Merge pull request 'Schema reconciliation + index hygiene, and the weekly base-image refresh' (#243) from dev into main

A structural sweep of the deployed schema, run AFTER 0088 got the models
and the chain to exact agreement. That agreement is what 0088 achieved,
and it is worth naming what it does not prove: a models-vs-chain diff
shows the two describe the same schema, not that the schema is right.
Everything here was wrong in BOTH.

The one that matters: image_tag has PRIMARY KEY (image_record_id, tag_id)
and no other index, so tag_id is unindexed. That is the gallery's tag
filter (tag_query.py builds `image_tag.c.tag_id == tid`) and the ON
DELETE CASCADE from tag, both scanning the largest table in the schema.
Six more FKs were unindexed on smaller tables; presentation_review.tag_id
also CASCADEs.

Dropped, on the other side: ix_image_record_sha256 was an exact duplicate
of the index uq_image_record_sha256 already builds — two btrees on the
same column of the highest-insert-rate table. The other six are
single-column indexes a later composite superseded without the narrow one
being retired; a btree on (a,b) already serves lookups on a.

0088 deliberately taught the models to declare BOTH sha256 indexes so
they would describe reality. This changes the reality instead, and the
models change with it — otherwise the next baseline.yml run reintroduces
exactly the drift 0088 removed.

CONCURRENTLY throughout, so building the image_tag index does not hold an
ACCESS EXCLUSIVE lock over every write for the duration. The cost is that
the migration cannot run in a transaction and so is not atomic: every
statement is IF NOT EXISTS / IF EXISTS, making a re-run after a partial
failure safe. The docstring carries the query for finding an INVALID
index left by an interrupted CONCURRENTLY build.

What the sweep found clean, for the record: all 43 tables have a primary
key; all 51 FKs declare an explicit ON DELETE, so none silently blocks a
delete; the three enum CHECKs match the code that writes them (rule 36).

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017QHszn9H8VBvx5Ke8x1hvw

.forgejo/workflows/baseline.yml

@@ -0,0 +1,370 @@
+
+# TEMPORARY — milestone 328 steps 1-2. Delete once the baseline is stamped.
+#
+# Squashing 87 alembic revisions into one baseline has exactly one dangerous
+# failure: the generated baseline does not reproduce the schema the chain
+# produced, `alembic stamp` writes a version string anyway (it validates
+# NOTHING), and the divergence surfaces on the next real migration against the
+# operator's live data.
+#
+# So this workflow does the comparison in CI, where a pgvector Postgres already
+# gets built from the chain on every integration run, and nothing is at risk.
+# It answers one question: does `upgrade head` on the collapsed chain produce a
+# byte-identical schema to `upgrade head` on the 87-revision chain?
+#
+# The chain is read from git rather than from the working tree, so this keeps
+# working AFTER the old revisions are deleted — `chain_ref` names a commit that
+# still has them. That is what makes this the proof for step 1 and the
+# pre-flight for step 2, rather than a one-shot script.
+#
+# While the chain is still present it also autogenerates a candidate baseline
+# from the models and prints it. That is a starting point, NOT the answer:
+# autogenerate reads SQLAlchemy metadata, and three things here do not live
+# there —
+#   * CREATE EXTENSION vector           (0001)
+#   * CREATE EXTENSION tsm_system_rows  (0004)
+#   * the HNSW index on image_record.siglip_embedding, which is raw SQL
+#     because alembic's create_index cannot express `USING hnsw (...)`  (0036)
+# plus any CHECK constraint or server_default that a migration added without
+# the model declaring it. Those must be hand-added, and the diff below is what
+# proves none were missed.
+name: Alembic baseline
+
+on:
+  workflow_dispatch:
+    inputs:
+      chain_ref:
+        description: 'Commit/tag carrying the full chain; blank = this ref (use a pinned commit only AFTER the collapse)'
+        type: string
+        default: ''
+      mode:
+        description: 'chain = compare against this tree''s migrations; models = compare against a schema built from the MODELS'
+        type: string
+        default: 'chain'
+
+jobs:
+  compare:
+    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
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          # Full history is the point: `chain_ref` is read out of git, so a
+          # shallow clone would not have the revisions to compare against.
+          fetch-depth: 0
+
+      - name: Resolve the Postgres service and install deps
+        run: |
+          set -eux
+          # Same service-IP dance as ci.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"
+          PG_IP=$(docker inspect -f '{{range .NetworkSettings.Networks}}{{.IPAddress}}{{end}}' "$PG")
+          test -n "$PG_IP"
+          echo "PG_CONTAINER=$PG" >> "$GITHUB_ENV"
+          echo "DB_HOST=$PG_IP" >> "$GITHUB_ENV"
+          for i in $(seq 1 60); do
+            (echo > "/dev/tcp/$PG_IP/5432") >/dev/null 2>&1 && break
+            sleep 2
+          done
+          if command -v uv >/dev/null 2>&1; then
+            uv pip install --system -r requirements.txt
+          else
+            pip install -r requirements.txt
+          fi
+
+      # DB 1: the 87-revision chain, read out of git at `chain_ref`.
+      #
+      # A git worktree rather than a checkout, so the current tree — which is
+      # what we are testing — is left completely alone.
+      - name: Build the schema the OLD chain produces
+        env:
+          CHAIN_REF: ${{ github.event.inputs.chain_ref }}
+          THIS_SHA: ${{ github.sha }}
+        run: |
+          set -eux
+          docker exec "$PG_CONTAINER" createdb -U fabledcurator fc_chain
+          # Blank means "the chain in this ref", which is what you want while
+          # the chain is still intact — comparing the models against a PINNED
+          # older commit reports every migration written since as a difference.
+          # Pin it only after the collapse, when the tree no longer has them.
+          git worktree add /tmp/chain "${CHAIN_REF:-$THIS_SHA}"
+          ls /tmp/chain/alembic/versions/*.py | wc -l
+          cd /tmp/chain
+          DB_NAME=fc_chain alembic upgrade head
+          cd -
+          docker exec "$PG_CONTAINER" pg_dump -U fabledcurator --schema-only \
+            --no-owner --no-privileges -d fc_chain > chain.sql
+          wc -l chain.sql
+          # Emit the dump itself, checksummed, for local analysis. Reconciling
+          # the models against the deployed schema (#3275) needs the ACTUAL
+          # schema, not an inference from a diff — parsing table context out of
+          # unified-diff hunks drops every table whose CREATE TABLE line falls
+          # outside a hunk, which silently under-reports.
+          #
+          # base64 + sha256 for the same reason as the candidate: a plain cat
+          # of a file this size was truncated mid-line by the runner with the
+          # step still green (run 4964).
+          set +x
+          B64=$(base64 -w 120 chain.sql)
+          echo "===== BEGIN CHAIN SCHEMA (base64) ====="
+          echo "$B64"
+          echo "===== END CHAIN SCHEMA ====="
+          echo "chain-sha256: $(sha256sum chain.sql | cut -d' ' -f1)"
+          echo "chain-bytes:  $(wc -c < chain.sql)"
+          set -x
+
+      # A candidate baseline, autogenerated from the models against an EMPTY
+      # database so every table shows up as a create. Printed for a human to
+      # finish — it will be missing the three raw-SQL items named at the top.
+      #
+      # Gated on the TREE, not on a workflow input. A `type: boolean` input
+      # read back as `github.event.inputs.generate == 'true'` silently
+      # evaluated false on this runner (run 4960 skipped this step entirely
+      # with no diagnostic) — the same `github.event.inputs` typing quirk
+      # build.yml already works around. The file count is the real question
+      # anyway: there is nothing to generate once the chain is collapsed.
+      - name: Autogenerate a candidate baseline
+        run: |
+          set -eux
+          if [ "$(ls alembic/versions/*.py | wc -l)" -le 1 ]; then
+            echo "already collapsed — nothing to generate"
+            exit 0
+          fi
+          docker exec "$PG_CONTAINER" createdb -U fabledcurator fc_gen
+          # Hide the existing revisions so alembic sees an empty history and
+          # emits the whole schema rather than a delta.
+          mkdir -p /tmp/versions_held
+          mv alembic/versions/*.py /tmp/versions_held/ 2>/dev/null || true
+          DB_NAME=fc_gen alembic revision --autogenerate -m "baseline" || true
+          # Printed rather than uploaded: ci-requirements.md records that this
+          # runner cannot do actions/upload-artifact@v4+, and the repo dropped
+          # the action entirely in 2026-05, so the job log is the retrieval
+          # channel actually proven here.
+          #
+          # base64, not the raw file. A plain `cat` of the ~33KB candidate was
+          # TRUNCATED MID-LINE by the runner on run 4964 — it stopped inside
+          # `sa.Column('mime', sa.String(length=128)` and carried straight on
+          # to the next traced command, with the step still green. A silent
+          # cut in the middle of a schema definition is the worst possible
+          # failure here, because the truncated text still looks like a
+          # plausible file.
+          #
+          # base64 at a fixed narrow width gives many short lines instead of
+          # few long ones, and — the actual point — a checksum and a line
+          # count that make truncation DETECTABLE rather than invisible.
+          set +x
+          F=$(ls alembic/versions/*.py | head -1)
+          B64=$(base64 -w 120 "$F")
+          echo "===== BEGIN CANDIDATE BASELINE (base64) ====="
+          echo "$B64"
+          echo "===== END CANDIDATE BASELINE ====="
+          echo "candidate-sha256: $(sha256sum "$F" | cut -d' ' -f1)"
+          echo "candidate-bytes:  $(wc -c < "$F")"
+          echo "candidate-b64-lines: $(echo "$B64" | wc -l)"
+          set -x
+          mkdir -p /tmp/candidate
+          cp alembic/versions/*.py /tmp/candidate/
+          # Put the tree back exactly as it was; this job never mutates state.
+          rm -f alembic/versions/*.py
+          mv /tmp/versions_held/*.py alembic/versions/ 2>/dev/null || true
+
+      # DB 2: what the CURRENT tree produces.
+      #
+      # `mode: models` applies the candidate autogenerated from the MODELS
+      # instead, which is what answers "do the models describe the schema?" —
+      # the question #3275 exists because nobody had ever asked it. Under that
+      # mode a clean diff means autogenerate is trustworthy again.
+      #
+      # The two extensions are created by hand first. They are database
+      # objects, not table metadata, so no model can carry them and their
+      # absence is not a model defect — it is simply outside what this
+      # comparison is asking about.
+      - name: Build the schema the CURRENT tree produces
+        env:
+          MODE: ${{ github.event.inputs.mode }}
+        run: |
+          set -eux
+          docker exec "$PG_CONTAINER" createdb -U fabledcurator fc_base
+          if [ "${MODE:-chain}" = "models" ]; then
+            docker exec "$PG_CONTAINER" psql -U fabledcurator -d fc_base \
+              -c "CREATE EXTENSION IF NOT EXISTS vector" \
+              -c "CREATE EXTENSION IF NOT EXISTS tsm_system_rows"
+            mkdir -p /tmp/held
+            mv alembic/versions/*.py /tmp/held/
+            cp /tmp/candidate/*.py alembic/versions/
+            # Autogenerate EMITS pgvector.sqlalchemy.vector.VECTOR(...) without
+            # importing it, so the file it writes cannot run:
+            #   NameError: name 'pgvector' is not defined
+            # Observed on run 4988, which is the proof rather than the theory.
+            # This is a defect in the GENERATOR, not in the models, so it is
+            # repaired here rather than counted as a schema difference — the
+            # comparison is about whether the models describe the schema.
+            sed -i '0,/^import sqlalchemy as sa$/s//import sqlalchemy as sa\nimport pgvector.sqlalchemy.vector/' alembic/versions/*.py
+            grep -n 'import pgvector' alembic/versions/*.py
+            # Second generator defect, same class as the missing import.
+            #
+            # base.py's naming convention includes %(constraint_name)s for ck,
+            # which — unlike uq/fk/ix — means the convention is applied even to
+            # a CheckConstraint that HAS a name. So a model declaring
+            # name="singleton" correctly becomes ck_ml_settings_singleton in
+            # the metadata. Autogenerate then writes that RENDERED name into
+            # the migration, and running the migration applies the convention a
+            # SECOND time: ck_ml_settings_ck_ml_settings_singleton.
+            #
+            # That is round-tripping damage done by the generator, not a claim
+            # the models make, so it is repaired here rather than counted as a
+            # schema difference. Undone by removing the ck_<table>_ prefix the
+            # convention will re-add — the exact inverse, and it only fires on
+            # a name that actually carries its own table's prefix.
+            python3 - alembic/versions/*.py <<'PYEOF'
+          import re, sys
+
+          table = None
+          for path in sys.argv[1:]:
+              out = []
+              for line in open(path):
+                  m = re.search(r"op\.create_table\(\s*[\"']([A-Za-z0-9_]+)[\"']", line)
+                  if m:
+                      table = m.group(1)
+                  if table and "CheckConstraint" in line:
+                      prefix = f"ck_{table}_"
+                      line = re.sub(
+                          r"(name=[\"'])" + re.escape(prefix),
+                          r"\1",
+                          line,
+                      )
+                  out.append(line)
+              open(path, "w").writelines(out)
+          PYEOF
+            grep -n 'CheckConstraint' alembic/versions/*.py || true
+            ls alembic/versions/*.py
+            DB_NAME=fc_base alembic upgrade head
+            rm -f alembic/versions/*.py
+            mv /tmp/held/*.py alembic/versions/
+          else
+            ls alembic/versions/*.py | wc -l
+            DB_NAME=fc_base alembic upgrade head
+          fi
+          docker exec "$PG_CONTAINER" pg_dump -U fabledcurator --schema-only \
+            --no-owner --no-privileges -d fc_base > baseline.sql
+          wc -l baseline.sql
+
+      # The verdict.
+      #
+      # pg_dump orders dumpable objects by name within type, not by creation
+      # order, so two schemas built by different routes are directly
+      # comparable. Normalisation is deliberately minimal, because a filter
+      # that hides a real difference is the one way this check passes when it
+      # should fail — blank lines, SQL comments, trailing whitespace, and:
+      #
+      #   \restrict / \unrestrict — a per-invocation RANDOM NONCE that newer
+      #   pg_dump emits to fence the dump against injection during restore. It
+      #   differs on every run by construction, so it is noise by definition,
+      #   not a schema difference. Measured on run 4960, the control: two dumps
+      #   of the SAME schema came back 1123 lines each and differed on exactly
+      #   these two lines and nothing else. That control is what licenses this
+      #   filter — it was observed to be the only false positive, rather than
+      #   assumed to be one.
+      # Column ORDER inside a CREATE TABLE is compared separately from column
+      # CONTENT, and only content is fatal.
+      #
+      # A table built by 87 migrations has its columns in ADD COLUMN order; the
+      # same table built in one shot has them in declaration order. That is a
+      # real and permanent difference which no baseline can erase — the
+      # operator's existing database keeps chain order forever, a fresh install
+      # gets model order — so a check that fails on it would never pass and
+      # would teach nothing. FC reaches every column through the ORM by name,
+      # and `SELECT *` ordering is not depended on anywhere.
+      #
+      # So the second pass SORTS the column lines within each CREATE TABLE
+      # rather than DELETING them. That distinction is the whole point: sorting
+      # cannot hide a column that exists on one side only, or one whose type,
+      # nullability or default differs — those still land in the diff. A filter
+      # could have hidden all three.
+      #
+      # Both diffs are reported. The ordered one is informational; the
+      # order-insensitive one is the verdict.
+      - name: Diff
+        run: |
+          set -eu
+          norm() {
+            grep -vE '^\s*(--|$)' "$1" \
+              | grep -vE '^\\(un)?restrict ' \
+              | sed 's/[[:space:]]*$//'
+          }
+          norm chain.sql > a.txt
+          norm baseline.sql > b.txt
+          echo "normalised: chain=$(wc -l < a.txt) lines, current=$(wc -l < b.txt) lines"
+
+          sort_table_columns() {
+            python3 - "$1" <<'PYEOF'
+          import re, sys
+
+          lines = open(sys.argv[1]).read().splitlines()
+          out, block = [], None
+          for line in lines:
+              if block is not None:
+                  # ');' on its own closes the CREATE TABLE body.
+                  if line.strip() == ");":
+                      out.extend(sorted(block))
+                      out.append(line)
+                      block = None
+                  else:
+                      # Drop the list comma before sorting. Only the LAST
+                      # column lacks one, so keeping it would make every
+                      # reordering look like a content change as well — the
+                      # comma is punctuation, and carries no schema meaning.
+                      block.append(line.rstrip().rstrip(","))
+                  continue
+              out.append(line)
+              if re.match(r"CREATE TABLE .*\($", line):
+                  block = []
+          if block is not None:  # unterminated body: emit it rather than drop it
+              out.extend(block)
+          print("\n".join(out))
+          PYEOF
+          }
+          sort_table_columns a.txt > a.sorted.txt
+          sort_table_columns b.txt > b.sorted.txt
+          test "$(wc -l < a.sorted.txt)" = "$(wc -l < a.txt)"
+          test "$(wc -l < b.sorted.txt)" = "$(wc -l < b.txt)"
+
+          if diff -u a.txt b.txt > schema.diff; then
+            echo "ORDERED DIFF: identical, column order included."
+          else
+            echo "ORDERED DIFF: $(grep -cE '^[+-]' schema.diff) changed lines (informational):"
+            cat schema.diff
+          fi
+          echo
+          echo "================================================================"
+          echo
+          if diff -u a.sorted.txt b.sorted.txt > sorted.diff; then
+            echo "SCHEMAS MATCH — every difference above is column ORDER alone."
+          else
+            echo "SCHEMAS DIFFER — $(grep -cE '^[+-]' sorted.diff) changed lines that are NOT ordering:"
+            cat sorted.diff
+            echo
+            echo "The baseline is wrong, not the database. Do not stamp."
+            exit 1
+          fi

