release.yml

name: Release

# ══════════════════════════════════════════════════════════════════════════════
# TWO LANES, ONE INVARIANT: ONLY A HUMAN PUBLISHES. (#6170)
# ══════════════════════════════════════════════════════════════════════════════
#
# Maintainer ruling, 2026-08-07 (verbatim, do not translate):
#
#   「刚才我也没提出要求,是哪个ai自己替我发了 rc.4,版本发布必须是人工的。
#     这个要写入规范。」
#
# WHAT THIS FILE USED TO DO, AND WHY IT MINTED TWO RELEASES NOBODY ASKED FOR
# --------------------------------------------------------------------------
# One job, triggered `on: push: branches: [main]`, so EVERY merge-queue landing
# started it. Inside it, two steps in sequence:
#
#   1. changesets/action@v1 with a `publish:` script. With pending changesets it
#      takes the version-PR path, which is (its own source, v1):
#           git checkout -b changeset-release/main
#           git reset --hard <github.context.sha>
#           pnpm run version                     # ← bumps every package.json
#           git add . && git commit -m 'chore: version packages'
#           git push origin HEAD:changeset-release/main --force
#      It never restores the workspace. The job therefore continues on the
#      FRESHLY VERSIONED tree, not on main's state.
#   2. "Ensure this version actually shipped" (`recover-publish`) read
#      `packages/cli/package.json` FROM THAT WORKSPACE. It documented itself as
#      "a no-op on the normal path, where main's version IS the last released
#      one" — but after step 1 the workspace carries the NEXT version, which is
#      ALWAYS absent from npm. So its repair branch fired and ran the real
#      publish: 69 packages to npm + an atomic tag push at a commit that only
#      ever existed on `changeset-release/main`.
#
# Twice, platform-stamped, with no human anywhere in the trigger chain:
#   17.0.0-rc.3 — 2026-08-03, version commit c6a52d3  (cleanup #6135 → #6149)
#   17.0.0-rc.4 — 2026-08-07, version commit a10cbc77 (cleanup #6169)
# Run 31146224227 is the rc.4 receipt: event `push`, actor
# `github-merge-queue[bot]`. The 4 quiet days in between are the same mechanism
# reporting green — the computed next version happened to already be on npm.
#
# HOW THE LANES ARE SPLIT NOW
# ---------------------------
#   schedule (6-hourly)  →  `version-pr`  keeps the "chore: version packages" PR
#   or a dispatch with                     (#4935) current. Carries NO publish
#   `refresh_version_pr`                   capability: the changesets step is
#                                          invoked WITHOUT a `publish-script:`,
#                                          so the action's publish branch is
#                                          unreachable by construction, not by
#                                          an `if:` someone can get wrong.
#                                          NOT on push — see the section below.
#   push to main  →  `release-integrity`   audits ONLY the version at
#                                          `github.sha`, and names the VERSION
#                                          COMMIT — the landing that brought
#                                          main to that version; when it queues
#                                          the publish, it also warns of every
#                                          changeset that commit carries but did
#                                          not consume. Never publishes,
#                                          never pushes a tag. May backfill GitHub
#                                          Releases / the ADR-0087 D4 asset /
#                                          the runtime image — but only for a
#                                          version ALREADY fully on npm, which
#                                          is repair that cannot mint anything.
#                 →  `publish`            the ONLY job that runs
#                                          `changeset publish` or pushes a
#                                          version tag. It starts only on the
#                                          push that CARRIES the version commit,
#                                          while that version is absent from
#                                          npm (ADR-0125 D1 as amended
#                                          2026-09-29), is then held, whole, at
#                                          `environment: release` until a
#                                          required reviewer approves it, and
#                                          checks out and publishes THE VERSION
#                                          COMMIT — never a later push's head.
#                 →  `stale-prompts`      cancels a push-lane publish prompt
#                                          still waiting for a version that is
#                                          already on npm. Mints nothing; never
#                                          touches a job that has started.
#                 →  `full-battery`       (#22998) reads the newest full-battery
#                                          runs of CI and Lint & Type Check at
#                                          or before the version commit;
#                                          `publish` needs it, so a red battery
#                                          queues no approval prompt at all.
#   workflow_dispatch → `publish`          the repair lane. Takes no version;
#   (WITHOUT                                audits main the same way, minus the
#    `refresh_version_pr`                   push range (a dispatch is one human
#    or `snapshot_next`)                    act, not one per landing), and
#                                          publishes the same version commit.
#                                          Same environment gate.
#   workflow_dispatch → `snapshot-next`    the `next` snapshot lane (ADR-0125,
#   WITH `snapshot_next`                   amendment of 2026-11-05). Runs NONE
#                                          of the stable jobs above. Cuts the
#                                          newest commit whose full battery is
#                                          green ON it, versions it in the
#                                          workspace only, and publishes under
#                                          `next` alone after the same
#                                          environment approval. No tag, no
#                                          Release, never `latest`.
#
# WHY `version-pr` LEFT THE PUSH TRIGGER (#11233, 2026-08-23)
# ----------------------------------------------------------
# changesets/action's version path is `git reset --hard <github.context.sha>` →
# re-version → `git push --force origin HEAD:changeset-release/main` (its source
# is quoted at the top of this file). On push that recomputes and force-pushes
# #4935 on EVERY landing, and main takes ~18 merges a working day. The standing
# Version Packages PR therefore never held still long enough for its own branch
# CI to finish: every run was superseded by the next force-push, so the PR could
# not converge and was ejected from the merge queue on entry. That is a
# structural property of (action algorithm × trigger frequency), not of the CLI
# version — this repo is already on @changesets/cli ^3.0.0 and the churn was
# unchanged. The fix is the one the changesets project documents for exactly
# this: refresh on a SCHEDULE instead of on every push.
#
# Between refreshes `changeset-release/main` is a static branch. Its CI
# converges, and it merges through the ordinary queue like any other PR. A stale
# window of up to six hours is the whole cost, and it is bounded on demand:
# dispatch with `refresh_version_pr` when you want it current NOW (immediately
# before a GA cut, say). The bookkeeping is not time-critical — the changesets
# are already committed on main; #4935 is only their rendering.
#
# ⛔ THE DISPATCH COLLISION, AND WHY THERE IS AN INPUT FOR IT
# ----------------------------------------------------------
# `workflow_dispatch` was already taken: it is the publish REPAIR lane (D4).
# One event name now has to start two lanes that must never start each other —
# a refresh that also queued the publish audit would park a WAITING DEPLOYMENT
# at the `release` environment on every routine refresh, i.e. an approval prompt
# a maintainer must open and dismiss to keep the real ones meaningful. Approval
# noise is how an approval stops being read, and this file's whole barrier is
# that approval (see below).
#
# So the event is split by an INPUT rather than by a second workflow file (the
# maintainer does not want another lane to maintain, and a second file would
# duplicate the publish invariants where they can drift apart):
#
#     dispatch WITH    `refresh_version_pr` → version-pr only, no deployment
#     dispatch WITH    `snapshot_next`      → the snapshot lane only (end of file)
#     dispatch WITHOUT either               → the repair lane, exactly as before
#     schedule                              → version-pr only
#     push to main                          → release-integrity (+ publish)
#
# Every job carries the half of that split it needs, in its own `if:`. No job
# infers its lane from another job's presence.
#
# ⚠️ The two inputs are INDEPENDENT, so `refresh_version_pr` + `force` is a
# reachable form, and it is refused rather than resolved: `publish`'s guard
# excludes any dispatch carrying `refresh_version_pr`, so that combination
# refreshes and publishes NOTHING. A dispatch that both refreshes bookkeeping
# and force-publishes is not a thing anyone means; the harmless reading is the
# one that runs.
#
# WHERE THE HUMAN IS, AFTER ADR-0125 (2026-08-20)
# -----------------------------------------------
# This file used to make the human confirmation a TYPED VERSION on a
# `workflow_dispatch` form, and its own comment called the dispatch event the
# guarantee: no push, no queue landing, no bot token, no schedule can synthesise
# it. That property is gone on purpose. The maintainer's ruling of 2026-08-20 is
# that merging the Version Packages PR is already the decision to release, and
# retyping the version afterwards confirms a decision they have just taken. So
# the two human acts are now:
#
#     1. merge the `chore: version packages` PR   ← the decision
#     2. approve the `release` environment        ← the authorisation
#
# The 2026-08-07 ruling 「版本发布必须是人工的」 is UNCHANGED and still binding.
# What changed is which act carries it.
#
# ⛔ THE GATE IS NOW A REPO SETTING, AND NOTHING HERE CAN CHECK IT.
# `environment: release` only creates the deployment gate. An environment with
# no protection rules passes AUTOMATICALLY and silently, and in the run log an
# unprotected gate is indistinguishable from an approved one. While the trigger
# was `workflow_dispatch` that was a weakness; now that the trigger is a push it
# is THE barrier — remove the reviewers and this file publishes 69 packages on
# every version-PR merge with nobody deciding, which is rc.3 / rc.4 exactly.
#     Settings → Environments → release → Required reviewers
# Confirmed configured by the maintainer on 2026-08-20. ⚠️ Verified by a human
# opening that page — not by this YAML, not by a CI gate, not by ADR-0125. If
# the reviewers are ever removed, revert `publish` to a `workflow_dispatch`
# trigger in the SAME change rather than leaving this running.
#
# WHAT IS DELIBERATELY STILL AUTOMATIC
# ------------------------------------
# Version-PR maintenance (this file's `version-pr` job) still runs unattended —
# harmless bookkeeping, and #4935 must keep regenerating — it just runs on a
# 6-hourly schedule instead of on every push (#11233). A `schedule` trigger
# reaching this job is not a loosening: the job has no publish capability by
# construction, so the event that starts it cannot change what it is able to do.
# Release/D4/image backfill for an already-published version stays on push runs —
# it is the #4900 repair, and it cannot mint a version. A stable `npm publish`
# and `git push --tags` still live in exactly one job, that job is reachable
# from `push` and from the repair dispatch ONLY — never from `schedule`, never
# from a refresh or a snapshot dispatch — and it cannot start without a human
# approving it. The one other publish, the `next` snapshot, lives in
# `snapshot-next`: dispatch-only, held at the same environment, no tag at all.

