
+clearly off". Seeding it on would make every fresh install reach HuggingFace.
+
+The caps start low on purpose. A cap that begins at the ceiling is a rubber
+stamp; starting at 4/2/2/1 means raising slots within the cap is ordinary and
+raising the cap is a deliberate act.
+
+## Existing installs
+
+Nothing is migrated FROM. The `CELERY_QUEUES` / `CELERY_CONCURRENCY` env vars
+stay exactly as they are and remain the baseline each lane boots at; these
+rows are the adjustment applied on top (step 3). So this migration changes no
+behaviour on a running stack — it only makes the numbers storable.
+
+Revision ID: 0103
+Revises: 0102
+Create Date: 2026-09-22
+
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0103"
+down_revision: Union[str, None] = "0102"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+# (name, slots, slots_cap, enabled) — see the docstring for why these are
+# literals and not an import.
+_SEED = (
+    ("worker", 1, 4, True),
+    ("scheduler", 1, 2, True),
+    ("maintenance_long", 1, 2, True),
+    ("ml", 0, 1, False),
+)
+
+
+def upgrade() -> None:
+    worker_lane = op.create_table(
+        "worker_lane",
+        sa.Column("name", sa.String(length=32), nullable=False),
+        sa.Column("slots", sa.Integer(), nullable=False),
+        sa.Column("slots_cap", sa.Integer(), nullable=False),
+        sa.Column("enabled", sa.Boolean(), nullable=False),
+        sa.Column(
+            "updated_at", sa.DateTime(timezone=True),
+            server_default=sa.text("now()"), nullable=False,
+        ),
+        sa.PrimaryKeyConstraint("name", name=op.f("pk_worker_lane")),
+        # Bare constraint names: Base.metadata's naming convention prepends
+        # ck_worker_lane_, and pre-prefixing doubles it — the defect alembic
+        # 0088 had to rename four constraints for (#3275). op.f() marks these
+        # as already-final so autogenerate does not propose renaming them.
+        sa.CheckConstraint("slots >= 0", name=op.f("ck_worker_lane_slots_non_negative")),
+        sa.CheckConstraint("slots_cap >= 0", name=op.f("ck_worker_lane_cap_non_negative")),
+        # The invariant that makes the cap mean anything, in the database
+        # rather than only in the service: a row violating it is not a
+        # rejected request, it is a lane that step 3's reconcile will drive UP
+        # to a number the operator capped.
+        sa.CheckConstraint("slots <= slots_cap", name=op.f("ck_worker_lane_slots_within_cap")),
+    )
+    # No index beyond the primary key, deliberately — four rows, forever. Same
+    # reasoning as service_seen, and the lesson of #3301, which removed seven
+    # indexes that were write cost buying nothing.
+    op.bulk_insert(
+        worker_lane,
+        [
+            {"name": name, "slots": slots, "slots_cap": cap, "enabled": enabled}
+            for name, slots, cap, enabled in _SEED
+        ],
+    )
+
+
+def downgrade() -> None:
+    # The rows go with the table. They are settings with shipped defaults, not
+    # operator data that predates this revision — a downgrade returns the stack
+    # to reading its concurrency from env, which is where it reads it from
+    # today anyway.
+    op.drop_table("worker_lane")

alembic/versions/0104_worker_lane_autoscale.py

@@ -0,0 +1,51 @@
+"""worker_lane.autoscale — may this lane grow itself?
+
+Milestone 422 step 7. One boolean, defaulting FALSE on every existing row and
+on every new one.
+
+## Why the default is false and not "sensible"
+
+This is the only part of the milestone that acts without anyone watching. The
+manual dial (step 4) and the reconcile (step 3) both do exactly what someone
+asked for; this one decides. Shipping it on would mean every install starts
+with a process that changes its own resource usage based on a heuristic tuned
+against nobody's workload.
+
+Off also makes the failure mode benign: if the signal is wrong, nothing
+happens until an operator opts a lane in, and they opted in while watching.
+
+## Why per lane and not one global switch
+
+The lanes are not alike in what a slot costs. A `worker` slot is a process;
+an `ml` slot is another copy of a ~3.5GB model. A global switch would enable
+growth on a lane whose behaviour under load nobody has observed, and the one
+it would hurt most is the one whose cost is least visible.
+
+Revision ID: 0104
+Revises: 0103
+Create Date: 2026-09-22
+
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0104"
+down_revision: Union[str, None] = "0103"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "worker_lane",
+        sa.Column(
+            "autoscale", sa.Boolean(),
+            server_default=sa.text("false"), nullable=False,
+        ),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("worker_lane", "autoscale")

alembic/versions/0105_worker_lane_one_cap.py

@@ -0,0 +1,133 @@
+"""worker_lane — one number: the cap. `slots`, `enabled` and `autoscale` go.
+
+Milestone 422, reshaped by the operator 2026-09-23:
+
+    "auto should be always on, not a setting, so that idle instances quiet
+    down when not running. the number that is visible and something the user
+    can tweak and manage should be the cap itself the number of running
+    workers is handled by the autoscaling function which is always on."
+
+## What each dropped column was, and why it is not needed
+
+**`slots`** — how many workers the lane should run. That is a MEASUREMENT,
+not a preference: the autoscaler moves the live pool between one and the cap
+according to the backlog, and reads it back from the worker every minute.
+Storing it made it look like something to keep in agreement with the cap,
+which is exactly what the operator had to do.
+
+**`autoscale`** — whether the lane was allowed to size itself. It gated the
+mechanism behind a per-lane opt-in, so a lane nobody enabled simply never
+gave its slots back. Always on now, which is the only way "idle instances
+quiet down" can be true of an install nobody has configured.
+
+**`enabled`** — whether the lane consumes its queues. Derived from `cap > 0`.
+It and `slots = 0` were two spellings of one fact and were free to disagree;
+this migration picks the one an operator can see.
+
+## Why the caps are rewritten rather than preserved
+
+The old defaults were 4 / 2 / 2 / 1, chosen when the number meant "the most
+you may raise SLOTS to" — a bound on a manual control, deliberately loose
+because moving within it was the ordinary act. The number now means "the most
+workers this lane may actually use", which is a different promise, and
+carrying the old figure over would silently quadruple the worker lane on
+every existing install at the moment this deploys.
+
+So every row is reset to the new defaults: **one for each required lane, zero
+for ML.** That loses whatever an operator had set — which is the honest
+trade, because what they set was an answer to a different question. The UI
+now tells a busy lane's operator to raise its cap, which is how the number
+gets back up on an install that needs it.
+
+ML at zero also keeps rule 164's carve-out intact: no consumers, so no model
+download until someone raises the cap.
+
+Revision ID: 0105
+Revises: 0104
+Create Date: 2026-09-23
+
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0105"
+down_revision: Union[str, None] = "0104"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+# (name, cap) — the same values `services/worker_lanes.LANES` declares. Seeded
+# here as literals rather than imported: a migration must describe the schema
+# at ITS point in history, and importing the live table would make this file
+# change meaning every time that table does.
+_CAPS = (
+    ("worker", 1),
+    ("scheduler", 1),
+    ("maintenance_long", 1),
+    ("ml", 0),
+)
+
+
+def upgrade() -> None:
+    # The constraints go first: they name `slots`, so dropping the column out
+    # from under them fails on Postgres.
+    #
+    # `op.f()` around each name, and it is load-bearing. Without it alembic
+    # runs the name through Base.metadata's naming convention, which prepends
+    # `ck_worker_lane_` to a string that already carries it — and the DROP
+    # goes looking for `ck_worker_lane_ck_worker_lane_slots_within_cap`, which
+    # no database has. That is #3275 exactly, from the other direction:
+    # alembic 0088 had to RENAME four constraints created with the same
+    # doubling. Caught here by the integration lane, run 7365.
+    op.drop_constraint(
+        op.f("ck_worker_lane_slots_within_cap"), "worker_lane", type_="check",
+    )
+    op.drop_constraint(
+        op.f("ck_worker_lane_slots_non_negative"), "worker_lane", type_="check",
+    )
+    op.drop_column("worker_lane", "slots")
+    op.drop_column("worker_lane", "enabled")
+    op.drop_column("worker_lane", "autoscale")
+
+    # Reset to the new meaning. See the docstring: the old value answered a
+    # different question, and carrying it over would raise every lane.
+    for name, cap in _CAPS:
+        op.execute(
+            sa.text("UPDATE worker_lane SET slots_cap = :cap WHERE name = :name")
+            .bindparams(cap=cap, name=name)
+        )
+
+    # A lane the old seed never wrote — or one an operator added by hand — is
+    # left alone rather than guessed at. `_rows_by_name` creates any missing
+    # row at the lane's default on first read.
+
+
+def downgrade() -> None:
+    op.add_column(
+        "worker_lane",
+        sa.Column("slots", sa.Integer(), nullable=False, server_default="1"),
+    )
+    op.add_column(
+        "worker_lane",
+        sa.Column(
+            "enabled", sa.Boolean(), nullable=False, server_default=sa.text("true"),
+        ),
+    )
+    op.add_column(
+        "worker_lane",
+        sa.Column(
+            "autoscale", sa.Boolean(), nullable=False, server_default=sa.text("false"),
+        ),
+    )
+    # Restore the pre-0105 invariants. `slots` comes back as 1 everywhere and
+    # the caps are 1/1/1/0, so a lane at cap 0 would violate `slots <= cap` —
+    # hence the clamp before the constraint is added.
+    op.execute(sa.text("UPDATE worker_lane SET slots = 0 WHERE slots_cap = 0"))
+    op.execute(sa.text("UPDATE worker_lane SET enabled = (slots_cap > 0)"))
+    op.create_check_constraint(
+        op.f("ck_worker_lane_slots_non_negative"), "worker_lane", "slots >= 0",
+    )
+    op.create_check_constraint(
+        op.f("ck_worker_lane_slots_within_cap"), "worker_lane", "slots <= slots_cap",
+    )

alembic/versions/0106_prune_phantom_roster_rows.py

@@ -0,0 +1,70 @@
+"""service_seen — delete the roster rows the fixed code can no longer write.
+
+Operator, 2026-09-23: *"clean up the stale service_seen rows"*. They were not
+stale. They were PHANTOMS, written on purpose by code that identified a celery
+worker from the queues it was consuming.
+
+A lane at cap 0 has its consumers cancelled, so it answers `active_queues()`
+with an empty list. The roster grouped on that empty set, wrote it under the
+key `celery:` and rendered `role_display_name(())` as the display name — a row
+called **`Worker ()`**, reported as running, beside the real lane's row going
+stale because nothing updated it any more.
+
+`worker_lanes.lane_for_node` fixes the cause: a worker is attributed by its
+NODE NAME, which survives having no consumers. Nothing will write `celery:`
+again.
+
+## Why a migration and not a retention sweep
+
+Lesson #4202: a guard that refuses to produce a bad value does not undo the
+bad value already stored. The row is the thing that has to change.
+
+And it must be deleted rather than aged out, because the roster deliberately
+NEVER forgets — *"anything that has run at least once stays listed, that is
+what lets a stopped one be noticed rather than simply vanishing"*. A row that
+merely goes quiet is exactly what the roster is for. Only a row that cannot
+correspond to anything real is safe to remove, and `celery:` is precisely
+that: the empty queue set, which no correctly-attributed worker can produce.
+
+## What is deliberately NOT deleted
+
+**Celery rows with a real but unmatched queue set.** A deployment slicing
+`CELERY_QUEUES` differently is supported and its rows are true. It is not this
+migration's business to decide that somebody else's worker is obsolete.
+
+**Agent rows, including a possible `agent:agent` from a build that omitted
+`agent_id`.** Nothing here can tell an abandoned agent id from a second agent
+that is currently down, and deleting a real one would hide a genuinely dead
+GPU agent — the one thing the roster exists to show. If such a row is present
+it needs a person to look at it, not a migration guessing.
+
+Revision ID: 0106
+Revises: 0105
+Create Date: 2026-09-23
+
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0106"
+down_revision: Union[str, None] = "0105"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Exactly the one key the empty queue set produced. Matched literally
+    # rather than by a LIKE or a prefix: `celery:` with nothing after it is
+    # the phantom, and `celery:ml` is a real lane.
+    op.execute(
+        sa.text("DELETE FROM service_seen WHERE key = :key").bindparams(key="celery:")
+    )
+
+
+def downgrade() -> None:
+    # Nothing. The row carried no information — an empty queue set and a
+    # timestamp — and the roster re-learns anything real on its next refresh.
+    # Re-creating it would put a phantom back.
+    pass

alembic/versions/0107_worker_lane_sample.py

@@ -0,0 +1,63 @@
+"""worker_lane_sample — where the sizing sweep leaves what it measured.
+
+Operator, 2026-09-23, on the System tab: *"there is a repull every time this
+page loads — is there a reason this info isn't being tracked in the
+background and stored in some way?"*
+
+`/api/system/workers` ran a full celery inspect on every call — four
+broadcasts on an eleven-second budget — and the page polls it every fifteen
+seconds. `size_worker_lanes` was already inspecting on a timer to decide pool
+sizes, computing exactly these numbers and discarding them. This table is
+where they land instead, and the endpoint becomes a plain read.
+
+## Why a new table rather than columns on `worker_lane`
+
+`worker_lane` holds the one number an operator sets. Putting a measurement
+beside it is the mistake alembic 0105 undid: `slots` sat next to `slots_cap`,
+and a measurement next to a preference reads as a second preference.
+
+No backfill. A row appears when the sweep first runs (within its period), and
+until then the lane reads as not-yet-measured, which is true.
+
+Revision ID: 0107
+Revises: 0106
+Create Date: 2026-09-23
+
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0107"
+down_revision: Union[str, None] = "0106"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "worker_lane_sample",
+        sa.Column("lane", sa.String(length=32), primary_key=True),
+        # Nullable=False with no server_default: the sweep writes every column
+        # on every upsert, so a row only ever exists complete.
+        sa.Column("present", sa.Boolean(), nullable=False),
+        sa.Column("replicas", sa.Integer(), nullable=False),
+        # Nullable on purpose — unknown, never zero. A worker that answered
+        # without reporting its pool, and a queue the broker did not answer
+        # for, must not be summed as empty.
+        sa.Column("pool", sa.Integer(), nullable=True),
+        sa.Column("active", sa.Integer(), nullable=False),
+        sa.Column("reserved", sa.Integer(), nullable=False),
+        sa.Column("queue_depth", sa.Integer(), nullable=True),
+        sa.Column(
+            "measured_at",
+            sa.DateTime(timezone=True),
+            nullable=False,
+            server_default=sa.func.now(),
+        ),
+    )
+
+
+def downgrade() -> None:
+    op.drop_table("worker_lane_sample")

alembic/versions/0108_download_revisit_days.py

@@ -0,0 +1,54 @@
+"""download_revisit_days — how far back a tick keeps looking for EDITED posts.
+
+Operator, 2026-09-23, pointing at a Floppystack post: *"this post has been
+updated as he implements hot fixes — any chance we have a way to scan for or
+see updated posts so we can update ours to match and pull the new attachments
+and pictures etc."*
+
+A tick stopped after 20 contiguous already-have-it items. That is the right
+instinct and the wrong unit: a post edited three days after publication sits
+well below twenty seen items, so the walk turned around before reaching it. The
+walk now needs BOTH a run of seen items and a post older than this many days
+before it stops.
+
+A settings row rather than a constant (rule 25) because the right window is a
+property of the CREATOR, not of FabledCurator — one artist appends hotfix
+builds for a fortnight, another never touches a post again. 0 turns the revisit
+off entirely and restores the pure count early-out.
+
+30 days is the operator's own number, 2026-09-23.
+
+Revision ID: 0108
+Revises: 0107
+Create Date: 2026-09-23
+
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0108"
+down_revision: Union[str, None] = "0107"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # server_default so the existing single settings row gets the window without
+    # a data migration — and so an install that predates this column reads 30
+    # rather than 0. 0 is a real, meaningful value here (revisit off), so the
+    # column must never be allowed to arrive at it by omission.
+    op.add_column(
+        "import_settings",
+        sa.Column(
+            "download_revisit_days",
+            sa.Integer(),
+            nullable=False,
+            server_default="30",
+        ),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("import_settings", "download_revisit_days")

alembic/versions/0109_discord_link_auto.py

@@ -0,0 +1,44 @@
+"""discord_link_auto — whether FC links a conclusive pair without asking.
+
+Operator, 2026-09-24: *"I don't want this to be manual that defeats the
+convenience that I'm going for."*
+
+Confirm-only was right while every signal was circumstantial. Time proximity
+and a body that mentions Discord can never be more than suggestive, so asking
+was the honest response. A shared working name is different in kind: when the
+creator's own name for a piece appears in exactly these two posts and nowhere
+else in their library, there is nothing left for the operator to adjudicate,
+and asking is just a chore FC invented for them.
+
+Defaults ON, which is a real change of posture and deliberate. It only governs
+the conclusive band — weaker evidence still queues — and a link is a row the
+operator can dismiss, so the reversal is a click rather than a migration.
+
+Revision ID: 0109
+Revises: 0108
+Create Date: 2026-09-24
+
+"""
+import sqlalchemy as sa
+from alembic import op
+
+revision = "0109"
+down_revision = "0108"
+branch_labels = None
+depends_on = None
+
+
+def upgrade():
+    op.add_column(
+        "import_settings",
+        sa.Column(
+            "discord_link_auto",
+            sa.Boolean(),
+            nullable=False,
+            server_default=sa.text("true"),
+        ),
+    )
+
+
+def downgrade():
+    op.drop_column("import_settings", "discord_link_auto")

alembic/versions/0110_unified_post_card.py

@@ -0,0 +1,59 @@
+"""The unified post card — fold window, family window, and who linked a pair.
+
+Milestone 388, #4402 and #4401. A Patreon teaser's card shows the Discord drop
+it announced, and the rest of that piece's variants, by REFERENCE: nothing is
+absorbed, nothing changes owner, and every Discord post keeps its own place.
+
+Three columns:
+
+* `import_settings.discord_link_fold_hours` — a linked drop leaves the feed
+  only when it is this close to its teaser (the same release, shown twice).
+* `import_settings.discord_family_window_days` — how far from the teaser the
+  card reaches for variants. 60 is measured: named families spread up to 44
+  days on artist 8, every collision found over 500.
+* `post_association.linked_by` — "fc" or "operator", so a link FC made by
+  itself can say so on the card and offer the undo the operator asked for.
+
+Revision ID: 0110
+Revises: 0109
+Create Date: 2026-09-24
+
+"""
+import sqlalchemy as sa
+from alembic import op
+
+revision = "0110"
+down_revision = "0109"
+branch_labels = None
+depends_on = None
+
+
+def upgrade():
+    op.add_column(
+        "import_settings",
+        sa.Column(
+            "discord_link_fold_hours",
+            sa.Float(),
+            nullable=False,
+            server_default=sa.text("24"),
+        ),
+    )
+    op.add_column(
+        "import_settings",
+        sa.Column(
+            "discord_family_window_days",
+            sa.Float(),
+            nullable=False,
+            server_default=sa.text("60"),
+        ),
+    )
+    op.add_column(
+        "post_association",
+        sa.Column("linked_by", sa.String(length=16), nullable=True),
+    )
+
+
+def downgrade():
+    op.drop_column("post_association", "linked_by")
+    op.drop_column("import_settings", "discord_family_window_days")
+    op.drop_column("import_settings", "discord_link_fold_hours")

alembic/versions/0111_discord_native_ledger.py

@@ -0,0 +1,75 @@
+"""Discord native ingester ledgers — seen and dead-letter, per source.
+
+Milestone 428, #4415. Discord moves off gallery-dl onto the native core, which
+keeps its memory of what a source has already fetched in these two tables
+instead of gallery-dl's archive. Same shape as the SubscribeStar pair.
+
+Revision ID: 0111
+Revises: 0110
+Create Date: 2026-09-24
+
+"""
+import sqlalchemy as sa
+from alembic import op
+
+revision = "0111"
+down_revision = "0110"
+branch_labels = None
+depends_on = None
+
+
+def upgrade():
+    op.create_table(
+        "discord_seen_media",
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("source_id", sa.Integer(), nullable=False),
+        sa.Column("filehash", sa.String(length=128), nullable=False),
+        sa.Column("post_id", sa.String(length=64), nullable=True),
+        sa.Column(
+            "seen_at", sa.DateTime(timezone=True),
+            server_default=sa.text("now()"), nullable=False,
+        ),
+        sa.ForeignKeyConstraint(
+            ["source_id"], ["source.id"],
+            name=op.f("fk_discord_seen_media_source_id_source"), ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_discord_seen_media")),
+        sa.UniqueConstraint("source_id", "filehash", name="uq_discord_seen_media_source_id"),
+    )
+    op.create_index(
+        op.f("ix_discord_seen_media_source_id"), "discord_seen_media", ["source_id"],
+    )
+    op.create_table(
+        "discord_failed_media",
+        sa.Column("id", sa.Integer(), nullable=False),
+        sa.Column("source_id", sa.Integer(), nullable=False),
+        sa.Column("filehash", sa.String(length=128), nullable=False),
+        sa.Column("attempts", sa.Integer(), server_default="1", nullable=False),
+        sa.Column("last_error", sa.Text(), nullable=True),
+        sa.Column(
+            "first_failed_at", sa.DateTime(timezone=True),
+            server_default=sa.text("now()"), nullable=False,
+        ),
+        sa.Column(
+            "last_failed_at", sa.DateTime(timezone=True),
+            server_default=sa.text("now()"), nullable=False,
+        ),
+        sa.ForeignKeyConstraint(
+            ["source_id"], ["source.id"],
+            name=op.f("fk_discord_failed_media_source_id_source"), ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id", name=op.f("pk_discord_failed_media")),
+        sa.UniqueConstraint(
+            "source_id", "filehash", name="uq_discord_failed_media_source_id",
+        ),
+    )
+    op.create_index(
+        op.f("ix_discord_failed_media_source_id"), "discord_failed_media", ["source_id"],
+    )
+
+
+def downgrade():
+    op.drop_index(op.f("ix_discord_failed_media_source_id"), table_name="discord_failed_media")
+    op.drop_table("discord_failed_media")
+    op.drop_index(op.f("ix_discord_seen_media_source_id"), table_name="discord_seen_media")
+    op.drop_table("discord_seen_media")

backend/app/api/__init__.py

@@ -41,6 +41,7 @@ def all_blueprints() -> list[Blueprint]:
     from .system_health import system_health_bp
     from .tags import tags_bp
     from .thumbnails import thumbnails_bp
+    from .workers import workers_bp
     return [
         api_bp,
         attachments_bp,
@@ -52,6 +53,7 @@ def all_blueprints() -> list[Blueprint]:
         showcase_bp,
         settings_bp,
         system_activity_bp,
+        workers_bp,
         system_health_bp,
         system_backup_bp,
         admin_bp,

backend/app/api/admin.py

@@ -475,6 +475,20 @@ async def trigger_reclaim_attachments():
     return _queued(async_result)
 
 
+@admin_bp.route("/maintenance/repair-discord-downloads", methods=["POST"])
+async def trigger_repair_discord_downloads():
+    """Clean re-download of the Discord files broken by the `None` naming
+    (#3999). Body {"dry_run": bool}; dry_run is the DEFAULT, because the apply
+    deletes files and makes gallery-dl forget every Discord download. Returns the
+    Celery task id — poll /maintenance/task-result/<id> for the summary."""
+    from ..tasks.admin import repair_discord_downloads_task
+
+    body = await request.get_json(silent=True) or {}
+    dry_run = bool(body.get("dry_run", True))
+    async_result = repair_discord_downloads_task.delay(dry_run=dry_run)
+    return _queued(async_result)
+
+
 @admin_bp.route("/maintenance/dedup-videos", methods=["POST"])
 async def trigger_dedup_videos():
     """Tier-1 video dedup (#871). Body {"dry_run": bool}: dry_run=true previews

backend/app/api/artists.py

@@ -65,6 +65,16 @@ async def autocomplete():
     ])
 
 
+@artists_bp.route("/names", methods=["GET"])
+async def names():
+    """Every artist, id + name + slug, alphabetical. For filter pickers that
+    list artists before anything is typed; `autocomplete` deliberately returns
+    nothing for an empty query."""
+    async with get_session() as session:
+        rows = await ArtistService(session).all_names()
+    return jsonify([{"id": i, "name": n, "slug": s} for i, n, s in rows])
+
+
 @artists_bp.route("/directory", methods=["GET"])
 async def directory():
     """FC-3f: cursor-paginated artists directory.

backend/app/api/extension.py

@@ -19,6 +19,7 @@ from ..models import AppSetting
 from ..services.extension_service import (
     ExtensionService,
     InvalidUrlError,
+    UnknownArtistError,
     UnknownPlatformError,
 )
 from ..services.source_service import KNOWN_PLATFORMS
@@ -87,10 +88,16 @@ async def probe_source():
     url = (request.args.get("url") or "").strip()
     if not url:
         return _bad("invalid_body", detail="url query parameter is required")
+    from .credentials import _get_crypto
+
     async with get_session() as session:
         if not await _ext_key_required(session):
             return _bad("unauthorized", status=401)
-        result = await ExtensionService(session).probe(url)
+        # crypto lets a Discord probe name the server and channel with the
+        # stored token; every other platform ignores it.
+        result = await ExtensionService(session, _get_crypto()).probe(
+            url, names=request.args.get("names") in ("1", "true"),
+        )
     return jsonify(result)
 
 
@@ -102,6 +109,18 @@ async def quick_add_source():
     url = body.get("url")
     if not isinstance(url, str) or not url.strip():
         return _bad("invalid_body", detail="url is required")
+    # Optional: connect the new source to an existing artist (artist_id) or to
+    # the artist of that name (artist_name). A Discord channel names no
+    # creator, so the extension's Add panel always sends one of them.
+    artist_id = body.get("artist_id")
+    if artist_id is not None and (isinstance(artist_id, bool) or not isinstance(artist_id, int)):
+        return _bad("invalid_body", detail="artist_id must be an integer")
+    artist_name = body.get("artist_name")
+    if artist_name is not None and not isinstance(artist_name, str):
+        return _bad("invalid_body", detail="artist_name must be a string")
+    # Patreon is canon: adding a Patreon source to an existing artist can take
+    # the creator's Patreon display name (name only; the slug never moves).
+    use_platform_name = body.get("use_platform_name") is True
 
     from .credentials import _get_crypto
 
@@ -109,9 +128,14 @@ async def quick_add_source():
         if not await _ext_key_required(session):
             return _bad("unauthorized", status=401)
         try:
-            # crypto lets a pixiv add resolve the artist's display name via the
-            # stored OAuth token (else it falls back to the numeric id). #130.
-            result = await ExtensionService(session, _get_crypto()).quick_add_source(url)
+            # crypto lets an add resolve the artist's display name via the
+            # stored credential (else it falls back to the URL handle). #130.
+            result = await ExtensionService(session, _get_crypto()).quick_add_source(
+                url, artist_id=artist_id, artist_name=artist_name,
+                use_platform_name=use_platform_name,
+            )
+        except UnknownArtistError as exc:
+            return _bad("not_found", detail=str(exc), status=404)
         except UnknownPlatformError as exc:
             return _bad(
                 "unknown_platform",

backend/app/api/gpu.py

@@ -245,6 +245,29 @@ async def errors_recover(image_id: int):
 
 # --- Agent (bearer token): lease / submit / heartbeat / fail ------------
 
+
+def _accel_detail(body: dict) -> dict:
+    """The agent's own report of which runtime got the GPU, kept on its roster
+    row so the System view can call a CPU-bound agent degraded (#4410).
+
+    Only a dict of {runtime: {device, error?}} is kept, and each value is
+    reduced to those two short strings: this is written on every lease, by a
+    client the server does not control. An agent that sends nothing (an
+    older build) simply has no `accel`, which reads as not-yet-reported.
+    """
+    raw = body.get("accel")
+    if not isinstance(raw, dict):
+        return {}
+    accel = {}
+    for name, entry in list(raw.items())[:4]:
+        if not isinstance(entry, dict):
+            continue
+        clean = {"device": str(entry.get("device") or "")[:16]}
+        if entry.get("error"):
+            clean["error"] = str(entry["error"])[:200]
+        accel[str(name)[:16]] = clean
+    return {"accel": accel} if accel else {}
+
 @gpu_bp.route("/jobs/lease", methods=["POST"])
 async def lease():
     body = await request.get_json(silent=True) or {}
@@ -267,7 +290,10 @@ async def lease():
             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)},
+            details={
+                "agent_id": agent_id, "last_call": "lease", "leased": len(jobs),
+                **_accel_detail(body),
+            },
         )
         ml = await MLSettings.load(session)
         # image rows for url/mime in one shot
@@ -347,7 +373,10 @@ async def heartbeat():
             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},
+            details={
+                "agent_id": agent_id, "last_call": "heartbeat", "extended": n,
+                **_accel_detail(body),
+            },
         )
         await session.commit()
     return jsonify({"extended": n})

backend/app/api/posts.py

@@ -207,6 +207,8 @@ async def rescan_associations():
     be within the window to exist at all); this is the button for a first run
     over a library that predates the feature."""
     async with get_session() as session:
-        result = await association_rescan(session)
+        # full=True: the button reaches the whole history, which the hourly
+        # sweep's 48-hour horizon never does.
+        result = await association_rescan(session, full=True)
         await session.commit()
     return jsonify(result)

backend/app/api/settings.py

@@ -36,6 +36,7 @@ _EDITABLE_FIELDS = (
     "download_validate_files",
     "download_schedule_default_seconds",
     "download_event_retention_days",
+    "download_revisit_days",
     "download_failure_warning_threshold",
     "series_suggest_enabled",
     "series_suggest_threshold",
@@ -43,6 +44,9 @@ _EDITABLE_FIELDS = (
     "discord_link_enabled",
     "discord_link_threshold",
     "discord_link_window_hours",
+    "discord_link_auto",
+    "discord_link_fold_hours",
+    "discord_family_window_days",
     "extdl_mega_enabled",
     "extdl_gdrive_enabled",
     "extdl_mediafire_enabled",
@@ -113,6 +117,12 @@ async def update_import_settings():
         v = body["download_schedule_default_seconds"]
         if not isinstance(v, int) or isinstance(v, bool) or v < 60 or v > 86400:
             return _bad_int("download_schedule_default_seconds", 60, 86400)
+    # 0 is a real value (revisit off), so the floor is 0, not 1 — and the
+    # ceiling is a year, past which a "tick" is a backfill wearing a hat.
+    if "download_revisit_days" in body:
+        v = body["download_revisit_days"]
+        if not isinstance(v, int) or isinstance(v, bool) or v < 0 or v > 365:
+            return _bad_int("download_revisit_days", 0, 365)
     if "download_event_retention_days" in body:
         v = body["download_event_retention_days"]
         if not isinstance(v, int) or isinstance(v, bool) or v < 1 or v > 3650:
@@ -158,6 +168,10 @@ async def update_import_settings():
         body["discord_link_enabled"], bool
     ):
         return jsonify({"error": "discord_link_enabled must be a boolean"}), 400
+    if "discord_link_auto" in body and not isinstance(
+        body["discord_link_auto"], bool
+    ):
+        return jsonify({"error": "discord_link_auto must be a boolean"}), 400
     if "discord_link_threshold" in body:
         v = body["discord_link_threshold"]
         if not isinstance(v, (int, float)) or isinstance(v, bool) or v < 0 or v > 1:
@@ -170,6 +184,12 @@ async def update_import_settings():
             return jsonify(
                 {"error": "discord_link_window_hours must be a positive number"}
             ), 400
+    # Zero is meaningful for both: fold nothing, or reference no variants.
+    for key in ("discord_link_fold_hours", "discord_family_window_days"):
+        if key in body:
+            v = body[key]
+            if not isinstance(v, (int, float)) or isinstance(v, bool) or v < 0:
+                return jsonify({"error": f"{key} must be a number >= 0"}), 400
     if "wip_title_tagging_enabled" in body and not isinstance(
         body["wip_title_tagging_enabled"], bool
     ):

backend/app/api/sources.py

@@ -201,6 +201,16 @@ async def set_backfill(source_id: int):
             rec = await SourceService(session).get(source_id)
             if rec is None:
                 return _bad("not_found", status=404)
+            # A disabled source must not be armable for a deep walk — the same
+            # rule /check has carried all along (see `source_disabled` below).
+            # Arming one anyway is how #4279 happened: the membership sweep had
+            # stopped Ebi77 as `former_patron`, a deep scan was armed twenty
+            # minutes later, the walk could not complete without access, and
+            # the recovery sweep stranded it with a failure count no surface
+            # could clear — a disabled source is never scheduled again, and
+            # Retry routes to /check, which refuses it.
+            if not rec.enabled:
+                return _bad("source_disabled", detail="enable the source first")
             native = uses_native_ingester(rec.platform)
             if native:
                 cred = CredentialService(session, _get_crypto())

backend/app/api/system_activity.py

@@ -21,18 +21,22 @@ from ..config import get_config
 from ..extensions import get_session
 from ..models import TaskRun
 from ..services.scheduler_service import scheduler_status
+from ..services.worker_lanes import LANES
 
 system_activity_bp = Blueprint(
     "system_activity", __name__, url_prefix="/api/system/activity",
 )
 
-# Canonical queue order — must match celery_app.task_routes. UI renders
-# in this order; queues with no LLEN response show as null rather than
-# absent.
-_QUEUE_NAMES = (
-    "default", "import", "thumbnail", "ml",
-    "download", "scan", "maintenance", "maintenance_long",
-)
+# Every queue, grouped by the lane that consumes it. DERIVED from
+# `worker_lanes.LANES` (milestone 422 step 1) rather than written out:
+# this was a hand-kept third copy of "which queues exist", alongside
+# celery_app.task_routes and service_roster.ROLE_NAMES, and its own comment
+# admitted the coupling — "must match celery_app.task_routes".
+#
+# The rendered ORDER changes with this: lane order rather than the previous
+# hand-chosen one. That is the better grouping for a lane-oriented UI, and
+# queues with no LLEN response still show as null rather than absent.
+_QUEUE_NAMES = tuple(q for lane in LANES for q in lane.queues)
 
 # Cache module-level so all requests share the cache between polls.
 # Tests can reset via direct dict mutation if needed.

backend/app/api/system_health.py

@@ -26,7 +26,6 @@ a true statement.
 from __future__ import annotations
 
 import asyncio
-import logging
 import time
 from datetime import UTC, datetime
 
@@ -36,9 +35,7 @@ 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__)
+from ..services.worker_lanes import SWEEP_PERIOD_SECONDS
 
 system_health_bp = Blueprint("system_health", __name__, url_prefix="/api/system")
 
@@ -53,14 +50,36 @@ system_health_bp = Blueprint("system_health", __name__, url_prefix="/api/system"
 STALE_AFTER_SECONDS = 90
 DOWN_AFTER_SECONDS = 300
 
+# The celery roster is written by `size_worker_lanes` and by nothing else, so
+# these thresholds are only meaningful against ITS cadence. Asserted at import
+# rather than left to a reader, because this is precisely the comparison that
+# was never made for the GPU agent: its lease poll backed off to 900s while
+# the roster called it stopped at 300s, and both numbers were individually
+# correct, in different directions, in different files (lesson #4355).
+#
+# Two clear sweeps before a part is even called STALE. One missed tick is
+# routine — the sweep rides the maintenance queue and does an inspect that can
+# take eleven seconds — and must not turn the page yellow.
+_SWEEPS_BEFORE_STALE = 2
+assert STALE_AFTER_SECONDS >= SWEEP_PERIOD_SECONDS * _SWEEPS_BEFORE_STALE, (
+    f"a {SWEEP_PERIOD_SECONDS}s sweep cannot keep a roster fresh against a "
+    f"{STALE_AFTER_SECONDS}s stale threshold: raise the threshold or shorten "
+    f"the sweep"
+)
+
 # 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"
+# Checking in, but working at a fraction of its speed: a GPU agent whose
+# runtimes fell back to the CPU (#4410). Below stale — a part that may have
+# stopped is the more urgent question — and above unknown, because this one
+# IS known to be wrong.
+_DEGRADED = "degraded"
 
 # Worst-first, so an overall verdict is just the max.
-_SEVERITY = {_OK: 0, _UNKNOWN: 1, _STALE: 2, _DOWN: 3}
+_SEVERITY = {_OK: 0, _UNKNOWN: 1, _DEGRADED: 2, _STALE: 3, _DOWN: 4}
 
 
 def _age_state(age_seconds: float) -> str:
@@ -86,6 +105,39 @@ def _describe_learned(name: str, state: str, age: float, details: dict) -> str:
     return f"{name} has not checked in for {ago} — treat it as stopped"
 
 
+def _cpu_runtimes(details: dict) -> list[str]:
+    """The runtimes an agent reported as NOT on the GPU, with why.
+
+    Both torch and onnxruntime fall back to the CPU without raising, so an
+    agent in that state leases, works and checks in exactly like a healthy
+    one. On 2026-09-24 one had been doing so since a driver update left a
+    stale CDI spec; the only sign was a line in the agent's own log.
+    """
+    accel = details.get("accel")
+    if not isinstance(accel, dict):
+        return []
+    out = []
+    for name, entry in sorted(accel.items()):
+        if not isinstance(entry, dict) or entry.get("device") == "cuda":
+            continue
+        why = entry.get("error") or entry.get("device") or "unknown"
+        out.append(f"{name} ({why})")
+    return out
+
+
+def _learned_state(name: str, state: str, age: float, details: dict) -> tuple[str, str]:
+    """A roster row's state and its sentence, degraded included."""
+    if state == _OK:
+        cpu = _cpu_runtimes(details)
+        if cpu:
+            return _DEGRADED, (
+                f"{name} is running on the CPU — not on the GPU: {'; '.join(cpu)}. "
+                "After a driver update, regenerate the agent host's CDI spec "
+                "(agent README)."
+            )
+    return state, _describe_learned(name, state, age, details)
+
+
 async def _probe_postgres(session) -> dict:
     started = time.monotonic()
     try:
@@ -151,26 +203,25 @@ async def system_health():
         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)
-
+            # A PURE READ since 2026-09-23. This used to refresh the celery
+            # roster here, rate-limited to once per 20s — so the roster only
+            # advanced while somebody had a browser open, and a broadcast rode
+            # on a request. `size_worker_lanes` writes it now, on a timer, and
+            # the assertion below is what keeps that cadence honest.
             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)
+                state, detail = _learned_state(
+                    row.display_name, _age_state(age), age, row.details or {},
+                )
                 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 {}),
+                    "detail": detail,
                     "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"},

backend/app/api/workers.py

@@ -0,0 +1,170 @@
+"""Worker lanes: what each is doing, and the dial that changes it.
+
+Milestone 422 step 2. The write half of a surface `api/system_activity.py`
+only reads.
+
+## Why this is a separate blueprint
+
+`system_activity` says in its own first line that it is read-only, and it
+answers a different question: its `/workers` is keyed on celery HOSTNAME and
+reports which nodes answered. That stays as it is — the existing
+SystemActivityTab consumes it.
+
+This is keyed on LANE, joins the stored cap to the live pool, and accepts
+writes. Two endpoints answering "which celery processes exist" and "how much
+work is each lane allowed to do" are not the same endpoint, and folding the
+second into the first would make a read-only module a write one.
+"""
+
+from __future__ import annotations
+
+from datetime import UTC, datetime
+from functools import partial
+
+from quart import Blueprint, current_app, jsonify, request
+
+from ..extensions import get_session
+from ..services.worker_control import (
+    LaneUpdateRefused,
+    lane_settings,
+    lane_view,
+    push_lane_cap,
+    store_lane_cap,
+)
+from ..services.worker_lanes import (
+    LANES_BY_NAME,
+    SWEEP_PERIOD_SECONDS,
+    Lane,
+    derived_ceiling,
+)
+from ._responses import error_response as _bad
+
+workers_bp = Blueprint("workers", __name__, url_prefix="/api/system/workers")
+
+
+@workers_bp.route("", methods=["GET"])
+async def list_lanes():
+    """Every lane: its cap, the ceiling above it, and what is live.
+
+    Response: {lanes: [...], fetched_at: iso8601}
+
+    One database read, and NO broker call. Operator, 2026-09-23: *"there is a
+    repull every time this page loads — is there a reason this info isn't
+    being tracked in the background and stored in some way?"*
+
+    It used to inspect the broker here, four broadcasts on an eleven-second
+    budget, four times a minute per open tab — while `size_worker_lanes` was
+    already inspecting on a timer and discarding the same numbers. The sweep
+    stores them now (`worker_lane_sample`) and this reads them.
+
+    So the live figures are up to `SWEEP_PERIOD_SECONDS` old, and each lane
+    carries the `measured_at` that says so. `sweep_period_seconds` is returned
+    alongside, so the UI can explain the age without hard-coding the cadence
+    in a second place.
+    """
+    async with get_session() as session:
+        settings = await lane_settings(session)
+    return jsonify({
+        "lanes": lane_view(settings),
+        "fetched_at": datetime.now(UTC).isoformat(),
+        "sweep_period_seconds": SWEEP_PERIOD_SECONDS,
+    })
+
+
+@workers_bp.route("/<name>", methods=["POST"])
+async def update_lane(name: str):
+    """Set a lane's cap. Stores it, answers, and makes the lane follow after.
+
+    ONE field, since 2026-09-23. It used to take `slots`, `slots_cap`,
+    `enabled` and `autoscale`; how many workers are running is now a
+    measurement the sizing pass owns, and `enabled` is `cap > 0`.
+
+    ## The reply does not wait for the lane
+
+    Operator, 2026-09-23: *"when the number is changed the change should be
+    queued so that it isn't blocking of the webui or the system itself. we
+    shouldn't have to wait for the validation live."*
+
+    So the request does exactly one thing that can be slow — a row update —
+    and hands the broker work to a background task. Turning a lane off is
+    four `cancel_consumer` messages and a resize; lowering a cap is an
+    `inspect` on an eleven-second budget. Both used to happen between the
+    click and the response, with the stepper disabled the whole time.
+
+    Nothing is lost by not waiting: the cap in the database is what the
+    system obeys, the sizing pass re-reads it every minute, and the table
+    polls, so the live columns catch up on their own. If the web process dies
+    before the background task runs, that sweep is the backstop — which is
+    the same guarantee the awaited version had, since a push could fail
+    there too.
+
+    Refusals still happen inline, because they are decided from the value and
+    the machine's ceiling alone and never touch the broker:
+
+    * **400** — the value is not allowed (negative, or above what this
+      container can hold). Nothing was stored. The body carries `detail`,
+      which is the sentence the UI shows; a refused control with no reason
+      reads as a bug.
+    """
+    lane = LANES_BY_NAME.get(name)
+    if lane is None:
+        return _bad("unknown_lane", detail=name, known=sorted(LANES_BY_NAME))
+
+    body = await request.get_json()
+    if not isinstance(body, dict):
+        return _bad("invalid_body", detail="body must be a JSON object")
+
+    if "slots_cap" not in body:
+        return _bad("invalid_body", detail="give slots_cap")
+    value = body["slots_cap"]
+    # Rejected rather than coerced: `True` is an int in Python, and silently
+    # reading it as a cap of 1 would be a control that appears to work and
+    # sets something nobody asked for.
+    if not isinstance(value, int) or isinstance(value, bool):
+        return _bad("invalid_body", detail="slots_cap must be an integer")
+
+    # Store, close the session, THEN hand off. The session must not be held
+    # across broker work — that is what made this page block the whole site
+    # (see `worker_control.LaneSettings`) — and now the request does not wait
+    # for that work either.
+    async with get_session() as session:
+        try:
+            was_cap = await store_lane_cap(session, lane, value)
+        except LaneUpdateRefused as exc:
+            return _bad("refused", detail=str(exc))
+
+    _schedule_push(lane, value, was_cap)
+
+    return jsonify({
+        "name": lane.name,
+        "slots_cap": value,
+        "ceiling": derived_ceiling(lane),
+        "enabled": value > 0,
+        # The value is stored; the live lane is being told separately. The UI
+        # patches its row from this and lets the next poll bring the live
+        # columns, rather than refetching and paying for an inspect it just
+        # avoided.
+        "queued": True,
+        # Raising the cap off zero is what downloads the model (step 6), and
+        # the background task does it. Reported here so the UI can say a
+        # download has started rather than leaving the operator to wonder why
+        # a lane they just turned on is busy.
+        "fetching_models": value > 0 and was_cap == 0 and bool(lane.models),
+    })
+
+
+def _schedule_push(lane: Lane, slots_cap: int, was_cap: int) -> None:
+    """Run the live push after the response has gone out.
+
+    A seam, not an abstraction: it is one call, and it exists so the tests can
+    hold the push still — a background task that outlived a test's patches
+    would reach the real broker during teardown.
+
+    Quart tracks the task on the app and awaits it at shutdown, so an
+    in-flight push survives a graceful restart. `partial` rather than passing
+    `was_cap=` through `add_background_task`, so nothing depends on how that
+    forwards keyword arguments.
+    """
+    current_app.add_background_task(
+        partial(push_lane_cap, lane, slots_cap, was_cap=was_cap)
+    )

backend/app/beat_scheduler.py

@@ -0,0 +1,77 @@
+"""Celery beat that remembers when each job last ran — from task_run, not a file.
+
+Celery's default PersistentScheduler keeps its memory in a shelve file in the
+working directory. Nothing mounts that directory, so every container recreate
+forgets it, and a scheduler that remembers nothing seeds every entry with
+`last_run_at = now`: each job waits a FULL interval after startup. A daily job
+therefore needs 24 hours without a redeploy to fire. Since the one-container
+image (172e33d) every redeploy restarts beat, and on 2026-09-24 no daily or
+weekly job had run since the 21st (#4408).
+
+task_run already records every task that starts (celery_signals), indexed on
+(task_name, started_at DESC), and prune_task_runs keeps the newest row of each
+task however old it is. So on startup each entry takes its last_run_at from
+there:
+  - a job that is overdue runs at once;
+  - a job that is not due waits only the remainder of its interval;
+  - a job that has never run is due now.
+
+Beat keeps last_run_at in memory from then on, as the default scheduler does;
+only the startup seed changes. If the database cannot be read, the entries keep
+Celery's own default rather than beat failing to start.
+"""
+
+from __future__ import annotations
+
+import logging
+from datetime import UTC, datetime
+
+from celery.beat import Scheduler
+from sqlalchemy import func, select
+
+log = logging.getLogger(__name__)
+
+# Seed for a job with no recorded run: far enough back that any interval or
+# crontab reads as due.
+NEVER = datetime(2000, 1, 1, tzinfo=UTC)
+
+
+def last_runs(session, task_names: list[str]) -> dict[str, datetime]:
+    """The latest recorded start of each task, by task name."""
+    from .models import TaskRun
+
+    if not task_names:
+        return {}
+    rows = session.execute(
+        select(TaskRun.task_name, func.max(TaskRun.started_at))
+        .where(TaskRun.task_name.in_(sorted(set(task_names))))
+        .group_by(TaskRun.task_name)
+    ).all()
+    return dict(rows)
+
+
+def seed(entries, last: dict[str, datetime]) -> None:
+    """Set each entry's last_run_at from `last`; a task with none is due now."""
+    for entry in entries:
+        entry.last_run_at = last.get(entry.task, NEVER)
+
+
+class TaskRunScheduler(Scheduler):
+    """An in-memory beat seeded from task_run history at startup."""
+
+    def setup_schedule(self):
+        super().setup_schedule()
+        try:
+            from .tasks._sync_engine import sync_session_factory
+
+            with sync_session_factory()() as session:
+                last = last_runs(session, [e.task for e in self.schedule.values()])
+        except Exception:
+            log.exception("beat: could not read task_run; every job waits a full interval")
+            return
+        seed(self.schedule.values(), last)
+        due = sum(1 for e in self.schedule.values() if e.is_due()[0])
+        log.info(
+            "beat: seeded %d job(s) from task_run, %d due now",
+            len(self.schedule), due,
+        )

backend/app/celery_app.py

@@ -14,6 +14,7 @@ Queues:
 from celery import Celery
 
 from .config import get_config
+from .services.worker_lanes import SWEEP_PERIOD_SECONDS
 
 
 def make_celery() -> Celery:
@@ -61,6 +62,13 @@ def make_celery() -> Celery:
             # can never starve the quick self-healing sweeps (operator-flagged
             # 2026-06-07: a 2h audit blocked vacuum/backup/normalize for hours).
             "backend.app.tasks.maintenance.*": {"queue": "maintenance"},
+            # The one long job in maintenance.py: a whole-library phash
+            # recompute (35 min hard limit; the library was cleared for
+            # re-hashing by migration 0098). On the quick lane it held a
+            # scheduler process for its whole run, and the minute ticks queued
+            # up behind it (2026-09-24: 7 waiting, "all workers busy for 18
+            # minutes"). An exact name wins over the glob above.
+            "backend.app.tasks.maintenance.backfill_phash": {"queue": "maintenance_long"},
             "backend.app.tasks.backup.*": {"queue": "maintenance_long"},
             "backend.app.tasks.admin.*": {"queue": "maintenance_long"},
             "backend.app.tasks.library_audit.*": {"queue": "maintenance_long"},
@@ -111,6 +119,34 @@ def make_celery() -> Celery:
                 "task": "backend.app.tasks.maintenance.recover_interrupted_tasks",
                 "schedule": 300.0,  # every 5 minutes
             },
+            "size-worker-lanes": {
+                "task": "backend.app.tasks.maintenance.size_worker_lanes",
+                "schedule": SWEEP_PERIOD_SECONDS,
+                #
+                # The number lives in `services/worker_lanes` because three
+                # places must agree on it: this schedule, the freshness of the
+                # sample the System tab reads, and the roster staleness
+                # thresholds in `api/system_health` — which now depend on this
+                # sweep rather than on a browser being open, and assert their
+                # headroom over it at import.
+                #
+                # ONE entry, replacing `autoscale-worker-lanes` (60s) and
+                # `reconcile-worker-lanes` (300s) on 2026-09-23. They were two
+                # sweeps over one number and most of the autoscaler's design
+                # existed to stop the reconcile undoing its work; with the
+                # stored `slots` gone there is nothing to disagree about.
+                #
+                # Fast enough to react to a BACKLOG — a five-minute reaction to
+                # a queue filling up is no reaction. It also carries what the
+                # reconcile was for: a worker restarted at its ENV concurrency
+                # is corrected on the next tick rather than after five.
+                #
+                # Cheap when settled: one inspect plus one LLEN sweep, and no
+                # control messages at all once every lane matches. It is also
+                # now the ONLY thing that inspects — nothing on a request path
+                # does — so this is the whole broker cost of the System tab,
+                # whether nobody or ten tabs are watching.
+            },
             "cleanup-old-tasks": {
                 "task": "backend.app.tasks.maintenance.cleanup_old_tasks",
                 "schedule": 86400.0,  # daily
@@ -120,6 +156,14 @@ def make_celery() -> Celery:
                 "schedule": 86400.0,  # daily — sweep .part/.partial left by a
                 # download/import killed mid-write (graceful-shutdown fallout)
             },
+            "backfill-phash-daily": {
+                "task": "backend.app.tasks.maintenance.backfill_phash",
+                "schedule": 86400.0,  # daily — NULL-only, so a no-op once the
+                # library is hashed. This is what makes migration 0098's
+                # re-hash happen on its own: 0098 NULLs every phash, and
+                # without a scheduled refill the library would sit
+                # dedup-disabled until someone ran a deep scan (#4223).
+            },
             "train-heads-nightly": {
                 "task": "backend.app.tasks.ml.scheduled_train_heads",
                 "schedule": 86400.0,  # passive cadence; manual retrain stays available
@@ -316,6 +360,9 @@ def make_celery() -> Celery:
             },
         },
         timezone="UTC",
+        # Beat's memory of when each job last ran comes from task_run, not a
+        # shelve file nothing persists — see beat_scheduler (#4408).
+        beat_scheduler="backend.app.beat_scheduler:TaskRunScheduler",
     )
     # FC-3i: register task_run signal handlers (side-effect import).
     from . import celery_signals  # noqa: F401

backend/app/frontend.py

@@ -46,6 +46,14 @@ async def serve_extension(filename: str):
 
     The application/x-xpinstall MIME tells Firefox to show its native
     install prompt instead of downloading the file as a blob.
+
+    Caching differs by name, and has to. A versioned name is one build's bytes
+    forever, so it can be cached for good. `fabledcurator-latest.xpi` is ONE
+    URL whose bytes change on every release, and Quart's default for a file
+    is `public, max-age=43200`: a browser that fetched it once reused those
+    bytes for 12 hours, so "install the latest" quietly reinstalled the
+    previous build (operator-flagged 2026-09-25). It is `no-cache` — the ETag
+    still makes an unchanged file a cheap 304.
     """
     if not _XPI_NAME_RE.fullmatch(filename):
         abort(404)
@@ -56,10 +64,11 @@ async def serve_extension(filename: str):
         if not xpis:
             abort(404)
         latest = xpis[-1]
-        return await send_file(
+        resp = await send_file(
             latest, mimetype="application/x-xpinstall",
             attachment_filename=latest.name,
         )
+        return _cache(resp, "no-cache")
     target = (XPI_DIR / filename).resolve()
     try:
         target.relative_to(XPI_DIR)
@@ -67,10 +76,19 @@ async def serve_extension(filename: str):
         abort(404)
     if not target.is_file():
         abort(404)
-    return await send_file(
+    resp = await send_file(
         target, mimetype="application/x-xpinstall",
         attachment_filename=filename,
     )
+    return _cache(resp, "public, max-age=31536000, immutable")
+
+
+def _cache(resp, policy: str):
+    """Set the XPI's Cache-Control, dropping the Expires send_file adds so the
+    two can never disagree."""
+    resp.headers["Cache-Control"] = policy
+    resp.headers.pop("Expires", None)
+    return resp
 
 
 @frontend_bp.route("/")

backend/app/models/__init__.py

@@ -8,6 +8,8 @@ from .backup_run import BackupRun
 from .base import Base
 from .character_prototype import CcipPrototypeState, CharacterPrototype
 from .credential import Credential
+from .discord_failed_media import DiscordFailedMedia
+from .discord_seen_media import DiscordSeenMedia
 from .download_event import DownloadEvent
 from .external_link import ExternalLink
 from .gpu_job import GpuJob
@@ -26,8 +28,6 @@ from .membership_sync import MembershipSync
 from .ml_settings import MLSettings
 from .patreon_failed_media import PatreonFailedMedia
 from .patreon_seen_media import PatreonSeenMedia
-from .pixiv_failed_media import PixivFailedMedia
-from .pixiv_seen_media import PixivSeenMedia
 from .platform_membership import PlatformMembership
 from .post import Post
 from .post_association import PostAssociation
@@ -46,6 +46,8 @@ from .tag_head import TagHead
 from .tag_positive_confirmation import TagPositiveConfirmation
 from .tag_suggestion_rejection import TagSuggestionRejection
 from .task_run import TaskRun
+from .worker_lane import WorkerLane
+from .worker_lane_sample import WorkerLaneSample
 
 __all__ = [
     "Base",
@@ -56,10 +58,10 @@ __all__ = [
     "BackupRun",
     "Source",
     "Credential",
+    "DiscordFailedMedia",
+    "DiscordSeenMedia",
     "PatreonFailedMedia",
     "PatreonSeenMedia",
-    "PixivFailedMedia",
-    "PixivSeenMedia",
     "SubscribeStarFailedMedia",
     "SubscribeStarSeenMedia",
     "Post",
@@ -98,4 +100,6 @@ __all__ = [
     "TagPositiveConfirmation",
     "TagSuggestionRejection",
     "TaskRun",
+    "WorkerLane",
+    "WorkerLaneSample",
 ]

backend/app/models/discord_failed_media.py

@@ -1,16 +1,9 @@
-"""PixivFailedMedia — per-source dead-letter ledger of Pixiv media that keeps
-failing to download/validate.
+"""DiscordFailedMedia — per-source dead-letter ledger of Discord files that
+keep failing to download or validate.
 