.forgejo/workflows/build.yml

@@ -45,6 +45,36 @@ on:
         type: boolean
         default: false
 
+  # The base-image refresh (milestone 326 step 4, #3154).
+  #
+  # Skip-if-exists is keyed on OUR source, so an artifact whose source stops
+  # moving stops picking up base-image updates. `agent/` last changed
+  # 2026-07-17; every push since has correctly declined to rebuild it, which
+  # also means it will serve that day's `nvidia/cuda` layers forever. Nothing
+  # is wrong until it has been unchanged for months, which is precisely why
+  # this is a calendar trigger and not a condition on the push path.
+  #
+  # Weekly, Sunday 06:00 UTC. Away from CI-runner's Monday security sweep so
+  # the two are never diagnosing each other, and on the quietest day so a
+  # surprise rebuild is not competing with a push.
+  schedule:
+    - cron: '0 6 * * 0'
+
+# Which branch a run BUILDS, as opposed to which one triggered it.
+#
+# They are the same thing on every trigger but `schedule`. Forgejo registers a
+# cron from the DEFAULT branch — `dev` here — so a scheduled run arrives with
+# `github.ref` pointing at dev, and a refresh that rebuilt `:dev` would be
+# refreshing the one channel that gets rebuilt constantly anyway. Production is
+# `main` (rule 147), and `:latest` is the tag that goes stale.
+#
+# So the ref is decided once, here, and every checkout in the file takes it.
+# Deriving it per job invites the two halves to disagree: sign-extension would
+# derive dev's extension version while build-web bundled main's, and the
+# release download would 404 on a version that exists perfectly well.
+env:
+  BUILD_REF: ${{ github.event_name == 'schedule' && 'main' || github.ref }}
+
 # Requires repo secret RELEASE_TOKEN — a Forgejo PAT with scopes:
 #   - write:package, read:package  (for docker push to git.fabledsword.com)
 #   - write:release                (for ext-<version> release asset cache)
@@ -88,12 +118,43 @@ jobs:
     steps:
       - uses: actions/checkout@v4
         with:
+          # Not the triggering ref — see the `env:` block at the top. On a
+          # scheduled refresh this is `main`; on everything else it is the ref
+          # that fired, so this is a no-op on every ordinary path.
+          ref: ${{ env.BUILD_REF }}
           # Full history is load-bearing, not a convenience: the version this
           # job signs is derived from the commit TIME of the newest packaged
           # extension change. A depth-1 clone sees one commit and derives a
           # wrong, too-low value rather than failing (ci-requirements.md).
           fetch-depth: 0
 
+      # BUILD_REF is what makes a scheduled run build `main` rather than the
+      # branch its cron fired from — and it is read through the `env` context
+      # inside `with:`, which this runner is NOT known to evaluate. If it does
+      # not, checkout silently falls back to the triggering ref and the weekly
+      # refresh publishes DEV's source to `:latest`, which is production.
+      # Every lane would stay green; the first sign of it would be production
+      # running code that was never merged.
+      #
+      # So assert the checkout instead of trusting the expression. A red
+      # weekly job is a fine outcome. Shipping dev to production is not.
+      #
+      # `if:` reads the `github` context, which the runner demonstrably does
+      # evaluate — this file already gates steps on it — so the guard cannot
+      # be disabled by the same uncertainty it exists to cover.
+      - name: Guard — a scheduled run must have checked out main
+        if: github.event_name == 'schedule'
+        run: |
+          set -eu
+          BRANCH=$(git rev-parse --abbrev-ref HEAD)
+          echo "schedule: HEAD is $BRANCH ($(git rev-parse --short HEAD))"
+          if [ "$BRANCH" != "main" ]; then
+            echo "schedule: expected main, got '$BRANCH'." >&2
+            echo "schedule: BUILD_REF was not honoured by the runner." >&2
+            echo "schedule: refusing to publish a channel tag from it." >&2
+            exit 1
+          fi
+
       # The version is DERIVED, not read from the repo (milestone 271 step 4,
       # cut over 2026-08-27). `packaging.sh version` returns `YYYY.M.D.HHMM`
       # UTC — the commit TIME of the newest change to a PACKAGED extension
@@ -364,12 +425,30 @@ jobs:
     steps:
       - uses: actions/checkout@v4
         with:
+          # Not the triggering ref — see the `env:` block at the top. On a
+          # scheduled refresh this is `main`; on everything else it is the ref
+          # that fired, so this is a no-op on every ordinary path.
+          ref: ${{ env.BUILD_REF }}
           # Full history: this job RE-DERIVES the extension version rather than
           # being handed it, and a depth-1 clone derives a wrong, too-low value
           # rather than failing — which would 404 the download of a release
           # that exists perfectly well under its real name.
           fetch-depth: 0
 
+      # See sign-extension's copy for why this guard exists.
+      - name: Guard — a scheduled run must have checked out main
+        if: github.event_name == 'schedule'
+        run: |
+          set -eu
+          BRANCH=$(git rev-parse --abbrev-ref HEAD)
+          echo "schedule: HEAD is $BRANCH ($(git rev-parse --short HEAD))"
+          if [ "$BRANCH" != "main" ]; then
+            echo "schedule: expected main, got '$BRANCH'." >&2
+            echo "schedule: BUILD_REF was not honoured by the runner." >&2
+            echo "schedule: refusing to publish a channel tag from it." >&2
+            exit 1
+          fi
+
       # --- derived values, one line (milestone 313) ------------------------
       # These stopped being shadow output at step 3. `revision` decides
       # whether the build below runs at all and `version` is what the image
@@ -429,8 +508,30 @@ jobs:
           # everywhere). Operator-flagged 2026-06-01 after the first :c-<sha>
           # main-push build failed at this step.
           SHORT_SHA=$(printf '%s' "$GITHUB_SHA" | cut -c1-7)
-          # Mirrors build-web's tag list; see the comment there.
+
-          if [ "${GITHUB_REF##*/}" = "main" ]; then
+          # A scheduled refresh publishes the CHANNEL and nothing else
+          # (#3154). :c-<sha> for main's HEAD already exists and names the
+          # bytes that commit actually built; re-pushing it over refreshed
+          # base layers would break the one tag rule 145 makes immutable —
+          # and it is the rollback unit, so the breakage would surface on the
+          # day somebody needed it.
+          #
+          # The accepted consequence: between a refresh and the next main
+          # push, :latest and :c-<sha> point at different manifests. That is
+          # the design, not drift. They RE-CONVERGE on that push — it hits
+          # reuse (a refresh does not move fc.revision, because it does not
+          # touch the source), and the repoint step then writes the new
+          # :c-<sha> from the refreshed :latest. So the rollback unit ends up
+          # naming the bytes production is actually running, which is the
+          # property that matters.
+          #
+          # Checked BEFORE the ref test, not after: a scheduled run's
+          # GITHUB_REF is the default branch (dev), so the main test would
+          # never fire on it.
+          if [ "${GITHUB_EVENT_NAME:-}" = "schedule" ]; then
+            echo "tags=git.fabledsword.com/bvandeusen/fabledcurator:latest" >> "$GITHUB_OUTPUT"
+            echo "channel=main" >> "$GITHUB_OUTPUT"
+          elif [ "${GITHUB_REF##*/}" = "main" ]; then
             echo "tags=git.fabledsword.com/bvandeusen/fabledcurator:latest,git.fabledsword.com/bvandeusen/fabledcurator:c-${SHORT_SHA}" >> "$GITHUB_OUTPUT"
             echo "channel=main" >> "$GITHUB_OUTPUT"
           else
@@ -524,6 +625,10 @@ jobs:
           # this runner is known to evaluate. Read through env rather than
           # interpolated into the run block, same rule as release.yml's TAG.
           FORCE: ${{ github.event.inputs.force_build }}
+          # A scheduled refresh has to bypass reuse by construction: it
+          # rebuilds the SAME source, so fc.revision always matches and the
+          # check would skip every refresh there has ever been.
+          EVENT: ${{ github.event_name }}
         run: |
           set -eu
           DERIVED=$(sh scripts/artifacts.sh revision web)
@@ -570,6 +675,9 @@ jobs:
           if [ "${FORCE:-false}" = "true" ]; then
             echo "hit=false" >> "$GITHUB_OUTPUT"
             echo "reuse: force_build set — building regardless"
+          elif [ "${EVENT:-}" = "schedule" ]; then
+            echo "hit=false" >> "$GITHUB_OUTPUT"
+            echo "reuse: scheduled base refresh — building regardless"
           elif [ -n "$PUBLISHED" ] && [ "$PUBLISHED" = "$DERIVED" ]; then
             echo "hit=true" >> "$GITHUB_OUTPUT"
             echo "reuse: already published — skipping the build"
@@ -661,6 +769,37 @@ jobs:
           context: .
           file: Dockerfile
           push: true
+          # Re-resolve the FROM references against the registry instead of
+          # trusting whatever digest the cache was built against. This is the
+          # whole mechanism of the scheduled refresh (#3154): if the base tag
+          # moved, the FROM layer's cache key changes, every layer above it
+          # invalidates, and the image genuinely rebuilds.
+          #
+          # MEASURED on the first real fire, run 4934 (#3265): when the base
+          # did NOT move, the build is ~13s and every content step reports
+          # CACHED — but the channel tag STILL gets a new manifest digest.
+          # buildkit mints a fresh image config each run, so identical layers
+          # are republished under a new config blob. All three images moved
+          # that way on 2026-08-30 with nothing whatsoever changed in them.
+          #
+          # So a refresh currently rewrites :latest every Sunday whether or
+          # not there is anything new in it, and :c-<sha> is handed a new
+          # manifest to diverge from on the same cadence. Layers are shared,
+          # so the storage cost is a config blob; the cost that matters is
+          # that a digest change no longer MEANS anything. Tracked in #3265 —
+          # the likely fix is a deterministic SOURCE_DATE_EPOCH, which would
+          # make "same source, same bytes" true and turn the no-op case into
+          # a genuine no-op.
+          #
+          # What `pull` does NOT catch either: a Debian package update inside
+          # the `apt-get install` layer while the base tag itself stands
+          # still. The official python/cuda images rebuild with those updates
+          # baked in, so this is a lag rather than a hole; closing it needs
+          # `no-cache: true`, which is a much larger version of the same
+          # churn #3265 is about.
+          #
+          # Only on the schedule. An ordinary push wants the cached base.
+          pull: ${{ github.event_name == 'schedule' }}
           # ONE tag, the channel's. Every other tag is written by the step
           # below, registry-side. buildx here pushes the first tag to the
           # registry and then re-pushes the rest through the DOCKER driver,
@@ -783,6 +922,12 @@ jobs:
             ARGS="$ARGS -t $t"
           done
           unset IFS
+          #
+          # This is also the whole of the scheduled refresh's tag handling
+          # (#3154): a refresh's tag list is the channel tag alone, so SOURCE
+          # is the only entry, it gets excluded, and this step correctly does
+          # nothing. No `if:` on the step and no schedule special-case —
+          # excluding the source was already the right rule.
           if [ -z "$ARGS" ]; then
             echo "repoint: $SOURCE is the only tag for this channel and"
             echo "repoint: already holds this revision — nothing to write."
@@ -799,6 +944,10 @@ jobs:
     steps:
       - uses: actions/checkout@v4
         with:
+          # Not the triggering ref — see the `env:` block at the top. On a
+          # scheduled refresh this is `main`; on everything else it is the ref
+          # that fired, so this is a no-op on every ordinary path.
+          ref: ${{ env.BUILD_REF }}
           # Full history: this job derives its artifact's version from the
           # commit its shipped files last changed in (milestone 313). A
           # depth-1 clone cannot see that commit — it either derives a wrong,
@@ -806,6 +955,20 @@ jobs:
           # the build would otherwise notice.
           fetch-depth: 0
 