on:
  push:
    branches:
      - main
  # Version-PR bookkeeping only (#11233). GitHub runs `schedule` on the DEFAULT
  # BRANCH exclusively, which is the ref `version-pr` may ever regenerate from,
  # so the trigger cannot reach a ref the job is not meant to touch. Six-hourly
  # is the stale-window budget: #4935 renders changesets that are already
  # committed on main, so lateness costs nothing that cannot be bought back on
  # demand with `refresh_version_pr` below.
  #
  # ⚠️ Scheduled runs are queued, not guaranteed on the minute — GitHub delays
  # or drops them under load, and disables them entirely after 60 days of
  # repository inactivity. Both are acceptable HERE and would not be on a
  # publishing lane: a refresh that arrives late leaves #4935 stale, which is
  # visible on the PR and fixable by one dispatch. This is a second reason the
  # publish lane must never be reachable from `schedule`.
  schedule:
    - cron: '0 */6 * * *'
  # The repair lane (ADR-0125 D4). Takes NO version: both lanes audit main the
  # same way. `force` exists for the one case the push lane's predicate cannot
  # see — a publish that died having already shipped the @objectstack/cli canary
  # but not every package in the fixed group. It is dispatch-only by
  # construction, so no push can set it, and it widens what may be ATTEMPTED,
  # never what may be published unattended: the `release` environment approval
  # below applies to this lane identically.
  workflow_dispatch:
    inputs:
      force:
        description: >-
          Publish even when @objectstack/cli is already on npm. Only for
          finishing a partial publish — changeset publish skips versions the
          registry already has, so this is a repair, never a duplicate.
        required: false
        default: false
        type: boolean
      # The on-demand half of #11233's schedule. Its ONLY effect is to move this
      # dispatch onto the bookkeeping lane: `version-pr` requires it, and both
      # `release-integrity` and `publish` refuse a dispatch that carries it. It
      # therefore cannot widen anything — it is strictly subtractive, the one
      # input in this file that can only ever cause LESS to run.
      refresh_version_pr:
        description: >-
          Regenerate the "chore: version packages" PR (#4935) now instead of
          waiting for the next 6-hourly refresh. Runs the bookkeeping lane ONLY:
          no audit, no publish, no deployment queued at the `release`
          environment. Leave unchecked to use the publish repair lane.
        required: false
        default: false
        type: boolean
      # The `next` snapshot lane (ADR-0125, amendment of 2026-11-05; ruling
      # record 6663894403, #29466 letter B). Its effect is to move this dispatch
      # onto the snapshot lane: `snapshot-next-preflight` and `snapshot-next`
      # require it, and `release-integrity`, `full-battery` and `publish` refuse
      # a dispatch that carries it, exactly as they refuse `refresh_version_pr`.
      # Toward the stable lane it is strictly subtractive. The snapshot lane
      # never reads `force`; beside `refresh_version_pr` it runs nothing (the
      # harmless reading, as for `refresh_version_pr` + `force` above).
      snapshot_next:
        description: >-
          Cut a `next` snapshot of main: the newest commit whose full battery is
          green ON it, versioned in the workspace only (never committed) and
          published under the `next` dist-tag alone after a maintainer approves
          the `release` environment. No stable job runs, no tag is pushed, no
          GitHub Release is cut and `latest` never moves.
        required: false
        default: false
        type: boolean

# ⛔ NO workflow-level concurrency — per-JOB groups below, deliberately
# (ADR-0125 D5).
#
# GitHub keeps at most ONE pending run per group: when a second run queues
# behind a running one, the older PENDING run is cancelled. The group this file
# used to carry was keyed on `github.event_name`, and the comment on it said the
# point was that "the two lanes can never displace each other". That key stopped
# separating anything the moment the publish lane moved onto `push` (D1) — both
# lanes are now the same event.
#
# Worse than not separating: a job waiting on the `release` environment approval
# holds its run IN PROGRESS for as long as the maintainer takes. Under one
# shared group every other run in that window would queue as pending and evict
# the one before it, so an hour spent deciding would silently stop the Version
# Packages PR from regenerating. Per-job groups keep the waiting publish from
# touching the bookkeeping lane at all. #11233 sharpened this rather than
# retiring it: the runs that would be evicted are now the 6-hourly refreshes and
# any on-demand one, and a refresh is exactly what someone reaches for while a
# release is mid-approval.