-Mirror of PatreonFailedMedia/SubscribeStarFailedMedia. Media that fails every
-walk (404'd pximg URL, deleted work, persistently-corrupt bytes) would
-otherwise re-error forever and re-burn backfill chunks. After ``attempts``
-reaches the dead-letter threshold the ingester skips it on routine
-tick/backfill walks (recovery still re-attempts). A later clean download
-clears the row.
-
-`filehash` is the same synthesized ``<illust_id>:p<num>`` /
-``<illust_id>:ugoira`` key the seen-ledger uses. UNIQUE (source_id, filehash)
-is the upsert key.
+Mirror of SubscribeStarFailedMedia. After `attempts` reaches the dead-letter
+threshold a routine walk skips the file (recovery still retries it); a later
+clean download clears the row. `filehash` is the seen-ledger's key.
 """
 
 from datetime import datetime
@@ -22,12 +15,10 @@ from sqlalchemy.types import DateTime
 from .base import Base
 
 
-class PixivFailedMedia(Base):
-    __tablename__ = "pixiv_failed_media"
+class DiscordFailedMedia(Base):
+    __tablename__ = "discord_failed_media"
     __table_args__ = (
-        UniqueConstraint(
-            "source_id", "filehash", name="uq_pixiv_failed_media_source_id"
-        ),
+        UniqueConstraint("source_id", "filehash", name="uq_discord_failed_media_source_id"),
     )
 
     id: Mapped[int] = mapped_column(Integer, primary_key=True)

backend/app/models/discord_seen_media.py

@@ -0,0 +1,34 @@
+"""DiscordSeenMedia — per-source ledger of Discord files already downloaded.
+
+Mirror of SubscribeStarSeenMedia for the native Discord ingester (milestone
+428). `filehash` holds the ingester's per-file key, `<message_id>:<media_id>`:
+the attachment id, or for an embed a hash of its URL path. Not the file's
+position in the message — an edit that removes a file renumbers the rest
+(see `discord_client.MediaItem`). The message record's own gate is the
+synthetic `message:<id>` key in the same column.
+"""
+
+from datetime import datetime
+
+from sqlalchemy import ForeignKey, Integer, String, UniqueConstraint, func
+from sqlalchemy.orm import Mapped, mapped_column
+from sqlalchemy.types import DateTime
+
+from .base import Base
+
+
+class DiscordSeenMedia(Base):
+    __tablename__ = "discord_seen_media"
+    __table_args__ = (
+        UniqueConstraint("source_id", "filehash", name="uq_discord_seen_media_source_id"),
+    )
+
+    id: Mapped[int] = mapped_column(Integer, primary_key=True)
+    source_id: Mapped[int] = mapped_column(
+        ForeignKey("source.id", ondelete="CASCADE"), nullable=False, index=True
+    )
+    filehash: Mapped[str] = mapped_column(String(128), nullable=False)
+    post_id: Mapped[str | None] = mapped_column(String(64), nullable=True)
+    seen_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), nullable=False, server_default=func.now()
+    )

backend/app/models/image_record.py

@@ -64,7 +64,11 @@ class ImageRecord(Base):
     # that 0001 also built was an exact duplicate of it — dropped in 0089
     # (#3301). Lookups by sha256 use the constraint's index.
     sha256: Mapped[str] = mapped_column(String(64), nullable=False)
-    phash: Mapped[str | None] = mapped_column(String(32), nullable=True, index=True)
+    # 64 hex chars = the 256-bit hash utils.phash emits at hash_size=16. Was
+    # String(32) (64-bit) until migration 0098; the narrow column was the
+    # reason for the undersized hash, and the undersized hash was collapsing
+    # variant artwork into one record (#4223).
+    phash: Mapped[str | None] = mapped_column(String(64), nullable=True, index=True)
     size_bytes: Mapped[int] = mapped_column(BigInteger, nullable=False)
     mime: Mapped[str] = mapped_column(String(64), nullable=False)
     width: Mapped[int | None] = mapped_column(Integer, nullable=True)

backend/app/models/import_settings.py

@@ -42,7 +42,12 @@ class ImportSettings(Base):
     single_color_threshold: Mapped[float] = mapped_column(Float, nullable=False, default=0.95, server_default="0.95")
     single_color_tolerance: Mapped[int] = mapped_column(Integer, nullable=False, default=30, server_default="30")
 
-    phash_threshold: Mapped[int] = mapped_column(Integer, nullable=False, default=10, server_default="10")
+    # Hamming distance over a 256-bit pHash (utils.phash, hash_size=16). The
+    # unit CHANGED in migration 0098 — it used to be bits out of 64 — so the
+    # old default of 10 is not this scale's 10, and 0098 resets every row.
+    # This is now the cheap PRE-FILTER: the aspect + pixel gates decide, which
+    # is what lets it be generous enough to catch a re-encoded rescale.
+    phash_threshold: Mapped[int] = mapped_column(Integer, nullable=False, default=24, server_default="24")
 
     # FC-3c downloader knobs
     download_rate_limit_seconds: Mapped[float] = mapped_column(
@@ -63,6 +68,16 @@ class ImportSettings(Base):
         Integer, nullable=False, default=90,
         server_default="90",
     )
+    # How far back a routine tick keeps looking after it has run out of new
+    # posts, so a creator who EDITS an older post to attach a hotfix build is
+    # still reached (ingest_core.DEFAULT_REVISIT_DAYS carries the reasoning).
+    # A knob rather than a constant because how long a creator keeps editing is
+    # a property of the creator, not of FabledCurator: 0 turns the revisit off
+    # and restores the pure count early-out.
+    download_revisit_days: Mapped[int] = mapped_column(
+        Integer, nullable=False, default=30,
+        server_default="30",
+    )
     download_failure_warning_threshold: Mapped[int] = mapped_column(
         Integer, nullable=False, default=5,
         server_default="5",
@@ -124,6 +139,50 @@ class ImportSettings(Base):
         server_default="24",
     )
 
+    # Whether FC links a CONCLUSIVE pair without asking.
+    #
+    # Operator, 2026-09-24: *"I don't want this to be manual that defeats the
+    # convenience that I'm going for."* Confirm-only was the right default
+    # while the only signals were circumstantial — proximity and a body that
+    # mentions Discord can never be more than suggestive, and asking was the
+    # honest response to that. A shared working name is different in kind: when
+    # the name appears in exactly these two posts and nowhere else in the
+    # artist's library, there is nothing for the operator to adjudicate.
+    #
+    # Only the conclusive band is affected (post_association_service.
+    # AUTO_LINK_FLOOR). Everything weaker still queues, and an accepted link is
+    # a row the operator can dismiss, so this is reversible in the UI rather
+    # than only in the database.
+    discord_link_auto: Mapped[bool] = mapped_column(
+        Boolean, nullable=False, default=True,
+        server_default="true",
+    )
+
+    # The unified card (#4402). A linked Discord drop is NOT absorbed into its
+    # teaser — it keeps its own post, date and provenance, and the teaser's card
+    # shows it by REFERENCE. Operator, 2026-09-24: *"discord 'posts' land as
+    # normal and only hidden from the post view they're posted the same day."*
+    #
+    # So the drop's own card leaves the feed only when it sits within this many
+    # hours of the teaser that references it — the adjacency that reads as the
+    # same thing twice. Hours rather than a calendar day: a teaser at 23:00 and
+    # its drop at 01:00 are one release, and "the same day" has no timezone
+    # the server can know.
+    discord_link_fold_hours: Mapped[float] = mapped_column(
+        Float, nullable=False, default=24.0,
+        server_default="24",
+    )
+    # How far from the teaser the card reaches for the rest of a piece's
+    # variants — the wips, alts and censor passes a creator trickles out under
+    # one working name (#4401). Measured on artist 8: named families spread a
+    # median 5 days and up to 44, while every name collision found spreads
+    # over 500. A reference, not a regrouping, so a generous value costs one
+    # extra thumbnail at worst — never a post moved or hidden.
+    discord_family_window_days: Mapped[float] = mapped_column(
+        Float, nullable=False, default=60.0,
+        server_default="60",
+    )
+
     # #830 off-platform file-host downloads — per-host enable lever (default on,
     # rule #26). Column names are extdl_<host>_enabled so the worker reads them
     # via getattr(settings, f"extdl_{host}_enabled", True).

backend/app/models/pixiv_seen_media.py

@@ -1,42 +0,0 @@
-"""PixivSeenMedia — per-source ledger of Pixiv media already
-downloaded+processed.
-
-Mirror of PatreonSeenMedia/SubscribeStarSeenMedia for the Pixiv native
-ingester (replacing gallery-dl). One queryable row per (source, media) so
-routine walks skip media we've already ingested; recovery mode bypasses the
-ledger to re-walk.
-
-Pixiv original URLs carry no content hash, so `filehash` is always the
-synthesized ``<illust_id>:p<num>`` (page) / ``<illust_id>:ugoira`` (frame
-zip) key — stable across any URL-shape drift. String(128) matches the sibling
-ledgers.
-"""
-
-from datetime import datetime
-
-from sqlalchemy import ForeignKey, Integer, String, UniqueConstraint, func
-from sqlalchemy.orm import Mapped, mapped_column
-from sqlalchemy.types import DateTime
-
-from .base import Base
-
-
-class PixivSeenMedia(Base):
-    __tablename__ = "pixiv_seen_media"
-    __table_args__ = (
-        # Dedup key the downloader upserts against: one ledger row per
-        # (source, media). A second sighting of the same media is a no-op.
-        UniqueConstraint(
-            "source_id", "filehash", name="uq_pixiv_seen_media_source_id"
-        ),
-    )
-
-    id: Mapped[int] = mapped_column(Integer, primary_key=True)
-    source_id: Mapped[int] = mapped_column(
-        ForeignKey("source.id", ondelete="CASCADE"), nullable=False, index=True
-    )
-    filehash: Mapped[str] = mapped_column(String(128), nullable=False)
-    post_id: Mapped[str | None] = mapped_column(String(64), nullable=True)
-    seen_at: Mapped[datetime] = mapped_column(
-        DateTime(timezone=True), nullable=False, server_default=func.now()
-    )

backend/app/models/post_association.py

@@ -91,6 +91,13 @@ class PostAssociation(Base):
     status: Mapped[str] = mapped_column(
         String(16), nullable=False, server_default="pending", index=True
     )
+    # WHO linked it: "fc" when the matcher linked a conclusive pair by itself
+    # (discord_link_auto), "operator" when a person accepted it. The card needs
+    # this to be honest — a link FC asserted on its own says so and offers an
+    # undo, which the operator chose over a silent merge (#4402). NULL on a row
+    # that is not linked, and on rows linked before the column existed, all of
+    # which an operator accepted: auto-linking shipped in the same release.
+    linked_by: Mapped[str | None] = mapped_column(String(16), nullable=True)
     created_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()
     )

backend/app/models/worker_lane.py

@@ -0,0 +1,85 @@
+"""worker_lane — the most workers the operator will let each lane use.
+
+Milestone 422 step 1, reshaped 2026-09-23. One row per lane in
+`services/worker_lanes.LANES`, and ONE COLUMN an operator sets.
+
+## Why there is only one number now
+
+There were three — `slots`, `slots_cap` and `autoscale` — because the manual
+dial was built first and the autoscaler arrived last, beside a control that
+already existed rather than in place of it.
+
+Operator: *"auto should be always on, not a setting, so that idle instances
+quiet down when not running. the number that is visible and something the
+user can tweak and manage should be the cap itself the number of running
+workers is handled by the autoscaling function which is always on."*
+
+So `slots` is gone. How many workers a lane is running right now is a
+MEASUREMENT — read live from the worker, moved by the autoscaler, never
+stored. Storing it made it look like a preference, which meant the operator
+had to keep two numbers in agreement and the autoscaler had to be told it was
+allowed to touch one of them.
+
+`autoscale` is gone for the same reason: it gated the mechanism behind a
+choice, and a lane nobody opted in simply never gave its slots back.
+
+`enabled` is gone too, and is now DERIVED: a cap of zero means no consumers.
+"Off" and "may use no workers" were two spellings of one fact, free to
+disagree.
+
+## What remains
+
+    1  <=  live pool  <=  slots_cap  <=  derived_ceiling
+           (autoscaler)   (this row)     (computed)
+
+The floor is one PROCESS, not zero: billiard will not run an empty pool, and
+the parked process is what `add_consumer` lands on when the cap goes back up.
+
+The DERIVED CEILING is deliberately absent from this table. It is computed
+from the container's cgroup limits on every read, so a row written on a 32GB
+host and later run in a 4GB container is bounded by the 4GB — a stored
+ceiling would quietly authorise what the box can no longer hold.
+"""
+
+from datetime import datetime
+
+from sqlalchemy import (
+    CheckConstraint,
+    DateTime,
+    Integer,
+    String,
+    func,
+)
+from sqlalchemy.orm import Mapped, mapped_column
+
+from .base import Base
+
+
+class WorkerLane(Base):
+    __tablename__ = "worker_lane"
+    __table_args__ = (
+        # Bare name — Base.metadata's naming convention prepends
+        # ck_worker_lane_. Pre-prefixing here doubles it, which is what
+        # alembic 0088 had to rename four constraints for (#3275).
+        CheckConstraint("slots_cap >= 0", name="cap_non_negative"),
+    )
+
+    # The lane name from services/worker_lanes.LANES — never a container
+    # hostname. See models/service_seen.py for why: celery's worker names here
+    # are `celery@<container id>`, minted fresh on every deploy.
+    name: Mapped[str] = mapped_column(String(32), primary_key=True)
+
+    # The most workers this lane may use. Zero means off — no consumers, so
+    # the lane takes no work and (for ML) downloads no model.
+    #
+    # There is no upper CHECK here, because the bound it would need is the
+    # derived ceiling, and no column holds that: it depends on the cgroup the
+    # container is running in right now. Enforced at write instead.
+    slots_cap: Mapped[int] = mapped_column(Integer, nullable=False)
+
+    updated_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True),
+        nullable=False,
+        server_default=func.now(),
+        onupdate=func.now(),
+    )

backend/app/models/worker_lane_sample.py

@@ -0,0 +1,83 @@
+"""worker_lane_sample — the last thing the sizing sweep measured about a lane.
+
+Milestone 422, 2026-09-23. A MEASUREMENT table, deliberately separate from
+`worker_lane`, which holds the one number an operator sets.
+
+## Why this exists
+
+Operator, 2026-09-23, looking at the System tab: *"there is a repull every
+time this page loads — is there a reason this info isn't being tracked in the
+background and stored in some way?"*
+
+There was not a good one. `/api/system/workers` ran a full celery inspect —
+four broadcast round trips on an eleven-second budget — on every call, and
+the page polls it every fifteen seconds. Meanwhile `size_worker_lanes` was
+already inspecting on a timer to decide pool sizes, computing exactly these
+numbers, using them, and throwing them away. The browser then asked the
+broker for them again.
+
+So the sweep writes what it saw here, and the endpoint reads this table. The
+request path makes no broker call at all any more.
+
+## Why NOT columns on `worker_lane`
+
+Because that is the mistake this milestone already made once and undid. That
+table used to carry `slots` — how many workers were running — beside
+`slots_cap`, and a measurement sitting next to a preference reads as a second
+preference: the operator had to keep two numbers in agreement, and the
+autoscaler had to be granted permission to move one of them.
+
+The distinction is the whole design, so it is a table boundary. Nothing an
+operator sets lives here; nothing here is ever an input to a decision about
+what they wanted.
+
+## Freshness is a value, not an assumption
+
+`measured_at` is returned to the UI, which says how old the reading is rather
+than implying it is live. A sample is a fact about a moment, and a page that
+presents a one-minute-old number as current is how an operator ends up
+mistrusting the whole surface.
+"""
+
+from datetime import datetime
+
+from sqlalchemy import Boolean, DateTime, Integer, String, func
+from sqlalchemy.orm import Mapped, mapped_column
+
+from .base import Base
+
+
+class WorkerLaneSample(Base):
+    __tablename__ = "worker_lane_sample"
+
+    # The lane name from services/worker_lanes.LANES. One row per lane,
+    # overwritten in place: this is the LATEST reading, not a history. A time
+    # series would be a different table with a different retention problem,
+    # and nothing has asked for one.
+    lane: Mapped[str] = mapped_column(String(32), primary_key=True)
+
+    # Whether anything answered for this lane. NOT the same as "zero workers"
+    # — an unswept absence is not a verdict (snippet #3969). False here means
+    # the inspect came back without this lane, so every count below is
+    # meaningless and the UI must say "not answering" rather than "0".
+    present: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False)
+
+    replicas: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+
+    # Pool size of ONE process, nullable because a worker that answered
+    # without reporting its pool is unknown rather than empty.
+    pool: Mapped[int | None] = mapped_column(Integer, nullable=True)
+
+    active: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    reserved: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+
+    # Redis LLEN across the lane's queues. Nullable for the same reason as
+    # `pool`: a queue the broker did not answer for is unknown, and summing it
+    # as zero would report a buried lane as idle.
+    queue_depth: Mapped[int | None] = mapped_column(Integer, nullable=True)
+
+    measured_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True),
+        nullable=False,
+        server_default=func.now(),
+    )

backend/app/scripts/gen_supervisord.py

@@ -0,0 +1,227 @@
+"""Emit a supervisord config for the single-container layout.
+
+Milestone 422 step 5. Writes to stdout; `entrypoint.sh all` redirects it to a
+file and execs supervisord against it.
+
+## Why this is generated and not a checked-in .conf
+
+A static config would spell out each lane's `-Q` list, and that would be a
+FIFTH hand-kept copy of the queue names — after `celery_app.task_routes`, and
+the three collapsed in steps 1, 2 and 4 (`service_roster.ROLE_NAMES`,
+`system_activity._QUEUE_NAMES`, and the Activity filter). Every one of those
+had already drifted by the time it was found.
+
+Generating from `worker_lanes.LANES` makes a stronger guarantee than "they
+match today": the processes this container runs and the lanes the application
+believes in are the same list, so a lane added to `LANES` gets a process
+without anyone remembering to add one, and a queue can never end up with no
+consumer because a config file was missed.
+
+## Why supervisord
+
+It is one pip dependency on an image that is already Python, and it does the
+four things this needs without being clever: restart a program that exits,
+give each one its OWN stop timeout, signal the process GROUP rather than the
+leader, and put every program's output on one stdout.
+
+The process-group part is not a detail. Celery's prefork pool forks children,
+and a TERM delivered only to the parent leaves them running — which is how a
+"graceful" shutdown turns into orphaned workers holding tasks. `stopasgroup`
+and `killasgroup` are both set for every program.
+
+s6-overlay is the other standard answer and would work; it needs a build-time
+download and a second mental model, and its advantage (correct PID-1 signal
+and zombie handling) is available here from `init: true` in compose, which
+puts tini in front of supervisord. Neither choice reaches the application —
+nothing in FC talks to the supervisor — so this is reversible without touching
+a line of product code.
+
+## Every lane, including ml
+
+Step 6 merged the images, so this one carries torch and the ML requirements
+and the `ml` lane gets a program like any other. It starts at one slot with
+its consumers CANCELLED — `enabled=false` in the seeded settings — so it
+holds a process and no model. That matters: `add_consumer` needs a running
+worker to reach, and without one the UI switch would have nothing to switch.
+
+Nothing is downloaded by starting it. The model fetch is enqueued when the
+lane is enabled, which is what lets rule 164 permit a runtime fetch at all.
+"""
+
+from __future__ import annotations
+
+import argparse
+import shlex
+import sys
+
+from ..services.worker_lanes import LANES, MIN_POOL_SLOTS, Lane
+
+# One number for the whole container, and it must cover the SLOWEST lane —
+# docker gives the container a single stop timeout, where compose today gives
+# each service its own (90/60/180/120s). `maintenance_long` is the 180s one:
+# DB backups, library audits and translation backfill. Anything less turns a
+# routine restart into a SIGKILL mid-backup.
+#
+# Per-program values below are the old per-service ones, preserved: supervisord
+# waits `stopwaitsecs` for each, and they stop in parallel, so the container's
+# own timeout needs to cover the max rather than the sum.
+STOP_WAIT_SECONDS: dict[str, int] = {
+    "worker": 90,
+    "scheduler": 60,
+    "maintenance_long": 180,
+    "ml": 120,
+}
+DEFAULT_STOP_WAIT = 60
+
+def _program(lane: Lane, *, slots: int) -> str:
+    """One [program:x] block.
+
+    `stdout_logfile=/dev/fd/1` with maxbytes 0 puts the lane's output straight
+    on the container's stdout unbuffered, so `docker logs` shows every lane
+    interleaved rather than supervisord swallowing them into rotated files.
+
+    The output is prefixed through `sed` so a line can be attributed to a lane
+    — four celery workers and hypercorn on one stream are otherwise
+    indistinguishable. The shell that the pipe requires is exactly why
+    `stopasgroup` matters: the signal has to reach the celery process, not the
+    `sh` holding the pipeline.
+    """
+    inner = f"./entrypoint.sh {lane.entrypoint_role}"
+    prefixed = f"{inner} 2>&1 | sed -u 's/^/[{lane.name}] /'"
+    stop_wait = STOP_WAIT_SECONDS.get(lane.name, DEFAULT_STOP_WAIT)
+    return "\n".join([
+        f"[program:{lane.name}]",
+        f"command=sh -c {shlex.quote(prefixed)}",
+        # QUOTED, and that is load-bearing. supervisord parses `environment`
+        # as a COMMA-separated KEY=VALUE list, so an unquoted queue list reads
+        # as CELERY_QUEUES=default followed by three malformed entries — and
+        # the lane would consume only its first queue. Silent: the worker
+        # starts, reports healthy, and simply never picks up `import`.
+        f'environment=CELERY_QUEUES="{",".join(lane.queues)}",'
+        f"CELERY_CONCURRENCY={slots},"
+        # A UNIQUE celery node name per lane, and the reason is not cosmetic.
+        # These processes share one hostname, so celery's default
+        # `celery@<hostname>` made all four the SAME node: inspect collapsed
+        # their replies, three lanes read as absent, and which three varied
+        # per call (run 7319). The healthcheck could never pass, and
+        # pool_grow's `destination` would have addressed an arbitrary lane.
+        f"CELERY_NODENAME={lane.name}",
+        "autostart=true",
+        "autorestart=true",
+        # A lane that dies instantly and repeatedly is a broken image, not a
+        # transient fault. Backing off stops it burning a core in a restart
+        # loop while still recovering from a one-off crash.
+        "startretries=3",
+        "startsecs=5",
+        f"stopwaitsecs={stop_wait}",
+        "stopasgroup=true",
+        "killasgroup=true",
+        "stdout_logfile=/dev/fd/1",
+        "stdout_logfile_maxbytes=0",
+        "redirect_stderr=true",
+        "",
+    ])
+
+
+# supervisord's control socket. /tmp for the same reason the generated config
+# lives there — writable by every role, and per-container by nature.
+SOCKET_PATH = "/tmp/supervisor.sock"
+
+
+def _web_program() -> str:
+    """hypercorn. Started FIRST (priority) because its role runs
+    `alembic upgrade head`, and a worker that boots against an un-migrated
+    schema fails in a way that looks like application breakage."""
+    prefixed = "./entrypoint.sh web 2>&1 | sed -u 's/^/[web] /'"
+    return "\n".join([
+        "[program:web]",
+        f"command=sh -c {shlex.quote(prefixed)}",
+        "priority=1",
+        "autostart=true",
+        "autorestart=true",
+        "startretries=3",
+        "startsecs=5",
+        # Short: HTTP requests and the occasional file download. Matches the
+        # 30s the operator's production stack gives the web service.
+        "stopwaitsecs=30",
+        "stopasgroup=true",
+        "killasgroup=true",
+        "stdout_logfile=/dev/fd/1",
+        "stdout_logfile_maxbytes=0",
+        "redirect_stderr=true",
+        "",
+    ])
+
+
+def render() -> str:
+    parts = [
+        "\n".join([
+            "[supervisord]",
+            # PID 1 in the container, so it must not daemonise.
+            "nodaemon=true",
+            # supervisord's OWN log. /dev/fd/1 keeps it on the container's
+            # stdout beside the programs rather than in a file nobody reads.
+            "logfile=/dev/fd/1",
+            "logfile_maxbytes=0",
+            "loglevel=info",
+            "",
+        ]),
+        # THE CONTROL SOCKET, and it is not optional furniture.
+        #
+        # Without these three sections supervisord runs perfectly and
+        # `supervisorctl` cannot talk to it at all:
+        #
+        #     Error: .ini file does not include supervisorctl section
+        #
+        # Which is the first thing anyone reaches for when a lane misbehaves
+        # in the consolidated container — `docker exec <c> supervisorctl
+        # status` to see which processes are up, or `restart ml` to bounce one
+        # without taking the whole application down with it. Consolidation
+        # took away `docker ps` as the way to see the lanes; this is what
+        # replaces it, and shipping without it would have left an operator
+        # with one container, five processes inside it, and no way to ask
+        # about any of them.
+        #
+        # Found by the smoke's own diagnostic line on run 7322, which printed
+        # this error instead of a process list. It was behind `|| true`, so it
+        # cost nothing and said so anyway — the argument for printing evidence
+        # even where nothing depends on it.
+        #
+        # /tmp, like the generated config itself: writable by every role
+        # without assuming a volume, and per-container state that must not
+        # outlive the container.
+        "\n".join([
+            "[unix_http_server]",
+            f"file={SOCKET_PATH}",
+            "chmod=0700",
+            "",
+            "[rpcinterface:supervisor]",
+            "supervisor.rpcinterface_factory = "
+            "supervisor.rpcinterface:make_main_rpcinterface",
+            "",
+            "[supervisorctl]",
+            f"serverurl=unix://{SOCKET_PATH}",
+            "",
+        ]),
+        _web_program(),
+    ]
+    # Lanes after web, in LANES order, so the log reads in a stable sequence.
+    for lane in LANES:
+        # A lane configured at zero slots still gets a PROCESS, at one slot
+        # with its consumers cancelled by the reconcile. Without a running
+        # worker there is nothing for `add_consumer` to reach, so enabling the
+        # lane from the UI could not work at all — the process has to exist for
+        # the switch to have something to switch.
+        parts.append(_program(lane, slots=MIN_POOL_SLOTS))
+    return "\n".join(parts)
+
+
+def main(argv: list[str] | None = None) -> int:
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.parse_args(argv)
+    sys.stdout.write(render())
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

backend/app/scripts/healthcheck.py

@@ -0,0 +1,181 @@
+"""The container's healthcheck. Picks the right check from the role it runs.
+
+Exit 0 healthy, non-zero unhealthy.
+
+## Why this is in the IMAGE and not in every compose file
+
+Because the image is the only thing that knows what it is running. A
+deployment had to declare a healthcheck per service, which meant every
+compose file, stack file and README repeated the same knowledge:
+
+    web         -> curl /api/health
+    worker      -> celery inspect ping -d celery@$HOSTNAME
+    all         -> both, for every lane
+
+Three checks, written out by hand, once per service, in every file anyone
+ever wrote — and none of them wrong until a role changed. Operator, 2026-09-23:
+*"why isn't the healthcheck built into the image or base on what command runs
+if one is passed in."* There was no reason. The role is already a fact the
+container holds; asking the deployment to restate it is the same duplication
+the lane table exists to remove one level down.
+
+So `entrypoint.sh` records the role it started, the Dockerfile declares ONE
+`HEALTHCHECK` that runs this, and a stack file says nothing at all. Declaring
+one anyway still works — docker lets a service override the image's — which
+is the escape hatch for a deployment that genuinely wants something else.
+
+## What each role is asked
+
+* **web** — hypercorn answers `/api/health`. No database: the endpoint is a
+  no-DB 200 that proves the app booted and is serving after `alembic upgrade
+  head`, which is what a rolling deploy needs to know.
+* **worker / scheduler / ml-worker** — THIS container's celery node answers a
+  ping over the broker. Not "some worker answered": the node name is pinned
+  to this container, or a healthy sibling would keep a dead one looking alive.
+* **all** — both halves, for every lane in the table. The failure mode
+  consolidation creates is that docker can no longer see the lanes as
+  separate services, so a web-only check reports a healthy container with
+  every worker dead.
+* **shell / alembic / anything else** — nothing to check. These are one-shot
+  or interactive; a liveness probe on them has no meaning, so it passes
+  rather than inventing a verdict.
+
+## An unrecorded role passes rather than failing
+
+If the role file is missing, the entrypoint did not run — someone used
+`--entrypoint` or ran a bare command. That is a debugging shape, and a
+healthcheck that cannot tell what it is looking at must not assert that the
+thing is broken (snippet #3969: an unswept read is not a verdict). It says so
+on stdout and exits 0.
+"""
+
+from __future__ import annotations
+
+import os
+import socket
+import sys
+import urllib.error
+import urllib.request
+
+# Written by entrypoint.sh at boot. /tmp because it is the one path writable
+# by every role without assuming a volume, and the value is per-container
+# state that must NOT survive into a new container.
+ROLE_FILE = os.environ.get("FC_ROLE_FILE", "/tmp/fc-role")
+
+WEB_URL = "http://localhost:8080/api/health"
+WEB_TIMEOUT = 5.0
+
+# A broker round trip, so it gets a deadline (rule 156). Generous relative to
+# `inspect`'s 2s elsewhere: this runs every 30s with retries, and a transient
+# blip flagging a worker unhealthy would roll back a deployment that is fine.
+PING_TIMEOUT = 10.0
+
+CELERY_ROLES = {"worker", "scheduler", "ml-worker"}
+# Roles with nothing to probe. Listed rather than treated as the default, so
+# an unknown role takes the "I cannot tell" path and says so.
+NO_CHECK_ROLES = {"shell", "bash", "alembic"}
+
+
+def current_role() -> str | None:
+    """The role this container was started with, or None if nothing recorded."""
+    env = os.environ.get("FC_ROLE")
+    if env:
+        return env.strip()
+    try:
+        with open(ROLE_FILE) as fh:
+            return fh.read().strip() or None
+    except OSError:
+        return None
+
+
+def _web_ok() -> tuple[bool, str]:
+    try:
+        with urllib.request.urlopen(WEB_URL, timeout=WEB_TIMEOUT) as resp:
+            if resp.status == 200:
+                return True, ""
+            return False, f"web returned {resp.status}"
+    except (urllib.error.URLError, OSError) as exc:
+        return False, f"web unreachable: {exc}"
+
+
+def _this_node_ok() -> tuple[bool, str]:
+    """Ping THIS container's celery node, by name.
+
+    Pinned to this node deliberately. A bare `ping()` is answered by any
+    worker on the broker, so in a stack with several replicas a dead one
+    would go on reporting healthy for as long as a sibling was alive — the
+    healthcheck would be measuring the cluster, not the container it is in.
+    """
+    from ..celery_app import celery as celery_app
+
+    node = f"{os.environ.get('CELERY_NODENAME', 'celery')}@{socket.gethostname()}"
+    try:
+        replies = celery_app.control.ping(destination=[node], timeout=PING_TIMEOUT)
+    except Exception as exc:  # noqa: BLE001 — a probe reports, never raises
+        return False, f"could not reach the broker: {exc}"
+    if not replies:
+        return False, f"{node} did not answer a ping"
+    return True, ""
+
+
+def _lanes_ok() -> tuple[bool, str]:
+    """Every lane in the table is answering.
+
+    Deliberately ignores whether a lane is ON: a lane at cap 0 still runs its
+    process with its consumers cancelled, so it answers `inspect` and is
+    healthy. Health is "is the process alive"; whether it should be consuming
+    is a settings question the sizing pass owns, and conflating them would
+    make turning a lane off mark the container unhealthy.
+
+    That was not merely a risk — it was happening. Until 2026-09-23 a worker
+    was attributed to its lane by the queues it was CONSUMING, and a lane with
+    its consumers cancelled reports none, so it read as absent and this check
+    failed. ML ships at cap 0, so a fresh install was permanently unhealthy
+    and Swarm restarts an unhealthy task forever. The docstring above said the
+    right thing while the code did the opposite; `worker_lanes.lane_for_node`
+    is what makes it true.
+    """
+    from ..services.worker_control import inspect_lanes_sync
+    from ..services.worker_lanes import LANES
+
+    live = inspect_lanes_sync()
+    missing = sorted(lane.name for lane in LANES if not live[lane.name].present)
+    if missing:
+        return False, "lanes not answering: " + ", ".join(missing)
+    return True, ""
+
+
+def main(argv: list[str] | None = None) -> int:
+    role = current_role()
+
+    if role is None:
+        # Not a failure. See the module docstring: the entrypoint did not run,
+        # so there is no role to check against and no basis for a verdict.
+        print("no role recorded; nothing to check")
+        return 0
+
+    if role in NO_CHECK_ROLES:
+        print(f"{role}: nothing to check")
+        return 0
+
+    checks = []
+    if role == "all":
+        checks = [_web_ok, _lanes_ok]
+    elif role == "web":
+        checks = [_web_ok]
+    elif role in CELERY_ROLES:
+        checks = [_this_node_ok]
+    else:
+        print(f"unknown role {role!r}; nothing to check")
+        return 0
+
+    for check in checks:
+        ok, detail = check()
+        if not ok:
+            print(detail, file=sys.stderr)
+            return 1
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

backend/app/scripts/wait_for_deps.py

@@ -0,0 +1,146 @@
+"""Block until Postgres and Redis accept connections. Exit 0 ready, 1 timed out.
+
+## Why the container has to do this itself
+
+Compose has `depends_on: {condition: service_healthy}`, and **Swarm ignores
+it**. `docker stack deploy` has no ordering primitive at all: every service in
+the stack starts at once, so FabledCurator races Postgres on every cold
+deploy and always has.
+
+The multi-service stack hid how sharp that is. `web` ran `alembic upgrade
+head`, failed against a Postgres that was still doing `initdb`, and the task
+died — but Swarm restarts a failed task forever, so the service came up a few
+seconds later and nobody saw a problem worth naming.
+
+Consolidation removes that safety net. supervisord gives each program
+`startretries=3`, so a web program that fails three times in the first
+seconds goes FATAL and **stays** FATAL: supervisord keeps running, the
+container keeps running, and the application never starts. The healthcheck
+catches it — but as a container that is permanently unhealthy for a reason
+that has nothing to do with the image, on a stack whose database simply took
+twenty seconds to initialise.
+
+Operator, 2026-09-23: *"it's a single container that need to connect
+successfully to redis and postgres before starting work shouldn't that simply
+be a check (with retries) at the start of the container."* Yes.
+
+## A TCP connect, not a query
+
+The same probe `build.yml`'s integration lane and the build smoke already use.
+It answers the question that is actually being asked — is something listening
+— and it cannot fail for a reason that retrying will never fix.
+
+A real query would be a stronger readiness signal and a worse gate: a wrong
+password or a missing database is not a transient condition, and a loop that
+waits for one to heal turns a five-second misconfiguration into a two-minute
+timeout with a misleading message. Those belong to alembic, which runs
+seconds later and says exactly what is wrong.
+
+The Postgres image is well behaved here: during `initdb` it serves on a unix
+socket only and opens TCP when it is ready for clients, so the connect is a
+good proxy for "ready" rather than merely "process exists".
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+import socket
+import sys
+import time
+from urllib.parse import urlparse
+
+# Long enough for a first-ever `initdb` on a slow disk, which is the worst
+# case this exists for and is measured in tens of seconds, not minutes. A
+# deploy that is genuinely misconfigured should fail while someone is still
+# watching it rather than hold the container open for a quarter of an hour.
+DEFAULT_TIMEOUT = 120.0
+CONNECT_TIMEOUT = 2.0
+RETRY_DELAY = 1.0
+# Progress every N attempts. `docker logs` on a container that is waiting must
+# say what it is waiting for — silence is indistinguishable from a hang.
+REPORT_EVERY = 5
+
+
+def _target(url: str | None, default_port: int) -> tuple[str, int] | None:
+    """(host, port) from a connection URL, or None if there is nothing to wait for."""
+    if not url:
+        return None
+    parsed = urlparse(url)
+    if not parsed.hostname:
+        return None
+    return parsed.hostname, parsed.port or default_port
+
+
+def targets() -> list[tuple[str, tuple[str, int]]]:
+    """What this container must reach, read from the same env the app reads.
+
+    Derived rather than passed in, so the wait cannot drift from what the
+    application will actually connect to — a gate that checks a different
+    host than the app uses is worse than no gate.
+    """
+    out: list[tuple[str, tuple[str, int]]] = []
+
+    host = os.environ.get("DB_HOST")
+    if host:
+        out.append(("postgres", (host, int(os.environ.get("DB_PORT") or 5432))))
+
+    broker = _target(os.environ.get("CELERY_BROKER_URL"), 6379)
+    if broker:
+        out.append(("redis", broker))
+
+    return out
+
+
+def _accepts(host: str, port: int) -> bool:
+    try:
+        with socket.create_connection((host, port), timeout=CONNECT_TIMEOUT):
+            return True
+    except OSError:
+        return False
+
+
+def wait(
+    name: str, host: str, port: int, deadline: float, now=time.monotonic,
+) -> bool:
+    attempt = 0
+    while True:
+        if _accepts(host, port):
+            print(f"[wait] {name} at {host}:{port} is accepting connections")
+            return True
+        attempt += 1
+        if now() >= deadline:
+            print(
+                f"[wait] TIMEOUT: {name} at {host}:{port} never accepted a "
+                f"connection ({attempt} attempts)",
+                file=sys.stderr,
+            )
+            return False
+        if attempt % REPORT_EVERY == 0:
+            left = int(deadline - now())
+            print(f"[wait] {name} at {host}:{port} not ready yet, {left}s left")
+        time.sleep(RETRY_DELAY)
+
+
+def main(argv: list[str] | None = None) -> int:
+    ap = argparse.ArgumentParser(description="Wait for Postgres and Redis.")
+    ap.add_argument("--timeout", type=float, default=DEFAULT_TIMEOUT)
+    args = ap.parse_args(argv)
+
+    wanted = targets()
+    if not wanted:
+        # Nothing configured to wait for. Not an error: `shell` and one-off
+        # runs are legitimate, and refusing to start would make this gate the
+        # reason a debugging container will not boot.
+        print("[wait] no database or broker configured; nothing to wait for")
+        return 0
+
+    deadline = time.monotonic() + args.timeout
+    for name, (host, port) in wanted:
+        if not wait(name, host, port, deadline):
+            return 1
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

backend/app/services/artist_service.py

@@ -300,21 +300,48 @@ class ArtistService:
         await self.session.commit()
         return artist
 
+    async def all_names(self) -> list[tuple[int, str, str]]:
+        """Every artist as (id, name, slug), alphabetical.
+
+        For pickers that should show a full list before anything is typed (the
+        Latest feed's artist filter). Three columns and no joins, so it stays
+        cheap on a library of thousands of artists.
+        """
+        rows = (await self.session.execute(
+            select(Artist.id, Artist.name, Artist.slug).order_by(func.lower(Artist.name))
+        )).all()
+        return [(r.id, r.name, r.slug) for r in rows]
+
     async def autocomplete(self, prefix: str, limit: int = 20) -> list[Artist]:
         cleaned = (prefix or "").strip()
         if not cleaned:
             return []
-        like = f"%{cleaned.lower()}%"
-        prefix_like = f"{cleaned.lower()}%"
-        # Rank: exact (0) < prefix (1) < substring (2).
+        low = cleaned.lower()
+        like = f"%{low}%"
+        prefix_like = f"{low}%"
+        # Spacing- and punctuation-insensitive too, so "Tamada Heijun" finds
+        # "TamadaHeijun" and "sabu_art" finds "Sabu Art" — the same creator is
+        # spelled differently on every platform, and the browser extension's
+        # Add panel matches a Discord server name against these (milestone 429).
+        # [[:alnum:]] keeps non-Latin letters; a query with none (all
+        # punctuation) skips this arm rather than matching every artist.
+        squashed = "".join(ch for ch in low if ch.isalnum())
+        name_squashed = func.regexp_replace(func.lower(Artist.name), "[^[:alnum:]]", "", "g")
+        matches = [func.lower(Artist.name).like(like)]
+        if squashed:
+            matches.append(name_squashed.like(f"%{squashed}%"))
+        # Rank: exact (0) < exact ignoring spacing (1) < prefix (2) <
+        # substring (3) < substring ignoring spacing (4).
         rank = case(
-            (func.lower(Artist.name) == cleaned.lower(), 0),
-            (func.lower(Artist.name).like(prefix_like), 1),
-            else_=2,
+            (func.lower(Artist.name) == low, 0),
+            (name_squashed == squashed, 1),
+            (func.lower(Artist.name).like(prefix_like), 2),
+            (func.lower(Artist.name).like(like), 3),
+            else_=4,
         ).label("rank")
         rows = (await self.session.execute(
             select(Artist, rank)
-            .where(func.lower(Artist.name).like(like))
+            .where(or_(*matches))
             .order_by(rank, Artist.name.asc())
             .limit(limit)
         )).all()

backend/app/services/backup_service.py

@@ -24,6 +24,25 @@ from pathlib import Path
 
 _BACKUPS_DIRNAME = "_backups"
 
+# Excluded from the images tarball, and each for its own reason (#4233, #4234):
+#
+#   _backups    — the archive would otherwise contain every previous archive.
+#                 This is not hypothetical: the 2026-05-23/24 runs, taken before
+#                 this exclude existed, grew 43G -> 107G -> ... -> 2123G as each
+#                 swallowed its predecessors, and cost 4.3T of the images
+#                 filesystem until they were reclaimed on 2026-09-21.
+#   _quarantine — holds files deliberately pulled OUT of the library.
+#   secrets     — `credential_key.b64`, the key that decrypts the stored
+#                 Patreon/SubscribeStar session credentials.
+#   cookies     — those session cookies themselves.
+#
+# The last two are the ones worth stating plainly: an images tarball is a media
+# archive, and a media archive that carries the key to the operator's accounts
+# is a credential leak wearing a backup's name. Encryption at rest buys nothing
+# when the key rides along in the same file. A restore therefore does NOT
+# re-establish credentials — you sign in again, which is the correct outcome.
+_IMAGES_EXCLUDED_DIRNAMES = ("_backups", "_quarantine", "secrets", "cookies")
+
 # Subprocess-level guardrails BEYOND the Celery soft_time_limit. The Celery
 # soft limit signals the Python process; subprocess.Popen in a blocking syscall
 # ignores that signal, so these bound the worst case directly. Each sits just
@@ -173,8 +192,10 @@ def backup_images(
         [
             "tar", "--zstd", "-cf", str(tar_path),
             "-C", str(images_root.parent), images_root.name,
-            f"--exclude={images_root.name}/_backups",
-            f"--exclude={images_root.name}/_quarantine",
+            *(
+                f"--exclude={images_root.name}/{name}"
+                for name in _IMAGES_EXCLUDED_DIRNAMES
+            ),
         ],
         _IMAGES_SUBPROCESS_TIMEOUT_S,
     )

backend/app/services/credential_crypto.py

@@ -24,6 +24,7 @@ rows undecryptable (recovery = delete the rows and re-upload).
 
 import logging
 import os
+import tempfile
 from pathlib import Path
 
 from cryptography.fernet import Fernet, InvalidToken
@@ -80,10 +81,59 @@ class CredentialCrypto:
         parent = self._key_path.parent
         parent.mkdir(parents=True, exist_ok=True)
         os.chmod(parent, 0o700)
+
+        # Written to a temp file and LINKED into place, not written directly.
+        #
+        # hypercorn starts several worker processes and each one builds the
+        # app, so on a first boot they all reach this at once. A plain
+        # `write_bytes` creates the file at size zero and fills it a moment
+        # later, which gives a second process an `exists()` of True and a
+        # `read_bytes()` of b"" — and the app dies with
+        #
+        #     ValueError: Fernet key must be 32 url-safe base64-encoded bytes.
+        #
+        # Seen on run 7368's smoke, and it is a race rather than a certainty:
+        # the same image had booted cleanly on the three runs before it. A
+        # first boot that fails one time in five is worse than one that fails
+        # every time, because it looks like the deployment rather than the code.
+        #
+        # `os.link` is the atomic part: it either creates the name or raises
+        # FileExistsError, and it cannot expose a half-written file. NOT
+        # `os.replace`, which would succeed — so two processes that both
+        # generated a key would each think they had won, and the loser's key
+        # would overwrite the one the winner had already handed to Fernet.
+        # `mkstemp`, not a pid-derived name. The first cut spelled the temp
+        # file `.credential_key.b64.<pid>.tmp`, which assumes one bootstrap per
+        # process — and the test that exercises this with eight THREADS shares
+        # one pid, so all eight raced the same filename and six died with
+        # FileNotFoundError when another had already unlinked it. The
+        # assumption held for hypercorn's workers and would have held in
+        # production; it was still an assumption the code did not need to make.
         key = Fernet.generate_key()
-        self._key_path.write_bytes(key)
-        os.chmod(self._key_path, 0o600)
-        return key
+        fd, tmp_name = tempfile.mkstemp(
+            dir=parent, prefix=f".{self._key_path.name}.", suffix=".tmp",
+        )
+        tmp = Path(tmp_name)
+        try:
+            with os.fdopen(fd, "wb") as fh:
+                fh.write(key)
+            os.chmod(tmp, 0o600)
+            try:
+                os.link(tmp, self._key_path)
+            except FileExistsError:
+                # Another process created it between our `exists()` check and
+                # here. Theirs is as good as ours, and using it is what keeps
+                # every worker on ONE key.
+                log.info(
+                    "another process created %s first; using that key",
+                    self._key_path,
+                )
+        finally:
+            tmp.unlink(missing_ok=True)
+        # Read back rather than returning `key`: on the losing branch the file
+        # holds somebody else's, and returning ours would leave this worker
+        # encrypting with a key no other worker can read.
+        return self._key_path.read_bytes()
 
     def encrypt(self, plaintext: str) -> bytes:
         return self._fernet.encrypt(plaintext.encode("utf-8"))

backend/app/services/db_helpers.py

@@ -16,7 +16,7 @@ from __future__ import annotations
 
 from collections.abc import Awaitable, Callable
 
-from sqlalchemy import Select
+from sqlalchemy import Select, and_
 from sqlalchemy.exc import IntegrityError
 from sqlalchemy.ext.asyncio import AsyncSession
 
@@ -67,12 +67,26 @@ async def get_or_create[T](
 
 
 def failing_sources_clause():
-    """A source is FAILING when its runs are actually erroring.
+    """A source is FAILING when it is ENABLED and its runs are erroring.
 
     Deliberately not `last_error IS NOT NULL` — a tier-limited source clears
     last_error and keeps a chip, and must never be counted as broken.
+
+    The `enabled` half was folded in 2026-09-21 (#4279). A disabled source is
+    one FC deliberately stopped — most often because the membership sweep saw
+    `former_patron` — and "stopped because you no longer subscribe" is not
+    "failing". Worse, it is a failure nobody can clear: a disabled source is
+    never scheduled, so no successful run ever resets the counter, and the
+    card's Retry button routes to `/check`, which refuses a disabled source
+    outright. Ebi77 sat in the banner for six days with no action available.
+
+    This also settles a disagreement the two callers already had. The
+    scheduler's status count paired this clause with `enabled.is_(True)`;
+    `SourceService.list(failing=True)` did not. One counted Ebi77, the other
+    did not — the exact drift the note above this function warns about, which
+    is why the `enabled` test belongs IN the predicate rather than beside it.
     """
-    return Source.consecutive_failures > 0
+    return and_(Source.enabled.is_(True), Source.consecutive_failures > 0)
 
 
 def no_access_sources_clause():