+      # See sign-extension's copy for why this guard exists.
+      - name: Guard — a scheduled run must have checked out main
+        if: github.event_name == 'schedule'
+        run: |
+          set -eu
+          BRANCH=$(git rev-parse --abbrev-ref HEAD)
+          echo "schedule: HEAD is $BRANCH ($(git rev-parse --short HEAD))"
+          if [ "$BRANCH" != "main" ]; then
+            echo "schedule: expected main, got '$BRANCH'." >&2
+            echo "schedule: BUILD_REF was not honoured by the runner." >&2
+            echo "schedule: refusing to publish a channel tag from it." >&2
+            exit 1
+          fi
+
       # --- derived values, one line (milestone 313) ------------------------
       # These stopped being shadow output at step 3. `revision` decides
       # whether the build below runs at all and `version` is what the image
@@ -843,8 +1006,12 @@ jobs:
           # everywhere). Operator-flagged 2026-06-01 after first :c-<sha>
           # main-push build failed at this step.
           SHORT_SHA=$(printf '%s' "$GITHUB_SHA" | cut -c1-7)
-          # Mirrors build-web's tag list; see the comment there.
+          # Mirrors build-web's tag list and its schedule handling; see
-          if [ "${GITHUB_REF##*/}" = "main" ]; then
+          # the comments there.
+          if [ "${GITHUB_EVENT_NAME:-}" = "schedule" ]; then
+            echo "tags=git.fabledsword.com/bvandeusen/fabledcurator-ml:latest" >> "$GITHUB_OUTPUT"
+            echo "channel=main" >> "$GITHUB_OUTPUT"
+          elif [ "${GITHUB_REF##*/}" = "main" ]; then
             echo "tags=git.fabledsword.com/bvandeusen/fabledcurator-ml:latest,git.fabledsword.com/bvandeusen/fabledcurator-ml:c-${SHORT_SHA}" >> "$GITHUB_OUTPUT"
             echo "channel=main" >> "$GITHUB_OUTPUT"
           else
@@ -921,6 +1088,10 @@ jobs:
           # this runner is known to evaluate. Read through env rather than
           # interpolated into the run block, same rule as release.yml's TAG.
           FORCE: ${{ github.event.inputs.force_build }}
+          # A scheduled refresh has to bypass reuse by construction: it
+          # rebuilds the SAME source, so fc.revision always matches and the
+          # check would skip every refresh there has ever been.
+          EVENT: ${{ github.event_name }}
         run: |
           set -eu
           DERIVED=$(sh scripts/artifacts.sh revision ml)
@@ -963,6 +1134,9 @@ jobs:
           if [ "${FORCE:-false}" = "true" ]; then
             echo "hit=false" >> "$GITHUB_OUTPUT"
             echo "reuse: force_build set — building regardless"
+          elif [ "${EVENT:-}" = "schedule" ]; then
+            echo "hit=false" >> "$GITHUB_OUTPUT"
+            echo "reuse: scheduled base refresh — building regardless"
           elif [ -n "$PUBLISHED" ] && [ "$PUBLISHED" = "$DERIVED" ]; then
             echo "hit=true" >> "$GITHUB_OUTPUT"
             echo "reuse: already published — skipping the build"
@@ -978,6 +1152,37 @@ jobs:
           context: .
           file: Dockerfile.ml
           push: true
+          # Re-resolve the FROM references against the registry instead of
+          # trusting whatever digest the cache was built against. This is the
+          # whole mechanism of the scheduled refresh (#3154): if the base tag
+          # moved, the FROM layer's cache key changes, every layer above it
+          # invalidates, and the image genuinely rebuilds.
+          #
+          # MEASURED on the first real fire, run 4934 (#3265): when the base
+          # did NOT move, the build is ~13s and every content step reports
+          # CACHED — but the channel tag STILL gets a new manifest digest.
+          # buildkit mints a fresh image config each run, so identical layers
+          # are republished under a new config blob. All three images moved
+          # that way on 2026-08-30 with nothing whatsoever changed in them.
+          #
+          # So a refresh currently rewrites :latest every Sunday whether or
+          # not there is anything new in it, and :c-<sha> is handed a new
+          # manifest to diverge from on the same cadence. Layers are shared,
+          # so the storage cost is a config blob; the cost that matters is
+          # that a digest change no longer MEANS anything. Tracked in #3265 —
+          # the likely fix is a deterministic SOURCE_DATE_EPOCH, which would
+          # make "same source, same bytes" true and turn the no-op case into
+          # a genuine no-op.
+          #
+          # What `pull` does NOT catch either: a Debian package update inside
+          # the `apt-get install` layer while the base tag itself stands
+          # still. The official python/cuda images rebuild with those updates
+          # baked in, so this is a lag rather than a hole; closing it needs
+          # `no-cache: true`, which is a much larger version of the same
+          # churn #3265 is about.
+          #
+          # Only on the schedule. An ordinary push wants the cached base.
+          pull: ${{ github.event_name == 'schedule' }}
           # ONE tag, the channel's. Every other tag is written by the step
           # below, registry-side. buildx here pushes the first tag to the
           # registry and then re-pushes the rest through the DOCKER driver,
@@ -1113,6 +1318,10 @@ jobs:
     steps:
       - uses: actions/checkout@v4
         with:
+          # Not the triggering ref — see the `env:` block at the top. On a
+          # scheduled refresh this is `main`; on everything else it is the ref
+          # that fired, so this is a no-op on every ordinary path.
+          ref: ${{ env.BUILD_REF }}
           # Full history: this job derives its artifact's version from the
           # commit its shipped files last changed in (milestone 313). A
           # depth-1 clone cannot see that commit — it either derives a wrong,
@@ -1120,6 +1329,20 @@ jobs:
           # the build would otherwise notice.
           fetch-depth: 0
 
+      # See sign-extension's copy for why this guard exists.
+      - name: Guard — a scheduled run must have checked out main
+        if: github.event_name == 'schedule'
+        run: |
+          set -eu
+          BRANCH=$(git rev-parse --abbrev-ref HEAD)
+          echo "schedule: HEAD is $BRANCH ($(git rev-parse --short HEAD))"
+          if [ "$BRANCH" != "main" ]; then
+            echo "schedule: expected main, got '$BRANCH'." >&2
+            echo "schedule: BUILD_REF was not honoured by the runner." >&2
+            echo "schedule: refusing to publish a channel tag from it." >&2
+            exit 1
+          fi
+
       # --- derived values, one line (milestone 313) ------------------------
       # These stopped being shadow output at step 3. `revision` decides
       # whether the build below runs at all and `version` is what the image
@@ -1152,8 +1375,12 @@ jobs:
         id: tag
         run: |
           SHORT_SHA=$(printf '%s' "$GITHUB_SHA" | cut -c1-7)
-          # Mirrors build-web's tag list; see the comment there.
+          # Mirrors build-web's tag list and its schedule handling; see
-          if [ "${GITHUB_REF##*/}" = "main" ]; then
+          # the comments there.
+          if [ "${GITHUB_EVENT_NAME:-}" = "schedule" ]; then
+            echo "tags=git.fabledsword.com/bvandeusen/fabledcurator-agent:latest" >> "$GITHUB_OUTPUT"
+            echo "channel=main" >> "$GITHUB_OUTPUT"
+          elif [ "${GITHUB_REF##*/}" = "main" ]; then
             echo "tags=git.fabledsword.com/bvandeusen/fabledcurator-agent:latest,git.fabledsword.com/bvandeusen/fabledcurator-agent:c-${SHORT_SHA}" >> "$GITHUB_OUTPUT"
             echo "channel=main" >> "$GITHUB_OUTPUT"
           else
@@ -1230,6 +1457,10 @@ jobs:
           # this runner is known to evaluate. Read through env rather than
           # interpolated into the run block, same rule as release.yml's TAG.
           FORCE: ${{ github.event.inputs.force_build }}
+          # A scheduled refresh has to bypass reuse by construction: it
+          # rebuilds the SAME source, so fc.revision always matches and the
+          # check would skip every refresh there has ever been.
+          EVENT: ${{ github.event_name }}
         run: |
           set -eu
           DERIVED=$(sh scripts/artifacts.sh revision agent)
@@ -1272,6 +1503,9 @@ jobs:
           if [ "${FORCE:-false}" = "true" ]; then
             echo "hit=false" >> "$GITHUB_OUTPUT"
             echo "reuse: force_build set — building regardless"
+          elif [ "${EVENT:-}" = "schedule" ]; then
+            echo "hit=false" >> "$GITHUB_OUTPUT"
+            echo "reuse: scheduled base refresh — building regardless"
           elif [ -n "$PUBLISHED" ] && [ "$PUBLISHED" = "$DERIVED" ]; then
             echo "hit=true" >> "$GITHUB_OUTPUT"
             echo "reuse: already published — skipping the build"
@@ -1287,6 +1521,37 @@ jobs:
           context: agent
           file: agent/Dockerfile
           push: true
+          # Re-resolve the FROM references against the registry instead of
+          # trusting whatever digest the cache was built against. This is the
+          # whole mechanism of the scheduled refresh (#3154): if the base tag
+          # moved, the FROM layer's cache key changes, every layer above it
+          # invalidates, and the image genuinely rebuilds.
+          #
+          # MEASURED on the first real fire, run 4934 (#3265): when the base
+          # did NOT move, the build is ~13s and every content step reports
+          # CACHED — but the channel tag STILL gets a new manifest digest.
+          # buildkit mints a fresh image config each run, so identical layers
+          # are republished under a new config blob. All three images moved
+          # that way on 2026-08-30 with nothing whatsoever changed in them.
+          #
+          # So a refresh currently rewrites :latest every Sunday whether or
+          # not there is anything new in it, and :c-<sha> is handed a new
+          # manifest to diverge from on the same cadence. Layers are shared,
+          # so the storage cost is a config blob; the cost that matters is
+          # that a digest change no longer MEANS anything. Tracked in #3265 —
+          # the likely fix is a deterministic SOURCE_DATE_EPOCH, which would
+          # make "same source, same bytes" true and turn the no-op case into
+          # a genuine no-op.
+          #
+          # What `pull` does NOT catch either: a Debian package update inside
+          # the `apt-get install` layer while the base tag itself stands
+          # still. The official python/cuda images rebuild with those updates
+          # baked in, so this is a lag rather than a hole; closing it needs
+          # `no-cache: true`, which is a much larger version of the same
+          # churn #3265 is about.
+          #
+          # Only on the schedule. An ordinary push wants the cached base.
+          pull: ${{ github.event_name == 'schedule' }}
           # ONE tag, the channel's. Every other tag is written by the step
           # below, registry-side. buildx here pushes the first tag to the
           # registry and then re-pushes the rest through the DOCKER driver,

