cut-rc.yml

name: Cut RC

# ══════════════════════════════════════════════════════════════════════════════
# ONE DISPATCH, ONE SNAPSHOT, ZERO FREEZES. (#7447)
# ══════════════════════════════════════════════════════════════════════════════
#
# Read `release.yml`'s header first — its 2026-08-07 ruling ("版本发布必须是人工的")
# and its lane split are binding context for this file, and this lane preserves
# both. Nothing here weakens them:
#
#   - `workflow_dispatch` is the ONLY trigger. No push, bot, merge-queue landing
#     or schedule can synthesise the event; a human types the exact version.
#   - `environment: release` gates it exactly as the publish lane is gated.
#   - The publish still only ever ships a commit that is ALREADY on main
#     (#6170): this lane PUSHES the version commit to main first and publishes
#     from it second. It never publishes from a ref that only exists in the run.
#
# WHAT THIS SOLVES THAT `release.yml` CANNOT
# ------------------------------------------
# The standing "Version Packages" PR (#6208) is force-refreshed on every push to
# main, so its CI can never converge while main is busy — on a repo with ~18
# merges a working day, cutting through it means asking two repositories to hold
# still. rc.6 was chased across four pin-bump laps
# (8aad9fd -> 9b9fa49 -> cfeb378 -> 1b6188d -> bcd3e02), every lap overtaken
# before its CI finished, and finishing would have needed ~40 minutes of
# coordinated freezes.
#
# The fix is not more discipline, it is a SNAPSHOT: this lane records the
# objectstack `main` sha it checked out and does every downstream step against
# that one frozen value. main may keep moving freely for the whole run.
#
# THE objectui PIN IS AN INPUT HERE, NEVER A DECISION (#10134)
# ------------------------------------------------------------
# Maintainer ruling, 2026-08-20, verbatim and untranslated:
#
#   > objectui 每次 pin 的时候，对应的 changeset 应该带过来，但是发版本的时候不需要去
#   > 扫描 objectui 仓库的最新版
#   >
#   > 更新到哪个版本是 objectstack 的 issue 自己决定的
#
# So this lane READS `.objectui-sha` and builds against it. It does not resolve
# objectui `main`, does not compare the pin to anything in another repo, and does
# not move the pin. The pin moves only on a manual bump PR whose target revision
# was decided in an objectstack issue (#10129 is the pattern).
#
# ⛔ Do not re-add a currency check here or anywhere on a release path. It is not
# that liveness is expensive — it is that "the pin is behind objectui main" is
# not a defect. #3340's actual invariant is "everything published is covered by
# the changeset record", and `scripts/bump-objectui.sh` +
# `scripts/objectui-changeset-digest.mjs` deliver it at BUMP time by carrying
# objectui's own changesets across OLD..NEW. That carry-over is the mechanism;
# scanning objectui mid-release was a second, redundant attempt at the same goal
# that additionally reached into another repo to second-guess a decision.
#
# CONSEQUENCE, STATED PLAINLY: an rc ships whatever console the committed pin
# names. If the pin is old, the rc's frontend is old, and nothing here will warn
# about it. That is the design, not an oversight — the pin is a decision.
#
# SCOPE — rc PRERELEASES ONLY
# ---------------------------
# The guard step below refuses anything that is not `X.Y.Z-rc.N`, and refuses to
# run at all unless `.changeset/pre.json` is in `mode: pre` with `tag: rc`. GA cuts
# keep the `release.yml` + Version-PR flow, where board-clearing, the #7275-A cut
# precondition and human changelog review live. See `docs/releases-maintenance.md`.
#
# THE RUNTIME IMAGE IS DELIBERATELY NOT BUILT HERE
# ------------------------------------------------
# This lane publishes to npm and pushes tags; it does not call `docker-publish.yml`.
# That is not an oversight and not a gap that stays open: `release.yml`'s
# `release-integrity` lane runs on EVERY push to main, and once this lane's version
# is on npm the next main push finds the image missing and requests the build — the
# #4900 repair path, doing exactly the job it was written for. On a repo with ~18
# merges a working day that is minutes, not days. For an image immediately, dispatch
# `docker-publish.yml` with the version. Duplicating the job here would be a second
# copy of a repair path that already exists.

on:
  # The human lane, and the only one. Deliberately no `push:`, no `schedule:`.
  workflow_dispatch:
    inputs:
      version:
        description: >-
          The EXPECTED resulting version, e.g. 17.0.0-rc.6 — what the version pass
          should compute from the changesets pending on main, not what main carries
          now. Typing it is the human confirmation the 2026-08-07 ruling requires:
          the run fails before anything irreversible if the computed version differs.
        required: true
        type: string
      dry_run:
        description: >-
          Stop after building the version commit locally and upload it as a patch
          artifact. Nothing is pushed and nothing is published. Run this first —
          it means the first REAL dispatch is not this workflow's first execution.
        required: false
        type: boolean
        default: false

# ONE group for the whole lane, so two cuts can never interleave two snapshots
# over one main. Note the eviction property `release.yml` documents: GitHub keeps
# at most one PENDING run per group, so dispatching a third cut while one runs and
# one waits silently drops the middle one. That is acceptable here and nowhere near
# the hazard it would be on a shared group — every run of this workflow is a
# deliberate human act, and a dropped one is visibly absent from the Actions list.
# No `cancel-in-progress`: a cut that has begun publishing must never be killed.
concurrency:
  group: cut-rc
  cancel-in-progress: false

permissions:
  contents: read

jobs:
  cut:
    name: Cut an rc from one snapshot
    runs-on: ubuntu-latest
    # ⚠️ An environment with NO protection rules passes AUTOMATICALLY and silently
    # — the same one-time setup note `release.yml`'s header carries applies here
    # verbatim (Settings -> Environments -> release -> Required reviewers). The
    # load-bearing guarantee until then is the `workflow_dispatch` trigger itself.
    environment: release
    permissions:
      # For `git push origin main` and the tag push. See "THE ADMIN PREREQUISITE"
      # on the checkout step: this permission is necessary and NOT sufficient when
      # main carries a protection ruleset.
      contents: write
    timeout-minutes: 120

    steps:
      # ────────────────────────────────────────────────────────────────────────
      # GUARDS THAT NEED NO CHECKOUT. Cheapest possible failure for the mistakes
      # that are most likely: wrong event, wrong branch, wrong version shape.
      # ────────────────────────────────────────────────────────────────────────
      - name: Guard the dispatch (event, ref, version shape)
        env:
          # Human-supplied text: read through env, never interpolated into the
          # shell. Same rule as `release.yml`'s guard step.
          REQUESTED: ${{ inputs.version }}
          DRY_RUN: ${{ inputs.dry_run }}
        run: |
          # Belt and braces against a FUTURE edit adding a second trigger to this
          # file. Today `workflow_dispatch` is the only one, so this cannot fire;
          # the day someone adds `push:` for convenience, it fires instead of
          # publishing. A structural guard is cheap; re-earning #6170 is not.
          if [ "${GITHUB_EVENT_NAME}" != "workflow_dispatch" ]; then
            echo "::error::cut-rc ran on event '${GITHUB_EVENT_NAME}'. This lane publishes to npm and may ONLY be started by a human dispatch (#6170, 2026-08-07 ruling). Remove whatever trigger produced this event."
            exit 1
          fi

          if [ "${GITHUB_REF}" != "refs/heads/main" ]; then
            echo "::error::dispatch cut-rc from main (got ${GITHUB_REF}). This workflow snapshots and pushes main; running it from another ref would execute that branch's workflow file against main's code."
            exit 1
          fi

