FabledCurator/.forgejo/workflows/build.yml

name: Build images

on:
  push:
    # `:dev` builds were dropped 2026-05-26 to save a docker build per dev
    # push, on the reasoning that "operator tests from `:latest` after
    # merge-to-main". Restored 2026-08-27: that is testing by shipping, and
    # family rules 146/147 now name it directly — `main` IS production, and a
    # channel that can only be refreshed by shipping is not a channel. The
    # pressure to merge in order to try something does not come from
    # carelessness; it comes from `:dev` being unable to carry the build.
    #
    # All three images build on dev, deliberately: a `:dev` web image paired
    # with a stale `:dev` ml or agent is a worse trap than no dev channel at
    # all, since the mismatch only shows up as a runtime failure.
    branches: [main, dev]
    #
    # NO tag trigger (milestone 318 step 2). A `v*` tag names a commit `main`
    # already built and published; rebuilding it produces the same source under
    # the same names and RE-PUSHES `:c-<sha>`, which rule 145 forbids even when
    # the bytes match — image configs carry timestamps, so "same source" does
    # not mean "same manifest". The release build was publishing nothing new
    # and violating an immutability rule to do it.
    #
    # Releases still happen (rule 148, on explicit request per rule 2). They
    # produce a changelog, not an image.

  # The escape hatch for the one thing skip-if-exists makes untestable: a
  # build that WOULD be skipped. `agent/` has not changed since 2026-07-17, so
  # every push since has correctly declined to build it — which also means the
  # agent build path has not run in six weeks and cannot be exercised on
  # demand. #3190 lives on exactly that path.
  #
  # Editing build.yml does not force one either, and that is deliberate: the
  # workflow is not shipped bytes, so it is in no artifact's path set. Putting
  # it in one would re-version every artifact for a comment change.
  #
  # ONE input, not one per artifact. Forcing all three is cheap once the
  # registry cache is warm (#3114), and three booleans is an interface nobody
  # remembers the meaning of.
  workflow_dispatch:
    inputs:
      force_build:
        description: 'Rebuild every image even if the published revision matches'
        type: boolean
        default: false
      refresh:
        description: 'Behave as the weekly base refresh: build main against fresh bases, publish through the candidate tag'
        type: boolean
        default: false

  # The base-image refresh (milestone 326 step 4, #3154).
  #
  # 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.
# IS THIS A BASE REFRESH? Asked in five places and previously spelled five
# ways — `github.event_name == 'schedule'` in an `if:`, `$GITHUB_EVENT_NAME` in
# one shell, an `EVENT:` env passed into another, and a bare expression on
# `pull:`. Five spellings of one fact is how half of them come to disagree
# after somebody adds a sixth trigger.
#
# The `refresh` dispatch input is here so this path can be EXERCISED. A weekly
# cron is otherwise testable once a week, which is not a cadence anything can
# be developed against — the same reason `force_build` exists (#3252, added to
# confirm #3190 was gone rather than wait for it to recur). It is also what
# makes the milestone-362 gate verifiable at all: a gate has to be watched
# rejecting something before anyone can believe it is wired up.
#
# Note this is a STRING comparison, not a boolean. Forgejo delivers
# workflow_dispatch inputs as strings, so `inputs.refresh` is 'true'/'false'
# and `&&` on it would treat the string 'false' as truthy.
env:
  IS_REFRESH: ${{ (github.event_name == 'schedule' || github.event.inputs.refresh == 'true') && 'true' || 'false' }}
  BUILD_REF: ${{ (github.event_name == 'schedule' || github.event.inputs.refresh == 'true') && 'main' || github.ref }}

# Requires repo secret RELEASE_TOKEN — a Forgejo PAT with scopes:
#   - write:package, read:package  (for docker push to git.fabledsword.com)
#   - write:release                (for ext-<version> release asset cache)
#   - write:issue                  (for future issue-management automation)
# The injected GITHUB_TOKEN cannot be used — it lacks write:package.

jobs:
  # Sign-or-fetch-from-cache: signs the extension via AMO if no ext-<version>
  # Forgejo release exists yet, otherwise downloads the cached signed XPI.
  # Result is uploaded as an Actions artifact for build-web to consume.
  #
  # Why this lives in build.yml (not a separate workflow): the image a push
  # publishes MUST carry the XPI. A separate sign workflow racing build.yml
  # leaves that image without one for ~5min (until the commit-back triggers
  # another build). Inline ordering eliminates the race.
  # Cache strategy: Forgejo Release Assets — picked 2026-05-25 over Generic
  # Packages (cleaner API surface) and commit-back-to-side-branch (no extra
  # branch to manage). AMO blocks re-signing the same version (returns 409),
  # so signing is intentionally one-shot per version.
  #
  # BOTH branches sign (milestone 271 step 6, 2026-08-27). Not two signatures:
  # the version is the commit TIME of the newest packaged-extension change, so
  # dev and main derive the SAME number for the same extension source. A dev
  # push that changes the extension signs it; the merge to main then finds the
  # ext-<version> release already there, hits the cache, and bundles the
  # byte-identical XPI into `:latest` with no second AMO call. One signature
  # per extension CHANGE, shared by both channels — that is what makes two
  # channels affordable, and it is why step 4 (derived version) had to land
  # first. Ungating this while the version was still the hand-set 1.0.11 would
  # have hit the existing ext-1.0.11 cache and bundled MAIN's stale XPI into
  # `:dev` — a dev channel confidently serving old code.
  #
  # Unconditional since milestone 318 step 2: main and dev are now the only
  # triggers, so the branch gate that used to exclude tag pushes matched
  # everything. A condition that is always true reads as if some path avoids
  # it, which is worse than no condition.
  sign-extension:
    runs-on: python-ci
    container:
      image: git.fabledsword.com/bvandeusen/ci-python:3.14
    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: env.IS_REFRESH == 'true'
        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
      # file, per family rule 148/149. Never a commit count, which orders by
      # branch rather than by recency.
      #
      # Unpadded, and only here: AMO's grammar rejects a leading zero, so the
      # extension renders rule 148's numbers without the family's padding
      # (milestone 318 step 8). Same value, one character narrower per segment;
      # ci.yml's extension-version lane checks the string against Mozilla's
      # published regex before this job ever calls AMO.
      #
      # The committed "version" in manifest.json / package.json decides NOTHING
      # — not even a MAJOR.MINOR prefix, which step 8 removed. The stamp step
      # below overwrites it in the working tree before web-ext ever reads it,
      # and it is deliberately NOT committed back: the commit carrying the bump
      # would itself be a change to the extension and would move the version
      # again. The repo holds the source; the build derives the label.
      - name: Derive extension version
        id: extver
        run: |
          set -eu
          VERSION=$(sh extension/scripts/packaging.sh version)
          echo "version=$VERSION" >> "$GITHUB_OUTPUT"
          echo "Derived extension version: $VERSION"

      # Firefox refuses a downgrade and AMO never releases a burned version,
      # so a version that moves BACKWARDS is unrecoverable: it strands every
      # install that already took the higher one. Two ways it could happen —
      # a checkout without full history (derives too low), or a rewritten
      # history that drops the newest packaged commit.
      #
      # The test is `derived < highest already signed`, strictly. Equality is
      # the ORDINARY case, not a fault: an unchanged extension derives the same
      # version it did last build, which is exactly what lets the ext-<version>
      # cache hit and holds AMO to one call per extension CHANGE. Only moving
      # backwards is a failure, so this runs on every path — cache hit
      # included — rather than only before a sign.
      # --- shadow mode (milestone 313, step 2) -----------------------------
      # Informational ONLY. Nothing reads this and it must never fail the
      # build — no `set -e`, and every derivation falls back to UNAVAILABLE.
      #
      # What to watch across pushes, because this is what step 3 will trust:
      #   * a push touching only agent/ moves the agent and leaves web and ml
      #     STILL. If web moves, its path set is too wide.
      #   * a push touching only docs moves nothing.
      #   * a push touching the extension moves the extension AND web, since
      #     web bakes in the XPI. If web does not move, its set is too narrow:
      #     the reuse check hits, and the channel serves a web image bundling
      #     the PREVIOUS XPI while the freshly signed one is orphaned (#3156).
      #   * dev and main derive the same values for the same source.
      - name: Shadow — derived artifact version (informational)
        run: |
          set -u
          A=extension
          V=$(sh scripts/artifacts.sh version "$A" 2>&1 || echo UNAVAILABLE)
          R=$(sh scripts/artifacts.sh revision "$A" 2>&1 || echo UNAVAILABLE)
          echo "derived: artifact=$A version=$V revision=$R sha=$GITHUB_SHA"

      - name: Guard — the derived version must never go backwards
        env:
          TOKEN: ${{ secrets.RELEASE_TOKEN }}
          DERIVED: ${{ steps.extver.outputs.version }}
        run: |
          python3 - <<'PY'
          import json, os, sys, urllib.request

          API = ("https://git.fabledsword.com/api/v1/repos/"
                 "bvandeusen/FabledCurator/releases")
          headers = {"Authorization": "token " + os.environ["TOKEN"]}

          # Paginated rather than first-page-only: ext-* releases share this
          # list with the v* release tags, so one page would start missing them
          # as those accumulate. The bound FAILS rather than silently scanning
          # part of the list and calling the highest it saw the highest there is.
          tags = []
          for page in range(1, 21):
              req = urllib.request.Request(
                  f"{API}?limit=50&page={page}", headers=headers)
              with urllib.request.urlopen(req, timeout=30) as resp:
                  batch = json.load(resp)
              if not batch:
                  break
              tags += [r.get("tag_name", "") for r in batch]
          else:
              sys.exit("guard: >1000 releases — pagination bound reached")

          def parse(v):
              try:
                  return tuple(int(part) for part in v.split("."))
              except ValueError:
                  return None

          derived_s = os.environ["DERIVED"]
          derived = parse(derived_s)
          if derived is None:
              sys.exit(f"guard: derived version {derived_s!r} is not numeric")

          signed = sorted(
              (v, t) for t in tags if t.startswith("ext-")
              for v in [parse(t[4:])] if v
          )
          if not signed:
              print("guard: no ext-* release yet — nothing to go backwards from")
              raise SystemExit(0)

          hi, hi_tag = signed[-1]
          print(f"guard: derived={derived_s}  highest already signed={hi_tag}")
          if derived < hi:
              sys.exit(
                  f"REFUSING TO SIGN: derived {derived_s} is OLDER than the "
                  f"already-signed {hi_tag}. Firefox would reject it as a "
                  f"downgrade, and AMO will not release the burned version. "
                  f"First thing to check: did this job check out with "
                  f"fetch-depth: 0?"
              )
          print("guard: ok")
          PY

      - name: Check Forgejo release-asset cache
        id: cache
        env:
          TOKEN: ${{ secrets.RELEASE_TOKEN }}
        run: |
          set -eu
          VERSION=${{ steps.extver.outputs.version }}
          STATUS=$(curl -s -o release.json -w "%{http_code}" \
            -H "Authorization: token $TOKEN" \
            "https://git.fabledsword.com/api/v1/repos/bvandeusen/FabledCurator/releases/tags/ext-$VERSION" || echo 000)
          echo "Tag lookup HTTP status: $STATUS"
          # JSON parsing via python (ci-python:3.14 has stdlib json; jq is
          # not in the image and adding it per ci-requirements.md is not
          # warranted for a single consumer — operator-flagged 2026-05-26
          # after a sign job failed with `jq: not found`).
          if [ "$STATUS" = "200" ]; then
            ASSET_ID=$(python3 -c "import json; r=json.load(open('release.json')); xpis=[a for a in r.get('assets', []) if a.get('name','').endswith('.xpi')]; print(xpis[0]['id'] if xpis else '')")
            if [ -n "$ASSET_ID" ]; then
              echo "cached=true" >> "$GITHUB_OUTPUT"
              echo "asset_id=$ASSET_ID" >> "$GITHUB_OUTPUT"
              echo "Cached XPI exists at ext-$VERSION (asset id $ASSET_ID); skipping AMO sign"
            else
              echo "cached=false" >> "$GITHUB_OUTPUT"
              echo "Release ext-$VERSION exists but has no .xpi asset; will re-sign + re-upload"
            fi
          else
            echo "cached=false" >> "$GITHUB_OUTPUT"
            echo "No release named ext-$VERSION; will sign via AMO and upload"
          fi

      # No "download cached XPI in sign-extension" step: build-web
      # fetches directly from the Forgejo ext-<version> release asset
      # (removed 2026-05-26 alongside the actions/upload-artifact
      # removal — sign-extension's job is just to ensure the cache
      # exists on Forgejo; the build-web side reads it independently).