alembic/versions/0088_reconcile_models_with_schema.py

@@ -0,0 +1,128 @@
+"""Reconcile the database with what the models have always claimed (#3275).
+
+Milestone 328 discovered ~130 places where the ORM models and the deployed
+schema disagreed. Almost all of them were the MODEL being wrong — missing
+`server_default`s, indexes and CHECK constraints that only ever existed in a
+migration — and those are fixed in the model files with no DDL at all, because
+the database already had them.
+
+This migration carries the remainder — the two places where DDL is actually
+needed, because the database is what is wrong.
+
+`tag.fandom_id` is declared `index=True` on the model, but no migration ever
+created that index. Every autogenerate run since would have proposed adding
+it; nobody ran one, so the model and the database simply drifted apart and
+stayed that way.
+
+Deliberately NOT in this migration: anything about `image_record.sha256`. An
+earlier draft of this file claimed sha256 was not unique in the database and
+that duplicate rows were therefore possible. That was WRONG, and it was wrong
+because it was read off `op.create_index("ix_image_record_sha256", ...)` at
+0001 line 151 without reading line 149 two lines above it:
+
+    sa.UniqueConstraint("sha256", name="uq_image_record_sha256"),
+
+Uniqueness has been enforced since the initial schema. The database simply
+expresses it as a CONSTRAINT plus a separate non-unique lookup index, where
+the model expressed it as one `unique=True, index=True` column — the same
+guarantee built from different objects, which is why the two schemas did not
+line up. The model now declares the constraint and the plain index separately,
+so it describes what is actually there. No DDL is needed for it.
+
+Also here: six CHECK constraints whose names carry their table prefix TWICE.
+
+`base.py`'s naming convention is `ck_%(table_name)s_%(constraint_name)s`, and
+unlike the uq/fk/ix entries it applies even to a constraint that already has a
+name. Six migrations passed an already-prefixed name, so the convention
+prefixed it again:
+
+    ck_external_link_ck_external_link_host
+    ck_external_link_ck_external_link_status
+    ck_import_settings_ck_import_settings_singleton
+    ck_ml_settings_ck_ml_settings_singleton
+    ck_post_ck_post_translation_override
+    ck_tag_ck_tag_fandom_requires_character
+
+Nothing reads a CHECK constraint by name, so this has never done any harm —
+but it is exactly the development-era residue the collapsed baseline exists to
+leave behind, and a public schema should not ship it. The models now declare
+bare names, which the convention renders into the single-prefix form; this
+renames the deployed constraints to match.
+
+RENAME CONSTRAINT is a catalog-only operation: no table scan, no rewrite, no
+validation of existing rows. It takes a brief ACCESS EXCLUSIVE lock and
+returns. That is why this is safe to do on `post` and `tag`, which are the two
+large tables in the schema.
+
+Revision ID: 0088
+Revises: 0087
+Create Date: 2026-08-30
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+
+revision: str = "0088"
+down_revision: Union[str, None] = "0087"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+# (table, doubled name, single-prefix name)
+#
+# Six, not the four a first read of the migrations turned up. The list that
+# settles it is the one extracted from the chain's pg_dump by matching
+# `ck_(\w+?)_ck_\1_` — reading the migrations by eye missed external_link
+# twice over, in the same way an earlier pass missed a UNIQUE constraint two
+# lines above the index it was looking at (see the sha256 note above).
+DOUBLED_CHECKS = (
+    ("external_link", "ck_external_link_ck_external_link_host",
+     "ck_external_link_host"),
+    ("external_link", "ck_external_link_ck_external_link_status",
+     "ck_external_link_status"),
+    ("import_settings", "ck_import_settings_ck_import_settings_singleton",
+     "ck_import_settings_singleton"),
+    ("ml_settings", "ck_ml_settings_ck_ml_settings_singleton",
+     "ck_ml_settings_singleton"),
+    ("post", "ck_post_ck_post_translation_override",
+     "ck_post_translation_override"),
+    ("tag", "ck_tag_ck_tag_fandom_requires_character",
+     "ck_tag_fandom_requires_character"),
+)
+
+
+def _rename_check(table: str, old: str, new: str) -> None:
+    # Guarded on pg_constraint rather than run bare: a database built from the
+    # models (a fresh install, or the CI integration schema) already has the
+    # single-prefix name, and this migration must be a no-op there rather than
+    # an error. Same reasoning as the CREATE INDEX IF NOT EXISTS below.
+    op.execute(
+        f"""
+        DO $$
+        BEGIN
+            IF EXISTS (
+                SELECT 1 FROM pg_constraint
+                WHERE conname = '{old}' AND conrelid = '{table}'::regclass
+            ) THEN
+                ALTER TABLE {table} RENAME CONSTRAINT {old} TO {new};
+            END IF;
+        END $$;
+        """
+    )
+
+
+def upgrade() -> None:
+    # IF NOT EXISTS because the index is what the model already asks for: any
+    # database built from metadata rather than from this chain will have it,
+    # and this migration must be a no-op there rather than an error.
+    op.execute("CREATE INDEX IF NOT EXISTS ix_tag_fandom_id ON tag (fandom_id)")
+
+    for table, old, new in DOUBLED_CHECKS:
+        _rename_check(table, old, new)
+
+
+def downgrade() -> None:
+    for table, old, new in DOUBLED_CHECKS:
+        _rename_check(table, new, old)
+
+    op.execute("DROP INDEX IF EXISTS ix_tag_fandom_id")

alembic/versions/0089_index_hygiene.py

@@ -0,0 +1,120 @@
+"""Index the seven unindexed FKs; drop the seven redundant indexes (#3300, #3301).
+
+Found by a structural sweep of the deployed schema done AFTER 0088 brought the
+models and the migration chain into exact agreement. That agreement is what
+0088 achieved, and it is worth being precise about what it does NOT prove: a
+models-vs-chain diff shows the two describe the same schema. It says nothing
+about whether that schema is right. Everything here was wrong in BOTH, which is
+exactly the class of problem the reconciliation could not see.
+
+## Added: seven FK indexes
+
+`image_tag.tag_id` is the one that matters. The table's only index is
+PRIMARY KEY (image_record_id, tag_id), which leads with the wrong column for
+the two hottest things done with it:
+
+  * the gallery's tag filter — services/tag_query.py builds
+    `image_tag.c.tag_id == tid` (and `.in_(tids)`) on every tag-scoped browse;
+  * ON DELETE CASCADE from `tag` — deleting or merging a tag makes Postgres
+    find that tag's rows before it can remove them.
+
+Both had to scan the largest table in the schema. The other six are the same
+shape on much smaller tables; `presentation_review.tag_id` is the notable one,
+since it also CASCADEs.
+
+## Dropped: seven redundant indexes
+
+`ix_image_record_sha256` was an exact duplicate. A UNIQUE constraint builds its
+own index, so `uq_image_record_sha256` already covered the column and
+`image_record` carried two btrees on `sha256` — on the highest-insert-rate
+table in the system.
+
+The other six are single-column indexes that a later composite superseded
+without the narrow one being retired. A btree on (a, b) already serves lookups
+on `a`, so each was pure write amplification. `task_run` and `backup_run` are
+append-heavy operational logs, which is where that cost lands hardest.
+
+Note for anyone reading 0088 next to this: 0088 deliberately taught the models
+to declare BOTH sha256 indexes, so they would describe reality. That was right.
+This migration changes the reality instead, and the models change with it.
+
+## CONCURRENTLY, and why this migration has no transaction
+
+`CREATE INDEX` takes an ACCESS EXCLUSIVE lock for the whole build, which on
+`image_tag` means stalling every write for as long as it takes. CONCURRENTLY
+builds without blocking writers, at the cost of two table passes and an
+inability to run inside a transaction — hence `autocommit_block()`.
+
+The consequence to know about: this migration is NOT atomic. If it fails
+partway, the work already done stays done. Every statement is therefore written
+IF NOT EXISTS / IF EXISTS so that re-running it after a failure is safe rather
+than an error.
+
+A failed CONCURRENTLY build also leaves an INVALID index behind — it is not
+used by the planner and not repaired automatically. Find them with:
+
+    SELECT c.relname FROM pg_index i
+    JOIN pg_class c ON c.oid = i.indexrelid
+    WHERE NOT i.indisvalid;
+
+Drop what that returns and re-run; nothing else is needed.
+
+Revision ID: 0089
+Revises: 0088
+Create Date: 2026-08-31
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+
+revision: str = "0089"
+down_revision: Union[str, None] = "0088"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+# (index name, table, column) — names match what the models render under
+# base.py's naming convention, so autogenerate stays quiet after this.
+MISSING_FK_INDEXES = (
+    ("ix_image_tag_tag_id", "image_tag", "tag_id"),
+    ("ix_presentation_review_tag_id", "presentation_review", "tag_id"),
+    ("ix_presentation_review_conflict_tag_id", "presentation_review", "conflict_tag_id"),
+    ("ix_import_task_result_image_id", "import_task", "result_image_id"),
+    ("ix_external_link_attachment_id", "external_link", "attachment_id"),
+    ("ix_character_prototype_region_id", "character_prototype", "region_id"),
+    ("ix_backup_run_restored_from_id", "backup_run", "restored_from_id"),
+)
+
+# (index name, table, column) — redundant; the second element of each pair in
+# the docstring is what still covers the column after the drop.
+REDUNDANT_INDEXES = (
+    ("ix_image_record_sha256", "image_record", "sha256"),
+    ("ix_backup_run_kind", "backup_run", "kind"),
+    ("ix_backup_run_status", "backup_run", "status"),
+    ("ix_task_run_queue", "task_run", "queue"),
+    ("ix_task_run_status", "task_run", "status"),
+    ("ix_task_run_task_name", "task_run", "task_name"),
+    ("ix_external_link_post_id", "external_link", "post_id"),
+)
+
+
+def upgrade() -> None:
+    with op.get_context().autocommit_block():
+        for name, table, column in MISSING_FK_INDEXES:
+            op.execute(
+                f"CREATE INDEX CONCURRENTLY IF NOT EXISTS {name} "
+                f"ON {table} ({column})"
+            )
+        for name, _table, _column in REDUNDANT_INDEXES:
+            op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {name}")
+
+
+def downgrade() -> None:
+    with op.get_context().autocommit_block():
+        for name, table, column in REDUNDANT_INDEXES:
+            op.execute(
+                f"CREATE INDEX CONCURRENTLY IF NOT EXISTS {name} "
+                f"ON {table} ({column})"
+            )
+        for name, _table, _column in MISSING_FK_INDEXES:
+            op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {name}")

backend/app/models/artist.py

@@ -27,10 +27,10 @@ class Artist(Base):
     notes: Mapped[str | None] = mapped_column(Text, nullable=True)
 
     # True once a Source is attached; flips false if all sources removed.
-    is_subscription: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False)
+    is_subscription: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False, server_default="false")
 
     # Per-artist scheduling overrides; null means "use global default".
-    auto_check: Mapped[bool] = mapped_column(Boolean, nullable=False, default=True)
+    auto_check: Mapped[bool] = mapped_column(Boolean, nullable=False, default=True, server_default="true")
     check_interval_seconds: Mapped[int | None] = mapped_column(Integer, nullable=True)
 
     created_at: Mapped[datetime] = mapped_column(

backend/app/models/backup_run.py

@@ -20,7 +20,7 @@ feedback_check_existing_enums):
 
 from datetime import datetime
 
-from sqlalchemy import JSON, BigInteger, DateTime, ForeignKey, Integer, String, Text
+from sqlalchemy import JSON, BigInteger, DateTime, ForeignKey, Index, Integer, String, Text, text
 from sqlalchemy.orm import Mapped, mapped_column
 
 from .base import Base
@@ -29,10 +29,21 @@ from .base import Base
 class BackupRun(Base):
     __tablename__ = "backup_run"
 
+
+    __table_args__ = (
+        # alembic 0017: reporting indexes, never declared on the model (#3275).
+        Index("ix_backup_run_kind_started", "kind", text("started_at DESC")),
+        Index("ix_backup_run_status_finished", "status", text("finished_at DESC")),
+        Index("ix_backup_run_tag_partial", "tag", postgresql_where=text("tag IS NOT NULL")),
+    )
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
-    kind: Mapped[str] = mapped_column(String(16), nullable=False, index=True)
+    # No index=True: ix_backup_run_kind_started (above) already leads with
+    # `kind`, so a single-column index on it was pure write cost (#3301).
+    kind: Mapped[str] = mapped_column(String(16), nullable=False)
     status: Mapped[str] = mapped_column(
-        String(16), nullable=False, default="pending", index=True,
+        # No index=True — ix_backup_run_status_finished leads with `status`.
+        String(16), nullable=False, default="pending",
+        server_default="pending",
     )
     tag: Mapped[str | None] = mapped_column(String(64), nullable=True, index=True)
     triggered_by: Mapped[str] = mapped_column(String(32), nullable=False)
@@ -49,7 +60,9 @@ class BackupRun(Base):
     manifest: Mapped[dict] = mapped_column(
         JSON, nullable=False, default=dict, server_default="{}",
     )
+    # Self-referential FK, unindexed until 0089 (#3300): SET NULL has to find
+    # the rows pointing at a deleted run before it can null them.
     restored_from_id: Mapped[int | None] = mapped_column(
         ForeignKey("backup_run.id", ondelete="SET NULL"),
-        nullable=True,
+        nullable=True, index=True,
     )

backend/app/models/character_prototype.py

@@ -40,8 +40,10 @@ class CharacterPrototype(Base):
     )
     # Provenance: the region this vector was copied from. SET NULL so pruning a
     # region doesn't delete the prototype mid-cycle (the next refresh reconciles).
+    # index=True added in 0089 — the FK was unindexed (#3300).
     region_id: Mapped[int | None] = mapped_column(
-        ForeignKey("image_region.id", ondelete="SET NULL"), nullable=True
+        ForeignKey("image_region.id", ondelete="SET NULL"), nullable=True,
+        index=True,
     )
 