          # rc ONLY. A GA cut goes through release.yml + the Version Packages PR,
          # where the human changelog review and the #7275-A precondition live.
          if ! printf '%s' "$REQUESTED" | grep -qE '^[0-9]+\.[0-9]+\.[0-9]+-rc\.[0-9]+$'; then
            echo "::error::'${REQUESTED}' is not an rc version. cut-rc cuts PRERELEASES only and expects the exact shape X.Y.Z-rc.N (e.g. 17.0.0-rc.6). For a GA release use the Release workflow and the 'chore: version packages' PR — see docs/releases-maintenance.md."
            exit 1
          fi

          echo "Requested: ${REQUESTED} (dry_run=${DRY_RUN}), dispatched by ${GITHUB_ACTOR}."

      # ────────────────────────────────────────────────────────────────────────
      # THE SNAPSHOT, HALF 1 — objectstack.
      #
      # `ref: main` is explicit rather than inherited: the guard above already
      # requires the dispatch ref to be main, and pinning it here means the two
      # can never drift apart in a later edit.
      #
      # `fetch-depth: 0` is required, not tidiness: the changeset gates below diff
      # against the snapshot sha, and `changesets` reads history.
      #
      # ⚠️ THE ADMIN PREREQUISITE — ONE-TIME, MAINTAINER-ONLY, AND THIS LANE
      # CANNOT CREATE IT FOR ITSELF.
      # This job pushes the version commit straight to `main`. If main carries a
      # protection ruleset (it does), the pushing identity must be on that
      # ruleset's BYPASS list, and `contents: write` alone does not put it there.
      # Either of these satisfies it:
      #
      #   (a) add the GitHub Actions app to the ruleset's bypass list, ideally
      #       scoped to this workflow, and leave RELEASE_PUSH_TOKEN unset; or
      #   (b) create a fine-grained PAT with `contents: write` on this repository
      #       for an account that IS on the bypass list, and store it as the
      #       repository secret RELEASE_PUSH_TOKEN.
      #
      # The token below picks (b) when the secret exists and (a) when it does not,
      # so configuring either one is enough and neither needs a workflow edit. The
      # push step names this prerequisite in its failure message, because a
      # rejected push is otherwise indistinguishable from a network fault.
      #
      # Whichever route is taken, the credential is persisted on `origin` here so
      # the version-commit push AND `release-publish.sh`'s atomic tag push both use
      # it — one credential, no second remote to keep in step.
      # ────────────────────────────────────────────────────────────────────────
      - name: Checkout main (full history)
        uses: actions/checkout@v7
        with:
          ref: main
          fetch-depth: 0
          token: ${{ secrets.RELEASE_PUSH_TOKEN || github.token }}

      - name: Record the objectstack snapshot + guard the release train
        id: snapshot
        env:
          REQUESTED: ${{ inputs.version }}
          # PRESENCE, not the value: `!= ''` is evaluated by the expression engine
          # and yields 'true'/'false', so the secret itself never reaches the
          # shell. Interpolating `secrets.X` into a `run:` block would put it in
          # the process's argv — masked in the log, but still the pattern
          # release.yml refuses ("read through env, never interpolated").
          HAS_PUSH_TOKEN: ${{ secrets.RELEASE_PUSH_TOKEN != '' }}
        run: |
          SNAPSHOT_SHA="$(git rev-parse HEAD)"
          echo "SNAPSHOT_SHA=${SNAPSHOT_SHA}" >> "$GITHUB_ENV"
          echo "snapshot-sha=${SNAPSHOT_SHA}" >> "$GITHUB_OUTPUT"

          # Report credential state on EVERY run, before anything can early-exit.
          # Otherwise a repo with the secret and one without look identical until a
          # push happens to be rejected, and "is it configured?" stays unanswerable
          # (the pattern cross-repo-issue-closer.yml uses). Presence only — the
          # value is never read into the log.
          if [ "$HAS_PUSH_TOKEN" = "true" ]; then
            echo "push credential: RELEASE_PUSH_TOKEN is configured (route b)."
          else
            echo "push credential: RELEASE_PUSH_TOKEN is NOT set — pushing as the Actions app (route a). That app must be on main's ruleset bypass list or the push below is rejected."
          fi

          # rc-only, asserted against committed state rather than trusted from the
          # version string. Pre mode is what makes `changeset version` compute
          # `-rc.N` at all; outside it this lane would silently cut a GA release
          # through a path that has none of GA's review steps.
          MODE="$(jq -r '.mode // empty' .changeset/pre.json 2>/dev/null || true)"
          TAG="$(jq -r '.tag // empty' .changeset/pre.json 2>/dev/null || true)"
          if [ "$MODE" != "pre" ] || [ "$TAG" != "rc" ]; then
            echo "::error::.changeset/pre.json is mode='${MODE:-<absent>}' tag='${TAG:-<absent>}', but cut-rc requires mode='pre' tag='rc'. Outside pre mode 'pnpm run version' computes a FINAL version, and this lane has none of the GA review steps. Enter pre mode ('changeset pre enter rc') or use the Release workflow."
            exit 1
          fi

          CURRENT="$(jq -r '.version' packages/cli/package.json)"
          if [ "$CURRENT" = "$REQUESTED" ]; then
            echo "::error::main already carries @objectstack/cli@${CURRENT}. cut-rc computes the NEXT version from the pending changesets — asking it for the version main already has means the cut has already happened. To (re)publish an existing version, use the Release workflow."
            exit 1
          fi

          echo "objectstack snapshot: ${SNAPSHOT_SHA}"
          echo "main currently carries @objectstack/cli@${CURRENT}; this run must compute ${REQUESTED}."