backend/app/services/discord_client.py

@@ -0,0 +1,562 @@
+"""Native Discord read client — the Discord counterpart to subscribestar_client.
+
+Mirrors gallery-dl 1.32.13's `extractor/discord.py` (rule 130: gallery-dl is the
+known-working base), adapted to the native core's client contract
+(`ingest_core` module docstring): `iter_posts` / `extract_media`, plus the
+post-first `post_record_key` and the `post_meta` date the revisit window reads.
+
+What is mirrored exactly, because drift in any of it changes what we fetch or
+where it lands on disk:
+
+  - API v10, `Authorization: <user token>` (a USER token, not a bot token).
+  - gallery-dl's request profile: its date-derived Firefox User-Agent,
+    `Accept: */*`, `Accept-Language`, `Referer: https://discord.com/`.
+  - `GET /channels/{id}/messages?limit=100&before=<last id>`, newest first,
+    stopping on a short page. Message types {0, 19, 21} only.
+  - The walk: a text/news channel's own messages then its threads, a forum's
+    threads, a category's children, a server's text/news/forum channels.
+  - Files: attachments, then embeds of type image/gifv/video (FC configures
+    `embeds: all`, which for files is the same three plus rich/link embeds that
+    carry an image), then forwarded `message_snapshots`, numbered from 1 across
+    the lot — the `num` in `{date}_{message_id}_{num}_{filename}`.
+  - Text: `content`, rich-embed author/title/description/fields/footer, poll.
+
+Two deliberate departures, both about the walk order, neither about content:
+
+  - Threads are walked newest-CREATED first (by id), not by last-message time.
+    A backfill resumes from a checkpointed channel; last-message order shifts
+    between chunks whenever someone posts, which can move an unwalked thread
+    above the resume point and skip it. Creation order only ever grows at the
+    front, where the next tick finds it.
+  - A 403 on a thread or a nested channel skips that feed instead of failing
+    the walk. gallery-dl skips only nested channels; one private thread the
+    token cannot read would otherwise stop every channel after it.
+
+FC runs on a plain-HTTP homelab; nothing here uses a secure-context Web API.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import logging
+import re
+import time
+from collections.abc import Iterator
+from dataclasses import dataclass
+from datetime import date
+from urllib.parse import unquote
+
+import requests
+
+from .native_ingest_common import (
+    NativeAuthError,
+    NativeDriftError,
+    NativeIngestError,
+    retry_after_seconds,
+)
+
+log = logging.getLogger(__name__)
+
+API_ROOT = "https://discord.com/api/v10"
+_ROOT = "https://discord.com"
+
+_TIMEOUT_SECONDS = 60.0
+_MESSAGES_BATCH = 100
+_THREADS_BATCH = 25
+# gallery-dl retries a 429 up to its default 4 retries, waiting
+# `request_interval_429` (60s) between them. Discord's Retry-After is exact, so
+# it is honoured when present; 60s is the fallback and the cap.
+_MAX_429_RETRIES = 4
+_429_WAIT_SECONDS = 60.0
+
+# https://discord.com/developers/docs/resources/message#message-object-message-types
+# DEFAULT, REPLY, CHAT_INPUT_COMMAND — the ones that carry user content.
+MESSAGE_TYPES = frozenset({0, 19, 21})
+# https://discord.com/developers/docs/resources/channel#channel-object-channel-types
+_TEXT = frozenset({0, 5})               # text, announcement: messages + threads
+_DIRECT = frozenset({1, 3, 10, 11, 12})  # DMs and threads: messages only
+_FORUM = frozenset({15, 16})            # forum, media: threads only
+_CATEGORY = 4
+_SERVER_WALK = _TEXT | _FORUM
+_EMBED_TYPES = frozenset({"image", "gifv", "video"})
+
+_URL_RE = re.compile(
+    r"^(?:https?://)?(?:www\.|ptb\.|canary\.)?discord(?:app)?\.com/channels/"
+    r"(?P<server>@me|\d+)(?:/(?:\d+/threads/)?(?P<channel>\d+))?(?P<rest>/.*)?/?$"
+)
+
+
+class DiscordAPIError(NativeIngestError):
+    """Base for native Discord client failures."""
+
+
+class DiscordAuthError(DiscordAPIError, NativeAuthError):
+    """401 (the token is invalid or expired) or a 403 on the channel the
+    source names. The fix is a new token, not a new client."""
+
+
+class DiscordDriftError(DiscordAPIError, NativeDriftError):
+    """A response did not have the shape the walk depends on."""
+
+
+def firefox_user_agent(today: date | None = None) -> str:
+    """gallery-dl's default User-Agent: a Firefox whose version advances every
+    four weeks (`util._ff_ver`, "147 on 2026-01-13"). Computed the same way so
+    the profile keeps matching the gallery-dl this replaced."""
+    ver = ((today or date.today()).toordinal() - 735_513) // 28
+    return (
+        f"Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:{ver}.0) "
+        f"Gecko/20100101 Firefox/{ver}.0"
+    )
+
+
+def nameext_from_url(url: str) -> tuple[str, str]:
+    """gallery-dl's `text.nameext_from_url`: the URL's last path segment,
+    unquoted, split at the last dot when the extension is at most 16 chars
+    (lowercased); otherwise the whole name and no extension."""
+    filename = unquote(url.partition("?")[0].rpartition("/")[2])
+    name, _, ext = filename.rpartition(".")
+    if name and len(ext) <= 16:
+        return name, ext.lower()
+    return filename, ""
+
+
+def parse_source_url(url: str) -> tuple[str | None, str | None]:
+    """`(server_id, channel_id)` from a Discord channel/server URL. `server_id`
+    is None for a DM (`@me`); `channel_id` is None for a whole server. Raises
+    DiscordAPIError for anything else, including a link to a single message —
+    a message is not something a source can subscribe to."""
+    m = _URL_RE.match((url or "").strip())
+    if not m or (m.group("rest") or "").strip("/"):
+        raise DiscordAPIError(
+            f"Not a Discord channel or server link: {url!r} "
+            "(expected https://discord.com/channels/<server>[/<channel>])"
+        )
+    server = m.group("server")
+    channel = m.group("channel")
+    if server == "@me":
+        if not channel:
+            raise DiscordAPIError(f"A DM link needs a channel id: {url!r}")
+        return None, channel
+    return server, channel
+
+
+def message_text(message: dict) -> str:
+    """gallery-dl's `extract_message_text`: the body plus the text of rich
+    embeds and polls, newline-joined, empties dropped."""
+    parts = [message.get("content") or ""]
+    for embed in message.get("embeds") or []:
+        if embed.get("type") != "rich":
+            continue
+        parts.append((embed.get("author") or {}).get("name") or "")
+        parts.append(embed.get("title") or "")
+        parts.append(embed.get("description") or "")
+        for fld in embed.get("fields") or []:
+            parts.append(fld.get("name") or "")
+            parts.append(fld.get("value") or "")
+        parts.append((embed.get("footer") or {}).get("text") or "")
+    poll = message.get("poll")
+    if poll:
+        parts.append(((poll.get("question") or {}).get("text")) or "")
+        for answer in poll.get("answers") or []:
+            parts.append(((answer.get("poll_media") or {}).get("text")) or "")
+    return "\n".join(p for p in parts if p)
+
+
+@dataclass
+class MediaItem:
+    """One file of a Discord message. `filename`/`extension` are gallery-dl's
+    split of the URL; `num` is its 1-based position across the message's files,
+    which is what names it on disk.
+
+    `media_id` is what the seen-ledger keys on, and it is deliberately NOT
+    `num`: an edit that removes a file renumbers the ones after it, and a
+    positional key would then call a different file seen. It is the
+    attachment's id, or for an embed (which has none) a hash of its URL path —
+    the query string is a signature that changes on every fetch. `filehash` is
+    always None; nothing in a signed CDN URL is a content hash."""
+
+    url: str
+    filename: str
+    extension: str
+    kind: str
+    post_id: str
+    num: int
+    media_id: str
+    filehash: str | None = None
+
+
+class DiscordClient:
+    """Synchronous Discord API v10 read client for one user token."""
+
+    def __init__(
+        self,
+        token: str | None,
+        *,
+        request_sleep: float = 0.0,
+        max_retries: int = _MAX_429_RETRIES,
+        session: requests.Session | None = None,
+    ):
+        self._session = session or requests.Session()
+        self._session.headers.update({
+            "User-Agent": firefox_user_agent(),
+            "Accept": "*/*",
+            "Accept-Language": "en-US,en;q=0.5",
+            "Referer": _ROOT + "/",
+        })
+        if token:
+            self._session.headers["Authorization"] = token
+        self._token = token
+        self._request_sleep = request_sleep or 0.0
+        self._max_retries = max_retries
+        self._server: dict = {}
+        self._channels: dict[str, dict] = {}
+        self._skip_feed = False
+
+    # -- request -----------------------------------------------------------
+
+    def _get(self, endpoint: str, params: dict | None = None):
+        if not self._token:
+            raise DiscordAuthError("No Discord token is configured for this source")
+        if self._request_sleep > 0:
+            time.sleep(self._request_sleep)
+        url = API_ROOT + endpoint
+        attempt = 0
+        while True:
+            try:
+                resp = self._session.get(url, params=params, timeout=_TIMEOUT_SECONDS)
+            except requests.RequestException as exc:
+                raise DiscordAPIError(f"Discord request failed ({endpoint}): {exc}") from exc
+            if resp.status_code == 429 and attempt < self._max_retries:
+                attempt += 1
+                delay = retry_after_seconds(
+                    resp, attempt, base=_429_WAIT_SECONDS, cap=_429_WAIT_SECONDS,
+                )
+                log.warning(
+                    "Discord 429 (%s) — waiting %.1fs (retry %d/%d)",
+                    endpoint, delay, attempt, self._max_retries,
+                )
+                time.sleep(delay)
+                continue
+            break
+        if resp.status_code == 401:
+            raise DiscordAuthError(
+                "Discord rejected the token (HTTP 401) — it is invalid or has "
+                "expired; copy a fresh one from the browser",
+                status_code=401,
+            )
+        if resp.status_code != 200:
+            raise DiscordAPIError(
+                f"Discord returned HTTP {resp.status_code} ({endpoint})",
+                status_code=resp.status_code,
+                retry_after=_retry_after(resp),
+            )
+        try:
+            return resp.json()
+        except ValueError as exc:
+            raise DiscordDriftError(
+                f"Discord returned non-JSON for {endpoint} ({len(resp.content)} bytes)"
+            ) from exc
+
+    # -- metadata (gallery-dl parse_server / parse_channel) -----------------
+
+    def _load_server(self, server_id: str) -> None:
+        server = self._get(f"/guilds/{server_id}")
+        if not isinstance(server, dict) or "id" not in server:
+            raise DiscordDriftError(f"Discord server {server_id} came back without an id")
+        self._server = {
+            "server": server.get("name") or "",
+            "server_id": str(server["id"]),
+            "owner_id": server.get("owner_id"),
+        }
+        channels = self._get(f"/guilds/{server_id}/channels")
+        if not isinstance(channels, list):
+            raise DiscordDriftError(f"Discord server {server_id} channel list is not a list")
+        # Categories first, so every child can name its parent.
+        for channel in sorted(channels, key=lambda ch: ch.get("type") != _CATEGORY):
+            self._parse_channel(channel)
+
+    def _parse_channel(self, channel: dict) -> dict:
+        parent_id = channel.get("parent_id")
+        meta = {
+            "channel": channel.get("name") or "",
+            "channel_id": str(channel.get("id")),
+            "channel_type": channel.get("type"),
+            "channel_topic": channel.get("topic") or "",
+            "parent_id": parent_id,
+            "is_thread": "thread_metadata" in channel,
+        }
+        parent = self._channels.get(parent_id) if parent_id else None
+        if parent:
+            meta["parent"] = parent["channel"]
+            meta["parent_type"] = parent["channel_type"]
+        if meta["channel_type"] in {1, 3}:
+            recipients = channel.get("recipients") or []
+            meta["channel"] = "DMs"
+            meta["recipients"] = [u.get("username") for u in recipients]
+            meta["recipients_id"] = [u.get("id") for u in recipients]
+        self._channels[meta["channel_id"]] = meta
+        return meta
+
+    def _channel_meta(self, channel_id: str) -> dict:
+        if channel_id not in self._channels:
+            self._parse_channel(self._get(f"/channels/{channel_id}"))
+        return self._channels[channel_id]
+
+    def _threads(self, channel_id: str) -> list[dict]:
+        """Every thread of a channel or forum, newest-created first (see the
+        module docstring for why not last-message order)."""
+        threads: list[dict] = []
+        offset = 0
+        while True:
+            data = self._get(f"/channels/{channel_id}/threads/search", {
+                "sort_by": "last_message_time",
+                "sort_order": "desc",
+                "limit": _THREADS_BATCH,
+                "offset": offset,
+            })
+            batch = (data.get("threads") or []) if isinstance(data, dict) else []
+            threads.extend(batch)
+            if len(batch) < _THREADS_BATCH:
+                break
+            offset += len(batch)
+        threads.sort(key=lambda t: int(t.get("id") or 0), reverse=True)
+        return threads
+
+    # -- the walk ------------------------------------------------------------
+
+    def _feeds(self, channel_id: str, *, safe: bool) -> Iterator[tuple[str, bool]]:
+        """`(channel_id, safe)` for every message feed under `channel_id`, in
+        gallery-dl's order. `safe` feeds are skipped on a 403."""
+        try:
+            ctype = self._channel_meta(channel_id)["channel_type"]
+        except DiscordAPIError as exc:
+            if exc.status_code != 403:
+                raise
+            if not safe:
+                raise DiscordAuthError(
+                    f"The Discord token cannot see channel {channel_id} (HTTP 403)",
+                    status_code=403,
+                ) from exc
+            log.info("Discord: no access to channel %s — skipped", channel_id)
+            return
+        if ctype in _TEXT or ctype in _DIRECT:
+            yield channel_id, safe
+        if ctype in _TEXT or ctype in _FORUM:
+            try:
+                threads = self._threads(channel_id)
+            except DiscordAPIError as exc:
+                if exc.status_code != 403:
+                    raise
+                log.info("Discord: cannot list threads of %s — skipped", channel_id)
+                threads = []
+            for thread in threads:
+                yield self._parse_channel(thread)["channel_id"], True
+        elif ctype == _CATEGORY:
+            for child in list(self._channels.values()):
+                if child.get("parent_id") == channel_id:
+                    yield from self._feeds(child["channel_id"], safe=True)
+        elif ctype not in _DIRECT and not safe:
+            raise DiscordAPIError(
+                f"Discord channel {channel_id} is of type {ctype}, which has no messages"
+            )
+
+    def _source_feeds(self, url: str) -> Iterator[tuple[str, bool]]:
+        server_id, channel_id = parse_source_url(url)
+        self._server, self._channels = {}, {}
+        if server_id is not None:
+            self._load_server(server_id)
+        if channel_id is not None:
+            yield from self._feeds(channel_id, safe=False)
+            return
+        for meta in list(self._channels.values()):
+            if meta["channel_type"] in _SERVER_WALK:
+                yield from self._feeds(meta["channel_id"], safe=True)
+
+    def skip_feed(self) -> None:
+        """Optional core seam (#4413): end the current channel and go on to the
+        next one. A tick's early-out means THIS channel has nothing new, not
+        that the server has nothing new."""
+        self._skip_feed = True
+
+    def iter_posts(
+        self, campaign_id: str, cursor: str | None = None
+    ) -> Iterator[tuple[dict, dict, str | None]]:
+        """Yield `(message, channel_meta, page_cursor)` for every content
+        message the source reaches, channel by channel, each newest first.
+
+        `campaign_id` is the source URL. The cursor is `<channel_id>:<before>`
+        — the channel and the `before` id that fetched the page (empty for a
+        channel's first page) — so a backfill resumes inside the right channel
+        and re-fetches the page it was cut in. A cursor naming a channel the
+        walk no longer reaches (a deleted thread) restarts from the top rather
+        than walking nothing.
+        """
+        resume_channel, _, resume_before = (cursor or "").partition(":")
+        resuming = bool(resume_channel)
+        feeds = list(self._source_feeds(campaign_id)) if resuming else None
+        if feeds is not None and resume_channel not in {cid for cid, _ in feeds}:
+            log.warning(
+                "Discord: resume channel %s is no longer in %s — restarting",
+                resume_channel, campaign_id,
+            )
+            resuming = False
+        for channel_id, safe in feeds if feeds is not None else self._source_feeds(campaign_id):
+            before = None
+            if resuming:
+                if channel_id != resume_channel:
+                    continue
+                resuming = False
+                before = resume_before or None
+            yield from self._iter_channel(channel_id, before, safe=safe)
+
+    def _iter_channel(
+        self, channel_id: str, before: str | None, *, safe: bool
+    ) -> Iterator[tuple[dict, dict, str | None]]:
+        self._skip_feed = False
+        meta = {**self._server, **self._channels.get(channel_id, {})}
+        while True:
+            page_cursor = f"{channel_id}:{before or ''}"
+            try:
+                messages = self._get(
+                    f"/channels/{channel_id}/messages",
+                    {"limit": _MESSAGES_BATCH, "before": before},
+                )
+            except DiscordAPIError as exc:
+                if exc.status_code != 403:
+                    raise
+                if not safe:
+                    raise DiscordAuthError(
+                        f"The Discord token cannot read channel {channel_id} (HTTP 403)",
+                        status_code=403,
+                    ) from exc
+                log.info("Discord: no access to messages of %s — skipped", channel_id)
+                return
+            if not isinstance(messages, list):
+                raise DiscordDriftError(
+                    f"Discord messages of {channel_id} came back as "
+                    f"{type(messages).__name__}, not a list"
+                )
+            for message in messages:
+                if message.get("type") not in MESSAGE_TYPES:
+                    continue
+                message["_meta"] = meta
+                yield message, meta, page_cursor
+                if self._skip_feed:
+                    return
+            if len(messages) < _MESSAGES_BATCH:
+                return
+            before = str(messages[-1]["id"])
+
+    # -- per-message -------------------------------------------------------
+
+    @staticmethod
+    def extract_media(post: dict, included: dict | None = None) -> list[MediaItem]:
+        """gallery-dl's file list for one message: attachments, then the first
+        of video/image/thumbnail `proxy_url` of each file-bearing embed, then
+        the same for every forwarded snapshot; numbered from 1 across them."""
+        mid = str(post.get("id") or "")
+        snapshots = [post] + [
+            (s or {}).get("message") or {}
+            for s in post.get("message_snapshots") or []
+            if ((s or {}).get("message") or {}).get("type", 0) in MESSAGE_TYPES
+        ]
+        found: list[tuple[str, str, str | None]] = []
+        for snap in snapshots:
+            for att in snap.get("attachments") or []:
+                if att.get("url"):
+                    aid = att.get("id")
+                    found.append((att["url"], "attachment", str(aid) if aid else None))
+            for embed in snap.get("embeds") or []:
+                if embed.get("type") not in _EMBED_TYPES:
+                    continue
+                for fld in ("video", "image", "thumbnail"):
+                    url = (embed.get(fld) or {}).get("proxy_url")
+                    if url:
+                        found.append((url, "embed", None))
+                        break
+        items = []
+        for num, (url, kind, fid) in enumerate(found, start=1):
+            name, ext = nameext_from_url(url)
+            if fid is None:
+                path = url.partition("?")[0].encode()
+                fid = "u" + hashlib.sha1(path, usedforsecurity=False).hexdigest()[:32]
+            items.append(MediaItem(
+                url=url, filename=name, extension=ext, kind=kind, post_id=mid,
+                num=num, media_id=fid,
+            ))
+        return items
+
+    @staticmethod
+    def post_meta(post: dict) -> dict:
+        """No title (Discord has none); `date` is the message timestamp, ISO
+        with an offset — what the core's revisit window reads."""
+        return {"title": None, "date": post.get("timestamp")}
+
+    @classmethod
+    def post_record_key(cls, post: dict) -> tuple[str, str] | None:
+        """`(message:<id>, <id>)` — gates the message record through the seen
+        ledger, like `post:<id>` on the other platforms.
+
+        None for a message with no files. gallery-dl wrote a sidecar only
+        beside a file, so a text-only chat line never became a post, and the
+        drop grouping (discord_grouping) is built on that: a channel's chatter
+        recorded as posts would bury the drops it exists to surface."""
+        mid = post.get("id")
+        mid = str(mid) if mid is not None else ""
+        if not mid or not cls.extract_media(post):
+            return None
+        return (f"message:{mid}", mid)
+
+    # -- verify ------------------------------------------------------------
+
+    def describe(self, server_id: str | None, channel_id: str | None) -> dict:
+        """The display names behind a server/channel pair, for the browser
+        extension's Add panel. Best-effort per name: one that can't be read
+        comes back None, and the other is still returned."""
+        out: dict = {"server": None, "channel": None, "parent": None}
+        if server_id:
+            try:
+                out["server"] = (self._get(f"/guilds/{server_id}") or {}).get("name") or None
+            except DiscordAPIError:
+                pass
+        if channel_id:
+            try:
+                meta = self._parse_channel(self._get(f"/channels/{channel_id}"))
+                out["channel"] = meta.get("channel") or None
+            except (DiscordAPIError, AttributeError):
+                pass
+        return out
+
+    def verify_auth(self, url: str) -> tuple[bool | None, str]:
+        """Is the token valid, and can it see what the source names?"""
+        try:
+            server_id, channel_id = parse_source_url(url)
+        except DiscordAPIError as exc:
+            return None, str(exc)
+        try:
+            me = self._get("/users/@me")
+            if channel_id is not None:
+                self._get(f"/channels/{channel_id}")
+            elif server_id is not None:
+                self._get(f"/guilds/{server_id}")
+        except DiscordAuthError as exc:
+            return False, f"Discord rejected the token — {exc}"
+        except DiscordAPIError as exc:
+            if exc.status_code in (403, 404):
+                return False, (
+                    "The token is valid, but its account cannot see "
+                    f"{'this channel' if channel_id else 'this server'} "
+                    f"(HTTP {exc.status_code})"
+                )
+            return None, f"Couldn't verify (network/HTTP issue): {exc}"
+        who = (me or {}).get("username") if isinstance(me, dict) else None
+        return True, f"Token valid{f' ({who})' if who else ''} — the source is readable."
+
+
+def _retry_after(resp: requests.Response) -> float | None:
+    hdr = resp.headers.get("Retry-After")
+    try:
+        return float(hdr) if hdr else None
+    except (TypeError, ValueError):
+        return None

backend/app/services/discord_downloader.py

@@ -0,0 +1,212 @@
+"""Native Discord media downloader — the Discord counterpart to
+subscribestar_downloader.
+
+Writes files exactly where gallery-dl wrote them, so a cutover finds every
+existing file on disk (`skipped_disk`) instead of fetching it again:
+
+    <images>/<artist>/discord/<channel>/<YYYYMMDD>_<message_id>_<NN>_<name>.<ext>
+
+That is what FC's gallery-dl config produced (directory `{channel}`, filename
+`{date:%Y%m%d}_{message_id}_{num:>02}_{filename}.{extension}`, under the
+per-source base directory `<images>/<artist>/<platform>`), retired from that
+config once Discord moved here; tests/test_discord_naming.py pins the match
+against a real gallery-dl sidecar. The name is cleaned the way gallery-dl cleans it
+on Linux — `/` becomes `_` and control characters are removed, nothing else
+(`path-restrict: auto`, `path-remove` defaults). It is NOT `sanitize_segment`,
+whose Windows set would turn a `:` in a channel or file name into `_` and miss
+the file gallery-dl wrote.
+
+Post-first (rule 120): each file gets a minimal sidecar named like it minus the
+extension (what `find_sidecar` pairs first), and the message itself gets one
+record, `<YYYYMMDD>_<message_id>_post.json`, carrying gallery-dl's metadata keys
+— `message_id` for the post id, `server_id`/`channel_id` for the permalink,
+`message` for the body, `date` — so `parse_sidecar` reads it exactly as it read
+the gallery-dl sidecars. Neither file carries an `id` or `post_id` key: both
+outrank `message_id` in the post-id chain (`platforms.base`).
+
+PURE: no DB; the seen-skip is an injected predicate.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import re
+import time
+from collections.abc import Callable
+from datetime import UTC, datetime
+from pathlib import Path
+
+import requests
+
+from .discord_client import firefox_user_agent, message_text
+from .native_ingest_common import (
+    BaseNativeDownloader,
+    MediaOutcome,
+    PostRecordOutcome,
+    make_session,
+)
+
+log = logging.getLogger(__name__)
+
+PLATFORM = "discord"
+_CONTROL = re.compile("[\x00-\x1f\x7f]")
+# gallery-dl falls back to the response's type for a URL with no extension;
+# we never see the response before naming, and such URLs do not occur for
+# Discord attachments or embed proxies in practice.
+_NO_EXTENSION = "bin"
+
+
+def gdl_clean(segment: str) -> str:
+    """One path segment as gallery-dl writes it on Linux."""
+    return _CONTROL.sub("", segment.replace("/", "_"))
+
+
+def message_date(post: dict) -> datetime | None:
+    raw = post.get("timestamp")
+    if not isinstance(raw, str) or not raw:
+        return None
+    try:
+        dt = datetime.fromisoformat(raw.replace("Z", "+00:00"))
+    except ValueError:
+        return None
+    return (dt if dt.tzinfo else dt.replace(tzinfo=UTC)).astimezone(UTC)
+
+
+def channel_dir(images_root: Path, artist_slug: str, post: dict) -> Path:
+    """gallery-dl's `{channel}` directory; an empty name adds no segment."""
+    base = Path(images_root) / artist_slug / PLATFORM
+    channel = gdl_clean(((post.get("_meta") or {}).get("channel") or "").strip())
+    return base / channel if channel else base
+
+
+def media_stem(post: dict, media) -> str:
+    """`<YYYYMMDD>_<message_id>_<NN>_<name>` — the file's name minus `.<ext>`."""
+    when = message_date(post)
+    day = f"{when:%Y%m%d}" if when else "None"
+    return gdl_clean(f"{day}_{post.get('id')}_{media.num:>02}_{media.filename}")
+
+
+class DiscordDownloader(BaseNativeDownloader):
+    """Download a message's files to gallery-dl's layout. The CDN gets
+    gallery-dl's browser profile and no token — gallery-dl sends the token only
+    to the API, and the CDN URLs are pre-signed."""
+
+    def __init__(
+        self,
+        images_root: Path,
+        cookies_path: str | None = None,
+        *,
+        validate: bool = True,
+        rate_limit: float = 0.0,
+        session: requests.Session | None = None,
+    ):
+        super().__init__(
+            images_root, None, platform=PLATFORM,
+            validate=validate, rate_limit=rate_limit,
+            session=session if session is not None else make_session(None, extra_headers={
+                "User-Agent": firefox_user_agent(),
+                "Accept-Language": "en-US,en;q=0.5",
+                "Referer": "https://discord.com/",
+            }),
+        )
+
+    def download_post(
+        self,
+        post: dict,
+        media_items: list,
+        artist_slug: str,
+        *,
+        is_seen: Callable[[object], bool] = lambda m: False,
+        should_stop: Callable[[], bool] = lambda: False,
+        recapture: bool = False,
+    ) -> list[MediaOutcome]:
+        """Every file of one message; per-file outcomes, one failure isolated."""
+        folder = channel_dir(self.images_root, artist_slug, post)
+        outcomes: list[MediaOutcome] = []
+        for media in media_items:
+            if should_stop():
+                break
+            try:
+                outcomes.append(self._download_one(
+                    post, media, folder, artist_slug, is_seen, recapture=recapture,
+                ))
+            except Exception as exc:  # resilient: isolate one item's failure
+                log.warning(
+                    "Discord media failed (message %s, file %d): %s",
+                    post.get("id"), media.num, exc,
+                )
+                outcomes.append(
+                    MediaOutcome(media=media, status="error", path=None, error=str(exc))
+                )
+        return outcomes
+
+    def _download_one(
+        self,
+        post: dict,
+        media,
+        folder: Path,
+        artist_slug: str,
+        is_seen: Callable[[object], bool],
+        *,
+        recapture: bool = False,
+    ) -> MediaOutcome:
+        seen = is_seen(media)
+        if seen and not recapture:
+            return MediaOutcome(media=media, status="skipped_seen", path=None, error=None)
+        stem = media_stem(post, media)
+        path = folder / f"{stem}.{media.extension or _NO_EXTENSION}"
+        if path.exists():  # tier-2: gallery-dl (or an earlier walk) wrote it
+            return MediaOutcome(media=media, status="skipped_disk", path=path, error=None)
+        if seen:  # recapture never re-fetches a seen file that is gone
+            return MediaOutcome(media=media, status="skipped_seen", path=None, error=None)
+
+        folder.mkdir(parents=True, exist_ok=True)
+        if self._rate_limit > 0:
+            time.sleep(self._rate_limit)
+        out = self._fetch_get(media.url, path)
+        reason, quarantined = self._validate_path(out, artist_slug, media.url)
+        if reason is not None:
+            return MediaOutcome(media=media, status="quarantined", path=quarantined, error=reason)
+        sidecar = {"category": PLATFORM, "message_id": str(post.get("id") or "")}
+        sidecar["source_url"] = media.url
+        (folder / f"{stem}.json").write_text(json.dumps(sidecar, indent=2))
+        return MediaOutcome(media=media, status="downloaded", path=out, error=None)
+
+    def write_post_record(
+        self, post: dict, artist_slug: str, *, revisit: bool = False,
+    ) -> PostRecordOutcome:
+        """The message record — the one writer of a Discord post's body, date
+        and permalink ids. `revisit` re-reads a message already captured (an
+        edit); an empty re-read writes nothing, so it never blanks a body."""
+        mid = str(post.get("id") or "")
+        body = message_text(post)
+        if not mid or (revisit and not body.strip()):
+            return PostRecordOutcome(path=None, post_type=None, title=None, body_chars=0)
+        meta = post.get("_meta") or {}
+        author = post.get("author") or {}
+        record = {
+            "category": PLATFORM,
+            "message_id": mid,
+            "server": meta.get("server"),
+            "server_id": meta.get("server_id"),
+            "channel": meta.get("channel"),
+            "channel_id": meta.get("channel_id") or post.get("channel_id"),
+            "parent": meta.get("parent"),
+            "is_thread": meta.get("is_thread"),
+            "author": author.get("username"),
+            "author_id": author.get("id"),
+            "message": body,
+            "date": post.get("timestamp"),
+        }
+        folder = channel_dir(self.images_root, artist_slug, post)
+        folder.mkdir(parents=True, exist_ok=True)
+        when = message_date(post)
+        day = f"{when:%Y%m%d}" if when else "None"
+        path = folder / f"{day}_{mid}_post.json"
+        path.write_text(json.dumps(
+            {k: v for k, v in record.items() if v is not None}, indent=2,
+        ))
+        return PostRecordOutcome(
+            path=path, post_type=None, title=None, body_chars=len(body),
+        )

backend/app/services/discord_grouping.py

@@ -59,14 +59,23 @@ from __future__ import annotations
 
 import logging
 import math
+from collections import Counter
 from dataclasses import dataclass, field
 from datetime import UTC, datetime, timedelta
 
-from sqlalchemy import Select, func, select, update
+from sqlalchemy import Select, delete, func, select, union, update
 from sqlalchemy.dialects.postgresql import insert as pg_insert
 from sqlalchemy.ext.asyncio import AsyncSession
 
-from ..models import ImageProvenance, ImageRecord, MLSettings, Post, Source
+from ..models import (
+    ImageProvenance,
+    ImageRecord,
+    MLSettings,
+    Post,
+    PostAssociation,
+    Source,
+)
+from .post_naming import FAMILY_MAX_POSTS, leading_name, rarity, token_frequencies
 
 log = logging.getLogger(__name__)
 
@@ -117,6 +126,31 @@ def cosine_distance(a, b) -> float:
     return 1.0 - (dot / (na * nb))
 
 
+def _message_images(posts):
+    """(post_id, image_id) for every image a message carries — owned AND re-posted.
+
+    A message owns an image through `primary_post_id`, but only the FIRST
+    message imported with a given file does. The same file posted again is a
+    provenance link, and the backfill runs newest-first, so it is usually the
+    ORIGINAL message that ends up owning nothing. Reading ownership alone left
+    101 of Yellowroom's messages (2018–2020 mostly) ungroupable: every image
+    they carried also sat in another message.
+
+    `posts` is a list of ids or a select of them; filtering both branches by it
+    keeps the union to the messages in hand rather than the whole library.
+    Callers pass MESSAGE posts only — a drop's own provenance rows would read
+    as images it carries.
+    """
+    owned = select(
+        ImageRecord.primary_post_id.label("post_id"), ImageRecord.id.label("image_id"),
+    ).where(ImageRecord.primary_post_id.in_(posts))
+    reposted = select(
+        ImageProvenance.post_id.label("post_id"),
+        ImageProvenance.image_record_id.label("image_id"),
+    ).where(ImageProvenance.post_id.in_(posts))
+    return union(owned, reposted).subquery()
+
+
 def _candidate_stmt(source_id: int, *, not_after: datetime) -> Select:
     """Ungrouped Discord message-posts, one representative image each, OLDEST
     FIRST — which is the order `build_groups` requires.
@@ -134,13 +168,15 @@ def _candidate_stmt(source_id: int, *, not_after: datetime) -> Select:
     take the OLDEST candidates instead of the lowest-numbered ones.
     """
     sort_key = func.coalesce(Post.post_date, Post.downloaded_at)
+    carried = _message_images(select(Post.id).where(Post.source_id == source_id))
     inner = (
         select(
             Post.id.label("post_id"),
             sort_key.label("occurred_at"),
             ImageRecord.siglip_embedding.label("embedding"),
         )
-        .join(ImageRecord, ImageRecord.primary_post_id == Post.id)
+        .join(carried, carried.c.post_id == Post.id)
+        .join(ImageRecord, ImageRecord.id == carried.c.image_id)
         .where(
             Post.source_id == source_id,
             # Never absorb a post FC wrote, and never re-absorb one already
@@ -286,9 +322,10 @@ async def _link_member_images(
     """
     if not member_ids:
         return 0
-    image_rows = (await session.execute(
-        select(ImageRecord.id).where(ImageRecord.primary_post_id.in_(member_ids))
-    )).scalars().all()
+    carried = _message_images(member_ids)
+    image_rows = sorted(set((await session.execute(
+        select(carried.c.image_id)
+    )).scalars().all()))
     if not image_rows:
         return 0
     await session.execute(
@@ -412,9 +449,11 @@ async def _group_seed(session: AsyncSession, post_id: int) -> list[float] | None
     free to disagree; this way there is one.
     """
     sort_key = func.coalesce(Post.post_date, Post.downloaded_at)
+    carried = _message_images(select(Post.id).where(Post.absorbed_by_post_id == post_id))
     return (await session.execute(
         select(ImageRecord.siglip_embedding)
-        .join(Post, ImageRecord.primary_post_id == Post.id)
+        .join(carried, carried.c.image_id == ImageRecord.id)
+        .join(Post, Post.id == carried.c.post_id)
         .where(
             Post.absorbed_by_post_id == post_id,
             ImageRecord.siglip_embedding.is_not(None),
@@ -600,6 +639,352 @@ async def join_open_groups(
     return joined
 
 
+# ---------------------------------------------------------------------------
+# #4390: a trickle is one drop — later drops of the same piece merge in.
+# ---------------------------------------------------------------------------
+#
+# Operator, 2026-09-24, on a feed of "Grouped from 1 Discord message" cards:
+# *"the groups are still single image even when they can clearly be seen as
+# group"*. 665 of Yellowroom's 714 drops were one message.
+#
+# Both paths above join on cosine distance to the group's SEED, and the stages
+# of one piece fail it: `svtt_wip4` did not join `svtt_wip3` from the day
+# before. A creator trickles a piece out as sketch -> wip -> wip -> release, and
+# each stage is nearest to the one before it, not to the first.
+#
+# Measured on artist 8 before any of this was written (#4390 log):
+#
+# * phash cannot see it. Stages of one piece sit 68-134 bits apart; unrelated
+#   pieces by the same artist sit at a median of 126, p5 110. Lesson #4400.
+# * The embedding's NEAREST neighbour can. Every stage of three real trickles
+#   had a sibling stage as its single nearest image in the artist's whole
+#   library, while siblings further along ranked 40-100 — which is exactly why
+#   seed distance fails. Negative control over all 137 recent Discord images:
+#   where the nearest neighbour was another message within 7 days, the two
+#   carried the same working name 53 times out of 53. Disagreements start past
+#   7 days.
+# * The working name sees it directly, when there is one.
+#
+# So a later drop merges into an earlier one when the two are within
+# `discord_group_close_after_hours` of each other (168h — the measured 7 days)
+# AND either they share a gated LEADING working name, or one's image has the
+# other's image as its nearest neighbour. A drop reaching SEVERAL earlier drops
+# pulls them all together — unless two of them are named as different pieces,
+# in which case nothing moves (see `_compatible`): leaving a drop alone is
+# recoverable, a wrong merge asserts that unrelated art belongs together.
+#
+# Chaining is permitted here and was forbidden above, deliberately. The seed
+# rule exists because tiny steps can drift from one piece to another; the
+# measured precision of nearest-neighbour inside 7 days is what bounds drift
+# for this route, and each link is between neighbours in time, never across a
+# quiet week.
+
+# How many unchecked drops one sweep examines per source. A first run over an
+# established library drains over successive sweeps, oldest first, rather than
+# issuing one nearest-neighbour query per image of the whole history at once.
+TRICKLE_BATCH = 300
+
+
+@dataclass
+class _Drop:
+    post: Post
+    members: set[int]
+    first_at: datetime
+    last_at: datetime
+    names: set[str]
+    images: list[int]
+    nearest: set[int] | None
+
+
+async def _nearest_message(
+    session: AsyncSession, *, artist_id: int, image_id: int, exclude: set[int],
+) -> int | None:
+    """The post that owns the nearest image in the artist's whole library.
+
+    The whole LIBRARY, not this source, because that is what was measured: a
+    neighbour that turns out to be a Patreon re-post simply yields no Discord
+    drop to merge into, which errs toward leaving things alone. `exclude` is
+    the drop's own messages — an image is always nearest to its own siblings
+    in the same drop, which says nothing.
+    """
+    embedding = (await session.execute(
+        select(ImageRecord.siglip_embedding).where(ImageRecord.id == image_id)
+    )).scalar_one_or_none()
+    if embedding is None:
+        return None
+    stmt = (
+        select(ImageRecord.primary_post_id)
+        .where(
+            ImageRecord.artist_id == artist_id,
+            ImageRecord.id != image_id,
+            ImageRecord.siglip_embedding.is_not(None),
+            ImageRecord.primary_post_id.is_not(None),
+        )
+        .order_by(ImageRecord.siglip_embedding.cosine_distance(embedding))
+        .limit(1)
+    )
+    if exclude:
+        stmt = stmt.where(ImageRecord.primary_post_id.not_in(exclude))
+    return (await session.execute(stmt)).scalar_one_or_none()
+
+
+async def _load_drops(session: AsyncSession, source: Source) -> list[_Drop]:
+    """Every live drop of this source, with what the merge rule reads, oldest first."""
+    posts = (await session.execute(
+        select(Post).where(
+            Post.source_id == source.id,
+            Post.synthesized_by == DROP_GROUPER,
+            Post.absorbed_by_post_id.is_(None),
+        )
+    )).scalars().all()
+    if not posts:
+        return []
+    by_id = {p.id: p for p in posts}
+
+    msg_at = func.coalesce(Post.post_date, Post.downloaded_at)
+    members: dict[int, set[int]] = {pid: set() for pid in by_id}
+    times: dict[int, list[datetime]] = {pid: [] for pid in by_id}
+    for mid, owner, at in (await session.execute(
+        select(Post.id, Post.absorbed_by_post_id, msg_at)
+        .where(Post.absorbed_by_post_id.in_(list(by_id)))
+    )).all():
+        members[owner].add(mid)
+        times[owner].append(at)
+
+    owner_of = {m: d for d, ms in members.items() for m in ms}
+    names: dict[int, set[str]] = {pid: set() for pid in by_id}
+    images: dict[int, list[int]] = {pid: [] for pid in by_id}
+    if owner_of:
+        carried = _message_images(list(owner_of))
+        for iid, message, path in (await session.execute(
+            select(ImageRecord.id, carried.c.post_id, ImageRecord.path)
+            .join(carried, carried.c.image_id == ImageRecord.id)
+            .order_by(ImageRecord.id)
+        )).all():
+            drop = owner_of[message]
+            if iid in images[drop]:
+                continue  # one file carried by two of the drop's messages
+            images[drop].append(iid)
+            if (name := leading_name(path)) is not None:
+                names[drop].add(name)
+
+    out = []
+    for pid, post in by_id.items():
+        if not times[pid]:
+            continue
+        stored = (post.synthesis_details or {}).get("nearest_message_ids")
+        out.append(_Drop(
+            post=post, members=members[pid],
+            first_at=min(times[pid]), last_at=max(times[pid]),
+            names=names[pid], images=images[pid],
+            nearest=set(stored) if stored is not None else None,
+        ))
+    return sorted(out, key=lambda d: (d.first_at, d.post.id))
+
+
+async def _name_posts(session: AsyncSession, artist_id: int) -> Counter[str]:
+    """Post-span counts of the artist's working names — the same corpus the
+    teaser card and the announcement matcher count against."""
+    by_post: dict[int, list[str]] = {}
+    for pid, path in (await session.execute(
+        select(ImageRecord.primary_post_id, ImageRecord.path).where(
+            ImageRecord.artist_id == artist_id,
+            ImageRecord.primary_post_id.is_not(None),
+        )
+    )).all():
+        by_post.setdefault(pid, []).append(path)
+    return token_frequencies(by_post.values())
+
+
+async def _repoint_associations(
+    session: AsyncSession, *, from_id: int, to_id: int,
+) -> None:
+    """Move announcement links from a drop about to merge onto the one it joins.
+
+    Without this the merge would silently undo a teaser link: the association's
+    payload FK cascades on delete. Where the teaser already points at the
+    surviving drop, the stronger claim is kept — a link over a proposal over a
+    dismissal — and the duplicate goes.
+    """
+    rank = {"linked": 2, "pending": 1, "dismissed": 0}
+    moving = (await session.execute(
+        select(PostAssociation).where(PostAssociation.payload_post_id == from_id)
+    )).scalars().all()
+    for a in moving:
+        existing = (await session.execute(
+            select(PostAssociation).where(
+                PostAssociation.announcement_post_id == a.announcement_post_id,
+                PostAssociation.payload_post_id == to_id,
+            )
+        )).scalar_one_or_none()
+        if existing is None:
+            a.payload_post_id = to_id
+            continue
+        if rank.get(a.status, 0) > rank.get(existing.status, 0):
+            existing.status = a.status
+            existing.linked_by = a.linked_by
+        await session.delete(a)
+    await session.flush()
+
+
+async def merge_trickles(
+    session: AsyncSession,
+    source: Source,
+    *,
+    gap: timedelta,
+    min_images: int,
+    cooldown: timedelta,
+    batch: int = TRICKLE_BATCH,
+) -> int:
+    """Fold later drops of the same piece into the earlier one. Returns merges."""
+    drops = await _load_drops(session, source)
+    if len(drops) < 2:
+        return 0
+    name_posts = await _name_posts(session, source.artist_id)
+
+    def gated(names: set[str]) -> set[str]:
+        return {n for n in names if rarity(name_posts.get(n, 0), FAMILY_MAX_POSTS) > 0}
+
+    alive: list[_Drop] = []
+    merged = 0
+    checked = 0
+    for drop in drops:
+        details = drop.post.synthesis_details or {}
+        if details.get("trickle_checked"):
+            alive.append(drop)
+            continue
+        if checked >= batch:
+            # Unchecked and out of budget: still a candidate for LATER drops'
+            # reverse edges, just not examined itself this run.
+            alive.append(drop)
+            continue
+        checked += 1
+
+        if drop.nearest is None:
+            found: set[int] = set()
+            for iid in drop.images:
+                pid = await _nearest_message(
+                    session, artist_id=source.artist_id, image_id=iid,
+                    exclude=drop.members,
+                )
+                if pid is not None:
+                    found.add(pid)
+            drop.nearest = found
+
+        mine = gated(drop.names)
+        targets: dict[int, tuple[_Drop, str]] = {}
+        for earlier in alive:
+            if drop.first_at - earlier.last_at > gap:
+                continue
+            shared = mine & gated(earlier.names)
+            if set(drop.images) & set(earlier.images):
+                # The creator posted the very same file again — the strongest
+                # evidence there is, and one nearest-neighbour cannot see: it
+                # skips the image itself, which is the one they share.
+                targets[earlier.post.id] = (earlier, "same_image")
+            elif shared:
+                targets[earlier.post.id] = (earlier, f"name:{min(shared)}")
+            elif drop.nearest & earlier.members or (earlier.nearest or set()) & drop.members:
+                targets[earlier.post.id] = (earlier, "nearest")
+
+        record = dict(details)
+        record["nearest_message_ids"] = sorted(drop.nearest)
+        record["trickle_checked"] = True
+        drop.post.synthesis_details = record
+
+        if not targets or not _compatible(
+            [gated(t.names) for t, _route in targets.values()] + [mine]
+        ):
+            alive.append(drop)
+            continue
+
+        # Every target is the same piece as this drop, so they are the same
+        # piece as each other: fold them all into the earliest, then this drop.
+        ordered = sorted(targets.values(), key=lambda tr: (tr[0].first_at, tr[0].post.id))
+        into = ordered[0][0]
+        for other, route in ordered[1:]:
+            await _merge_drop(
+                session, into=into, drop=other, route=route,
+                min_images=min_images, cooldown=cooldown,
+            )
+            alive.remove(other)
+            merged += 1
+        await _merge_drop(
+            session, into=into, drop=drop, route=ordered[0][1],
+            min_images=min_images, cooldown=cooldown,
+        )
+        merged += 1
+    return merged
+
+
+def _compatible(name_sets: list[set[str]]) -> bool:
+    """May drops carrying these working names become one post?
+
+    Refused only when two of them are NAMED AS DIFFERENT PIECES — both carry a
+    gated name, and they share none. An unnamed drop (a canvas screenshot)
+    fits anywhere, which is the whole of the Marin case: two early stages both
+    nearest to the same later one are one trickle, not an ambiguity.
+
+    What it does NOT refuse is a drop the creator made two pieces in
+    themselves. Measured on artist 8: one November message carries both
+    `AdL01_wip4` and `Year_20k_wip_z4`, so its drop holds both names, and a
+    later drop of either piece joins it on its own name. That is the creator's
+    co-posting carried forward — Discord shows those two together too — not a
+    bridge FC built.
+    """
+    named = [n for n in name_sets if n]
+    return all(a & b for i, a in enumerate(named) for b in named[i + 1:])
+
+
+async def _merge_drop(
+    session: AsyncSession,
+    *,
+    into: _Drop,
+    drop: _Drop,
+    route: str,
+    min_images: int,
+    cooldown: timedelta,
+) -> None:
+    """Absorb `drop`'s messages into `into`, carry its links over, delete it.
+
+    Growth is stamped at the merged messages' OWN time, not the wall clock.
+    Merging history must not drag a two-year-old drop to the top of the feed,
+    and the time the group actually grew is when those messages arrived.
+    """
+    await _repoint_associations(session, from_id=drop.post.id, to_id=into.post.id)
+    grew_before = into.post.last_grew_at
+    await _absorb_into(
+        session, group=into.post, member_ids=sorted(drop.members),
+        source_id=into.post.source_id, now=drop.last_at,
+        min_images=min_images, cooldown=cooldown,
+    )
+    # Never backwards: a group that already grew later than these messages
+    # keeps that later date.
+    if grew_before is not None and grew_before > drop.last_at:
+        into.post.last_grew_at = grew_before
+    details = dict(into.post.synthesis_details or {})
+    if grew_before is not None and grew_before > drop.last_at:
+        details["last_grew_at"] = grew_before.isoformat()
+    # The honesty rule, extended: a grouping FC invented says what it was
+    # built from, and a merge says WHY — "name:svtt" or "nearest".
+    details["merged"] = [
+        *details.get("merged", []),
+        {"post_id": drop.post.id, "route": route, "message_ids": sorted(drop.members)},
+    ]
+    details["nearest_message_ids"] = sorted((into.nearest or set()) | (drop.nearest or set()))
+    into.post.synthesis_details = details
+
+    into.members |= drop.members
+    into.names |= drop.names
+    into.images += drop.images
+    into.nearest = (into.nearest or set()) | (drop.nearest or set())
+    into.last_at = max(into.last_at, drop.last_at)
+
+    await session.execute(delete(ImageProvenance).where(ImageProvenance.post_id == drop.post.id))
+    await session.delete(drop.post)
+    await session.flush()
+
+
 async def sweep(session: AsyncSession, *, now: datetime | None = None) -> dict:
     """Group every enabled Discord source. No-op when the switch is off.
 
@@ -612,6 +997,7 @@ async def sweep(session: AsyncSession, *, now: datetime | None = None) -> dict:
     if not settings.discord_grouping_enabled:
         return {
             "enabled": False, "sources": 0, "posts_created": 0, "images_joined": 0,
+            "drops_merged": 0,
         }
 
     sources = (await session.execute(
@@ -626,6 +1012,7 @@ async def sweep(session: AsyncSession, *, now: datetime | None = None) -> dict:
 
     created = 0
     joined = 0
+    merged = 0
     for source in sources:
         joined += await join_open_groups(
             session, source,
@@ -644,12 +1031,20 @@ async def sweep(session: AsyncSession, *, now: datetime | None = None) -> dict:
             window_minutes=window_minutes,
             now=now,
         )
+        # Last, so the drops the two passes above just wrote are merged in
+        # the same sweep rather than showing as singletons for an hour.
+        merged += await merge_trickles(
+            session, source,
+            gap=timedelta(hours=float(settings.discord_group_close_after_hours)),
+            min_images=int(settings.discord_group_resurface_min_images),
+            cooldown=timedelta(hours=float(settings.discord_group_resurface_cooldown_hours)),
+        )
     log.info(
         "discord drop grouping: %d source(s), %d synthetic post(s) created, "
-        "%d image(s) joined to open groups",
-        len(sources), created, joined,
+        "%d image(s) joined to open groups, %d trickle drop(s) merged",
+        len(sources), created, joined, merged,
     )
     return {
         "enabled": True, "sources": len(sources),
-        "posts_created": created, "images_joined": joined,
+        "posts_created": created, "images_joined": joined, "drops_merged": merged,
     }

backend/app/services/discord_ingester.py

@@ -0,0 +1,85 @@
+"""Native Discord ingester — the Discord ADAPTER over `ingest_core.Ingester`.
+
+Thin counterpart to subscribestar_ingester (milestone 428). The walk's modes,
+both ledgers, cursor checkpointing and the post-first capture live in the core;
+this wires in the Discord client, downloader, ledger models and key.
+
+Two things differ from the cookie platforms:
+
+  - Discord authenticates with a user TOKEN, so `auth_token` is the credential
+    here rather than an argument accepted and ignored.
+  - The body canary is off. It fails a walk whose first 30+ captured posts all
+    came back without text, on the theory that a creator nearly always writes
+    something; a Discord drop is routinely files and nothing else, so on
+    Discord that is an ordinary backfill, not a broken parser.
+
+`campaign_id` is the source URL (a server, channel, thread or category link).
+FC runs on a plain-HTTP homelab; nothing here uses a secure-context Web API.
+"""
+
+from __future__ import annotations
+
+import asyncio
+from collections.abc import Callable
+from pathlib import Path
+
+from ..models import DiscordFailedMedia, DiscordSeenMedia
+from .discord_client import DiscordAPIError, DiscordClient, MediaItem
+from .discord_downloader import DiscordDownloader
+from .ingest_core import Ingester
+
+_LEDGER_KEY_MAX = 128
+
+
+def _ledger_key(media: MediaItem) -> str:
+    """`<message_id>:<media_id>` — stable across edits (see MediaItem)."""
+    return f"{media.post_id}:{media.media_id}"[:_LEDGER_KEY_MAX]
+
+
+class DiscordIngester(Ingester):
+    """Walk a Discord source's channels, download unseen files, return a
+    `DownloadResult`. `client` / `downloader` are injectable for tests."""
+
+    def __init__(
+        self,
+        images_root: Path,
+        cookies_path: str | None,
+        session_factory: Callable[[], object],
+        *,
+        validate: bool = True,
+        rate_limit: float = 0.0,
+        request_sleep: float = 0.0,
+        auth_token: str | None = None,
+        client: DiscordClient | None = None,
+        downloader: DiscordDownloader | None = None,
+    ):
+        del cookies_path  # Discord authenticates by token (uniform signature)
+        self.images_root = Path(images_root)
+        super().__init__(
+            client=client if client is not None else DiscordClient(
+                auth_token, request_sleep=request_sleep,
+            ),
+            downloader=downloader if downloader is not None else DiscordDownloader(
+                self.images_root, validate=validate, rate_limit=rate_limit,
+            ),
+            session_factory=session_factory,
+            seen_model=DiscordSeenMedia,
+            failed_model=DiscordFailedMedia,
+            seen_constraint="uq_discord_seen_media_source_id",
+            failed_constraint="uq_discord_failed_media_source_id",
+            ledger_key=_ledger_key,
+            platform="discord",
+            error_base=DiscordAPIError,
+            drift_label="Discord API",
+            body_canary=False,
+        )
+
+
+async def verify_discord_credential(url: str, auth_token: str | None) -> tuple[bool | None, str]:
+    """The uniform `(ok, message)` probe: is the token valid, and can its
+    account see the channel or server the source names?"""
+    if not auth_token:
+        return False, "No Discord token is saved — add one under Credentials."
+    client = DiscordClient(auth_token)
+    loop = asyncio.get_running_loop()
+    return await loop.run_in_executor(None, client.verify_auth, url)

backend/app/services/discord_repair.py

@@ -0,0 +1,183 @@
+"""Repair the Discord downloads made before the naming fix (issue #3999).
+
+Until dc840fe, gallery-dl's Discord patterns asked for keys the extractor never
+emits, so every Discord download landed as
+`<artist>/discord/None/<date>_None_<original name>`, next to a sidecar named
+after the attachment's ORIGINAL name. That broke two things:
+
+* **No Post, no date.** `find_sidecar` can never pair those names, so these
+  files were imported as loose images with no Post, and a card shows the
+  download time.
+* **No trustworthy metadata to relink from.** Every `image.png` in a channel
+  wrote the same `image.json`, so the surviving sidecar describes whichever
+  message was written last. The message id is gone from the filename too.
+
+The operator chose a clean re-download (2026-09-13) over relinking in place:
+delete the broken files and their records, make gallery-dl forget it fetched
+them, and backfill every Discord source again under the fixed naming.
+
+## Why the archive is cleared for ALL of Discord
+
+gallery-dl records a download as `discord{message_id}_{num}` (upstream
+`DiscordExtractor.archive_fmt`, prefixed with the category). The broken files
+lost their message ids, so there is no way to forget one source's entries and
+not another's. Every Discord download made before the fix is broken, so
+forgetting all of them is exactly right. Anything downloaded AFTER the fix still
+exists on disk under its correct name, and gallery-dl's `skip` sees the file and
+does not fetch it again.
+
+## What it does not touch
+
+Discord posts FC grouped itself (#388 E2) are built from Posts, and these files
+never had one, so there is nothing grouped to unwind. Images outside a
+`discord/None/` directory are never selected: the path filter requires both the
+`None` directory and the `_None_` filename, the pair only the bug produced.
+
+Operator-triggered only (Settings, preview first). Never on a beat.
+"""
+
+from __future__ import annotations
+
+import logging
+import sqlite3
+from pathlib import Path
+
+from sqlalchemy import func, select
+from sqlalchemy.orm import Session
+
+from ..models import ImageRecord, Source
+from .cleanup_service import delete_images
+from .gallery_dl import archive_path
+from .source_service import arm_backfill
+
+log = logging.getLogger(__name__)
+
+# `%` and `_` are LIKE wildcards, so the literal underscores around None are
+# escaped. `________` is the eight-digit date prefix the old pattern wrote.
+_BROKEN_PATH_LIKE = r"%/discord/None/________\_None\_%"
+
+# Upstream keys asset downloads as `asset_{server_id}_{id}`. FC never fetches
+# server assets, but excluding them keeps this to exactly the message
+# attachments the bug mangled.
+_ARCHIVE_SQL_MATCH = r"entry LIKE 'discord%' AND entry NOT LIKE 'discordasset\_%' ESCAPE '\'"
+_COUNT_SQL = "SELECT COUNT(*) FROM archive WHERE " + _ARCHIVE_SQL_MATCH
+_DELETE_SQL = "DELETE FROM archive WHERE " + _ARCHIVE_SQL_MATCH
+
+
+def broken_directories(images_root: Path) -> list[Path]:
+    """Every `<artist>/discord/None` directory. Artist folders that differ only by
+    case (`Conto` and `conto`) are separate directories and both are found."""
+    return sorted(d for d in Path(images_root).glob("*/discord/None") if d.is_dir())
+
+
+def count_archive_entries(archive: Path) -> int:
+    return _archive(archive, delete=False)
+
+
+def forget_archive_entries(archive: Path) -> int:
+    return _archive(archive, delete=True)
+
+
+def _archive(archive: Path, *, delete: bool) -> int:
+    if not archive.is_file():
+        return 0
+    # A download running at the same moment holds this file briefly. Waiting
+    # 30s for its lock beats failing the repair over a transient contention.
+    conn = sqlite3.connect(str(archive), timeout=30)
+    try:
+        has_table = conn.execute(
+            "SELECT 1 FROM sqlite_master WHERE type='table' AND name='archive'"
+        ).fetchone()
+        if not has_table:
+            return 0
+        if not delete:
+            return conn.execute(_COUNT_SQL).fetchone()[0]
+        cur = conn.execute(_DELETE_SQL)
+        conn.commit()
+        return cur.rowcount
+    finally:
+        conn.close()
+
+
+def _sweep_directory(directory: Path) -> tuple[int, bool]:
+    """Remove what the record deletes left behind: the collided sidecars, plus any
+    file that never became a record (a quarantined or rejected download). Then
+    the directory itself, if it is empty. Returns (files removed, dir removed)."""
+    removed = 0
+    for f in directory.iterdir():
+        if f.is_file():
+            try:
+                f.unlink()
+                removed += 1
+            except OSError as exc:
+                log.warning("discord repair: could not remove %s: %s", f, exc)
+    try:
+        directory.rmdir()
+        return removed, True
+    except OSError:
+        return removed, False
+
+
+def repair_discord_downloads(
+    session: Session, *, images_root: Path, dry_run: bool,
+) -> dict:
+    images_root = Path(images_root)
+    archive = archive_path(images_root)
+
+    broken = select(ImageRecord.id, ImageRecord.size_bytes).where(
+        ImageRecord.path.like(_BROKEN_PATH_LIKE, escape="\\")
+    )
+    rows = session.execute(broken).all()
+    image_ids = [r.id for r in rows]
+    directories = broken_directories(images_root)
+    sources = session.execute(
+        select(Source).where(Source.platform == "discord")
+    ).scalars().all()
+
+    summary = {
+        "images": len(image_ids),
+        "bytes": sum(r.size_bytes or 0 for r in rows),
+        "directories": len(directories),
+        "sources": len(sources),
+        "enabled_sources": sum(1 for s in sources if s.enabled),
+    }
+
+    if dry_run:
+        summary["archive_entries"] = count_archive_entries(archive)
+        return summary
+
+    deleted = delete_images(session, image_ids=image_ids, images_root=images_root)
+
+    swept = 0
+    directories_removed = 0
+    for d in directories:
+        n, gone = _sweep_directory(d)
+        swept += n
+        directories_removed += int(gone)
+
+    # Only after the files are gone. Forgetting first and failing half way would
+    # leave gallery-dl free to re-fetch into a directory still full of the old
+    # copies.
+    forgotten = forget_archive_entries(archive)
+
+    for source in sources:
+        arm_backfill(source)
+    session.commit()
+
+    remaining = session.execute(
+        select(func.count(ImageRecord.id)).where(
+            ImageRecord.path.like(_BROKEN_PATH_LIKE, escape="\\")
+        )
+    ).scalar_one()
+
+    summary.update(
+        images_deleted=deleted["images_deleted"],
+        files_failed=deleted["files_failed"],
+        leftover_files_removed=swept,
+        directories_removed=directories_removed,
+        archive_entries=forgotten,
+        backfills_started=len(sources),
+        remaining=remaining,
+    )
+    log.info("discord repair applied: %s", summary)
+    return summary

backend/app/services/download_backends.py

@@ -23,18 +23,22 @@ from __future__ import annotations
 import asyncio
 from pathlib import Path
 
+from .discord_ingester import DiscordIngester
 from .gallery_dl import DownloadResult, ErrorType
+from .ingest_core import DEFAULT_REVISIT_DAYS
 from .patreon_ingester import PatreonIngester
 from .patreon_resolver import extract_vanity, resolve_campaign_id_for_source
-from .pixiv_client import user_id_from_url
-from .pixiv_ingester import PixivIngester
 from .platforms import known_platform_keys
 from .subscribestar_ingester import SubscribeStarIngester
 
 # Platforms whose download + verify go through the native ingester rather than
-# gallery-dl. gallery-dl still serves the rest (hentaifoundry, discord) until
-# they migrate too. pixiv left this set when it was retired (milestone #406).
-NATIVE_INGESTER_PLATFORMS = frozenset({"patreon", "subscribestar"})
+# gallery-dl. gallery-dl still serves the rest (hentaifoundry) until it
+# migrates too. Discord joined in milestone 428.
+NATIVE_INGESTER_PLATFORMS = frozenset({"patreon", "subscribestar", "discord"})
+
+# Native platforms whose feed id IS the source URL, so there is nothing to
+# resolve: SubscribeStar's creator page, Discord's server/channel link.
+_URL_IS_FEED = frozenset({"subscribestar", "discord"})
 
 
 def _unsupported_platform_message(platform: str) -> str | None:
@@ -67,8 +71,8 @@ def _native_ingester_cls(platform: str):
     dispatch pick up the replacement."""
     return {
         "patreon": PatreonIngester,
-        "pixiv": PixivIngester,
         "subscribestar": SubscribeStarIngester,
+        "discord": DiscordIngester,
     }[platform]
 
 
@@ -86,6 +90,7 @@ async def run_download(
     mode: str | None,
     gdl,
     sync_session_factory,
+    revisit_days: int = DEFAULT_REVISIT_DAYS,
 ) -> tuple[DownloadResult, str | None]:
     """Uniform download across backends — the download counterpart to
     `verify_source_credential`, so this module is the ONE place that knows how
@@ -109,7 +114,7 @@ async def run_download(
         ), None
     if uses_native_ingester(platform):
         return await _run_native_ingester(
-            ctx, source_config, mode, gdl, sync_session_factory
+            ctx, source_config, mode, gdl, sync_session_factory, revisit_days
         )
     result = await gdl.download(
         url=ctx["url"],
@@ -127,25 +132,17 @@ async def _resolve_native_campaign_id(
     platform: str, url: str, cookies_path: str | None, overrides: dict,
 ) -> tuple[str | None, str | None]:
     """`(campaign_id, resolved_campaign_id)` for a native source. SubscribeStar's
-    feed id IS the creator URL; Pixiv's is the numeric user id parsed straight
-    from it (no lookup → resolved None either way). Patreon resolves the
-    campaign id from the vanity URL (resolved non-None when a lookup actually ran,
-    so phase 3 caches it)."""
-    if platform == "subscribestar":
+    and Discord's feed id IS the source URL (no lookup → resolved None). Patreon
+    resolves the campaign id from the vanity URL (resolved non-None when a lookup
+    actually ran, so phase 3 caches it)."""
+    if platform in _URL_IS_FEED:
         return url, None
-    if platform == "pixiv":
-        return user_id_from_url(url), None
     return await resolve_campaign_id_for_source(url, cookies_path, overrides)
 
 
 def _campaign_resolution_error(platform: str, url: str) -> str:
     """Operator-facing message for a native source whose campaign id could not
     be resolved — names the platform's own lookup mechanism."""
-    if platform == "pixiv":
-        return (
-            f"Could not extract a pixiv user id. source_url={url!r} — expected "
-            "a URL like https://www.pixiv.net/users/<id>."
-        )
     vanity = extract_vanity(url)
     return (
         f"Could not resolve Patreon campaign id. source_url={url!r}; "
@@ -157,6 +154,7 @@ def _campaign_resolution_error(platform: str, url: str) -> str:
 
 async def _run_native_ingester(
     ctx: dict, source_config, mode: str | None, gdl, sync_session_factory,
+    revisit_days: int = DEFAULT_REVISIT_DAYS,
 ) -> tuple[DownloadResult, str | None]:
     """Run the native ingester for a native platform in a worker thread (sync
     requests/subprocess). Patreon resolves a campaign id from the vanity URL;
@@ -172,8 +170,8 @@ async def _run_native_ingester(
         platform, ctx["url"], ctx["cookies_path"], overrides
     )
     if not campaign_id:
-        # Patreon: vanity lookup failed. Pixiv: no numeric user id in the URL.
-        # (SubscribeStar's campaign id is the URL itself — never lands here.)
+        # Patreon: vanity lookup failed. (SubscribeStar's campaign id is the
+        # URL itself — never lands here.)
         url = ctx["url"]
         return (
             DownloadResult(
@@ -205,7 +203,7 @@ async def _run_native_ingester(
         validate=gdl._validate_files,
         rate_limit=rate_limit,
         request_sleep=request_sleep,
-        # Uniform across adapters: token platforms (pixiv) authenticate with
+        # Uniform across adapters: a token platform would authenticate with
         # it, cookie platforms accept-and-ignore — so this construction stays
         # platform-agnostic.
         auth_token=ctx["auth_token"],
@@ -221,6 +219,9 @@ async def _run_native_ingester(
             mode=mode,
             resume_cursor=source_config.resume_cursor,
             time_budget_seconds=source_config.timeout,
+            # How far back a tick keeps looking for EDITED posts. The ingester
+            # applies it to ticks only; a backfill ignores it.
+            revisit_days=revisit_days,
             posts_base=int(overrides.get("_backfill_posts", 0)),
             # plan #709: live progress writes to this running event mid-walk.
             event_id=ctx.get("event_id"),
@@ -252,16 +253,15 @@ async def verify_source_credential(
     if uses_native_ingester(platform):
         # Native ingester platforms verify via their own lightweight auth probe.
         # SubscribeStar's probe takes the creator URL directly; Patreon's
-        # resolves the campaign id first; Pixiv's is one OAuth refresh (the
-        # exact call that fails when the token is bad — no feed walk).
+        # resolves the campaign id first.
         if platform == "subscribestar":
             from .subscribestar_ingester import verify_subscribestar_credential
 
             return await verify_subscribestar_credential(url, cookies_path, config_overrides)
-        if platform == "pixiv":
-            from .pixiv_ingester import verify_pixiv_credential
+        if platform == "discord":
+            from .discord_ingester import verify_discord_credential
 
-            return await verify_pixiv_credential(auth_token)
+            return await verify_discord_credential(url, auth_token)
         from .patreon_ingester import verify_patreon_credential
 
         return await verify_patreon_credential(url, cookies_path, config_overrides)

backend/app/services/download_service.py

@@ -39,6 +39,7 @@ from .gallery_dl import (
     walk_completed,
 )
 from .importer import Importer
+from .ingest_core import DEFAULT_REVISIT_DAYS
 from .platforms import auth_type_for
 from .scheduler_service import set_platform_cooldown
 
@@ -60,6 +61,7 @@ class DownloadService:
         importer: Importer,
         cred_service: CredentialService,
         sync_session_factory=None,
+        revisit_days: int = DEFAULT_REVISIT_DAYS,
     ):
         self.async_session = async_session
         self.sync_session = sync_session
@@ -71,6 +73,12 @@ class DownloadService:
         # the multi-minute walk — see PatreonIngester). Only the patreon branch
         # of phase 2 uses it; gallery-dl sources leave it None.
         self.sync_session_factory = sync_session_factory
+        # ImportSettings.download_revisit_days — how far back a tick keeps
+        # looking for EDITED posts (ingest_core.DEFAULT_REVISIT_DAYS). Passed in
+        # rather than read here because the task already loads the settings row
+        # for rate_limit/validate_files, and a second load on every download
+        # would be the same row twice for one number.
+        self.revisit_days = revisit_days
 
     async def download_source(self, source_id: int) -> int:
         """Returns DownloadEvent.id. Idempotent: in-flight events are returned as-is."""
@@ -178,6 +186,7 @@ class DownloadService:
         return await run_download(
             ctx=ctx, source_config=source_config, skip_value=skip_value, mode=mode,
             gdl=self.gdl, sync_session_factory=self.sync_session_factory,
+            revisit_days=self.revisit_days,
         )
 
     async def _phase1_setup(self, source_id: int) -> dict[str, Any]:
@@ -414,6 +423,13 @@ class DownloadService:
 
             await loop.run_in_executor(None, _upsert)
 
+        # Only now is it safe to call this walk's media seen: every file above
+        # has been through the importer. Had the run died before here they stay
+        # unmarked, and the next walk imports them from disk (ingest_core).
+        mark_seen = getattr(dl_result, "mark_seen_after_import", None)
+        if mark_seen is not None:
+            await loop.run_in_executor(None, mark_seen)
+
         # #830 recapture: backfill source_filehash on EXISTING on-disk images so
         # their post-body inline <img src=CDN> remaps to the local copy. A
         # SEPARATE non-deleting channel (NOT the import list — that would unlink

backend/app/services/extension_service.py

@@ -21,6 +21,9 @@ from .source_service import BACKFILL_MAX_CHUNKS
 
 log = logging.getLogger(__name__)
 
+# The probe runs while the chip is drawing; names that take longer are skipped.
+_NAME_LOOKUP_SECONDS = 6.0
+
 
 class UnknownPlatformError(Exception):
     """URL didn't match any platform pattern."""
@@ -30,6 +33,10 @@ class InvalidUrlError(Exception):
     """URL was empty or missing a scheme."""
 