 

backend/app/models/download_event.py

@@ -25,8 +25,8 @@ class DownloadEvent(Base):
         DateTime(timezone=True), nullable=False, server_default=func.now()
     )
     finished_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True), nullable=True)
-    bytes_downloaded: Mapped[int] = mapped_column(BigInteger, nullable=False, default=0)
+    bytes_downloaded: Mapped[int] = mapped_column(BigInteger, nullable=False, default=0, server_default="0")
-    files_count: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    files_count: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
     error: Mapped[str | None] = mapped_column(Text, nullable=True)
     metadata_: Mapped[dict] = mapped_column(
         "metadata", JSONB, nullable=False, default=dict,

backend/app/models/external_link.py

@@ -16,6 +16,7 @@ doesn't delete the link record).
 from datetime import datetime
 
 from sqlalchemy import (
+    CheckConstraint,
     DateTime,
     Float,
     ForeignKey,
@@ -38,15 +39,33 @@ STATUSES = ("pending", "downloading", "downloaded", "failed", "skipped", "dead")
 class ExternalLink(Base):
     __tablename__ = "external_link"
     __table_args__ = (
+        # alembic 0028 enum CHECKs. Rule 36 territory: a new host or status value
+        # needs its constraint swapped in the same migration (#3275).
+        CheckConstraint(
+            "host IN ('mega', 'gdrive', 'mediafire', 'dropbox', 'pixeldrain')",
+            # Bare name: Base.metadata's naming convention prepends
+            # ck_<table>_. Pre-prefixing it here doubles the prefix — see
+            # alembic 0088, which renames the four constraints that shipped
+            # that way (#3275).
+            name="host",
+        ),
+        CheckConstraint(
+            "status IN ('pending', 'downloading', 'downloaded', 'failed', 'skipped', 'dead')",
+            name="status",
+        ),
         # One row per (post, url). The full url (incl. #fragment) is the identity
         # — the same file linked twice in a post collapses to one row.
         Index("uq_external_link_post_url", "post_id", "url", unique=True),
         Index("ix_external_link_status", "status"),
+        # Unindexed FK (#3300).
+        Index("ix_external_link_attachment_id", "attachment_id"),
     )
 
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
+    # No index=True: uq_external_link_post_url (post_id, url) already leads
+    # with post_id (#3301).
     post_id: Mapped[int] = mapped_column(
-        ForeignKey("post.id", ondelete="CASCADE"), nullable=False, index=True
+        ForeignKey("post.id", ondelete="CASCADE"), nullable=False
     )
     artist_id: Mapped[int | None] = mapped_column(
         ForeignKey("artist.id", ondelete="SET NULL"), nullable=True, index=True

backend/app/models/gpu_job.py

@@ -50,7 +50,8 @@ class GpuJob(Base):
     # What to compute, e.g. 'ccip' (detect figures + CCIP-embed) or 'siglip_region'.
     task: Mapped[str] = mapped_column(String(32), nullable=False)
     status: Mapped[str] = mapped_column(
-        String(16), nullable=False, default="pending", index=True
+        String(16), nullable=False, default="pending", index=True,
+        server_default="pending",
     )
     # pending | leased | done | error
     lease_token: Mapped[str | None] = mapped_column(String(64), nullable=True)
@@ -60,7 +61,7 @@ class GpuJob(Base):
     lease_expires_at: Mapped[datetime | None] = mapped_column(
         DateTime(timezone=True), nullable=True
     )
-    attempts: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    attempts: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
     error: Mapped[str | None] = mapped_column(Text, nullable=True)
     # Triage verdict for an ERRORED job (#125): NULL = not yet probed;
     # 'defect' = the integrity probe says the FILE itself is bad (surfaced for

backend/app/models/head_auto_apply_run.py

@@ -24,10 +24,11 @@ class HeadAutoApplyRun(Base):
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
     # dry_run=True is a PREVIEW: scores + counts what WOULD apply, writes nothing
     # (preview/apply parity, rule 93).
-    dry_run: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False)
+    dry_run: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False, server_default="false")
     params: Mapped[dict[str, Any]] = mapped_column(JSONB, nullable=False)
     status: Mapped[str] = mapped_column(
-        String(16), nullable=False, default="running", index=True
+        String(16), nullable=False, default="running", index=True,
+        server_default="running",
     )
     # running | ready | error
     started_at: Mapped[datetime] = mapped_column(

backend/app/models/head_metric.py

@@ -24,9 +24,9 @@ class HeadMetric(Base):
         ForeignKey("tag.id", ondelete="CASCADE"), primary_key=True
     )
     # An auto-applied (source='head_auto') tag the operator later REMOVED.
-    n_misfires: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    n_misfires: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
     # A tag with a head that the operator added by HAND (the head missed it).
-    n_underfires: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    n_underfires: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
     updated_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()
     )

backend/app/models/head_metrics_snapshot.py

@@ -19,8 +19,14 @@ class HeadMetricsSnapshot(Base):
     __tablename__ = "head_metrics_snapshot"
 
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
-    tag_id: Mapped[int] = mapped_column(
+    # Nullable, matching alembic 0060, which declared this column without
-        ForeignKey("tag.id", ondelete="CASCADE"), index=True
+    # `nullable=False`. The model had it as `Mapped[int]` — NOT NULL — which
+    # was simply never true of the database (#3275). Left nullable rather than
+    # tightened: a snapshot of a tag that is later hard-deleted is a row worth
+    # keeping, and the FK is ON DELETE CASCADE, so tightening it would only
+    # change behaviour, not correct a bug.
+    tag_id: Mapped[int | None] = mapped_column(
+        ForeignKey("tag.id", ondelete="CASCADE"), nullable=True, index=True
     )
     # Denormalized so a snapshot stays readable even if the tag is later renamed.
     name: Mapped[str] = mapped_column(String(255), nullable=False)
@@ -28,9 +34,9 @@ class HeadMetricsSnapshot(Base):
         DateTime(timezone=True), nullable=False, server_default=func.now(), index=True
     )
     # Current count of source='head_auto' applications still standing.
-    n_auto_applied: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    n_auto_applied: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
-    n_misfires: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    n_misfires: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
-    n_underfires: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    n_underfires: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
     # The head's measured quality at snapshot time (null if no head exists).
     ap: Mapped[float | None] = mapped_column(Float, nullable=True)
     precision_cv: Mapped[float | None] = mapped_column(Float, nullable=True)

backend/app/models/head_training_run.py

@@ -24,7 +24,8 @@ class HeadTrainingRun(Base):
     # Training parameters: {min_positives, neg_ratio, precision_target, ...}.
     params: Mapped[dict[str, Any]] = mapped_column(JSONB, nullable=False)
     status: Mapped[str] = mapped_column(
-        String(16), nullable=False, default="running", index=True
+        String(16), nullable=False, default="running", index=True,
+        server_default="running",
     )
     # running | ready | error
     started_at: Mapped[datetime] = mapped_column(

backend/app/models/image_provenance.py

@@ -47,8 +47,15 @@ class ImageProvenance(Base):
     # attachment on the post. NULL for loose downloads and pre-backfill rows.
     # SET NULL so deleting the archive attachment never destroys the (image,
     # post) edge — it just forgets which archive it came from.
+    # FK named explicitly: the convention renders this
+    # `fk_image_provenance_from_attachment_id_post_attachment`, but alembic
+    # 0055 created it as `fk_image_provenance_from_attachment` (#3275).
     from_attachment_id: Mapped[int | None] = mapped_column(
-        ForeignKey("post_attachment.id", ondelete="SET NULL"),
+        ForeignKey(
+            "post_attachment.id",
+            ondelete="SET NULL",
+            name="fk_image_provenance_from_attachment",
+        ),
         nullable=True, index=True,
     )
     captured_metadata: Mapped[dict | None] = mapped_column(JSON, nullable=True)

backend/app/models/image_record.py

@@ -14,10 +14,13 @@ from sqlalchemy import (
     Enum,
     Float,
     ForeignKey,
+    Index,
     Integer,
     String,
     Text,
+    UniqueConstraint,
     func,
+    text,
 )
 from sqlalchemy.orm import Mapped, mapped_column
 
@@ -29,11 +32,38 @@ ORIGIN_CHOICES = ("downloaded", "imported_filesystem", "uploaded")
 class ImageRecord(Base):
     __tablename__ = "image_record"
 
+
+    __table_args__ = (
+        # alembic 0001. The database enforces sha256 uniqueness with a
+        # CONSTRAINT and carries a SEPARATE non-unique btree index; the model
+        # said `unique=True, index=True`, which collapses both into a single
+        # UNIQUE index under a different name. Same guarantee either way, but
+        # not the same objects, so autogenerate saw a drop and an add (#3275).
+        UniqueConstraint("sha256", name="uq_image_record_sha256"),
+        # alembic 0036, and the last thing in this schema that lived only in a
+        # migration. SQLAlchemy CAN express an hnsw index with an operator
+        # class, so there is no reason for it to be invisible to the models —
+        # and its absence was the quietest failure of the lot: everything
+        # works, similarity search just silently stops using an index.
+        Index(
+            "ix_image_record_siglip_hnsw",
+            "siglip_embedding",
+            postgresql_using="hnsw",
+            postgresql_ops={"siglip_embedding": "vector_cosine_ops"},
+        ),
+        # alembic 0035/0071: the date-ordered browse indexes (#3275).
+        Index("ix_image_record_effective_date", text("effective_date DESC"), text("id DESC")),
+        Index("ix_image_record_earliest_post_date", text("earliest_post_date DESC"), text("id DESC")),
+    )
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
 
     # On-disk identity
     path: Mapped[str] = mapped_column(Text, nullable=False, unique=True)
-    sha256: Mapped[str] = mapped_column(String(64), nullable=False, unique=True, index=True)
+    # Neither unique= nor index=: uq_image_record_sha256 in __table_args__
+    # above creates its own index, and the separate ix_image_record_sha256
+    # 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)
     size_bytes: Mapped[int] = mapped_column(BigInteger, nullable=False)
     mime: Mapped[str] = mapped_column(String(64), nullable=False)
@@ -47,7 +77,8 @@ class ImageRecord(Base):
     # Integrity verification status. FC-2e populates this; FC-2a leaves rows at 'unknown'.
     # Values: 'unknown' (default), 'ok', 'corrupt', 'failed_verification'.
     integrity_status: Mapped[str] = mapped_column(
-        String(24), nullable=False, default="unknown", index=True
+        String(24), nullable=False, default="unknown", index=True,
+        server_default="unknown",
     )
 
     # Thumbnail (populated by FC-2)
@@ -72,8 +103,15 @@ class ImageRecord(Base):
     )
     # FC-2d-vii-c: canonical per-image artist (the single source of truth
     # for attribution; provenance posts remain lineage detail).
+    # FK named explicitly: the naming convention renders this
+    # `fk_image_record_artist_id_artist`, but alembic 0008 created it as
+    # `fk_image_record_artist_id` (#3275).
     artist_id: Mapped[int | None] = mapped_column(
-        ForeignKey("artist.id", ondelete="SET NULL"), nullable=True, index=True
+        ForeignKey(
+            "artist.id", ondelete="SET NULL", name="fk_image_record_artist_id"
+        ),
+        nullable=True,
+        index=True,
     )
 
     # ML fields (populated by the ml-worker / GPU agent). 1152 = SigLIP-so400m

backend/app/models/import_batch.py

@@ -21,17 +21,17 @@ class ImportBatch(Base):
     )
     finished_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True), nullable=True)
 
-    total_files: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    total_files: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
-    imported: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    imported: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
-    skipped: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    skipped: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
-    failed: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    failed: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
-    attachments: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    attachments: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
     # Deep-scan only: count of already-imported files whose sidecar metadata
     # got re-applied this run (post/source/provenance upsert). Stays 0 on
     # quick-scan batches. See `Importer.import_one(deep_scan=True)`.
-    refreshed: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    refreshed: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
 
-    status: Mapped[str] = mapped_column(String(16), nullable=False, default="running", index=True)
+    status: Mapped[str] = mapped_column(String(16), nullable=False, default="running", index=True, server_default="running")
     # running | complete | cancelled
 
     tasks = relationship("ImportTask", back_populates="batch", cascade="all, delete-orphan")

backend/app/models/import_settings.py

@@ -4,7 +4,15 @@ Enforced as a single row via a CHECK (id = 1) constraint. The application
 always SELECTs id=1 and never inserts/deletes after the initial migration.
 """
 