      # web-ext signs whatever manifest.json says, so the derived value has to
      # reach the tree before signing. package.json is written too: the two are
      # required to agree (ci.yml's guard), and a local `npm run build` reads
      # it. Working tree only — never committed, per the note on the derive
      # step.
      - name: Stamp the derived version into manifest.json + package.json
        env:
          DERIVED: ${{ steps.extver.outputs.version }}
        run: |
          python3 - <<'PY'
          import json, os

          version = os.environ["DERIVED"]
          for path in ("extension/manifest.json", "extension/package.json"):
              with open(path) as fh:
                  doc = json.load(fh)
              doc["version"] = version
              with open(path, "w") as fh:
                  json.dump(doc, fh, indent=2)
                  fh.write("\n")
              print(f"{path}: version -> {version}")
          PY

      - name: Sign via AMO (cache miss)
        if: steps.cache.outputs.cached != 'true'
        run: |
          cd extension && npm install --no-save --no-audit --no-fund && npm run sign
        env:
          WEB_EXT_API_KEY: ${{ secrets.MOZILLA_AMO_JWT_KEY }}
          WEB_EXT_API_SECRET: ${{ secrets.MOZILLA_AMO_JWT_SECRET }}

      - name: Upload signed XPI to ext-<version> release (cache miss)
        if: steps.cache.outputs.cached != 'true'
        env:
          TOKEN: ${{ secrets.RELEASE_TOKEN }}
        run: |
          set -eux
          VERSION=${{ steps.extver.outputs.version }}
          # AMO renames signed XPIs with its internal addon-id-safe-string;
          # canonicalize to fabledcurator-<version>.xpi so the FC server's
          # whitelist (backend/app/frontend.py expects 'fabledcurator-*.xpi')
          # keeps working.
          SIGNED=$(ls extension/web-ext-artifacts/*.xpi | head -1)
          XPI="extension/web-ext-artifacts/fabledcurator-$VERSION.xpi"
          cp "$SIGNED" "$XPI"
          # Find-or-create the ext-<version> release. Track whether WE
          # created it so an upload failure below can roll back (don't
          # leave an empty release tombstone that the next run's
          # cache-check mistakes for a partial-failure state).
          #
          # target_commitish is the signing commit, not a branch name: since
          # step 6 either branch can create this release, and hard-coding
          # `main` would tag a dev-signed XPI against a main commit that may
          # not even contain the extension source it was built from.
          STATUS=$(curl -s -o release.json -w "%{http_code}" \
            -H "Authorization: token $TOKEN" \
            "https://git.fabledsword.com/api/v1/repos/bvandeusen/FabledCurator/releases/tags/ext-$VERSION" || echo 000)
          if [ "$STATUS" = "200" ]; then
            CREATED_BY_US=false
          else
            curl -s -X POST -H "Authorization: token $TOKEN" -H "Content-Type: application/json" \
              -d "{\"tag_name\":\"ext-$VERSION\",\"name\":\"Extension $VERSION (signed XPI cache)\",\"body\":\"Internal cache for the signed XPI consumed by build.yml's build-web job. Not a user-facing FC release.\",\"target_commitish\":\"$GITHUB_SHA\"}" \
              -o release.json \
              "https://git.fabledsword.com/api/v1/repos/bvandeusen/FabledCurator/releases"
            CREATED_BY_US=true
          fi
          RELEASE_ID=$(python3 -c "import json; print(json.load(open('release.json'))['id'])")
          test -n "$RELEASE_ID"
          # Rollback-on-failure: if the asset upload fails AND we just
          # created the release in this run, delete it. Prevents an empty
          # ext-<version> release from poisoning the next workflow run
          # (operator-flagged 2026-05-26 — without rollback the next run
          # saw 'release exists, no asset → cache miss → sign' which AMO
          # then rejected with 409 'Version already exists').
          rollback_if_we_created() {
            if [ "$CREATED_BY_US" = "true" ]; then
              echo "Rolling back: deleting just-created release $RELEASE_ID"
              curl -s -X DELETE -H "Authorization: token $TOKEN" \
                "https://git.fabledsword.com/api/v1/repos/bvandeusen/FabledCurator/releases/$RELEASE_ID" || true
              curl -s -X DELETE -H "Authorization: token $TOKEN" \
                "https://git.fabledsword.com/api/v1/repos/bvandeusen/FabledCurator/tags/ext-$VERSION" || true
            fi
          }
          trap 'rollback_if_we_created' EXIT
          HTTP_CODE=$(curl -s -X POST -H "Authorization: token $TOKEN" \
            -F "attachment=@$XPI" \
            -o /dev/null -w "%{http_code}" \
            "https://git.fabledsword.com/api/v1/repos/bvandeusen/FabledCurator/releases/$RELEASE_ID/assets?name=fabledcurator-$VERSION.xpi")
          if [ "$HTTP_CODE" != "201" ] && [ "$HTTP_CODE" != "200" ]; then
            echo "Asset upload failed with HTTP $HTTP_CODE"
            exit 1
          fi
          # Upload succeeded — clear the rollback trap.
          trap - EXIT
          echo "Uploaded fabledcurator-$VERSION.xpi to ext-$VERSION release"

      # No actions/upload-artifact step: Forgejo Actions (and our
       # act_runner) doesn't support upload-artifact@v4+ (GHES limitation
       # surfaced 2026-05-26). Instead build-web reads the signed XPI
       # straight from the ext-<version> Forgejo release we just uploaded
       # to. Same source of truth; no double-store.

  build-web:
    # A plain `needs` — no `always()`. That expression existed to let a
    # SKIPPED sign-extension through on a tag push while still blocking a
    # FAILED one. With no tag trigger, sign-extension always runs, so the
    # default behaviour is exactly what we want: a failed sign skips build-web
    # rather than shipping an image without its XPI.
    needs: [sign-extension]
    runs-on: python-ci
    container:
      image: git.fabledsword.com/bvandeusen/ci-python:3.14
    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: env.IS_REFRESH == 'true'
        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
      # reports about itself; the load-bearing steps each print only the one
      # they use, so this is the only place the pair appears together. When a
      # build is skipped, this is the line that says what the commit derived.
      #
      # Still diagnostic, so it still must not fail the build — no `set -e`,
      # and every derivation falls back to UNAVAILABLE. A broken echo must
      # never be the reason an image does not ship.
      #
      # What it should say:
      #   * a push touching only agent/ moves the agent and leaves web and ml
      #     STILL. If web moves, its path set is too wide.
      #   * a push touching only docs moves nothing.
      #   * a push touching the extension moves the extension AND web, since
      #     web bakes in the XPI. If web does not move, its set is too narrow:
      #     the reuse check hits, and the channel serves a web image bundling
      #     the PREVIOUS XPI while the freshly signed one is orphaned (#3156).
      #   * dev and main derive the same values for the same source.
      - name: Report the derived artifact version
        run: |
          set -u
          A=web
          V=$(sh scripts/artifacts.sh version "$A" 2>&1 || echo UNAVAILABLE)
          R=$(sh scripts/artifacts.sh revision "$A" 2>&1 || echo UNAVAILABLE)
          echo "derived: artifact=$A version=$V revision=$R sha=$GITHUB_SHA"

      - name: Determine tag
        id: tag
        run: |
          # Two trigger shapes, and between them they publish three tags:
          #   main → :latest  (production, moving — rule 147: main IS production)
          #          :c-<sha> (immutable, the rollback unit — rule 145)
          #   dev  → :dev     (the rolling test channel — rule 146)
          #
          # That is the whole list. No :<version>, and no :main — rule 145,
          # narrowed 2026-08-28 once it was verified that nothing pins:
          # "a third name for the same thing is upkeep for a model we do not
          # run." The date tag published between milestone 313 step 3 and
          # milestone 318 was exactly that; :main was a second moving name for
          # whatever :latest already pointed at.
          #
          # `dev` gets no :c-<sha> deliberately. On a channel whose entire
          # contract is that it moves, a per-push immutable tag is a rollback
          # target nobody has ever pulled, accumulating forever. The accepted
          # cost: on dev there is no rollback but the previous :dev, which is
          # gone — recovery is revert-on-git plus a CI cycle.
          #
          # Reinstating :<version> is a real decision, not a default. It earns
          # its place when something genuinely pins: a second instance held on
          # a known-good build, or a deliberately frozen window. Tag at the
          # moment you decide to freeze; no back-catalogue is needed.
          #
          # POSIX-safe substring (the runner shell is dash/BusyBox sh, not
          # bash — `${var:0:7}` errors with "Bad substitution"; cut works
          # everywhere). Operator-flagged 2026-06-01 after the first :c-<sha>
          # main-push build failed at this step.
          SHORT_SHA=$(printf '%s' "$GITHUB_SHA" | cut -c1-7)

          # 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 [ "${IS_REFRESH:-}" = "true" ]; then
            echo "tags=git.fabledsword.com/bvandeusen/fabledcurator:latest" >> "$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" >> "$GITHUB_OUTPUT"
          else
            echo "tags=git.fabledsword.com/bvandeusen/fabledcurator:dev" >> "$GITHUB_OUTPUT"
            echo "channel=dev" >> "$GITHUB_OUTPUT"
          fi

      # A shell step, not docker/login-action@v3, because the action's shared
      # cache races itself (#3118). act_runner caches a remote action under one
      # /root/.cache/act/<hash> per runner, and build-web, build-ml and
      # build-agent all start in the same second and all want this same action.
      # One job re-clones the directory — which empties and repopulates it —
      # while another is walking it to copy into its container, and the walker
      # lstat()s a file that has just vanished. It failed twice on 2026-08-27,
      # naming a DIFFERENT missing file each time (`eslint.config.mjs`, then
      # `jest.config.ts`), which is what rules out a corrupt cache and points at
      # a race. The loser dies with MODULE_NOT_FOUND on dist/index.js before the
      # action runs at all, so the secret is never even reached.
      #
      # Nothing is lost by dropping it: logging in is one command, the docker
      # CLI is already in the CI image (ci-requirements.md), and the same
      # reasoning as family rule 5 applies — a marketplace action buys nothing
      # when the tool is baked into the image the workflow already selected.
      #
      # Password on stdin, never as an argument: an argument lands in the
      # process table and draws docker's own deprecation warning.
      - name: Login to Forgejo registry
        env:
          TOKEN: ${{ secrets.RELEASE_TOKEN }}
          ACTOR: ${{ github.actor }}
        run: echo "$TOKEN" | docker login git.fabledsword.com -u "$ACTOR" --password-stdin