 
+class UnknownArtistError(Exception):
+    """quick-add named an `artist_id` that does not exist."""
+
+
 # Mirrored byte-for-byte from extension/lib/platforms.js
 # PLATFORM_ARTIST_PATTERNS. Keep these two copies in sync by hand —
 # reviewers catch drift.
@@ -55,39 +62,127 @@ _PLATFORM_PATTERNS: list[tuple[str, re.Pattern[str]]] = [
         r"^https?://(?:www\.)?hentai-foundry\.com/user/(?P<slug>[^/?#]+)",
         re.IGNORECASE,
     )),
+    # A Discord URL names a server or a channel, never a creator, so the slug is
+    # `<server>` or `<server>/<channel>` and the artist is chosen, not derived.
+    # A trailing message id (a jump link) still names its channel. DMs (`@me`)
+    # are not sources; thread links (`/threads/`) are left to the manual form.
+    ("discord", re.compile(
+        r"^https?://(?:www\.|ptb\.|canary\.)?discord\.com/channels/"
+        r"(?P<slug>\d+(?:/\d+)?)(?:/\d+)?/?(?:[?#].*)?$",
+        re.IGNORECASE,
+    )),
 ]
 
+DISCORD = "discord"
+
+
+def canonical_source_url(platform: str, url: str, slug: str) -> str:
+    """The URL a new source is stored under. Discord's is rebuilt from the ids
+    — the form the manual Add form and the ingester use — so a jump link, a
+    ptb/canary host or a trailing slash never makes a second source for the
+    same channel. Every other platform keeps the URL as given."""
+    if platform == DISCORD:
+        return f"https://discord.com/channels/{slug}"
+    return url
+
+
+def _discord_ids(url: str) -> tuple[str | None, str | None] | None:
+    """`(server_id, channel_id)` of a stored Discord source URL, None if it
+    does not parse (a DM or thread link, or an old malformed row)."""
+    from .discord_client import DiscordAPIError, parse_source_url
+    try:
+        return parse_source_url(url)
+    except DiscordAPIError:
+        return None
+
 
 class ExtensionService:
     def __init__(self, session: AsyncSession, crypto=None) -> None:
         self.session = session
-        # Optional decryptor for resolving a token-auth platform's display name
-        # (pixiv) at add-time. None → skip resolution, fall back to the handle.
+        # Optional decryptor for resolving a platform's display name at
+        # add-time. None → skip resolution, fall back to the handle.
         self._crypto = crypto
 
-    async def quick_add_source(self, url: str) -> dict:
+    async def quick_add_source(
+        self,
+        url: str,
+        *,
+        artist_id: int | None = None,
+        artist_name: str | None = None,
+        use_platform_name: bool = False,
+    ) -> dict:
+        """Add `url` as a source. `artist_id` connects it to an existing
+        artist, `artist_name` to that artist (created if new); with neither,
+        the artist is resolved from the platform as before.
+
+        `use_platform_name` applies the operator's convention that the Patreon
+        name is canon: a Patreon source added to an existing artist renames
+        that artist to the creator's Patreon display name. Name only — the
+        slug, and every path keyed off it, never moves (#130). Ignored on every
+        other platform, and when the name can't be read."""
         platform, raw_slug = self._derive(url)
+        url = canonical_source_url(platform, url, raw_slug)
+        renamed_from = None
         # Identity by SOURCE handle (#130): an existing (platform, url) source
         # keeps its artist on re-add — even if that artist was since renamed (its
-        # frozen slug no longer matches the current name). Only a genuinely new
-        # source resolves/creates an artist.
-        existing = (await self.session.execute(
-            select(Source).where(Source.platform == platform, Source.url == url)
-        )).scalar_one_or_none()
+        # frozen slug no longer matches the current name), and even when the
+        # add named a different artist. Only a genuinely new source
+        # resolves/creates an artist.
+        existing = await self._existing_source(platform, url)
         if existing is not None:
             artist = (await self.session.execute(
                 select(Artist).where(Artist.id == existing.artist_id)
             )).scalar_one()
             return self._shape(existing, artist, created_source=False, created_artist=False)
 
-        # New source → name the artist properly by resolving the real display
-        # name from the platform (falls back to the URL handle).
-        name = await self._resolve_artist_name(platform, raw_slug, url)
-        artist, created_artist = await self._find_or_create_artist(name)
+        if artist_id is not None:
+            artist = (await self.session.execute(
+                select(Artist).where(Artist.id == artist_id)
+            )).scalar_one_or_none()
+            if artist is None:
+                raise UnknownArtistError(f"no artist with id {artist_id}")
+            created_artist = False
+            if use_platform_name and platform == "patreon":
+                renamed_from = await self._adopt_patreon_name(artist, raw_slug, url)
+        else:
+            name = (artist_name or "").strip()
+            if not name:
+                # Name the artist properly by resolving the real display name
+                # from the platform (falls back to the URL handle).
+                name = await self._resolve_artist_name(platform, raw_slug, url)
+            artist, created_artist = await self._find_or_create_artist(name)
         source, created_source = await self._find_or_create_source(
             artist_id=artist.id, platform=platform, url=url,
         )
-        return self._shape(source, artist, created_source, created_artist)
+        shaped = self._shape(source, artist, created_source, created_artist)
+        if renamed_from is not None:
+            shaped["renamed_from"] = renamed_from
+        return shaped
+
+    async def _adopt_patreon_name(self, artist, raw_slug: str, url: str) -> str | None:
+        """Rename `artist` to the Patreon display name; the old name when it
+        changed, else None. Unreadable name → no rename, never the handle."""
+        name = await self._platform_display_name("patreon", raw_slug, url)
+        if not name or name == artist.name:
+            return None
+        old = artist.name
+        artist.name = name
+        await self.session.commit()
+        return old
+
+    async def _existing_source(self, platform: str, url: str) -> Source | None:
+        """The source this URL already is, whichever artist owns it. Discord
+        compares ids, not strings, so a row stored before canonicalisation (a
+        ptb host, a trailing slash) is still found."""
+        if platform != DISCORD:
+            return (await self.session.execute(
+                select(Source).where(Source.platform == platform, Source.url == url)
+            )).scalars().first()
+        want = _discord_ids(url)
+        rows = (await self.session.execute(
+            select(Source).where(Source.platform == DISCORD).order_by(Source.id)
+        )).scalars().all()
+        return next((s for s in rows if _discord_ids(s.url) == want), None)
 
     @staticmethod
     def _shape(source, artist, created_source: bool, created_artist: bool) -> dict:
@@ -112,31 +207,38 @@ class ExtensionService:
         self, platform: str, raw_slug: str, url: str
     ) -> str:
         """The real display name for a new artist, resolved from the platform at
-        add-time (#130). Our native platforms each have a name source — pixiv the
-        app API (token), patreon the campaigns API, subscribestar the profile
-        page (both cookies). Other platforms (and any failure — no credential,
-        network error) fall back to the URL handle, which is already readable.
+        add-time (#130). Our native platforms each have a name source — patreon
+        the campaigns API, subscribestar the profile page (both cookies). Other
+        platforms (and any failure — no credential, network error) fall back to
+        the URL handle, which is already readable.
         The resolvers are sync, so they run in an executor."""
-        if self._crypto is None or platform not in ("pixiv", "patreon", "subscribestar"):
-            return raw_slug
+        if platform == DISCORD:
+            # The server's name: what the operator knows the community as.
+            server_id = raw_slug.split("/", 1)[0]
+            names = await self._discord_names(server_id, None)
+            return names.get("server") or f"Discord {server_id}"
+        return await self._platform_display_name(platform, raw_slug, url) or raw_slug
+
+    async def _platform_display_name(
+        self, platform: str, raw_slug: str, url: str
+    ) -> str | None:
+        """The creator's display name as Patreon or SubscribeStar shows it, read
+        with the stored cookies; None when it can't be read (no credential, a
+        network error, a slow answer, any other platform). None, not the handle,
+        so a caller can tell a real name from a fallback — a rename to the
+        Patreon name must never rename to a URL handle instead."""
+        if self._crypto is None or platform not in ("patreon", "subscribestar"):
+            return None
         import asyncio
 
         from .credential_service import CredentialService
         cred = CredentialService(self.session, self._crypto)
         loop = asyncio.get_running_loop()
         try:
-            if platform == "pixiv":
-                token = await cred.get_token("pixiv")
-                if not token:
-                    return raw_slug
-                from .pixiv_client import PixivClient
-                name = await loop.run_in_executor(
-                    None, PixivClient(token).resolve_display_name, raw_slug
-                )
-            elif platform == "patreon":
+            if platform == "patreon":
                 cookies = await cred.get_cookies_path("patreon")
                 from .patreon_resolver import resolve_display_name
-                name = await loop.run_in_executor(
+                call = loop.run_in_executor(
                     None, resolve_display_name, raw_slug,
                     str(cookies) if cookies else None,
                 )
@@ -144,15 +246,14 @@ class ExtensionService:
                 cookies = await cred.get_cookies_path("subscribestar")
                 from .subscribestar_client import SubscribeStarClient
                 client = SubscribeStarClient(str(cookies) if cookies else None)
-                name = await loop.run_in_executor(
-                    None, client.resolve_display_name, url
-                )
+                call = loop.run_in_executor(None, client.resolve_display_name, url)
+            name = await asyncio.wait_for(call, timeout=_NAME_LOOKUP_SECONDS)
         except Exception as exc:      # resolution is best-effort — never block the add
             log.warning("artist display-name resolution failed (%s): %s", platform, exc)
-            return raw_slug
-        return name or raw_slug
+            return None
+        return (name or "").strip() or None
 
-    async def probe(self, url: str) -> dict:
+    async def probe(self, url: str, *, names: bool = False) -> dict:
         """Read-only resolution of a creator-page URL against the FC DB.
         Returns one of:
           - {state: 'unknown_platform'}            — URL didn't match any
@@ -168,21 +269,30 @@ class ExtensionService:
                                                    — exact (artist, platform,
             url) Source already exists
 
-        Side-effect-free: two SELECTs at most.
+        `names` (the Add panel asks, the chip does not) adds `display_name`:
+        the creator's name as Patreon/SubscribeStar shows it, or None. It costs
+        a request to the platform, so a plain page view never pays it.
+
+        Side-effect-free: two SELECTs at most, plus that one lookup.
         """
         try:
             platform, raw_slug = self._derive(url)
         except (UnknownPlatformError, InvalidUrlError):
             return {"state": "unknown_platform"}
+        if platform == DISCORD:
+            return await self._probe_discord(raw_slug)
 
         slug = slugify(raw_slug)
+        result: dict = {"platform": platform, "slug": slug}
+        if names:
+            result["display_name"] = await self._platform_display_name(
+                platform, raw_slug, url,
+            )
         artist = (await self.session.execute(
             select(Artist).where(Artist.slug == slug)
         )).scalar_one_or_none()
         if artist is None:
-            return {"state": "new", "platform": platform, "slug": slug}
-
-        artist_payload = {"id": artist.id, "name": artist.name, "slug": artist.slug}
+            return {"state": "new", **result}
 
         source = (await self.session.execute(
             select(Source).where(
@@ -192,26 +302,114 @@ class ExtensionService:
             )
         )).scalar_one_or_none()
         if source is None:
-            return {
-                "state": "artist_match",
-                "platform": platform,
-                "slug": slug,
-                "artist": artist_payload,
-            }
+            return {"state": "artist_match", **result, "artist": self._artist_payload(artist)}
 
         return {
             "state": "source_match",
-            "platform": platform,
-            "slug": slug,
-            "artist": artist_payload,
-            "source": {
-                "id": source.id,
-                "artist_id": source.artist_id,
-                "platform": source.platform,
-                "url": source.url,
-                "enabled": source.enabled,
+            **result,
+            "artist": self._artist_payload(artist),
+            "source": self._source_payload(source),
+        }
+
+    async def _probe_discord(self, raw_slug: str) -> dict:
+        """probe for a Discord server or channel. The states mean what they
+        mean elsewhere, but the artist is never read off the URL:
+
+          - source_match: this channel is a source — or the whole server is
+            (`covered_by_server`), which already walks every channel;
+          - artist_match: another source on this server belongs to an artist,
+            the one this channel most likely belongs to too (a suggestion the
+            Add panel preselects, not a decision);
+          - new: nothing on this server yet.
+
+        `discord` carries the ids, both canonical URLs and the display names,
+        read with the stored token; a name that can't be read is None."""
+        server_id, _, channel_id = raw_slug.partition("/")
+        channel_id = channel_id or None
+        rows = (await self.session.execute(
+            select(Source, Artist)
+            .join(Artist, Artist.id == Source.artist_id)
+            .where(Source.platform == DISCORD)
+            .order_by(Source.id)
+        )).all()
+        exact = server_whole = on_server = None
+        for source, artist in rows:
+            ids = _discord_ids(source.url)
+            if ids is None or ids[0] != server_id:
+                continue
+            if ids[1] == channel_id and exact is None:
+                exact = (source, artist)
+            elif ids[1] is None and server_whole is None:
+                server_whole = (source, artist)
+            if on_server is None:
+                on_server = (source, artist)
+
+        names = await self._discord_names(server_id, channel_id)
+        base = f"https://discord.com/channels/{server_id}"
+        result: dict = {
+            "platform": DISCORD,
+            "slug": raw_slug,
+            "discord": {
+                "server_id": server_id,
+                "channel_id": channel_id,
+                "server_name": names.get("server"),
+                "channel_name": names.get("channel"),
+                "server_url": base,
+                "channel_url": f"{base}/{channel_id}" if channel_id else None,
             },
         }
+        hit = exact or server_whole
+        if hit is not None:
+            source, artist = hit
+            result.update(
+                state="source_match",
+                artist=self._artist_payload(artist),
+                source=self._source_payload(source),
+                covered_by_server=exact is None,
+            )
+        elif on_server is not None:
+            result.update(state="artist_match", artist=self._artist_payload(on_server[1]))
+        else:
+            result["state"] = "new"
+        return result
+
+    async def _discord_names(self, server_id: str | None, channel_id: str | None) -> dict:
+        """Server/channel display names via the stored Discord token. Never
+        raises and never waits out a rate limit: it runs while the operator
+        looks at a page, so a slow or missing answer just means no names."""
+        if self._crypto is None:
+            return {}
+        import asyncio
+
+        from .credential_service import CredentialService
+        from .discord_client import DiscordClient
+        try:
+            token = await CredentialService(self.session, self._crypto).get_token(DISCORD)
+            if not token:
+                return {}
+            client = DiscordClient(token, max_retries=0)
+            loop = asyncio.get_running_loop()
+            return await asyncio.wait_for(
+                loop.run_in_executor(None, client.describe, server_id, channel_id),
+                timeout=_NAME_LOOKUP_SECONDS,
+            )
+        except Exception as exc:      # names are decoration — never fail the call
+            log.info("Discord name lookup failed: %s", exc)
+            return {}
+
+    @staticmethod
+    def _artist_payload(artist) -> dict:
+        return {"id": artist.id, "name": artist.name, "slug": artist.slug}
+
+    @staticmethod
+    def _source_payload(source) -> dict:
+        return {
+            "id": source.id,
+            "artist_id": source.artist_id,
+            "platform": source.platform,
+            "url": source.url,
+            "enabled": source.enabled,
+        }
 
     def _derive(self, url: str) -> tuple[str, str]:
         if not isinstance(url, str) or not url.strip():

backend/app/services/gallery_dl.py

@@ -17,6 +17,7 @@ import subprocess
 import sys
 import tempfile
 import time
+from collections.abc import Callable
 from dataclasses import dataclass, field
 from datetime import UTC, datetime
 from enum import StrEnum
@@ -94,6 +95,16 @@ BACKFILL_CHUNK_SECONDS = 600
 _DEFAULT_GDL_TIMEOUT_SECONDS = 870
 
 
+def archive_path(images_root: Path) -> Path:
+    """gallery-dl's download archive: the record of what it has already fetched.
+
+    One definition, because the Discord repair (services/discord_repair.py) has
+    to find the same file the downloader writes, without constructing a service
+    whose __init__ creates directories.
+    """
+    return Path(images_root) / ".gallery-dl" / "archive.sqlite3"
+
+
 @dataclass
 class SourceConfig:
     """Per-source overrides loaded from Source.config_overrides JSON.
@@ -178,6 +189,13 @@ class DownloadResult:
     # the platform cooldown matches the hint instead of a flat default. None when
     # unknown (no header, or not a rate-limit failure).
     retry_after_seconds: float | None = None
+    # Native ingester only: marks this walk's fetched media seen in its ledger.
+    # Phase 3 calls it AFTER the import loop, never before — a file marked seen
+    # but not yet imported is invisible to every later walk, so a run killed in
+    # between orphaned it for good (TamadaHeijun's 12PCG post lost 7 of 13
+    # images to a stranded run, 2026-09-24). Unmarked, the next walk finds the
+    # file on disk with no ImageRecord and imports it. None on gallery-dl.
+    mark_seen_after_import: Callable[[], None] | None = None
 
 
 def extract_errors_warnings(stderr: str) -> str:
@@ -365,24 +383,16 @@ class GalleryDLService:
     # (services/patreon_ingester.py), not gallery-dl.
     PLATFORM_DEFAULTS = {
         # subscribestar removed — native-ingester platform now (#71); pixiv
-        # removed likewise (#129); deviantart removed at #3069 as a dropped
-        # platform, not a migrated one. The remaining entries are the
-        # gallery-dl platforms not yet migrated.
+        # removed likewise (#129); discord likewise (milestone 428, whose
+        # downloader keeps this config's on-disk naming); deviantart removed at
+        # #3069 as a dropped platform, not a migrated one. HentaiFoundry is the
+        # one platform left here, by the operator's choice not to migrate it.
         "hentaifoundry": {
             "content_types": ["all"],
             "directory": [],
             "filename": "{category}_{index:>03}_{title[:50]}.{extension}",
             "include": "all",
         },
-        "discord": {
-            "content_types": ["all"],
-            "directory": ["{channel[name]}"],
-            "filename": "{date:%Y%m%d}_{id}_{filename}.{extension}",
-            "embeds": "all",
-            "stickers": True,
-            "reactions": False,
-            "threads": True,
-        },
     }
 
     def __init__(
@@ -402,7 +412,7 @@ class GalleryDLService:
         config = {
             "extractor": {
                 "base-directory": str(self.images_root),
-                "archive": str(self._config_dir / "archive.sqlite3"),
+                "archive": str(archive_path(self.images_root)),
                 "skip": True,
                 "sleep": self._rate_limit,
                 "sleep-request": max(0.5, self._rate_limit / 4),
@@ -744,9 +754,6 @@ class GalleryDLService:
 
         if cookies_path:
             config["extractor"]["cookies"] = cookies_path
-        if auth_token and platform == "discord":
-            config["extractor"].setdefault("discord", {})
-            config["extractor"]["discord"]["token"] = auth_token
 
         with tempfile.NamedTemporaryFile(
             mode="w", suffix=".json", delete=False, dir=str(self._config_dir),
@@ -930,8 +937,6 @@ class GalleryDLService:
         config = self._build_config_for_source(platform, source_config, artist_slug)
         if cookies_path:
             config["extractor"]["cookies"] = cookies_path
-        if auth_token and platform == "discord":
-            config["extractor"].setdefault("discord", {})["token"] = auth_token
 
         with tempfile.NamedTemporaryFile(
             mode="w", suffix=".json", delete=False, dir=str(self._config_dir),

backend/app/services/gallery_service.py

@@ -45,7 +45,7 @@ from .tag_query import (
 # provenance (filesystem imports). Returned by facets() as a null-valued
 # bucket; the frontend maps that null back to this sentinel in the URL so the
 # bucket is selectable. Underscore-wrapped so it can't collide with a real
-# gallery-dl platform name (patreon/pixiv/...).
+# gallery-dl platform name (patreon/hentaifoundry/...).
 UNSOURCED_PLATFORM = "__unsourced__"
 
 
@@ -322,7 +322,7 @@ def _gallery_images(rows, artists: dict[int, dict]) -> list[GalleryImage]:
     ]
 
 
-def _diversify_similar(src, rows, limit, *, dup_threshold=8, lam=0.40):
+def _diversify_similar(src, rows, limit, *, dup_threshold=32, lam=0.40):
     """Trim a nearest-cosine candidate pool down to `limit` diverse picks.
 
     1. pHash collapse: drop any candidate whose perceptual hash is within
@@ -338,6 +338,11 @@ def _diversify_similar(src, rows, limit, *, dup_threshold=8, lam=0.40):
     2026-07-01 — dropped 0.55→0.40, dup 6→8, paired with a wider pool in
     `similar()`).
 
+    `dup_threshold` counts Hamming bits, so it moved 8→32 when the pHash went
+    from 64 to 256 bits (#4223, migration 0098) — the same fraction of the
+    hash, i.e. the tuning the operator chose, unchanged. This collapse is
+    DISPLAY-only: it hides a near-dup from one rail, it never drops a record.
+
     Falls back to nearest-order (`rows[:limit]`) on any failure or a small pool.
     """
     if len(rows) <= 1:

backend/app/services/importer.py

@@ -33,13 +33,14 @@ from ..models import (
 )
 from ..utils import safe_probe
 from ..utils.paths import (
+    canonical_subdir,
     derive_subdir,
     derive_top_level_artist,
     filehash_from_url,
     hash_suffixed_name,
     safe_ext,
 )
-from ..utils.phash import compute_phash, find_similar
+from ..utils.phash import compute_phash, find_similar, fingerprint_path, fingerprints_match
 from ..utils.sidecar import find_sidecar, parse_sidecar
 from ..utils.slug import slugify
 from .archive_extractor import extract_archive, is_archive
@@ -234,6 +235,38 @@ class Importer:
                 (phash, width or 0, height or 0, image_id)
             )
 
+    def _pixel_confirmer(self, source: Path):
+        """Build `find_similar`'s gate-3 callback for an incoming file.
+
+        pHash proposes; this accepts. A candidate is a duplicate only if its
+        file really is the same picture as `source` at a different size —
+        which is the only merge the operator asked for (#4223). Everything
+        else (a missing file, an unreadable one, a deleted row) returns
+        False: the destructive outcomes here are dropping a download and
+        overwriting a kept file, so an unanswerable question must not read
+        as "yes".
+
+        Both sides' fingerprints are computed lazily and cached, so an
+        import that matches nothing costs no I/O at all and an archive
+        member that keeps hitting the same candidate pays for it once.
+        """
+        new_fp: list = []
+        cand_fps: dict[int, object] = {}
+
+        def confirm(candidate_id: int) -> bool:
+            if not new_fp:
+                new_fp.append(fingerprint_path(source))
+            if new_fp[0] is None:
+                return False
+            if candidate_id not in cand_fps:
+                rec = self.session.get(ImageRecord, candidate_id)
+                cand_fps[candidate_id] = (
+                    fingerprint_path(Path(rec.path)) if rec and rec.path else None
+                )
+            return fingerprints_match(new_fp[0], cand_fps[candidate_id])
+
+        return confirm
+
     def _get_or_create(self, stmt, factory):
         """Race-safe find-or-create. Run `stmt` (scalar_one_or_none); if a
         row exists, return it. Otherwise open a savepoint and INSERT
@@ -862,6 +895,7 @@ class Importer:
                 rel, match_id = find_similar(
                     phash, width or 0, height or 0,
                     candidates, self.settings.phash_threshold,
+                    confirm=self._pixel_confirmer(source),
                 )
                 if rel == "larger_exists":
                     # Enrich-on-duplicate (parity with attach_in_place).
@@ -911,7 +945,7 @@ class Importer:
                 )
                 return ImportResult(status="superseded", image_id=match_id)
 
-        dest = self._copy_to_library(source, sha, attribution_path)
+        dest = self._copy_to_library(source, sha, attribution_path, path_artist)
 
         record = ImageRecord(
             path=str(dest),
@@ -1241,6 +1275,7 @@ class Importer:
                 rel, match_id = find_similar(
                     phash, width or 0, height or 0,
                     candidates, self.settings.phash_threshold,
+                    confirm=self._pixel_confirmer(path),
                 )
                 if rel == "larger_exists":
                     # Enrich-on-duplicate: link the near-dup's post to the
@@ -1566,7 +1601,8 @@ class Importer:
         )
 
     def _copy_to_library(
-        self, source: Path, sha: str, attribution_path: Path
+        self, source: Path, sha: str, attribution_path: Path,
+        artist: Artist | None = None,
     ) -> Path:
         """Copy `source` to its final library path. Returns the destination.
 
@@ -1574,8 +1610,18 @@ class Importer:
         _import_media (filesystem scan) and _supersede (when new_path is
         not passed). FC-3c's attach_in_place skips this helper entirely
         — the file is already at its final home.
+
+        `artist`, when resolved, decides the top-level directory: the
+        library is keyed on the Artist row's slug, NOT on however the
+        import folder happened to be capitalised. Without that, an import
+        from `/import/Conto/` and a download for the same artist write to
+        `Conto/` and `conto/` respectively and the library grows a second
+        home for one artist (milestone #421).
         """
-        subdir = derive_subdir(attribution_path, self.import_root)
+        subdir = canonical_subdir(
+            derive_subdir(attribution_path, self.import_root),
+            artist.slug if artist else None,
+        )
         dest_dir = self.images_root / subdir if subdir else self.images_root
         dest_dir.mkdir(parents=True, exist_ok=True)
         dest_name = hash_suffixed_name(source.stem, sha, source.suffix)
@@ -1610,7 +1656,16 @@ class Importer:
         that path (FC-3c attach_in_place case) — skip the copy step.
         Otherwise the file is copied via _copy_to_library."""
         if new_path is None:
-            dest = self._copy_to_library(source, sha, source)
+            # The KEPT row's artist decides the destination, not the incoming
+            # file's folder — a supersede rewrites `existing.path`, so writing
+            # it anywhere but that artist's canonical directory would move a
+            # row OUT of the tree milestone #421 is consolidating. ImageRecord
+            # carries `artist_id` with no relationship attribute, so this is a
+            # lookup rather than `existing.artist`.
+            kept_artist = artist
+            if kept_artist is None and existing.artist_id is not None:
+                kept_artist = self.session.get(Artist, existing.artist_id)
+            dest = self._copy_to_library(source, sha, source, kept_artist)
         else:
             dest = new_path
 

backend/app/services/ingest_core.py

@@ -31,10 +31,12 @@ import json
 import logging
 import time
 from collections.abc import Callable
+from datetime import UTC, datetime, timedelta
 
 from sqlalchemy import delete, func, select, text
 from sqlalchemy.dialects.postgresql import insert as pg_insert
 
+from ..models import ImageRecord
 from .gallery_dl import (
     DownloadResult,
     ErrorType,
@@ -51,6 +53,33 @@ log = logging.getLogger(__name__)
 # per-file HEADs. Headroom against paywalled/undownloadable items interleaving.
 _TICK_SEEN_THRESHOLD = 20
 
+# How far back a tick keeps looking even once everything is already-have-it —
+# the REVISIT WINDOW. Operator, 2026-09-23, holding up a Floppystack post:
+# *"this post has been updated as he implements hot fixes — any chance we have a
+# way to scan for or see updated posts so we can update ours to match and pull
+# the new attachments and pictures etc."*
+#
+# A creator who edits a three-day-old post to append a hotfix build was
+# structurally unreachable: that post sits twenty-odd already-seen items down
+# the feed, so the count early-out above fired before the walk ever got to it.
+# Not a bug in the early-out — a COUNT cannot express "recent".
+#
+# So the early-out now needs BOTH conditions: the run of already-seen items AND
+# a post published before the horizon. Strictly a widening. Two properties this
+# shape has and a plain "walk the last N days" would not:
+#
+#   * window 0 is exactly the old behaviour, so the feature has an off switch
+#     that costs nothing to reason about;
+#   * no window can make a tick stop EARLIER than it used to. A source paused
+#     for months has an unseen backlog stretching well past any horizon, and
+#     the walk still runs to the end of it — the horizon is a FLOOR on how far
+#     to look, never a ceiling.
+#
+# The live value is `ImportSettings.download_revisit_days` (rule 25 — an
+# operator tuning how far back their creators edit should not need a redeploy).
+# This is the fallback for a caller that passes none.
+DEFAULT_REVISIT_DAYS = 30
+
 # plan #705 #7: after this many failed download/validate attempts a media is
 # "dead-lettered" and skipped on routine tick/backfill walks (recovery still
 # re-attempts it). Stops a permanently-broken media re-erroring forever.
@@ -77,6 +106,44 @@ _LIVE_PROGRESS_INTERVAL = 5.0
 # recapture (the operator's schema-test flow) reaches the sample.
 _CANARY_MIN_SAMPLE = 30
 
+# The walk's time budget covers only the walk, but phase 3 runs in the SAME
+# Celery task, under the same soft limit (tasks/download.py: 1350s). A walk that
+# finds a lot of work for phase 3 must stop early and leave it to the next chunk,
+# or phase 3 is killed mid-import: TamadaHeijun's recapture, 2026-09-24, walked
+# for ~2 min and then spent 20 min importing 431 orphans and relinking ~3000
+# on-disk files, and died at the soft limit.
+#
+# So the walk also stops when its elapsed time PLUS phase 3's estimated cost
+# would pass CHUNK_TOTAL_SECONDS. Costs measured on the live instance: 431
+# imports took 976s (~2.3s each: hash, pHash, sidecar, provenance); a relink is
+# a sha256 over NFS, 0.15s for the 8.8 MB average file, plus a lookup.
+# test_download_source_task pins CHUNK_TOTAL_SECONDS under the soft limit.
+CHUNK_TOTAL_SECONDS = 1200.0
+PHASE3_IMPORT_SECONDS = 2.5
+PHASE3_RELINK_SECONDS = 0.25
+
+
+def _parse_published(raw: object) -> datetime | None:
+    """An ISO-8601 post date from either native client, as aware UTC.
+
+    Patreon's `published_at` is tz-aware with a `Z` or `+00:00` offset;
+    SubscribeStar's is NAIVE (`_parse_ss_datetime` renders a parsed local
+    timestamp with no zone). A naive value is read as UTC — the alternative is
+    discarding it, and a post whose date we refuse to read is a post the revisit
+    window can never reach.
+
+    Anything unparseable returns None, which reads downstream as "not provably
+    recent" and leaves the walk on its pre-revisit behaviour. Never raises: a
+    date we cannot read must not fail a walk that is otherwise working.
+    """
+    if not isinstance(raw, str) or not raw.strip():
+        return None
+    try:
+        parsed = datetime.fromisoformat(raw.strip().replace("Z", "+00:00"))
+    except ValueError:
+        return None
+    return parsed if parsed.tzinfo else parsed.replace(tzinfo=UTC)
+
 
 class Ingester:
     """Generic native-ingest orchestration. Subclass with a platform adapter
@@ -113,9 +180,9 @@ class Ingester:
         # (e.g. "Patreon API", "SubscribeStar markup").
         self._drift_label = drift_label or platform
         # #862 canary opt-out: platforms whose posts legitimately have empty
-        # bodies across large samples (pixiv — caption-less artists are common)
-        # would false-positive the zero-bodies-means-drift alarm; their clients
-        # catch drift structurally (response-shape checks) instead. The
+        # bodies across large samples would false-positive the
+        # zero-bodies-means-drift alarm; their clients catch drift structurally
+        # (response-shape checks) instead. The
         # "bodies X/N" summary line still surfaces the ratio either way.
         self._body_canary = body_canary
 
@@ -132,11 +199,17 @@ class Ingester:
         resume_cursor: str | None = None,
         time_budget_seconds: float = 870.0,
         seen_threshold: int = _TICK_SEEN_THRESHOLD,
+        revisit_days: int = DEFAULT_REVISIT_DAYS,
         posts_base: int = 0,
         event_id: int | None = None,
     ) -> DownloadResult:
         """Walk + download for one source, returning a gallery-dl-shaped result.
 
+        `revisit_days` is the tick's revisit window (see DEFAULT_REVISIT_DAYS):
+        inside it a tick neither early-outs nor trusts the post-record gate, so
+        a post edited after we first captured it is re-read and its new
+        attachments downloaded. 0 turns the window off.
+
         `mode` is "tick" | "backfill" | "recovery" | "recapture". Recovery
         bypasses the tier-1 seen-ledger AND the dead-letter ledger (tier-2 disk
         still skips kept files). Recapture (#830) is the cheap "re-grab post
@@ -179,6 +252,29 @@ class Ingester:
         # no media download, no post-record stub. Absent on stub/not-yet-migrated
         # clients → nothing is ever treated as gated.
         post_is_gated = getattr(self.client, "post_is_gated", None)
+        # The revisit window (see DEFAULT_REVISIT_DAYS). `post_meta` is an
+        # existing client seam — both native clients already implement it, for
+        # a preview sample whose caller has since gone, so this needed no new
+        # contract, only a live consumer for one. Absent seam, an unreadable
+        # date or a window of 0 → `horizon` never matches and the walk behaves
+        # exactly as it did before 2026-09-23.
+        #
+        # The window applies to TICKS only. A backfill is gated on purpose
+        # (capture each post once) and `recapture` mode already exists for the
+        # operator-driven "re-read every body" pass; a horizon there would be a
+        # third overlapping answer to a question that has two.
+        post_meta = getattr(self.client, "post_meta", None)
+        # #4413: optional client seam for a source that is several feeds walked
+        # one after another (a Discord server: every channel and thread). The
+        # tick early-out means "this feed has nothing new", and without the
+        # seam it ends the WHOLE walk — so the first quiet channel would hide
+        # every channel after it. With it, the early-out asks the client to
+        # move on and the walk continues. Absent → the early-out ends the walk,
+        # exactly as before (Patreon and SubscribeStar are one feed each).
+        skip_feed = getattr(self.client, "skip_feed", None)
+        horizon: datetime | None = None
+        if mode == "tick" and revisit_days > 0 and post_meta is not None:
+            horizon = datetime.now(UTC) - timedelta(days=revisit_days)
         start = time.monotonic()
         last_live = start  # plan #709: last live-progress write timestamp
         log_lines: list[str] = []
@@ -190,6 +286,9 @@ class Ingester:
         # source_filehash and (b) link the on-disk image to its Post (#1288) —
         # WITHOUT re-downloading or unlinking the file. Empty outside recapture.
         relink: list[tuple[str, str, str]] = []
+        # Media handed to phase 3 for import. Marked seen by phase 3 once the
+        # import has run (`mark_seen_after_import`), not here — see there.
+        fetched: list[tuple[str, str]] = []
         downloaded = 0
         errors = 0
         quarantined = 0
@@ -210,11 +309,18 @@ class Ingester:
         # absolute across chunks instead of an inflating sum. posts_processed
         # stays the gross per-chunk count used for the run summary.
         chunk_new_posts = 0
+        # Posts inside the revisit window that we had already captured, and the
+        # media those revisits turned up. Reported in the run summary — the
+        # operator's ask was to SEE the updated posts, not only to end up with
+        # their files ("so we can update ours to match").
+        revisited = 0
+        revisit_downloads = 0
         consecutive_seen = 0
         emitted_cursor: str | None = None
         reached_bottom = False
         budget_hit = False
         early_out = False
+        feeds_caught_up = 0  # #4413: feeds a tick left early via skip_feed
         stopped = False  # plan #708 B4: operator hit Stop mid-walk
         cancel_armed = False  # latched once we observe a live "running" state
 
@@ -236,6 +342,7 @@ class Ingester:
                 written_paths=written,
                 post_record_paths=list(post_records),
                 relink_source_paths=list(relink),
+                mark_seen_after_import=lambda: self._mark_seen(source_id, fetched),
                 stdout="\n".join(log_lines),
                 stderr="",
                 return_code=return_code,
@@ -311,7 +418,17 @@ class Ingester:
 
                 # Time-box check at the post boundary (coarse, like a gallery-dl
                 # chunk). Backfill/recovery resume from emitted_cursor next chunk.
-                if time.monotonic() - start >= time_budget_seconds:
+                # The second half is phase 3's share of the task — see
+                # CHUNK_TOTAL_SECONDS. A mid-page stop resumes the same page.
+                elapsed = time.monotonic() - start
+                phase3 = (
+                    len(written) * PHASE3_IMPORT_SECONDS
+                    + len(relink) * PHASE3_RELINK_SECONDS
+                )
+                if (
+                    elapsed >= time_budget_seconds
+                    or elapsed + phase3 >= CHUNK_TOTAL_SECONDS
+                ):
                     budget_hit = True
                     break
 
@@ -322,6 +439,20 @@ class Ingester:
                 # resume_cursor None, so everything counts.
                 if not (resume_cursor and page_cursor == resume_cursor):
                     chunk_new_posts += 1
+                # Inside the revisit window? Computed per post rather than
+                # "stop once one post is old" because the feed is only MOSTLY
+                # date-ordered — a pinned or re-pinned post can sit above older
+                # ones, and one such post must not end the walk.
+                in_window = False
+                if horizon is not None:
+                    published = _parse_published((post_meta(post) or {}).get("date"))
+                    in_window = published is not None and published >= horizon
+                # Set by the post-record block below when this post was already
+                # captured on an earlier walk. Stays False when the platform has
+                # no post-record seam, so the revisit accounting simply reports
+                # nothing rather than guessing.
+                post_already_recorded = False
+                downloaded_before = downloaded
                 # Tier-gated post (#874): the account can't fully view it, so
                 # Patreon serves only blurred locked-preview media. Skip it
                 # ENTIRELY — no media download AND no post-record stub (operator
@@ -353,11 +484,31 @@ class Ingester:
                             set() if recapture_records
                             else self._seen_keys(source_id, [pkey])
                         )
-                        if pkey not in already:
-                            rec = write_post_record(post, artist_slug)
-                            posts_recorded += 1
-                            if rec.body_chars:
-                                posts_with_body += 1
+                        post_already_recorded = pkey in already
+                        # A post inside the revisit window is re-read even
+                        # though the gate has it: that gate's whole job is to
+                        # stop us paying for a post twice, and an EDITED post is
+                        # not the same post. `revisit=True` keeps the cost at
+                        # zero requests — the downloader re-reads the body from
+                        # the feed response already in hand and declines to
+                        # write at all if that body came back empty, so a
+                        # detail-fetched body is never overwritten by a blank.
+                        if not post_already_recorded or in_window:
+                            rec = write_post_record(
+                                post, artist_slug, revisit=post_already_recorded,
+                            )
+                            if not post_already_recorded:
+                                # FIRST captures only feed the #862 body canary.
+                                # A revisit legitimately comes back empty — a
+                                # post whose body only ever arrived from the
+                                # detail endpoint has none in the feed, and the
+                                # downloader declines to write it. Counting
+                                # those into the sample would walk the canary
+                                # toward firing on healthy ticks, which is the
+                                # one thing a drift alarm must never do.
+                                posts_recorded += 1
+                                if rec.body_chars:
+                                    posts_with_body += 1
                             if rec.path is not None:
                                 post_records.append(str(rec.path))
                                 self._mark_seen(source_id, [(pkey, ppid)])
@@ -367,7 +518,8 @@ class Ingester:
                             # a 0-char body is the "why is this one empty" answer.
                             log_lines.append(
                                 f"  post {ppid} [{rec.post_type or '?'}] "
-                                f"body: {rec.body_chars} chars"
+                                + ("re-read, " if post_already_recorded else "")
+                                + f"body: {rec.body_chars} chars"
                                 + ("" if rec.body_chars else " — EMPTY")
                                 + (f" — {rec.title}" if rec.title else "")
                             )
@@ -400,6 +552,13 @@ class Ingester:
                     recapture=recapture,
                 )
 
+                # An on-disk file is only "done" if something imported it. One
+                # with no ImageRecord at its path was written by a run that died
+                # before phase 3 — it goes to import, not to the ledger.
+                imported_paths = self._recorded_paths([
+                    str(o.path) for o in outcomes
+                    if o.status == "skipped_disk" and o.path is not None
+                ])
                 to_mark: list[tuple[str, str]] = []
                 to_clear: list[str] = []  # recovered → drop any dead-letter row
                 to_fail: list[tuple[str, str, str]] = []  # (key, post_id, error)
@@ -411,11 +570,29 @@ class Ingester:
                         downloaded += 1
                         if outcome.path is not None:
                             written.append(str(outcome.path))
-                        to_mark.append((key, media_item.post_id))
+                        fetched.append((key, media_item.post_id))
                         to_clear.append(key)
                         consecutive_seen = 0
+                    elif (
+                        outcome.status == "skipped_disk"
+                        and outcome.path is not None
+                        and str(outcome.path) not in imported_paths
+                    ):
+                        # On disk, never imported: a prior run wrote it and died
+                        # before phase 3. Import it now. Safe to feed to
+                        # attach_in_place because no record owns this path —
+                        # the unlink below is about a file that IS the record.
+                        written.append(str(outcome.path))
+                        fetched.append((key, media_item.post_id))
+                        to_clear.append(key)
+                        skipped_count += 1
+                        consecutive_seen += 1
+                        log_lines.append(
+                            f"  post {media_item.post_id} — on disk but never "
+                            f"imported: {outcome.path.name}"
+                        )
                     elif outcome.status == "skipped_disk":
-                        # Already on disk (a prior run). Reconcile the ledger so a
+                        # Already on disk and imported. Reconcile the ledger so a
                         # later tick skips it at tier-1 without a disk stat, but
                         # do NOT re-feed it to phase 3 — attach_in_place would see
                         # the duplicate sha256 and unlink the on-disk copy.
@@ -451,7 +628,15 @@ class Ingester:
                         to_fail.append((key, media_item.post_id, outcome.error or "error"))
                         # An error neither advances nor resets the run-of-seen.
 
-                    if mode == "tick" and consecutive_seen >= seen_threshold:
+                    # `not in_window` is the revisit window's half of the
+                    # early-out: a run of already-seen items is only permission
+                    # to stop once the walk is BELOW the horizon. Both halves,
+                    # never either alone — see DEFAULT_REVISIT_DAYS.
+                    if (
+                        mode == "tick"
+                        and not in_window
+                        and consecutive_seen >= seen_threshold
+                    ):
                         early_out = True
                         break
 
@@ -465,6 +650,21 @@ class Ingester:
                 if to_fail:
                     self._record_failures(source_id, to_fail)
 
+                # An already-captured post that yielded NEW media is an edited
+                # post — the operator's Floppystack case, and the one thing in
+                # this walk worth naming individually in the run log. The media
+                # half needed no new detection: `extract_media` reads the media
+                # list off the live feed response, so a hotfix build appended
+                # last night is simply a ledger key we have never seen.
+                new_here = downloaded - downloaded_before
+                if post_already_recorded and new_here:
+                    revisited += 1
+                    revisit_downloads += new_here
+                    log_lines.append(
+                        f"  post {post.get('id')} — updated: "
+                        f"{new_here} new file(s)"
+                    )
+
                 # plan #709: time-throttled live progress to the running event so
                 # the Downloads view ticks ~every 5s, independent of page size.
                 now = time.monotonic()
@@ -483,9 +683,15 @@ class Ingester:
                     })
 
                 if early_out:
-                    break
+                    if skip_feed is None:
+                        break
+                    skip_feed()
+                    feeds_caught_up += 1
+                    early_out = False
+                    consecutive_seen = 0
             else:
-                reached_bottom = True
+                # A walk that left feeds early did not read to their ends.
+                reached_bottom = not feeds_caught_up
         except self._error_base as exc:
             # The platform's client-error base — _failure_result (adapter)
             # maps it to a typed error.
@@ -527,6 +733,13 @@ class Ingester:
             # visible in the Raw stdout (e.g. "bodies 3/180" reads as off).
             + (f", bodies {posts_with_body}/{posts_recorded}" if posts_recorded else "")
             + (f", {gated_skipped} gated-skipped" if gated_skipped else "")
+            # Only when it happened: on a quiet tick this is 0 and saying so
+            # every run would bury the times it is not.
+            + (
+                f", {revisited} post(s) updated ({revisit_downloads} new file(s))"
+                if revisited else ""
+            )
+            + (f", {feeds_caught_up} feed(s) caught up" if feeds_caught_up else "")
             + (", reached end" if reached_bottom else "")
             + (", time-boxed" if budget_hit else "")
         )
@@ -540,7 +753,13 @@ class Ingester:
             # next chunk resumes from the emitted cursor. No progress → TIMEOUT,
             # which feeds download_service's backfill stall-guard. rc<0 mirrors
             # subprocess TimeoutExpired so completion detection stays false.
-            made_progress = downloaded > 0 or emitted_cursor != resume_cursor
+            # Work handed to phase 3 is progress too: a recapture chunk that
+            # stopped for its imports downloaded nothing, and may not have left
+            # its first page.
+            made_progress = (
+                downloaded > 0 or bool(written) or bool(relink)
+                or emitted_cursor != resume_cursor
+            )
             if made_progress:
                 return _result(
                     success=False, return_code=-1,
@@ -746,6 +965,16 @@ class Ingester:
             )
             session.commit()
 
+    def _recorded_paths(self, paths: list[str]) -> set[str]:
+        """Which of `paths` an ImageRecord already points at."""
+        if not paths:
+            return set()
+        with self.session_factory() as session:
+            rows = session.execute(
+                select(ImageRecord.path).where(ImageRecord.path.in_(paths))
+            ).scalars().all()
+        return set(rows)
+
     def _mark_seen(self, source_id: int, items: list[tuple[str, str]]) -> None:
         """Idempotent upsert of (filehash, post_id) seen-ledger rows for a page.
 

backend/app/services/membership_reconcile.py

@@ -8,9 +8,10 @@ trustworthy enough to act on.
 1. `subscribed_not_tracked` — you pay for this and FC does not follow it. The
    adoption win, and the only bucket carrying an action.
 2. `tracked_not_subscribed` — FC follows this and the roster does not show you
-   paying for it. REPORT ONLY, by the operator's decision (2026-09-11): it says
-   what it sees and links to the existing Subscriptions row, and offers no
-   one-click disable.
+   paying for it. No longer shown on the card: the operator reversed the
+   2026-09-11 "report only" call on 2026-09-13. The lapsed half of it now ACTS,
+   in `apply_membership_lapses` below (#3995). The absent half still only
+   reports, because absence proves nothing.
 3. `matched` — the healthy set. Counted, not listed loudly.
 4. `unidentified` — sources this join cannot speak to at all. Reported as
    exactly that, because the alternative is filing them under a verdict.
@@ -39,7 +40,7 @@ rendered as lapsed. That is the whole reason it returns a tri-state.
 
 from __future__ import annotations
 
-from datetime import datetime
+from datetime import UTC, datetime
 
 from sqlalchemy import select
 from sqlalchemy.ext.asyncio import AsyncSession
@@ -47,12 +48,12 @@ from sqlalchemy.ext.asyncio import AsyncSession
 from ..models import Artist, MembershipSync, PlatformMembership, Source
 from .membership_roster import (
     get_sync_state,
-    has_paid_access,
     identity_keys_for_source,
     pair_sources_with_memberships,
     roster_is_fresh,
     url_tail,
 )
+from .native_ingest_common import has_paid_access
 
 # Why a source appears in `tracked_not_subscribed`. Ordered strongest first —
 # the UI renders a different sentence per basis, because collapsing them into
@@ -233,3 +234,126 @@ async def reconcile_all(session: AsyncSession, now: datetime | None = None) -> d
             await reconcile(session, platform=p, now=now) for p in sorted(platforms)
         ]
     }
+
+
+# ---------------------------------------------------------------------------
+# Stop pulling what the account no longer pays for (#3995)
+# ---------------------------------------------------------------------------
+#
+# Operator decision, 2026-09-13, reversing the 2026-09-11 "report only" call
+# for this direction: "if I kill a subscription on patreon I would like the
+# pulling to stop on curator as well", with automatic resume on resubscribing.
+#
+# This is a SOURCE-level action taken by the daily sweep, visible on the source
+# row and reversible there. It is not a fetch-path decision. The line C5 draws,
+# that the roster never decides a POST is inaccessible, still holds: nothing
+# here reads per-post access, and no download path reads the roster
+# (`test_no_fetch_path_can_read_the_roster`). The scheduler keeps selecting on
+# `enabled` alone.
+#
+# Acts ONLY on positive evidence. A source whose matched membership says access
+# has ended is stopped. A source with NO matched membership is left alone,
+# because absence has innocent causes: a creator rename, a source never walked
+# so no id is cached, a membership the platform stopped listing. Stopping on
+# absence would switch off things the operator still pays for.
+#
+# Two app-managed config_overrides keys carry the state. The `_` prefix is
+# already the "FC writes this, an operator edit preserves it" family.
+#   _membership_stopped  set when the sweep stops a source; the sweep resumes
+#                        ONLY sources carrying it, so a source the operator
+#                        switched off by hand is never switched back on
+#   _membership_kept     set by SourceService.update when the operator turns a
+#                        stopped source back ON: a deliberate choice to keep
+#                        pulling a lapsed creator, which the next sweep must
+#                        not undo. Cleared when the membership is paid again.
+STOPPED_KEY = "_membership_stopped"
+KEPT_KEY = "_membership_kept"
+
+
+def _access_expires_at(m: PlatformMembership) -> datetime | None:
+    """When paid access actually ends, if the platform says.
+
+    Patreon keeps a cancelled membership's access until the end of the billing
+    period and reports that date (`member.access_expires_at`, note #3992).
+    SubscribeStar's page gives no such date, so a cancelled SubscribeStar
+    membership stops at once. Returns None when there is no usable date.
+    """
+    details = m.details or {}
+    raw = details.get("access_expires_at") or (details.get("member") or {}).get("access_expires_at")
+    if not isinstance(raw, str) or not raw:
+        return None
+    try:
+        parsed = datetime.fromisoformat(raw.replace("Z", "+00:00"))
+    except ValueError:
+        return None
+    return parsed if parsed.tzinfo else parsed.replace(tzinfo=UTC)
+
+
+async def apply_membership_lapses(
+    session: AsyncSession, *, platform: str, now: datetime | None = None,
+) -> dict:
+    """Stop sources whose paid access has ended; resume the ones this stopped.
+
+    Refuses to act on a roster that isn't fresh, for the same reason C4 refuses
+    to draw conclusions from one.
+    """
+    now = now or datetime.now(UTC)
+    state = await get_sync_state(session, platform)
+    if not roster_is_fresh(state, now=now):
+        return {"platform": platform, "skipped": "roster not fresh", "stopped": 0, "resumed": 0}
+
+    memberships = (await session.execute(
+        select(PlatformMembership).where(PlatformMembership.platform == platform)
+    )).scalars().all()
+    sources = (await session.execute(
+        select(Source).where(Source.platform == platform)
+    )).scalars().all()
+    pairs = pair_sources_with_memberships(list(sources), list(memberships))
+
+    stopped: list[int] = []
+    resumed: list[int] = []
+    for source in sources:
+        pair = pairs.get(source.id)
+        if pair is None:
+            continue  # absence is never acted on, see above
+        m, _kind = pair
+        paid = has_paid_access(
+            m.platform, m.status,
+            is_free_member=bool((m.details or {}).get("is_free_member")),
+        )
+        co = dict(source.config_overrides or {})
+
+        if paid is True:
+            changed = co.pop(KEPT_KEY, None) is not None
+            if STOPPED_KEY in co:
+                co.pop(STOPPED_KEY)
+                source.enabled = True
+                resumed.append(source.id)
+                changed = True
+            if changed:
+                source.config_overrides = co
+            continue
+
+        # Unknown status: never a reason to stop something (has_paid_access's
+        # tri-state exists for exactly this).
+        if paid is None:
+            continue
+        if not source.enabled or co.get(KEPT_KEY):
+            continue
+        expires = _access_expires_at(m)
+        if expires is not None and expires > now:
+            continue  # still inside the paid-through period
+
+        co[STOPPED_KEY] = {"at": now.isoformat(), "status": m.status}
+        source.config_overrides = co
+        source.enabled = False
+        # The same clean slate a manual disable gives (SourceService.update,
+        # #1285), so a stopped source doesn't linger as failing or gated.
+        source.last_error = None
+        source.error_type = None
+        source.consecutive_failures = 0
+        stopped.append(source.id)
+
+    await session.commit()
+    return {"platform": platform, "stopped": len(stopped), "resumed": len(resumed)}
+

backend/app/services/membership_roster.py

@@ -20,8 +20,10 @@ own word — `active_patron`, not some normalised FC value. The mapping from
 those words to FC's meaning is a read-site concern and belongs in code that can
 be corrected without a migration, because the vocabulary comes from whatever
 each platform says and will be discovered per platform rather than designed up
-front. `MEMBERSHIP_STATUS` below is a place for that knowledge to accumulate as
-platforms are characterised; it is deliberately empty of guesses today.
+front. `native_ingest_common.MEMBERSHIP_STATUS` is where that knowledge
+accumulates as platforms are characterised, and it holds no guesses. It lives
+there rather than here because platform clients need it, and a client may not
+import this module (test_gated_reason.py).
 """
 
 from __future__ import annotations
@@ -35,78 +37,10 @@ from sqlalchemy.dialects.postgresql import insert as pg_insert
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from ..models import MembershipSync, PlatformMembership, Source
+from .native_ingest_common import has_paid_access
 
 log = logging.getLogger(__name__)
 