-from sqlalchemy import Boolean, CheckConstraint, Float, Integer, Text, select
+from sqlalchemy import (
+    Boolean,
+    CheckConstraint,
+    Float,
+    Integer,
+    Text,
+    select,
+    text,
+)
 from sqlalchemy.orm import Mapped, mapped_column
 
 from .base import Base
@@ -14,63 +22,79 @@ class ImportSettings(Base):
     __tablename__ = "import_settings"
     # Bare constraint name — Base.metadata's naming convention applies the
     # ck_<table>_<name> prefix, producing the final ck_import_settings_singleton.
+    # Bare name — Base.metadata's naming convention prepends ck_<table>_,
+    # producing ck_import_settings_singleton. The chain shipped the DOUBLED
+    # ck_import_settings_ck_import_settings_singleton, because the migration
+    # pre-prefixed the name and the convention prefixed it again; alembic
+    # 0088 renames it to what this line has always produced (#3275).
     __table_args__ = (CheckConstraint("id = 1", name="singleton"),)
 
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
-    import_scan_path: Mapped[str] = mapped_column(Text, nullable=False, default="/import")
+    import_scan_path: Mapped[str] = mapped_column(Text, nullable=False, default="/import", server_default="/import")
 
-    min_width: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    min_width: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
-    min_height: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    min_height: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
 
-    skip_transparent: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False)
+    skip_transparent: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False, server_default="false")
-    transparency_threshold: Mapped[float] = mapped_column(Float, nullable=False, default=0.9)
+    transparency_threshold: Mapped[float] = mapped_column(Float, nullable=False, default=0.9, server_default="0.9")
 
-    skip_single_color: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False)
+    skip_single_color: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False, server_default="false")
-    single_color_threshold: Mapped[float] = mapped_column(Float, nullable=False, default=0.95)
+    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)
+    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)
+    phash_threshold: Mapped[int] = mapped_column(Integer, nullable=False, default=10, server_default="10")
 
     # FC-3c downloader knobs
     download_rate_limit_seconds: Mapped[float] = mapped_column(
-        Float, nullable=False, default=3.0
+        Float, nullable=False, default=3.0,
+        server_default="3",
     )
     download_validate_files: Mapped[bool] = mapped_column(
-        Boolean, nullable=False, default=True
+        Boolean, nullable=False, default=True,
+        server_default="true",
     )
 
     # FC-3d scheduling knobs
     download_schedule_default_seconds: Mapped[int] = mapped_column(
-        Integer, nullable=False, default=28800
+        Integer, nullable=False, default=28800,
+        server_default="28800",
     )
     download_event_retention_days: Mapped[int] = mapped_column(
-        Integer, nullable=False, default=90
+        Integer, nullable=False, default=90,
+        server_default="90",
     )
     download_failure_warning_threshold: Mapped[int] = mapped_column(
-        Integer, nullable=False, default=5
+        Integer, nullable=False, default=5,
+        server_default="5",
     )
 
     # FC-3h backup knobs.
     backup_db_nightly_enabled: Mapped[bool] = mapped_column(
         Boolean, nullable=False, default=False,
+        server_default="false",
     )
     backup_db_nightly_hour_utc: Mapped[int] = mapped_column(
         Integer, nullable=False, default=3,
+        server_default="3",
     )
     backup_db_keep_last_n: Mapped[int] = mapped_column(
         Integer, nullable=False, default=14,
+        server_default="14",
     )
     backup_images_keep_last_n: Mapped[int] = mapped_column(
         Integer, nullable=False, default=3,
+        server_default="3",
     )
 
     # FC-6.3 series continuation matcher. enabled gates the rescan; threshold is
     # the weighted-score cut-off (0..1) above which a pending suggestion is made.
     series_suggest_enabled: Mapped[bool] = mapped_column(
         Boolean, nullable=False, default=True,
+        server_default="true",
     )
     series_suggest_threshold: Mapped[float] = mapped_column(
         Float, nullable=False, default=0.5,
+        server_default="0.5",
     )
 
     # #830 off-platform file-host downloads — per-host enable lever (default on,
@@ -113,7 +137,9 @@ class ImportSettings(Base):
     # English (e.g. "… WIP Part 1") as a European language at ~0.86. CJK stays
     # trusted regardless (script-detected). Per-post overrides handle the misses.
     translation_min_confidence: Mapped[float] = mapped_column(
-        Float, nullable=False, default=0.9, server_default="0.9",
+        # text() because alembic 0084 used sa.text(); see ml_settings for why
+        # the form matters and why it is per-column (#3275).
+        Float, nullable=False, default=0.9, server_default=text("0.9"),
     )
 
     # Title-based WIP auto-tagging (task #1458). When a freshly-imported post's

backend/app/models/import_task.py

@@ -13,10 +13,12 @@ from sqlalchemy import (
     Boolean,
     DateTime,
     ForeignKey,
+    Index,
     Integer,
     String,
     Text,
     func,
+    text,
 )
 from sqlalchemy.orm import Mapped, mapped_column, relationship
 
@@ -26,6 +28,12 @@ from .base import Base
 class ImportTask(Base):
     __tablename__ = "import_task"
 
+
+    __table_args__ = (
+        Index("ix_import_task_created_at_desc", text("created_at DESC")),
+        # Unindexed FK (#3300).
+        Index("ix_import_task_result_image_id", "result_image_id"),
+    )
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
     batch_id: Mapped[int] = mapped_column(
         ForeignKey("import_batch.id", ondelete="CASCADE"), nullable=False, index=True
@@ -33,14 +41,14 @@ class ImportTask(Base):
 
     source_path: Mapped[str] = mapped_column(Text, nullable=False)
     task_type: Mapped[str] = mapped_column(String(16), nullable=False)  # media|archive
-    status: Mapped[str] = mapped_column(String(16), nullable=False, default="pending", index=True)
+    status: Mapped[str] = mapped_column(String(16), nullable=False, default="pending", index=True, server_default="pending")
 
     # Poison-pill circuit breaker (alembic 0026). recovery_count tracks
     # how many times the stuck-task sweep has re-queued this row; after
     # the cap it's failed with a diagnostic instead of looping. refetched
     # bounds the one-shot re-download remediation to a single attempt.
-    recovery_count: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    recovery_count: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
-    refetched: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False)
+    refetched: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False, server_default="false")
 
     result_image_id: Mapped[int | None] = mapped_column(
         ForeignKey("image_record.id", ondelete="SET NULL"), nullable=True

backend/app/models/library_audit_run.py

@@ -8,7 +8,7 @@ reads it and routes through cleanup_service.delete_images.
 from datetime import datetime
 from typing import Any
 
-from sqlalchemy import DateTime, Integer, String, Text, func
+from sqlalchemy import DateTime, Integer, String, Text, func, text
 from sqlalchemy.dialects.postgresql import JSONB
 from sqlalchemy.orm import Mapped, mapped_column
 
@@ -23,6 +23,7 @@ class LibraryAuditRun(Base):
     params: Mapped[dict[str, Any]] = mapped_column(JSONB, nullable=False)
     status: Mapped[str] = mapped_column(
         String(16), nullable=False, default="running", index=True,
+        server_default="running",
     )
     # running | ready | applied | cancelled | error
     started_at: Mapped[datetime] = mapped_column(
@@ -31,14 +32,16 @@ class LibraryAuditRun(Base):
     finished_at: Mapped[datetime | None] = mapped_column(
         DateTime(timezone=True), nullable=True,
     )
-    scanned_count: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    scanned_count: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
-    matched_count: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    matched_count: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
-    matched_ids: Mapped[list[int]] = mapped_column(JSONB, nullable=False, default=list)
+    matched_ids: Mapped[list[int]] = mapped_column(
+        JSONB, nullable=False, default=list, server_default=text("'[]'::jsonb")
+    )
     error: Mapped[str | None] = mapped_column(Text, nullable=True)
     # Chunked-scan state (alembic 0039): keyset cursor the next chunk resumes
     # from, and the last time a chunk made progress (so the recovery sweep can
     # tell a progressing multi-chunk audit from a stuck one).
-    resume_after_id: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    resume_after_id: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
     last_progress_at: Mapped[datetime | None] = mapped_column(
         DateTime(timezone=True), nullable=True,
     )

backend/app/models/ml_settings.py

@@ -11,6 +11,7 @@ from sqlalchemy import (
     String,
     func,
     select,
+    text,
 )
 from sqlalchemy.orm import Mapped, mapped_column
 
@@ -20,7 +21,10 @@ from .base import Base
 class MLSettings(Base):
     __tablename__ = "ml_settings"
     # Bare name — Base.metadata's naming convention prepends ck_<table>_,
-    # producing the final ck_ml_settings_singleton (matches migration 0003).
+    # producing ck_ml_settings_singleton. The chain shipped the DOUBLED
+    # ck_ml_settings_ck_ml_settings_singleton, because the migration
+    # pre-prefixed the name and the convention prefixed it again; alembic
+    # 0088 renames it to what this line has always produced (#3275).
     __table_args__ = (CheckConstraint("id = 1", name="singleton"),)
 