      # A REAL buildx builder, not the default `docker` driver (#3114, #3190).
      #
      # The default driver builds through the local dockerd. It cannot export a
      # registry cache at all — which is why the agent rebuilds a ~6.3 GB CUDA
      # + torch image from scratch whenever the runner's local cache is cold,
      # measured at 9m26s against 7s warm. It is also #3190's leading suspect:
      # after a registry-direct push it resolves image metadata against a local
      # store the push never filled, and reports `No such image` on an image
      # that published perfectly well three seconds earlier.
      #
      # These jobs run INSIDE a container against a mounted docker socket, so
      # the buildkit container this starts is a SIBLING of the job container,
      # not a child. That works over the socket mount; it had never been tried
      # here before milestone 326 step 1.
      - name: Set up buildx
        uses: docker/setup-buildx-action@v3

      # --- reuse-if-published (milestone 313, step 4) ----------------------
      # Does the image the channel tag already points at carry THIS commit's
      # revision? If so the bytes this job would produce are already published
      # and the build is pure waste: the remaining tags get repointed at that
      # existing manifest instead, registry-side, in seconds.
      #
      # Keyed on an `fc.revision` LABEL rather than on a tag of its own
      # (milestone 318 step 3). A tag would be a name minted per build that one
      # thing reads — what rule 145 narrowed against — and would be prunable
      # under the registry's keep_pattern (#3157), silently expiring the cache.
      # A label rides inside a tag that has to exist anyway.
      #
      # An image with no such label reads as a miss and rebuilds. That is the
      # migration, not a fault: labels cannot be backfilled, since the reuse
      # path copies a manifest and config labels are not manifest annotations.
      # Each artifact pays one rebuild, once.
      #
      # This is what stops a push that touched only `agent/` from rebuilding
      # web and ml, and a merge to main from rebuilding what dev already built.
      #
      # The failure direction is deliberate. An inspect that errors for ANY
      # reason — network, auth, a registry hiccup — reads as a miss and the
      # build runs. Only a genuine 200 skips one, so there is no path here
      # that skips a build that was actually needed; the worst case is paying
      # for a build we could have avoided.
      #
      # BASE-IMAGE FRESHNESS: an artifact whose source stops moving stops
      # picking up base-image updates. Milestone 318 removed the argument this
      # used to need rather than answering it — with no version tags there is
      # no immutable name a refresh could contradict, and rule 145 already
      # allows a rebuild with different contents to republish a MOVING tag.
      # So a refresh is just a build. A scheduled channel-only one is tracked
      # separately (#3154); it does not belong in the push path.
      - name: Is this content already published?
        id: reuse
        env:
          IMAGE: git.fabledsword.com/bvandeusen/fabledcurator
          CHANNEL: ${{ steps.tag.outputs.channel }}
          # Empty on a push; the string "true" only from a workflow_dispatch
          # that asked for it. `github.event.inputs` rather than the `inputs`
          # context — release.yml already uses that form, and it is the one
          # this runner is known to evaluate. Read through env rather than
          # interpolated into the run block, same rule as release.yml's TAG.
          FORCE: ${{ github.event.inputs.force_build }}
          # A scheduled refresh has to bypass reuse by construction: it
          # rebuilds the SAME source, so fc.revision always matches and the
          # check would skip every refresh there has ever been.
        run: |
          set -eu
          DERIVED=$(sh scripts/artifacts.sh revision web)
          echo "revision=$DERIVED" >> "$GITHUB_OUTPUT"
          # Baked into the web image as FC_VERSION and reported by /api/health.
          # A pure function of the revision — same commit, same string — so it
          # adds no variability the reuse check would have to account for.
          echo "version=$(sh scripts/artifacts.sh version web)" >> "$GITHUB_OUTPUT"

          # The build clock, pinned to the same commit (#3265). Without it
          # buildkit stamps the image config with the wall clock of the build,
          # so identical layers republish under a new config blob and the
          # channel tag gets a new manifest digest for no reason. Derived from
          # `newest()` like revision and version, so all three name one commit
          # and cannot drift apart.
          echo "epoch=$(sh scripts/artifacts.sh epoch web)" >> "$GITHUB_OUTPUT"

          # The moving tag for this channel. Which tag we ask IS the channel —
          # that is why the revision needs no -main/-dev qualifier any more.
          if [ "$CHANNEL" = "main" ]; then T=latest; else T=dev; fi
          echo "channel_ref=$IMAGE:$T" >> "$GITHUB_OUTPUT"

          # WHERE THE BUILD PUBLISHES, which is not always the channel — and
          # whether the channel then has to be written separately.
          #
          # On a push the build writes the channel tag directly: the bytes came
          # from a commit, and a commit is the thing CI tests. Nothing to hold
          # it behind.
          #
          # On the scheduled refresh it writes a CANDIDATE tag instead. A
          # refresh rebuilds against freshly resolved base images, and the web
          # image's runtime is a line of UNPINNED Debian packages (ffmpeg,
          # libjpeg62-turbo, libpq5, megatools…) re-resolved on every build.
          # Nothing in ci.yml can see that: its lanes run on ci-python:3.14 and
          # install requirements.txt, and a base bump changes neither. So
          # refreshed bytes have to be proven before :latest names them, and
          # proving needs a moment between "built" and "published" to occupy.
          # This is that moment; :latest goes on naming the build that works
          # until something says otherwise.
          #
          # `:refresh-candidate` is one moving ref per image, overwritten in
          # place, holding a build nobody is told to pull — the shape rule 145
          # already allows for :buildcache, not the per-build tag family that
          # milestone 318 withdrew.
          #
          # Both values are decided HERE, beside `hit`, for the reason the
          # force/schedule branch below gives: one step decides what this job
          # does. A promote condition derived independently could disagree with
          # the tag the build actually wrote.
          if [ "${IS_REFRESH:-}" = "true" ]; then
            echo "build_ref=$IMAGE:refresh-candidate" >> "$GITHUB_OUTPUT"
            echo "promote=true" >> "$GITHUB_OUTPUT"
          else
            echo "build_ref=$IMAGE:$T" >> "$GITHUB_OUTPUT"
            echo "promote=false" >> "$GITHUB_OUTPUT"
          fi

          # Compare VALUES, never exit codes. Measured on buildx v0.36.1
          # (run 4732): a missing key returns an empty string and exits 0, so
          # branching on the exit code would read "no label yet" as success.
          # An unreachable tag also lands here as empty via the `|| echo`.
          # Empty never equals a 12-char revision, so every uncertain case
          # falls through to a build — the safe direction, with no special
          # casing for it.
          #
          # Read the SPECIFIC key. The map also carries whatever the base image
          # set, and `org.opencontainers.image.version` sits right beside ours
          # looking like a plausible answer (it reads 24.04 on the agent).
          PUBLISHED=$(docker buildx imagetools inspect "$IMAGE:$T" \
            --format '{{ index .Image.Config.Labels "fc.revision" }}' \
            2>/dev/null || echo "")
          echo "reuse: $IMAGE:$T carries fc.revision=${PUBLISHED:-<none>}; derived=$DERIVED"
          if [ -z "$PUBLISHED" ] && docker buildx imagetools inspect "$IMAGE:$T" >/dev/null 2>&1; then
            # The tag resolves but carries no readable label. Expected exactly
            # once per artifact, during the migration onto labels. If it recurs
            # every push, something is rewriting the channel tag as a manifest
            # index — see the repoint step's note.
            echo "reuse: NOTE $IMAGE:$T exists but has no readable fc.revision."
            echo "reuse: NOTE Fine once, while migrating. Every push means the"
            echo "reuse: NOTE tag is being index-wrapped and reuse is dead."
          fi

          # FORCE is checked here rather than in the build step's `if:`, so
          # that one decision drives everything downstream. The repoint step
          # keys off `hit` too, and a force that bypassed only the build would
          # leave the two disagreeing about what just happened.
          if [ "${FORCE:-false}" = "true" ]; then
            echo "hit=false" >> "$GITHUB_OUTPUT"
            echo "reuse: force_build set — building regardless"
          elif [ "${IS_REFRESH:-}" = "true" ]; then
            echo "hit=false" >> "$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"
          else
            echo "hit=false" >> "$GITHUB_OUTPUT"
            echo "reuse: not published — building"
          fi

      - name: Download signed XPI from Forgejo release asset
        # dev and main each bundle the XPI their own sign-extension just
        # published — the point of the channel work (milestone 271 step 6): the
        # dev image carries the extension being developed, rather than
        # requiring a merge to try it.
        #
        # The 10-minute polling loop that used to live here is gone with the
        # tag trigger (milestone 318 step 2). It existed for one shape only: a
        # release cut fired the tag build and the main build together, the tag
        # build skipped sign-extension and raced straight here, and it lost
        # every time (operator-flagged 2026-05-27 after v26.05.27.0). Polling
        # was the fix for a build that should not have been running.
        #
        # sign-extension is a `needs` dependency and it succeeded, so the
        # release exists. A single fetch is correct, and a 404 now means a real
        # disagreement about the derived version rather than a race — which is
        # exactly what should fail loudly instead of being slept through.
        #
        # Still gated on the reuse miss: a published image already contains its
        # XPI, so this would fetch a file nothing then reads.
        if: steps.reuse.outputs.hit != 'true'
        env:
          TOKEN: ${{ secrets.RELEASE_TOKEN }}
        run: |
          set -eux
          # Re-derived, not read from the repo: sign-extension published
          # ext-<derived>, and the committed version has been inert since
          # milestone 271 step 4. Both jobs run `packaging.sh version` over the
          # same commit, so they agree by construction — and if they ever
          # didn't, this download 404s and the build fails loudly instead of
          # shipping a stale XPI.
          VERSION=$(sh extension/scripts/packaging.sh version)
          # One fetch, no retry. sign-extension ran to success in this same
          # workflow and published ext-$VERSION; both jobs derive $VERSION from
          # the same commit, so they agree by construction. A 404 here means
          # they did NOT agree, and sleeping on that would only delay the
          # report.
          STATUS=$(curl -s -o release.json -w "%{http_code}" \
            -H "Authorization: token $TOKEN" \
            "https://git.fabledsword.com/api/v1/repos/bvandeusen/FabledCurator/releases/tags/ext-$VERSION" || echo 000)
          if [ "$STATUS" != "200" ]; then
            echo "ERROR: ext-$VERSION release not found (HTTP $STATUS)."
            echo "sign-extension succeeded in this run, so it published some"
            echo "other version — the two jobs derived different values for one"
            echo "commit. Check that both checked out with fetch-depth: 0."
            exit 1
          fi
          # Extract the .xpi asset's browser_download_url (Forgejo's
          # /releases/assets/<id> endpoint returns ASSET METADATA, not
          # the binary blob — operator-flagged 2026-05-26: my prior
          # code curl'd the metadata endpoint without -f and wrote the
          # resulting 404-page-not-found text into fabledcurator-*.xpi,
          # which Firefox then rejected as "corrupt").
          # browser_download_url is the canonical binary endpoint and
          # is also publicly accessible (no token needed) but we pass
          # the token anyway for symmetry with private-repo support.
          DOWNLOAD_URL=$(python3 -c "import json; r=json.load(open('release.json')); xpis=[a for a in r.get('assets', []) if a.get('name','').endswith('.xpi')]; print(xpis[0]['browser_download_url'])")
          test -n "$DOWNLOAD_URL"
          echo "Downloading XPI from: $DOWNLOAD_URL"
          mkdir -p frontend/public/extension
          DEST="frontend/public/extension/fabledcurator-$VERSION.xpi"
          # -f = fail on HTTP error (prevents silent corruption like the
          # 2026-05-26 incident); -L = follow redirects.
          curl -sfL -H "Authorization: token $TOKEN" -o "$DEST" "$DOWNLOAD_URL"
          # Sanity check: the binary should start with the ZIP magic (PK\x03\x04).
          # If it's anything else, the next docker build will ship a corrupt XPI.
          MAGIC=$(head -c 2 "$DEST" | od -An -c | tr -d ' \n')
          if [ "$MAGIC" != "PK" ]; then
            echo "ERROR: downloaded XPI does not start with ZIP magic 'PK' (got '$MAGIC')"
            echo "File contents preview:"
            head -c 200 "$DEST"
            exit 1
          fi
          cp "$DEST" "frontend/public/extension/fabledcurator-latest.xpi"
          ls -la frontend/public/extension/