      # ────────────────────────────────────────────────────────────────────────
      # THE objectui HALF — READ FROM THE COMMITTED PIN, NOT RESOLVED (#10134).
      #
      # `.objectui-sha` is the whole input. Everything downstream reads
      # $OBJECTUI_SHA, which is that file and nothing else, so objectui may merge
      # freely for the rest of the run and this cut neither notices nor cares —
      # not because a snapshot froze a moving target, but because the target was
      # never objectui `main`. See "THE objectui PIN IS AN INPUT HERE" up top.
      # ────────────────────────────────────────────────────────────────────────
      - name: Read the committed objectui pin
        run: |
          # Shape-checked, not trusted: `.objectui-sha` is hand-edited by
          # bump-objectui.sh and by humans, and a truncated or empty file would
          # otherwise surface much later as an unhelpful clone/build failure.
          OBJECTUI_SHA="$(tr -d '[:space:]' < .objectui-sha)"
          if ! printf '%s' "$OBJECTUI_SHA" | grep -qE '^[0-9a-f]{40}$'; then
            echo "::error::.objectui-sha does not contain a 40-hex commit id (got '${OBJECTUI_SHA}'). Refusing to cut against an unreadable frontend revision."
            exit 1
          fi
          echo "OBJECTUI_SHA=${OBJECTUI_SHA}" >> "$GITHUB_ENV"
          echo "objectui pin (committed, NOT resolved): ${OBJECTUI_SHA}. This cut ships the console at that revision, however old it is — moving the pin is a separate, deliberate PR."

      - name: Clone objectui at full depth, at the pin
        run: |
          OBJECTUI_ROOT="${RUNNER_TEMP}/objectui"
          # FULL clone, not shallow, and still a measured requirement rather than
          # caution. Two reasons, both surviving #10134:
          #   - the check below asks whether the PIN is reachable from objectui
          #     main; a tip-only shallow clone cannot answer that and would answer
          #     "no" for every pin older than the tip. Necessary, not sufficient:
          #     a full clone carries every branch, main's and otherwise, so it
          #     makes the question answerable without answering it;
          #   - build-console.sh builds from THIS clone (it honours
          #     $OBJECTUI_ROOT) by adding a worktree at the pin, which needs the
          #     pin's tree present.
          # The #4731 lesson — a degraded changeset record and a complete one look
          # alike, so never let history quietly go missing — now applies on the
          # bump PR, which is where objectui-changeset-digest.mjs walks OLD..NEW.
          git clone --no-tags https://github.com/objectstack-ai/objectui.git "$OBJECTUI_ROOT"
          echo "OBJECTUI_ROOT=${OBJECTUI_ROOT}" >> "$GITHUB_ENV"

          # THE PIN MUST BE ON objectui MAIN — WHICH IS NOT WHAT OBJECT PRESENCE
          # ANSWERS (#9450). This was one `cat-file -e`, and the sentence it
          # printed on failure — "not reachable from main (unmerged branch, or
          # main was rewritten)" — named a case the test cannot see. Measured on
          # a fresh `--no-tags` clone of objectui, 2026-08-21: 291 commits across
          # 118 branch tips are present and NOT reachable from main, and
          # `cat-file -e` says yes to every one of them — including the "branch
          # that never merged" the message claimed to catch. The clone being FULL
          # is what OPENS that gap rather than closing it: `git clone` fetches
          # every branch head, so the more complete the clone, the more non-main
          # revisions it can vouch for.
          #
          # Nothing upstream closes it either — bump-objectui.sh pins
          # `git rev-parse HEAD` of a local objectui checkout without asking which
          # branch that is. "The operator happened to be on main" is the whole of
          # the protection, so ask the question here rather than assume it.
          #
          # Three questions, three exits, in this order because the later ones
          # cannot be asked until the earlier ones hold: `merge-base --is-ancestor`
          # exits 128 on an absent object — an error, not a verdict — and with no
          # origin/main it would report "the pin left main", which is this block's
          # own overclaim wearing a new message.
          if ! git -C "$OBJECTUI_ROOT" rev-parse --verify --quiet origin/main >/dev/null; then
            echo "::error::the objectui clone has no origin/main ref, so \"is the pin on main\" cannot be answered in it. Refusing to cut rather than assuming the answer."
            exit 1
          fi
          if ! git -C "$OBJECTUI_ROOT" cat-file -e "${OBJECTUI_SHA}^{commit}" 2>/dev/null; then
            echo "::error::the committed pin ${OBJECTUI_SHA} is not present in a fresh full clone of objectui at all — no branch carries it. It was never pushed, its branch was deleted, or main was rewritten. Fix .objectui-sha in its own PR; do not cut against it."
            exit 1
          fi
          if ! git -C "$OBJECTUI_ROOT" merge-base --is-ancestor "$OBJECTUI_SHA" origin/main; then
            echo "::error::the committed pin ${OBJECTUI_SHA} exists in objectui but is NOT reachable from objectui main — it names a revision on a branch that never merged. Cutting against it would publish @objectstack/console built from code that is not on main. Fix .objectui-sha in its own PR; do not cut against it."
            exit 1
          fi
          echo "objectui pin ${OBJECTUI_SHA:0:12}: present in the clone, and reachable from objectui main."

          # Detach the clone AT THE PIN so nothing downstream can accidentally
          # read a working tree that is objectui main. build-console.sh builds
          # from its own worktree at the pin regardless; this makes the intent
          # true of the clone itself rather than incidental.
          git -C "$OBJECTUI_ROOT" checkout --detach --quiet "$OBJECTUI_SHA"
          echo "objectui clone detached at ${OBJECTUI_SHA:0:12}."

      # ────────────────────────────────────────────────────────────────────────
      # Toolchain. Mirrors release.yml's publish job step for step.
      # ────────────────────────────────────────────────────────────────────────
      - name: Setup Node.js
        uses: actions/setup-node@v7
        with:
          # Cannot go below 22 — the downstream hotcrm smoke below clones
          # hotcrm@v1.2.0, whose manifest pins engines.node >=22.
          node-version: '22'

      - name: Setup pnpm
        uses: ./.github/actions/setup-pnpm

      - name: Get pnpm store directory
        shell: bash
        run: |
          echo "STORE_PATH=$(pnpm store path --silent)" >> $GITHUB_ENV

      - name: Setup pnpm cache
        uses: actions/cache@v6
        with:
          path: ${{ env.STORE_PATH }}
          key: ${{ runner.os }}-pnpm-store-v3-${{ hashFiles('**/pnpm-lock.yaml') }}
          restore-keys: |
            ${{ runner.os }}-pnpm-store-v3-