jobs:
  # ══════════════════════════════════════════════════════════════════════════
  # PUSH LANE 1 — version-PR bookkeeping. Structurally cannot publish.
  # ══════════════════════════════════════════════════════════════════════════
  version-pr:
    name: Version PR maintenance
    # ⛔ NOT `push` (#11233). On push this job force-pushed #4935 on every one of
    # main's ~18 daily landings, so the PR's own CI could never converge and the
    # PR could never merge. The scheduled tick is the refresh; the dispatch input
    # is the same refresh on demand.
    #
    # `inputs.refresh_version_pr` is guarded by the event test rather than read
    # bare: the `inputs` context exists only on `workflow_dispatch`, so on a
    # `schedule` run it is null — and `null` is falsy, which would be the right
    # answer by accident. Say which event we are on, so the guard states the lane
    # split instead of leaning on a context's emptiness.
    if: >-
      github.event_name == 'schedule' ||
      (github.event_name == 'workflow_dispatch' && inputs.refresh_version_pr)
    runs-on: ubuntu-latest
    # Run wall of the 23 schedule and dispatch runs among the 400 most recent
    # (2026-10-22..27), the only events this job runs on: p99 2.4 min, max 2.4. The bound
    # leaves room for a cold pnpm store on the install step.
    timeout-minutes: 20
    # Serialise against itself so two refreshes cannot race the force-push to
    # `changeset-release/main`; never cancel in progress. The races it covers
    # changed with the trigger (#11233) but did not go away: a scheduled tick can
    # still overlap the previous one if a refresh runs long, and an on-demand
    # dispatch is most likely to be fired precisely when someone is impatient
    # with a tick already in flight. An evicted PENDING run is harmless here —
    # this job regenerates the PR from scratch, so the newest run's result is the
    # one that was wanted anyway.
    concurrency:
      group: release-version-pr-${{ github.ref }}
      cancel-in-progress: false
    permissions:
      contents: write
      pull-requests: write
    steps:
      - name: Checkout repository
        uses: actions/checkout@v7

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

      # Decides WHAT the version pass will bump, so it belongs on this lane too
      # (the publish lane runs it again — a gate on one lane is not a gate on
      # the other).
      - name: Verify Changesets "fixed" group covers every public package
        run: pnpm check:changeset-fixed

      # ══════════════════════════════════════════════════════════════════════
      # POST-VERSION VALIDATION — THE ONLY GATE THE RELEASE COMMIT EVER PASSES
      # (#11945)
      # ══════════════════════════════════════════════════════════════════════
      # The "chore: version packages" PR runs NO CI and cannot be made to.
      # GitHub does not start workflow runs from events raised by the default
      # `GITHUB_TOKEN`, and the step below opens and force-pushes that PR with
      # exactly that token — so it gets no `pull_request` build, and because the
      # branch is force-pushed there is not even a stable head for a human to
      # re-run checks against. Two files in this repository already state the
      # mechanism in their own words rather than inheriting it from a sibling:
      #
      #   lint.yml, on check:docs-image-tag — "`changeset version` bumps
      #   packages/cli on a release PR that (per sync-template-versions.mjs's
      #   own header) gets NO CI because changesets/action opens it with the
      #   default GITHUB_TOKEN — so the bump merged green and the gate reddened
      #   on the NEXT ordinary PR, naming files that author never touched."
      #
      #   scripts/sync-protocol-version.mjs — "the lockstep guard
      #   (protocol-version.test.ts) exists, but release PRs opened by
      #   changesets/action with the default GITHUB_TOKEN do not trigger CI
      #   (GitHub's anti-recursion rule), so the guard only fired AFTER the
      #   merge. Fixing the value at version time is the only spot that cannot
      #   be skipped."
      #
      # That last sentence is the treadmill. Every version-time output has so
      # far had to buy its own correctness with its own rewriter and its own
      # self-test — sync-protocol-version.mjs (#2769), sync-template-versions.mjs,
      # sync-docs-image-tags.mjs (#9064) — because nothing validated the tree as
      # a whole. Maintainer ruling on the sibling card (objectui#5397,
      # 2026-08-22, 「接受所有」) chose option A: validate the POST-VERSION tree
      # HERE, after the version step and before the action opens or updates the
      # PR. NOT by handing the PR a PAT / GitHub App token (widens the
      # supply-chain trust surface, adds rotation obligations), and NOT one
      # artifact at a time — per-artifact guarantees do not generalise, which is
      # the half this repository has already paid for three times.
      #
      # THE SURFACE IS MEASURED HERE, NOT INHERITED (2026-08-25, at 3689991d2)
      # ---------------------------------------------------------------------
      # The sibling's scope conclusion does NOT transfer. There the version step
      # moves no source byte at all. Here `pnpm run version` is FOUR rewriters
      # and one of them writes TypeScript. Rendered against this repository's
      # real tree — 158 pending changesets, @objectstack/cli 17.2.0 -> 17.3.0 —
      # `pnpm run version` moved 313 paths, in four classes and no others:
      #
      #     158  .changeset/*.md      deleted (consumed by the version)
      #      76  */package.json       152 changed lines, every one `"version":`
      #                               — not one dependency range moved
      #      76  */CHANGELOG.md       generated from the changeset bodies
      #       3  doc surfaces         8 concrete image-tag / npm-pin lines, in
      #                               content/docs/deployment/self-hosting.mdx,
      #                               content/docs/upgrading.mdx, docker/README.md
      #
      # Two more classes exist and did NOT move on that train, because both are
      # major-boundary only: packages/spec/src/kernel/protocol-version.ts, and
      # per bundled template objectstack.config.ts / objectstack.manifest.json
      # (with the template's own package.json, whose `^N.0.0` ranges move too).
      # They stayed dormant while check-changeset-no-major.mjs held; pre mode
      # stands that guard aside, and the v18 opening woke them. RE-MEASURED on
      # that first boundary train (2026-10-07: 3d9188502e plus the opening's
      # `.changeset/pre.json` and `major` changeset, 37 changesets,
      # @objectstack/cli 17.7.0 -> 18.0.0-next.0): 236 paths, 0 outside the
      # surface, 4 major-only — the three blank-template files and
      # protocol-version.ts. The template half is now judged by the gates below;
      # protocol-version.ts is still refused, for the measured reason given at
      # the refusal.
      #
      # WHAT IS CHECKED, AND WHY EACH ONE (7 s total, measured on the rendered
      # tree in this job's own state: install, NO build. With the major-boundary
      # gates below, the whole step held 44 s on a shared dev container,
      # 2026-10-07, on a 17.8.0 train — every gate green)
      # ---------------------------------------------------------------------
      # Two halves, because the classes above fail in two different ways.
      #
      #   SHAPE — every moved path must fall inside the reviewed surface, and
      #   the doc and template halves of that surface are RESOLVED AT RUN TIME
      #   from the very declarations the rewriters read (`SURFACES` in
      #   check-docs-image-tag.mjs, `stampedPaths()` in
      #   sync-template-versions.mjs), on the principle those two declarations
      #   exist for: one list, N consumers, because "a fourth literal
      #   is a fourth contract". This is the half that ends the treadmill: a
      #   fifth rewriter, or an existing one growing an output, reddens this
      #   lane on the day it lands instead of shipping unvalidated.
      #
      #   CONTENT — the gates whose corpus is what the version step actually
      #   writes:
      #     check:docs-image-tag        the 3 doc surfaces against packages/cli's
      #                                 NEW version. #9064's gate, and the one
      #                                 that structurally could not fire here.
      #     check:docs-image-tag-sync   that rewriter's self-test — it just ran
      #                                 on this tree and a rewriter that has
      #                                 nothing to do cannot be observed working.
      #     check:template-version-sync the template stamper's, for the same
      #                                 reason.
      #     check:release-index-currency-sync
      #                                 the release-index stamper's (#15332), for the
      #                                 same reason again — on a train whose index is
      #                                 already current it rewrites nothing, and a
      #                                 rewriter with nothing to do cannot be observed
      #                                 working.
      #     check:nul-bytes             lint.yml's one UNCONDITIONAL gate whose
      #                                 corpus is every byte in the tree. The 76
      #                                 generated CHANGELOGs are the only prose
      #                                 the version step writes that it did not
      #                                 author, and this commit never runs
      #                                 lint.yml at all.
      #     check:release-notes, check:release-page-status,
      #     check-release-section-coverage.mjs, check:published-readme-links
      #                                 the four gates that READ CHANGELOG.md.
      #
      #   And the gates of the two MAJOR-boundary classes, run on EVERY refresh
      #   rather than only on the train that moves them: each reads a version
      #   this step writes on every train, and a gate that runs once a major
      #   rots unseen between majors. All existing gates; none needs a build.
      #     create-objectstack's suite  the bundled templates RENDERED: the
      #                                 scaffolder's own copy and identity
      #                                 rewrite, from a template packed the way
      #                                 npm ships it, plus the ratchets that
      #                                 judge all three stamps against
      #                                 create-objectstack's NEW major
      #                                 (template-consistency.test.ts).
      #                                 Measured on the boundary train: 16
      #                                 files, 249 tests, green, 23 s.
      #     check:template-manifests   each stamped manifest still parses as
      #                                 TemplateManifestSchema (2 s, green).
      #     protocol-version.test.ts   PROTOCOL_VERSION in lockstep with spec's
      #                                 NEW major. ONE file, 5 s — reachable
      #                                 without the whole spec suite, which
      #                                 this comment used to say it was not.
      #     check:spec-changes, check:upgrade-guide
      #                                 the two projections DERIVED from
      #                                 PROTOCOL_MAJOR (2-3 s each). Before the
      #                                 #22449 B′ ruling they compared committed
      #                                 copies, and were measured RED on the
      #                                 boundary train (the version pass moves
      #                                 the constant and regenerated neither).
      #                                 They now generate both in memory, since
      #                                 the publish job generates the shipped
      #                                 copies: red only when the post-version
      #                                 tree cannot generate them.
      #
      # ⛔ WHAT IS DELIBERATELY NOT RUN, AND WHY IT IS NOT AN OMISSION.
      #   The whole derived farm. Fed this 313-path change set,
      #   scripts/pm/dispatch-gates.mjs names 49 families; 44 of them run
      #   without a build, in 113 s, and every one is green on the post-version
      #   tree. The control that makes that number mean something: the identical
      #   sweep on the PRE-version tree returns the identical verdicts, so the
      #   version step changes no gate's mind. It is mostly restatement — a
      #   family is derived because a package's `package.json` moved, and what
      #   moved inside it was the `"version"` key, while the gate reads that
      #   package's SOURCE, which ci.yml judged on this very commit minutes ago.
      #   Running the selector here was drafted and withdrawn: it needs
      #   `scripts/check-self-test-wired.mjs` to grow a new ledger shape, and
      #   check-dispatch-gates.mjs's header already records the reviewed
      #   position that "there is no verdict in [the live derivation] for CI to
      #   hold". That is a maintainer's call, not a rider on this card.
      #   The five build-requiring families in that farm (check:i18n,
      #   check:i18n-coverage, check:dev-prereqs, check:doc-formula-expressions,
      #   check:doc-security-posture) would in any case re-add to this lane the
      #   ~9 minutes of build the comment below records as deliberately moved to
      #   the publish job, to gate a manifest's `"version"` key. What they read
      #   is not what the version step wrote: the two doc gates judge formula
      #   fences and security-posture prose, and the only content/** lines this
      #   step can move are the 8 concrete pins above.
      #
      # FAILURE SEMANTICS. A red validation fails the job, so the step below
      # never runs and the standing PR keeps its last VALIDATED content instead
      # of being force-pushed to a broken one. If the tree stays broken the PR
      # goes stale — but loudly, on a red run every six hours, which is the
      # opposite of this card's defect. Nothing here can reach the publish lane:
      # this whole job is `schedule` / `workflow_dispatch` only.
      - name: Render the post-version tree
        id: post_version
        run: |
          set -euo pipefail
          shopt -s nullglob