      - name: Build and push web image
        if: steps.reuse.outputs.hit != 'true'
        # Read by buildx out of the ENVIRONMENT, not passed as a build-arg —
        # it normalises the image config's `created` field and the history
        # timestamps rather than being consumed by the Dockerfile. See #3265
        # and the reuse step's `epoch` output.
        env:
          SOURCE_DATE_EPOCH: ${{ steps.reuse.outputs.epoch }}
        uses: docker/build-push-action@v5
        with:
          context: .
          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 was ~13s with every content step CACHED —
          # and the channel tag STILL got a new manifest digest, because
          # buildkit stamps a fresh image config per run and republishes the
          # identical layers under it. All three images moved that way on
          # 2026-08-30 with nothing whatsoever changed in them.
          #
          # SOURCE_DATE_EPOCH (below) is the fix: pinned to the commit the
          # content came from, the config is byte-identical across runs, so
          # the manifest digest is too and the push is a registry no-op. A
          # digest change means the content changed again, which is the only
          # thing a digest is any use for.
          #
          # 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: ${{ env.IS_REFRESH == 'true' }}
          # ONE tag, the channel's. Every other tag is written by the step
          # below, registry-side. buildx here pushes the first tag to the
          # registry and then re-pushes the rest through the DOCKER driver,
          # out of a local image store a registry-direct build never filled —
          # #3190, which cost `main` its :c-<sha> on 2026-08-29 while :latest
          # published perfectly well.
          tags: ${{ steps.reuse.outputs.build_ref }}
          # The reuse key. Read back off the channel tag on the next push to
          # decide whether that push needs to build at all, so this is not
          # decoration — an unstamped image is one that will always rebuild.
          labels: |
            fc.revision=${{ steps.reuse.outputs.revision }}
          # LOAD-BEARING, not a preference. On the default docker driver these
          # were no-ops; on the docker-container driver above,
          # build-push-action@v5 defaults provenance to TRUE when pushing.
          # Provenance attaches an attestation manifest, which makes the pushed
          # tag a manifest INDEX — and `.Image.Config.Labels` does not resolve
          # through an index.
          #
          # The label directly above IS the reuse key. Wrap the channel tag in
          # an index and the next push reads fc.revision=<none>, misses, and
          # rebuilds. Then so does the one after that, forever. Nothing fails,
          # nothing goes red, and the only symptom is the bill. That is #3183
          # arriving through a different door, and note #3127 §4 records the
          # same shape for `platforms:`.
          provenance: false
          sbom: false
          # The ONLY cache this driver can have. `docker-container` gets a
          # FRESH buildkit instance per job, so unlike the default docker
          # driver it has no local layer store to fall back on — measured on
          # run 4896, the first builds after the driver change: web 3m44s
          # (was 2m23s), ml 3m49s (was 3m20s), agent 11m12s (was 9m26s). The
          # driver change ALONE is a regression; this is the other half of it.
          #
          # mode=max so intermediate stages cache too. web's frontend-builder
          # stage and the agent's two ~150s pip layers are the whole cost, and
          # they are exactly what a min-mode cache would drop.
          #
          # A `:buildcache` tag is NOT the withdrawn tag scheme coming back.
          # Rule 145 narrowed against names NOTHING reads; this one is read by
          # every build that runs, is one moving ref per image rather than one
          # per build, holds cache blobs rather than a shippable artifact, and
          # is overwritten in place rather than accumulating. It is closer to
          # :dev than to the :2026.8.28 tags milestone 318 deleted. (#3114.)
          cache-from: type=registry,ref=git.fabledsword.com/bvandeusen/fabledcurator:buildcache
          cache-to: type=registry,ref=git.fabledsword.com/bvandeusen/fabledcurator:buildcache,mode=max
          # Only the web image carries these: it is the one with a UI and an
          # HTTP surface to report them on. The ml and agent images have
          # nothing to tell.
          build-args: |
            FC_CHANNEL=${{ steps.tag.outputs.channel }}
            FC_VERSION=${{ steps.reuse.outputs.version }}

      # Point the channel tag at the candidate the refresh just built.
      #
      # Unconditional TODAY, so this milestone never leaves the refresh in a
      # state where it builds and publishes nothing. Step 4 wraps it in the
      # smoke suite's verdict; until then the scheduled path behaves exactly
      # as it did, just via two operations instead of one.
      #
      # NOT `imagetools create`. That wraps its source in an INDEX, and an
      # indexed channel tag is the one thing this pipeline cannot survive:
      # `.Image.Config.Labels` does not resolve through an index, so the
      # fc.revision the reuse check reads off the channel tag would come back
      # empty, every subsequent push would miss and rebuild, and nothing would
      # go red. That is #3183, observed on run 4751 — reuse worked exactly once
      # and the only symptom was the bill. The repoint step below excludes its
      # own source tag for precisely this reason; a promote that re-introduced
      # the wrap through a different door would undo that care.
      #
      # A manifest PUT is what "make this tag name that image" means at the
      # registry level: the same bytes under the same media type, so the digest
      # is identical, the media type is preserved, and no layer moves.
      - name: Promote the refresh candidate to the channel
        if: steps.reuse.outputs.promote == 'true'
        env:
          IMAGE: git.fabledsword.com/bvandeusen/fabledcurator
          CHANNEL_REF: ${{ steps.reuse.outputs.channel_ref }}
          TOKEN: ${{ secrets.RELEASE_TOKEN }}
          ACTOR: ${{ github.actor }}
        run: |
          set -eu
          REPO=${IMAGE#git.fabledsword.com/}
          TAG=${CHANNEL_REF##*:}

          # Registry auth is its own token exchange — the `docker login` above
          # authenticates the docker client, not curl. Deadline on every call
          # (rule 156): a registry that stops answering must fail this step,
          # not hang the weekly refresh until the job times out.
          BEARER=$(curl -fsS --max-time 30 -u "$ACTOR:$TOKEN" \
            "https://git.fabledsword.com/v2/token?scope=repository:$REPO:pull,push&service=git.fabledsword.com" \
            | python3 -c 'import sys,json; print(json.load(sys.stdin)["token"])')

          # Ask for the image manifest media types ONLY. Offering the index
          # types too would let the registry hand back an index if one ever
          # existed at this tag, and we would faithfully copy the thing we are
          # trying not to create.
          ACCEPT='application/vnd.oci.image.manifest.v1+json, application/vnd.docker.distribution.manifest.v2+json'
          CT=$(curl -fsS --max-time 60 -o manifest.json -D headers.txt \
                 -H "Authorization: Bearer $BEARER" -H "Accept: $ACCEPT" \
                 "https://git.fabledsword.com/v2/$REPO/manifests/refresh-candidate" \
               && tr -d '\r' < headers.txt | awk -F': ' '/^[Cc]ontent-[Tt]ype:/{print $2}')
          test -n "$CT"
          SRC_DIGEST=$(tr -d '\r' < headers.txt | awk -F': ' '/^[Dd]ocker-[Cc]ontent-[Dd]igest:/{print $2}')
          echo "promote: candidate is $SRC_DIGEST ($CT)"

          curl -fsS --max-time 120 -X PUT \
            -H "Authorization: Bearer $BEARER" -H "Content-Type: $CT" \
            --data-binary @manifest.json \
            "https://git.fabledsword.com/v2/$REPO/manifests/$TAG"

          # Read it back. A PUT that returned 2xx but landed something else is
          # exactly the silent-and-plausible failure this pipeline keeps
          # producing, and the check costs one request.
          NOW=$(curl -fsS --max-time 30 -o /dev/null -D - \
                  -H "Authorization: Bearer $BEARER" -H "Accept: $ACCEPT" \
                  "https://git.fabledsword.com/v2/$REPO/manifests/$TAG" \
                | tr -d '\r' | awk -F': ' '/^[Dd]ocker-[Cc]ontent-[Dd]igest:/{print $2}')
          if [ "$NOW" != "$SRC_DIGEST" ]; then
            echo "promote: $IMAGE:$TAG is $NOW, expected $SRC_DIGEST" >&2
            exit 1
          fi
          echo "promote: $IMAGE:$TAG now names $NOW"

      # Every tag but the channel's own is written HERE, registry-side,
      # whether or not a build ran. Each -t becomes another reference to the
      # SAME manifest the channel tag holds, so :c-<sha> is byte-identical to
      # what is published rather than a lookalike rebuild.
      #
      # Owning the build path too is #3190's fix, not a tidy-up:
      #
      #   #27 pushing …/fabledcurator:latest           DONE 15.8s
      #   #28 pushing …/fabledcurator:c-0e15c44 with docker
      #   #28 ERROR: tag does not exist: …:c-0e15c44
      #
      # Intermittent — build-ml made the identical two-tag push seconds later
      # and succeeded — and worse than it looks. `:latest` had already
      # published, so production was correct while the immutable rollback tag
      # rule 145 requires of every main push simply did not exist. Nothing but
      # the red job would ever have noticed: a missing :c-<sha> has no
      # consumer that fails, so it surfaces when somebody needs to roll back.
      #
      # `imagetools create` is a registry-side manifest copy — no layer
      # transfer, no local daemon, nothing that can be absent. The reuse case
      # has always gone this way, so this puts the build case on the code that
      # was already proven rather than on a second path.
      #
      # Running on every path also keeps family rule 146 true: a rolling
      # channel refreshes itself, so skipping a build must never leave :dev or
      # :latest pointing at something older than the commit just pushed.
      #
      # The cost, accepted knowingly: `imagetools create` wraps its source in
      # an index, so :c-<sha> becomes an index and fc.revision does not
      # resolve through it. Nothing reads that label off :c-<sha> — the reuse
      # check only ever inspects the CHANNEL tag — and the index names the
      # same manifest, so a pull is byte-identical. The reuse path already
      # produced :c-<sha> this way; this only makes it uniform.
      - name: Write the remaining tags from the published image
        env:
          IMAGE: git.fabledsword.com/bvandeusen/fabledcurator
          SOURCE: ${{ steps.reuse.outputs.channel_ref }}
          TAGS: ${{ steps.tag.outputs.tags }}
        run: |
          set -euf
          # The source tag is EXCLUDED from the targets, and that is load-
          # bearing rather than an optimisation.
          #
          # `imagetools create` wraps the source manifest in an INDEX. Point it
          # at the channel tag with that same tag as a target and the tag stops
          # being a plain image — after which `.Image.Config.Labels` no longer
          # resolves through it and the fc.revision label reads as absent. The
          # next push then misses and rebuilds, so reuse worked exactly once
          # and every subsequent push paid full price. Observed on run 4751:
          # ml:dev reported fc.revision=<none> one push after run 4749 had read
          # a7e626a67a79 off it. Nothing failed; the savings just evaporated.
          #
          # Excluding the source means the channel tag is only ever written
          # by a real build, so it stays a plain image and stays readable.
          # On dev that leaves nothing to do either way: the build pushed :dev
          # itself, or the hit established it was already right. On main it
          # leaves :c-<sha>, which rule 145 requires of every main push whether
          # or not a build ran.
          #
          # steps.tag emits ONE comma-separated list; imagetools wants a -t per
          # ref. (That list used to feed docker/build-push-action directly —
          # which is exactly what #3190 made unsafe.)
          ARGS=""
          IFS=,
          for t in $TAGS; do
            [ "$t" = "$SOURCE" ] && continue
            ARGS="$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."
            exit 0
          fi
          # shellcheck disable=SC2086
          docker buildx imagetools create $ARGS "$SOURCE"
          echo "repointed from $SOURCE:$ARGS"