      # Restore-only, like ci.yml's Console Pin Gate: this workflow runs a handful
      # of times a month so its own namespace is almost always cold, and the
      # build-core fallbacks seeded from main are the ones that actually hit.
      # Saving is lint.yml's job.
      - name: Restore Turbo cache
        uses: actions/cache/restore@v6
        with:
          path: .turbo/cache
          key: ${{ runner.os }}-turbo-${{ github.job }}-${{ github.ref_name }}-${{ github.sha }}
          restore-keys: |
            ${{ runner.os }}-turbo-${{ github.job }}-${{ github.ref_name }}-
            ${{ runner.os }}-turbo-${{ github.job }}-
            ${{ runner.os }}-turbo-build-core-${{ github.ref_name }}-
            ${{ runner.os }}-turbo-build-core-

      - name: Install dependencies
        run: pnpm install --frozen-lockfile

      # A fresh workspace must build local dists BEFORE the console build: the
      # console's vite alias table (objectui#4103) resolves @objectstack/client to
      # this tree's packages/client/dist, whose DTS needs @objectstack/core built,
      # and build-console.sh's own fallback is a bare `pnpm build` inside
      # packages/client that tsup cannot finish on its own. @objectstack/spec is
      # named explicitly because generated-artifact reads need it even where the
      # client closure would already have pulled it in. Same step, same reason, as
      # ci.yml's Console Pin Gate and showcase-smoke.yml.
      - name: Build the @objectstack/client and @objectstack/spec closures
        run: pnpm exec turbo run build --filter=@objectstack/client... --filter=@objectstack/spec... --concurrency=4

      # ────────────────────────────────────────────────────────────────────────
      # ⛔ NO PIN BUMP HAPPENS HERE (#10134).
      #
      # This lane used to run `bash scripts/bump-objectui.sh "$OBJECTUI_SHA"` and
      # then assert `.objectui-sha` equalled the sha it had just written — an
      # assertion that was trivially true, and that existed only because the sha
      # came from outside the repo. Both are gone: the pin is committed input, so
      # there is nothing to bump to and nothing to reconcile.
      #
      # The `@objectstack/console` changeset for a pin range is emitted by the
      # manual bump PR, which is the only place the pin moves. A cut therefore
      # publishes a console the release record already describes, or it publishes
      # the same console as last time — never a console nobody wrote a record for.
      # ────────────────────────────────────────────────────────────────────────

      - name: Build the vendored Console SPA at the committed pin
        run: pnpm objectui:build

      - name: Verify the Console dist stamp matches the pin
        run: pnpm check:console-sha

      # ────────────────────────────────────────────────────────────────────────
      # ADR-0082 D4 declaration-parity ratchet — the mandatory SECOND HALF of
      # every pin move (#5960), run here against the pin as COMMITTED.
      #
      # ⚠️ Since #10134 this lane no longer moves the pin, so this is no longer
      # "the second half of a move this run performed" — the move, and the
      # ratchet that belongs to it, happen on the manual bump PR. Kept anyway,
      # and deliberately: the ratchet is an on-demand gate by decision (ADR-0082
      # addendum), which means "unrun" is its live failure mode, and a bump PR
      # that skipped it leaves a divergence that reaches a release unnoticed.
      # This is the last place before publish that can still say so. It writes to
      # a gitignored path and asserts against the committed baseline, so it adds
      # no file to the version commit; on a pin that already ratcheted it simply
      # passes.
      #
      # Installing a Playwright browser is fine HERE. The #5960 ruling that keeps
      # the ratchet off CI is about not putting an objectui build plus a browser
      # download on every matching PR; this workflow runs a handful of times a
      # month, on purpose, and has already built the console two steps above.
      # ────────────────────────────────────────────────────────────────────────
      - name: Install a Playwright browser for the manifest dump
        run: |
          # The dump runs INSIDE objectui's build tree (gen-sdui-manifest.sh does
          # `pushd $BUILD_ROOT` and drives objectui's own playwright), so the
          # browser must be installed against THAT workspace. `pnpm exec
          # playwright` from the framework root resolves nothing — playwright is
          # not a framework dependency.
          BUILD_ROOT=".cache/objectui-${OBJECTUI_SHA:0:12}"
          if [ ! -d "$BUILD_ROOT" ]; then
            echo "::error::expected objectui build tree at ${BUILD_ROOT} (created by 'pnpm objectui:build'). Cannot install the browser the ratchet needs."
            exit 1
          fi
          pnpm --dir "$BUILD_ROOT" exec playwright install chromium-headless-shell

      - name: 'Declaration-parity ratchet at the committed pin (ADR-0082 D4)'
        run: pnpm sdui:manifest

      # ────────────────────────────────────────────────────────────────────────
      # PRE-VERSION GATES. Every one of these reads `.changeset/*`, so all of them
      # must run BEFORE `pnpm run version` consumes it.
      #
      # ON THE `--base` VALUE: these three are FORWARD-ONLY DIFF gates — they judge
      # what a change INTRODUCES, never the standing stock (their `--list` modes
      # are audits that always exit 0, and `--audit-stock` says in its own header
      # that a non-zero exit there would make it a gate, which it is not). So the
      # honest base here is $SNAPSHOT_SHA, and the diff they judge is this run's
      # OWN contribution.
      #
      # ⚠️ Since #10134 that contribution is EMPTY by design — the lane no longer
      # bumps the pin, so it emits no @objectstack/console changeset of its own,
      # and these three assert exactly that. They are kept rather than deleted for
      # two reasons: "this cut introduces nothing before versioning" is a real
      # invariant to hold (it is what makes the allowlist assertion below a
      # statement about `pnpm run version` alone), and each still runs its
      # `--self-test`, so a broken gate is caught here rather than on the PR that
      # broke it. The risk they used to cover on this lane — an objectui `major`
      # promoting all ~70 packages, or a digest bug emitting empty frontmatter and
      # stalling the release silently and greenly (#4898) — now lands on the
      # manual bump PR, which is where the digest runs and where these same gates
      # judge a non-empty diff.
      #
      # $SNAPSHOT_SHA is a pinned sha, which #6129 warns against — but that warning
      # is about pr-automation.yml's FROZEN `base.sha`, a value that goes stale as
      # main moves under a PR. Here the snapshot is by construction the checkout
      # this run is judging, so it IS the merge base and cannot go stale.
      # ────────────────────────────────────────────────────────────────────────
      - name: Gate — Changesets "fixed" group covers every public package
        run: node scripts/check-changeset-fixed.mjs

      - name: Gate — no control bytes anywhere in the tree
        run: pnpm check:nul-bytes

      - name: Gate — this cut introduces no empty-frontmatter changeset
        run: |
          node scripts/check-empty-changeset.mjs --self-test
          node scripts/check-empty-changeset.mjs --base "$SNAPSHOT_SHA"

      - name: Gate — this cut introduces no unregistered breaking change (ADR-0087)
        run: |
          node scripts/check-adr-0087-registration.mjs --self-test
          node scripts/check-adr-0087-registration.mjs --base "$SNAPSHOT_SHA"