          # Whether `.changeset/pre/*.md` is pending depends on the prerelease
          # mode. @changesets/cli v3 moves every changeset a prerelease version
          # pass consumes into `.changeset/pre/` (the layout the 18.0.0-next.0
          # cut created, #23282). In `pre` mode those are CONSUMED: the version
          # pass skips them. On `pre exit` the GA version consumes them all
          # again. That is the CLI's own rule (`getRelevantChangesets`,
          # @changesets/assemble-release-plan), and the PR step's reader applies
          # the same split. Counting `pre/` in `pre` mode made a tree with no
          # new changeset read as hundreds pending: the version pass then moved
          # nothing and the validation below went red on a lane with nothing
          # to do. Measured on the post-version tree: 539 counted, 0 moved.
          pre_mode=''
          if [ -f .changeset/pre.json ]; then
            pre_mode="$(node -e 'process.stdout.write(String(JSON.parse(require("node:fs").readFileSync(".changeset/pre.json", "utf8")).mode ?? ""))')"
          fi

          # `/^README\.md$/i` is how `@changesets/read` spells its exclusion, so
          # the count is case-folded rather than matching one spelling.
          count_pending() {
            local n=0 file
            local globs=(.changeset/*.md)
            if [ "${pre_mode}" = 'exit' ]; then
              globs+=(.changeset/pre/*.md)
            fi
            for file in "${globs[@]}"; do
              if [ "$(printf '%s' "${file##*/}" | tr '[:upper:]' '[:lower:]')" = 'readme.md' ]; then
                continue
              fi
              n=$((n + 1))
            done
            printf '%s' "$n"
          }

          pending_before=$(count_pending)
          echo "pending_before=${pending_before}" >> "$GITHUB_OUTPUT"

          # The version PR was just merged, or nothing has landed since. The
          # action will take its no-changesets branch and do nothing, so there is
          # no post-version tree to render. Said out loud, because a validation
          # that silently does nothing is this card's own failure class.
          if [ "${pending_before}" -eq 0 ]; then
            echo 'No pending changesets — the version step would move nothing. Nothing to render, nothing to validate.'
            {
              echo '### Post-version validation'
              echo
              echo '- pending changesets: `0` — nothing to render, nothing to validate.'
            } >> "$GITHUB_STEP_SUMMARY"
            exit 0
          fi

          # Everything below reads the tree as the version step's OUTPUT, and
          # that reading is only true while nothing else is dirty first.
          # Asserted rather than assumed: measured locally, a single unrelated
          # edit left in the tree turned 313 moved paths into 314 and would have
          # been validated as though the version pass had written it.
          dirty_before=$(git status --porcelain | wc -l | tr -d ' ')
          if [ "${dirty_before}" -ne 0 ]; then
            git status --porcelain
            echo "::error::the working tree is already dirty before the version step (${dirty_before} path(s), listed above). The post-version change set is read as everything dirty, so a pre-existing change would be validated as though the version pass had written it."
            exit 1
          fi

          pnpm run version

          # Tracked moves (deletions included) plus new untracked files. Read
          # this way rather than by classifying `git status --porcelain` letters:
          # a rename arrives as a delete/add pair whose halves are both named
          # here, so no status code has to be interpreted.
          {
            git diff --name-only HEAD
            git ls-files --others --exclude-standard
          } | sort -u > "${RUNNER_TEMP}/post-version-paths.txt"
          moved=$(wc -l < "${RUNNER_TEMP}/post-version-paths.txt" | tr -d ' ')