  build-ml:
    runs-on: python-ci
    container:
      image: git.fabledsword.com/bvandeusen/ci-python:3.14
    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,
          # too-low value or finds nothing at all, and neither is a failure
          # 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: env.IS_REFRESH == 'true'
        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
      # reports about itself; the load-bearing steps each print only the one
      # they use, so this is the only place the pair appears together. When a
      # build is skipped, this is the line that says what the commit derived.
      #
      # Still diagnostic, so it still must not fail the build — no `set -e`,
      # and every derivation falls back to UNAVAILABLE. A broken echo must
      # never be the reason an image does not ship.
      #
      # What it should say:
      #   * a push touching only agent/ moves the agent and leaves web and ml
      #     STILL. If web moves, its path set is too wide.
      #   * a push touching only docs moves nothing.
      #   * a push touching the extension moves the extension AND web, since
      #     web bakes in the XPI. If web does not move, its set is too narrow:
      #     the reuse check hits, and the channel serves a web image bundling
      #     the PREVIOUS XPI while the freshly signed one is orphaned (#3156).
      #   * dev and main derive the same values for the same source.
      - name: Report the derived artifact version
        run: |
          set -u
          A=ml
          V=$(sh scripts/artifacts.sh version "$A" 2>&1 || echo UNAVAILABLE)
          R=$(sh scripts/artifacts.sh revision "$A" 2>&1 || echo UNAVAILABLE)
          echo "derived: artifact=$A version=$V revision=$R sha=$GITHUB_SHA"

      - name: Determine tag
        id: tag
        run: |
          # Mirrors build-web's tag list; see the comment there.
          # POSIX-safe substring (the runner shell is dash/BusyBox sh, not
          # bash — `${var:0:7}` errors with "Bad substitution"; cut works
          # everywhere). Operator-flagged 2026-06-01 after first :c-<sha>
          # main-push build failed at this step.
          SHORT_SHA=$(printf '%s' "$GITHUB_SHA" | cut -c1-7)
          # Mirrors build-web's tag list and its schedule handling; see
          # the comments there.
          if [ "${IS_REFRESH:-}" = "true" ]; then
            echo "tags=git.fabledsword.com/bvandeusen/fabledcurator-ml:latest" >> "$GITHUB_OUTPUT"
            echo "channel=main" >> "$GITHUB_OUTPUT"
          elif [ "${GITHUB_REF##*/}" = "main" ]; then
            echo "tags=git.fabledsword.com/bvandeusen/fabledcurator-ml:latest,git.fabledsword.com/bvandeusen/fabledcurator-ml:c-${SHORT_SHA}" >> "$GITHUB_OUTPUT"
            echo "channel=main" >> "$GITHUB_OUTPUT"
          else
            echo "tags=git.fabledsword.com/bvandeusen/fabledcurator-ml:dev" >> "$GITHUB_OUTPUT"
            echo "channel=dev" >> "$GITHUB_OUTPUT"
          fi

      # Shell step rather than docker/login-action — see build-web's note on
      # the shared action-cache race (#3118).
      - name: Login to Forgejo registry
        env:
          TOKEN: ${{ secrets.RELEASE_TOKEN }}
          ACTOR: ${{ github.actor }}
        run: echo "$TOKEN" | docker login git.fabledsword.com -u "$ACTOR" --password-stdin

      # A REAL buildx builder, not the default `docker` driver (#3114, #3190).
      #
      # The default driver builds through the local dockerd. It cannot export a
      # registry cache at all — which is why the agent rebuilds a ~6.3 GB CUDA
      # + torch image from scratch whenever the runner's local cache is cold,
      # measured at 9m26s against 7s warm. It is also #3190's leading suspect:
      # after a registry-direct push it resolves image metadata against a local
      # store the push never filled, and reports `No such image` on an image
      # that published perfectly well three seconds earlier.
      #
      # These jobs run INSIDE a container against a mounted docker socket, so
      # the buildkit container this starts is a SIBLING of the job container,
      # not a child. That works over the socket mount; it had never been tried
      # here before milestone 326 step 1.
      - name: Set up buildx
        uses: docker/setup-buildx-action@v3

      # --- reuse-if-published (milestone 313, step 4) ----------------------
      # Does the image the channel tag already points at carry THIS commit's
      # revision? If so the bytes this job would produce are already published
      # and the build is pure waste: the remaining tags get repointed at that
      # existing manifest instead, registry-side, in seconds.
      #
      # Keyed on an `fc.revision` LABEL rather than on a tag of its own
      # (milestone 318 step 3). A tag would be a name minted per build that one
      # thing reads — what rule 145 narrowed against — and would be prunable
      # under the registry's keep_pattern (#3157), silently expiring the cache.
      # A label rides inside a tag that has to exist anyway.
      #
      # An image with no such label reads as a miss and rebuilds. That is the
      # migration, not a fault: labels cannot be backfilled, since the reuse
      # path copies a manifest and config labels are not manifest annotations.
      # Each artifact pays one rebuild, once.
      #
      # This is what stops a push that touched only `agent/` from rebuilding
      # web and ml, and a merge to main from rebuilding what dev already built.
      #
      # The failure direction is deliberate. An inspect that errors for ANY
      # reason — network, auth, a registry hiccup — reads as a miss and the
      # build runs. Only a genuine 200 skips one, so there is no path here
      # that skips a build that was actually needed; the worst case is paying
      # for a build we could have avoided.
      #
      # BASE-IMAGE FRESHNESS: an artifact whose source stops moving stops
      # picking up base-image updates. Milestone 318 removed the argument this
      # used to need rather than answering it — with no version tags there is
      # no immutable name a refresh could contradict, and rule 145 already
      # allows a rebuild with different contents to republish a MOVING tag.
      # So a refresh is just a build. A scheduled channel-only one is tracked
      # separately (#3154); it does not belong in the push path.
      - name: Is this content already published?
        id: reuse
        env:
          IMAGE: git.fabledsword.com/bvandeusen/fabledcurator-ml
          CHANNEL: ${{ steps.tag.outputs.channel }}
          # Empty on a push; the string "true" only from a workflow_dispatch
          # that asked for it. `github.event.inputs` rather than the `inputs`
          # context — release.yml already uses that form, and it is the one
          # this runner is known to evaluate. Read through env rather than
          # interpolated into the run block, same rule as release.yml's TAG.
          FORCE: ${{ github.event.inputs.force_build }}
          # A scheduled refresh has to bypass reuse by construction: it
          # rebuilds the SAME source, so fc.revision always matches and the
          # check would skip every refresh there has ever been.
        run: |
          set -eu
          DERIVED=$(sh scripts/artifacts.sh revision ml)
          echo "revision=$DERIVED" >> "$GITHUB_OUTPUT"

          # The build clock, pinned to the same commit (#3265). Without it
          # buildkit stamps the image config with the wall clock of the build,
          # so identical layers republish under a new config blob and the
          # channel tag gets a new manifest digest for no reason. Derived from
          # `newest()` like revision and version, so all three name one commit
          # and cannot drift apart.
          echo "epoch=$(sh scripts/artifacts.sh epoch ml)" >> "$GITHUB_OUTPUT"

          # The moving tag for this channel. Which tag we ask IS the channel —
          # that is why the revision needs no -main/-dev qualifier any more.
          if [ "$CHANNEL" = "main" ]; then T=latest; else T=dev; fi
          echo "channel_ref=$IMAGE:$T" >> "$GITHUB_OUTPUT"

          # WHERE THE BUILD PUBLISHES, which is not always the channel — and
          # whether the channel then has to be written separately.
          #
          # On a push the build writes the channel tag directly: the bytes came
          # from a commit, and a commit is the thing CI tests. Nothing to hold
          # it behind.
          #
          # On the scheduled refresh it writes a CANDIDATE tag instead. A
          # refresh rebuilds against freshly resolved base images, and the web
          # image's runtime is a line of UNPINNED Debian packages (ffmpeg,
          # libjpeg62-turbo, libpq5, megatools…) re-resolved on every build.
          # Nothing in ci.yml can see that: its lanes run on ci-python:3.14 and
          # install requirements.txt, and a base bump changes neither. So
          # refreshed bytes have to be proven before :latest names them, and
          # proving needs a moment between "built" and "published" to occupy.
          # This is that moment; :latest goes on naming the build that works
          # until something says otherwise.
          #
          # `:refresh-candidate` is one moving ref per image, overwritten in
          # place, holding a build nobody is told to pull — the shape rule 145
          # already allows for :buildcache, not the per-build tag family that
          # milestone 318 withdrew.
          #
          # Both values are decided HERE, beside `hit`, for the reason the
          # force/schedule branch below gives: one step decides what this job
          # does. A promote condition derived independently could disagree with
          # the tag the build actually wrote.
          if [ "${IS_REFRESH:-}" = "true" ]; then
            echo "build_ref=$IMAGE:refresh-candidate" >> "$GITHUB_OUTPUT"
            echo "promote=true" >> "$GITHUB_OUTPUT"
          else
            echo "build_ref=$IMAGE:$T" >> "$GITHUB_OUTPUT"
            echo "promote=false" >> "$GITHUB_OUTPUT"
          fi

          # Compare VALUES, never exit codes. Measured on buildx v0.36.1
          # (run 4732): a missing key returns an empty string and exits 0, so
          # branching on the exit code would read "no label yet" as success.
          # An unreachable tag also lands here as empty via the `|| echo`.
          # Empty never equals a 12-char revision, so every uncertain case
          # falls through to a build — the safe direction, with no special
          # casing for it.
          #
          # Read the SPECIFIC key. The map also carries whatever the base image
          # set, and `org.opencontainers.image.version` sits right beside ours
          # looking like a plausible answer (it reads 24.04 on the agent).
          PUBLISHED=$(docker buildx imagetools inspect "$IMAGE:$T" \
            --format '{{ index .Image.Config.Labels "fc.revision" }}' \
            2>/dev/null || echo "")
          echo "reuse: $IMAGE:$T carries fc.revision=${PUBLISHED:-<none>}; derived=$DERIVED"
          if [ -z "$PUBLISHED" ] && docker buildx imagetools inspect "$IMAGE:$T" >/dev/null 2>&1; then
            # The tag resolves but carries no readable label. Expected exactly
            # once per artifact, during the migration onto labels. If it recurs
            # every push, something is rewriting the channel tag as a manifest
            # index — see the repoint step's note.
            echo "reuse: NOTE $IMAGE:$T exists but has no readable fc.revision."
            echo "reuse: NOTE Fine once, while migrating. Every push means the"
            echo "reuse: NOTE tag is being index-wrapped and reuse is dead."
          fi