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
@@ -31,17 +35,20 @@ class MLSettings(Base):
     # queueing embed work nothing will consume (the daily GPU 'embed' backfill
     # covers those images instead).
     cpu_embed_enabled: Mapped[bool] = mapped_column(
-        Boolean, nullable=False, default=True
+        Boolean, nullable=False, default=True,
+        server_default="true",
     )
     # Video embedding (#747). Sample one frame every N seconds (fixed CADENCE, not
     # a fixed count) so coverage reflects real screen time regardless of length;
     # cap the total so a long video can't explode into hundreds of embeds. The
     # per-frame SigLIP embeddings are mean-pooled. Operator-tunable.
     video_frame_interval_seconds: Mapped[float] = mapped_column(
-        Float, nullable=False, default=4.0
+        Float, nullable=False, default=4.0,
+        server_default="4",
     )
     video_max_frames: Mapped[int] = mapped_column(
-        Integer, nullable=False, default=64
+        Integer, nullable=False, default=64,
+        server_default="64",
     )
     # Tagging-v2 head training (#114). The head is the suggestion source that
     # LEARNS from the operator's tags (replacing Camie + centroid). A concept
@@ -49,10 +56,12 @@ class MLSettings(Base):
     # head_auto_apply_precision is the precision bar a head must clear (at some
     # operating point) to "graduate" into earned auto-apply. Operator-tunable.
     head_min_positives: Mapped[int] = mapped_column(
-        Integer, nullable=False, default=8
+        Integer, nullable=False, default=8,
+        server_default="8",
     )
     head_auto_apply_precision: Mapped[float] = mapped_column(
-        Float, nullable=False, default=0.97
+        Float, nullable=False, default=0.97,
+        server_default="0.97",
     )
     # Earned auto-apply (#114). A graduated head fires (tags images without a
     # human) when this master switch is on AND the head has at least
@@ -61,29 +70,34 @@ class MLSettings(Base):
     # default (operator-asked 2026-06-29: opt-OUT, not opt-in); the support +
     # measured-precision gates keep it safe, and every auto-tag is reversible.
     head_auto_apply_enabled: Mapped[bool] = mapped_column(
-        Boolean, nullable=False, default=True
+        Boolean, nullable=False, default=True,
+        server_default="true",
     )
     head_auto_apply_min_positives: Mapped[int] = mapped_column(
         # Support floor raised 30→50 (operator-asked 2026-07-06): a head needs
         # more human labels before it may fire without a human.
-        Integer, nullable=False, default=50
+        Integer, nullable=False, default=50,
+        server_default="30",
     )
     # CCIP character-match cosine cut (#114). 0.85 default — the v1 flat 0.75
     # over-fired (high-reference characters matched a scatter of images); 0.85
     # keeps the confident single-character matches. Tunable from the agent card.
     ccip_match_threshold: Mapped[float] = mapped_column(
-        Float, nullable=False, default=0.85
+        Float, nullable=False, default=0.85,
+        server_default="0.85",
     )
     # CCIP auto-apply (#114). Confident matches (>= ccip_auto_apply_threshold,
     # above the suggest cut) auto-tag on a daily sweep. ON by default (opt-out);
     # single-character references + the high bar keep it safe, every tag reversible.
     ccip_auto_apply_enabled: Mapped[bool] = mapped_column(
-        Boolean, nullable=False, default=True
+        Boolean, nullable=False, default=True,
+        server_default="true",
     )
     ccip_auto_apply_threshold: Mapped[float] = mapped_column(
         # Raised 0.92→0.95 (operator-asked 2026-07-06) so only very confident
         # character matches auto-tag.
-        Float, nullable=False, default=0.95
+        Float, nullable=False, default=0.95,
+        server_default="0.92",
     )
     # -- Presentation chrome auto-hide (#141) -------------------------------
     # `banner` (chrome — clusters on UI, not content) auto-applies on the sweep
@@ -95,13 +109,21 @@ class MLSettings(Base):
     # (opt-out); every auto-tag is reversible. NOTE (#1464): `wip` + `editor
     # screenshot` are no longer chrome — they went to the PROCESS path below.
     presentation_auto_apply_enabled: Mapped[bool] = mapped_column(
-        Boolean, nullable=False, default=True
+        Boolean, nullable=False, default=True,
+        server_default="true",
     )
     presentation_auto_apply_threshold: Mapped[float] = mapped_column(
-        Float, nullable=False, default=0.90
+        Float, nullable=False, default=0.90,
+        # text(), not a string, because alembic 0082 used sa.text(): a bare
+        # string renders DEFAULT '0.90'::double precision while text() renders
+        # DEFAULT 0.90, and the chain is MIXED — some migrations used one,
+        # some the other. Same value, different stored expression, so each
+        # column here mirrors whichever form its own migration used (#3275).
+        server_default=text("0.90"),
     )
     presentation_conflict_threshold: Mapped[float] = mapped_column(
-        Float, nullable=False, default=0.50
+        Float, nullable=False, default=0.50,
+        server_default=text("0.50"),
     )
     # -- Process auto-apply (#1464) ----------------------------------------
     # `wip` / `editor screenshot` are PROCESS art — unfinished pieces + program
@@ -115,24 +137,29 @@ class MLSettings(Base):
     # (PresentationReview, mode='process') rather than silently marked. OFF by
     # default — a new whole-library auto-tagger is opt-in; every auto-tag reversible.
     process_auto_apply_enabled: Mapped[bool] = mapped_column(
-        Boolean, nullable=False, default=False
+        Boolean, nullable=False, default=False,
+        server_default="false",
     )
     process_auto_apply_threshold: Mapped[float] = mapped_column(
-        Float, nullable=False, default=0.90
+        Float, nullable=False, default=0.90,
+        server_default="0.90",
     )
     process_conflict_threshold: Mapped[float] = mapped_column(
-        Float, nullable=False, default=0.50
+        Float, nullable=False, default=0.50,
+        server_default="0.50",
     )
     # Default = SigLIP 2 (so400m, 512px) for new installs (migration 0069);
     # existing libraries keep their stored value until the operator re-embeds.
     embedder_model_version: Mapped[str] = mapped_column(
-        String(128), nullable=False, default="siglip2-so400m-patch16-512"
+        String(128), nullable=False, default="siglip2-so400m-patch16-512",
+        server_default="siglip2-so400m-patch16-512",
     )
     # The HF model NAME the embedder loads (server CPU embed + announced to the
     # GPU agent in the lease). Operator-settable so the embedder is a choice, not
     # a hardcode (#1190): set name + version together, then re-embed + retrain.
     embedder_model_name: Mapped[str] = mapped_column(
-        String(128), nullable=False, default="google/siglip2-so400m-patch16-512"
+        String(128), nullable=False, default="google/siglip2-so400m-patch16-512",
+        server_default="google/siglip2-so400m-patch16-512",
     )
     # -- Crop proposers / detectors (#1202, #134) --------------------------
     # WHERE-to-crop YOLO detectors feeding the crop→SigLIP bag + CCIP. Config
@@ -145,20 +172,24 @@ class MLSettings(Base):
     # person: general COCO figure detector for Western/realistic art the anime
     # person-detector misses → NMS-merged with imgutils → CCIP + concept.
     detector_person_enabled: Mapped[bool] = mapped_column(
-        Boolean, nullable=False, default=True
+        Boolean, nullable=False, default=True,
+        server_default="true",
     )
     detector_person_weights: Mapped[str] = mapped_column(
-        String(512), nullable=False, default="yolo11n.pt"
+        String(512), nullable=False, default="yolo11n.pt",
+        server_default="yolo11n.pt",
     )
     detector_person_conf: Mapped[float] = mapped_column(
-        Float, nullable=False, default=0.35
+        Float, nullable=False, default=0.35,
+        server_default=text("0.35"),
     )
     # anatomy: booru_yolo anime/furry/NSFW torso components → concept crops.
     # Default = yolov11m_aa22 (26 classes, best mAP50-95 0.96), committed in the
     # upstream repo so the URL resolves. License UNSTATED — fine for a private
     # homelab (operator accepted #1202).
     detector_anatomy_enabled: Mapped[bool] = mapped_column(
-        Boolean, nullable=False, default=True
+        Boolean, nullable=False, default=True,
+        server_default="true",
     )
     detector_anatomy_weights: Mapped[str] = mapped_column(
         String(512), nullable=False,
@@ -166,37 +197,47 @@ class MLSettings(Base):
             "https://github.com/aperveyev/booru_yolo/raw/main/models/"
             "yolov11m_aa22.pt"
         ),
+        server_default="https://github.com/aperveyev/booru_yolo/raw/main/models/yolov11m_aa22.pt",
     )
     detector_anatomy_conf: Mapped[float] = mapped_column(
-        Float, nullable=False, default=0.30
+        Float, nullable=False, default=0.30,
+        server_default=text("0.30"),
     )
     # panel: comic page → panel regions → concept crops (Apache-2.0, YOLOv12x).
     detector_panel_enabled: Mapped[bool] = mapped_column(
-        Boolean, nullable=False, default=True
+        Boolean, nullable=False, default=True,
+        server_default="true",
     )
     detector_panel_weights: Mapped[str] = mapped_column(
         String(512), nullable=False,
         default="mosesb/best-comic-panel-detection::best.pt",
+        server_default="mosesb/best-comic-panel-detection::best.pt",
     )
     detector_panel_conf: Mapped[float] = mapped_column(
-        Float, nullable=False, default=0.30
+        Float, nullable=False, default=0.30,
+        server_default=text("0.30"),
     )
     # Per-frame caps bound the crop→embed explosion; max_regions is the hard
     # per-job backstop; dedupe_iou drops near-duplicate crops before the embed.
     detector_max_figures: Mapped[int] = mapped_column(
-        Integer, nullable=False, default=8
+        Integer, nullable=False, default=8,
+        server_default="8",
     )
     detector_max_components: Mapped[int] = mapped_column(
-        Integer, nullable=False, default=8
+        Integer, nullable=False, default=8,
+        server_default="8",
     )
     detector_max_panels: Mapped[int] = mapped_column(
-        Integer, nullable=False, default=8
+        Integer, nullable=False, default=8,
+        server_default="8",
     )
     detector_max_regions: Mapped[int] = mapped_column(
-        Integer, nullable=False, default=128
+        Integer, nullable=False, default=128,
+        server_default="128",
     )
     detector_dedupe_iou: Mapped[float] = mapped_column(
-        Float, nullable=False, default=0.85
+        Float, nullable=False, default=0.85,
+        server_default=text("0.85"),
     )
     # -- CCIP character prototypes (#1317) ---------------------------------
     # The per-character reference set is precomputed + refreshed INCREMENTALLY
@@ -208,7 +249,8 @@ class MLSettings(Base):
         String(128), nullable=True
     )
     ccip_prototype_cap: Mapped[int] = mapped_column(
-        Integer, nullable=False, default=64
+        Integer, nullable=False, default=64,
+        server_default="64",
     )
     updated_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()

backend/app/models/patreon_failed_media.py

@@ -35,7 +35,7 @@ class PatreonFailedMedia(Base):
         ForeignKey("source.id", ondelete="CASCADE"), nullable=False, index=True
     )
     filehash: Mapped[str] = mapped_column(String(128), nullable=False)
-    attempts: Mapped[int] = mapped_column(Integer, nullable=False, default=1)
+    attempts: Mapped[int] = mapped_column(Integer, nullable=False, default=1, server_default="1")
     last_error: Mapped[str | None] = mapped_column(Text, nullable=True)
     first_failed_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()

backend/app/models/pixiv_failed_media.py

@@ -35,7 +35,7 @@ class PixivFailedMedia(Base):
         ForeignKey("source.id", ondelete="CASCADE"), nullable=False, index=True
     )
     filehash: Mapped[str] = mapped_column(String(128), nullable=False)
-    attempts: Mapped[int] = mapped_column(Integer, nullable=False, default=1)
+    attempts: Mapped[int] = mapped_column(Integer, nullable=False, default=1, server_default="1")
     last_error: Mapped[str | None] = mapped_column(Text, nullable=True)
     first_failed_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()

backend/app/models/post.py

@@ -13,11 +13,13 @@ from sqlalchemy import (
     CheckConstraint,
     DateTime,
     ForeignKey,
+    Index,
     Integer,
     String,
     Text,
     UniqueConstraint,
     func,
+    text,
 )
 from sqlalchemy.orm import Mapped, mapped_column
 
@@ -27,6 +29,10 @@ from .base import Base
 class Post(Base):
     __tablename__ = "post"
     __table_args__ = (
+        # alembic 0030. The comment above described this index; nothing declared
+        # it, so autogenerate proposed dropping it (#3275).
+        Index("uq_post_artist_external_id_null_source", "artist_id", "external_post_id",
+              unique=True, postgresql_where=text("source_id IS NULL")),
         # Source-bound dedup. Postgres treats NULL != NULL so rows
         # with source_id IS NULL aren't deduped by this constraint;
         # the partial unique index `uq_post_artist_external_id_null_source`
@@ -35,7 +41,11 @@ class Post(Base):
         UniqueConstraint("source_id", "external_post_id", name="uq_post_source_external_id"),
         CheckConstraint(
             "translation_override IN ('auto', 'force', 'original')",
-            name="ck_post_translation_override",
+            # Bare name: Base.metadata's naming convention prepends
+            # ck_<table>_. Pre-prefixing it here doubles the prefix — see
+            # alembic 0088, which renames the four constraints that shipped
+            # that way (#3275).
+            name="translation_override",
         ),
     )
 

backend/app/models/presentation_review.py

@@ -11,7 +11,7 @@ are pruned by retention.
 
 from datetime import datetime
 
-from sqlalchemy import DateTime, Float, ForeignKey, String, func
+from sqlalchemy import DateTime, Float, ForeignKey, Index, String, func
 from sqlalchemy.orm import Mapped, mapped_column
 
 from .base import Base
@@ -20,6 +20,14 @@ from .base import Base
 class PresentationReview(Base):
     __tablename__ = "presentation_review"
 
+
+    __table_args__ = (
+        Index("ix_presentation_review_resolved_at", "resolved_at"),
+        # Both FKs to tag were unindexed (#3300); tag_id CASCADEs, so a tag
+        # delete had to scan this table to find its rows.
+        Index("ix_presentation_review_tag_id", "tag_id"),
+        Index("ix_presentation_review_conflict_tag_id", "conflict_tag_id"),
+    )
     image_record_id: Mapped[int] = mapped_column(
         ForeignKey("image_record.id", ondelete="CASCADE"), primary_key=True
     )

backend/app/models/series_chapter.py

@@ -16,7 +16,14 @@ title is the optional chapter name; stated_part is the optional operator-facing
 
 from datetime import datetime
 
-from sqlalchemy import DateTime, ForeignKey, Integer, Text, func
+from sqlalchemy import (
+    DateTime,
+    ForeignKey,
+    Integer,
+    Text,
+    UniqueConstraint,
+    func,
+)
 from sqlalchemy.orm import Mapped, mapped_column
 
 from .base import Base
@@ -25,14 +32,26 @@ from .base import Base
 class SeriesChapter(Base):
     __tablename__ = "series_chapter"
 
+    __table_args__ = (
+        # alembic 0047 named the UNIQUE `uq_series_chapter_anchor_page`, not
+        # the `uq_series_chapter_anchor_page_id` a bare `unique=True` would
+        # render (#3275).
+        UniqueConstraint("anchor_page_id", name="uq_series_chapter_anchor_page"),
+    )
+
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
     series_tag_id: Mapped[int] = mapped_column(
         ForeignKey("tag.id", ondelete="CASCADE"), nullable=False, index=True
     )
+    # Both the UNIQUE (above) and the FK carry the names 0047 gave them; the
+    # convention would render the FK `fk_series_chapter_anchor_page_id_series_page`.
     anchor_page_id: Mapped[int] = mapped_column(
-        ForeignKey("series_page.id", ondelete="CASCADE"),
+        ForeignKey(
+            "series_page.id",
+            ondelete="CASCADE",
+            name="fk_series_chapter_anchor_page",
+        ),
         nullable=False,
-        unique=True,
     )
     title: Mapped[str | None] = mapped_column(Text, nullable=True)
     stated_part: Mapped[int | None] = mapped_column(Integer, nullable=True)

backend/app/models/series_page.py

@@ -14,7 +14,14 @@ number parsed from the source post, nullable when unknown.
 
 from datetime import datetime
 
-from sqlalchemy import DateTime, ForeignKey, Integer, String, func
+from sqlalchemy import (
+    DateTime,
+    ForeignKey,
+    Integer,
+    String,
+    UniqueConstraint,
+    func,
+)
 from sqlalchemy.orm import Mapped, mapped_column
 
 from .base import Base
@@ -23,14 +30,22 @@ from .base import Base
 class SeriesPage(Base):
     __tablename__ = "series_page"
 
+    __table_args__ = (
+        # alembic 0005 named this `uq_series_page_image`; a bare `unique=True`
+        # on the column renders `uq_series_page_image_id` under the naming
+        # convention, which is a different object from the one the database
+        # has (#3275).
+        UniqueConstraint("image_id", name="uq_series_page_image"),
+    )
+
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
     series_tag_id: Mapped[int] = mapped_column(
         ForeignKey("tag.id", ondelete="CASCADE"), nullable=False, index=True
     )
+    # UNIQUE lives in __table_args__ above, under the name 0005 gave it.
     image_id: Mapped[int] = mapped_column(
         ForeignKey("image_record.id", ondelete="CASCADE"),
         nullable=False,
-        unique=True,
     )
     # 'placed' = in the series-global run (page_number set); 'pending' = staged
     # from a post awaiting the operator's sort (page_number NULL). (#789 P2)

backend/app/models/source.py

@@ -5,7 +5,16 @@ Multiple sources per artist support creators with cross-platform presence.
 