      - name: Gate — this cut introduces no major bump
        run: |
          node scripts/check-changeset-no-major.mjs --self-test
          node scripts/check-changeset-no-major.mjs --base "$SNAPSHOT_SHA"

      # ────────────────────────────────────────────────────────────────────────
      # VERSION. `pnpm run version`, never a bare `changeset version`: the repo
      # script is `changeset version && sync-protocol-version.mjs &&
      # sync-template-versions.mjs`, and the two sync steps are what keep
      # PROTOCOL_VERSION and the scaffolder templates in lockstep with the number
      # being cut. Pre mode (`.changeset/pre.json`, tag rc) is committed state and
      # applies automatically — the guard step already proved it is active.
      # ────────────────────────────────────────────────────────────────────────
      - name: Version the workspace
        run: pnpm run version

      - name: Assert the computed version is the one that was dispatched
        env:
          REQUESTED: ${{ inputs.version }}
        run: |
          COMPUTED="$(jq -r '.version' packages/cli/package.json)"
          if [ "$COMPUTED" != "$REQUESTED" ]; then
            echo "::error::you dispatched '${REQUESTED}' but the pending changesets compute @objectstack/cli@${COMPUTED}. Nothing has been pushed or published. Re-dispatch with ${COMPUTED} if that is the intended cut, or find the changeset that moved the number."
            exit 1
          fi
          echo "computed version ${COMPUTED} matches the dispatch."

      # ────────────────────────────────────────────────────────────────────────
      # THE SINGLE VERSION COMMIT.
      #
      # Squashed onto $SNAPSHOT_SHA so main receives exactly one commit for the
      # cut: the pin bump and the version output are one atomic unit, and there is
      # never a main commit carrying a bumped pin but an unversioned tree.
      #
      # WHAT MAY BE IN IT — re-measured under @changesets/cli v3 on the
      # 17.0.0 -> 17.1.0-rc.0 train (a full local `pnpm run version` in a throwaway
      # clone over 209 pending changesets), not assumed. 365 staged paths:
      #   76 modified package.json, 76 modified CHANGELOG.md, 209 changesets MOVED
      #   into `.changeset/pre/`, `.changeset/pre.json`, and the 3 doc surfaces
      #   below. `.changeset/pre.json` is UNTRACKED on the first cut of a train
      #   (`changeset pre enter` just created it) and modified on later ones;
      #   `git add -A` covers both.
      #   The 3 major-boundary paths below did NOT appear on this train and are
      #   allowed for the train that does move the major — as is the blank
      #   template's own package.json, which is why the v2 measurement this block
      #   used to carry counted 77 package.json against this one's 76: that train
      #   was 17.0.0 -> 18.0.0-rc.0, so `sync-template-versions.mjs` had to rewrite
      #   the template's `^17.0.0` pins. On a minor train it logs "already pins" and
      #   writes nothing. The counts move with the SHAPE of the train; the
      #   allowlist below is what does not have to.
      #
      #   THE CONSUMED CHANGESETS MOVE — new in v3 (changesets#2190), and the one
      #   thing about a cut that this file's prose used to get wrong. v2 left every
      #   consumed `.changeset/*.md` on disk and recorded it in `pre.json`; v3 moves
      #   it to `.changeset/pre/NAME.md` VERBATIM (measured: a moved file diffs
      #   identical to its original, and git reports all 209 as `R100` renames), and
      #   `pre.json` is `{"mode","tag"}` only — `changeset version` never rewrites
      #   it. So in the worktree a cut is 209 deletions plus a new untracked
      #   `.changeset/pre/` directory, and the claim this block used to make — "a
      #   cut never races a lane PR over a `.changeset/*.md` file it wants to keep"
      #   — is FALSE under v3. The cut moves tracked changeset paths, so a lane PR
      #   that edits or deletes one of them conflicts; the rebase guard further down
      #   is what catches it, and it already names `.changeset` in its conflict
      #   surface. What does NOT change is this allowlist: `^\.changeset/` covers
      #   `.changeset/pre/…` and `git add -A -- … .changeset …` stages the deletions
      #   and the new directory together (measured: all 365 staged paths accepted,
      #   nothing tracked left unstaged).
      # Three more paths can appear at a major boundary and are allowed for that
      # reason, all written by `sync-protocol-version.mjs` /
      # `sync-template-versions.mjs`: packages/spec/src/kernel/protocol-version.ts
      # and, per bundled template, objectstack.config.ts (engines.protocol) and
      # objectstack.manifest.json (specVersion). protocol-version.ts is named
      # literally because that one file IS the whole of its surface. The template
      # paths are not, and the reason is the one the doc half states below.
      #
      # THE DOC SURFACES ARE READ, NOT RESTATED. `pnpm run version` ends with
      # `sync-docs-image-tags.mjs`, which rewrites the doc surfaces that pin a
      # concrete image tag / CLI version. That rewriter does not own the list — it
      # imports `SURFACES` from `scripts/check-docs-image-tag.mjs`, the gate whose
      # findings it exists to clear, on the stated principle "one list, two
      # consumers". This step is the THIRD consumer and joins on the same terms:
      # it resolves `SURFACES` at run time rather than copying the paths, because a
      # fourth literal is a fourth contract and the two drifting apart is exactly
      # the defect that reached this assertion in the first place — the rewriter
      # joined the `version` script while this allowlist kept describing a train
      # measured before it existed, and every cut since would have refused to push.
      # Note what this deliberately is NOT: a wildcard over `content/docs/**`. Only
      # the paths a reviewed declaration names are allowed, so a version pass that
      # writes anywhere else — a new sync script, or this rewriter growing an
      # output outside `SURFACES` — still fails loudly here, which is the whole
      # point of the assertion. Onboarding a surface stays a one-line edit to
      # `SURFACES` that a reviewer sees in the diff.
      #
      # THE TEMPLATE SURFACES ARE READ ON THE SAME TERMS. They used to be two
      # literals, both hard-coding the template name `blank`, and this block used
      # to record why: `sync-template-versions.mjs` declared its targets (a
      # `TEXT_STAMPS` table plus a walk of `src/templates/`) but exported none of
      # them and ran the sync at module scope, so importing it to ask "which files
      # does the version pass stamp?" would have rewritten the templates instead of
      # answering. #9648 gave it the export and the entry-point guard this block
      # asked for, so the restatement is retired here rather than re-seeded:
      # `stampedPaths()` derives its answer from the same walk and the same table
      # the stamper's own `main()` uses.
      #
      # This is a COLLAPSE OF TWO COPIES OF ONE LIST, not a widening. The literals
      # were never a deliberately narrower allowlist — they were the same set,
      # spelled by hand, and they equal it only while `blank` is the only bundled
      # template. `findTemplateDirs()` exists BECAUSE the template set is not
      # curated ("a template added tomorrow is covered on the day it lands"), so
      # the day a second template ships, the walk stamps it, the literal pair does
      # not cover it, and the unstaged-files assertion below refuses the cut — with
      # nothing red until someone attempts a release. `stampedPaths()` also reports
      # each template's own package.json, which `'*package.json'` already permits,
      # so the resolved list is a superset of the two literals and never a
      # narrowing.
      #
      # The allowlist is ENFORCED rather than described. Only allowlisted paths are
      # staged, and then every staged path is re-checked against the same rule and
      # the worktree is asserted clean. If `pnpm run version` ever grows a new
      # output, this fails loudly here instead of pushing a surprise to main — the
      # "disjoint from lane PRs by construction" claim is only worth anything if
      # something actually checks it.
      # ────────────────────────────────────────────────────────────────────────
      - name: Build the single version commit
        id: commit
        env:
          REQUESTED: ${{ inputs.version }}
          DRY_RUN: ${{ inputs.dry_run }}
        run: |
          # The committer identity. It used to be set by the pin-bump step, which
          # #10134 deleted; `actions/checkout` does not set one, so `git commit`
          # below would fail with "empty ident name" without this.
          git config user.name 'github-actions[bot]'
          git config user.email '41898282+github-actions[bot]@users.noreply.github.com'