-# Platform word -> whether the account currently has paid access.
-#
-# Every entry here must come from a CHARACTERISED response, never from API docs
-# or a plausible guess — project rule 130, and inventing a status before seeing
-# it in a real payload is exactly the failure it names.
-#
-# patreon: from a live capture of the operator's own session, 2026-09-10
-# (Scribe note #3886). Only two values were OBSERVED in `patron_status` and
-# only those two are here.
-#
-# `declined_patron` is deliberately ABSENT even though it looks obviously
-# right. It appears in the request's `filter[membership_type]`, and the capture
-# proved that filter is NOT the same vocabulary as the attribute — a row
-# selected by the filter as `free_member` came back with
-# `patron_status: former_patron`, a word the filter does not contain. Reading
-# the filter as an enum is the specific mistake the capture caught; adding
-# `declined_patron` on the strength of it would be repeating that mistake one
-# step later.
-#
-# Unknown words are NOT an error: an unrecognised status means the roster
-# records evidence it cannot yet interpret, which is a better state than
-# dropping the row or asserting a meaning for it.
-#
-# subscribestar: from a live capture of the account's /subscriptions page,
-# 2026-09-13 (Scribe note #3989). SubscribeStar gives NO per-row status word —
-# a membership's state is which of two tables it sits in — so the "word" stored
-# is the table card's own `data-identifier`, verbatim. Those two identifiers are
-# the whole vocabulary; there is nothing further to characterise later.
-MEMBERSHIP_STATUS: dict[str, dict[str, bool]] = {
-    "patreon": {
-        "active_patron": True,
-        "former_patron": False,
-    },
-    "subscribestar": {
-        "active_subscriptions": True,
-        "cancelled_subscriptions": False,
-    },
-}
-
-
-def has_paid_access(
-    platform: str, status: str | None, *, is_free_member: bool = False,
-) -> bool | None:
-    """Does this membership mean the account currently PAYS for access?
-
-    Returns None for a status this code has not been taught, which callers must
-    treat as "unknown" rather than as False. The difference matters: False says
-    the operator has lost access, and asserting that from an unrecognised word
-    would tell them to cancel a source they are still paying for.
-
-    `is_free_member` is a second axis, not a status, and that is Patreon's
-    design rather than ours: the capture shows a free follow expressed as a
-    boolean alongside `patron_status`, so a "current" membership can still be
-    one nobody is paying for. Taking status alone would report a free follower
-    as a paying patron, and C4 would then never offer to clean it up.
-
-    (Honest limit: the capture contains no ACTIVE free member, so it cannot
-    demonstrate the two axes coming apart. The separation is what the payload's
-    shape says; the sample only shows it is possible, not that it happens.)
-    """
-    if status is None:
-        return None
-    known = MEMBERSHIP_STATUS.get(platform, {}).get(status)
-    if known is None:
-        return None
-    if not known:
-        return False
-    return not is_free_member
-
 
 async def touch_membership(
     session: AsyncSession,

backend/app/services/ml/ccip.py

@@ -35,6 +35,58 @@ DEFAULT_SIM_THRESHOLD = 0.85
 _FIGURE_KINDS = ("face", "figure")
 
 
+# How many cosine scores to hold in memory at once, per matmul block.
+# 4M float32 is 16 MB — small enough to stay in cache-friendly territory on the
+# shared ml lane, large enough that the per-call overhead stops mattering.
+_MAX_SCORE_ELEMS = 4_000_000
+
+
+def char_maxima(q_by_image, allref, seg, np, *, max_elems=_MAX_SCORE_ELEMS):
+    """(n_images, n_chars) — each image's best cosine to each character.
+
+    `q_by_image` is one L2-normalised `(n_figures, dim)` array per image, in
+    the order the answer comes back in. `allref` is every character's
+    prototypes stacked, and `seg` their per-character start offsets into it.
+
+    ## Why this is batched, and why that is safe
+
+    `scheduled_ccip_auto_apply` did this one image at a time — a `(nq, dim) @
+    (dim, total)` product per image, over every image in the library on every
+    run. At ~119k images that is 119k separate matmuls, each too small to pay
+    for its own BLAS setup, and on 2026-09-23 the daily sweep hit its 1800s
+    soft limit on the operator's instance.
+
+    Batching changes no arithmetic. The score a character gets for an image is
+    a max over that image's figures AND over that character's prototypes, and
+    max does not care in what order or grouping it is taken — so reducing the
+    prototype axis first (per row, inside a block) and the figure axis after
+    (per image, across blocks) gives exactly what the per-image loop gave.
+    That equivalence is what `test_char_maxima_matches_the_per_image_loop`
+    pins, against the naive form written out longhand.
+
+    Blocked by ROWS rather than done in one product, because the full score
+    matrix is (all figures in the chunk x every prototype) and that grows with
+    the library on both axes. The block bound is on elements, so the memory
+    this uses stays flat as either axis grows.
+    """
+    counts = [len(q) for q in q_by_image]
+    rows = np.vstack(q_by_image)
+    total = max(int(allref.shape[0]), 1)
+    block = max(1, max_elems // total)
+
+    per_row = np.empty((rows.shape[0], len(seg)), dtype=np.float32)
+    for a in range(0, rows.shape[0], block):
+        scores = rows[a:a + block] @ allref.T
+        per_row[a:a + block] = np.maximum.reduceat(scores, seg, axis=1)
+
+    # Start offset of each image's rows. Every image has at least one figure —
+    # it is in `q_by_image` because a region produced it — so these strictly
+    # increase, which is what `reduceat` needs to reduce rather than pass a row
+    # through untouched.
+    starts = np.cumsum([0] + counts[:-1])
+    return np.maximum.reduceat(per_row, starts, axis=0)
+
+
 async def _settings_threshold(session: AsyncSession) -> float:
     val = (
         await session.execute(

backend/app/services/ml/embedder.py

@@ -11,12 +11,21 @@ from pathlib import Path
 import numpy as np
 from PIL import Image, ImageFile
 
+from ..worker_lanes import LANES_BY_NAME
+
 ImageFile.LOAD_TRUNCATED_IMAGES = True
 
 # Cap torch's intra-op threads so each ml-worker replica is a bounded core
-# consumer on a shared node (torch otherwise uses all cores). Keep
-# N_replicas × this within the cores allotted to ML to avoid oversubscription.
-_INTRA_OP_THREADS = 4
+# consumer on a shared node (torch otherwise uses all cores).
+#
+# Read from the lane rather than restated here. This was a literal 4 beside a
+# comment reading "keep N_replicas x this within the cores allotted to ML" —
+# a constraint written where nothing could act on it, and nothing did: the ML
+# ceiling came from memory alone, offered the operator ~49 slots on a
+# large-memory host, and the lane spent 2026-09-23 with ~200 torch threads on
+# it. `derived_ceiling` now divides the cores by this number, which only means
+# anything while the two are the same number.
+_INTRA_OP_THREADS = LANES_BY_NAME["ml"].threads_per_slot
 
 DEFAULT_MODEL_NAME = os.environ.get(
     "SIGLIP_MODEL_NAME", "google/siglip-so400m-patch14-384"

backend/app/services/native_ingest_common.py

@@ -231,7 +231,7 @@ class Membership:
 
     `status` carries the PLATFORM's own word, verbatim and unmapped
     (`active_patron`, `former_patron`, ...). Deciding what it means is the read
-    site's job — `membership_roster.has_paid_access` — precisely so an
+    site's job — `has_paid_access`, below — precisely so an
     unrecognised word records as evidence rather than as a decision.
 
     `is_free_member` is SEPARATE from status and must stay that way. Patreon
@@ -396,3 +396,81 @@ class BaseNativeDownloader:
         sidecar_path = media_path.with_suffix(".json")
         sidecar_path.write_text(json.dumps(data, indent=2))
         return sidecar_path
+
+# --- membership status vocabulary (#387) ------------------------------------
+#
+# Lives here, beside `Membership`, rather than in `membership_roster`. It is
+# pure platform knowledge with no database behind it, and the platform clients
+# need it too. Patreon's must tell a lapsed membership to a deleted creator
+# (skippable) from a paid one it cannot attribute (drift), and a client may not
+# import `membership_roster`: test_gated_reason.py forbids any fetch path from
+# reaching the roster, so the roster can explain a skip but never cause one.
+#
+# Platform word -> whether the account currently has paid access.
+#
+# Every entry here must come from a CHARACTERISED response, never from API docs
+# or a plausible guess — project rule 130, and inventing a status before seeing
+# it in a real payload is exactly the failure it names.
+#
+# patreon: from a live capture of the operator's own session, 2026-09-10
+# (Scribe note #3886). Only two values were OBSERVED in `patron_status` and
+# only those two are here.
+#
+# `declined_patron` is deliberately ABSENT even though it looks obviously
+# right. It appears in the request's `filter[membership_type]`, and the capture
+# proved that filter is NOT the same vocabulary as the attribute — a row
+# selected by the filter as `free_member` came back with
+# `patron_status: former_patron`, a word the filter does not contain. Reading
+# the filter as an enum is the specific mistake the capture caught; adding
+# `declined_patron` on the strength of it would be repeating that mistake one
+# step later.
+#
+# Unknown words are NOT an error: an unrecognised status means the roster
+# records evidence it cannot yet interpret, which is a better state than
+# dropping the row or asserting a meaning for it.
+#
+# subscribestar: from a live capture of the account's /subscriptions page,
+# 2026-09-13 (Scribe note #3989). SubscribeStar gives NO per-row status word —
+# a membership's state is which of two tables it sits in — so the "word" stored
+# is the table card's own `data-identifier`, verbatim. Those two identifiers are
+# the whole vocabulary; there is nothing further to characterise later.
+MEMBERSHIP_STATUS: dict[str, dict[str, bool]] = {
+    "patreon": {
+        "active_patron": True,
+        "former_patron": False,
+    },
+    "subscribestar": {
+        "active_subscriptions": True,
+        "cancelled_subscriptions": False,
+    },
+}
+
+
+def has_paid_access(
+    platform: str, status: str | None, *, is_free_member: bool = False,
+) -> bool | None:
+    """Does this membership mean the account currently PAYS for access?
+
+    Returns None for a status this code has not been taught, which callers must
+    treat as "unknown" rather than as False. The difference matters: False says
+    the operator has lost access, and asserting that from an unrecognised word
+    would tell them to cancel a source they are still paying for.
+
+    `is_free_member` is a second axis, not a status, and that is Patreon's
+    design rather than ours: the capture shows a free follow expressed as a
+    boolean alongside `patron_status`, so a "current" membership can still be
+    one nobody is paying for. Taking status alone would report a free follower
+    as a paying patron, and C4 would then never offer to clean it up.
+
+    (Honest limit: the capture contains no ACTIVE free member, so it cannot
+    demonstrate the two axes coming apart. The separation is what the payload's
+    shape says; the sample only shows it is possible, not that it happens.)
+    """
+    if status is None:
+        return None
+    known = MEMBERSHIP_STATUS.get(platform, {}).get(status)
+    if known is None:
+        return None
+    if not known:
+        return False
+    return not is_free_member

backend/app/services/patreon_client.py

@@ -58,6 +58,7 @@ from .native_ingest_common import (
     NativeDriftError,
     NativeIngestError,
     basename_from_url,
+    has_paid_access,
     make_session,
     retry_after_seconds,
 )
@@ -482,8 +483,14 @@ class PatreonClient:
 
     @staticmethod
     def post_meta(post: dict) -> dict:
-        """Title + published date for a post — for the preview sample (plan #708
-        B4). Part of the client contract `ingest_core.Ingester.preview` calls."""
+        """Title + published date for a post. Part of the client contract.
+
+        Written for a preview sample (plan #708 B4) whose caller has since gone;
+        as of 2026-09-23 its consumer is the core's REVISIT WINDOW, which needs
+        a post's date to know whether a tick is still inside it. Both native
+        clients answer in the same shape — an ISO-8601 string under `date`, or
+        None — so the core reads a date without knowing the platform.
+        """
         attrs = post.get("attributes") or {}
         title = attrs.get("title")
         published = attrs.get("published_at")
@@ -631,7 +638,25 @@ class PatreonClient:
                 "cannot tell a complete roster from a truncated one"
             )
 
-    def _membership(self, member: dict, index: dict) -> Membership:
+    def _membership(self, member: dict, index: dict) -> Membership | None:
+        """One member row as a Membership, or None for a row the roster can skip.
+
+        The one skippable row is a LAPSED membership whose creator no longer
+        exists. The live roster (note #3886, CORRECTION 3) returned 104 rows,
+        because FC sends no membership-type filter and so gets lapses going back
+        years. One of them, a membership that ended in 2017, carried no
+        `campaign` relationship at all: the key is absent, not null, and its
+        reward names no campaign either. The creator's page is gone.
+
+        Raising on that row made the whole roster unusable over one membership
+        nobody can act on. Skipping it changes no conclusion. A lapsed
+        membership already means "not paying", absence means the same, and no
+        Source can be matched to a campaign that no longer has an id.
+
+        The refusal stays for every other row. An active or unrecognised
+        membership without a creator is something FC cannot vouch for, and
+        dropping it would read downstream as a cancellation.
+        """
         attrs = member.get("attributes") or {}
         if "patron_status" not in attrs:
             raise PatreonDriftError(
@@ -640,6 +665,17 @@ class PatreonClient:
 
         campaign_ids = self._related_ids(member, "campaign")
         if not campaign_ids:
+            paid = has_paid_access(
+                "patreon", attrs.get("patron_status"),
+                is_free_member=bool(attrs.get("is_free_member")),
+            )
+            if paid is False:
+                log.info(
+                    "Patreon roster: skipping a lapsed membership with no campaign "
+                    "(creator deleted); status=%s access_expires_at=%s",
+                    attrs.get("patron_status"), attrs.get("access_expires_at"),
+                )
+                return None
             raise PatreonDriftError(
                 "Patreon member resource has no campaign relationship — a "
                 "membership we cannot attribute to a creator is not usable"
@@ -700,7 +736,9 @@ class PatreonClient:
             index = self._transform(response)
             rows = [m for m in (response.get("data") or []) if isinstance(m, dict)]
             for member in rows:
-                yield self._membership(member, index)
+                membership = self._membership(member, index)
+                if membership is not None:
+                    yield membership
 
             seen += len(rows)
             total = int(response["meta"]["pagination"]["total"] or 0)

backend/app/services/patreon_downloader.py

@@ -340,7 +340,7 @@ class PatreonDownloader(BaseNativeDownloader):
 
     def _write_sidecar_data(
         self, post: dict, sidecar_path: Path, *, source_url: str | None = None,
-        minimal: bool = False,
+        minimal: bool = False, detail_fetch: bool = True,
     ) -> Path:
         """Serialize the post's metadata to `sidecar_path`. The post-only record
         (`write_post_record`) writes the FULL post (body/title/date/url); the
@@ -364,7 +364,13 @@ class PatreonDownloader(BaseNativeDownloader):
         # dict — so a multi-image post fetches detail at most once, the post-record
         # body-length read reuses it, and a fully-seen post (no fresh download → no
         # sidecar write) never pays the extra GET.
-        if (not content or not content.strip()) and self._content_fetcher:
+        # `detail_fetch=False` on a REVISIT (a post inside the tick's revisit
+        # window that we already captured): re-read the body from the feed
+        # response we are holding and pay nothing. Without this a 30-day window
+        # would buy one detail GET per body-less post per tick, forever — a
+        # per-creator cost that grows with how prolific they are, to re-fetch a
+        # body we already stored.
+        if (not content or not content.strip()) and self._content_fetcher and detail_fetch:
             fetched = self._content_fetcher(str(post.get("id") or ""))
             if fetched:
                 content = fetched
@@ -386,7 +392,9 @@ class PatreonDownloader(BaseNativeDownloader):
         sidecar_path.write_text(json.dumps(data, indent=2))
         return sidecar_path
 
-    def write_post_record(self, post: dict, artist_slug: str) -> PostRecordOutcome:
+    def write_post_record(
+        self, post: dict, artist_slug: str, *, revisit: bool = False,
+    ) -> PostRecordOutcome:
         """Write a post-ONLY sidecar (no media file) for a media-less post, so
         the importer can still upsert the Post + its body — text posts often hold
         the only copy of an external <a href> link. Named `_post.json`: the
@@ -397,6 +405,18 @@ class PatreonDownloader(BaseNativeDownloader):
         Returns a PostRecordOutcome (path None when the post has no id) carrying
         the captured body's shape — post_type + final char count — so the engine
         can log per-post handling without re-reading the post itself.
+
+        `revisit=True` is the tick re-reading a post it already captured
+        (ingest_core's revisit window, #...). Two differences, both about not
+        making an update cost more than it is worth:
+
+        * no detail-fetch — the body comes from the feed response already in
+          hand, so a revisit costs zero requests;
+        * a body that comes back empty writes NOTHING and returns `path=None`.
+          On a first capture an empty body is the truth about the post; on a
+          revisit it usually just means this post's body only ever came from
+          the detail endpoint we just declined to call, and writing it would
+          blank a stored body to say something we never learned.
         """
         attrs = post.get("attributes") or {}
         title = attrs.get("title") if isinstance(attrs.get("title"), str) else None
@@ -406,9 +426,17 @@ class PatreonDownloader(BaseNativeDownloader):
             return PostRecordOutcome(
                 path=None, post_type=post_type, title=title, body_chars=0,
             )
+        if revisit:
+            feed_body = post_body_html(attrs)
+            if not (isinstance(feed_body, str) and feed_body.strip()):
+                return PostRecordOutcome(
+                    path=None, post_type=post_type, title=title, body_chars=0,
+                )
         post_dir = self.images_root / artist_slug / "patreon" / post_dir_name(post)
         post_dir.mkdir(parents=True, exist_ok=True)
-        path = self._write_sidecar_data(post, post_dir / "_post.json")
+        path = self._write_sidecar_data(
+            post, post_dir / "_post.json", detail_fetch=not revisit,
+        )
         # _write_sidecar_data has by now memoized any detail-fetched body onto
         # post["attributes"]["content"], so re-read it for the FINAL char count.
         body = (post.get("attributes") or {}).get("content")

backend/app/services/pixiv_client.py

@@ -1,579 +0,0 @@
-"""Native Pixiv client — the Pixiv adapter's read path.
-
-Pixiv has a real (if unofficial) API: the mobile app API gallery-dl drives
-(`PixivAppAPI`). Per the downloader ground rule — gallery-dl is the
-known-working base — this client mirrors gallery-dl 1.32.5's request profile
-EXACTLY: the same iOS app headers on every request, the same OAuth
-refresh-token dance against oauth.secure.pixiv.net (X-Client-Time +
-X-Client-Hash), and the same `/v1/user/illusts` walk paginated by `next_url`.
-Deviating from that profile is how the SubscribeStar/Patreon spikes broke, so
-any change here should be diffed against gallery-dl's extractor first.
-
-Feed shape (characterized from gallery-dl 1.32.5, extractor/pixiv.py):
-  - `GET /v1/user/illusts?user_id=<id>` returns `{"illusts": [work...],
-    "next_url": "https://app-api...?user_id=..&offset=30" | null}`.
-  - Pagination: re-issue the SAME endpoint with `next_url`'s query params. The
-    query string doubles as our resumable page cursor (re-fetching it re-serves
-    the same page — the ingest-core resume contract).
-  - A work carries id/title/type(illust|manga|ugoira)/caption(HTML)/
-    create_date(ISO+09:00)/tags[{name,translated_name}]/user/page_count/
-    x_restrict/series/total_view/total_bookmarks/meta_single_page/meta_pages.
-  - Files: multi-page → meta_pages[].image_urls.original; single page →
-    meta_single_page.original_image_url; ugoira → `/v1/ugoira/metadata` zip
-    (600x600 → 1920x1080 URL swap, gallery-dl's default non-original mode).
-
-`campaign_id` for Pixiv is the numeric user id (extracted from the source URL
-by `user_id_from_url` — no network resolver needed).
-
-Gated works: pixiv serves a `https://s.pximg.net/common/images/limit_*.png`
-placeholder as the "original" when a work is blocked for this account
-(sanity-level filter, my-pixiv lock, deleted). gallery-dl's fallback for those
-is a web-AJAX scrape that needs PHPSESSID browser cookies — FC stores only the
-OAuth refresh token, so (exactly like our previous gallery-dl configuration,
-which warned "No PHPSESSID cookie set") those works are skipped, via the
-post_is_gated seam. Auth failures are loud (rotate the refresh token); a
-response missing the fields we depend on is DRIFT (update this client).
-
-FC runs on a plain-HTTP homelab; nothing here uses a secure-context Web API.
-"""
-
-from __future__ import annotations
-
-import hashlib
-import logging
-import time
-from collections.abc import Iterator
-from dataclasses import dataclass
-from datetime import UTC, datetime
-from urllib.parse import parse_qsl, urlsplit
-
-import requests
-
-from ..utils.paths import safe_ext
-from .native_ingest_common import (
-    _MAX_429_RETRIES,
-    NativeAuthError,
-    NativeDriftError,
-    NativeIngestError,
-    make_session,
-    retry_after_seconds,
-)
-
-log = logging.getLogger(__name__)
-
-_TIMEOUT_SECONDS = 30.0
-_API_ROOT = "https://app-api.pixiv.net"
-_OAUTH_URL = "https://oauth.secure.pixiv.net/auth/token"
-
-# gallery-dl's public Pixiv-app credentials (PixivAppAPI, also pixivpy's) —
-# these identify the official iOS app to the API, NOT the operator; the
-# operator's identity is the OAuth refresh token.
-_CLIENT_ID = "MOBrBDS8blbauoSck0ZfDbtuzpyT"
-_CLIENT_SECRET = "lsACyCD94FhDUtGTXi3QzcFE2uU1hqtDaKeqrdwj"
-_HASH_SECRET = (
-    "28c1fdd170a5204386cb1313c7077b34"
-    "f83e4aaf4aa829ce78c231e05b0bae2c"
-)
-
-# The exact header set gallery-dl 1.32.5 installs on its session — the proven
-# app-API request profile. The Referer also unlocks i.pximg.net media GETs
-# (403 without it), so the downloader reuses this constant.
-PIXIV_APP_HEADERS = {
-    "App-OS": "ios",
-    "App-OS-Version": "16.7.2",
-    "App-Version": "7.19.1",
-    "User-Agent": "PixivIOSApp/7.19.1 (iOS 16.7.2; iPhone12,8)",
-    "Referer": "https://app-api.pixiv.net/",
-}
-
-# Placeholder image prefix pixiv serves instead of a blocked work's original
-# (limit_sanity_level / limit_mypixiv / limit_unknown variants).
-_LIMIT_URL = "https://s.pximg.net/common/images/limit_"
-
-# The app API reports rate-limiting as an error MESSAGE (often on HTTP 403),
-# not only as HTTP 429. gallery-dl sleeps 300s in-walk; sleeping that long
-# inside our time-boxed chunk would eat the whole budget, so we surface it as
-# a typed 429 and let download_service's cooldown machinery honor the wait.
-_RATE_LIMIT_RETRY_AFTER = 300.0
-
-_TITLE_MAX = 50  # gallery-dl pixiv filename template: {title[:50]}
-
-_RATINGS = {0: "General", 1: "R-18", 2: "R-18G"}
-
-
-class PixivAPIError(NativeIngestError):
-    """Base for native Pixiv client failures. status_code / retry_after are
-    inherited from NativeIngestError."""
-
-
-class PixivAuthError(PixivAPIError, NativeAuthError):
-    """Auth failure — missing/expired/revoked OAuth refresh token. Fix =
-    rotate the credential (Settings → Credentials → Pixiv), not update the
-    client. Maps to error_type 'auth_error'."""
-
-
-class PixivDriftError(PixivAPIError, NativeDriftError):
-    """A response did not match the shape this client depends on (missing
-    `illusts`, un-parseable JSON where JSON was promised). Fail loud so the
-    run flags 'the Pixiv app API changed' instead of silently importing
-    nothing. Maps to API_DRIFT."""
-
-
-@dataclass
-class MediaItem:
-    """One resolved downloadable file belonging to a Pixiv work.
-
-    Fields mirror the other native clients' MediaItem so the downloader and
-    ledger are structurally the same. Pixiv original URLs carry no content
-    hash, so `filehash` is always None and the ledger keys on
-    `<post_id>:<media_id>` where media_id is `p<num>` (page) or `ugoira`
-    (the frame zip) — stable across URL-shape drift.
-    """
-
-    url: str
-    filename: str
-    kind: str
-    filehash: str | None
-    post_id: str
-    media_id: str
-
-
-def user_id_from_url(url: str) -> str | None:
-    """The numeric pixiv user id from a source URL, or None.
-
-    Handles the modern forms FC accepts as sources
-    (https://www.pixiv.net/users/<id>, /en/users/<id>) plus the legacy
-    member.php?id=<id>. This IS the campaign id — no network resolver.
-    """
-    parts = urlsplit(url or "")
-    if "pixiv.net" not in parts.netloc:
-        return None
-    segs = [s for s in parts.path.split("/") if s]
-    if segs and segs[0] == "en":
-        segs = segs[1:]
-    if len(segs) >= 2 and segs[0] == "users" and segs[1].isdigit():
-        return segs[1]
-    if segs and segs[0] == "member.php":
-        qid = dict(parse_qsl(parts.query)).get("id", "")
-        if qid.isdigit():
-            return qid
-    return None
-
-
-def _work_filename(work: dict, num: int, url: str) -> str:
-    """gallery-dl layout parity: `{id}_{title[:50]}_{num:>02}.{extension}`
-    (the downloader sanitizes the final segment)."""
-    title = work.get("title")
-    title50 = (title if isinstance(title, str) else "")[:_TITLE_MAX]
-    ext = safe_ext(urlsplit(url).path.rsplit("/", 1)[-1])
-    return f"{work.get('id')}_{title50}_{num:02d}{ext}"
-
-
-class PixivClient:
-    """Synchronous Pixiv app-API read client. Construct with the operator's
-    OAuth refresh token (the same token-type Credential the gallery-dl path
-    consumed as `extractor.pixiv.refresh-token`)."""
-
-    def __init__(
-        self,
-        refresh_token: str | None,
-        *,
-        request_sleep: float = 0.0,
-        max_retries: int = _MAX_429_RETRIES,
-        session: requests.Session | None = None,
-    ):
-        self.refresh_token = refresh_token
-        self._request_sleep = request_sleep or 0.0
-        self._max_retries = max_retries
-        # No cookies — the app API authenticates via the Bearer token _login
-        # installs. make_session still supplies the retry/UA plumbing; the
-        # extra_headers overwrite its browser UA with the app profile.
-        self._session = (
-            session if session is not None
-            else make_session(None, extra_headers=PIXIV_APP_HEADERS)
-        )
-        self._authed_user: dict = {}
-        # Monotonic deadline after which the access token must be refreshed;
-        # 0 forces a refresh on first use.
-        self._token_deadline = 0.0
-
-    # -- auth ----------------------------------------------------------------
-
-    def _login(self) -> None:
-        """Exchange the refresh token for a Bearer access token (gallery-dl's
-        `_login_impl`, including the X-Client-Time/X-Client-Hash pair the
-        endpoint validates). No-op while the current token is still fresh."""
-        if time.monotonic() < self._token_deadline:
-            return
-        if not self.refresh_token:
-            raise PixivAuthError(
-                "No Pixiv refresh token configured — add the OAuth refresh "
-                "token as the Pixiv credential (token type)."
-            )
-        # gallery-dl stamps naive-UTC with a literal +00:00 suffix.
-        now = datetime.now(UTC).strftime("%Y-%m-%dT%H:%M:%S+00:00")
-        headers = {
-            "X-Client-Time": now,
-            "X-Client-Hash": hashlib.md5(
-                (now + _HASH_SECRET).encode()
-            ).hexdigest(),
-        }
-        data = {
-            "client_id": _CLIENT_ID,
-            "client_secret": _CLIENT_SECRET,
-            "grant_type": "refresh_token",
-            "refresh_token": self.refresh_token,
-            "get_secure_url": "1",
-        }
-        try:
-            resp = self._session.post(
-                _OAUTH_URL, data=data, headers=headers,
-                timeout=_TIMEOUT_SECONDS,
-            )
-        except requests.RequestException as exc:
-            raise PixivAPIError(f"Pixiv OAuth request failed: {exc}") from exc
-        if resp.status_code >= 400:
-            raise PixivAuthError(
-                "Pixiv rejected the refresh token (HTTP "
-                f"{resp.status_code}) — rotate the Pixiv credential.",
-                status_code=resp.status_code,
-            )
-        try:
-            payload = resp.json()["response"]
-            access = payload["access_token"]
-        except (ValueError, KeyError, TypeError) as exc:
-            raise PixivDriftError(
-                f"Pixiv OAuth response shape changed: {exc}"
-            ) from exc
-        self._authed_user = payload.get("user") or {}
-        self._session.headers["Authorization"] = f"Bearer {access}"
-        # expires_in is 3600 today; refresh 60s early so a long walk never
-        # rides an expiring token into a spurious 400.
-        expires_in = payload.get("expires_in")
-        lifetime = float(expires_in) if isinstance(expires_in, (int, float)) else 3600.0
-        self._token_deadline = time.monotonic() + max(60.0, lifetime - 60.0)
-
-    # -- request -------------------------------------------------------------
-
-    def _call(self, endpoint: str, params: dict) -> dict:
-        """Authenticated app-API GET → parsed JSON body, with the shared 429
-        backoff and the loud auth/drift/rate-limit mapping."""
-        self._login()
-        if self._request_sleep > 0:
-            time.sleep(self._request_sleep)
-        url = _API_ROOT + endpoint
-        attempt = 0
-        while True:
-            try:
-                resp = self._session.get(
-                    url, params=params, timeout=_TIMEOUT_SECONDS
-                )
-            except requests.RequestException as exc:
-                raise PixivAPIError(
-                    f"Pixiv request failed ({endpoint}): {exc}"
-                ) from exc
-            if resp.status_code == 429 and attempt < self._max_retries:
-                attempt += 1
-                delay = retry_after_seconds(resp, attempt)
-                log.warning(
-                    "Pixiv 429 (%s) — backing off %.1fs (retry %d/%d)",
-                    endpoint, delay, attempt, self._max_retries,
-                )
-                time.sleep(delay)
-                continue
-            break
-
-        try:
-            body = resp.json()
-        except ValueError as exc:
-            raise PixivDriftError(
-                f"Pixiv returned non-JSON for {endpoint} "
-                f"(HTTP {resp.status_code})"
-            ) from exc
-
-        error = body.get("error") if isinstance(body, dict) else None
-        message = ""
-        if isinstance(error, dict):
-            message = str(
-                error.get("user_message") or error.get("message") or ""
-            )
-        # Rate limiting first: the app API reports it as an error MESSAGE
-        # (often on HTTP 403), which must not be mistaken for an auth failure.
-        if resp.status_code == 429 or "rate limit" in message.lower():
-            raise PixivAPIError(
-                f"Pixiv rate limit hit ({endpoint}): {message or 'HTTP 429'}",
-                status_code=429,
-                retry_after=_RATE_LIMIT_RETRY_AFTER,
-            )
-        if resp.status_code in (400, 401, 403):
-            # Invalid/expired access token surfaces as 400 invalid_grant-style
-            # errors on the app API; 401/403 are straight auth rejections.
-            raise PixivAuthError(
-                f"Pixiv rejected the request ({endpoint}, HTTP "
-                f"{resp.status_code}): {message or 'auth rejected'} — "
-                "rotate the Pixiv refresh token.",
-                status_code=resp.status_code,
-            )
-        if resp.status_code >= 400:
-            raise PixivAPIError(
-                f"Pixiv API error ({endpoint}, HTTP {resp.status_code}): "
-                f"{message or 'unknown error'}",
-                status_code=resp.status_code,
-            )
-        if error:
-            # HTTP 200 carrying an error object — unexpected, but never
-            # silently treat it as data.
-            raise PixivAPIError(
-                f"Pixiv API error ({endpoint}): {message or error}",
-                status_code=resp.status_code,
-            )
-        return body
-
-    # -- normalization -------------------------------------------------------
-
-    @staticmethod
-    def _normalize(work: dict) -> dict:
-        """Wrap an app-API work in the `{"id", "attributes", ...}` post shape
-        the platform-agnostic core and shared helpers read. The raw work rides
-        along under `_work` for extract_media / the post record."""
-        title = work.get("title")
-        caption = work.get("caption")
-        wtype = work.get("type")
-        return {
-            "id": work.get("id"),
-            "attributes": {
-                "title": title if isinstance(title, str) else "",
-                "content": caption if isinstance(caption, str) else "",
-                "published_at": work.get("create_date"),
-                "post_type": wtype if isinstance(wtype, str) else "illust",
-            },
-            "_work": work,
-        }
-
-    # -- post-first seams ----------------------------------------------------
-
-    @staticmethod
-    def post_record_key(post: dict) -> tuple[str, str] | None:
-        """`(ledger_key, post_id)` gating post-record capture through the seen
-        ledger (`post:<id>` synthetic key), or None when the work has no id."""
-        pid = post.get("id")
-        pid = str(pid) if pid is not None else ""
-        if not pid:
-            return None
-        return (f"post:{pid}", pid)
-
-    @staticmethod
-    def post_meta(post: dict) -> dict:
-        attrs = post.get("attributes") or {}
-        return {"title": attrs.get("title") or None, "date": attrs.get("published_at")}
-
-    @staticmethod
-    def post_is_gated(post: dict) -> bool:
-        """True when this account cannot fetch the work's real files: pixiv
-        substitutes a `limit_*` placeholder for the original (sanity-level
-        filter / my-pixiv lock / deleted), or zeroes the author (deleted
-        account). Mirrors #874 semantics: gated content leaves NO trace — a
-        placeholder thumbnail and an empty stub would only pollute the
-        archive. (gallery-dl's PHPSESSID web-scrape fallback for these is out
-        of scope: FC holds no pixiv browser cookies — module docstring.)"""
-        work = post.get("_work") or {}
-        user = work.get("user") or {}
-        if not user.get("id"):
-            return True
-        if work.get("meta_pages"):
-            return False
-        single = work.get("meta_single_page") or {}
-        original = single.get("original_image_url")
-        return isinstance(original, str) and original.startswith(_LIMIT_URL)
-
-    # -- media ---------------------------------------------------------------
-
-    def extract_media(self, post: dict, included_index: dict) -> list[MediaItem]:
-        """Resolve a work's downloadable files (gallery-dl's `_extract_files`):
-        multi-page originals, the single-page original, or the ugoira frame
-        zip. `included_index` is unused (pixiv works are self-contained)."""
-        work = post.get("_work") or {}
-        pid = str(post.get("id") or "")
-        if not pid or self.post_is_gated(post):
-            return []
-
-        if work.get("type") == "ugoira":
-            return self._ugoira_media(work, pid)
-
-        meta_pages = work.get("meta_pages") or []
-        if meta_pages:
-            items = []
-            for num, page in enumerate(meta_pages):
-                urls = page.get("image_urls") or {}
-                url = urls.get("original")
-                if not isinstance(url, str) or not url:
-                    continue
-                items.append(
-                    MediaItem(
-                        url=url,
-                        filename=_work_filename(work, num, url),
-                        kind="image",
-                        filehash=None,
-                        post_id=pid,
-                        media_id=f"p{num}",
-                    )
-                )
-            return items
-
-        single = work.get("meta_single_page") or {}
-        url = single.get("original_image_url")
-        if not isinstance(url, str) or not url or url.startswith(_LIMIT_URL):
-            return []
-        return [
-            MediaItem(
-                url=url,
-                filename=_work_filename(work, 0, url),
-                kind="image",
-                filehash=None,
-                post_id=pid,
-                media_id="p0",
-            )
-        ]
-
-    def _ugoira_meta(self, work: dict, pid: str) -> dict | None:
-        """Fetch + memoize the ugoira metadata (frames + zip urls) for a work.
-
-        Idempotent and cached on the work dict, so the post record and the
-        media extraction share ONE `/v1/ugoira/metadata` call regardless of
-        which runs first (the core writes the post record BEFORE it extracts
-        media). Returns None — and caches the miss — on a non-auth failure
-        (matching gallery-dl's downgrade); auth failures stay loud."""
-        if "_ugoira_meta" in work:
-            return work["_ugoira_meta"]
-        try:
-            body = self._call("/v1/ugoira/metadata", {"illust_id": pid})
-            meta = body["ugoira_metadata"]
-        except PixivAuthError:
-            raise
-        except (PixivAPIError, KeyError, TypeError) as exc:
-            log.warning("Pixiv ugoira metadata failed for %s: %s", pid, exc)
-            work["_ugoira_meta"] = None
-            return None
-        work["_ugoira_meta"] = meta
-        # Frame delays: a future ugoira→video conversion needs the timings (the
-        # zip alone has none), so the post record captures them.
-        work["_ugoira_frames"] = meta.get("frames") or []
-        return meta
-
-    def fetch_ugoira_frames(self, post: dict) -> None:
-        """Populate `post['_work']['_ugoira_frames']` for an ugoira post (no-op
-        otherwise). The core writes the post record BEFORE extract_media, so
-        without this the frame timings would never reach the record; this
-        fetches (and memoizes, so extract_media reuses it) the metadata. Injected
-        into the downloader by the ingester, mirroring Patreon's content_fetcher.
-        Auth errors propagate; other failures leave frames unset."""
-        work = post.get("_work") or {}
-        if work.get("type") != "ugoira":
-            return
-        pid = str(post.get("id") or "")
-        if pid:
-            self._ugoira_meta(work, pid)
-
-    def _ugoira_media(self, work: dict, pid: str) -> list[MediaItem]:
-        """The ugoira frame zip (gallery-dl's default non-original mode):
-        `/v1/ugoira/metadata` → zip_urls.medium with the 600x600→1920x1080
-        swap. A metadata failure downgrades to 'no media' with a warning
-        (matching gallery-dl) instead of failing the walk — except auth
-        failures, which stay loud."""
-        meta = self._ugoira_meta(work, pid)
-        if meta is None:
-            return []
-        try:
-            zip_url = meta["zip_urls"]["medium"]
-        except (KeyError, TypeError) as exc:
-            log.warning("Pixiv ugoira zip url missing for %s: %s", pid, exc)
-            return []
-        url = zip_url.replace("_ugoira600x600", "_ugoira1920x1080", 1)
-        return [
-            MediaItem(
-                url=url,
-                filename=_work_filename(work, 0, url),
-                kind="ugoira",
-                filehash=None,
-                post_id=pid,
-                media_id="ugoira",
-            )
-        ]
-
-    # -- iteration -----------------------------------------------------------
-
-    def iter_posts(
-        self, campaign_id: str, cursor: str | None = None
-    ) -> Iterator[tuple[dict, dict, str | None]]:
-        """Yield (post, {}, page_cursor) for every work in the user's feed.
-
-        `campaign_id` is the numeric pixiv user id. `cursor` is the query
-        string of the app API's `next_url` (offset pagination); None fetches
-        page 1. The yielded `page_cursor` is the cursor that FETCHED this
-        work's page, so the core checkpoints a value that re-serves the same
-        page on resume (the shared cursor contract)."""
-        if not str(campaign_id or "").isdigit():
-            raise PixivDriftError(
-                f"Pixiv campaign id must be a numeric user id, got "
-                f"{campaign_id!r}"
-            )
-        current = cursor
-        while True:
-            page_cursor = current
-            if current is None:
-                params: dict = {"user_id": campaign_id}
-            else:
-                params = dict(parse_qsl(current))
-            data = self._call("/v1/user/illusts", params)
-            works = data.get("illusts")
-            if not isinstance(works, list):
-                raise PixivDriftError(
-                    "Pixiv user-illusts response had no 'illusts' list "
-                    f"(keys: {sorted(data)[:8]})"
-                )
-            for work in works:
-                if not isinstance(work, dict):
-                    continue
-                yield self._normalize(work), {}, page_cursor
-            next_url = data.get("next_url")
-            if not next_url:
-                return
-            current = str(next_url).rpartition("?")[2]
-
-    # -- user detail ---------------------------------------------------------
-
-    def resolve_display_name(self, user_id: str) -> str | None:
-        """The pixiv user's display name via `/v1/user/detail` (gallery-dl's
-        user_detail) — used to name the Artist when a source is added by numeric
-        id. None on any failure (the caller falls back to the id)."""
-        try:
-            body = self._call("/v1/user/detail", {"user_id": str(user_id)})
-        except PixivAPIError:
-            return None
-        name = (body.get("user") or {}).get("name") if isinstance(body, dict) else None
-        return name if isinstance(name, str) and name.strip() else None
-
-    # -- verify --------------------------------------------------------------
-
-    def verify_auth(self) -> tuple[bool | None, str]:
-        """Cheap credential probe: run the OAuth refresh (the thing that fails
-        when the token is bad) without walking any feed."""
-        try:
-            self._token_deadline = 0.0  # force a real refresh
-            self._login()
-        except PixivAuthError as exc:
-            return False, f"Pixiv rejected the credential — {exc}"
-        except PixivAPIError as exc:
-            return None, f"Couldn't verify (network/HTTP issue): {exc}"
-        account = self._authed_user.get("account") or self._authed_user.get("name")
-        suffix = f" as {account}" if account else ""
-        return True, f"Credentials valid — Pixiv OAuth refresh succeeded{suffix}."
-
-
-def rating_label(x_restrict) -> str | None:
-    """Human rating from pixiv's x_restrict (0/1/2) — written into the post
-    record so the archive keeps the R-18 flag without the reader needing to
-    know pixiv's numeric scheme."""
-    if isinstance(x_restrict, bool) or not isinstance(x_restrict, int):
-        return None
-    return _RATINGS.get(x_restrict)

backend/app/services/pixiv_downloader.py

@@ -1,276 +0,0 @@
-"""Native Pixiv media downloader — the Pixiv counterpart to
-patreon_downloader / subscribestar_downloader.
-
-Given a normalized Pixiv work and its resolved `MediaItem`s
-(pixiv_client.extract_media), download the originals to gallery-dl's on-disk
-layout (so pre-cutover gallery-dl downloads are recognized on disk and not
-re-fetched), write the post-first sidecars the importer consumes, and report
-per-media outcomes.
-
-On-disk layout (matches FC's gallery-dl pixiv config, PLATFORM_DEFAULTS:
-base-directory `<images_root>/<artist_slug>/pixiv` + `directory:
-["{category}"]` + filename `{id}_{title[:50]}_{num:>02}.{extension}`):
-
-    <images_root>/<artist_slug>/pixiv/pixiv/<id>_<title50>_<NN>.<ext>
-
-— note the intentional DOUBLE `pixiv` segment: gallery-dl appended
-`{category}` under a base-directory that already ended in the platform name,
-and tier-2 disk-skip parity requires reproducing that exactly. The layout is
-FLAT (no per-post directory), so the post-first record is `_post_<id>.json`
-in the same directory (the id suffix prevents the collisions a bare
-`_post.json` would have here; phase 3 receives explicit post_record_paths, so
-the name is a convention, not a discovery key).
-
-Simpler than Patreon (no Mux/yt-dlp video branch) — the one special file is
-the ugoira frame zip, downloaded as-is; FC's archive-containment import
-extracts the frames, and the frame DELAYS ride the post record (the zip
-carries none — a future ugoira→video conversion needs them).
-
-PURE: no DB; the seen-skip is an injected predicate. FC runs on a plain-HTTP
-homelab; nothing here uses a secure-context Web API.
-"""
-
-from __future__ import annotations
-
-import json
-import logging
-import re
-import time
-from collections.abc import Callable
-from pathlib import Path
-
-import requests
-
-from .native_ingest_common import (
-    BaseNativeDownloader,
-    MediaOutcome,
-    PostRecordOutcome,
-)
-from .pixiv_client import PIXIV_APP_HEADERS, rating_label
-
-log = logging.getLogger(__name__)
-
-# Control chars (0x00–0x1f + 0x7f DEL) — gallery-dl's default `path-remove`.
-_GDL_PATH_REMOVE_RE = re.compile(r"[\x00-\x1f\x7f]")
-
-
-def gdl_clean_filename(name: str) -> str:
-    """Reproduce gallery-dl's on-disk filename EXACTLY as it wrote it on this
-    Linux host, so the tier-2 disk-skip recognizes pre-cutover files instead of
-    re-downloading them.
-
-    gallery-dl's PathFormat.build_filename is `clean_path(clean_segment(name))`.
-    On Linux (verified against gallery-dl 1.32.5 path.py) the defaults resolve to:
-      - path-restrict "auto" → "/"          → clean_segment replaces ONLY "/" → "_"
-      - path-remove   "\\x00-\\x1f\\x7f"       → clean_path DELETES control chars
-      - path-strip    "auto" → ""           → NO trailing dot/space stripping
-    Crucially it does NOT touch the Windows-forbidden set (<>:"|?*) — those stay
-    raw in titles on disk. A stricter sanitizer here would rename any such title,
-    miss the on-disk match, and re-pull the whole work. Order mirrors gallery-dl
-    (segment inner, path outer); for these disjoint char sets it's commutative.
-    """
-    return _GDL_PATH_REMOVE_RE.sub("", name.replace("/", "_"))
-
-# Enrichment keys copied verbatim from the app-API work dict into the post
-# record (they're already JSON scalars/objects). Everything lands in
-# Post.raw_metadata via the importer, so the archive keeps pixiv's stats and
-# structure without a schema change.
-_WORK_PASSTHROUGH_KEYS = (
-    "type",
-    "page_count",
-    "width",
-    "height",
-    "total_view",
-    "total_bookmarks",
-    "total_comments",
-    "is_bookmarked",
-    "illust_ai_type",
-    "series",
-)
-
-
-class PixivDownloader(BaseNativeDownloader):
-    """Download resolved Pixiv media to gallery-dl's on-disk layout.
-    Subclasses BaseNativeDownloader for the shared streaming GET
-    (transient-retry + Range-resume) and validation/quarantine. PURE: no DB."""
-
-    def __init__(
-        self,
-        images_root: Path,
-        cookies_path: str | None = None,
-        *,
-        validate: bool = True,
-        rate_limit: float = 0.0,
-        session: requests.Session | None = None,
-        ugoira_frames_fetcher: Callable[[dict], None] | None = None,
-    ):
-        super().__init__(
-            images_root, cookies_path, platform="pixiv",
-            validate=validate, rate_limit=rate_limit, session=session,
-        )
-        # Injected by the ingester (client.fetch_ugoira_frames) so write_post_record
-        # can populate frame timings — which extract_media memoizes, but the core
-        # writes the post record FIRST. Mirrors Patreon's content_fetcher.
-        self._ugoira_frames_fetcher = ugoira_frames_fetcher
-        if session is None:
-            # i.pximg.net 403s any GET without the app Referer; mirror the
-            # client's full app-header profile (gallery-dl serves media off
-            # the same session it drives the API with). An injected session
-            # (tests) owns its own headers.
-            self.session.headers.update(PIXIV_APP_HEADERS)
-
-    # -- public ------------------------------------------------------------
-
-    def download_post(
-        self,
-        post: dict,
-        media_items: list,
-        artist_slug: str,
-        *,
-        is_seen: Callable[[object], bool] = lambda m: False,
-        should_stop: Callable[[], bool] = lambda: False,
-        recapture: bool = False,
-    ) -> list[MediaOutcome]:
-        """Download every media item of one work; return per-item outcomes.
-        Mirrors SubscribeStarDownloader.download_post (two-tier skip, mid-post
-        time-box, recapture surfacing)."""
-        flat_dir = self._flat_dir(artist_slug)
-        outcomes: list[MediaOutcome] = []
-        for media in media_items:
-            if should_stop():
-                break
-            try:
-                outcomes.append(
-                    self._download_one(
-                        post, media, flat_dir, artist_slug, is_seen,
-                        recapture=recapture,
-                    )
-                )
-            except Exception as exc:  # resilient: isolate one item's failure
-                log.warning(
-                    "Pixiv media failed (work %s, %s): %s",
-                    post.get("id"), getattr(media, "media_id", "?"), exc,
-                )
-                outcomes.append(
-                    MediaOutcome(media=media, status="error", path=None, error=str(exc))
-                )
-        return outcomes
-
-    def _flat_dir(self, artist_slug: str) -> Path:
-        # Double platform segment — gallery-dl layout parity (module docstring).
-        return self.images_root / artist_slug / "pixiv" / "pixiv"
-
-    # -- per-item ----------------------------------------------------------
-
-    def _download_one(
-        self,
-        post: dict,
-        media,
-        flat_dir: Path,
-        artist_slug: str,
-        is_seen: Callable[[object], bool],
-        *,
-        recapture: bool = False,
-    ) -> MediaOutcome:
-        seen = is_seen(media)
-        if seen and not recapture:
-            return MediaOutcome(media=media, status="skipped_seen", path=None, error=None)
-
-        # The client's filename already carries the {id}_{title50}_{NN} shape
-        # (raw title, gallery-dl-template order); clean it to the byte-exact
-        # name gallery-dl wrote on disk so tier-2 disk-skip matches (else a
-        # re-download of the whole work). See gdl_clean_filename.
-        media_path = flat_dir / gdl_clean_filename(media.filename)
-
-        if media_path.exists():  # tier-2: already on disk
-            return MediaOutcome(
-                media=media, status="skipped_disk", path=media_path, error=None
-            )
-        # recapture: a seen item not on disk is NOT re-downloaded (recovery's job).
-        if seen:
-            return MediaOutcome(media=media, status="skipped_seen", path=None, error=None)
-
-        flat_dir.mkdir(parents=True, exist_ok=True)
-        if self._rate_limit > 0:
-            time.sleep(self._rate_limit)
-
-        out_path = self._fetch_get(media.url, media_path)
-        reason, quarantine_dest = self._validate_path(out_path, artist_slug, media.url)
-        if reason is not None:
-            return MediaOutcome(
-                media=media, status="quarantined", path=quarantine_dest, error=reason,
-            )
-        self._write_minimal_sidecar(post, out_path, source_url=media.url)
-        return MediaOutcome(media=media, status="downloaded", path=out_path, error=None)
-
-    # -- post record ---------------------------------------------------------
-
-    def write_post_record(self, post: dict, artist_slug: str) -> PostRecordOutcome:
-        """Write the post-first `_post_<id>.json` — the sole writer of the post
-        body/metadata on the native path. Beyond the standard body fields, the
-        record carries pixiv's own structure (tags + EN translations, rating,
-        series, view/bookmark counts, AI flag, dimensions, author, ugoira frame
-        delays) so the archive keeps what the platform knows about the work."""
-        attrs = post.get("attributes") or {}
-        work = post.get("_work") or {}
-        title = attrs.get("title") if isinstance(attrs.get("title"), str) else None
-        post_type = attrs.get("post_type") if isinstance(attrs.get("post_type"), str) else None
-        pid = str(post.get("id") or "")
-        if not pid:
-            return PostRecordOutcome(
-                path=None, post_type=post_type, title=title, body_chars=0,
-            )
-
-        content = attrs.get("content")
-        content = content if isinstance(content, str) else ""
-        data: dict = {
-            "category": "pixiv",
-            "id": pid,
-            "title": title or "",
-            "content": content,
-            "published_at": attrs.get("published_at"),
-            # The post permalink is synthesized by platforms/pixiv.py
-            # derive_post_url from `id` at parse time — no url key here.
-            "rating": rating_label(work.get("x_restrict")),
-        }
-        for key in _WORK_PASSTHROUGH_KEYS:
-            if key in work:
-                data[key] = work[key]
-        tags = work.get("tags")
-        if isinstance(tags, list):
-            data["tags"] = [
-                {
-                    "name": t.get("name"),
-                    "translated_name": t.get("translated_name"),
-                }
-                for t in tags
-                if isinstance(t, dict)
-            ]
-        user = work.get("user")
-        if isinstance(user, dict):
-            data["user"] = {
-                "id": user.get("id"),
-                "account": user.get("account"),
-                "name": user.get("name"),
-            }
-        # Ugoira frame timings. extract_media memoizes these, but the core writes
-        # the post record BEFORE extracting media, so fetch them here (shared +
-        # idempotent via the client's memoization) so the record actually keeps
-        # them — the zip carries no timings.
-        if (
-            work.get("type") == "ugoira"
-            and not work.get("_ugoira_frames")
-            and self._ugoira_frames_fetcher is not None
-        ):
-            self._ugoira_frames_fetcher(post)
-        frames = work.get("_ugoira_frames")
-        if frames:
-            data["ugoira_frames"] = frames
-
-        flat_dir = self._flat_dir(artist_slug)
-        flat_dir.mkdir(parents=True, exist_ok=True)
-        path = flat_dir / f"_post_{pid}.json"
-        path.write_text(json.dumps(data, indent=2, ensure_ascii=False))
-        return PostRecordOutcome(
-            path=path, post_type=post_type, title=title, body_chars=len(content),
-        )

backend/app/services/pixiv_ingester.py

@@ -1,121 +0,0 @@
-"""Native Pixiv ingester — the Pixiv ADAPTER over the platform-agnostic core
-(`ingest_core.Ingester`).
-
-Thin counterpart to patreon_ingester / subscribestar_ingester: wires the Pixiv
-client/downloader/ledger models/constraints/key into the core and supplies the
-Pixiv failure mapping. The modes (tick / backfill / recovery / recapture), the
-seen + dead-letter ledgers, cursor checkpointing, and the post-first capture
-all live in the core. `download_service.download_source` drives
-`PixivIngester.run` exactly as it drives the other two.
-
-`campaign_id` is the numeric pixiv user id (download_backends extracts it from
-the source URL — no network resolver). Auth is the operator's OAuth refresh
-token (the token-type Credential), passed as `auth_token` — pixiv is the first
-native platform authenticating by token rather than cookies, so the uniform
-constructor accepts both and ignores what it doesn't need.
-
-FC runs on a plain-HTTP homelab; nothing here uses a secure-context Web API.
-"""
-
-from __future__ import annotations
-
-import asyncio
-import logging
-from collections.abc import Callable
-from pathlib import Path
-
-from ..models import PixivFailedMedia, PixivSeenMedia
-from .ingest_core import DEAD_LETTER_THRESHOLD, Ingester
-from .pixiv_client import MediaItem, PixivAPIError, PixivClient
-from .pixiv_downloader import PixivDownloader
-
-__all__ = [
-    "DEAD_LETTER_THRESHOLD",
-    "PixivIngester",
-    "_ledger_key",
-    "verify_pixiv_credential",
-]
-
-log = logging.getLogger(__name__)
-
-_LEDGER_KEY_MAX = 128
-
-
-def _ledger_key(media: MediaItem) -> str:
-    """Stable per-media identity for the cross-run seen-ledger. Pixiv original
-    URLs carry no content hash, so the key is the page/zip identity scoped to
-    its work: `<illust_id>:p<num>` / `<illust_id>:ugoira`. Bounded to the
-    column width."""
-    if media.filehash:
-        return media.filehash
-    return f"{media.post_id}:{media.media_id}"[:_LEDGER_KEY_MAX]
-
-
-class PixivIngester(Ingester):
-    """Walk a pixiv user's works, download unseen originals, return a
-    `DownloadResult`. A thin adapter over `ingest_core.Ingester`; `client` /
-    `downloader` are injectable seams so unit tests run without network."""
-
-    def __init__(
-        self,
-        images_root: Path,
-        cookies_path: str | None,
-        session_factory: Callable[[], object],
-        *,
-        validate: bool = True,
-        rate_limit: float = 0.0,
-        request_sleep: float = 0.0,
-        auth_token: str | None = None,
-        client: PixivClient | None = None,
-        downloader: PixivDownloader | None = None,
-    ):
-        self.images_root = Path(images_root)
-        self.cookies_path = str(cookies_path) if cookies_path else None
-        resolved_client = (
-            client
-            if client is not None
-            else PixivClient(auth_token, request_sleep=request_sleep)
-        )
-        resolved_downloader = (
-            downloader
-            if downloader is not None
-            else PixivDownloader(
-                self.images_root, cookies_path, validate=validate, rate_limit=rate_limit,
-                # write_post_record runs before extract_media in the core, so it
-                # fetches ugoira frame timings via the SAME client (shared,
-                # memoized) — else the record's ugoira_frames stays empty.
-                ugoira_frames_fetcher=resolved_client.fetch_ugoira_frames,
-            )
-        )
-        super().__init__(
-            client=resolved_client,
-            downloader=resolved_downloader,
-            session_factory=session_factory,
-            seen_model=PixivSeenMedia,
-            failed_model=PixivFailedMedia,
-            seen_constraint="uq_pixiv_seen_media_source_id",
-            failed_constraint="uq_pixiv_failed_media_source_id",
-            ledger_key=_ledger_key,
-            platform="pixiv",
-            error_base=PixivAPIError,
-            # API_DRIFT message phrasing; the base Ingester._failure_result owns
-            # the auth/drift/HTTP→error_type mapping (shared across platforms).
-            drift_label="Pixiv app API",
-            # Captions are legitimately empty for many pixiv artists, so the
-            # zero-bodies #862 canary would false-positive here; the client's
-            # response-shape checks (missing `illusts` → drift) cover the same
-            # failure class structurally.
-            body_canary=False,
-        )
-
-
-async def verify_pixiv_credential(
-    auth_token: str | None,
-) -> tuple[bool | None, str]:
-    """Native Pixiv credential probe — one OAuth refresh via
-    PixivClient.verify_auth (the exact call that fails when the token is
-    bad; no feed walk). Returns the uniform `(ok, message)` contract so
-    download_backends.verify_source_credential treats it like the others."""
-    client = PixivClient(auth_token)
-    loop = asyncio.get_running_loop()
-    return await loop.run_in_executor(None, client.verify_auth)

backend/app/services/platform_lock.py

@@ -23,8 +23,10 @@ log = logging.getLogger(__name__)
 # Platforms walked one-at-a-time. gallery-dl platforms are intentionally NOT
 # here: each runs as a self-pacing subprocess and they're lower-volume. The
 # native-ingester platforms are serialized (one paced scrape/API walk at a time).
-# Add a platform here to cap it to a single concurrent walk.
-SERIALIZED_PLATFORMS = frozenset({"patreon", "subscribestar"})
+# Add a platform here to cap it to a single concurrent walk. Discord most of
+# all: every source walks on the operator's ONE user token, and parallel walks
+# on a user account are both how its rate limit trips and what gets it flagged.
+SERIALIZED_PLATFORMS = frozenset({"patreon", "subscribestar", "discord"})
 