 from datetime import datetime
 
-from sqlalchemy import JSON, Boolean, DateTime, ForeignKey, Integer, String, Text
+from sqlalchemy import (
+    JSON,
+    Boolean,
+    DateTime,
+    ForeignKey,
+    Integer,
+    String,
+    Text,
+    UniqueConstraint,
+)
 from sqlalchemy.orm import Mapped, mapped_column, relationship
 
 from .base import Base
@@ -14,13 +23,27 @@ from .base import Base
 class Source(Base):
     __tablename__ = "source"
 
+    __table_args__ = (
+        # alembic 0010. One row per (artist, platform, url): re-adding a source
+        # the artist already has is an update, not a second row. The model had
+        # never declared it (#3275), so autogenerate would have proposed
+        # DROPPING it — the guarantee existed only in the migration chain.
+        #
+        # Named explicitly because the naming convention would render this
+        # `uq_source_artist_id` (uq keys off column_0_name), which is both
+        # wrong about the shape and not what the database actually has.
+        UniqueConstraint(
+            "artist_id", "platform", "url", name="uq_source_artist_platform_url"
+        ),
+    )
+
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
     artist_id: Mapped[int] = mapped_column(
         ForeignKey("artist.id", ondelete="CASCADE"), nullable=False, index=True
     )
     platform: Mapped[str] = mapped_column(String(64), nullable=False)
     url: Mapped[str] = mapped_column(Text, nullable=False)
-    enabled: Mapped[bool] = mapped_column(Boolean, nullable=False, default=True)
+    enabled: Mapped[bool] = mapped_column(Boolean, nullable=False, default=True, server_default="true")
 
     config_overrides: Mapped[dict | None] = mapped_column(JSON, nullable=True)
 
@@ -32,7 +55,7 @@ class Source(Base):
     # by _update_source_health alongside last_error; cleared on 'ok'.
     error_type: Mapped[str | None] = mapped_column(String(32), nullable=True, index=True)
     check_interval_override: Mapped[int | None] = mapped_column(Integer, nullable=True)
-    consecutive_failures: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    consecutive_failures: Mapped[int] = mapped_column(Integer, nullable=False, default=0, server_default="0")
 
     # alembic 0031: sticky deep-scan budget. When > 0, the next N download
     # runs use gallery-dl's full-walk config (skip: True + 1800s timeout);

backend/app/models/subscribestar_failed_media.py

@@ -34,7 +34,7 @@ class SubscribeStarFailedMedia(Base):
         ForeignKey("source.id", ondelete="CASCADE"), nullable=False, index=True
     )
     filehash: Mapped[str] = mapped_column(String(128), nullable=False)
-    attempts: Mapped[int] = mapped_column(Integer, nullable=False, default=1)
+    attempts: Mapped[int] = mapped_column(Integer, nullable=False, default=1, server_default="1")
     last_error: Mapped[str | None] = mapped_column(Text, nullable=True)
     first_failed_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()

backend/app/models/tag.py

@@ -15,11 +15,13 @@ from sqlalchemy import (
     Column,
     DateTime,
     ForeignKey,
+    Index,
     Integer,
     String,
     Table,
     false,
     func,
+    text,
 )
 from sqlalchemy import (
     Enum as SQLEnum,
@@ -67,17 +69,31 @@ image_tag = Table(
         primary_key=True,
     ),
     Column("tag_id", ForeignKey("tag.id", ondelete="CASCADE"), primary_key=True),
-    Column("source", String(32), nullable=False, default="manual"),
+    Column("source", String(32), nullable=False, default="manual", server_default="manual"),
     Column("created_at", DateTime(timezone=True), nullable=False, server_default=func.now()),
+    # The PK is (image_record_id, tag_id), which leads with the WRONG column
+    # for the two things that matter most here (#3300): the gallery's tag
+    # filter (tag_query.py builds `image_tag.c.tag_id == tid`) and the
+    # ON DELETE CASCADE from tag, which has to find a tag's rows to remove
+    # them. Without this index both scan the largest table in the schema.
+    Index("ix_image_tag_tag_id", "tag_id"),
 )
 
 
 class Tag(Base):
     __tablename__ = "tag"
     __table_args__ = (
+        # alembic 0002. An EXPRESSION index — COALESCE cannot be expressed as a
+        # UniqueConstraint, which is why it only ever existed in a migration (#3275).
+        Index("uq_tag_name_kind_fandom", "name", "kind", text("COALESCE(fandom_id, 0)"),
+              unique=True),
         CheckConstraint(
             "(fandom_id IS NULL) OR (kind = 'character')",
-            name="ck_tag_fandom_requires_character",
+            # Bare name: Base.metadata's naming convention prepends
+            # ck_<table>_. Pre-prefixing it here doubles the prefix — see
+            # alembic 0088, which renames the four constraints that shipped
+            # that way (#3275).
+            name="fandom_requires_character",
         ),
     )
 
@@ -87,6 +103,7 @@ class Tag(Base):
         SQLEnum(TagKind, name="tag_kind", values_callable=lambda e: [m.value for m in e]),
         nullable=False,
         default=TagKind.general,
+        server_default="general",
     )
     fandom_id: Mapped[int | None] = mapped_column(
         ForeignKey("tag.id", ondelete="SET NULL"), nullable=True, index=True

backend/app/models/tag_alias.py

@@ -5,7 +5,7 @@ in image_prediction stay unmolested.
 
 from datetime import datetime
 
-from sqlalchemy import DateTime, ForeignKey, String, func
+from sqlalchemy import DateTime, ForeignKey, Index, String, func
 from sqlalchemy.orm import Mapped, mapped_column
 
 from .base import Base
@@ -14,10 +14,17 @@ from .base import Base
 class TagAlias(Base):
     __tablename__ = "tag_alias"
 
+
+    __table_args__ = (
+        # Named explicitly: the database calls this ix_tag_alias_canonical, while
+        # a bare index=True on the column would generate ix_tag_alias_canonical_tag_id
+        # and silently propose a drop+create on the next autogenerate (#3275).
+        Index("ix_tag_alias_canonical", "canonical_tag_id"),
+    )
     alias_string: Mapped[str] = mapped_column(String(255), primary_key=True)
     alias_category: Mapped[str] = mapped_column(String(32), primary_key=True)
     canonical_tag_id: Mapped[int] = mapped_column(
-        ForeignKey("tag.id", ondelete="CASCADE"), nullable=False, index=True
+        ForeignKey("tag.id", ondelete="CASCADE"), nullable=False
     )
     created_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()

backend/app/models/tag_suggestion_rejection.py

@@ -5,7 +5,7 @@ Prevents re-suggestion AND prevents allowlist auto-apply on that image.
 
 from datetime import datetime
 
-from sqlalchemy import DateTime, ForeignKey, func
+from sqlalchemy import DateTime, ForeignKey, Index, func
 from sqlalchemy.orm import Mapped, mapped_column
 
 from .base import Base
@@ -14,11 +14,24 @@ from .base import Base
 class TagSuggestionRejection(Base):
     __tablename__ = "tag_suggestion_rejection"
 
+
+    __table_args__ = (
+        # Named explicitly; see tag_alias for why (#3275).
+        Index("ix_tag_suggestion_rejection_tag", "tag_id"),
+    )
+    # Both FKs named explicitly. alembic 0003 used a hand-shortened `tsr`
+    # prefix; the convention would render the full table name (#3275).
     image_record_id: Mapped[int] = mapped_column(
-        ForeignKey("image_record.id", ondelete="CASCADE"), primary_key=True
+        ForeignKey(
+            "image_record.id",
+            ondelete="CASCADE",
+            name="fk_tsr_image_record_id_image_record",
+        ),
+        primary_key=True,
     )
     tag_id: Mapped[int] = mapped_column(
-        ForeignKey("tag.id", ondelete="CASCADE"), primary_key=True, index=True
+        ForeignKey("tag.id", ondelete="CASCADE", name="fk_tsr_tag_id_tag"),
+        primary_key=True,
     )
     rejected_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, server_default=func.now()

backend/app/models/task_run.py

@@ -15,7 +15,7 @@ backend.app.tasks.maintenance.recover_stalled_task_runs (Beat 5 min).
 
 from datetime import datetime
 
-from sqlalchemy import DateTime, Integer, String, Text
+from sqlalchemy import DateTime, Index, Integer, String, Text, text
 from sqlalchemy.orm import Mapped, mapped_column
 
 from .base import Base
@@ -24,12 +24,21 @@ from .base import Base
 class TaskRun(Base):
     __tablename__ = "task_run"
 
+
+    __table_args__ = (
+        # alembic 0016: the three task-history indexes (#3275).
+        Index("ix_task_run_name_started", "task_name", text("started_at DESC")),
+        Index("ix_task_run_queue_started", "queue", text("started_at DESC")),
+        Index("ix_task_run_status_started", "status", text("started_at DESC")),
+    )
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
     celery_task_id: Mapped[str] = mapped_column(
         String(64), nullable=False, index=True,
     )
-    queue: Mapped[str] = mapped_column(String(32), nullable=False, index=True)
+    # Neither carries index=True: ix_task_run_queue_started and
-    task_name: Mapped[str] = mapped_column(String(128), nullable=False, index=True)
+    # ix_task_run_name_started already lead with these columns (#3301).
+    queue: Mapped[str] = mapped_column(String(32), nullable=False)
+    task_name: Mapped[str] = mapped_column(String(128), nullable=False)
     target_id: Mapped[int | None] = mapped_column(Integer, nullable=True)
     started_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), nullable=False, index=True,
@@ -39,7 +48,9 @@ class TaskRun(Base):
     )
     duration_ms: Mapped[int | None] = mapped_column(Integer, nullable=True)
     status: Mapped[str] = mapped_column(
-        String(16), nullable=False, default="running", index=True,
+        # No index=True — ix_task_run_status_started leads with `status`.
+        String(16), nullable=False, default="running",
+        server_default="running",
     )
     error_type: Mapped[str | None] = mapped_column(String(128), nullable=True)
     error_message: Mapped[str | None] = mapped_column(Text, nullable=True)

ci-requirements.md

@@ -167,6 +167,48 @@ per `docs/process.md`'s "add deps to the image when used by >1 project".
   `github.event.inputs` into an env var rather than interpolated into a run
   block, and it is checked inside the reuse step so that one decision drives
   both the build and the repoint.
+- **A weekly `schedule` rebuilds all three images against fresh base layers**
+  (Sunday 06:00 UTC, milestone 326 step 4, #3154). Skip-if-exists is keyed on
+  OUR source, so an artifact whose source stops moving stops picking up base
+  updates — `agent/` has not changed since 2026-07-17 and would otherwise serve
+  that day's `nvidia/cuda` layers forever. Four things make it work:
+  - It **builds `main`, not the branch that triggered it.** Forgejo registers a
+    cron from the DEFAULT branch (`dev` here), so a scheduled run arrives with
+    `github.ref` on dev. The ref is decided once in a top-level `env:
+    BUILD_REF` that every checkout in the file takes, rather than per job —
+    otherwise `sign-extension` would derive dev's extension version while
+    `build-web` bundled main's, and the release download would 404 on a version
+    that exists perfectly well. Every job then ASSERTS its checkout is `main`
+    before doing anything, because `env` inside `with:` is not a context this
+    runner is known to evaluate — if it silently resolved to empty, checkout
+    would fall back to the triggering ref and the refresh would publish dev's
+    source to `:latest` with every lane green.
+  - It **publishes only `:latest`.** `:c-<sha>` for main's HEAD already names
+    the bytes that commit built; re-pushing it over refreshed layers would
+    break the one tag rule 145 makes immutable, and it is the rollback unit.
+    The repoint step needs no schedule case for this — the tag list is the
+    channel tag alone, so SOURCE is the only entry, it is excluded as always,
+    and the step correctly does nothing.
+  - **`:latest` and `:c-<sha>` therefore diverge between a refresh and the next
+    `main` push, by design.** They re-converge on that push: it hits reuse (a
+    refresh does not move `fc.revision`, because it does not touch the source),
+    and the repoint writes the NEW `:c-<sha>` from the refreshed `:latest`. The
+    push path needed no change for this, because the repoint already excluded
+    the source tag — the same rule that keeps the label readable also keeps a
+    refresh from being undone.
+  - **`pull: true` on the scheduled path only** is the mechanism: a moved base
+    tag changes the `FROM` layer's cache key and everything above it rebuilds.
+    **It does not currently make the unmoved case free.** Measured on the first
+    real fire (run 4934, 2026-08-30): every content step reported `CACHED` and
+    the bases resolved to unchanged digests, yet all three `:latest` tags got a
+    NEW manifest digest, because buildkit mints a fresh image config per run and
+    republishes identical layers under it. So `:latest` is rewritten weekly
+    whether or not anything changed, and `:c-<sha>` is handed a new manifest to
+    diverge from on the same cadence — a digest change stops meaning anything.
+    Tracked as #3265; the likely fix is a deterministic `SOURCE_DATE_EPOCH`.
+    Separately not caught: a Debian package update inside the `apt-get install`
+    layer while the base tag stands still — a lag rather than a hole, since the
+    official python/cuda images rebuild with those updates baked in.
 - **`FC_CHANNEL` and `FC_VERSION` are build args, not runtime settings.**
   `build.yml` passes them to the web image only — the ml and agent images have
   nothing to report them to. `/api/health` returns both, the foot of Settings