          # FORCE is checked here rather than in the build step's `if:`, so
          # that one decision drives everything downstream. The repoint step
          # keys off `hit` too, and a force that bypassed only the build would
          # leave the two disagreeing about what just happened.
          if [ "${FORCE:-false}" = "true" ]; then
            echo "hit=false" >> "$GITHUB_OUTPUT"
            echo "reuse: force_build set — building regardless"
          elif [ "${IS_REFRESH:-}" = "true" ]; then
            echo "hit=false" >> "$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"
          else
            echo "hit=false" >> "$GITHUB_OUTPUT"
            echo "reuse: not published — building"
          fi

      - name: Build and push ml image
        if: steps.reuse.outputs.hit != 'true'
        # Read by buildx out of the ENVIRONMENT, not passed as a build-arg —
        # it normalises the image config's `created` field and the history
        # timestamps rather than being consumed by the Dockerfile. See #3265
        # and the reuse step's `epoch` output.
        env:
          SOURCE_DATE_EPOCH: ${{ steps.reuse.outputs.epoch }}
        uses: docker/build-push-action@v5
        with:
          context: .
          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 was ~13s with every content step CACHED —
          # and the channel tag STILL got a new manifest digest, because
          # buildkit stamps a fresh image config per run and republishes the
          # identical layers under it. All three images moved that way on
          # 2026-08-30 with nothing whatsoever changed in them.
          #
          # SOURCE_DATE_EPOCH (below) is the fix: pinned to the commit the
          # content came from, the config is byte-identical across runs, so
          # the manifest digest is too and the push is a registry no-op. A
          # digest change means the content changed again, which is the only
          # thing a digest is any use for.
          #
          # 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: ${{ env.IS_REFRESH == 'true' }}
          # ONE tag, the channel's. Every other tag is written by the step
          # below, registry-side. buildx here pushes the first tag to the
          # registry and then re-pushes the rest through the DOCKER driver,
          # out of a local image store a registry-direct build never filled —
          # #3190, which cost `main` its :c-<sha> on 2026-08-29 while :latest
          # published perfectly well.
          tags: ${{ steps.reuse.outputs.build_ref }}
          # The reuse key. Read back off the channel tag on the next push to
          # decide whether that push needs to build at all, so this is not
          # decoration — an unstamped image is one that will always rebuild.
          labels: |
            fc.revision=${{ steps.reuse.outputs.revision }}
          # LOAD-BEARING, not a preference. On the default docker driver these
          # were no-ops; on the docker-container driver above,
          # build-push-action@v5 defaults provenance to TRUE when pushing.
          # Provenance attaches an attestation manifest, which makes the pushed
          # tag a manifest INDEX — and `.Image.Config.Labels` does not resolve
          # through an index.
          #
          # The label directly above IS the reuse key. Wrap the channel tag in
          # an index and the next push reads fc.revision=<none>, misses, and
          # rebuilds. Then so does the one after that, forever. Nothing fails,
          # nothing goes red, and the only symptom is the bill. That is #3183
          # arriving through a different door, and note #3127 §4 records the
          # same shape for `platforms:`.
          provenance: false
          sbom: false
          # The ONLY cache this driver can have. `docker-container` gets a
          # FRESH buildkit instance per job, so unlike the default docker
          # driver it has no local layer store to fall back on — measured on
          # run 4896, the first builds after the driver change: web 3m44s
          # (was 2m23s), ml 3m49s (was 3m20s), agent 11m12s (was 9m26s). The
          # driver change ALONE is a regression; this is the other half of it.
          #
          # mode=max so intermediate stages cache too. web's frontend-builder
          # stage and the agent's two ~150s pip layers are the whole cost, and
          # they are exactly what a min-mode cache would drop.
          #
          # A `:buildcache` tag is NOT the withdrawn tag scheme coming back.
          # Rule 145 narrowed against names NOTHING reads; this one is read by
          # every build that runs, is one moving ref per image rather than one
          # per build, holds cache blobs rather than a shippable artifact, and
          # is overwritten in place rather than accumulating. It is closer to
          # :dev than to the :2026.8.28 tags milestone 318 deleted. (#3114.)
          cache-from: type=registry,ref=git.fabledsword.com/bvandeusen/fabledcurator-ml:buildcache
          cache-to: type=registry,ref=git.fabledsword.com/bvandeusen/fabledcurator-ml:buildcache,mode=max

      # Point the channel tag at the candidate the refresh just built.
      #
      # Unconditional TODAY, so this milestone never leaves the refresh in a
      # state where it builds and publishes nothing. Step 4 wraps it in the
      # smoke suite's verdict; until then the scheduled path behaves exactly
      # as it did, just via two operations instead of one.
      #
      # NOT `imagetools create`. That wraps its source in an INDEX, and an
      # indexed channel tag is the one thing this pipeline cannot survive:
      # `.Image.Config.Labels` does not resolve through an index, so the
      # fc.revision the reuse check reads off the channel tag would come back
      # empty, every subsequent push would miss and rebuild, and nothing would
      # go red. That is #3183, observed on run 4751 — reuse worked exactly once
      # and the only symptom was the bill. The repoint step below excludes its
      # own source tag for precisely this reason; a promote that re-introduced
      # the wrap through a different door would undo that care.
      #
      # A manifest PUT is what "make this tag name that image" means at the
      # registry level: the same bytes under the same media type, so the digest
      # is identical, the media type is preserved, and no layer moves.
      - name: Promote the refresh candidate to the channel
        if: steps.reuse.outputs.promote == 'true'
        env:
          IMAGE: git.fabledsword.com/bvandeusen/fabledcurator-ml
          CHANNEL_REF: ${{ steps.reuse.outputs.channel_ref }}
          TOKEN: ${{ secrets.RELEASE_TOKEN }}
          ACTOR: ${{ github.actor }}
        run: |
          set -eu
          REPO=${IMAGE#git.fabledsword.com/}
          TAG=${CHANNEL_REF##*:}

          # Registry auth is its own token exchange — the `docker login` above
          # authenticates the docker client, not curl. Deadline on every call
          # (rule 156): a registry that stops answering must fail this step,
          # not hang the weekly refresh until the job times out.
          BEARER=$(curl -fsS --max-time 30 -u "$ACTOR:$TOKEN" \
            "https://git.fabledsword.com/v2/token?scope=repository:$REPO:pull,push&service=git.fabledsword.com" \
            | python3 -c 'import sys,json; print(json.load(sys.stdin)["token"])')

          # Ask for the image manifest media types ONLY. Offering the index
          # types too would let the registry hand back an index if one ever
          # existed at this tag, and we would faithfully copy the thing we are
          # trying not to create.
          ACCEPT='application/vnd.oci.image.manifest.v1+json, application/vnd.docker.distribution.manifest.v2+json'
          CT=$(curl -fsS --max-time 60 -o manifest.json -D headers.txt \
                 -H "Authorization: Bearer $BEARER" -H "Accept: $ACCEPT" \
                 "https://git.fabledsword.com/v2/$REPO/manifests/refresh-candidate" \
               && tr -d '\r' < headers.txt | awk -F': ' '/^[Cc]ontent-[Tt]ype:/{print $2}')
          test -n "$CT"
          SRC_DIGEST=$(tr -d '\r' < headers.txt | awk -F': ' '/^[Dd]ocker-[Cc]ontent-[Dd]igest:/{print $2}')
          echo "promote: candidate is $SRC_DIGEST ($CT)"

          curl -fsS --max-time 120 -X PUT \
            -H "Authorization: Bearer $BEARER" -H "Content-Type: $CT" \
            --data-binary @manifest.json \
            "https://git.fabledsword.com/v2/$REPO/manifests/$TAG"

          # Read it back. A PUT that returned 2xx but landed something else is
          # exactly the silent-and-plausible failure this pipeline keeps
          # producing, and the check costs one request.
          NOW=$(curl -fsS --max-time 30 -o /dev/null -D - \
                  -H "Authorization: Bearer $BEARER" -H "Accept: $ACCEPT" \
                  "https://git.fabledsword.com/v2/$REPO/manifests/$TAG" \
                | tr -d '\r' | awk -F': ' '/^[Dd]ocker-[Cc]ontent-[Dd]igest:/{print $2}')
          if [ "$NOW" != "$SRC_DIGEST" ]; then
            echo "promote: $IMAGE:$TAG is $NOW, expected $SRC_DIGEST" >&2
            exit 1
          fi
          echo "promote: $IMAGE:$TAG now names $NOW"

      # Every tag but the channel's own is written HERE, registry-side,
      # whether or not a build ran. Each -t becomes another reference to the
      # SAME manifest the channel tag holds, so :c-<sha> is byte-identical to
      # what is published rather than a lookalike rebuild.
      #
      # Owning the build path too is #3190's fix, not a tidy-up:
      #
      #   #27 pushing …/fabledcurator:latest           DONE 15.8s
      #   #28 pushing …/fabledcurator:c-0e15c44 with docker
      #   #28 ERROR: tag does not exist: …:c-0e15c44
      #
      # Intermittent — build-ml made the identical two-tag push seconds later
      # and succeeded — and worse than it looks. `:latest` had already
      # published, so production was correct while the immutable rollback tag
      # rule 145 requires of every main push simply did not exist. Nothing but
      # the red job would ever have noticed: a missing :c-<sha> has no
      # consumer that fails, so it surfaces when somebody needs to roll back.
      #
      # `imagetools create` is a registry-side manifest copy — no layer
      # transfer, no local daemon, nothing that can be absent. The reuse case
      # has always gone this way, so this puts the build case on the code that
      # was already proven rather than on a second path.
      #
      # Running on every path also keeps family rule 146 true: a rolling
      # channel refreshes itself, so skipping a build must never leave :dev or
      # :latest pointing at something older than the commit just pushed.
      #
      # The cost, accepted knowingly: `imagetools create` wraps its source in
      # an index, so :c-<sha> becomes an index and fc.revision does not
      # resolve through it. Nothing reads that label off :c-<sha> — the reuse
      # check only ever inspects the CHANNEL tag — and the index names the
      # same manifest, so a pull is byte-identical. The reuse path already
      # produced :c-<sha> this way; this only makes it uniform.
      - name: Write the remaining tags from the published image
        env:
          IMAGE: git.fabledsword.com/bvandeusen/fabledcurator-ml
          SOURCE: ${{ steps.reuse.outputs.channel_ref }}
          TAGS: ${{ steps.tag.outputs.tags }}
        run: |
          set -euf
          # The source tag is EXCLUDED from the targets, and that is load-
          # bearing rather than an optimisation.
          #
          # `imagetools create` wraps the source manifest in an INDEX. Point it
          # at the channel tag with that same tag as a target and the tag stops
          # being a plain image — after which `.Image.Config.Labels` no longer
          # resolves through it and the fc.revision label reads as absent. The
          # next push then misses and rebuilds, so reuse worked exactly once
          # and every subsequent push paid full price. Observed on run 4751:
          # ml:dev reported fc.revision=<none> one push after run 4749 had read
          # a7e626a67a79 off it. Nothing failed; the savings just evaporated.
          #
          # Excluding the source means the channel tag is only ever written
          # by a real build, so it stays a plain image and stays readable.
          # On dev that leaves nothing to do either way: the build pushed :dev
          # itself, or the hit established it was already right. On main it
          # leaves :c-<sha>, which rule 145 requires of every main push whether
          # or not a build ran.
          #
          # steps.tag emits ONE comma-separated list; imagetools wants a -t per
          # ref. (That list used to feed docker/build-push-action directly —
          # which is exactly what #3190 made unsafe.)
          ARGS=""
          IFS=,
          for t in $TAGS; do
            [ "$t" = "$SOURCE" ] && continue
            ARGS="$ARGS -t $t"
          done
          unset IFS
          if [ -z "$ARGS" ]; then
            echo "repoint: $SOURCE is the only tag for this channel and"
            echo "repoint: already holds this revision — nothing to write."
            exit 0
          fi
          # shellcheck disable=SC2086
          docker buildx imagetools create $ARGS "$SOURCE"
          echo "repointed from $SOURCE:$ARGS"