 _LOCK_PREFIX = "fc:download_lock:"
 

backend/app/services/platforms/__init__.py

@@ -13,9 +13,8 @@ URL patterns match GS exactly so the existing browser extension
 hits FC unmodified. deviantart was dropped at #3069 (2026-08-27) —
 FC downloaders are art-dedicated services only. pixiv was retired at
 milestone #406 (2026-09-13, rule #171): unregistered here first, which
-switches it off everywhere this registry is consulted; `pixiv.py` and the
-pixiv client/downloader/ingester stay in the tree, uncalled, until the
-milestone's phase 2 deletes them.
+switched it off everywhere this registry is consulted, then removed from
+the tree entirely in the milestone's phase 2 (2026-09-21).
 """
 
 from .base import (

backend/app/services/platforms/base.py

@@ -24,8 +24,8 @@ from typing import Literal
 # external_post_id chain: `post_id` MUST come before `id` because
 # SubscribeStar gallery-dl puts the per-attachment id in `id` and the
 # actual post id in `post_id`; picking `id` first fragments
-# multi-image SubscribeStar posts into N Post rows. Patreon/Pixiv have
-# no `post_id` so `id` still wins for them; HF uses `index`, Discord
+# multi-image SubscribeStar posts into N Post rows. Patreon has
+# no `post_id` so `id` still wins for it; HF uses `index`, Discord
 # uses `message_id` — all reached via the remaining chain entries.
 # (Banked 2026-05-27 during the sidecar audit.)
 DEFAULT_EXTERNAL_POST_ID_KEYS: tuple[str, ...] = (
@@ -62,7 +62,7 @@ class PlatformInfo:
     # --- Behavioral hooks ---
     # Synthesize a post permalink from sidecar data. Required when
     # gallery-dl's `url` field is the file/CDN URL rather than the post
-    # permalink (subscribestar/pixiv/hf/discord). None = trust the bare
+    # permalink (subscribestar/hf/discord). None = trust the bare
     # `url` field (patreon).
     derive_post_url: Callable[[dict], str | None] | None = None
 

backend/app/services/platforms/pixiv.py

@@ -1,38 +0,0 @@
-"""Pixiv — one quirk.
-
-post_url: the sidecar's `url` (legacy gallery-dl era) is the image URL
-on `i.pximg.net`, and the native post record (#129) writes no url key
-at all — the permalink is synthesized from `id` here either way:
-/artworks/<id>. external_post_id (= `id`) was already correct, so no
-override there.
-
-Downloads run through the native ingester (pixiv_ingester.py), not
-gallery-dl; this registry entry still owns URL validation, sidecar
-parsing, and the credential surface (the OAuth refresh token).
-"""
-
-from .base import GD_DEFAULTS, PlatformInfo, str_id_value
-
-
-def derive_post_url(data: dict) -> str | None:
-    pid = str_id_value(data.get("id"))
-    if pid:
-        return f"https://www.pixiv.net/artworks/{pid}"
-    return None
-
-
-INFO = PlatformInfo(
-    key="pixiv",
-    name="Pixiv",
-    description="Download artwork from Pixiv artists",
-    auth_type="token",
-    requires_auth=True,
-    url_pattern=r"^https?://(www\.)?pixiv\.net/",
-    url_examples=[
-        "https://www.pixiv.net/users/12345678",
-        "https://www.pixiv.net/en/users/12345678",
-    ],
-    default_config={**GD_DEFAULTS, "content_types": ["all"]},
-    notes="Requires OAuth refresh token. Run `gallery-dl oauth:pixiv` to obtain one.",
-    derive_post_url=derive_post_url,
-)

backend/app/services/post_association_service.py

@@ -15,24 +15,56 @@ where they already were, a wrong one actively misinforms and then propagates
 into whatever reads the association. So the matcher's job is to make a SHORT
 list worth reading, not a long list worth trusting.
 
-## Signals, and the one deliberately NOT built
+## Two routes, because the evidence is of two different kinds
+
+CIRCUMSTANTIAL evidence says two things happened near each other. It is
+additive, weighted, and no single one of its signals may reach the threshold:
 
 1. **Time proximity.** The Patreon post exists in order to announce the drop,
    so the two are minutes-to-hours apart. Nearly free, and strong.
 2. **The post says so.** These announcements routinely name Discord or carry
    an invite link, which is close to a declaration.
+3. **A shared marker.** The creator's own tie-back — `🍈🍈` in the Patreon
+   title and `@everyone 🍈 🍈` in the Discord message — gated on how rare that
+   marker is in THIS artist's posts, because a habitual emoji is punctuation.
 
-3. **Crop-to-source matching is HELD, on the plan's own instruction** — it is
-   real work with real false-positive risk, and it is only worth building once
-   1 and 2 are shown to be insufficient against the operator's actual artists.
-   Nothing here should be read as evidence it is unnecessary; it is deferred,
-   and the thing that would justify it is an empty review queue on a pair the
-   operator can see with their own eyes.
+IDENTITY evidence says two things are the same thing, and it gets its own
+route (see `IDENTITY_FLOOR`). Two signals, and the stronger one stands rather
+than them being summed — saying "the same piece" twice is not more true:
 
-   Note also that a naive whole-image SigLIP similarity is NOT that signal. A
-   cropped teaser and its full version are exactly the pair a whole-image
-   comparison handles worst, so adding one as a "bonus" would mostly add noise
-   while looking like progress.
+4. **A shared working name.** The creator exports the teaser and the release
+   from one file, and the internal name survives into both platforms
+   untouched. Measured on the operator's artist: `ConnFront` ↔ `ConnFront`.
+   This is the only signal that reaches a pair 23.8 hours apart, which
+   proximity scores at 0.005.
+5. **The drop contains the teaser's image.** Rare, and near-certain when it
+   happens. It is the one signal needing no cooperation from the creator: it
+   works on a teaser called `Screenshot 2026-08-13`, and on a creator whose
+   two platforms share no naming convention.
+
+## The one deliberately NOT built
+
+**Crop-to-source matching stays held, and now for a measured reason rather
+than a cautious one.**
+
+It was deferred until the cheap signals could be shown insufficient. They can:
+of artist 8's 27 teasers with a drop inside a day, 11 still go unlinked, and
+five of those are screenshot teasers carrying no working name at all.
+
+So it was tried, on those exact pairs. Every teaser image was correlated
+against every window of every nearby drop image at five scales, with the pairs
+the working name independently confirms as ground truth and unrelated
+same-artist posts a month away as a control. **It does not separate.** True
+pairs score as low as 0.401 while the control reaches 0.605 — the two
+distributions overlap, and no threshold divides them.
+
+The reason is the reason the naive version was rejected in the first place,
+and it turns out to apply just as hard to the sophisticated one: one artist's
+work is stylistically homogeneous, so any whole-image comparison between two
+of their pieces is high whether or not it is the same piece. Signal 5 above is
+what survived that experiment — it asks a narrower question ("is this the same
+image") that the measurement shows is answerable, instead of a broader one
+("is this a crop of that") that it shows is not.
 
 ## Creator identity comes free, so E4 is not actually a prerequisite
 
@@ -56,14 +88,27 @@ from __future__ import annotations
 
 import logging
 import re
+from collections import Counter
+from dataclasses import dataclass
 from datetime import UTC, datetime, timedelta
 
-from sqlalchemy import func, or_, select
+from sqlalchemy import and_, func, or_, select
 from sqlalchemy.ext.asyncio import AsyncSession
 
-from ..models import ImportSettings, Post, PostAssociation
+from ..models import ImageRecord, ImportSettings, Post, PostAssociation
+from ..utils.phash import hamming, hash_bits
 from ..utils.text import html_to_plain
 from .discord_grouping import DROP_GROUPER
+from .post_naming import (
+    IDENTITY_FLOOR,
+    MAX_TOKEN_POSTS,
+    marker_frequencies,
+    marker_overlap,
+    rarity,
+    shared_identity,
+    token_frequencies,
+    working_name_tokens,
+)
 
 log = logging.getLogger(__name__)
 
@@ -76,18 +121,83 @@ log = logging.getLogger(__name__)
 # enough" arithmetic rather than aspirational: on a busy day an artist posts
 # several times, and a matcher that could pair on proximity alone would turn
 # every busy day into false pairs. A guard test pins this.
-WEIGHTS = {"proximity": 0.55, "declared": 0.45}
+WEIGHTS = {"proximity": 0.45, "declared": 0.35, "marker": 0.20}
 
 # A Discord INVITE in the body is close to a declaration; the bare word is
 # weaker but still meaningful, because these posts are short and on-topic.
+#
+# "the server" and its possessives are here because the word `discord` is NOT
+# how these creators actually write. Measured across 20,558 Patreon bodies:
+# `discord` appears in 486 and `the server` in 37 — but the distribution is the
+# point, not the totals. For the artist this step was built for, 21 of 42 posts
+# say `discord` and 7 say `the server`, and it is the RECENT ones that say the
+# latter: the phrasing drifted once the audience already knew where the server
+# was. A vocabulary list written from old posts silently stops matching.
 _INVITE = re.compile(r"discord\.(?:gg|com/invite)/", re.I)
-_MENTION = re.compile(r"\bdiscord\b", re.I)
+_MENTION = re.compile(r"\b(?:discord|(?:the|our|my)\s+server)\b", re.I)
 
 DECLARED_INVITE = 1.0
 DECLARED_MENTION = 0.6
 
 MAX_CANDIDATES = 25
 
+# How many just-grouped drops one sweep will look around. A ceiling on the
+# `or_` the sweep builds, not a policy — a backfill that authors thousands of
+# drops at once should not turn one sweep into a full-library rescan, which is
+# the manual button's job.
+MAX_RECENT_DROPS = 200
+
+# What a shared name must reach before FC links a pair WITHOUT asking.
+#
+# 1.0, which under post_naming's post-span counting means the name appears in
+# exactly these two posts and nowhere else in the artist's library. That is not
+# "strong evidence" — within the library it is conclusive, and the remaining
+# ways to be wrong are a mis-parse or the creator reusing a name for a genuinely
+# different piece on the same day.
+#
+# Deliberately above IDENTITY_FLOOR, which is what a name needs to PROPOSE.
+# The gap between them is the review queue: real evidence, not certain enough
+# for FC to act on by itself. Measured on artist 8, 15 name-sharing pairs: 11
+# are conclusive, 2 more propose, 2 fall short of both.
+AUTO_LINK_FLOOR = 1.0
+
+# When the drop simply CONTAINS the teaser's image — a pHash within this many
+# of 256 bits.
+#
+# 32, the same number and unit `gallery_service._diversify_similar` already
+# calls a near-duplicate. Measured on artist 8, comparing every teaser against
+# every drop within a day: pairs the working name independently confirms score
+# 0, 0 and 20, and the nearest unrelated same-artist pair in a 29-sample
+# control scores **108**. A 76-bit gap, so the threshold is not finely tuned
+# and does not need to be.
+#
+# `utils/phash` warns that the hash alone must not decide a MERGE, because
+# variants of one piece collide at this distance. That warning does not invert
+# here, it is the point: merging destroys a file, so a variant colliding with
+# its original is a loss, while this is asking whether two POSTS are about the
+# same piece — and a variant of the drop's image is exactly that. Nothing is
+# deleted either way, so no pixel confirm is needed to accept.
+DUPLICATE_MAX_DISTANCE = 32
+
+
+@dataclass(frozen=True)
+class _Corpus:
+    """One artist's rare-token evidence, gathered once rather than per pair.
+
+    Both rare-token signals are scoped to a single artist — a working name and
+    a marker belong to the person who chose them — so the counts are useless
+    across artists and expensive to rebuild per candidate. A sweep touches an
+    artist's posts many times over; this is loaded on the first touch and kept
+    for the life of the service.
+    """
+
+    tokens_by_post: dict[int, set[str]]
+    token_posts: Counter[str]
+    text_by_post: dict[int, str]
+    marker_posts: Counter[str]
+    hashes_by_post: dict[int, list[int]]
+    hash_posts: Counter[int]
+
 
 def proximity_signal(gap: timedelta, window: timedelta) -> float:
     """1.0 when the two posts are simultaneous, decaying linearly to 0 at the
@@ -123,6 +233,40 @@ def declared_signal(description: str | None) -> float:
     return 0.0
 
 
+def shared_image(
+    left: list[int],
+    right: list[int],
+    hash_posts: Counter[int],
+    *,
+    max_distance: int = DUPLICATE_MAX_DISTANCE,
+    max_frequency: int = MAX_TOKEN_POSTS,
+) -> float:
+    """Strength in [0, 1] that the drop contains the teaser's own image.
+
+    IDENTITY evidence, and the only one of the three that needs no cooperation
+    from the creator — it works on a teaser named `Screenshot 2026-08-13`, and
+    on a creator whose two platforms share no naming convention at all. Where
+    it fires it is close to certain; it is simply quiet most of the time,
+    because a teaser is usually a crop rather than a copy.
+
+    Rarity-gated on POSTS like the other two: an image the creator puts on many
+    posts is a banner, not a piece.
+    """
+    if not left or not right:
+        return 0.0
+    best = None
+    for a in left:
+        for b in right:
+            d = hamming(a, b)
+            if d is None or d > max_distance:
+                continue
+            span = max(hash_posts.get(a, 1), hash_posts.get(b, 1), 1)
+            strength = rarity(span, max_frequency)
+            if best is None or strength > best:
+                best = strength
+    return round(best, 4) if best is not None else 0.0
+
+
 def weighted_score(signals: dict) -> float:
     return round(sum(WEIGHTS[k] * signals.get(k, 0.0) for k in WEIGHTS), 4)
 
@@ -134,81 +278,298 @@ def _post_time(post: Post) -> datetime:
 class PostAssociationService:
     def __init__(self, session: AsyncSession):
         self.session = session
+        self._corpora: dict[int, _Corpus] = {}
 
-    async def _decided(self, announcement_id: int) -> set[int]:
-        """Payload posts already proposed for this announcement, in ANY status.
+    async def _corpus(self, artist_id: int) -> _Corpus:
+        if artist_id in self._corpora:
+            return self._corpora[artist_id]
 
-        Dismissed pairs are included deliberately: re-proposing a pair the
-        operator has already rejected on every subsequent scan is the single
-        behaviour that makes a review queue get ignored.
+        paths_by_post: dict[int, list[str]] = {}
+        hashes_by_post: dict[int, list[int]] = {}
+        # An image is counted under the post a reader SEES it on: a Discord
+        # message absorbed into a drop contributes to the drop, not to itself.
+        #
+        # Keyed on `primary_post_id` alone until 2026-09-24, which on the live
+        # instance meant a drop never had a name or a hash at all — its images
+        # are owned by its member messages, and the drop only claims them
+        # through provenance. The identity route therefore never fired there;
+        # the tests missed it because they attached images to the drop itself,
+        # which discord_grouping never does.
+        #
+        # Counting by the drop also makes the span honest: five wips of one
+        # piece posted as five messages and grouped into one drop are ONE post
+        # as far as "how many posts carry this name" is concerned.
+        owner = Post.__table__.alias("owner")
+        rows = await self.session.execute(
+            select(
+                func.coalesce(owner.c.absorbed_by_post_id, ImageRecord.primary_post_id),
+                ImageRecord.path, ImageRecord.phash,
+            )
+            .select_from(ImageRecord)
+            .join(owner, owner.c.id == ImageRecord.primary_post_id)
+            .where(
+                ImageRecord.artist_id == artist_id,
+                ImageRecord.primary_post_id.is_not(None),
+            )
+        )
+        for post_id, path, phash in rows:
+            paths_by_post.setdefault(post_id, []).append(path)
+            bits = hash_bits(phash)
+            if bits is not None:
+                hashes_by_post.setdefault(post_id, []).append(bits)
+
+        text_by_post: dict[int, str] = {}
+        rows = await self.session.execute(
+            select(Post.id, Post.post_title, Post.description).where(
+                Post.artist_id == artist_id
+            )
+        )
+        for post_id, title, description in rows:
+            text_by_post[post_id] = "\n".join(
+                part for part in (title, html_to_plain(description) or "") if part
+            )
+
+        corpus = _Corpus(
+            tokens_by_post={
+                pid: {t for path in paths for t in working_name_tokens(path)}
+                for pid, paths in paths_by_post.items()
+            },
+            # Both counts take POSTS, which is why they are built from these
+            # groupings rather than from flat lists — see post_naming.
+            token_posts=token_frequencies(paths_by_post.values()),
+            text_by_post=text_by_post,
+            marker_posts=marker_frequencies(text_by_post.values()),
+            hashes_by_post=hashes_by_post,
+            # An image the creator puts on many posts — a banner, a watermark
+            # plate, a recurring title card — is a habit exactly as a character
+            # name is, and gets gated the same way. Counted on the EXACT hash,
+            # which is what a reused file produces.
+            hash_posts=Counter(
+                h for hs in hashes_by_post.values() for h in set(hs)
+            ),
+        )
+        self._corpora[artist_id] = corpus
+        return corpus
+
+    async def _decided(self, announcement_id: int) -> dict[int, PostAssociation]:
+        """Pairs already recorded for this announcement, keyed by payload.
+
+        Linked and dismissed pairs are never touched again: re-proposing a pair
+        the operator has already rejected on every subsequent scan is the
+        single behaviour that makes a review queue get ignored.
+
+        A PENDING pair is different — nobody has decided it — so the caller
+        re-scores it. Otherwise a pair queued by an older, weaker matcher sits
+        in the queue forever even once the evidence is conclusive, which is the
+        chore the operator asked FC not to hand them.
         """
         rows = (await self.session.execute(
-            select(PostAssociation.payload_post_id)
+            select(PostAssociation)
             .where(PostAssociation.announcement_post_id == announcement_id)
         )).scalars().all()
-        return set(rows)
+        return {a.payload_post_id: a for a in rows}
 
     async def _candidate_groups(
         self, announcement: Post, *, window: timedelta,
-    ) -> list[Post]:
-        """Synthetic Discord groupings by the SAME artist, inside the window.
+    ) -> list[tuple[Post, datetime]]:
+        """Synthetic Discord groupings by the SAME artist with a message inside
+        the window — each with the time of its message CLOSEST to the teaser.
 
         Same-artist is the identity signal and it is free (see the module
         docstring on E4). It is also a hard filter rather than a scored one:
         two different creators posting minutes apart is a coincidence, not
         evidence, and letting it score at all would mean a busy hour across the
         library could out-vote everything else.
+
+        Measured on the MESSAGES, not the drop's own date. A drop is dated by
+        its first message, and since #4390 merges a creator's trickle into one
+        drop, that can be days before the release the teaser announces —
+        "Very early Marin" on Sep 7, `MarinaraSauce_base` on Sep 11. Matching
+        on the drop's date would put every merged trickle outside the window.
         """
         at = _post_time(announcement)
-        sort_key = func.coalesce(Post.post_date, Post.downloaded_at)
-        return (await self.session.execute(
-            select(Post)
+        member = Post.__table__.alias("member")
+        member_at = func.coalesce(member.c.post_date, member.c.downloaded_at)
+        rows = list((await self.session.execute(
+            select(member.c.absorbed_by_post_id, member_at)
             .where(
+                member.c.artist_id == announcement.artist_id,
+                member.c.absorbed_by_post_id.is_not(None),
+                member_at >= at - window,
+                member_at <= at + window,
+            )
+        )).all())
+        # The drop's own date counts too — the first message's, so it adds
+        # nothing for a real drop, but it keeps a drop with no member rows
+        # (hand-built, or one whose messages were removed) matchable.
+        own_at = func.coalesce(Post.post_date, Post.downloaded_at)
+        rows += (await self.session.execute(
+            select(Post.id, own_at).where(
                 Post.artist_id == announcement.artist_id,
                 Post.synthesized_by == DROP_GROUPER,
+                own_at >= at - window,
+                own_at <= at + window,
+            )
+        )).all()
+        closest: dict[int, datetime] = {}
+        for group_id, when in rows:
+            if group_id not in closest or abs(when - at) < abs(closest[group_id] - at):
+                closest[group_id] = when
+        if not closest:
+            return []
+        groups = (await self.session.execute(
+            select(Post).where(
+                Post.id.in_(list(closest)),
+                Post.synthesized_by == DROP_GROUPER,
                 Post.id != announcement.id,
-                sort_key >= at - window,
-                sort_key <= at + window,
             )
-            .order_by(sort_key)
-            .limit(MAX_CANDIDATES)
         )).scalars().all()
+        ranked = sorted(groups, key=lambda g: (abs(closest[g.id] - at), g.id))
+        return [(g, closest[g.id]) for g in ranked[:MAX_CANDIDATES]]
+
+    async def _claimed(self, announcement_id: int, payload_id: int) -> bool:
+        """Is either end of this pair already spoken for by an accepted link?
+
+        An auto-link is FC asserting something the operator never saw, so it
+        only happens where there is nothing to contradict. A drop already
+        linked to a different announcement is exactly such a contradiction, and
+        resolving it is a judgement about which one is right — which is the
+        operator's, not FC's.
+        """
+        return (await self.session.execute(
+            select(PostAssociation.id).where(
+                PostAssociation.status == "linked",
+                or_(
+                    PostAssociation.payload_post_id == payload_id,
+                    PostAssociation.announcement_post_id == announcement_id,
+                ),
+            ).limit(1)
+        )).scalar() is not None
 
     async def match_post(
         self, announcement_id: int, *, threshold: float, window_hours: float,
-    ) -> int:
-        """Score one announcement against nearby groupings. Returns proposals made."""
+        auto_link: bool = False,
+    ) -> tuple[int, int]:
+        """Score one announcement against nearby groupings.
+
+        Returns `(proposed, linked)` — how many pairs were written, and how
+        many of those were linked outright rather than queued.
+        """
         announcement = await self.session.get(Post, announcement_id)
         if announcement is None or announcement.synthesized_by is not None:
             # A synthetic post cannot announce anything — FC wrote it.
-            return 0
+            return 0, 0
 
         window = timedelta(hours=window_hours)
         declared = declared_signal(announcement.description)
         already = await self._decided(announcement_id)
+        corpus = await self._corpus(announcement.artist_id)
+        here = corpus.tokens_by_post.get(announcement.id, set())
+        here_text = corpus.text_by_post.get(announcement.id, "")
+        here_hashes = corpus.hashes_by_post.get(announcement.id, [])
 
         made = 0
-        for group in await self._candidate_groups(announcement, window=window):
-            if group.id in already:
+        scored: list[tuple[Post, float, dict, float]] = []
+        for group, group_at in await self._candidate_groups(announcement, window=window):
+            prior = already.get(group.id)
+            if prior is not None and prior.status != "pending":
                 continue
-            signals = {
+            named, token = shared_identity(
+                here,
+                corpus.tokens_by_post.get(group.id, set()),
+                corpus.token_posts,
+            )
+            # The two identity signals answer the same question by different
+            # means, so the stronger one stands rather than them being summed:
+            # a name and a shared image both say "the same piece", and saying
+            # it twice is not more true.
+            copied = shared_image(
+                here_hashes,
+                corpus.hashes_by_post.get(group.id, []),
+                corpus.hash_posts,
+            )
+            identity = max(named, copied)
+            circumstantial = {
                 "proximity": proximity_signal(
-                    _post_time(group) - _post_time(announcement), window,
+                    group_at - _post_time(announcement), window,
                 ),
                 "declared": declared,
+                "marker": marker_overlap(
+                    here_text,
+                    corpus.text_by_post.get(group.id, ""),
+                    corpus.marker_posts,
+                ),
             }
-            score = weighted_score(signals)
+            score = weighted_score(circumstantial)
+            # THE TWO ROUTES, and why identity is not simply a fourth weight.
+            #
+            # Circumstance and identity answer different questions. Proximity
+            # and a declaration say two things happened near each other and
+            # that one of them mentioned Discord; a working name the creator
+            # uses on these two posts and nowhere else says they are the same
+            # piece. Averaging those makes the threshold uninterpretable, and
+            # it costs both: adding identity as a weight dilutes the others
+            # enough that measured teaser/drop pairs an hour apart stop
+            # proposing, while capping identity's contribution at its weight
+            # means the strongest evidence available can never carry a pair on
+            # its own.
+            #
+            # So identity may override, never dilute. Below the floor it is
+            # recorded for the operator to read and moves nothing — which is
+            # the conservative direction, since a wrong link asserts that two
+            # different pieces are one.
+            if identity >= IDENTITY_FLOOR:
+                score = max(score, identity)
             if score < threshold:
                 continue
+            signals = {**circumstantial, "identity": identity}
+            if copied:
+                signals["identity_image"] = copied
+            if token and named >= copied:
+                # Carried so the queue can say WHY. A review queue that cannot
+                # explain itself is one the operator learns to click through.
+                signals["identity_token"] = token
+            scored.append((group, score, signals, identity))
+
+        # Who, if anyone, FC links without asking.
+        #
+        # EXACTLY ONE candidate may be conclusive. Two drops sharing a name
+        # with one teaser at full strength is not a tie to be broken by score —
+        # it means the name identifies something other than what FC thinks it
+        # does, and the right response is to queue both and say nothing.
+        auto_id = None
+        if auto_link:
+            conclusive = [c for c in scored if c[3] >= AUTO_LINK_FLOOR]
+            if len(conclusive) == 1 and not await self._claimed(
+                announcement.id, conclusive[0][0].id
+            ):
+                auto_id = conclusive[0][0].id
+
+        linked = 0
+        for group, score, signals, _identity in scored:
+            status = "linked" if group.id == auto_id else "pending"
+            if status == "linked":
+                linked += 1
+            prior = already.get(group.id)
+            if prior is not None:
+                # Re-scored in place: a pending pair keeps its row (and id),
+                # and only an upgrade to linked counts as news.
+                prior.score = score
+                prior.signals = signals
+                if status == "linked":
+                    prior.status = "linked"
+                    prior.linked_by = "fc"
+                continue
             self.session.add(PostAssociation(
                 announcement_post_id=announcement.id,
                 payload_post_id=group.id,
                 score=score,
                 signals=signals,
-                status="pending",
+                status=status,
+                linked_by="fc" if status == "linked" else None,
             ))
             made += 1
-        return made
+        return made, linked
 
     async def list_pending(self) -> list[dict]:
         rows = (await self.session.execute(
@@ -232,14 +593,19 @@ class PostAssociationService:
         if a is None:
             return None
         a.status = "linked"
+        a.linked_by = "operator"
         return {"id": a.id, "status": a.status}
 
     async def dismiss(self, association_id: int) -> dict | None:
         a = await self.session.get(PostAssociation, association_id)
         if a is None:
             return None
-        # Kept, not deleted — the row is what remembers the rejection.
+        # Kept, not deleted — the row is what remembers the rejection. It is
+        # also the undo for a link FC made itself (#4402): the unified card
+        # dismisses the pair, and the dismissed row stops the next sweep from
+        # linking it straight back.
         a.status = "dismissed"
+        a.linked_by = None
         return {"id": a.id, "status": a.status}
 
     async def linked_for(self, post_ids: list[int]) -> dict[int, list[dict]]:
@@ -273,36 +639,124 @@ class PostAssociationService:
         return out
 
 
-async def rescan(session: AsyncSession, *, now: datetime | None = None) -> dict:
-    """Score every recent non-synthetic post against nearby groupings."""
+async def rescan(
+    session: AsyncSession, *, now: datetime | None = None, full: bool = False,
+) -> dict:
+    """Score recent non-synthetic posts against nearby groupings.
+
+    `full=True` scores EVERY post by an artist who has Discord drops at all —
+    the manual button's job, for history that predates the feature or that a
+    trickle merge (#4390) has just rearranged. It used to share the sweep's
+    48-hour horizon, so the button described as "a first run over a library
+    that predates the feature" could not reach that library.
+    """
     settings = await ImportSettings.load(session)
     if not settings.discord_link_enabled:
-        return {"enabled": False, "scanned": 0, "proposed": 0}
+        return {"enabled": False, "scanned": 0, "proposed": 0, "linked": 0}
 
     now = now or datetime.now(UTC)
     window_hours = float(settings.discord_link_window_hours)
+    window = timedelta(hours=window_hours)
     # Only look at announcements that could still have a partner in range —
     # a full-library rescan is the manual button's job, not the sweep's.
     horizon = now - timedelta(hours=window_hours * 2)
     sort_key = func.coalesce(Post.post_date, Post.downloaded_at)
-    ids = (await session.execute(
+    if full:
+        with_drops = select(Post.artist_id).where(
+            Post.synthesized_by == DROP_GROUPER
+        ).distinct()
+        ids = set((await session.execute(
+            select(Post.id).where(
+                Post.synthesized_by.is_(None),
+                Post.absorbed_by_post_id.is_(None),
+                Post.artist_id.in_(with_drops),
+            )
+        )).scalars().all())
+        return await _score(session, settings, ids, window_hours)
+    ids = set((await session.execute(
         select(Post.id).where(
             Post.synthesized_by.is_(None),
             Post.absorbed_by_post_id.is_(None),
             sort_key >= horizon,
         )
-    )).scalars().all()
+    )).scalars().all())
 
+    # ...and announcements sitting next to a drop FC has only JUST authored.
+    #
+    # A drop's `post_date` is backdated to its first message, but FC cannot
+    # write the drop until the message has an embedding and the hourly grouper
+    # has run — so a drop created this minute can land weeks back in the feed.
+    # Its neighbours were last swept before it existed, and a sweep keyed only
+    # on how recent the ANNOUNCEMENT is will never look at them again.
+    #
+    # That is #4392's third cause, and it is the one that left a measured 0.800
+    # pair with an empty review queue on the live instance. The other two were
+    # about scoring; this one meant nothing was scored at all.
+    #
+    # The times are the drops' MESSAGES, not the drops' own dates: a drop that
+    # grew today by a trickle merge (#4390) is dated by its first stage, days
+    # earlier, while the teaser sits beside the message that just joined.
+    recent = (
+        select(Post.id).where(
+            Post.synthesized_by == DROP_GROUPER,
+            func.coalesce(Post.last_grew_at, Post.downloaded_at) >= horizon,
+        )
+        .order_by(func.coalesce(Post.last_grew_at, Post.downloaded_at).desc())
+        .limit(MAX_RECENT_DROPS)
+    )
+    member = Post.__table__.alias("member")
+    drop_times = set((await session.execute(
+        select(func.coalesce(member.c.post_date, member.c.downloaded_at))
+        .where(member.c.absorbed_by_post_id.in_(recent))
+    )).scalars().all())
+    # Plus each drop's own date — its first message's, so a duplicate for a
+    # real drop, but what a drop with no member rows is placed by.
+    drop_times |= set((await session.execute(
+        select(sort_key).where(Post.id.in_(recent))
+    )).scalars().all())
+    # The interval arithmetic is done in Python rather than SQL: a handful of
+    # literal ranges is portable, and `now - INTERVAL` is not.
+    ranges = [
+        and_(sort_key >= at - window, sort_key <= at + window)
+        for at in drop_times
+        if at is not None
+    ]
+    if ranges:
+        ids |= set((await session.execute(
+            select(Post.id).where(
+                Post.synthesized_by.is_(None),
+                Post.absorbed_by_post_id.is_(None),
+                or_(*ranges),
+            )
+        )).scalars().all())
+
+    return await _score(session, settings, ids, window_hours)
+
+
+async def _score(
+    session: AsyncSession, settings: ImportSettings, ids: set[int], window_hours: float,
+) -> dict:
     svc = PostAssociationService(session)
     proposed = 0
-    for pid in ids:
-        proposed += await svc.match_post(
+    linked = 0
+    # Sorted because `ids` is now a union of two queries: set iteration order
+    # is arbitrary, and a sweep that visits posts in a different order each
+    # run is one whose failures cannot be reproduced.
+    for pid in sorted(ids):
+        made, auto = await svc.match_post(
             pid,
             threshold=float(settings.discord_link_threshold),
             window_hours=window_hours,
+            auto_link=bool(settings.discord_link_auto),
         )
+        proposed += made
+        linked += auto
     log.info(
-        "discord announcement matcher: scanned %d post(s), proposed %d pair(s)",
-        len(ids), proposed,
+        "discord announcement matcher: scanned %d post(s), proposed %d pair(s), "
+        "linked %d outright",
+        len(ids), proposed, linked,
     )
-    return {"enabled": True, "scanned": len(ids), "proposed": proposed}
+    return {
+        "enabled": True, "scanned": len(ids), "proposed": proposed,
+        "linked": linked,
+    }

backend/app/services/post_feed_service.py

@@ -19,6 +19,7 @@ from ..models import (
     ExternalLink,
     ImageProvenance,
     ImageRecord,
+    ImportSettings,
     Post,
     PostAttachment,
     Source,
@@ -118,6 +119,15 @@ class PostFeedService:
         # from. `around` and `get_post` deliberately do NOT apply this: reaching
         # a member by id is how you inspect a grouping.
         stmt = stmt.where(Post.absorbed_by_post_id.is_(None))
+        # A linked Discord drop is shown ON its teaser's card by reference
+        # (#4402), so its own card sitting beside that teaser is the same
+        # release twice. Only then is it left out — an older drop keeps its
+        # place, because a reference does not take anything out of history.
+        fold_hours = await self._fold_hours()
+        if fold_hours > 0:
+            from .post_unification import fold_clause
+
+            stmt = stmt.where(fold_clause(fold_hours))
         if artist_id is not None:
             stmt = stmt.where(Post.artist_id == artist_id)
         if platform is not None:
@@ -169,9 +179,12 @@ class PostFeedService:
         thumbs_map = await self._thumbnails_for(post_ids)
         atts_map = await self._attachments_for(post_ids)
         links_map = await self._links_for(post_ids)
+        unified_map = await self._unified_for([p for p, _, _ in rows])
 
         items = [
-            self._to_dict(post, artist, source, thumbs_map, atts_map, links_map)
+            self._to_dict(
+                post, artist, source, thumbs_map, atts_map, links_map, unified_map,
+            )
             for post, artist, source in rows
         ]
         return {"items": items, "next_cursor": next_cursor}
@@ -213,6 +226,7 @@ class PostFeedService:
         anchor_item = self._to_dict(
             anchor_post, anchor_artist, anchor_source, thumbs_map, atts_map,
             await self._links_for([anchor_post.id]),
+            await self._unified_for([anchor_post]),
         )
         return {
             "items": newer["items"] + [anchor_item] + older["items"],
@@ -239,6 +253,7 @@ class PostFeedService:
         item = self._to_dict(
             post, artist, source, thumbs_map, atts_map,
             await self._links_for([post.id]),
+            await self._unified_for([post]),
         )
         item["description_full"] = html_to_plain(post.description)
         # Full (uncapped) translated description for the detail view (#143).
@@ -385,9 +400,27 @@ class PostFeedService:
 
         return await PostAssociationService(self.session).linked_for(post_ids)
 
+    async def _unified_for(self, posts: list[Post]) -> dict[int, dict]:
+        """The reference set each teaser's card shows (#4402). Local import for
+        the same reason as `_links_for`: it reaches the association service."""
+        from .post_unification import PostUnificationService
+
+        return await PostUnificationService(self.session).unified_for(posts)
+
+    async def _fold_hours(self) -> float:
+        """`discord_link_fold_hours`, read without assuming the row exists.
+
+        The feed is the one surface that must not fail on a settings row the
+        caller never needed, so a missing row folds nothing rather than
+        raising — which is also exactly how the feed behaved before this.
+        """
+        settings = await self.session.get(ImportSettings, 1)
+        return float(settings.discord_link_fold_hours) if settings is not None else 0.0
+
     def _to_dict(
         self, post: Post, artist: Artist, source: Source | None,
         thumbs_map: dict, atts_map: dict, links_map: dict | None = None,
+        unified_map: dict | None = None,
     ) -> dict:
         plain_full = html_to_plain(post.description) if post.description else None
         if plain_full is None:
@@ -436,6 +469,11 @@ class PostFeedService:
             # post is the drop). Always a list so the UI never branches on
             # absence.
             "associations": (links_map or {}).get(post.id, []),
+            # #4402. On a teaser with a linked drop: what the card shows BY
+            # REFERENCE — the drop's images, the piece's older variants, and
+            # the text of each — plus who made each link, so one FC made by
+            # itself can say so and offer the undo. None on every other post.
+            "unified": (unified_map or {}).get(post.id),
             # Non-null on a chat message a synthetic post absorbed. The feed
             # filters these out, but `around`/`get_post` still reach them, and
             # the UI uses this to explain why a post it linked to is not in the

backend/app/services/post_naming.py

@@ -0,0 +1,426 @@
+"""The creator's own working name for a piece, recovered from a filename.
+
+Milestone 388. Pure functions, no DB and no ML — the whole point is that this
+signal is free.
+
+## What this is for
+
+Two of the operator's artists post a censored or cropped teaser on Patreon and
+the real release in their Discord. Matching those by IMAGE is the pair a
+whole-image comparison handles worst: the teaser is a crop with a censor bar,
+which is exactly the local edit that moves a perceptual hash and blurs a
+semantic embedding.
+
+But the creator names both exports after the same internal working title, and
+that name survives into both platforms untouched. Measured on the live instance
+2026-09-24, artist 8:
+
+    01_((0-k           <->  0-k_base          (1.3h apart)
+    01_680LC           <->  680LC_Border      (21.0h apart)
+    01_cnni18x         <->  cnni18x           (21.5h apart)
+
+Three pairs, no false positives, and **two of them are 21 hours apart** — far
+enough that time proximity scores them ~0.10 and could never propose them. The
+naming signal is orthogonal to the timing one: each finds pairs the other
+cannot, which is why both are kept rather than one being tuned to cover both.
+
+## Why a filename and not a perceptual hash
+
+A shared working-name token is IDENTITY evidence — `680lc` appearing on both
+platforms is not a coincidence. Proximity is CIRCUMSTANTIAL: it says two things
+happened near each other, never that they are the same thing. The distinction
+drives the weighting in `post_association_service`, and it is why a rare enough
+token is allowed to carry a proposal on its own while no amount of circumstance
+is.
+
+## The one false-positive class found, and why the fix is shaped this way
+
+A first pass matched `01_Screenshot 2026-08-13 000004` to
+`Screenshot_2026-08-13_032144` on the token `2026-08-13`, twice.
+
+A screenshot filename is a camera artifact. It carries no working name, and the
+date inside it collides across platforms on the same day BY CONSTRUCTION — the
+teaser and the release are posted the same day, so their screenshot names
+always share a date token. That is a signal that fires precisely when it is
+least informative.
+
+So a filename with no working name contributes NOTHING, rather than the
+plausible-looking date match it could be squeezed for. Re-run with that rule:
+the same three true pairs, zero false. Half of this creator's recent teasers
+are screenshots, and those pairs are simply out of this signal's reach — which
+is where crop-to-source matching earns its cost, and nowhere else.
+"""
+
+from __future__ import annotations
+
+import re
+from collections import Counter
+from collections.abc import Iterable
+from pathlib import PurePosixPath
+
+# A screenshot name, on either platform. Patreon's importer writes
+# `01_Screenshot 2026-08-13 000004`; gallery-dl's Discord naming writes
+# `Screenshot_2026-09-22_003651`. Matched after the index/message prefixes are
+# stripped, so both shapes reach this as a bare `Screenshot ...`.
+# NOT `\b` after "shot": `\b` needs a word/non-word transition and `_` is a
+# WORD character, so `Screenshot_2026-08-13_032144` — gallery-dl's Discord
+# spelling — sailed straight past the guard while the space-separated Patreon
+# spelling was caught. Found by running this against the live library rather
+# than by reading it. Assert the next character is not a letter instead.
+_SCREENSHOT = re.compile(r"^(?:screen[ _-]?shot(?![a-z])|\u30b9\u30af\u30ea\u30fc\u30f3\u30b7\u30e7\u30c3\u30c8|\u30b9\u30af\u30b7\u30e7)", re.I)
+
+# The importer's per-post media index: `01_`, `02_`. Not part of any name.
+_MEDIA_INDEX = re.compile(r"^\d{1,3}_")
+
+# gallery-dl's Discord filename pattern (#3999):
+# `{date:%Y%m%d}_{message_id}_{num:>02}_{filename}`.
+_DISCORD_PREFIX = re.compile(r"^\d{8}_\d{6,}_\d{1,3}_")
+
+# The legacy era (#4002): images sit FLAT at the artist root as
+# `<post id>_media_<media id>_<name>`. Stripping it is not cosmetic — the
+# SCREENSHOT guard below matches from the start of the stem, so while this
+# prefix was left on, a legacy screenshot never looked like one. Measured on
+# tamadaheijun: `109078417_media_334848471_Screenshot 2025-07-27 182450ab`
+# sailed through and contributed `2025-07-27`, which is precisely the
+# same-day date collision this module was built to refuse.
+_LEGACY_PREFIX = re.compile(r"^\d+_media_\d+_")
+
+# The importer's content-hash suffix, `__<10 hex>`, sometimes doubled on files
+# that went through an older import era.
+_HASH_SUFFIX = re.compile(r"(?:__[0-9a-f]{10})+$")
+
+# Generic export decorations. Stripped as SUFFIXES so the stem survives:
+# `cnni18x_wip3` and `cnni18x` must yield the same token, or a work-in-progress
+# would never match the piece it became.
+_DECORATION = re.compile(
+    r"(?:[_-]?(?:wip|base|final|alt|alts|edit|edits|border|clean|raw|hd|full|"
+    r"censored|uncensored|nsfw|sfw|ver|v)\d*)+$",
+    re.I,
+)
+
+# Tokens that carry no identity even when they survive the rules above.
+_STOPWORDS = frozenset({
+    "img", "image", "untitled", "new", "test", "page", "final", "copy",
+    "post", "media", "file", "avatar", "cover", "banner", "icon", "splash",
+    # Each of these was MEASURED carrying a false match in the 3..6 frequency
+    # band, where the rarity gate still admits a token: `capture` across six
+    # unrelated knuxy posts, `the` and `patreon` out of legacy title-derived
+    # names, `main` out of `Anya Main CST`, `timeline` and `gif` out of
+    # tamadaheijun's exports.
+    "the", "gif", "main", "patreon", "capture", "timeline", "screenshot",
+    # Literally the string "None": issue #3999's Discord naming rendered
+    # `{user[name]}` as it for ~1,600 files, so it is the single most common
+    # "name" in the library and identifies nothing.
+    "none",
+    # gallery-dl's fallback when a Discord attachment has no filename of its
+    # own. Measured on artist 8: four unrelated images across 1,974 days, and
+    # the one false family the leading-name rule admitted inside 60 days.
+    "image0",
+})
+
+# A bare year: still needed for the TEXT signal, where words and numbers are
+# tokenised separately.
+_YEAR = re.compile(r"^(?:19|20)\d{2}$")
+
+# An identity token must contain a LETTER. This replaces separate "all digits"
+# and "bare year" rules with the property behind both, and it is the rule that
+# holds once hyphens are kept inside tokens for `0-k`'s sake: without it
+# `2025-07-27` and `3-0002` survive as single tokens, and both were measured
+# linking unrelated posts — the second across three of them, out of
+# tamadaheijun's `timeline 3-0002` exports.
+#
+# `0-k`, `680lc`, `cnni18x` and `p59` all keep a letter and are unaffected.
+_HAS_LETTER = re.compile(r"[A-Za-z]")
+
+MIN_TOKEN_LEN = 3
+
+# A token appearing in more than this many of ONE ARTIST's POSTS is a habit,
+# not an identity — a character name, a series tag, a recurring export preset.
+#
+# POSTS, not files, and the difference is not bookkeeping. Counting files
+# punishes a piece for having many exports, which is the one thing a working
+# name is GUARANTEED to do. Measured across the operator's four dual-platform
+# artists: knuxy's comic pages carry `p217` on four files spread over exactly
+# two posts — the Patreon post and the Discord drop — and file-counting scored
+# every one of ~200 such tokens at half strength for it. Counting posts scores
+# them 1.00 while still catching the real habits, which span many posts:
+# tamadaheijun's `comic2` spans 8, conto's `seth2` 5, `maid` 4.
+#
+# Six rather than two, although two posts IS the shape of a teaser and its
+# drop, because a creator legitimately revisits one working name: a wip post,
+# then an alt, then the release. Measured on artist 8, `cnni18x` and `680lc`
+# each span four posts and both are genuine. IDENTITY_FLOOR below is what
+# decides how much span a link may carry on its own.
+MAX_TOKEN_POSTS = 6
+
+# A shared name at or above this strength is enough to propose a link with NO
+# corroboration — it is identity evidence, and the whole reason this module
+# exists is that identity survives where circumstance does not.
+#
+# 0.75 is a token spanning three posts or fewer. Measured on artist 8: of the
+# 15 same-artist pairs that share a name, 13 clear this bar, including the
+# operator's own example (`0-k`, three posts, 1.3h apart). The two that do not
+# — `680lc` and `cnni18x`, four posts each, 21h apart — are real pairs this
+# signal will not carry alone; they are the measured cost of not admitting the
+# four-post band, where conto's `illustration9` and `maid` also sit.
+IDENTITY_FLOOR = 0.75
+
+# A LEADING name spanning this many of ONE ARTIST's posts is a habit — a
+# character the creator returns to — not one piece's trickle. Used wherever a
+# name gathers a FAMILY: the teaser card's variants (#4401) and the Discord
+# grouper's trickle merge (#4390), so the two cannot disagree about what a
+# family is.
+#
+# Its own value rather than MAX_TOKEN_POSTS (6), which is calibrated for
+# PAIRING two posts and measured too tight for a family. On artist 8,
+# `tentacooler` spans 6 posts over 7 days and `0-k1` 6 posts over 10: both real
+# families, both gated out at 6. At 8, `anya` (7) passes the cap — and has no
+# pair inside any family window, which is what the window is for.
+FAMILY_MAX_POSTS = 8
+
+
+def _strip_prefixes(stem: str) -> str:
+    """Remove the framing each platform's importer adds around the real name."""
+    stem = _DISCORD_PREFIX.sub("", stem)
+    stem = _LEGACY_PREFIX.sub("", stem)
+    stem = _MEDIA_INDEX.sub("", stem)
+    return _HASH_SUFFIX.sub("", stem)
+
+
+def _ordered_tokens(path: str) -> list[str]:
+    """The identity-bearing tokens of one filename, in the order written.
+
+    The one tokenizer both public readings share, so the set of names and the
+    leading name cannot disagree about what counts as a name.
+    """
+    stem = _strip_prefixes(PurePosixPath(path).stem)
+    if _SCREENSHOT.match(stem.strip()):
+        return []
+
+    out: list[str] = []
+    # Hyphens are kept INSIDE tokens — `0-k` is a real working name on the live
+    # instance, and splitting on hyphen would reduce it to a single character
+    # and then discard it for being too short.
+    for raw in re.split(r"[^0-9A-Za-z-]+", stem.lower()):
+        tok = _DECORATION.sub("", raw).strip("-")
+        if len(tok) < MIN_TOKEN_LEN:
+            continue
+        if tok in _STOPWORDS or not _HAS_LETTER.search(tok):
+            continue
+        if tok not in out:
+            out.append(tok)
+    return out
+
+
+def working_name_tokens(path: str) -> set[str]:
+    """The identity-bearing tokens in one image's filename.
+
+    Returns an EMPTY set for a name that carries no working title — a
+    screenshot, a bare number, a stopword. Empty means "no evidence", which the
+    caller must treat as silence rather than as a weak match; see the module
+    docstring for the false positive that rule exists for.
+    """
+    return set(_ordered_tokens(path))
+
+
+def leading_name(path: str) -> str | None:
+    """The FIRST identity token of a filename — the piece, not its decoration.
+
+    Creators lead with what the piece is and trail with what this export of it
+    is: `Year_20k_wip1`, `not_sombra_21-cumpeen`, `Tentacooler_c_ins`. Content
+    words sit at the tail, and they span too FEW posts for any frequency cap
+    to catch — `nude`, `cum` and `top` are on three of artist 8's posts each.
+    Position is what separates them from a name.
+
+    Measured on artist 8, Discord messages 2-60 days apart sharing a gated
+    token: 121 pairs. The 106 sharing the leading name all read as one piece's
+    trickle; of the 15 sharing only a trailing word, 14 are sibling pieces
+    (`Bea_Machamp_Shiny_*` / `Bea_Machoke_Shiny_*`) and one is a plain
+    collision (`Undyne_insert_bottom_only-C` / `Lichgalclc_Lingerie_Bottom_21`).
+
+    None when the name carries no identity at all — see `working_name_tokens`.
+    """
+    tokens = _ordered_tokens(path)
+    return tokens[0] if tokens else None
+
+
+def token_frequencies(posts: Iterable[Iterable[str]]) -> Counter[str]:
+    """How many of ONE ARTIST's POSTS each working-name token appears in.
+
+    Takes posts — each an iterable of that post's image paths — rather than a
+    flat list of paths, because the unit of the count is the post. See
+    MAX_TOKEN_POSTS for what that buys; the short version is that a piece with
+    six exports in one post has used its name once.
+
+    Scoped to the ARTIST, not the library: a working name belongs to the person
+    who chose it, and the same string can be one creator's piece and another's
+    boilerplate. Built once per artist per sweep, not per candidate pair.
+    """
+    counts: Counter[str] = Counter()
+    for paths in posts:
+        counts.update({t for path in paths for t in working_name_tokens(path)})
+    return counts
+
+
+def rarity(freq: int, max_frequency: int) -> float:
+    """Rarity of one token within an artist's own corpus, in [0, 1].
+
+    Shared by EVERY rarity-gated signal deliberately. They carried one each
+    until 2026-09-24, and the copies drifted: the filename signal grew a
+    frequency gate and the marker signal never did, so a creator's habitual
+    emoji scored the same 1.00 as a marker they had used twice. One
+    definition cannot drift from itself.
+
+    Full strength at 2 rather than 1: a genuine match means the token is on
+    at least two things, so demanding uniqueness would reject every real
+    pair. Decays to zero AT the cap rather than falling off it, so nothing
+    sits on a cliff edge.
+    """
+    if freq <= 2:
+        return 1.0
+    if freq >= max_frequency:
+        return 0.0
+    return (max_frequency - freq) / (max_frequency - 2)
+
+
+def shared_identity(
+    left: Iterable[str],
+    right: Iterable[str],
+    frequencies: Counter[str],
+    *,
+    max_frequency: int = MAX_TOKEN_POSTS,
+) -> tuple[float, str | None]:
+    """Strength in [0, 1] that two sets of filenames name the SAME piece.
+
+    Returns `(strength, token)` — the token is carried back so the proposal can
+    say WHY it was made. A review queue that cannot explain itself is one the
+    operator learns to click through without reading.
+
+    Strength is a function of the winning token's rarity within the artist's
+    own library, not of how many tokens matched. One decisive token beats three
+    vague ones, and a token appearing across forty of this artist's posts is a
+    habit rather than an identity however exactly it matches.
+
+    `frequencies` must be the POST counts from `token_frequencies`.
+    """
+    shared = {t for t in set(left) & set(right) if frequencies.get(t, 0) <= max_frequency}
+    if not shared:
+        return 0.0, None
+
+    # The rarest shared token decides — one decisive token beats three vague
+    # ones.
+    token = min(shared, key=lambda t: (frequencies.get(t, 0), -len(t), t))
+    strength = round(rarity(max(frequencies.get(token, 1), 1), max_frequency), 4)
+    # A token sitting exactly ON the cap decays to zero, and naming it anyway
+    # would hand the review queue a reason that carries no weight — "matched on
+    # loislanetb2", with nothing behind it. Measured: that token is on 6 of this
+    # artist's images. Report a token only when it is doing work.
+    return (strength, token) if strength > 0 else (0.0, None)
+
+
+# --- the body/title signal ---------------------------------------------------
+#
+# The same idea applied to TEXT. The operator's example pair carries `🍈🍈` in
+# the Patreon title and `@everyone 🍈 🍈` in the Discord message — a marker the
+# creator uses to tie the two together, which no vocabulary list would predict.
+#
+# Rarity-gated, exactly as the filename signal is, and the gate is here because
+# the first pass did NOT have one. Measured on artist 8, 300 posts:
+#
+#     💦 11 posts (4%)   🫴 6   🌰 5   🍗 5   🫣 4
+#
+# 💦 is punctuation for this creator — about one post in twenty-five. Ungated it
+# scored a full 1.00 and was the DECIDING term in a proposal that proximity
+# alone (0.441) could not carry. A habitual marker riding along with proximity
+# is just proximity wearing a hat, which is the exact failure the matcher's
+# threshold sits above 0.55 to prevent. The operator's 🍈🍈 is the opposite
+# case: two posts, and they are the pair itself.
+
+_WORD = re.compile(r"[0-9A-Za-z]{3,}")
+# Anything outside the Basic Multilingual Plane's text ranges: emoji, symbols,
+# kaomoji parts. These are the tokens creators actually use as markers, and
+# they are rare enough in prose to be evidence on their own.
+# U+1F000-1FAFF is the emoji planes; U+2190-2BFF covers arrows, dingbats and
+# the miscellaneous-symbol blocks, which already contains U+2600-27BF.
+_SYMBOL = re.compile(r"[\U0001F000-\U0001FAFF\u2190-\u2BFF]")
+
+_COMMON_TEXT = frozenset({
+    "the", "and", "for", "you", "new", "out", "now", "this", "that", "with",
+    "everyone", "here", "post", "all", "art", "one", "get", "has", "are",
+})
+
+# A marker in more than this many of ONE ARTIST's posts is a signature, not a
+# tie-back. A marker tying an announcement to its drop lands on two posts —
+# the two.
+#
+# Tighter than MAX_TOKEN_POSTS. Both count posts, so the numbers are directly
+# comparable and the gap between them is the claim being made: a working name
+# is the creator's private label for one piece and may honestly recur as they
+# revisit it, while a marker is public decoration and stops being evidence the
+# moment it is reused. Measured on artist 8: 💦 spans 13 posts, the word
+# "like" 43, and 🌗 — a real tie-back — exactly 2.
+MAX_MARKER_POSTS = 4
+
+
+def text_markers(text: str | None) -> set[str]:
+    """Distinctive tokens in a post body or title: symbols, and rare-ish words.
+
+    Symbols count individually rather than as a run, so `🍈🍈` and `🍈 🍈` —
+    which is how the same marker appears on the two platforms — reduce to the
+    same token. Spacing is a platform's rendering, not the creator's intent.
+    """
+    if not text:
+        return set()
+    out = {m.group(0) for m in _SYMBOL.finditer(text)}
+    out |= {
+        w.lower() for w in _WORD.findall(text)
+        if w.lower() not in _COMMON_TEXT and not _YEAR.match(w)
+    }
+    return out
+
+
+def marker_frequencies(texts: Iterable[str | None]) -> Counter[str]:
+    """How many of ONE ARTIST's posts each marker appears in.
+
+    Per POST, not per occurrence: a creator who repeats an emoji six times in
+    one body has used it once as far as identity goes. Scoped to the artist for
+    the same reason `token_frequencies` is — a marker is a personal habit, and
+    one creator's signature is another's whole vocabulary.
+    """
+    counts: Counter[str] = Counter()
+    for t in texts:
+        counts.update(text_markers(t))
+    return counts
+
+
+def marker_overlap(
+    left: str | None,
+    right: str | None,
+    frequencies: Counter[str],
+    *,
+    max_frequency: int = MAX_MARKER_POSTS,
+) -> float:
+    """Strength in [0, 1] that two texts share a DELIBERATE marker.
+
+    `frequencies` is required rather than defaulted to "no gate". An ungated
+    call is the bug this signature exists to make impossible to write by
+    accident, and a default would have kept it one keyword away.
+
+    Symbols weigh full and words a quarter, because prose shares words by
+    accident: a creator who writes "commission" in both posts on a Tuesday has
+    told us nothing that the timestamps did not already say.
+
+    There is no divisor. An earlier pass halved the total so that a long body
+    could not out-vote a short one, which the rarity gate now does properly —
+    and halving meant the operator's own 🍈🍈 pair, a marker on exactly two
+    posts, could reach only 0.5. One marker the creator uses nowhere else is
+    the whole signal, not half of it.
+    """
+    shared = text_markers(left) & text_markers(right)
+    if not shared:
+        return 0.0
+    score = sum(
+        (1.0 if _SYMBOL.match(t) else 0.25) * rarity(frequencies.get(t, 1), max_frequency)
+        for t in shared
+    )
+    return round(min(1.0, score), 4)