          {
            echo '### Post-version validation'
            echo
            echo "- rendered from \`${GITHUB_SHA}\`"
            echo "- pending changesets consumed: \`${pending_before}\`"
            echo "- paths the version step moved: \`${moved}\`"
          } >> "$GITHUB_STEP_SUMMARY"

      - name: Validate the post-version tree
        if: steps.post_version.outputs.pending_before != '0'
        run: |
          set -euo pipefail

          MOVED_FILE="${RUNNER_TEMP}/post-version-paths.txt"
          moved=$(wc -l < "${MOVED_FILE}" | tr -d ' ')
          if [ "${moved}" -eq 0 ]; then
            echo "::error::changesets were pending but the version step moved no path. Either the version script did nothing, or it wrote outside the working tree — both are a broken version lane, never an empty diff."
            exit 1
          fi

          # ── SHAPE ────────────────────────────────────────────────────────
          # The doc surfaces the rewriter writes, read from the same declaration
          # the rewriter itself reads. Import-safe by construction: that module
          # carries an entry-point guard added for exactly this kind of consumer.
          SURFACE_LIST="${RUNNER_TEMP}/post-version-doc-surfaces.txt"
          if ! node --input-type=module \
                 -e 'import { SURFACES } from "./scripts/check-docs-image-tag.mjs"; for (const s of SURFACES) console.log(s.file);' \
                 > "${SURFACE_LIST}"; then
            echo "::error::could not resolve SURFACES from scripts/check-docs-image-tag.mjs, so the doc half of the post-version surface is unknown. Refusing to call this tree validated."
            exit 1
          fi
          # Unknown is a failure, never an empty allowlist: an empty list would
          # make every doc the version pass rewrites read as an unexpected path.
          if [ ! -s "${SURFACE_LIST}" ]; then
            echo "::error::SURFACES in scripts/check-docs-image-tag.mjs resolved EMPTY, so the doc surfaces the version pass rewrites would read as unexpected paths. Refusing to call this tree validated."
            exit 1
          fi

          # The template surfaces the stamper writes, on the same terms, from the
          # same walk and table its own main() uses.
          TEMPLATE_LIST="${RUNNER_TEMP}/post-version-template-surfaces.txt"
          if ! node --input-type=module \
                 -e 'import { stampedPaths } from "./scripts/sync-template-versions.mjs"; for (const p of stampedPaths()) console.log(p);' \
                 > "${TEMPLATE_LIST}"; then
            echo "::error::could not resolve stampedPaths() from scripts/sync-template-versions.mjs, so the template half of the post-version surface is unknown. Refusing to call this tree validated."
            exit 1
          fi
          if [ ! -s "${TEMPLATE_LIST}" ]; then
            echo "::error::stampedPaths() in scripts/sync-template-versions.mjs resolved EMPTY. It THROWS on a moved or empty template directory, so an empty file here is a resolution that reported nothing while still exiting 0. Refusing to call this tree validated."
            exit 1
          fi

          # The release-index surface the currency stamper writes, on the same terms
          # again: resolved from `syncedPaths()` in the rewriter, which derives it
          # from `INDEX_PATH` in the gate whose finding it clears. Third rewriter,
          # third resolved list, zero restated literals.
          INDEX_LIST="${RUNNER_TEMP}/post-version-release-index-surfaces.txt"
          if ! node --input-type=module \
                 -e 'import { syncedPaths } from "./scripts/sync-release-index-currency.mjs"; for (const p of syncedPaths()) console.log(p);' \
                 > "${INDEX_LIST}"; then
            echo "::error::could not resolve syncedPaths() from scripts/sync-release-index-currency.mjs, so the release-index half of the post-version surface is unknown. Refusing to call this tree validated."
            exit 1
          fi
          if [ ! -s "${INDEX_LIST}" ]; then
            echo "::error::syncedPaths() in scripts/sync-release-index-currency.mjs resolved EMPTY, so the release index the version pass rewrites would read as an unexpected path. Refusing to call this tree validated."
            exit 1
          fi

          # The SKILL.md stamps the skill-compatibility rewriter writes (the root
          # `version` script's last step, `--fix`), on the same terms: resolved from
          # `stampedSkillPaths()`, derived from the walk its `--fix` writes from, and
          # judged by `check:skill-compatibility` in the content half below.
          SKILL_LIST="${RUNNER_TEMP}/post-version-skill-surfaces.txt"
          if ! node --input-type=module \
                 -e 'import { stampedSkillPaths } from "./scripts/check-skill-compatibility-version.mjs"; for (const p of stampedSkillPaths()) console.log(p);' \
                 > "${SKILL_LIST}"; then
            echo "::error::could not resolve stampedSkillPaths() from scripts/check-skill-compatibility-version.mjs, so the skill half of the post-version surface is unknown. Refusing to call this tree validated."
            exit 1
          fi
          if [ ! -s "${SKILL_LIST}" ]; then
            echo "::error::stampedSkillPaths() in scripts/check-skill-compatibility-version.mjs resolved EMPTY. It THROWS on a moved or empty skills directory, so an empty file here is a resolution that reported nothing while still exiting 0. Refusing to call this tree validated."
            exit 1
          fi

          # Five filters over the reviewed surface: the fixed
          # release paths by pattern, then the four declared lists by WHOLE-LINE
          # EXACT match, so no derived filter can accept a path its declaration
          # does not name and none needs regex-escaping.
          UNEXPECTED="$(grep -vE '(^|/)package\.json$|(^|/)CHANGELOG\.md$|^\.changeset/|^packages/spec/src/kernel/protocol-version\.ts$' "${MOVED_FILE}" \
            | grep -vxF -f "${TEMPLATE_LIST}" \
            | grep -vxF -f "${SURFACE_LIST}" \
            | grep -vxF -f "${INDEX_LIST}" \
            | grep -vxF -f "${SKILL_LIST}" || true)"
          if [ -n "${UNEXPECTED}" ]; then
            printf '%s\n' "${UNEXPECTED}" | sed 's/^/::error::  unexpected: /'
            echo "::error::the version pass wrote outside the reviewed post-version surface (paths above). This is the treadmill guard: a new version-time output must arrive together with the gate that judges it. Add the surface to the declaration its rewriter reads, and its gate to the content half below — deliberately, in one reviewed diff."
            exit 1
          fi