          # Structural guarantee that the version commit's parent is the snapshot.
          # Since #10134 no step before this one commits anything, so this is a
          # no-op in the normal case — kept because it is what makes "one commit
          # on top of the snapshot" true by construction rather than by review.
          git reset --soft "$SNAPSHOT_SHA"

          # The doc surfaces `sync-docs-image-tags.mjs` rewrites, read from the same
          # declaration the rewriter reads. Import-safe by construction: the module
          # carries an entry-point guard added for exactly this kind of consumer.
          SURFACE_LIST="${RUNNER_TEMP:-/tmp}/cut-rc-docs-image-tag-surfaces.txt"
          if ! node --input-type=module \
                 -e 'import { SURFACES } from "./scripts/check-docs-image-tag.mjs"; for (const s of SURFACES) console.log(s.file);' \
                 > "$SURFACE_LIST"; then
            echo "::error::could not resolve SURFACES from scripts/check-docs-image-tag.mjs, so the doc half of the release file surface is unknown. Refusing to push."
            exit 1
          fi
          # An empty list would silently re-open the hole this replaced: the docs the
          # version pass rewrites would go unstaged and the assertion below would
          # blame them. Unknown is a failure, never an empty allowlist.
          if [ ! -s "$SURFACE_LIST" ]; then
            echo "::error::SURFACES in scripts/check-docs-image-tag.mjs resolved EMPTY, so no doc surface would be staged even though the version pass rewrites them. Refusing to push."
            exit 1
          fi
          mapfile -t DOCS_SURFACES < "$SURFACE_LIST"
          echo "doc surfaces declared by SURFACES (${#DOCS_SURFACES[@]}):"
          sed 's/^/  /' "$SURFACE_LIST"

          # The template surfaces `sync-template-versions.mjs` stamps, read from the
          # same walk and the same table the stamper itself uses. Import-safe by
          # construction: that module carries the entry-point guard and the exports
          # added for exactly this consumer.
          TEMPLATE_LIST="${RUNNER_TEMP:-/tmp}/cut-rc-template-version-surfaces.txt"
          if ! node --input-type=module \
                 -e 'import { stampedPaths } from "./scripts/sync-template-versions.mjs"; for (const p of stampedPaths()) console.log(p);' \
                 > "$TEMPLATE_LIST"; then
            echo "::error::could not resolve stampedPaths() from scripts/sync-template-versions.mjs, so the template half of the release file surface is unknown. Refusing to push."
            exit 1
          fi
          # Same rule as the doc half: unknown is a failure, never an empty
          # allowlist. `stampedPaths()` THROWS rather than returning [] on a moved
          # or empty template directory, so an empty file here means a resolution
          # that reported nothing while still exiting 0 — which would stage no
          # template path at all and then blame the files it left unstaged.
          if [ ! -s "$TEMPLATE_LIST" ]; then
            echo "::error::stampedPaths() in scripts/sync-template-versions.mjs resolved EMPTY, so no template surface would be staged even though the version pass stamps them. Refusing to push."
            exit 1
          fi
          mapfile -t TEMPLATE_SURFACES < "$TEMPLATE_LIST"
          echo "template surfaces declared by stampedPaths() (${#TEMPLATE_SURFACES[@]}):"
          sed 's/^/  /' "$TEMPLATE_LIST"

          git add -A -- \
            '*package.json' \
            '*CHANGELOG.md' \
            .changeset \
            .objectui-sha \
            packages/spec/src/kernel/protocol-version.ts \
            "${TEMPLATE_SURFACES[@]}" \
            "${DOCS_SURFACES[@]}"

          STAGED="$(git diff --cached --name-only)"
          if [ -z "$STAGED" ]; then
            echo "::error::the version pass produced no changes to commit. Nothing to cut."
            exit 1
          fi

          # Re-check every staged path against the allowlist. The pathspec above is
          # convenience; THIS is the guarantee. Three filters, same allowlist the
          # pathspec used: the fixed release paths by pattern, then the declared
          # template and doc surfaces by WHOLE-LINE EXACT match (`-xF`) against the
          # very lists that were staged — so neither derived filter can accept a path
          # its declaration does not name, and neither needs regex-escaping of the
          # paths to stay exact.
          BAD="$(printf '%s\n' "$STAGED" \
            | grep -vE '(^|/)package\.json$|(^|/)CHANGELOG\.md$|^\.changeset/|^\.objectui-sha$|^packages/spec/src/kernel/protocol-version\.ts$' \
            | grep -vxF -f "$TEMPLATE_LIST" \
            | grep -vxF -f "$SURFACE_LIST" || true)"
          if [ -n "$BAD" ]; then
            echo "::error::the version commit would carry paths outside the release file surface. Refusing to push. Offending paths follow; if the version pass legitimately grew a new output, widen the allowlist in this workflow deliberately."
            printf '%s\n' "$BAD" | sed 's/^/::error::  unexpected: /'
            exit 1
          fi

          # Nothing tracked may be left behind: a modified-but-unstaged file means
          # the version pass wrote somewhere the allowlist does not cover, and the
          # commit would silently ship a partial cut.
          LEFT="$(git diff --name-only)"
          if [ -n "$LEFT" ]; then
            echo "::error::tracked files were modified but fall outside the release file surface, so the version commit would be incomplete. Refusing to push."
            printf '%s\n' "$LEFT" | sed 's/^/::error::  unstaged: /'
            exit 1
          fi