backend/app/services/post_unification.py

@@ -0,0 +1,417 @@
+"""The unified post card — a teaser shows what it points at (#4402, #4401).
+
+Milestone 388. A Patreon teaser is a POINTER: a cropped, censored fragment
+whose job is to say "the full set is in Discord". Until this module the card
+rendered the fragment and a text link, and the reader had to make the join FC
+had already made.
+
+Operator, 2026-09-24: *"the teaser from the patreon post doesn't show the items
+that it's supposed to reference so I'm trying to unify the teaser post with the
+content it's meant to draw attention to."*
+
+## A reference, never an absorption
+
+`discord_grouping` folds chat messages into a synthetic post by transferring
+ownership (`absorbed_by_post_id`). That is the wrong primitive here, and the
+operator said so directly: *"the nested items on the unified post are a
+duplicate or reference of existing content. that's why they can show similar
+items and not erase or invalidate the way the discord items landed."*
+
+So nothing here writes. The Discord posts keep their own rows, dates and
+places in the feed; the teaser's card DISPLAYS them. That is also what makes
+reaching back for older variants safe at all: a wrong reference shows one
+extra thumbnail in one place, where a wrong regrouping would move content.
+
+## What a teaser references
+
+1. The Discord drops a `linked` PostAssociation joins it to (#4392) — accepted
+   by the operator, or linked by FC on a conclusive name match.
+2. The rest of that piece's VARIANT FAMILY (#4401): the wips, alts and censor
+   passes a creator trickles out under one working name, days or weeks apart.
+
+Families are found by the creator's LEADING working name, not by any shared
+token and not by image similarity — see `post_naming.leading_name` for the
+measurement, and lesson #4400 for why a whole-image comparison between two
+works by one artist cannot separate "same piece" from "same artist".
+"""
+from __future__ import annotations
+
+from collections import Counter
+from collections.abc import Iterable
+from dataclasses import dataclass
+from datetime import datetime, timedelta
+
+from sqlalchemy import and_, exists, extract, func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.orm import aliased
+
+from ..models import (
+    ImageProvenance,
+    ImageRecord,
+    ImportSettings,
+    Post,
+    PostAssociation,
+    Source,
+)
+from ..utils.phash import hamming, hash_bits
+from ..utils.text import html_to_plain, truncate_at_word
+from .discord_grouping import PLATFORM as DISCORD
+from .gallery_service import thumbnail_url
+from .post_association_service import DUPLICATE_MAX_DISTANCE
+from .post_naming import FAMILY_MAX_POSTS, leading_name, rarity, token_frequencies
+
+# The text each referenced post contributes to the card, per post. The card
+# clamps it again; this keeps a long Discord thread from making the feed
+# payload the size of the thread.
+TEXT_LIMIT = 280
+
+
+@dataclass(frozen=True)
+class Candidate:
+    """One image as the family search sees it."""
+
+    image_id: int
+    post_id: int
+    path: str
+    phash: int | None
+    at: datetime
+
+
+def family(
+    seed: Iterable[Candidate],
+    pool: Iterable[Candidate],
+    name_posts: Counter[str],
+    hash_posts: Counter[int],
+    *,
+    anchor: datetime,
+    window: timedelta,
+    max_posts: int = FAMILY_MAX_POSTS,
+) -> list[Candidate]:
+    """The images in `pool` that belong to the same piece as `seed`.
+
+    A member shares a seed image's LEADING working name, or is a perceptual
+    near-duplicate of one (the same file re-posted), and lies within `window`
+    of `anchor` — the teaser's date. Oldest first, so the card reads as the
+    trickle it was.
+
+    ONE hop from the seed, never transitive. Every measured family is one hop
+    from any of its members, because the members share the name; chaining is
+    what lets a family drift from `Year_20k` to whatever `Year_20k_Base`'s
+    other tokens happen to touch.
+
+    Both identity routes are rarity-gated against `max_posts`, exactly as the
+    matcher gates them. A leading name the creator uses across many posts is a
+    character, and a hash on many posts is a banner.
+    """
+    seed = list(seed)
+    names = {
+        name for c in seed
+        if (name := leading_name(c.path)) is not None
+        and rarity(name_posts.get(name, 0), max_posts) > 0
+    }
+    hashes = [
+        c.phash for c in seed
+        if c.phash is not None and rarity(hash_posts.get(c.phash, 0), max_posts) > 0
+    ]
+    taken = {c.image_id for c in seed}
+
+    out: list[Candidate] = []
+    for c in pool:
+        if c.image_id in taken or abs(c.at - anchor) > window:
+            continue
+        named = leading_name(c.path) in names
+        copied = c.phash is not None and any(
+            (d := hamming(c.phash, h)) is not None and d <= DUPLICATE_MAX_DISTANCE
+            for h in hashes
+        )
+        if named or copied:
+            out.append(c)
+            taken.add(c.image_id)
+    return sorted(out, key=lambda c: (c.at, c.image_id))
+
+
+def _when(post: Post) -> datetime:
+    return post.post_date or post.downloaded_at
+
+
+def _text(post: Post) -> str | None:
+    plain = html_to_plain(post.description) if post.description else None
+    if not plain or not plain.strip():
+        return None
+    return truncate_at_word(plain.strip(), TEXT_LIMIT)[0]
+
+
+@dataclass
+class _Artist:
+    """Everything the family search needs about one artist, loaded once."""
+
+    rows: dict[int, tuple]          # image_id -> (post_id, path, phash, sha, mime, thumb)
+    posts: dict[int, Post]
+    platform: dict[int, str | None]  # post_id -> platform
+    name_posts: Counter[str]
+    hash_posts: Counter[int]
+
+
+class PostUnificationService:
+    def __init__(self, session: AsyncSession):
+        self.session = session
+        self._artists: dict[int, _Artist] = {}
+
+    async def _artist(self, artist_id: int) -> _Artist:
+        if artist_id in self._artists:
+            return self._artists[artist_id]
+
+        posts: dict[int, Post] = {}
+        platform: dict[int, str | None] = {}
+        for post, plat in (await self.session.execute(
+            select(Post, Source.platform)
+            .outerjoin(Source, Post.source_id == Source.id)
+            .where(Post.artist_id == artist_id)
+        )).all():
+            posts[post.id] = post
+            platform[post.id] = plat
+
+        rows: dict[int, tuple] = {}
+        paths_by_post: dict[int, list[str]] = {}
+        hashes_by_post: dict[int, set[int]] = {}
+        for img_id, post_id, path, phash, sha, mime, thumb in (await self.session.execute(
+            select(
+                ImageRecord.id, ImageRecord.primary_post_id, ImageRecord.path,
+                ImageRecord.phash, ImageRecord.sha256, ImageRecord.mime,
+                ImageRecord.thumbnail_path,
+            ).where(
+                ImageRecord.artist_id == artist_id,
+                ImageRecord.primary_post_id.is_not(None),
+            )
+        )).all():
+            bits = hash_bits(phash)
+            rows[img_id] = (post_id, path, bits, sha, mime, thumb)
+            paths_by_post.setdefault(post_id, []).append(path)
+            if bits is not None:
+                hashes_by_post.setdefault(post_id, set()).add(bits)
+
+        # Counted over EVERY post the artist has, exactly as the matcher's
+        # corpus counts them — a family is judged against the whole library,
+        # not against the slice inside the window, or a character name would
+        # look rare in any quiet month.
+        found = _Artist(
+            rows=rows,
+            posts=posts,
+            platform=platform,
+            name_posts=token_frequencies(paths_by_post.values()),
+            hash_posts=Counter(h for hs in hashes_by_post.values() for h in hs),
+        )
+        self._artists[artist_id] = found
+        return found
+
+    async def _drop_images(self, drop_ids: list[int]) -> dict[int, list[int]]:
+        """drop post id -> its image ids, through provenance as the feed reads them.
+
+        A synthetic drop owns no image outright: its images belong to the
+        member messages, and `discord_grouping` gives the drop a provenance row
+        for each. The primary_post_id arm keeps any image that has one and no
+        row, the same union `PostFeedService._thumbnails_for` takes.
+        """
+        out: dict[int, list[int]] = {pid: [] for pid in drop_ids}
+        if not drop_ids:
+            return out
+        links = (
+            select(
+                ImageProvenance.image_record_id.label("image_id"),
+                ImageProvenance.post_id.label("post_id"),
+            )
+            .where(ImageProvenance.post_id.in_(drop_ids))
+            .union(
+                select(
+                    ImageRecord.id.label("image_id"),
+                    ImageRecord.primary_post_id.label("post_id"),
+                ).where(ImageRecord.primary_post_id.in_(drop_ids))
+            )
+            .subquery()
+        )
+        for img_id, pid in (await self.session.execute(
+            select(links.c.image_id, links.c.post_id).order_by(links.c.image_id)
+        )).all():
+            out[pid].append(img_id)
+        return out
+
+    async def unified_for(self, posts: Iterable[Post]) -> dict[int, dict]:
+        """post id -> the card's reference set, for each post that HAS one.
+
+        Only teasers get one: a post with at least one `linked` association on
+        the announcing side. Every other post is absent from the result, and
+        the card renders exactly as it did before this module existed.
+        """
+        teasers = {p.id: p for p in posts if p.synthesized_by is None}
+        if not teasers:
+            return {}
+        links = (await self.session.execute(
+            select(PostAssociation)
+            .where(
+                PostAssociation.status == "linked",
+                PostAssociation.announcement_post_id.in_(list(teasers)),
+            )
+            .order_by(PostAssociation.id)
+        )).scalars().all()
+        if not links:
+            return {}
+
+        settings = await self.session.get(ImportSettings, 1)
+        window = timedelta(days=float(
+            settings.discord_family_window_days if settings is not None else 60.0
+        ))
+        drop_images = await self._drop_images(
+            sorted({a.payload_post_id for a in links})
+        )
+
+        by_teaser: dict[int, list[PostAssociation]] = {}
+        for a in links:
+            by_teaser.setdefault(a.announcement_post_id, []).append(a)
+
+        out: dict[int, dict] = {}
+        for teaser_id, assocs in by_teaser.items():
+            teaser = teasers[teaser_id]
+            artist = await self._artist(teaser.artist_id)
+            out[teaser_id] = self._compose(teaser, assocs, drop_images, artist, window)
+        return out
+
+    def _compose(
+        self,
+        teaser: Post,
+        assocs: list[PostAssociation],
+        drop_images: dict[int, list[int]],
+        artist: _Artist,
+        window: timedelta,
+    ) -> dict:
+        def candidate(img_id: int) -> Candidate | None:
+            row = artist.rows.get(img_id)
+            if row is None:
+                return None
+            post_id, path, bits, *_ = row
+            post = artist.posts.get(post_id)
+            if post is None:
+                return None
+            return Candidate(img_id, post_id, path, bits, _when(post))
+
+        drop_ids = [a.payload_post_id for a in assocs]
+        shown = [i for d in drop_ids for i in drop_images.get(d, [])]
+        own = [i for i, row in artist.rows.items() if row[0] == teaser.id]
+        seed = [c for i in own + shown if (c := candidate(i)) is not None]
+
+        # Variants come from Discord only. That is where a creator trickles
+        # them out, it is the corpus the family rule was measured on, and it
+        # keeps one teaser from pulling a DIFFERENT teaser's crop onto its card.
+        pool = [
+            c for i, row in artist.rows.items()
+            if artist.platform.get(row[0]) == DISCORD
+            and (c := candidate(i)) is not None
+        ]
+        variants = family(
+            seed, pool, artist.name_posts, artist.hash_posts,
+            anchor=_when(teaser), window=window,
+        )
+
+        def thumb(img_id: int, post_id: int, role: str) -> dict | None:
+            row = artist.rows.get(img_id)
+            if row is None:
+                return None
+            _pid, _path, _bits, sha, mime, tp = row
+            return {
+                "image_id": img_id,
+                "thumbnail_url": thumbnail_url(tp, sha, mime),
+                "mime": mime,
+                "post_id": post_id,
+                "role": role,
+            }
+
+        own_ids = set(own)
+        thumbnails: list[dict] = []
+        seen: set[int] = set(own_ids)
+        for drop_id in drop_ids:
+            for img_id in drop_images.get(drop_id, []):
+                if img_id in seen:
+                    continue
+                if (t := thumb(img_id, drop_id, "drop")) is not None:
+                    thumbnails.append(t)
+                    seen.add(img_id)
+        for c in variants:
+            if c.image_id in seen:
+                continue
+            if (t := thumb(c.image_id, c.post_id, "variant")) is not None:
+                thumbnails.append(t)
+                seen.add(c.image_id)
+
+        # The text of every item the card unifies — the operator's *"the
+        # unified card should also contain the text for any of the items
+        # unified on it"*. A drop's own description already joins its member
+        # messages, so a variant's text is its MESSAGE, read off the member
+        # post that owns the image. A line said twice (`@everyone 🍈🍈` on
+        # every message of a drop) is shown once.
+        texts: list[dict] = []
+        said: set[str] = set()
+
+        def add_text(post: Post | None, role: str) -> None:
+            if post is None:
+                return
+            text = _text(post)
+            if text is None or text in said:
+                return
+            said.add(text)
+            texts.append({
+                "post_id": post.id,
+                "role": role,
+                "date": _when(post).isoformat(),
+                "text": text,
+            })
+
+        for drop_id in drop_ids:
+            add_text(artist.posts.get(drop_id), "drop")
+        for post_id in dict.fromkeys(c.post_id for c in variants):
+            add_text(artist.posts.get(post_id), "variant")
+
+        return {
+            "links": [
+                {
+                    "association_id": a.id,
+                    "post_id": a.payload_post_id,
+                    # "fc" | "operator" | None (linked before the column
+                    # existed — an operator accept, every one of them).
+                    "linked_by": a.linked_by,
+                    "token": (a.signals or {}).get("identity_token"),
+                }
+                for a in assocs
+            ],
+            "thumbnails": thumbnails,
+            "variant_count": sum(1 for t in thumbnails if t["role"] == "variant"),
+            "texts": texts,
+        }
+
+
+def fold_clause(fold_hours: float):
+    """WHERE clause: this post is NOT a linked drop sitting beside its teaser.
+
+    Operator: *"discord 'posts' land as normal and only hidden from the post
+    view they're posted the same day."* Everything else stays — an older
+    variant the teaser also references is history, and a reference does not
+    remove it from history.
+
+    Built on `post_date`/`downloaded_at`, not the feed's `resurfaced_at`-led
+    sort key: whether two posts are the same release is a question about when
+    they were published, not about where the feed has since moved one.
+    """
+    teaser = aliased(Post)
+    # The SQL-standard EXTRACT(epoch FROM …), which every Postgres accepts.
+    gap = func.abs(extract(
+        "epoch",
+        func.coalesce(Post.post_date, Post.downloaded_at)
+        - func.coalesce(teaser.post_date, teaser.downloaded_at),
+    ))
+    return ~exists(
+        select(PostAssociation.id)
+        .join(teaser, teaser.id == PostAssociation.announcement_post_id)
+        .where(and_(
+            PostAssociation.payload_post_id == Post.id,
+            PostAssociation.status == "linked",
+            gap <= fold_hours * 3600,
+        ))
+    )

backend/app/services/scheduler_service.py

@@ -115,6 +115,32 @@ async def active_platform_cooldowns(session: AsyncSession) -> dict[str, datetime
     return active
 
 
+def backfill_ready(source: Source) -> bool:
+    """A deep walk the operator started, with budget left and no failure
+    backing it off — due NOW rather than at its next scheduled check.
+
+    A backfill runs one time-boxed chunk per download (plan #693), and nothing
+    queued the next chunk: each waited for the source's regular interval. At
+    the 8-hour default a freshly armed backfill sat untouched until the next
+    check (the operator armed one on 2026-09-25 and saw nothing happen) and a
+    five-chunk walk took most of two days. The tick's in-flight guard keeps
+    one chunk at a time per source and the platform lock one walk per
+    platform, so "due every tick" means "next chunk as soon as the last one
+    ends".
+
+    The failure gate is what keeps a broken source from retrying every
+    minute: any failed chunk raises `consecutive_failures`, which drops the
+    source back onto its backed-off interval. A chunk that fails to progress
+    twice marks the walk stalled (download_service), which ends it here too.
+    """
+    co = source.config_overrides or {}
+    return (
+        co.get("_backfill_state") == "running"
+        and (source.backfill_runs_remaining or 0) > 0
+        and not (source.consecutive_failures or 0)
+    )
+
+
 async def select_due_sources(session: AsyncSession) -> list[Source]:
     """Sources where (enabled, artist.auto_check) and now >= last_checked_at + effective_interval.
 
@@ -123,6 +149,9 @@ async def select_due_sources(session: AsyncSession) -> list[Source]:
     cooldown is the preventive half of the burst-prevention pair (per-source
     consecutive_failures backoff handles the offending source itself).
 
+    A running backfill (`backfill_ready`) is due on every tick, and whether
+    or not its artist is on auto-check — the operator started it by hand.
+
     Ordering: last_checked_at ASC NULLS FIRST, then id. Never-checked
     sources go first, then the longest-since-checked, so the most overdue
     sources hit Celery's FIFO download queue first. Anti-starvation: if
@@ -135,7 +164,6 @@ async def select_due_sources(session: AsyncSession) -> list[Source]:
         .options(selectinload(Source.artist))
         .join(Artist, Source.artist_id == Artist.id)
         .where(Source.enabled.is_(True))
-        .where(Artist.auto_check.is_(True))
         .order_by(Source.last_checked_at.asc().nulls_first(), Source.id)
     )).scalars().all()
 
@@ -147,6 +175,11 @@ async def select_due_sources(session: AsyncSession) -> list[Source]:
     for s in rows:
         if s.platform in cooldowns:
             continue
+        if backfill_ready(s):
+            due.append(s)
+            continue
+        if not s.artist.auto_check:
+            continue
         interval = compute_effective_interval(s, s.artist, settings)
         if s.last_checked_at is None:
             due.append(s)
@@ -163,6 +196,8 @@ def compute_next_check_at(
     """Return the projected datetime of the next check, or None if never checked."""
     if source.last_checked_at is None:
         return None
+    if backfill_ready(source):
+        return datetime.now(UTC)
     interval = compute_effective_interval(source, artist, settings)
     return source.last_checked_at + timedelta(seconds=interval)
 
@@ -229,7 +264,7 @@ async def scheduler_status(session: AsyncSession) -> dict:
     # links to cannot disagree about what they are counting.
     failing_sources = (await session.execute(
         select(func.count()).select_from(Source)
-        .where(Source.enabled.is_(True), failing_sources_clause())
+        .where(failing_sources_clause())
     )).scalar_one()
     no_access_sources = (await session.execute(
         select(func.count()).select_from(Source)

backend/app/services/service_roster.py

@@ -31,18 +31,20 @@ from __future__ import annotations
 import asyncio
 import logging
 
-from sqlalchemy import func, select
+from sqlalchemy import func
 from sqlalchemy.dialects.postgresql import insert as pg_insert
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from ..models import ServiceSeen
+from .worker_lanes import LANES, lane_for_node
 
 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
+# There is no refresh TTL any more. It existed because the HEALTH REQUEST
+# refreshed the roster, rate-limited to 20s so that a page open in two tabs
+# did not inspect twice as often. `size_worker_lanes` owns the refresh now, on
+# `SWEEP_PERIOD_SECONDS`, so the cadence is a schedule rather than a side
+# effect of someone looking.
 
 # 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
@@ -50,16 +52,44 @@ REFRESH_TTL_SECONDS = 20.0
 # page that exists to explain it.
 INSPECT_TIMEOUT_SECONDS = 2.0
 
+# How many broadcast round trips `_inspect_celery_sync` makes. Named, because
+# the wrapper's budget is derived from it and the two must not drift.
+#
+# `active_queues()` and `active()` are separate broadcasts, and a broadcast
+# with no `destination` cannot know how many replies to expect — so each one
+# waits out its full timeout rather than returning on the last reply. The sync
+# call therefore costs ~2 x INSPECT_TIMEOUT_SECONDS in the ordinary case, not
+# once.
+INSPECT_ROUND_TRIPS = 2
+
+# Slack for the thread handoff. `asyncio.to_thread` hands work to the default
+# executor, and on a loaded web process — the operator's showcase page pulling
+# ninety thumbnails a second — the thread may not even be scheduled inside the
+# budget, let alone finish.
+#
+# This exists because the wrapper used to allow `INSPECT_TIMEOUT_SECONDS * 2`,
+# which LOOKS like a safety factor and is exactly the worst case with nothing
+# left over. Observed on the operator's first consolidated deploy, 2026-09-23:
+# a TimeoutError traceback per refresh while the two inspect calls were
+# working perfectly. A budget equal to the work is a budget that fails under
+# any load at all.
+INSPECT_SLACK_SECONDS = 3.0
+
 # 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
+# DERIVED from `worker_lanes.LANES` (milestone 422 step 1) rather than written
+# out here. It was a hand-kept second copy of the same fact, and it had already
+# drifted: `maintenance_long` is a live lane with four task routes pointing at
+# it and a dedicated worker in the operator's stack, and this map did not know
+# it — so the System tab labelled it `Worker (maintenance_long)`. One list of
+# lanes now names them everywhere.
+#
+# A deployment that slices CELERY_QUEUES differently still falls through to the
+# raw queue list rather than being given a name this code 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",
+    lane.queue_key: lane.display_name for lane in LANES
 }
 
 
@@ -80,12 +110,29 @@ def _inspect_celery_sync() -> dict[tuple[str, ...], dict]:
     from ..celery_app import celery as celery_app
 
     insp = celery_app.control.inspect(timeout=INSPECT_TIMEOUT_SECONDS)
+    # TWO broadcasts, each waiting out its own timeout — see
+    # INSPECT_ROUND_TRIPS, which the caller's budget is derived from. Adding a
+    # third call here without updating that constant puts the wrapper back
+    # under the work it is waiting for.
     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}))
+        # Keyed on the LANE's queue set when the node name identifies one, so
+        # a lane keeps the same roster row whether or not it is consuming.
+        #
+        # Grouping on the ACTIVE queues alone meant a lane at cap 0 — which
+        # cancels its consumers — reported an empty set, landed under the key
+        # `celery:`, and rendered as a phantom row named `Worker ()` while its
+        # real row went stale beside it. Both symptoms on the operator's
+        # screen, 2026-09-23, from this one line.
+        #
+        # Deriving the key from `lane.queue_key` rather than inventing a new
+        # one keeps every existing row: it is the same string the lane already
+        # had while it was running.
+        lane = lane_for_node(hostname)
+        key = lane.queue_key if lane else 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, []))
@@ -94,10 +141,13 @@ def _inspect_celery_sync() -> dict[tuple[str, ...], dict]:
     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.
+def touch_service_stmt(*, key: str, kind: str, display_name: str, details: dict):
+    """The upsert that records a check-in, as a statement.
+
+    Built here rather than inline so the async caller (an agent lease, over
+    the API) and the sync one (the sizing sweep, in a celery task) run the
+    SAME write. Two spellings of one upsert is the kind of duplication that
+    stays correct right up until one of them gains a column.
 
     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
@@ -108,7 +158,7 @@ async def touch_service(
     stmt = pg_insert(ServiceSeen).values(
         key=key, kind=kind, display_name=display_name, details=details,
     )
-    stmt = stmt.on_conflict_do_update(
+    return stmt.on_conflict_do_update(
         index_elements=[ServiceSeen.key],
         set_={
             "kind": stmt.excluded.kind,
@@ -117,7 +167,75 @@ async def touch_service(
             "last_seen_at": func.now(),
         },
     )
-    await session.execute(stmt)
+
+
+async def touch_service(
+    session: AsyncSession, *, key: str, kind: str, display_name: str, details: dict
+) -> None:
+    """Record that a part checked in just now."""
+    await session.execute(touch_service_stmt(
+        key=key, kind=kind, display_name=display_name, details=details,
+    ))
+
+
+def _roster_rows(grouped: dict[tuple[str, ...], dict]) -> list[dict]:
+    """The `touch_service` arguments for everything that answered.
+
+    Split from the write so the async and sync refreshes below share the
+    mapping as well as the statement — what a roster row IS should not depend
+    on which kind of session is writing it.
+    """
+    return [
+        {
+            "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"],
+            },
+        }
+        for queues, entry in grouped.items()
+    ]
+
+
+def refresh_celery_roster_sync(session) -> None:
+    """The roster refresh, from the sizing sweep's sync session.
+
+    ## Why the sweep owns this now
+
+    It used to run on the request path, rate-limited to once every 20s by the
+    newest celery row. So the roster only advanced while somebody had a
+    browser open — the liveness of the workers was a function of whether
+    anyone was looking at them, which is the observer-effect version of the
+    bug this roster exists to prevent.
+
+    Operator, 2026-09-23: *"there is a repull every time this page loads — is
+    there a reason this info isn't being tracked in the background and stored
+    in some way?"*
+
+    Now a timer writes it and the page only reads. The cadence is
+    `SWEEP_PERIOD_SECONDS`, and `api/system_health` asserts it leaves headroom
+    under the staleness thresholds — because a sweep period and a stale
+    threshold chosen in different files and never compared is exactly how the
+    idle GPU agent came to read as stopped (lesson #4355).
+
+    Never raises. A failure means the roster does not advance, and the rows
+    going stale is then a TRUE report about a broker nobody can reach.
+    """
+    try:
+        grouped = _inspect_celery_sync()
+    except Exception:
+        log.warning(
+            "service roster: celery inspect failed; roster not refreshed",
+            exc_info=True,
+        )
+        return
+    for row in _roster_rows(grouped):
+        session.execute(touch_service_stmt(**row))
+    session.commit()
 
 
 async def refresh_celery_roster(session: AsyncSession) -> None:
@@ -131,45 +249,14 @@ async def refresh_celery_roster(session: AsyncSession) -> None:
     try:
         grouped = await asyncio.wait_for(
             asyncio.to_thread(_inspect_celery_sync),
-            timeout=INSPECT_TIMEOUT_SECONDS * 2,
+            timeout=(
+                INSPECT_TIMEOUT_SECONDS * INSPECT_ROUND_TRIPS
+                + INSPECT_SLACK_SECONDS
+            ),
         )
     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)
+    for row in _roster_rows(grouped):
+        await touch_service(session, **row)

backend/app/services/source_service.py

@@ -18,7 +18,9 @@ from ..models import (
     Source,
 )
 from .db_helpers import failing_sources_clause
+from .download_backends import uses_native_ingester
 from .gallery_dl import ErrorType
+from .membership_reconcile import KEPT_KEY, STOPPED_KEY
 from .membership_roster import gated_reasons_for_sources
 from .platforms import known_platform_keys
 from .scheduler_service import compute_next_check_at
@@ -124,6 +126,11 @@ class SourceRecord:
             "backfill_posts": self.backfill_posts,
             "tier_gated_count": self.tier_gated_count,
             "gated_reason": self.gated_reason,
+            # Recover / recapture exist only on the native ingester. Sent so the
+            # UI asks the backend's own predicate instead of keeping a copy of
+            # the platform list — the copy said "patreon, subscribestar" for a
+            # day after Discord went native (milestone 428).
+            "native_ingester": uses_native_ingester(self.platform),
         }
 
 
@@ -158,6 +165,38 @@ def _is_app_managed(key: str) -> bool:
 BACKFILL_MAX_CHUNKS = 200
 
 
+def arm_backfill(source: Source) -> None:
+    """Arm a fresh run-until-done backfill on `source` (plan #693). Mutation
+    only — the caller commits. Shared by `SourceService.start_backfill` and the
+    sync Discord repair, so both clear exactly the same resume state."""
+    co = dict(source.config_overrides or {})
+    co["_backfill_state"] = "running"
+    for k in ("_backfill_cursor", "_backfill_cursor_stalls", "_backfill_chunks",
+              "_backfill_posts"):
+        co.pop(k, None)
+    source.config_overrides = co
+    source.backfill_runs_remaining = BACKFILL_MAX_CHUNKS
+
+
+def _record_manual_enable_choice(source: Source, *, enabled: bool) -> None:
+    """Keep the membership sweep (#3995) from overriding the operator.
+
+    Turning a source the sweep STOPPED back on is a deliberate choice to keep
+    pulling a lapsed creator, so it is marked kept and the next sweep leaves it
+    alone. Turning a source off by hand drops any sweep marker, so the sweep
+    never switches back on something the operator switched off themselves.
+    """
+    co = dict(source.config_overrides or {})
+    if enabled and STOPPED_KEY in co:
+        co.pop(STOPPED_KEY)
+        co[KEPT_KEY] = True
+    elif not enabled:
+        co.pop(STOPPED_KEY, None)
+    else:
+        return
+    source.config_overrides = co
+
+
 class SourceService:
     def __init__(self, session: AsyncSession):
         self.session = session
@@ -415,6 +454,9 @@ class SourceService:
         for key, value in fields.items():
             setattr(source, key, value)
 
+        if "enabled" in fields:
+            _record_manual_enable_choice(source, enabled=bool(fields["enabled"]))
+
         if url_changed:
             # Repointing a source at a different creator makes a cached campaign
             # id WRONG, not merely stale, and `patreon_resolver` consults that
@@ -473,13 +515,7 @@ class SourceService:
         )).scalar_one_or_none()
         if source is None:
             raise LookupError(f"source id={source_id} not found")
-        co = dict(source.config_overrides or {})
-        co["_backfill_state"] = "running"
-        for k in ("_backfill_cursor", "_backfill_cursor_stalls", "_backfill_chunks",
-                  "_backfill_posts"):
-            co.pop(k, None)
-        source.config_overrides = co
-        source.backfill_runs_remaining = BACKFILL_MAX_CHUNKS
+        arm_backfill(source)
         await self.session.commit()
         return await self._row_to_record(source)
 
@@ -521,8 +557,8 @@ class SourceService:
         whole source); the two flags are mutually exclusive, so arming recapture
         clears bypass_seen. Clears prior cursor/chunk/stall state so it walks
         fresh from the top. The flag is cleared on completion (download_service)
-        and on stop. Recapture is Patreon-only (the native ingester's post-record
-        capture); inert elsewhere. The UI gates the action to Patreon sources."""
+        and on stop. Recapture needs the native ingester's post-record capture,
+        so the UI offers it on native sources only (`native_ingester`)."""
         source = (await self.session.execute(
             select(Source).where(Source.id == source_id)
         )).scalar_one_or_none()

backend/app/services/subscribestar_client.py

@@ -316,7 +316,7 @@ _ROSTER_URL = f"{_ROSTER_BASE}/subscriptions"
 # `data-identifier`, the one vocabulary that names a state: the table class
 # inside the cancelled card says `for-unsubscribed_users`, a different word for
 # the same list (note #3989, CORRECTION 1). The identifier is stored verbatim as
-# Membership.status and mapped in membership_roster.MEMBERSHIP_STATUS.
+# Membership.status and mapped in native_ingest_common.MEMBERSHIP_STATUS.
 _ROSTER_ACTIVE = "active_subscriptions"
 _ROSTER_CANCELLED = "cancelled_subscriptions"
 
@@ -742,8 +742,15 @@ class SubscribeStarClient:
 
     @staticmethod
     def post_meta(post: dict) -> dict:
-        """Title + date for the preview sample. Title is synthesized from the body
-        (SubscribeStar has no title field)."""
+        """Title + date. Title is None — SubscribeStar has no title field, and
+        the importer synthesizes one from the body.
+
+        `date` is the contract the core's REVISIT WINDOW reads (2026-09-23):
+        ISO-8601 or None. NAIVE here, because `_parse_ss_datetime` renders a
+        parsed local timestamp with no zone; the core reads a naive date as UTC
+        rather than discarding it, since a date it refuses to read is a post
+        the window can never reach.
+        """
         attrs = post.get("attributes") or {}
         return {"title": None, "date": attrs.get("published_at")}
 

backend/app/services/subscribestar_downloader.py

@@ -184,10 +184,20 @@ class SubscribeStarDownloader(BaseNativeDownloader):
         sidecar_path.write_text(json.dumps(data, indent=2))
         return sidecar_path
 
-    def write_post_record(self, post: dict, artist_slug: str) -> PostRecordOutcome:
+    def write_post_record(
+        self, post: dict, artist_slug: str, *, revisit: bool = False,
+    ) -> PostRecordOutcome:
         """Write the post-first `_post.json` (body/links/metadata) — the sole
         writer of the post record on the native path. SubscribeStar's body is
-        already in the feed HTML, so no detail-fetch is needed."""
+        already in the feed HTML, so no detail-fetch is needed.
+
+        `revisit=True` is the tick re-reading a post it already captured
+        (ingest_core's revisit window). The no-detail-fetch half of that
+        contract is free here — there is no detail endpoint — but the
+        don't-blank-a-stored-body half still applies: a chunk that parsed with
+        no content must not overwrite a body we already have. Same guarantee as
+        the Patreon downloader, for the same reason, so a walk behaves the same
+        on both platforms."""
         attrs = post.get("attributes") or {}
         title = attrs.get("title") if isinstance(attrs.get("title"), str) else None
         post_type = attrs.get("post_type") if isinstance(attrs.get("post_type"), str) else None
@@ -196,6 +206,12 @@ class SubscribeStarDownloader(BaseNativeDownloader):
             return PostRecordOutcome(
                 path=None, post_type=post_type, title=title, body_chars=0,
             )
+        if revisit:
+            body = attrs.get("content")
+            if not (isinstance(body, str) and body.strip()):
+                return PostRecordOutcome(
+                    path=None, post_type=post_type, title=title, body_chars=0,
+                )
         post_dir = self.images_root / artist_slug / "subscribestar" / post_dir_name(post)
         post_dir.mkdir(parents=True, exist_ok=True)
         path = self._write_sidecar_data(post, post_dir / "_post.json")