          # The two MAJOR-boundary classes. Both are INSIDE the surface above —
          # protocol-version.ts by the pattern, the template stamps by
          # `stampedPaths()` — so the assertion accepts them and they reach this
          # check with a diagnosis of their own. Measured: with the pattern
          # missing, protocol-version.ts fell into the "wrote outside the
          # reviewed surface" branch instead, which is red for the right reason
          # and wrong about why.
          #
          # The TEMPLATE stamps are validated, not refused: the create-objectstack
          # suite and check:template-manifests in the content half judge them
          # (see the header). They are counted here only for the summary.
          #
          # protocol-version.ts is still REFUSED, and not for want of a gate: its
          # three gates run below. What moving it does is wider than all three.
          # The ADR-0087 D1 handshake (`assertProtocolCompat` on the app LOAD
          # seam, packages/runtime/src/app-plugin.ts) refuses every manifest whose
          # `engines.protocol` excludes the new major, and this repository's own
          # example apps declare the old one. Measured with the boundary train's
          # constant: `^17` -> OS_PROTOCOL_INCOMPATIBLE, `^18` -> ok (control on
          # the pre-version tree: the reverse). No gate this lane can afford boots
          # them, so the refusal stays until a reviewed change says how the
          # protocol major moves on this lane. It is collected rather than exited
          # on, so one boundary run names every gate verdict as well.
          MAJOR_TEMPLATES="$(grep -xF -f "${TEMPLATE_LIST}" "${MOVED_FILE}" || true)"
          UNVALIDATED=''
          if grep -qxF 'packages/spec/src/kernel/protocol-version.ts' "${MOVED_FILE}"; then
            UNVALIDATED='packages/spec/src/kernel/protocol-version.ts'
          fi

          # ── CONTENT ──────────────────────────────────────────────────────
          failed=()
          run_gate() {
            echo "::group::$*"
            if "$@"; then
              echo "green: $*"
            else
              failed+=("$*")
            fi
            echo "::endgroup::"
          }

          run_gate pnpm check:docs-image-tag
          run_gate pnpm check:docs-image-tag-sync
          run_gate pnpm check:template-version-sync
          run_gate pnpm check:release-index-currency-sync
          run_gate pnpm check:skill-compatibility
          run_gate pnpm check:nul-bytes
          run_gate pnpm check:release-notes
          run_gate pnpm check:release-page-status
          run_gate node scripts/check-release-section-coverage.mjs
          run_gate pnpm check:published-readme-links
          # The MAJOR-boundary classes' gates. `--fail-if-no-match`: a filter that
          # matches no package otherwise exits 0 having run nothing.
          run_gate pnpm --filter create-objectstack --fail-if-no-match test
          run_gate pnpm --filter @objectstack/spec --fail-if-no-match check:template-manifests
          run_gate pnpm --filter @objectstack/spec --fail-if-no-match exec vitest run --project local src/kernel/protocol-version.test.ts
          run_gate pnpm --filter @objectstack/spec --fail-if-no-match check:spec-changes
          run_gate pnpm --filter @objectstack/spec --fail-if-no-match check:upgrade-guide

          template_moved=$(printf '%s\n' "${MAJOR_TEMPLATES}" | sed '/^$/d' | wc -l | tr -d ' ')
          {
            echo "- post-version surface: \`${moved}\` path(s), all inside the reviewed shape"
            echo "- major-boundary template stamps moved: \`${template_moved}\` (judged by the content gates)"
            echo "- content gates failed: \`${#failed[@]}\`"
            echo "- major-boundary surfaces left unvalidated: \`$([ -n "${UNVALIDATED}" ] && echo 1 || echo 0)\`"
          } >> "$GITHUB_STEP_SUMMARY"

          if [ "${#failed[@]}" -ne 0 ]; then
            printf '::error::the post-version tree fails %s gate(s) that no CI run would ever have judged: %s\n' \
              "${#failed[@]}" "$(printf '%s; ' "${failed[@]}")"
          fi
          if [ -n "${UNVALIDATED}" ]; then
            echo "::error::  unvalidated: ${UNVALIDATED}"
            echo "::error::this version pass moved the protocol major. Its lockstep and the two artifacts derived from it are judged above, but the move also makes the protocol handshake refuse every manifest still declaring the old major in engines.protocol — this repository's example apps among them — and no gate on this lane boots them. A reviewed change must decide how the protocol major moves before this refresh goes through."
          fi
          if [ "${#failed[@]}" -ne 0 ] || [ -n "${UNVALIDATED}" ]; then
            exit 1
          fi
          echo 'Post-version tree validated.'

      # ══════════════════════════════════════════════════════════════════════
      # THE READ REHEARSAL — THE PR STEP'S OWN READER, ON THE TREE THE CUT LANDS
      # (#23282)
      # ══════════════════════════════════════════════════════════════════════
      # changesets/action does not use this repository's @changesets/cli to
      # decide whether anything is pending. It bundles its own @changesets/read,
      # so two readers of `.changeset/` meet on this lane, released on two
      # schedules. They disagreed for the first time at the 18.0.0-next.0 cut:
      # that version pass moved 507 consumed changesets into `.changeset/pre/`,
      # and the reader bundled in v1.9.0 (what `@v1` resolved to) takes any
      # directory under `.changeset/` for a legacy v1 folder changeset, so it
      # opened `.changeset/pre/changes.md` and failed with ENOENT on every
      # refresh after the cut. Nothing failed BEFORE the cut, because no tree
      # main had carried until then held `.changeset/pre/`.
      #
      # The tree rendered above IS the tree the next cut lands on main. So the
      # PR step's reader (same action, same pinned commit) reads it here, every
      # refresh, and must answer "nothing left to version". A reader that
      # cannot read what the CLI just wrote now fails while the Version
      # Packages PR is still open, instead of on the first refresh after it
      # merges. Measured against the next.0-shaped tree: v1.9.0 fails here with
      # the same ENOENT; v2.1.2 answers `has-changesets` false.
      #
      # ⛔ WHY IT CANNOT WRITE ANYTHING. The rehearsal runs on a COPY of the
      # changeset inputs outside every git work tree, and is handed no
      # credential. The only path out of a "pending" read is `runVersion`,
      # whose first act is `git checkout` (src/run.ts) and which fails there,
      # before any push or API call. Measured on the pre-version tree:
      # `has-changesets` true, then exit 1 at that checkout, zero API requests.
      # The copy carries the root `package.json` and `node_modules` so the
      # action's own Changesets CLI generation check runs against the
      # installed CLI too.
      - name: Stage the post-version changesets for the read rehearsal
        if: steps.post_version.outputs.pending_before != '0'
        run: |
          set -euo pipefail
          REHEARSAL="${RUNNER_TEMP}/changeset-read-rehearsal"
          rm -rf "${REHEARSAL}"
          mkdir -p "${REHEARSAL}"
          cp package.json pnpm-workspace.yaml "${REHEARSAL}/"
          cp -R .changeset "${REHEARSAL}/.changeset"
          ln -s "${GITHUB_WORKSPACE}/node_modules" "${REHEARSAL}/node_modules"
          if git -C "${REHEARSAL}" rev-parse --is-inside-work-tree > /dev/null 2>&1; then
            echo "::error::${REHEARSAL} is inside a git work tree, so the read rehearsal could reach a branch. Refusing to run it."
            exit 1
          fi
          echo "Staged $(find "${REHEARSAL}/.changeset" -name '*.md' | wc -l | tr -d ' ') changeset file(s) outside any git work tree: ${REHEARSAL}"

      - name: Rehearse the version-PR changeset read on the post-version tree
        id: read_rehearsal
        if: steps.post_version.outputs.pending_before != '0'
        # ⛔ Keep this ref identical to the PR step's below. A rehearsal of a
        # different reader proves nothing about the one that runs.
        uses: changesets/action@ae32849d5ba541f9ae29e40e22a623bc13562f51 # v2.1.2
        with:
          cwd: ${{ runner.temp }}/changeset-read-rehearsal
          # Not a credential. The rehearsal only reads files.
          github-token: no-credential-the-read-rehearsal-only-reads

      - name: Assert the read rehearsal found nothing left to version
        if: steps.post_version.outputs.pending_before != '0'
        env:
          HAS_CHANGESETS: ${{ steps.read_rehearsal.outputs.has-changesets }}
        run: |
          set -euo pipefail
          # `false` is the only passing answer. `true` on a zero exit is the
          # action's "all changesets are empty" branch: after the cut the PR
          # step would then refresh nothing, forever, with nothing red. Empty
          # means the output was renamed, so the pin no longer reads anything.
          if [ "${HAS_CHANGESETS}" != 'false' ]; then
            echo "::error::the version PR step's changeset reader, run on the post-version tree, answered has-changesets='${HAS_CHANGESETS}' where only 'false' is right: once this tree lands on main the PR step would misread it. Re-verify the pinned changesets/action against the changeset layout @changesets/cli writes before this version PR merges."
            exit 1
          fi
          echo "The PR step's reader reads the post-version tree as fully versioned: has-changesets=${HAS_CHANGESETS}."