  # The desktop GPU agent (#114) — published so the operator pulls + runs it on
  # the GPU machine instead of building locally. Independent of web/ml (its own
  # CUDA + onnxruntime-gpu image, context = agent/). Same tag cadence.
  build-agent:
    runs-on: python-ci
    container:
      image: git.fabledsword.com/bvandeusen/ci-python:3.14
    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,
          # too-low value or finds nothing at all, and neither is a failure
          # 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: env.IS_REFRESH == 'true'
        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
      # reports about itself; the load-bearing steps each print only the one
      # they use, so this is the only place the pair appears together. When a
      # build is skipped, this is the line that says what the commit derived.
      #
      # Still diagnostic, so it still must not fail the build — no `set -e`,
      # and every derivation falls back to UNAVAILABLE. A broken echo must
      # never be the reason an image does not ship.
      #
      # What it should say:
      #   * a push touching only agent/ moves the agent and leaves web and ml
      #     STILL. If web moves, its path set is too wide.
      #   * a push touching only docs moves nothing.
      #   * a push touching the extension moves the extension AND web, since
      #     web bakes in the XPI. If web does not move, its set is too narrow:
      #     the reuse check hits, and the channel serves a web image bundling
      #     the PREVIOUS XPI while the freshly signed one is orphaned (#3156).
      #   * dev and main derive the same values for the same source.
      - name: Report the derived artifact version
        run: |
          set -u
          A=agent
          V=$(sh scripts/artifacts.sh version "$A" 2>&1 || echo UNAVAILABLE)
          R=$(sh scripts/artifacts.sh revision "$A" 2>&1 || echo UNAVAILABLE)
          echo "derived: artifact=$A version=$V revision=$R sha=$GITHUB_SHA"

      - name: Determine tag
        id: tag
        run: |
          SHORT_SHA=$(printf '%s' "$GITHUB_SHA" | cut -c1-7)
          # Mirrors build-web's tag list and its schedule handling; see
          # the comments there.
          if [ "${IS_REFRESH:-}" = "true" ]; then
            echo "tags=git.fabledsword.com/bvandeusen/fabledcurator-agent:latest" >> "$GITHUB_OUTPUT"
            echo "channel=main" >> "$GITHUB_OUTPUT"
          elif [ "${GITHUB_REF##*/}" = "main" ]; then
            echo "tags=git.fabledsword.com/bvandeusen/fabledcurator-agent:latest,git.fabledsword.com/bvandeusen/fabledcurator-agent:c-${SHORT_SHA}" >> "$GITHUB_OUTPUT"
            echo "channel=main" >> "$GITHUB_OUTPUT"
          else
            echo "tags=git.fabledsword.com/bvandeusen/fabledcurator-agent:dev" >> "$GITHUB_OUTPUT"
            echo "channel=dev" >> "$GITHUB_OUTPUT"
          fi

      # Shell step rather than docker/login-action — see build-web's note on
      # the shared action-cache race (#3118).
      - name: Login to Forgejo registry
        env:
          TOKEN: ${{ secrets.RELEASE_TOKEN }}
          ACTOR: ${{ github.actor }}
        run: echo "$TOKEN" | docker login git.fabledsword.com -u "$ACTOR" --password-stdin

      # A REAL buildx builder, not the default `docker` driver (#3114, #3190).
      #
      # The default driver builds through the local dockerd. It cannot export a
      # registry cache at all — which is why the agent rebuilds a ~6.3 GB CUDA
      # + torch image from scratch whenever the runner's local cache is cold,
      # measured at 9m26s against 7s warm. It is also #3190's leading suspect:
      # after a registry-direct push it resolves image metadata against a local
      # store the push never filled, and reports `No such image` on an image
      # that published perfectly well three seconds earlier.
      #
      # These jobs run INSIDE a container against a mounted docker socket, so
      # the buildkit container this starts is a SIBLING of the job container,
      # not a child. That works over the socket mount; it had never been tried
      # here before milestone 326 step 1.
      - name: Set up buildx
        uses: docker/setup-buildx-action@v3

      # --- reuse-if-published (milestone 313, step 4) ----------------------
      # Does the image the channel tag already points at carry THIS commit's
      # revision? If so the bytes this job would produce are already published
      # and the build is pure waste: the remaining tags get repointed at that
      # existing manifest instead, registry-side, in seconds.
      #
      # Keyed on an `fc.revision` LABEL rather than on a tag of its own
      # (milestone 318 step 3). A tag would be a name minted per build that one
      # thing reads — what rule 145 narrowed against — and would be prunable
      # under the registry's keep_pattern (#3157), silently expiring the cache.
      # A label rides inside a tag that has to exist anyway.
      #
      # An image with no such label reads as a miss and rebuilds. That is the
      # migration, not a fault: labels cannot be backfilled, since the reuse
      # path copies a manifest and config labels are not manifest annotations.
      # Each artifact pays one rebuild, once.
      #
      # This is what stops a push that touched only `agent/` from rebuilding
      # web and ml, and a merge to main from rebuilding what dev already built.
      #
      # The failure direction is deliberate. An inspect that errors for ANY
      # reason — network, auth, a registry hiccup — reads as a miss and the
      # build runs. Only a genuine 200 skips one, so there is no path here
      # that skips a build that was actually needed; the worst case is paying
      # for a build we could have avoided.
      #
      # BASE-IMAGE FRESHNESS: an artifact whose source stops moving stops
      # picking up base-image updates. Milestone 318 removed the argument this
      # used to need rather than answering it — with no version tags there is
      # no immutable name a refresh could contradict, and rule 145 already
      # allows a rebuild with different contents to republish a MOVING tag.
      # So a refresh is just a build. A scheduled channel-only one is tracked
      # separately (#3154); it does not belong in the push path.
      - name: Is this content already published?
        id: reuse
        env:
          IMAGE: git.fabledsword.com/bvandeusen/fabledcurator-agent
          CHANNEL: ${{ steps.tag.outputs.channel }}
          # Empty on a push; the string "true" only from a workflow_dispatch
          # that asked for it. `github.event.inputs` rather than the `inputs`
          # context — release.yml already uses that form, and it is the one
          # this runner is known to evaluate. Read through env rather than
          # interpolated into the run block, same rule as release.yml's TAG.
          FORCE: ${{ github.event.inputs.force_build }}
          # A scheduled refresh has to bypass reuse by construction: it
          # rebuilds the SAME source, so fc.revision always matches and the
          # check would skip every refresh there has ever been.
        run: |
          set -eu
          DERIVED=$(sh scripts/artifacts.sh revision agent)
          echo "revision=$DERIVED" >> "$GITHUB_OUTPUT"

          # The build clock, pinned to the same commit (#3265). Without it
          # buildkit stamps the image config with the wall clock of the build,
          # so identical layers republish under a new config blob and the
          # channel tag gets a new manifest digest for no reason. Derived from
          # `newest()` like revision and version, so all three name one commit
          # and cannot drift apart.
          echo "epoch=$(sh scripts/artifacts.sh epoch agent)" >> "$GITHUB_OUTPUT"

          # The moving tag for this channel. Which tag we ask IS the channel —
          # that is why the revision needs no -main/-dev qualifier any more.
          if [ "$CHANNEL" = "main" ]; then T=latest; else T=dev; fi
          echo "channel_ref=$IMAGE:$T" >> "$GITHUB_OUTPUT"

          # WHERE THE BUILD PUBLISHES, which is not always the channel — and
          # whether the channel then has to be written separately.
          #
          # On a push the build writes the channel tag directly: the bytes came
          # from a commit, and a commit is the thing CI tests. Nothing to hold
          # it behind.
          #
          # On the scheduled refresh it writes a CANDIDATE tag instead. A
          # refresh rebuilds against freshly resolved base images, and the web
          # image's runtime is a line of UNPINNED Debian packages (ffmpeg,
          # libjpeg62-turbo, libpq5, megatools…) re-resolved on every build.
          # Nothing in ci.yml can see that: its lanes run on ci-python:3.14 and
          # install requirements.txt, and a base bump changes neither. So
          # refreshed bytes have to be proven before :latest names them, and
          # proving needs a moment between "built" and "published" to occupy.
          # This is that moment; :latest goes on naming the build that works
          # until something says otherwise.
          #
          # `:refresh-candidate` is one moving ref per image, overwritten in
          # place, holding a build nobody is told to pull — the shape rule 145
          # already allows for :buildcache, not the per-build tag family that
          # milestone 318 withdrew.
          #
          # Both values are decided HERE, beside `hit`, for the reason the
          # force/schedule branch below gives: one step decides what this job
          # does. A promote condition derived independently could disagree with
          # the tag the build actually wrote.
          if [ "${IS_REFRESH:-}" = "true" ]; then
            echo "build_ref=$IMAGE:refresh-candidate" >> "$GITHUB_OUTPUT"
            echo "promote=true" >> "$GITHUB_OUTPUT"
          else
            echo "build_ref=$IMAGE:$T" >> "$GITHUB_OUTPUT"
            echo "promote=false" >> "$GITHUB_OUTPUT"
          fi

          # Compare VALUES, never exit codes. Measured on buildx v0.36.1
          # (run 4732): a missing key returns an empty string and exits 0, so
          # branching on the exit code would read "no label yet" as success.
          # An unreachable tag also lands here as empty via the `|| echo`.
          # Empty never equals a 12-char revision, so every uncertain case
          # falls through to a build — the safe direction, with no special
          # casing for it.
          #
          # Read the SPECIFIC key. The map also carries whatever the base image
          # set, and `org.opencontainers.image.version` sits right beside ours
          # looking like a plausible answer (it reads 24.04 on the agent).
          PUBLISHED=$(docker buildx imagetools inspect "$IMAGE:$T" \
            --format '{{ index .Image.Config.Labels "fc.revision" }}' \
            2>/dev/null || echo "")
          echo "reuse: $IMAGE:$T carries fc.revision=${PUBLISHED:-<none>}; derived=$DERIVED"
          if [ -z "$PUBLISHED" ] && docker buildx imagetools inspect "$IMAGE:$T" >/dev/null 2>&1; then
            # The tag resolves but carries no readable label. Expected exactly
            # once per artifact, during the migration onto labels. If it recurs
            # every push, something is rewriting the channel tag as a manifest
            # index — see the repoint step's note.
            echo "reuse: NOTE $IMAGE:$T exists but has no readable fc.revision."
            echo "reuse: NOTE Fine once, while migrating. Every push means the"
            echo "reuse: NOTE tag is being index-wrapped and reuse is dead."
          fi

          # FORCE is checked here rather than in the build step's `if:`, so
          # that one decision drives everything downstream. The repoint step
          # keys off `hit` too, and a force that bypassed only the build would
          # leave the two disagreeing about what just happened.
          if [ "${FORCE:-false}" = "true" ]; then
            echo "hit=false" >> "$GITHUB_OUTPUT"
            echo "reuse: force_build set — building regardless"
          elif [ "${IS_REFRESH:-}" = "true" ]; then
            echo "hit=false" >> "$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"
          else
            echo "hit=false" >> "$GITHUB_OUTPUT"
            echo "reuse: not published — building"
          fi