backend/app/services/worker_control.py

@@ -0,0 +1,996 @@
+"""Read and change a lane's live pool, over the broker.
+
+Milestone 422 step 2. The half of the milestone that does something.
+
+## No docker socket is involved, and that is the point
+
+Milestone 365 put "acting on the state" out of scope because restarting a
+dead worker needs a docker socket the web container deliberately does not
+have. That is true of RESTARTING a container. It is not true of changing how
+much work a RUNNING worker does: celery's remote control sends a message over
+the broker and the worker resizes its own pool. Same Redis the app already
+uses, no new privilege, no new surface.
+
+    pool_grow / pool_shrink   how many slots a lane runs
+    add_consumer / cancel_consumer   whether it consumes its queues at all
+
+The operator ruled the socket out independently (2026-09-22: *"this feature
+is a very invasive idea in my mind and I'd like to avoid it"*), and nothing
+here raises the question.
+
+## The setting is PER PROCESS, not per lane total
+
+`pool_grow(n, destination=[...])` adds n slots to EACH destination it names.
+While the stack still runs several containers per lane — the operator's
+production `worker` is `replicas: 2` — a single delta applied to a lane's
+total would be wrong for every replica.
+
+So `slots` means what `CELERY_CONCURRENCY` means: the pool size of one
+process. The reconcile below drives EACH replica to that number
+independently, computing its own delta from that replica's current pool, so
+replicas that have drifted apart (one restarted, one was grown) converge
+rather than being moved in lockstep from a shared baseline.
+
+After step 5 there is one process per lane and the distinction disappears.
+It matters now, and getting it wrong now would be invisible — the totals
+would simply be double what the UI claimed.
+
+## Why reserved() is read alongside the queue depth
+
+Celery PREFETCHES: a worker pulls more messages than it can run and holds
+them in memory. Those have already left the Redis list, so `LLEN` — which is
+what `/api/system/activity/queues` reports — can read 0 while thirty tasks
+are waiting inside a worker. Any judgement about backlog that uses only LLEN
+under-reports, which matters for the UI and is disqualifying for step 7's
+autoscaler.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+from dataclasses import dataclass, field
+from datetime import UTC, datetime
+
+from sqlalchemy import func, select
+from sqlalchemy.dialects.postgresql import insert as pg_insert
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from ..models import TaskRun, WorkerLane, WorkerLaneSample
+from .worker_lanes import (
+    LANES,
+    LANES_BY_QUEUE_KEY,
+    MIN_POOL_SLOTS,
+    Lane,
+    derived_ceiling,
+    lane_for_node,
+)
+
+log = logging.getLogger(__name__)
+
+# celery control is a broker round trip on a request path, so it gets a
+# deadline (rule 156) — the same reasoning and the same budget as
+# service_roster's inspect. A broker that stopped answering must make this
+# report "not present", which is true, rather than hang the page.
+CONTROL_TIMEOUT_SECONDS = 2.0
+
+
+# The WORST case of `inspect_lanes_sync`, for callers that need a deadline.
+#
+# One broadcast plus three targeted reads. The targeted three normally return
+# as soon as the named nodes answer; each can still cost a full timeout if a
+# node disappears mid-read, so the bound stays four.
+CONTROL_ROUND_TRIPS = 4
+
+# Slack for the `asyncio.to_thread` handoff. A budget equal to the work is a
+# budget that fails under load — the roster carried exactly that bug into the
+# operator's first consolidated deploy and logged a TimeoutError per refresh
+# while the inspect calls underneath were working fine.
+CONTROL_SLACK_SECONDS = 3.0
+
+INSPECT_BUDGET_SECONDS = (
+    CONTROL_TIMEOUT_SECONDS * CONTROL_ROUND_TRIPS + CONTROL_SLACK_SECONDS
+)
+
+
+@dataclass
+class LaneLiveState:
+    """What `celery inspect` says about one lane right now.
+
+    `present=False` is NOT "zero slots" — it is "nothing answered". A lane
+    whose worker is restarting, or whose broker is unreachable, must read as
+    unknown rather than as stopped: an unswept absence is not a verdict
+    (snippet #3969). The reconcile in step 3 skips an absent lane rather than
+    correcting it, which is only safe because this distinction is kept.
+    """
+
+    present: bool = False
+    replicas: int = 0
+    active: int = 0
+    reserved: int = 0
+    hostnames: list[str] = field(default_factory=list)
+    # The queues this lane is actually consuming right now, across replicas.
+    # Distinct from the lane's CONFIGURED queues: `cancel_consumer` stops a
+    # worker consuming one without changing what it was started with, which
+    # is how `enabled=false` is implemented. The reconcile needs this to tell
+    # "already disabled" from "needs disabling" — without it, it would re-send
+    # add_consumer for every queue on every tick forever (lesson #4183).
+    consuming: set[str] = field(default_factory=set)
+    # Pool size PER HOSTNAME, not aggregated. The resize below computes each
+    # replica's own delta from its own current pool, so replicas that have
+    # drifted apart converge instead of being moved in lockstep from a shared
+    # baseline — which is what an aggregate here would silently reintroduce.
+    pools: dict[str, int] = field(default_factory=dict)
+
+    @property
+    def pool(self) -> int | None:
+        """One number for the UI. `max` rather than a sum: `slots` means the
+        pool size of ONE process (see the module docstring), so the largest
+        replica is the honest answer to "what is this lane set to". None when
+        no replica reported — unknown, never zero."""
+        return max(self.pools.values()) if self.pools else None
+
+    @property
+    def capacity(self) -> int:
+        """Total slots across replicas — how many tasks this lane can run at
+        once. Distinct from `pool`, and the two must not be confused: `pool`
+        is the DIAL (one process's size, what grow/shrink move), `capacity` is
+        the CAPABILITY. Asking "is this lane saturated" compares `active`,
+        which is summed across replicas, against this — against `pool` it
+        would call two half-busy replicas of 4 saturated at 4 active."""
+        return sum(self.pools.values())
+
+
+def _lane_for_queues(queues: tuple[str, ...]) -> Lane | None:
+    return LANES_BY_QUEUE_KEY.get(tuple(sorted(queues)))
+
+
+def inspect_lanes_sync() -> dict[str, LaneLiveState]:
+    """Live state per lane name. Sync — callers wrap in asyncio.to_thread.
+
+    Never raises. Every lane is present in the result; ones nothing answered
+    for carry `present=False`, so a caller cannot accidentally read a missing
+    lane as an empty one by iterating only what came back.
+    """
+    out = {lane.name: LaneLiveState() for lane in LANES}
+    try:
+        from ..celery_app import celery as celery_app
+
+        # ONE broadcast, then three TARGETED reads.
+        #
+        # A broadcast with no `destination` cannot know how many replies to
+        # expect, so it waits out its whole timeout rather than returning on
+        # the last one. Four of those is four full timeouts — about eight
+        # seconds — and `lane_view` sits on the Settings card, so that was the
+        # load time of the Worker lanes page every time it was opened.
+        #
+        # Naming the destinations lets celery stop as soon as those nodes have
+        # answered, which for workers in this same container is milliseconds.
+        # The worst case is unchanged: a node that vanishes between the
+        # broadcast and the targeted reads costs a full timeout waiting for a
+        # reply that is not coming.
+        insp = celery_app.control.inspect(timeout=CONTROL_TIMEOUT_SECONDS)
+        active_queues = insp.active_queues() or {}
+
+        # Nothing answered — and the three reads below exist only to describe
+        # what did. Returning here also makes the broker-down case FAST
+        # (one timeout, not four), which is exactly when the healthcheck and
+        # the card need an answer rather than a long wait.
+        if not active_queues:
+            return out
+
+        targeted = celery_app.control.inspect(
+            destination=sorted(active_queues),
+            timeout=CONTROL_TIMEOUT_SECONDS,
+        )
+        stats = targeted.stats() or {}
+        active = targeted.active() or {}
+        reserved = targeted.reserved() or {}
+    except Exception:
+        log.warning("worker_control: celery inspect failed", exc_info=True)
+        return out
+
+    for hostname, queues in active_queues.items():
+        # The NODE NAME first — see `lane_for_node`. A lane at cap 0 has its
+        # consumers cancelled and answers with an empty queue list, which
+        # matches no lane, so attributing by queues alone dropped every lane
+        # the operator had turned off and reported it as "not answering".
+        lane = lane_for_node(hostname) or _lane_for_queues(
+            tuple(q["name"] for q in queues)
+        )
+        if lane is None:
+            # A deployment slicing CELERY_QUEUES differently. Reported by the
+            # roster under its raw queue list; it simply has no lane row to
+            # control, which is honest rather than an error.
+            continue
+        state = out[lane.name]
+        state.present = True
+        state.replicas += 1
+        state.hostnames.append(hostname)
+        state.active += len(active.get(hostname, []))
+        state.reserved += len(reserved.get(hostname, []))
+        state.consuming.update(q["name"] for q in queues)
+
+        # Absent on a worker whose stats did not answer, which leaves
+        # pool=None — unknown, not zero.
+        pool = pool_size((stats.get(hostname) or {}).get("pool") or {})
+        if pool is not None:
+            state.pools[hostname] = pool
+
+    for state in out.values():
+        state.hostnames.sort()
+    return out
+
+
+def pool_size(pool_stats: dict) -> int | None:
+    """How many processes a prefork pool is running NOW, from `inspect stats`.
+
+    The length of `processes`, not `max-concurrency`. Celery reports
+    `max-concurrency` as the pool's `limit`, set once at boot; `pool_grow` and
+    `pool_shrink` hand straight to billiard and never touch it (celery 5.6
+    `concurrency/prefork.py`: `self.grow = P.grow`). So it read 1 forever on a
+    lane booted at 1, and the sizing sweep, computing `target - 1` on every
+    tick, grew a scheduler capped at 2 to six processes while the System tab
+    showed "6 / 1" (2026-09-24) — and could never shrink one, since 1 - 1 is 0.
+
+    `max-concurrency` is the fallback only for a pool that lists no processes
+    (a non-prefork pool), where it is the best number there is.
+    """
+    procs = pool_stats.get("processes")
+    if isinstance(procs, list):
+        return len(procs)
+    limit = pool_stats.get("max-concurrency")
+    return limit if isinstance(limit, int) else None
+
+
+def effective_slots(target: int) -> int:
+    """What a pool can actually be set to. Never below one process.
+
+    Used wherever a target is COMPARED as well as wherever one is sent: a
+    reconcile that compares against the unclamped number sees a difference
+    that no control message can ever close, and re-sends it every tick.
+    """
+    return max(MIN_POOL_SLOTS, target)
+
+
+def set_lane_slots_sync(
+    lane: Lane, target: int, live: LaneLiveState | None = None,
+) -> tuple[bool, str | None]:
+    """Drive every replica of `lane` to `target` slots. Returns (applied, err).
+
+    Per-replica deltas rather than one shared delta: see the module docstring.
+    A replica already at the target is issued nothing at all, which is what
+    makes step 3's periodic reconcile converge instead of re-sending a grow of
+    zero forever (lesson #4183 — an enforcer without a reachable fixed point
+    re-does its own work every tick).
+
+    `applied=False` is not a failure of the SETTING. The caller has already
+    stored the value; this says only that the live push did not land, and the
+    reconcile will carry it when the lane answers again.
+    """
+    target = effective_slots(target)
+    try:
+        from ..celery_app import celery as celery_app
+
+        if live is None:
+            live = inspect_lanes_sync()[lane.name]
+        if not live.present:
+            return False, "lane is not running"
+        if not live.pools:
+            return False, "worker did not report its pool size"
+
+        control = celery_app.control
+        unreported = [h for h in live.hostnames if h not in live.pools]
+        for hostname, current in live.pools.items():
+            delta = target - current
+            if delta > 0:
+                control.pool_grow(delta, destination=[hostname])
+            elif delta < 0:
+                control.pool_shrink(-delta, destination=[hostname])
+        if unreported:
+            # Resized what could be resized, and said which could not. Silence
+            # here would leave a replica running at a size the UI claims it is
+            # not, with nothing anywhere recording the gap.
+            return False, f"no pool size reported by {', '.join(sorted(unreported))}"
+        return True, None
+    except Exception as exc:  # noqa: BLE001 — reported, never raised at a caller
+        log.warning("worker_control: could not resize %s", lane.name, exc_info=True)
+        return False, str(exc)
+
+
+def set_lane_enabled_sync(
+    lane: Lane, enabled: bool, live: LaneLiveState | None = None,
+) -> tuple[bool, str | None]:
+    """Start or stop `lane` consuming its queues, without killing the process.
+
+    `cancel_consumer` rather than a shutdown: a stopped consumer keeps its
+    worker alive and answering `inspect`, so a disabled lane stays visible and
+    can be turned back on. A killed worker would read as absent, which is the
+    same signal as a crash — and the whole point of the roster (#365) is that
+    those two must not look alike.
+    """
+    try:
+        from ..celery_app import celery as celery_app
+
+        if live is None:
+            live = inspect_lanes_sync()[lane.name]
+        if not live.present:
+            return False, "lane is not running"
+        control = celery_app.control
+        for queue in lane.queues:
+            if enabled:
+                control.add_consumer(queue, destination=live.hostnames)
+            else:
+                control.cancel_consumer(queue, destination=live.hostnames)
+        return True, None
+    except Exception as exc:  # noqa: BLE001
+        log.warning(
+            "worker_control: could not %s %s",
+            "enable" if enabled else "disable", lane.name, exc_info=True,
+        )
+        return False, str(exc)
+
+
+# --- the settings half, which is async ----------------------------------------
+#
+# Sync celery control above, async DB below, in one module. Same split
+# `service_roster` already runs (`_inspect_celery_sync` beside `touch_service`)
+# — the boundary is the transport, not the concern, and "control the workers"
+# is one concern.
+
+
+async def _rows_by_name(session: AsyncSession) -> dict[str, WorkerLane]:
+    """Every lane's row, creating any that are missing from its LANES defaults.
+
+    Self-heals rather than depending on a migration having run for a lane
+    added later: alembic 0103 seeded the four that existed on 2026-09-22, and
+    a fifth added to LANES afterwards gets its row the first time anything
+    asks. Without this, a new lane would read as absent and the UI would
+    simply not show it.
+    """
+    rows = {
+        row.name: row
+        for row in (await session.execute(select(WorkerLane))).scalars()
+    }
+    missing = [lane for lane in LANES if lane.name not in rows]
+    for lane in missing:
+        row = WorkerLane(name=lane.name, slots_cap=lane.default_slots_cap)
+        session.add(row)
+        rows[lane.name] = row
+    if missing:
+        await session.commit()
+    return rows
+
+
+@dataclass(frozen=True)
+class LaneSample:
+    """What the sizing sweep last measured about one lane.
+
+    The same fields `LaneLiveState` carries, plus the queue depth and WHEN —
+    because this one is read from a table rather than from the broker, and a
+    reading with no timestamp invites being presented as current.
+
+    `measured_at=None` means no sweep has written this lane yet: a fresh
+    install inside its first period, or a stack whose beat is not running.
+    Distinct from `present=False` (something asked, nothing answered), and the
+    UI says different things about the two.
+    """
+
+    present: bool = False
+    replicas: int = 0
+    pool: int | None = None
+    active: int = 0
+    reserved: int = 0
+    queue_depth: int | None = None
+    measured_at: datetime | None = None
+
+
+def _lane_depth(lane: Lane, depths: dict[str, int | None]) -> int | None:
+    """A lane's backlog across its queues — None when NOTHING answered.
+
+    A queue the broker did not answer for must not be summed as zero: an
+    unknown depth is not an empty one, and reporting a buried lane as idle is
+    the direction that matters.
+    """
+    known = [depths.get(q) for q in lane.queues]
+    if not any(d is not None for d in known):
+        return None
+    return sum(d for d in known if d is not None)
+
+
+def store_lane_samples_sync(session, live: dict[str, LaneLiveState], depths) -> None:
+    """Write what the sweep just measured. SYNC — the celery task owns a sync
+    session, and this is the only place these rows are written.
+
+    Upsert per lane, last writer wins, same shape as `service_roster`'s
+    `touch_service`: two processes sweeping at once is a benign race that
+    needs no coordination, because both are recording what they actually saw.
+
+    A lane that did not answer is STILL written, with `present=False`. Skipping
+    it would leave the previous reading in place and let the page go on showing
+    a pool that is no longer there — the stale row would read as a current one
+    (lesson #4202: the row is the thing that has to change).
+    """
+    now = datetime.now(UTC)
+    for lane in LANES:
+        state = live.get(lane.name) or LaneLiveState()
+        values = {
+            "lane": lane.name,
+            "present": state.present,
+            "replicas": state.replicas,
+            "pool": state.pool,
+            "active": state.active,
+            "reserved": state.reserved,
+            "queue_depth": _lane_depth(lane, depths),
+            "measured_at": now,
+        }
+        stmt = pg_insert(WorkerLaneSample).values(**values)
+        session.execute(stmt.on_conflict_do_update(
+            index_elements=[WorkerLaneSample.lane],
+            set_={k: v for k, v in values.items() if k != "lane"},
+        ))
+    session.commit()
+
+
+@dataclass
+class LaneSettings:
+    """What the DATABASE knows about the lanes — read and finished with before
+    anything touches the broker.
+
+    This exists because holding a Postgres connection across a celery round
+    trip is what made the System tab block the whole site (operator,
+    2026-09-23: *"something about changing the cap number is blocking to the
+    website"*).
+
+    `lane_view` used to take the session and keep it open through an inspect
+    whose budget is eleven seconds — and that page polls every fifteen. With a
+    lane not answering, every inspect ran to nearly its full budget, so each
+    poll pinned a connection for ten seconds. SQLAlchemy's default pool is
+    five connections plus ten overflow; a couple of browser tabs, the health
+    endpoint doing the same thing, and a cap change adding two more inspects
+    exhausts that, and every OTHER request then waits on a connection.
+
+    Nothing was slow in itself. The slowness was a scarce resource held across
+    it, which is why it surfaced as the whole site stalling rather than as one
+    slow page.
+    """
+
+    caps: dict[str, int]
+    oldest_by_queue: dict[str, datetime]
+    # The sizing sweep's last reading per lane. Since 2026-09-23 this is where
+    # the live numbers come from: the endpoint no longer inspects at all.
+    samples: dict[str, LaneSample] = field(default_factory=dict)
+
+
+async def lane_settings(session: AsyncSession) -> LaneSettings:
+    """Every DB read the lane view needs, in one short-lived session.
+
+    Which is now ALL of them. `lane_view` below takes what this returns and
+    talks to nothing.
+    """
+    rows = await _rows_by_name(session)
+    samples = {
+        row.lane: LaneSample(
+            present=row.present,
+            replicas=row.replicas,
+            pool=row.pool,
+            active=row.active,
+            reserved=row.reserved,
+            queue_depth=row.queue_depth,
+            measured_at=row.measured_at,
+        )
+        for row in (
+            await session.execute(select(WorkerLaneSample))
+        ).scalars()
+    }
+    return LaneSettings(
+        caps={name: row.slots_cap for name, row in rows.items()},
+        oldest_by_queue=await _oldest_running_by_queue(session),
+        samples=samples,
+    )
+
+
+def lane_view(settings: LaneSettings) -> list[dict]:
+    """Every lane: what is configured, what was last measured, what it may
+    grow to. NO broker call, and no database — `settings` is the whole input.
+
+    ## It used to inspect, on every request
+
+    Four broadcast round trips on an eleven-second budget, on a page that
+    polls every fifteen seconds. Operator, 2026-09-23: *"there is a repull
+    every time this page loads — is there a reason this info isn't being
+    tracked in the background and stored in some way?"*
+
+    There was one, and it had expired. The docstring here used to say the
+    endpoint was deliberately uncached because *"this is the surface an
+    operator watches while dragging a stepper, and a cached reply would show
+    them the value from before their own change"*. True while a cap change
+    refetched the table — and that refetch is exactly what was removed in
+    `1353d34`, so the UI now patches its own row from the write's reply and
+    nothing depends on this being live.
+
+    Meanwhile `size_worker_lanes` was already inspecting on a timer to decide
+    pool sizes: the same numbers, computed, used, and discarded, while the
+    browser asked the broker for them again four times a minute.
+
+    So the sweep writes `worker_lane_sample` and this reads it. The reading is
+    up to `SWEEP_PERIOD_SECONDS` old, and `measured_at` travels with it so the
+    UI can say so rather than implying it is current.
+
+    `pending` is still the honest backlog — depth PLUS reserved — because
+    celery prefetches and LLEN alone reads 0 while a worker holds tasks in
+    memory.
+    """
+    oldest = settings.oldest_by_queue
+    now = datetime.now(UTC)
+    out = []
+    for lane in LANES:
+        cap = settings.caps[lane.name]
+        # A lane with no row yet is not-measured, which is distinct from
+        # measured-as-absent. The default carries `measured_at=None`, and the
+        # UI says "not measured yet" rather than "not answering".
+        sample = settings.samples.get(lane.name) or LaneSample()
+        depth = sample.queue_depth
+        out.append({
+            "name": lane.name,
+            "display_name": lane.display_name,
+            "queues": list(lane.queues),
+            "slots_cap": cap,
+            "ceiling": derived_ceiling(lane),
+            # DERIVED, never stored. A cap of zero means no consumers, so
+            # "off" and "may use no workers" cannot disagree.
+            "enabled": cap > 0,
+            "memory_bound": lane.memory_bound,
+            "optional": lane.optional,
+            # What raising this lane's cap will download, so the UI can say
+            # WHICH model and how big BEFORE the first slot is asked for
+            # rather than after a multi-GB fetch has started. `measured`
+            # travels with the numbers: the UI must not present an estimate
+            # as a fact.
+            "models": [
+                {
+                    "repo": m.repo,
+                    "download_bytes": m.approx_download_bytes,
+                    "resident_bytes": m.approx_resident_bytes,
+                    "measured": m.measured,
+                }
+                for m in lane.models
+            ],
+            "live": {
+                "present": sample.present,
+                "replicas": sample.replicas,
+                "pool": sample.pool,
+                "active": sample.active,
+                "reserved": sample.reserved,
+            },
+            "queue_depth": depth,
+            "pending": None if depth is None else depth + sample.reserved,
+            # When the numbers above were read. Per lane rather than one for
+            # the response, because a lane whose row has never been written
+            # has no reading at all and must not borrow another lane's.
+            "measured_at": (
+                sample.measured_at.isoformat() if sample.measured_at else None
+            ),
+            # How long the oldest still-running task on this lane has been
+            # going, in minutes. Read from `task_run`, not from the sweep, so
+            # this one IS current. The operator asked for a trigger here —
+            # grow a lane whose tasks run past some duration — and it stayed a
+            # REPORT: a long task does not finish sooner because the lane
+            # gained a slot, so scaling on it would spend memory to change
+            # nothing. Shown so they can see a lane wedged on one slow job,
+            # which is the genuinely useful half of the idea.
+            "oldest_running_minutes": _minutes_since(
+                min(
+                    (oldest[q] for q in lane.queues if q in oldest),
+                    default=None,
+                ),
+                now,
+            ),
+        })
+    return out
+
+
+async def _oldest_running_by_queue(session: AsyncSession) -> dict[str, datetime]:
+    """When the longest-running unfinished task on each queue started.
+
+    Read from `task_run`, which is OUR OWN table on OUR OWN wall clock, and
+    deliberately not from celery's `inspect active()`. Those entries carry a
+    `time_start` taken from the WORKER's `time.monotonic()` — a clock with an
+    arbitrary origin per process. Subtracting it from this process's wall
+    clock produces a number that looks like a duration and is meaningless, and
+    it would be meaningless in the direction that matters: plausible.
+
+    `task_run` also already carries the per-queue staleness thresholds the
+    recovery sweep uses, so a row still `running` here is one the system
+    itself considers legitimately in flight rather than abandoned.
+    """
+    result = await session.execute(
+        select(TaskRun.queue, func.min(TaskRun.started_at))
+        .where(TaskRun.status == "running", TaskRun.finished_at.is_(None))
+        .group_by(TaskRun.queue)
+    )
+    return {queue: started for queue, started in result if started is not None}
+
+
+def _minutes_since(started: datetime | None, now: datetime) -> int | None:
+    """Whole minutes, or None when nothing is running. Never negative: a row
+    written by a container whose clock is a few seconds ahead must read as 0
+    rather than as a task that starts in the future."""
+    if started is None:
+        return None
+    return max(0, int((now - started).total_seconds() // 60))
+
+
+def _queue_depths_sync() -> dict[str, int | None]:
+    """Redis LLEN per queue. None for one that did not answer — see lane_view.
+
+    Sync; the caller threads it. A per-queue try/except so one bad queue does
+    not cost the whole report, matching `api/system_activity._read_queues_sync`.
+    """
+    import redis
+
+    from ..config import get_config
+
+    out: dict[str, int | None] = {}
+    try:
+        client = redis.Redis.from_url(get_config().celery_broker_url)
+    except Exception:
+        log.warning("worker_control: no broker for queue depths", exc_info=True)
+        return {q: None for lane in LANES for q in lane.queues}
+    for lane in LANES:
+        for queue in lane.queues:
+            try:
+                out[queue] = int(client.llen(queue))
+            except Exception:  # noqa: BLE001 — a hiccup must not break the UI
+                out[queue] = None
+    return out
+
+
+class LaneUpdateRefused(ValueError):
+    """A requested value is outside what the lane may hold. Carries the reason
+    the UI shows — a greyed control with no explanation reads as a bug."""
+
+
+async def store_lane_cap(
+    session: AsyncSession, lane: Lane, slots_cap: int,
+) -> int:
+    """Validate and store the cap. Returns the PREVIOUS cap. DB only.
+
+    Split from the live push for the reason `LaneSettings` gives at length: a
+    Postgres connection must not be held across a celery round trip. Everything
+    here is fast and finished with before `push_lane_cap` starts.
+    """
+    rows = await _rows_by_name(session)
+    row = rows[lane.name]
+
+    ceiling = derived_ceiling(lane)
+    if slots_cap < 0:
+        raise LaneUpdateRefused("a cap cannot be negative")
+    if slots_cap > ceiling:
+        raise LaneUpdateRefused(
+            f"a cap of {slots_cap} is above what this container can hold "
+            f"({ceiling} for {lane.display_name})"
+        )
+
+    was_cap = row.slots_cap
+    row.slots_cap = slots_cap
+    await session.commit()
+    # The previous value, because the push needs the DIRECTION: lowering a cap
+    # has to reach the running lane now, and raising one has nothing to say.
+    return was_cap
+
+
+async def push_lane_cap(lane: Lane, slots_cap: int, *, was_cap: int) -> dict:
+    """Make the running lane obey a cap that is already stored. NO database.
+
+    Runs OFF the request path since 2026-09-23 — the endpoint stores the cap,
+    answers, and hands this to a background task (operator: *"the change
+    should be queued so that it isn't blocking of the webui"*). Nothing here
+    changed as a result except who waits for it: the return value is now read
+    by the log rather than by a browser, and every branch below already
+    treated failure as "the sizing pass will carry it".
+
+    ## What is pushed, and what is not
+
+    Consumers follow the cap immediately in BOTH directions: zero means off,
+    and off must take effect when it is asked for rather than up to a minute
+    later.
+
+    The pool is only ever pushed DOWNWARD. Raising a cap is permission, not a
+    request — growing on permission would put workers on a lane with nothing
+    to do — so the sizing pass spends it on its next tick if there is work.
+    That also makes the common case (raising a cap) free: no broker round trip
+    AT ALL, which is the difference between a control that answers instantly
+    and one that takes ten seconds. Keyed on the previous cap rather than on
+    "is it on" — the first cut only knew on/off, so it inspected on every
+    raise to find out whether the pool needed lowering, and the control it was
+    meant to make instant still waited out an inspect.
+
+    A failed push is not a failed setting. The value is already stored and the
+    sizing pass carries it within a minute; `applied: false` with a reason
+    lets the UI say "saved, not yet live" rather than "that didn't work"
+    (lesson #4202 — a live change that does not survive, with nothing saying
+    so).
+    """
+    was_on, now_on = was_cap > 0, slots_cap > 0
+    applied, error = True, None
+    if now_on != was_on:
+        applied, error = await asyncio.to_thread(set_lane_enabled_sync, lane, now_on)
+    if applied and not now_on:
+        # Down to the floor at once. The pool cannot be emptied, so "off" is
+        # one parked process with its consumers cancelled.
+        applied, error = await asyncio.to_thread(
+            set_lane_slots_sync, lane, MIN_POOL_SLOTS,
+        )
+    elif applied and now_on and slots_cap < was_cap:
+        # LOWERED on a running lane. Only this direction needs a message, and
+        # only when the pool is actually above the new cap — so it reads the
+        # live pool rather than resizing blind. A raise never reaches here.
+        #
+        # Bounded (rule 156): `to_thread` on its own is an await with no
+        # deadline, and this runs in a background task where a hang would be
+        # silent rather than visible as a slow page. On a timeout the lane is
+        # simply not resized here and the sizing sweep carries it.
+        try:
+            live = await asyncio.wait_for(
+                asyncio.to_thread(inspect_lanes_sync),
+                timeout=INSPECT_BUDGET_SECONDS,
+            )
+        except TimeoutError:
+            log.warning(
+                "worker_control: inspect exceeded %ss lowering %s; leaving the "
+                "pool to the sizing pass", INSPECT_BUDGET_SECONDS, lane.name,
+            )
+            return _cap_result(lane, slots_cap, now_on, applied, error, False)
+        current = live[lane.name].pool
+        if current is not None and current > slots_cap:
+            applied, error = await asyncio.to_thread(
+                set_lane_slots_sync, lane, slots_cap, live=live[lane.name],
+            )
+
+    # Raising the cap off zero is what triggers the model download (milestone
+    # 422 step 6). Never at boot: that made every start of the ML role reach
+    # HuggingFace for ~3.5GB, and rule 164 permits a runtime fetch only for a
+    # feature that is optional and clearly OFF.
+    #
+    # On the TRANSITION, so re-saving a cap on a lane already running does not
+    # re-enqueue.
+    #
+    # NOT gated on the consumer change having landed, which it was until
+    # 2026-09-23. The reasoning then was that enqueueing onto a queue nothing
+    # consumes leaves the task pending — true, and it is the right place for
+    # it to wait. Gated, a cap raised while the lane was restarting stored the
+    # cap, let the sizing pass start the consumers a minute later, and left
+    # the lane running with no model, because nothing else ever asks for one.
+    # A task parked on the `ml` queue is picked up the moment that happens.
+    fetching = False
+    if now_on and not was_on and lane.models:
+        fetching = _enqueue_model_fetch()
+
+    # Nobody is waiting on this any more, so the log is where a push that did
+    # not land has to be visible. Not an error: the value is stored and the
+    # sizing pass carries it within a minute.
+    if not applied:
+        log.info(
+            "worker_control: %s cap %s stored, not pushed (%s); "
+            "the sizing pass will carry it",
+            lane.name, slots_cap, error,
+        )
+
+    return _cap_result(lane, slots_cap, now_on, applied, error, fetching)
+
+
+def _cap_result(
+    lane: Lane, slots_cap: int, now_on: bool, applied: bool,
+    error: str | None, fetching: bool,
+) -> dict:
+    """The push's outcome. One builder, because `push_lane_cap` has two exits
+    and a second literal would be free to disagree with the first."""
+    return {
+        "name": lane.name,
+        "slots_cap": slots_cap,
+        "ceiling": derived_ceiling(lane),
+        "enabled": now_on,
+        "applied": applied,
+        "apply_error": error,
+        # Tells the UI to say a download has started rather than leaving the
+        # operator to wonder why a lane they just turned on is busy.
+        "fetching_models": fetching,
+    }
+
+
+def _enqueue_model_fetch() -> bool:
+    """Queue the model download. Returns whether it was accepted.
+
+    Import inside the function: `backend.app.tasks.ml` pulls in torch, and web
+    must not pay that import cost on a module that every settings request
+    touches.
+
+    Never raises. A broker that will not take the task is worth reporting, but
+    the SETTING has already been stored and the lane is already enabled — so
+    failing the whole request here would roll back nothing and tell the
+    operator their change did not happen when it did.
+    """
+    try:
+        from ..tasks.ml import ensure_models
+
+        ensure_models.delay()
+        return True
+    except Exception:  # noqa: BLE001 — reported, never raised at a caller
+        log.warning("worker_control: could not enqueue the model fetch", exc_info=True)
+        return False
+
+
+
+# --- the sizing pass: one sweep, always on ------------------------------------
+#
+# This replaced BOTH `reconcile_lanes_sync` (step 3) and `autoscale_lanes_sync`
+# (step 7) on 2026-09-23. They were two enforcers over one number, and the
+# whole of step 7's hardest reasoning — a stored value that is a FLOOR, a
+# target of `max(stored, current)` so the reconcile does not undo what the
+# autoscaler added — existed only to stop them fighting. Delete one of them and
+# the problem is not solved, it is absent.
+#
+# Operator: *"auto should be always on, not a setting, so that idle instances
+# quiet down when not running. the number that is visible and something the
+# user can tweak and manage should be the cap itself the number of running
+# workers is handled by the autoscaling function which is always on."*
+#
+# So there is one pass, it runs every minute, it reads the live pool rather
+# than any stored number, and the only thing it obeys is the cap.
+#
+# It also subsumes what the reconcile existed for. `pool_grow` is not durable:
+# a worker restarted by its supervisor comes back at its ENV concurrency,
+# silently below what the lane should be running. This pass reads the live
+# pool every minute and sizes from the backlog, so that worker is corrected on
+# the next tick — sooner than the five-minute reconcile managed, and without a
+# second sweep that could disagree with this one.
+
+# How much work justifies a slot. `pending` is depth + reserved, so it already
+# counts what celery has prefetched into worker memory — one task, one slot.
+#
+# Growth is IMMEDIATE and shrink is one slot per tick, deliberately asymmetric.
+# A backlog of four thousand should not take an hour to reach the cap, and a
+# lane that idles for one minute should not drop every process it has: the
+# cost of being one slot too large for a minute is a sleeping process, and the
+# cost of being too small is work not happening. For ML the asymmetry matters
+# most — every new slot reloads a multi-GB model, so the slow shrink is what
+# stops a quiet patch from paying that cost again a minute later.
+SHRINK_STEP = 1
+
+
+@dataclass
+class LaneSizing:
+    """What the pass did to one lane, and why — in the operator's terms.
+
+    A reason on every outcome including "held", because a sizing pass that
+    only speaks when it acts is one nobody can debug when it does not.
+    """
+
+    lane: str
+    action: str  # "grew" | "shrank" | "held" | "skipped"
+    slots: int
+    reason: str
+
+
+def wanted_slots(cap: int, active: int, pending: int | None) -> int:
+    """How many workers this lane has work for right now, within its cap.
+
+    One slot per task in flight or waiting, floored at one process and
+    ceilinged by the cap. `pending` of None means the broker did not answer
+    for this lane's queues — an unknown backlog is not an empty one (snippet
+    #3969), so it contributes nothing rather than being read as zero.
+
+    A cap of zero still returns one: billiard cannot run an empty pool, and
+    the parked process is what `add_consumer` lands on when the cap goes back
+    up. "Off" is expressed by cancelling consumers, not by emptying the pool.
+    """
+    if cap <= 0:
+        return MIN_POOL_SLOTS
+    return max(MIN_POOL_SLOTS, min(cap, active + (pending or 0)))
+
+
+def size_lanes_sync(
+    caps: dict[str, int],
+    *,
+    live: dict[str, LaneLiveState] | None = None,
+    depths: dict[str, int | None] | None = None,
+) -> list[LaneSizing]:
+    """Size every lane to its backlog, within the cap. The whole control loop.
+
+    `caps` is lane name -> slots_cap, read from the database by the caller.
+    This function touches no database: the celery task that schedules it owns
+    the session, and keeping the DB out of here is what lets it be called from
+    anywhere that already knows the caps.
+
+    `live` and `depths` are the measurements. Passing them in is not an
+    optimisation — it is how the caller gets to KEEP them. The sweep now
+    stores what it measured (`worker_lane_sample`) so the System tab reads a
+    table instead of inspecting on every page load, and that is only possible
+    if the same reading serves both purposes. Measured here when not given, so
+    every existing caller and test is unaffected.
+
+    ## It must converge and then go quiet
+
+    One `inspect` for all lanes, and `set_lane_slots_sync` issues nothing to a
+    replica already at its target. A settled system therefore performs one
+    broker round trip plus one LLEN sweep per tick and sends no control
+    messages at all — the reachable fixed point lesson #4183 is about. An
+    enforcer that re-sent a grow of zero every tick would churn forever and
+    bury a real correction in its own noise.
+
+    ## An absent lane is SKIPPED, not corrected
+
+    `present=False` means nothing answered — a worker restarting, or an
+    unreachable broker. It does NOT mean zero slots. Deciding from that would
+    be a verdict drawn from an unswept read, and here it is worse than
+    useless: there is nothing to send the message to.
+    """
+    if live is None:
+        live = inspect_lanes_sync()
+    if depths is None:
+        depths = _queue_depths_sync()
+    out: list[LaneSizing] = []
+
+    for lane in LANES:
+        cap = caps.get(lane.name)
+        if cap is None:
+            continue
+        state = live[lane.name]
+        if not state.present:
+            out.append(LaneSizing(lane.name, "skipped", 0, "lane is not answering"))
+            continue
+
+        # Consumers first, and only when they DISAGREE. Sending add_consumer
+        # for every queue on every tick of a settled system is the exact churn
+        # above, and invisible: add_consumer on a queue already consumed is
+        # harmless and reports success.
+        should_consume = cap > 0
+        if should_consume != state.consuming.issuperset(lane.queues):
+            ok, err = set_lane_enabled_sync(lane, should_consume, live=state)
+            if not ok:
+                out.append(LaneSizing(
+                    lane.name, "held", state.pool or 0,
+                    f"could not {'start' if should_consume else 'stop'} "
+                    f"consuming: {err}",
+                ))
+                continue
+
+        current = state.pool
+        if current is None:
+            out.append(LaneSizing(
+                lane.name, "held", 0, "worker did not report its pool size",
+            ))
+            continue
+
+        depth = _lane_depth(lane, depths)
+        pending = None if depth is None else depth + state.reserved
+        want = wanted_slots(cap, state.active, pending)
+
+        if want > current:
+            new = want
+            verb = "grew"
+        elif want < current:
+            # One at a time on the way down. See SHRINK_STEP.
+            new = max(want, current - SHRINK_STEP)
+            verb = "shrank"
+        else:
+            out.append(LaneSizing(
+                lane.name, "held", current,
+                f"{pending if pending is not None else '?'} waiting, "
+                f"{state.active} busy, cap {cap}",
+            ))
+            continue
+
+        ok, err = set_lane_slots_sync(lane, new, live=state)
+        if not ok:
+            out.append(LaneSizing(
+                lane.name, "held", current, f"could not resize: {err}",
+            ))
+            continue
+        out.append(LaneSizing(
+            lane.name, verb, new,
+            f"{pending if pending is not None else '?'} waiting, "
+            f"{state.active} busy, cap {cap}",
+        ))
+    return out

backend/app/services/worker_lanes.py

@@ -0,0 +1,464 @@
+"""The worker lanes: what they are, and how many slots each may be given.
+
+Milestone 422 step 1. This module is the ONE place that knows the lane set;
+`models/worker_lane.py` holds only what the operator can change about them.
+
+## Why the queues are here and not in the table
+
+A lane's queue set is not a preference — it is decided by `celery_app.py`'s
+`task_routes`, which is what puts a backup on `maintenance_long` and a
+thumbnail on `thumbnail`. An operator cannot move a task to another lane, so
+storing the queues as settings would create a row that can disagree with the
+routing table, and nothing would notice until a queue had no consumer.
+
+So: queues and display names are code, slots and caps are data. The table
+stores three numbers and a flag, and nothing that could contradict celery.
+
+This also collapses a duplicate rather than adding one.
+`service_roster.ROLE_NAMES` was a second copy of "queue set -> the name an
+operator recognises", and it had already drifted: `maintenance_long` is a
+live lane with four task routes pointing at it, and the roster did not know
+its name, so the System tab rendered it as `Worker (maintenance_long)`. That
+map is now derived from `LANES` below, so a lane added here is named
+everywhere at once.
+
+## Why the ceiling is derived rather than configured
+
+Consolidating the stack into one container (step 5) widens the OOM blast
+radius: today an ml-worker that exhausts memory is killed by Docker on its
+own, and web keeps serving. In one container the kernel picks a victim from
+the whole cgroup, and it may pick hypercorn — so a tagging task can take the
+UI down with it, on exactly the modest hardware least able to spare the
+memory.
+
+Operator, 2026-09-22: *"ram isn't an issue for me but some users might run
+this on weaker hardware and I don't want it to kill their servers."*
+
+So the maximum is computed from what the container actually has, and the
+operator's own `slots_cap` must fit under it. Three numbers, not two, and the
+ordering is the point:
+
+    slots  <=  slots_cap  <=  derived_ceiling
+    (live)     (operator)     (this module)
+
+The operator can always lower their cap. They cannot raise it past what the
+box can hold. The derived ceiling is never stored — a row that outlived a
+change in container limits must not carry a stale one.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+from dataclasses import dataclass
+from pathlib import Path
+
+log = logging.getLogger(__name__)
+
+# --- the lanes ---------------------------------------------------------------
+
+
+GIB = 1024 ** 3
+
+
+@dataclass(frozen=True)
+class ModelRequirement:
+    """A model a lane must download before it can do anything.
+
+    Surfaced to the UI so the operator is told WHICH model, how big, and what
+    it costs to hold — before they turn the lane on, not after a multi-GB
+    download has already started. The lane is optional and its cost is not
+    obvious from its name, which is the whole reason this is structured data
+    rather than a sentence in a component.
+
+    `measured=False` means the numbers are ESTIMATES and the UI must say so.
+    They come from the checkpoint's parameter count and dtype, not from a
+    build — and a number presented as fact decides whether someone's server
+    survives, so it is labelled rather than rounded confidently.
+    """
+
+    # The Hugging Face repo id, which is the honest answer to "which model".
+    repo: str
+    # Roughly what the download costs, for the operator's bandwidth and disk.
+    approx_download_bytes: int
+    # Roughly what ONE slot holds while running. Prefork forks a child per
+    # slot and each loads its own copy, so this multiplies.
+    approx_resident_bytes: int
+    measured: bool = False
+
+
+# SigLIP so400m — the only model FabledCurator itself downloads.
+#
+# What it is for, which is NOT obvious from the lane's name: it produces the
+# image embeddings that back similarity search, duplicate grouping and the
+# tag heads. WD14 tagging is the GPU AGENT's job, not this lane's — the
+# comment in celery_app.py naming both is stale since B3 (#1238), when the
+# agent took over and this lane was left as the CPU embed fallback for stacks
+# running no agent at all (see MLSettings.cpu_embed_enabled).
+#
+# Both numbers are ESTIMATES, derived from the checkpoint rather than from a
+# build: ~877M parameters at fp32 is ~3.5GB of weights, and holding them plus
+# activations and the torch runtime is what the resident figure covers. They
+# err high. Replace them with measurements — download the repo and read its
+# size; run one embed and read the worker child's VmHWM — and set
+# `measured=True` when you do.
+SIGLIP_MODEL = ModelRequirement(
+    repo="google/siglip-so400m-patch14-384",
+    approx_download_bytes=3_500_000_000,
+    approx_resident_bytes=4 * GIB,
+    measured=False,
+)
+
+
+@dataclass(frozen=True)
+class Lane:
+    """A worker lane. `name` is the stable key the settings row is keyed on.
+
+    Keyed on a lane NAME rather than a container hostname for the reason
+    `models/service_seen.py` gives at length: celery's worker names here are
+    `celery@<container id>` and are minted fresh on every deploy, so anything
+    keyed on them records a death and a birth every time the stack updates.
+    """
+
+    name: str
+    display_name: str
+    queues: tuple[str, ...]
+    # Which `entrypoint.sh` role starts this lane. NOT always the lane name:
+    # `maintenance_long` is the plain `worker` role pointed at a different
+    # queue, exactly as docker-compose starts it today (`command: ["worker"]`
+    # with CELERY_QUEUES=maintenance_long). Recorded here so the generated
+    # supervisord config and the compose file cannot disagree about it.
+    entrypoint_role: str
+    # THE cap a lane starts with — and, since 2026-09-23, the only number an
+    # operator sets for it. How many workers actually run is the autoscaler's
+    # job; this is the most it may use. Zero means the lane is off.
+    #
+    # One, and zero for ML. Deliberately far below the operator's own
+    # production numbers, which are tuned for their hardware and are not a
+    # sane first boot for a stranger — and low enough that a busy instance
+    # tells them to raise it rather than quietly consuming the machine.
+    default_slots_cap: int
+    # True when a slot costs a copy of the ML model rather than just a process.
+    # Such a lane is bounded by memory AS WELL AS by cores, never instead of.
+    memory_bound: bool = False
+    # CPU threads ONE slot uses. More than one for a lane whose work is an
+    # inference library with its own thread pool: `services/ml/embedder.py`
+    # calls `torch.set_num_threads` with this number, so a slot is four cores'
+    # worth of demand rather than one process's.
+    #
+    # It lives here because the CEILING has to know it. It was a private
+    # constant in the embedder with a comment saying "keep N_replicas x this
+    # within the cores allotted to ML" — a rule stated where nothing could
+    # enforce it. Nothing did: the ML ceiling was computed from memory alone,
+    # so a large-memory host offered ~49 slots, the operator took them, and
+    # 2026-09-23's log shows ~200 torch threads fighting over the box —
+    # embeds at 107-246s each, and the daily CCIP sweep sharing that pool
+    # timing out at 1800s.
+    threads_per_slot: int = 1
+    # Models this lane downloads the first time it is enabled. Empty for every
+    # lane that needs none, which is how the UI knows whether to warn at all.
+    models: tuple[ModelRequirement, ...] = ()
+    # An optional lane is one the product works without. Shown as such, so
+    # nobody turns on a multi-GB download believing it is required.
+    optional: bool = False
+
+    @property
+    def queue_key(self) -> tuple[str, ...]:
+        """The sorted queue set, which is how `service_seen` identifies a
+        running worker. The join between what is configured here and what
+        `celery inspect` reports."""
+        return tuple(sorted(self.queues))
+
+
+# ONE CAP PER LANE, and that is the whole of what an operator sets.
+#
+# Operator, 2026-09-23: *"auto should be always on, not a setting, so that
+# idle instances quiet down when not running. the number that is visible and
+# something the user can tweak and manage should be the cap itself the number
+# of running workers is handled by the autoscaling function which is always
+# on."*
+#
+# Until then a lane had THREE operator values — `slots`, `slots_cap` and
+# `autoscale` — because the manual dial was built first (steps 2-4) and the
+# autoscaler arrived last (step 7) as an opt-in beside a control that already
+# existed. Nothing ever asked whether the dial should still exist once
+# something could move it automatically. It should not: "how many are running
+# right now" is a measurement, not a preference.
+#
+# One of each, and ML at zero. ML at zero is also rule 164's carve-out: a cap
+# of zero means no consumers, so a fresh install never loads a model or
+# reaches HuggingFace, and raising the cap is what triggers the fetch.
+#
+# These are far below the operator's own production numbers, and deliberately
+# so — they are what a stranger's first boot should do, not what a tuned
+# machine can. The UI is what closes that gap: a lane sitting at its cap with
+# a backlog says so, and says raising the cap is the fix. Without that a
+# conservative default is just a slow instance nobody knows how to speed up.
+LANES: tuple[Lane, ...] = (
+    Lane(
+        name="worker",
+        display_name="Worker",
+        queues=("default", "import", "thumbnail", "download"),
+        entrypoint_role="worker",
+        default_slots_cap=1,
+    ),
+    Lane(
+        name="scheduler",
+        display_name="Scheduler",
+        queues=("maintenance", "scan"),
+        entrypoint_role="scheduler",
+        default_slots_cap=1,
+    ),
+    Lane(
+        name="maintenance_long",
+        display_name="Long maintenance",
+        queues=("maintenance_long",),
+        entrypoint_role="worker",
+        default_slots_cap=1,
+    ),
+    Lane(
+        name="ml",
+        display_name="ML tagging",
+        queues=("ml",),
+        entrypoint_role="ml-worker",
+        default_slots_cap=0,
+        memory_bound=True,
+        threads_per_slot=4,
+        models=(SIGLIP_MODEL,),
+        optional=True,
+    ),
+)
+
+LANES_BY_NAME: dict[str, Lane] = {lane.name: lane for lane in LANES}
+LANES_BY_QUEUE_KEY: dict[tuple[str, ...], Lane] = {
+    lane.queue_key: lane for lane in LANES
+}
+
+
+# --- what the container actually has -----------------------------------------
+
+# cgroup v2 first, then v1. A container started without an explicit memory
+# limit reports "max" on v2 and a sentinel near 2**63 on v1; both mean "no
+# limit", and the answer then is the host's RAM.
+_CGROUP_V2_MEMORY = Path("/sys/fs/cgroup/memory.max")
+_CGROUP_V1_MEMORY = Path("/sys/fs/cgroup/memory/memory.limit_in_bytes")
+_CGROUP_V2_CPU = Path("/sys/fs/cgroup/cpu.max")
+_CGROUP_V1_CPU_QUOTA = Path("/sys/fs/cgroup/cpu/cpu.cfs_quota_us")
+_CGROUP_V1_CPU_PERIOD = Path("/sys/fs/cgroup/cpu/cpu.cfs_period_us")
+
+# A v1 "unlimited" is PAGE_SIZE-aligned LONG_MAX, not a round number, so it is
+# recognised by magnitude rather than by equality. Anything claiming more than
+# a petabyte is a sentinel, not a machine.
+_UNLIMITED_ABOVE = 1 << 50
+
+# DERIVED from the model requirement above, never restated. The ceiling and
+# the number shown to the operator before they enable the lane have to be the
+# same figure, or the UI promises something the cap will then refuse.
+ML_BYTES_PER_SLOT = SIGLIP_MODEL.approx_resident_bytes
+
+# Held back for hypercorn and the non-ML lanes before any ML slot is offered.
+# In the consolidated container these share one cgroup with ML, and they are
+# the processes an OOM kill must not take (see the module docstring).
+RESERVED_BYTES = 2 * GIB
+
+# The smallest pool a lane can actually run: ONE process, never zero.
+#
+# billiard refuses to remove the last worker in a pool, so a lane asked to
+# shrink to nothing gets `ValueError("Can't shrink pool. All processes
+# busy!")` and the sizing pass re-sends the doomed message forever. Found on
+# the operator's live deploy, 2026-09-23.
+#
+# It is also what makes "off" expressible: a lane at cap 0 keeps this one
+# parked process with its consumers cancelled, so it still answers `inspect`
+# (and so reads as present rather than crashed), and `add_consumer` has
+# something to reach when the cap goes back up.
+#
+# Lives HERE rather than in `worker_control` because `gen_supervisord` needs
+# it at container boot and must not import the models package to get it.
+MIN_POOL_SLOTS = 1
+
+# The floor a cores-derived ceiling never goes below. A single-core box still
+# needs to be able to run its lanes; the ceiling exists to stop absurd values,
+# not to make a small machine unusable.
+MIN_CEILING = 1
+
+# How often `size_worker_lanes` runs — the beat schedule, and the freshness of
+# everything the System tab shows.
+#
+# It is here, in the import-light module, because three places have to agree
+# about it and they are in different packages: the beat entry in `celery_app`,
+# the sample the sweep writes (`worker_lane_sample`), and the roster's
+# staleness thresholds in `api/system_health`, which now depend on this sweep
+# rather than on a browser being open.
+#
+# 30s, down from 60s, because the sweep became the ONLY writer of the celery
+# roster on 2026-09-23. A part is called stale after 90s of silence, so a
+# 60-second sweep left one missed tick between "normal" and "everything is
+# yellow". That is the shape of lesson #4355 — a reader's threshold and an
+# emitter's cadence chosen in different files and never compared — and the
+# fix is headroom plus a test that asserts it, not a number that happens to
+# work today.
+#
+# The cost is one inspect every 30s instead of every 60s; the saving is every
+# inspect that used to run on a request path, which with a single tab open
+# was roughly four a minute against this two. Consequence worth knowing: the
+# pass also SHRINKS an idle lane by one slot per tick, so an idle lane now
+# gives its workers back twice as fast. That is the direction the operator
+# asked for — *"idle instances quiet down when not running"*.
+SWEEP_PERIOD_SECONDS = 30.0
+
+# What an unreadable limit yields. Low rather than unlimited, on purpose: not
+# knowing how much memory there is must never read as "plenty". An unswept
+# absence is not a verdict.
+UNKNOWN_CEILING = 1
+
+
+def _read_int(path: Path) -> int | None:
+    try:
+        raw = path.read_text().strip()
+    except OSError:
+        return None
+    if raw == "max":
+        return None
+    try:
+        return int(raw)
+    except ValueError:
+        return None
+
+
+def container_memory_bytes() -> int | None:
+    """The memory this container may use, or None when it cannot be read.
+
+    None means UNKNOWN, never UNLIMITED. Every caller must treat it as the
+    conservative case — the whole point of the ceiling is to protect a machine
+    whose size we are unsure of.
+    """
+    for path in (_CGROUP_V2_MEMORY, _CGROUP_V1_MEMORY):
+        value = _read_int(path)
+        if value is not None and value < _UNLIMITED_ABOVE:
+            return value
+        if value is not None:
+            # A sentinel: the cgroup exists but sets no limit, so the real
+            # bound is the host's.
+            break
+    try:
+        return os.sysconf("SC_PHYS_PAGES") * os.sysconf("SC_PAGE_SIZE")
+    except (ValueError, OSError, AttributeError):
+        return None
+
+
+def container_cpu_count() -> int | None:
+    """Effective cores, honouring a cgroup CPU quota.
+
+    `os.cpu_count()` reports the HOST's cores from inside a container, so a
+    quota of 2.0 on a 32-core host would otherwise offer 32 slots. The
+    operator's own stack sets `cpus: '4.0'` on ml-worker, so this is a real
+    configuration here and not a hypothetical.
+    """
+    quota: float | None = None
+    try:
+        raw = _CGROUP_V2_CPU.read_text().strip().split()
+        if raw and raw[0] != "max":
+            quota = int(raw[0]) / int(raw[1])
+    except (OSError, ValueError, IndexError, ZeroDivisionError):
+        pass
+    if quota is None:
+        q = _read_int(_CGROUP_V1_CPU_QUOTA)
+        p = _read_int(_CGROUP_V1_CPU_PERIOD)
+        if q is not None and p and q > 0:
+            quota = q / p
+    if quota is not None and quota > 0:
+        return max(1, int(quota))
+    return os.cpu_count()
+
+
+def lane_for_node(hostname: str) -> Lane | None:
+    """`ml@7f3c9a1b` -> the ml lane. None for a node this build did not name.
+
+    ## Why the node name, and not the queues it is consuming
+
+    Because a lane that is OFF is consuming nothing, and "nothing" identifies
+    no lane at all.
+
+    Both the roster and `inspect_lanes_sync` used to map a worker to its lane
+    through `active_queues()`. That is exact while the lane is running and
+    useless the moment it is not: a lane at cap 0 has its consumers cancelled,
+    so it answers the broadcast with an EMPTY queue list, matches no lane, and
+    is dropped. Three things followed, and the operator saw all three at once
+    on 2026-09-23:
+
+    1. The lanes table showed the lane as **not answering** — which is the
+       signal for a crashed worker, not for one the operator turned off.
+    2. The roster grew a phantom row called **`Worker ()`**, the empty queue
+       set rendered as a display name, "running" beside the real lane's row
+       going stale.
+    3. **The container went unhealthy.** `healthcheck._lanes_ok` requires
+       every lane in the table to be present, and its docstring asserted the
+       opposite of what the code did — *"a disabled lane still runs its
+       process with its consumers cancelled, so it answers inspect and is
+       healthy"*. It answers; it is not attributed. ML ships at cap 0, so a
+       fresh install would have been permanently unhealthy, and Swarm
+       restarts an unhealthy task forever.
+
+    The node name survives all of that: `gen_supervisord` sets
+    `CELERY_NODENAME={lane.name}` per program and the entrypoint passes it to
+    `celery -n`, so the identity travels with the PROCESS rather than with
+    what it happens to be doing. Falls back to the queue set for a deployment
+    that sets no node name — the multi-service compose stack, where every node
+    is `celery@<host>`.
+    """
+    return LANES_BY_NAME.get(hostname.split("@", 1)[0])
+
+
+def _cpu_bound_slots(lane: Lane) -> int:
+    """How many slots this container's cores can feed, at `threads_per_slot`.
+
+    Never zero: a machine with fewer cores than one slot wants still runs the
+    lane, just slowly. That is a real trade an operator may want, and refusing
+    to offer the lane at all on a small box would make ML unreachable there —
+    unlike the memory bound, where the honest answer IS zero, because the
+    first task would OOM the container rather than merely be slow.
+    """
+    cores = container_cpu_count()
+    if cores is None:
+        return UNKNOWN_CEILING
+    return max(MIN_CEILING, cores // lane.threads_per_slot)
+
+
+def derived_ceiling(lane: Lane) -> int:
+    """The most slots `lane` may be given on this container.
+
+    Never stored. Recomputed on every read so a container whose limits changed
+    is bounded by what it has NOW rather than by what it had when its row was
+    written.
+    """
+    by_cpu = _cpu_bound_slots(lane)
+    if not lane.memory_bound:
+        return by_cpu
+
+    total = container_memory_bytes()
+    if total is None:
+        log.warning(
+            "worker_lanes: cannot read a memory limit; capping %s at %d",
+            lane.name, UNKNOWN_CEILING,
+        )
+        return UNKNOWN_CEILING
+    usable = total - RESERVED_BYTES
+    if usable < ML_BYTES_PER_SLOT:
+        # Honestly zero. A box that cannot hold one model alongside the web
+        # process must be told it cannot run tagging, not sold a slot that
+        # will OOM the container the first time it is used.
+        return 0
+
+    # BOTH bounds, whichever binds first. Memory alone was the whole answer
+    # until 2026-09-23, and on a large-memory host that is the wrong one: RAM
+    # said ~49 slots, and each of those slots wants `threads_per_slot` cores.
+    # The operator raised the cap to what the dial offered and the lane
+    # starved itself — a control is not allowed to offer a number the machine
+    # cannot feed.
+    return min(int(usable // ML_BYTES_PER_SLOT), by_cpu)
+
+
+def ceilings() -> dict[str, int]:
+    """Every lane's ceiling, for the settings API and the UI."""
+    return {lane.name: derived_ceiling(lane) for lane in LANES}

backend/app/tasks/admin.py

@@ -122,6 +122,28 @@ def prune_missing_file_records_task(self) -> dict:
     return {"checked": checked, "missing": len(missing_ids), "deleted": deleted}
 
 
+@celery.task(
+    name="backend.app.tasks.admin.repair_discord_downloads_task",
+    bind=True,
+    autoretry_for=(OperationalError, DBAPIError),
+    retry_backoff=15, retry_backoff_max=180, max_retries=1,
+    soft_time_limit=1800, time_limit=2400,  # 30 min / 40 min
+)
+def repair_discord_downloads_task(self, dry_run: bool = True) -> dict:
+    """Clean re-download of the Discord files broken by the `None` naming
+    (#3999). dry_run (the default) returns the projection; apply deletes the
+    broken images and their files, makes gallery-dl forget every Discord
+    download, and restarts every Discord source's backfill. Defaults to the SAFE
+    preview because the apply deletes files. Operator-triggered only."""
+    from ..services.discord_repair import repair_discord_downloads
+
+    SessionLocal = _sync_session_factory()
+    with SessionLocal() as session:
+        return repair_discord_downloads(
+            session, images_root=IMAGES_ROOT, dry_run=dry_run,
+        )
+
+
 @celery.task(
     name="backend.app.tasks.admin.dedup_videos_task",
     bind=True,

backend/app/tasks/download.py

@@ -177,6 +177,10 @@ def download_source(self, source_id: int, _serialize_waits: int = 0) -> int:
                 settings = ImportSettings.load_sync(sync_session)
                 rate_limit = settings.download_rate_limit_seconds
                 validate_files = settings.download_validate_files
+                # How far back a tick keeps looking for EDITED posts. Read here
+                # with the other downloader knobs, off the row this block is
+                # already holding open.
+                revisit_days = settings.download_revisit_days
 
             gdl = GalleryDLService(
                 images_root=IMAGES_ROOT,
@@ -207,6 +211,7 @@ def download_source(self, source_id: int, _serialize_waits: int = 0) -> int:
                         # the walk). Same factory the importer's sync session
                         # comes from — a different DB connection per checkout.
                         sync_session_factory=SyncFactory,
+                        revisit_days=revisit_days,
                     )
                     return await svc.download_source(source_id)
         finally:

backend/app/tasks/maintenance.py

@@ -473,6 +473,10 @@ def prune_task_runs() -> dict:
       (recover_stalled_task_runs) is the mechanism that flips them to
       terminal state; prune doesn't touch in-flight state.
     - 'retry' rows: treated as failures (>7d).
+    - The NEWEST row of each task is never deleted, whatever its age: it is
+      what the beat scheduler reads to know when a job last ran (#4408).
+      Without it a weekly job's last run would be pruned after a day, and beat
+      would think it had never run and fire it on every restart.
 
     Returns dict of how many rows were deleted in each bucket.
     """
@@ -480,16 +484,19 @@ def prune_task_runs() -> dict:
     now = datetime.now(UTC)
     ok_cutoff = now - timedelta(seconds=TASK_RUN_KEEP_OK_SECONDS)
     fail_cutoff = now - timedelta(seconds=TASK_RUN_KEEP_FAILURE_SECONDS)
+    newest = select(func.max(TaskRun.id)).group_by(TaskRun.task_name)
     with SessionLocal() as session:
         ok_deleted = session.execute(
             delete(TaskRun)
             .where(TaskRun.status == "ok")
             .where(TaskRun.finished_at < ok_cutoff)
+            .where(TaskRun.id.not_in(newest))
         ).rowcount or 0
         fail_deleted = session.execute(
             delete(TaskRun)
             .where(TaskRun.status.in_(["error", "timeout", "retry"]))
             .where(TaskRun.finished_at < fail_cutoff)
+            .where(TaskRun.id.not_in(newest))
         ).rowcount or 0
         session.commit()
         return {"ok_deleted": ok_deleted, "failures_deleted": fail_deleted}
@@ -502,10 +509,16 @@ def prune_task_runs() -> dict:
     soft_time_limit=1800, time_limit=2100,
 )
 def backfill_phash() -> int:
-    """Recompute phash for stored images that have none (imported before
-    FC-2d-i+ii). Keyset-paginated by id (restart-safe), NULL-only fill,
-    idempotent. Videos legitimately keep phash NULL. A missing/unreadable
-    file is logged and left NULL — never fails the task."""
+    """Recompute phash for stored images that have none. Keyset-paginated by
+    id (restart-safe), NULL-only fill, idempotent. Videos legitimately keep
+    phash NULL. A missing/unreadable file is logged and left NULL — never
+    fails the task.
+
+    Two sources of NULLs: images imported before FC-2d-i+ii, and migration
+    0098, which cleared every phash so the library could be re-hashed at
+    hash_size=16 (#4223). The daily beat entry exists for the second — until
+    a row is refilled it takes no part in dedup, which is why this runs on a
+    schedule rather than waiting for a deep scan."""
     SessionLocal = _sync_session_factory()
     updated = 0
     last_id = 0
@@ -1169,7 +1182,7 @@ def group_discord_drops() -> str:
         return "disabled"
     return (
         f"sources={res['sources']} created={res['posts_created']} "
-        f"joined={res['images_joined']}"
+        f"joined={res['images_joined']} merged={res['drops_merged']}"
     )
 
 
@@ -1178,12 +1191,16 @@ def group_discord_drops() -> str:
     soft_time_limit=900, time_limit=1200,
 )
 def match_post_associations() -> str:
-    """Milestone 388 E5: propose which Patreon post announced which Discord drop.
+    """Milestone 388 E5: which Patreon post announced which Discord drop.
 
-    Proposes only — every pair lands in a review queue and nothing is linked
-    until the operator accepts. Maintenance lane for the same reason as the
-    grouper: no inference, no ML library, and it must not depend on the
-    optional ml-worker being present.
+    A CONCLUSIVE pair — one the creator's own working name identifies, where
+    that name appears in these two posts and nowhere else in their library — is
+    linked outright when `discord_link_auto` is on, because there is nothing
+    there for the operator to adjudicate. Everything weaker lands in the review
+    queue and stays unlinked until they accept it.
+
+    Maintenance lane for the same reason as the grouper: no inference, no ML
+    library, and it must not depend on the optional ml-worker being present.
     """
     import asyncio
 
@@ -1247,6 +1264,7 @@ def sync_memberships() -> str:
     from ..services.artist_membership_service import rescan as membership_rescan
     from ..services.credential_crypto import CredentialCrypto
     from ..services.credential_service import CredentialService
+    from ..services.membership_reconcile import apply_membership_lapses
     from ..services.membership_roster import roster_user_id, sync_platform
     from ..services.patreon_client import PatreonClient
     from ..services.subscribestar_client import SubscribeStarClient
@@ -1298,9 +1316,18 @@ def sync_memberships() -> str:
                     )
 
                 async with async_factory() as session:
-                    results.append(
-                        await sync_platform(session, platform=platform, fetch=fetch)
-                    )
+                    result = await sync_platform(session, platform=platform, fetch=fetch)
+                results.append(result)
+
+                # #3995: stop pulling sources whose paid access has ended, and
+                # resume the ones this stopped once they are paid again. Only
+                # right after a successful sync, so it always acts on the roster
+                # just written, never on a stale one.
+                if result.get("ok"):
+                    async with async_factory() as session:
+                        result["lapses"] = await apply_membership_lapses(
+                            session, platform=platform,
+                        )
 
             # #388 E4: offer the freshly-synced roster to the artists FC already
             # tracks. Chained here rather than given its own beat entry because
@@ -1321,9 +1348,100 @@ def sync_memberships() -> str:
         if "skipped" in r:
             parts.append(f"{r['platform']}=skipped({r['skipped']})")
         elif r.get("ok"):
-            parts.append(f"{r['platform']}={r['count']}")
+            lapses = r.get("lapses") or {}
+            detail = (
+                f"(stopped={lapses['stopped']},resumed={lapses['resumed']})"
+                if lapses.get("stopped") or lapses.get("resumed") else ""
+            )
+            parts.append(f"{r['platform']}={r['count']}{detail}")
         else:
             parts.append(f"{r['platform']}=FAILED({r['error']})")
     if res.get("suggested") is not None:
         parts.append(f"suggested={res['suggested']}")
     return " ".join(parts) or "no platforms"
+
+@celery.task(name="backend.app.tasks.maintenance.size_worker_lanes")
+def size_worker_lanes() -> dict:
+    """Size every lane to its backlog, within the cap the operator set.
+
+    ONE sweep, replacing `reconcile_worker_lanes` and `autoscale_worker_lanes`
+    (2026-09-23). They were two enforcers over one number: the reconcile drove
+    the pool to a stored `slots`, the autoscaler moved it away from that same
+    value, and most of the autoscaler's design existed to keep the reconcile
+    from undoing its work. Deleting the stored number deletes the conflict.
+
+    It still does what the reconcile existed for. `pool_grow` is not durable —
+    a worker restarted by its supervisor comes back at its ENV concurrency,
+    silently below what the lane should run — and this reads the LIVE pool
+    every minute, so that worker is corrected on the next tick rather than
+    after five.
+
+    Returns every lane's outcome INCLUDING the ones it held, each with a
+    reason. A pass that only speaks when it acts cannot be debugged on the day
+    it does not.
+
+    ## It is also the only thing that MEASURES, since 2026-09-23
+
+    It always inspected the broker to decide pool sizes, and then threw the
+    reading away — while `/api/system/workers` ran the same inspect on every
+    page load and the System tab polls it four times a minute. Operator:
+    *"there is a repull every time this page loads — is there a reason this
+    info isn't being tracked in the background and stored in some way?"*
+
+    So one inspect now feeds three things: the sizing decision, the stored
+    sample the System tab reads, and the celery roster. No request path
+    touches the broker any more.
+
+    Order matters. The sample is stored BEFORE the roster refresh, because
+    that refresh does its own broadcast and a broker that has just started
+    failing must not cost us the reading we already have.
+    """
+    from ..models import WorkerLane
+    from ..services.service_roster import refresh_celery_roster_sync
+    from ..services.worker_control import (
+        _queue_depths_sync,
+        inspect_lanes_sync,
+        size_lanes_sync,
+        store_lane_samples_sync,
+    )
+
+    # Read INSIDE the session. Reading a column off a detached instance
+    # happens to work while the attribute is still loaded and stops working
+    # the moment anything expires it — a failure that would appear long after
+    # this line, in a sweep nobody is watching.
+    with _sync_session_factory()() as session:
+        caps = {
+            row.name: row.slots_cap
+            for row in session.execute(select(WorkerLane)).scalars()
+        }
+    if not caps:
+        # Migration 0103/0105 seed these, so an empty table means they have
+        # not run yet. Nothing to assert — and inventing defaults here would
+        # let this task disagree with the seed it is meant to be enforcing.
+        return {"sized": []}
+
+    # Measured ONCE, here, and then used three times. Passing them down is
+    # what makes the reading keepable rather than an implementation detail of
+    # a function that returns decisions.
+    live = inspect_lanes_sync()
+    depths = _queue_depths_sync()
+
+    sized = size_lanes_sync(caps, live=live, depths=depths)
+
+    with _sync_session_factory()() as session:
+        store_lane_samples_sync(session, live, depths)
+        refresh_celery_roster_sync(session)
+
+    for d in sized:
+        if d.action not in ("held", "skipped"):
+            log.info(
+                "worker lanes: %s %s to %s slots — %s",
+                d.lane, d.action, d.slots, d.reason,
+            )
+    return {
+        "sized": [
+            {"lane": d.lane, "action": d.action, "slots": d.slots,
+             "reason": d.reason}
+            for d in sized
+        ],
+    }

backend/app/tasks/ml.py

@@ -488,7 +488,7 @@ def scheduled_ccip_auto_apply() -> str:
 
     from ..models import ImageRegion, MLSettings, Tag, TagKind
     from ..models.tag import image_tag
-    from ..services.ml.ccip import _FIGURE_KINDS
+    from ..services.ml.ccip import _FIGURE_KINDS, char_maxima
     from ..services.ml.training_data import _applied_or_rejected, _l2norm
 
     SessionLocal = _sync_session_factory()
@@ -553,11 +553,22 @@ def scheduled_ccip_auto_apply() -> str:
             by_img: dict[int, list] = {}
             for iid, vec in rows:
                 by_img.setdefault(iid, []).append(vec)
-            for iid, vecs in by_img.items():
-                q = _l2norm(np.asarray(vecs, dtype=np.float32), np)  # (nq, 768)
-                colmax = (q @ allref.T).max(axis=0)              # (total,)
-                charmax = np.maximum.reduceat(colmax, seg)       # (n_chars,)
-                for ci in np.where(charmax >= thr)[0]:
+            if not by_img:
+                continue
+
+            # One matmul per BLOCK of figures, not one per image. This loop ran
+            # over every image in the library on every daily run and did a
+            # matmul too small to pay for itself each time; it hit the 1800s
+            # soft limit on the operator's instance on 2026-09-23. Same
+            # arithmetic — see `char_maxima`.
+            iids = list(by_img)
+            charmax = char_maxima(
+                [_l2norm(np.asarray(by_img[i], dtype=np.float32), np) for i in iids],
+                allref, seg, np,
+            )                                                    # (n_img, n_chars)
+
+            for row, iid in enumerate(iids):
+                for ci in np.where(charmax[row] >= thr)[0]:
                     t = ref_tags[int(ci)]
                     if iid in skip[t]:
                         continue
@@ -668,3 +679,31 @@ def scheduled_retract_auto_tags() -> str:
     with SessionLocal() as session:
         n_ccip = retract_auto_applied_ccip(session)
     return f"head={n_head} ccip={n_ccip}"
+
+
+@celery.task(name="backend.app.tasks.ml.ensure_models", bind=True)
+def ensure_models(self) -> dict:
+    """Fetch the models this lane needs, if they are not already present.
+
+    Milestone 422 step 6. This used to run in `entrypoint.sh` before celery
+    started, which made every boot of the ML role reach HuggingFace for
+    ~3.5GB — a startup dependency on a third party, for a feature the operator
+    may never use. Rule 164 permits a runtime fetch only for something
+    "optional and clearly off", so it moved here: enqueued the moment the lane
+    is ENABLED, never at boot.
+
+    Being a task rather than a startup step is what makes it visible: it gets
+    a TaskRun row like any other, so the download shows in Activity with a
+    duration and a status, and a failure is something the operator can see and
+    retry rather than a container that quietly never became useful.
+
+    Idempotent — `download_models` fetches only what is missing — so enabling
+    an already-provisioned lane costs one no-op task rather than a re-download.
+    That matters because the reconcile may enqueue it again.
+    """
+    from ..scripts.download_models import main as download
+
+    rc = download()
+    if rc != 0:
+        raise RuntimeError(f"model download failed with exit code {rc}")
+    return {"ok": True}