      # ⚠️ LOAD-BEARING — DO NOT DROP THIS STEP. The steps above render the
      # version into the runner's working tree, and changesets/action picks its
      # branch from that tree. With `.changeset/` already consumed it finds
      # `hasChangesets` false, and this lane deliberately passes no
      # `publish-script:` input, so `hasPublishScript` is false too — the pinned
      # v2.1.2 then takes `case !hasChangesets && !hasPublishScript`
      # (src/index.ts), logs "No changesets present or were removed by merging
      # version PR" and RETURNS.
      # The refresh becomes a permanent no-op and #4935 fossilises with nothing
      # red anywhere: this file's own failure class, one lane over. So the tree
      # is put back, and the restoration is ASSERTED rather than assumed.
      - name: Restore the pre-version tree
        if: always()
        env:
          PENDING_BEFORE: ${{ steps.post_version.outputs.pending_before }}
        run: |
          set -euo pipefail
          shopt -s nullglob

          if [ -z "${PENDING_BEFORE}" ] || [ "${PENDING_BEFORE}" = '0' ]; then
            echo 'Nothing was rendered, so there is nothing to restore.'
            exit 0
          fi

          git checkout -- .
          # Narrow on purpose, and measured: those are the only roots the version
          # pass writes into. A `git clean` on a lane that can force-push a
          # branch should never be able to reach further than the thing it undoes.
          git clean -fdq -- .changeset packages examples content docker

          # The render step's count, read the same way and only now: `pre exit`
          # versions delete `.changeset/pre.json`, so the mode is read from the
          # RESTORED tree.
          pre_mode=''
          if [ -f .changeset/pre.json ]; then
            pre_mode="$(node -e 'process.stdout.write(String(JSON.parse(require("node:fs").readFileSync(".changeset/pre.json", "utf8")).mode ?? ""))')"
          fi
          count_pending() {
            local n=0 file
            local globs=(.changeset/*.md)
            if [ "${pre_mode}" = 'exit' ]; then
              globs+=(.changeset/pre/*.md)
            fi
            for file in "${globs[@]}"; do
              if [ "$(printf '%s' "${file##*/}" | tr '[:upper:]' '[:lower:]')" = 'readme.md' ]; then
                continue
              fi
              n=$((n + 1))
            done
            printf '%s' "$n"
          }

          pending_after=$(count_pending)
          dirty=$(git status --porcelain --untracked-files=no | wc -l | tr -d ' ')

          if [ "${pending_after}" != "${PENDING_BEFORE}" ] || [ "${dirty}" -ne 0 ]; then
            echo "::error::the pre-version tree was NOT restored (${PENDING_BEFORE} pending changeset(s) before, ${pending_after} after; ${dirty} tracked path(s) still modified). changesets/action would read this tree, find nothing pending and no publish script, and return without refreshing the PR — a silent permanent no-op (#11945)."
            exit 1
          fi
          echo "Pre-version tree restored: ${pending_after} pending changeset(s), working tree clean."

      # `pnpm run version` = changeset version + sync-protocol-version +
      # sync-template-versions + sync-docs-image-tags — FOUR rewriters, read
      # from the root `version` script, not three. The fourth joined it in #9064
      # and this comment did not follow. Corrected here because the block above
      # prices what that script moves, and a stale count of the rewriters is a
      # stale statement of the surface being priced. All four are pure file
      # rewrites; none reads a built artifact. The workspace build, the vendored Console build and the
      # live hotcrm smoke that used to run here were PRE-PUBLISH gates — they
      # moved to the `publish` job below, where the publish they gate now lives.
      # Leaving them here would gate nothing and cost ~9 minutes of every main
      # push.
      # Quoted because the name embeds `: ` — YAML would otherwise read it as a
      # nested mapping (caught by check:workflow-status-functions' self-test).
      - name: 'Create or update the "chore: version packages" PR'
        id: changesets
        # PINNED BY COMMIT (#23282). v2 is the major upstream pairs with
        # @changesets/cli v3: its bundled reader reads `.changeset/pre/` and, in
        # `pre` mode, leaves those consumed changesets out of `hasChangesets`
        # (src/readChangesetState.ts), and its CLI generation check refuses a
        # CLI v2 project. v1 was held on 2026-08-19 (#9208) because its
        # replacement's publish property had not been read, and the hold named
        # how to reopen it: read v2's source for that property and for the
        # bundled reader. Both were read at this commit, and the bundle was run
        # end to end on the next.0-shaped tree. A commit, not `@v2`: `v2` is a
        # BRANCH upstream, so a moving ref would change the bundled reader with
        # no diff here, and the read rehearsal above runs this same commit.
        uses: changesets/action@ae32849d5ba541f9ae29e40e22a623bc13562f51 # v2.1.2
        with:
          # ⛔ THERE IS NO `publish-script:` INPUT HERE, AND THAT IS THE FIX. ⛔
          #
          # Not an oversight and not a style choice — it is what makes this lane
          # structurally unable to publish, per #6170. The pinned v2.1.2
          # branches on `hasPublishScript = !!publishScript` (src/index.ts),
          # where `publishScript = core.getInput("publish-script")`:
          #
          #     case !hasChangesets && !hasPublishScript:
          #       core.info("No changesets present or were removed by merging
          #                  version PR. Not publishing because publish-script
          #                  is not set.");
          #       return;
          #     case !hasChangesets && hasPublishScript:
          #       await runPublish({...});   // unreachable: no publish-script
          #     case hasChangesets:
          #       await runVersion({...});   // version PR only
          #
          # With no publish script, `runPublish` is not reachable from any input
          # state the action can observe. Adding one back here re-arms the exact
          # lane that minted rc.3 and rc.4 without a human. v1's input names
          # (`publish:`, `version:`, `commit:`, `title:`) are not accepted
          # silently by v2: it throws on each one by name.
          github-token: ${{ secrets.GITHUB_TOKEN }}
          version-script: pnpm run version
          commit-message: 'chore: version packages'
          pr-title: 'chore: version packages'
          # v1's `commitMode: git-cli` default, spelled the v2 way. v2 otherwise
          # pushes the version commit through the GitHub API, a different path
          # whose commit and size behaviour this lane has never been measured on.
          push-with-git-cli: true
          # Both no-ops without a publish script. Kept so that re-adding one can
          # never silently resurrect #4900: the action posts each package's raw
          # CHANGELOG section as the Release body, and @objectstack/spec's
          # section for a single v17 RC is ~343k characters against the API's
          # 125,000 limit. Releases are created by scripts/release-github-releases.mjs,
          # and version tags are pushed only by the publish job below. v2 also
          # pushes tags on its own unless told not to, so that one is off too.
          create-github-releases: false
          push-git-tags: false