          # Separate -m flags rather than one multi-line string: inside a YAML
          # block scalar every continuation line carries the block's indentation,
          # which would land verbatim in the commit body.
          git commit -q \
            -m "chore(release): ${REQUESTED}" \
            -m "objectui pin (unchanged by this cut): ${OBJECTUI_SHA}" \
            -m "cut from ${SNAPSHOT_SHA} by ${GITHUB_ACTOR} via cut-rc (#7447)"

          COMMIT_SHA="$(git rev-parse HEAD)"
          COUNT="$(printf '%s\n' "$STAGED" | wc -l | tr -d ' ')"
          echo "commit-sha=${COMMIT_SHA}" >> "$GITHUB_OUTPUT"
          echo "file-count=${COUNT}" >> "$GITHUB_OUTPUT"
          echo "built ${COMMIT_SHA} on top of ${SNAPSHOT_SHA}: ${COUNT} file(s)."
          {
            echo "## Cut ${REQUESTED}"
            echo
            echo "| | |"
            echo "|---|---|"
            echo "| objectstack snapshot | \`${SNAPSHOT_SHA}\` |"
            echo "| objectui pin (committed) | \`${OBJECTUI_SHA}\` |"
            echo "| version commit | \`${COMMIT_SHA}\` |"
            echo "| files in commit | ${COUNT} |"
            echo "| dry run | ${DRY_RUN} |"
          } >> "$GITHUB_STEP_SUMMARY"

      # ────────────────────────────────────────────────────────────────────────
      # DRY RUN STOPS HERE.
      #
      # It stops AFTER the local commit rather than after the version step, which
      # is strictly more of the pipeline than #7447's acceptance criterion asks
      # for: the squash, the allowlist enforcement and the version assertion have
      # all really executed, and the artifact is the exact commit a real run would
      # push. Everything downstream of this point is irreversible.
      # ────────────────────────────────────────────────────────────────────────
      - name: Export the would-be version commit (dry run)
        if: ${{ inputs.dry_run }}
        env:
          REQUESTED: ${{ inputs.version }}
        run: |
          mkdir -p "${RUNNER_TEMP}/cut-rc"
          git format-patch -1 --stdout > "${RUNNER_TEMP}/cut-rc/version-commit.patch"
          git show --stat --oneline HEAD > "${RUNNER_TEMP}/cut-rc/version-commit.stat.txt"
          {
            echo "# cut-rc dry run — ${REQUESTED}"
            echo
            echo "- objectstack snapshot: ${SNAPSHOT_SHA}"
            echo "- objectui pin:         ${OBJECTUI_SHA} (committed, unchanged)"
            echo "- version commit:       $(git rev-parse HEAD)"
            echo
            echo "Nothing was pushed and nothing was published."
            echo "version-commit.patch is byte-for-byte what a real dispatch would push to main."
          } > "${RUNNER_TEMP}/cut-rc/README.md"

      - name: Upload the dry-run artifact
        if: ${{ inputs.dry_run }}
        uses: actions/upload-artifact@v7
        with:
          name: cut-rc-${{ inputs.version }}-dry-run
          path: ${{ runner.temp }}/cut-rc
          if-no-files-found: error

      - name: Stop (dry run)
        if: ${{ inputs.dry_run }}
        run: |
          echo "::notice::dry run complete — the version commit was built and verified locally, then discarded. Nothing was pushed, nothing was published. Re-dispatch with dry_run unchecked to cut for real."
          echo "- **dry run**: stopped before the push. Nothing was published." >> "$GITHUB_STEP_SUMMARY"

      # ────────────────────────────────────────────────────────────────────────
      # LAND ON MAIN FIRST, PUBLISH SECOND (#6170).
      #
      # rc.3 and rc.4 tagged commits that lived only on `changeset-release/main`,
      # so main kept a stale version and every later release recomputed an
      # npm-occupied number. The ordering below is the structural fix: by the time
      # anything reaches npm, the exact commit it was built from is on main.
      #
      # THE DRY-RUN GUARD IS WRITTEN TWICE, ON PURPOSE. The step-level `if:` is the
      # primary control; the in-shell assert is there because a dropped `if:` in a
      # future edit would otherwise publish from a run a human asked NOT to
      # publish. Two languages, one decision — the same belt-and-braces reasoning
      # release.yml applies to its own publish path.
      # ────────────────────────────────────────────────────────────────────────
      - name: Push the version commit to main
        if: ${{ !inputs.dry_run }}
        env:
          DRY_RUN: ${{ inputs.dry_run }}
          REQUESTED: ${{ inputs.version }}
        run: |
          if [ "$DRY_RUN" = "true" ]; then
            echo "::error::refusing to push: this run was dispatched as a dry run. A step-level 'if:' has been dropped in an edit to this workflow."
            exit 1
          fi

          # Bounded fetch-rebase-retry. main is busy by assumption — that is the
          # premise of this whole workflow — so losing the push race is expected,
          # not exceptional.
          ATTEMPTS=5
          for attempt in $(seq 1 "$ATTEMPTS"); do
            if git push origin HEAD:main; then
              echo "pushed $(git rev-parse HEAD) to main on attempt ${attempt}."
              echo "PUSHED_SHA=$(git rev-parse HEAD)" >> "$GITHUB_ENV"
              exit 0
            fi

            if [ "$attempt" -eq "$ATTEMPTS" ]; then
              echo "::error::could not push the version commit to main after ${ATTEMPTS} attempts. Nothing has been published, so this is safe to retry."
              echo "::error::If the rejection was 'protected branch' / 'refusing to allow ... to update', this is the ONE-TIME ADMIN PREREQUISITE and no re-run will clear it: the pushing identity must be on main's ruleset BYPASS list. Either add the GitHub Actions app to the ruleset bypass list (ideally scoped to this workflow), or create a fine-grained PAT with contents:write for an account already on that list and store it as the repository secret RELEASE_PUSH_TOKEN. See this workflow's checkout step for the full note."
              exit 1
            fi

            echo "push rejected (attempt ${attempt}/${ATTEMPTS}) — refreshing main and rebasing the version commit."
            git fetch origin main
            if ! git rebase origin/main; then
              git rebase --abort || true
              echo "::error::the version commit conflicts with main. Nothing has been published. A commit landing during the cut touched the release file surface (package.json / CHANGELOG.md / .changeset / .objectui-sha) — resolve it and re-dispatch."
              exit 1
            fi