      - name: Build and push agent image
        if: steps.reuse.outputs.hit != 'true'
        # Read by buildx out of the ENVIRONMENT, not passed as a build-arg —
        # it normalises the image config's `created` field and the history
        # timestamps rather than being consumed by the Dockerfile. See #3265
        # and the reuse step's `epoch` output.
        env:
          SOURCE_DATE_EPOCH: ${{ steps.reuse.outputs.epoch }}
        uses: docker/build-push-action@v5
        with:
          context: agent
          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 was ~13s with every content step CACHED —
          # and the channel tag STILL got a new manifest digest, because
          # buildkit stamps a fresh image config per run and republishes the
          # identical layers under it. All three images moved that way on
          # 2026-08-30 with nothing whatsoever changed in them.
          #
          # SOURCE_DATE_EPOCH (below) is the fix: pinned to the commit the
          # content came from, the config is byte-identical across runs, so
          # the manifest digest is too and the push is a registry no-op. A
          # digest change means the content changed again, which is the only
          # thing a digest is any use for.
          #
          # 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: ${{ env.IS_REFRESH == 'true' }}
          # ONE tag, the channel's. Every other tag is written by the step
          # below, registry-side. buildx here pushes the first tag to the
          # registry and then re-pushes the rest through the DOCKER driver,
          # out of a local image store a registry-direct build never filled —
          # #3190, which cost `main` its :c-<sha> on 2026-08-29 while :latest
          # published perfectly well.
          tags: ${{ steps.reuse.outputs.build_ref }}
          # The reuse key. Read back off the channel tag on the next push to
          # decide whether that push needs to build at all, so this is not
          # decoration — an unstamped image is one that will always rebuild.
          labels: |
            fc.revision=${{ steps.reuse.outputs.revision }}
          # LOAD-BEARING, not a preference. On the default docker driver these
          # were no-ops; on the docker-container driver above,
          # build-push-action@v5 defaults provenance to TRUE when pushing.
          # Provenance attaches an attestation manifest, which makes the pushed
          # tag a manifest INDEX — and `.Image.Config.Labels` does not resolve
          # through an index.
          #
          # The label directly above IS the reuse key. Wrap the channel tag in
          # an index and the next push reads fc.revision=<none>, misses, and
          # rebuilds. Then so does the one after that, forever. Nothing fails,
          # nothing goes red, and the only symptom is the bill. That is #3183
          # arriving through a different door, and note #3127 §4 records the
          # same shape for `platforms:`.
          provenance: false
          sbom: false
          # The ONLY cache this driver can have. `docker-container` gets a
          # FRESH buildkit instance per job, so unlike the default docker
          # driver it has no local layer store to fall back on — measured on
          # run 4896, the first builds after the driver change: web 3m44s
          # (was 2m23s), ml 3m49s (was 3m20s), agent 11m12s (was 9m26s). The
          # driver change ALONE is a regression; this is the other half of it.
          #
          # mode=max so intermediate stages cache too. web's frontend-builder
          # stage and the agent's two ~150s pip layers are the whole cost, and
          # they are exactly what a min-mode cache would drop.
          #
          # A `:buildcache` tag is NOT the withdrawn tag scheme coming back.
          # Rule 145 narrowed against names NOTHING reads; this one is read by
          # every build that runs, is one moving ref per image rather than one
          # per build, holds cache blobs rather than a shippable artifact, and
          # is overwritten in place rather than accumulating. It is closer to
          # :dev than to the :2026.8.28 tags milestone 318 deleted. (#3114.)
          cache-from: type=registry,ref=git.fabledsword.com/bvandeusen/fabledcurator-agent:buildcache
          cache-to: type=registry,ref=git.fabledsword.com/bvandeusen/fabledcurator-agent:buildcache,mode=max

      # Point the channel tag at the candidate the refresh just built.
      #
      # Unconditional TODAY, so this milestone never leaves the refresh in a
      # state where it builds and publishes nothing. Step 4 wraps it in the
      # smoke suite's verdict; until then the scheduled path behaves exactly
      # as it did, just via two operations instead of one.
      #
      # NOT `imagetools create`. That wraps its source in an INDEX, and an
      # indexed channel tag is the one thing this pipeline cannot survive:
      # `.Image.Config.Labels` does not resolve through an index, so the
      # fc.revision the reuse check reads off the channel tag would come back
      # empty, every subsequent push would miss and rebuild, and nothing would
      # go red. That is #3183, observed on run 4751 — reuse worked exactly once
      # and the only symptom was the bill. The repoint step below excludes its
      # own source tag for precisely this reason; a promote that re-introduced
      # the wrap through a different door would undo that care.
      #
      # A manifest PUT is what "make this tag name that image" means at the
      # registry level: the same bytes under the same media type, so the digest
      # is identical, the media type is preserved, and no layer moves.
      - name: Promote the refresh candidate to the channel
        if: steps.reuse.outputs.promote == 'true'
        env:
          IMAGE: git.fabledsword.com/bvandeusen/fabledcurator-agent
          CHANNEL_REF: ${{ steps.reuse.outputs.channel_ref }}
          TOKEN: ${{ secrets.RELEASE_TOKEN }}
          ACTOR: ${{ github.actor }}
        run: |
          set -eu
          REPO=${IMAGE#git.fabledsword.com/}
          TAG=${CHANNEL_REF##*:}

          # Registry auth is its own token exchange — the `docker login` above
          # authenticates the docker client, not curl. Deadline on every call
          # (rule 156): a registry that stops answering must fail this step,
          # not hang the weekly refresh until the job times out.
          BEARER=$(curl -fsS --max-time 30 -u "$ACTOR:$TOKEN" \
            "https://git.fabledsword.com/v2/token?scope=repository:$REPO:pull,push&service=git.fabledsword.com" \
            | python3 -c 'import sys,json; print(json.load(sys.stdin)["token"])')

          # Ask for the image manifest media types ONLY. Offering the index
          # types too would let the registry hand back an index if one ever
          # existed at this tag, and we would faithfully copy the thing we are
          # trying not to create.
          ACCEPT='application/vnd.oci.image.manifest.v1+json, application/vnd.docker.distribution.manifest.v2+json'
          CT=$(curl -fsS --max-time 60 -o manifest.json -D headers.txt \
                 -H "Authorization: Bearer $BEARER" -H "Accept: $ACCEPT" \
                 "https://git.fabledsword.com/v2/$REPO/manifests/refresh-candidate" \
               && tr -d '\r' < headers.txt | awk -F': ' '/^[Cc]ontent-[Tt]ype:/{print $2}')
          test -n "$CT"
          SRC_DIGEST=$(tr -d '\r' < headers.txt | awk -F': ' '/^[Dd]ocker-[Cc]ontent-[Dd]igest:/{print $2}')
          echo "promote: candidate is $SRC_DIGEST ($CT)"

          curl -fsS --max-time 120 -X PUT \
            -H "Authorization: Bearer $BEARER" -H "Content-Type: $CT" \
            --data-binary @manifest.json \
            "https://git.fabledsword.com/v2/$REPO/manifests/$TAG"

          # Read it back. A PUT that returned 2xx but landed something else is
          # exactly the silent-and-plausible failure this pipeline keeps
          # producing, and the check costs one request.
          NOW=$(curl -fsS --max-time 30 -o /dev/null -D - \
                  -H "Authorization: Bearer $BEARER" -H "Accept: $ACCEPT" \
                  "https://git.fabledsword.com/v2/$REPO/manifests/$TAG" \
                | tr -d '\r' | awk -F': ' '/^[Dd]ocker-[Cc]ontent-[Dd]igest:/{print $2}')
          if [ "$NOW" != "$SRC_DIGEST" ]; then
            echo "promote: $IMAGE:$TAG is $NOW, expected $SRC_DIGEST" >&2
            exit 1
          fi
          echo "promote: $IMAGE:$TAG now names $NOW"

      # Every tag but the channel's own is written HERE, registry-side,
      # whether or not a build ran. Each -t becomes another reference to the
      # SAME manifest the channel tag holds, so :c-<sha> is byte-identical to
      # what is published rather than a lookalike rebuild.
      #
      # Owning the build path too is #3190's fix, not a tidy-up:
      #
      #   #27 pushing …/fabledcurator:latest           DONE 15.8s
      #   #28 pushing …/fabledcurator:c-0e15c44 with docker
      #   #28 ERROR: tag does not exist: …:c-0e15c44
      #
      # Intermittent — build-ml made the identical two-tag push seconds later
      # and succeeded — and worse than it looks. `:latest` had already
      # published, so production was correct while the immutable rollback tag
      # rule 145 requires of every main push simply did not exist. Nothing but
      # the red job would ever have noticed: a missing :c-<sha> has no
      # consumer that fails, so it surfaces when somebody needs to roll back.
      #
      # `imagetools create` is a registry-side manifest copy — no layer
      # transfer, no local daemon, nothing that can be absent. The reuse case
      # has always gone this way, so this puts the build case on the code that
      # was already proven rather than on a second path.
      #
      # Running on every path also keeps family rule 146 true: a rolling
      # channel refreshes itself, so skipping a build must never leave :dev or
      # :latest pointing at something older than the commit just pushed.
      #
      # The cost, accepted knowingly: `imagetools create` wraps its source in
      # an index, so :c-<sha> becomes an index and fc.revision does not
      # resolve through it. Nothing reads that label off :c-<sha> — the reuse
      # check only ever inspects the CHANNEL tag — and the index names the
      # same manifest, so a pull is byte-identical. The reuse path already
      # produced :c-<sha> this way; this only makes it uniform.
      - name: Write the remaining tags from the published image
        env:
          IMAGE: git.fabledsword.com/bvandeusen/fabledcurator-agent
          SOURCE: ${{ steps.reuse.outputs.channel_ref }}
          TAGS: ${{ steps.tag.outputs.tags }}
        run: |
          set -euf
          # The source tag is EXCLUDED from the targets, and that is load-
          # bearing rather than an optimisation.
          #
          # `imagetools create` wraps the source manifest in an INDEX. Point it
          # at the channel tag with that same tag as a target and the tag stops
          # being a plain image — after which `.Image.Config.Labels` no longer
          # resolves through it and the fc.revision label reads as absent. The
          # next push then misses and rebuilds, so reuse worked exactly once
          # and every subsequent push paid full price. Observed on run 4751:
          # ml:dev reported fc.revision=<none> one push after run 4749 had read
          # a7e626a67a79 off it. Nothing failed; the savings just evaporated.
          #
          # Excluding the source means the channel tag is only ever written
          # by a real build, so it stays a plain image and stays readable.
          # On dev that leaves nothing to do either way: the build pushed :dev
          # itself, or the hit established it was already right. On main it
          # leaves :c-<sha>, which rule 145 requires of every main push whether
          # or not a build ran.
          #
          # steps.tag emits ONE comma-separated list; imagetools wants a -t per
          # ref. (That list used to feed docker/build-push-action directly —
          # which is exactly what #3190 made unsafe.)
          ARGS=""
          IFS=,
          for t in $TAGS; do
            [ "$t" = "$SOURCE" ] && continue
            ARGS="$ARGS -t $t"
          done
          unset IFS
          if [ -z "$ARGS" ]; then
            echo "repoint: $SOURCE is the only tag for this channel and"
            echo "repoint: already holds this revision — nothing to write."
            exit 0
          fi
          # shellcheck disable=SC2086
          docker buildx imagetools create $ARGS "$SOURCE"
          echo "repointed from $SOURCE:$ARGS"