            # The rebase moved the commit onto newer main, so the version
            # assertion must be re-earned rather than inherited. If another
            # version commit landed underneath us the number is now wrong, and
            # that must stop the cut before npm, not after.
            COMPUTED="$(jq -r '.version' packages/cli/package.json)"
            if [ "$COMPUTED" != "$REQUESTED" ]; then
              echo "::error::after rebasing onto refreshed main the tree carries @objectstack/cli@${COMPUTED}, not the dispatched ${REQUESTED}. Something else versioned main during this cut. Nothing has been published."
              exit 1
            fi

            sleep $(( attempt * 5 ))
          done

      # From here the commit IS on main, so everything published is a commit main
      # already carries — #6170's invariant, kept to the letter. If the loop above
      # rebased, main also carries whatever landed during the cut; those commits
      # passed their own CI, and the publish below builds from this exact tree, so
      # the artifacts and the main commit still describe each other exactly.
      - name: Confirm the pushed commit is main
        if: ${{ !inputs.dry_run }}
        run: |
          git fetch origin main
          REMOTE="$(git rev-parse origin/main)"
          LOCAL="$(git rev-parse HEAD)"
          if [ "$REMOTE" != "$LOCAL" ]; then
            echo "::warning::main has already moved past the version commit (${LOCAL} -> ${REMOTE}). That is normal on a busy repository; the version commit itself is on main, which is what the publish requires."
            if ! git merge-base --is-ancestor "$LOCAL" "$REMOTE"; then
              echo "::error::the version commit ${LOCAL} is NOT an ancestor of origin/main. Refusing to publish a commit main does not carry (#6170)."
              exit 1
            fi
          fi
          echo "@objectstack/cli@$(jq -r '.version' packages/cli/package.json) is on main at ${LOCAL}."

      # Pre-publish gate #2035, kept in step with release.yml's publish job rather
      # than dropped for speed. In pre mode (which the guard step proved is active,
      # so ALWAYS on this lane) the #3600 amendment makes it advisory: a major
      # train exists to ship deliberate surface removals, and a migrated hotcrm
      # cannot exist until the rc artifacts it would migrate against are published.
      # It reports; it cannot block.
      - name: Downstream backward-compat smoke (live hotcrm, advisory in pre mode)
        if: ${{ !inputs.dry_run }}
        env:
          HOTCRM_REF: v2.1.0
        run: |
          if bash scripts/downstream-smoke.sh; then
            echo "::notice::hotcrm@${HOTCRM_REF} is still compatible with the pre-release train."
          else
            echo "::warning::hotcrm@${HOTCRM_REF} is incompatible with the pre-release train — expected for this window's deliberate removals. Ship a migrated hotcrm release and bump HOTCRM_REF in release.yml before 'changeset pre exit' re-arms that gate."
          fi

      # ────────────────────────────────────────────────────────────────────────
      # PUBLISH. `pnpm run release` = build + build-console + release-publish.sh,
      # which is `changeset publish` followed by ONE atomic `git push origin --tags`
      # (#2191: the changesets action's concurrent per-tag pushes raced GitHub's
      # ref backend and lost ~half the tags).
      #
      # changesets/action is NOT used, exactly as in release.yml: handed a
      # workspace with pending changesets it would take the VERSION path and mint a
      # commit. `changeset publish` can only publish what the checked-out
      # package.json files already declare — which is what main now carries.
      # ────────────────────────────────────────────────────────────────────────
      - name: Publish to npm + push version tags
        id: publish
        if: ${{ !inputs.dry_run }}
        env:
          NPM_TOKEN: ${{ secrets.NPM_TOKEN }}
          DRY_RUN: ${{ inputs.dry_run }}
          VERSION: ${{ inputs.version }}
        run: |
          if [ "$DRY_RUN" = "true" ]; then
            echo "::error::refusing to publish: this run was dispatched as a dry run. A step-level 'if:' has been dropped in an edit to this workflow."
            exit 1
          fi

          printf '//registry.npmjs.org/:_authToken=%s\n' "$NPM_TOKEN" >> "$HOME/.npmrc"

          pnpm run release

          # `changeset publish` skips versions already on the registry, so a
          # re-dispatch over a partially-published release is a repair rather than
          # a duplicate. What is NOT optional is that the version is on npm when
          # this step ends.
          if ! npm view "@objectstack/cli@$VERSION" version >/dev/null 2>&1; then
            echo "::error::publish ran but @objectstack/cli@$VERSION is still not on npm"
            exit 1
          fi
          echo "published=true" >> "$GITHUB_OUTPUT"

      - name: Create GitHub Releases (bodies truncated to the API limit)
        # `!cancelled()` rather than the implicit success(): npm is already public
        # by the time this runs, so a failure upstream must not be the reason the
        # release record stays empty (#4900).
        if: ${{ !cancelled() && steps.publish.outputs.published == 'true' }}
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          # The fixed group releases every public package at one version, so the
          # version alone drives the whole publishable workspace.
          RELEASE_VERSION: ${{ inputs.version }}
        run: node scripts/release-github-releases.mjs

      - name: Attach spec-changes.json to the GitHub Release (ADR-0087 D4)
        # Ordering is load-bearing: `gh release upload` needs the Release the step
        # above created.
        if: ${{ !cancelled() && steps.publish.outputs.published == 'true' }}
        env:
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          RELEASE_VERSION: ${{ inputs.version }}
        run: bash scripts/release-spec-changes.sh

      - name: Summarise the cut
        if: ${{ !cancelled() && steps.publish.outputs.published == 'true' }}
        env:
          REQUESTED: ${{ inputs.version }}
        run: |
          {
            echo
            echo "**${REQUESTED} is published.**"
            echo
            echo "- npm: every package in the fixed group, dist-tag \`rc\`"
            echo "- git: version tags pushed atomically"
            echo "- main: carries the version commit (nothing was published that main does not have)"
            echo
            echo "### Two things that are expected, not defects"
            echo
            echo "1. **The standing \`chore: version packages (rc)\` PR (#6208) may look stale.**"
            echo "   Its changesets were consumed by this cut and moved to \`.changeset/pre/\` (v3)."
            echo "   Whether it refreshes now depends on the push credential: pushes made with the"
            echo "   Actions \`GITHUB_TOKEN\` trigger no workflow runs (GitHub's recursion guard), so"
            echo "   \`version-pr\` does not fire until some later push to main; with a"
            echo "   \`RELEASE_PUSH_TOKEN\` PAT it fires now and the PR self-corrects. **Do not fix it"
            echo "   by hand** — see docs/releases-maintenance.md, 'What happens to the standing"
            echo "   Version Packages PR'."
            echo "2. **No runtime image yet.** This lane does not build one; \`release-integrity\` in"
            echo "   \`release.yml\` requests it on a later push to main, or dispatch"
            echo "   \`docker-publish.yml\` with version \`${REQUESTED}\` for one immediately."
          } >> "$GITHUB_STEP_SUMMARY"