  # ══════════════════════════════════════════════════════════════════════════
  # PUSH LANE 2 — release integrity audit. Never mints. Reads github.sha and the
  # version commit it selects, both out of the object database; every backfill
  # builds from that version commit's tree, as the publish job does.
  # ══════════════════════════════════════════════════════════════════════════
  release-integrity:
    name: Release integrity (audit + no-mint backfill)
    # Runs on both RELEASE events (ADR-0125 D1): it is the single place that
    # reads what main carries and asks npm whether that version exists, so the
    # push lane and the repair lane converge on ONE predicate and one guard
    # instead of two code paths that can drift.
    #
    # This `if:` is new with #11233 and is the reason that sentence still holds.
    # The job used to carry no `if:` at all, which meant "every event this file
    # has" — correct while the file had exactly the two release events, and
    # wrong the moment a third arrived. Without it the 6-hourly tick and every
    # on-demand refresh would run a full release audit, and each one that found
    # main's version absent from npm would queue a deployment at the `release`
    # environment for a maintainer to dismiss. The audit mints nothing, so this
    # is not a safety guard; it is the noise guard the collision section above
    # argues for. Bookkeeping events do not get a release audit.
    #
    # `!inputs.snapshot_next` is the snapshot lane's half (ADR-0125, amendment
    # of 2026-11-05), read exactly as `refresh_version_pr` is: a snapshot
    # dispatch gets no stable audit, so it can never queue a stable publish
    # prompt — one a dispatch would park in the publish group beyond
    # `stale-prompts`' reach (it never cancels a dispatch run).
    if: >-
      github.event_name == 'push' ||
      (github.event_name == 'workflow_dispatch' && !inputs.refresh_version_pr && !inputs.snapshot_next)
    runs-on: ubuntu-latest
    # Run wall of the 376 push runs among the 400 most recent (2026-10-22..27), an upper
    # bound for this job: p99 4.6 min, max 6.9. The bound leaves room for the backfill
    # path (install plus Release writes), which none of them took.
    timeout-minutes: 20
    # Serialised: its backfills create GitHub Releases and push a runtime image,
    # and two runs doing that at once is not a state worth reasoning about.
    #
    # ⚠️ Residual race, stated rather than hidden: GitHub keeps one PENDING run
    # per group, so if the version-PR landing's audit is itself queued behind an
    # earlier one and a THIRD push arrives, this job is cancelled — and `publish`
    # needs it, so the release quietly does not queue. It is narrow (the landing
    # run has to be the pending one, not the running one) and it is visible (no
    # approval request arrives, and every later push's audit says in its notice
    # that main's version is unpublished and was not queued THERE) and it is
    # recoverable without any special handling: the version is still absent
    # from npm, so the `workflow_dispatch` repair lane re-audits and queues the
    # same deployment, on the same version commit. ⚠️ Until 2026-09-29 the next
    # landing re-queued it by accident — that accident was the defect the
    # amended D1 removes (every landing queued a publish of its OWN head), so
    # the dispatch is now the only recovery, by design.
    concurrency:
      group: release-integrity-${{ github.ref }}
      cancel-in-progress: false
    permissions:
      # `contents: write` is for GitHub Releases, never for refs: this job runs
      # no `git push` of any kind.
      contents: write
      # `actions: read` is the Releases backfill's guard: the audit reads this
      # workflow's runs and their jobs to ask whether the same version's
      # publish is in flight in another run. Reads only, never a cancel.
      actions: read
    outputs:
      # "the docker job must build" — set only when npm ALREADY has this
      # version's WHOLE fixed group and its runtime image is missing.
      published: ${{ steps.audit.outputs.image-missing }}
      cli-version: ${{ steps.audit.outputs.version }}
      # The commit `publish` checks out: the newest first-parent commit at which
      # @objectstack/cli's version differs from its parent's — the landing that
      # brought main to the version above (ADR-0125 D1 as amended 2026-09-29).
      # Computed on every audited event, because it is what the approval
      # authorises on both lanes.
      version-commit: ${{ steps.audit.outputs.version-commit }}
      # THE release predicate (ADR-0125 D1, amended 2026-09-29): on a push,
      # 'true' exactly when THIS push carries the version commit and that
      # version is absent from npm; on the repair dispatch, exactly when the
      # version is absent from npm. The old predicate — "absent from npm" alone
      # — stayed true on every landing until the publish finished, so each one
      # queued its own publish of its own head (17.5.0 shipped 8 PRs past its
      # version commit). A later landing now queues nothing and evicts nothing.
      publish-pending: ${{ steps.audit.outputs.publish-pending }}
    steps:
      # `fetch-depth: 0` is load-bearing: the version commit is found by walking
      # main's first-parent history, and the push range by ancestry against
      # `github.event.before`. On a shallow clone the graft boundary would read
      # as "the version changed here" — scripts/release-pending-publish.mjs
      # refuses a shallow clone rather than answer from one.
      - name: Checkout repository
        uses: actions/checkout@v7
        with:
          fetch-depth: 0

      - name: Setup Node.js
        uses: actions/setup-node@v7
        with:
          node-version-file: .nvmrc

      # ──────────────────────────────────────────────────────────────────────
      # R2 — the probe may only ever see the version main ACTUALLY carries.
      #
      # This job deliberately does not contain the changesets action, so no step
      # can re-version its workspace. Belt and braces on top of that: the version
      # is read out of the OBJECT DATABASE at `github.sha`, not off disk, and a
      # tripwire fails the run if the two ever disagree. Had this shape existed
      # on 2026-08-03 the run would have gone red instead of publishing rc.3.
      # ──────────────────────────────────────────────────────────────────────
      - name: Audit the release that main actually carries
        id: audit
        env:
          SHA: ${{ github.sha }}
          EVENT: ${{ github.event_name }}
          # On a push, main's tip before this push landed (all zeros on a branch
          # creation); empty on a dispatch, which has no push range.
          BEFORE: ${{ github.event.before }}
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: |
          version=$(git show "${SHA}:packages/cli/package.json" | jq -r '.version')
          if [ -z "$version" ] || [ "$version" = "null" ]; then
            echo "::error::could not read @objectstack/cli version at ${SHA}"
            exit 1
          fi

          # Tripwire, not decoration: this is the exact assertion the old
          # recover-publish step lacked. A workspace that disagrees with
          # github.sha means something re-versioned the tree, and that is the
          # #6170 mechanism — refuse to act on it rather than probe it.
          tree_version=$(jq -r '.version' packages/cli/package.json)
          if [ "$tree_version" != "$version" ]; then
            echo "::error::workspace carries @objectstack/cli@${tree_version} but ${SHA} carries ${version} — something re-versioned this workspace (#6170). Refusing to audit a version main does not have."
            exit 1
          fi

          echo "version=$version" >> "$GITHUB_OUTPUT"
          echo "main (${SHA}) carries @objectstack/cli@${version}"

          # ── the version commit, and whether THIS event queues its publish ─
          # ADR-0125 D1 as amended 2026-09-29. The logic, and the pins that hold
          # it (a version commit then two landings, a merge-queue batch, a
          # published version, a force-push, a shallow clone), live in
          # scripts/release-pending-publish.mjs, whose --self-test lint.yml
          # runs. It also asks npm (present / absent / unknown), so the npm
          # branch below reads its answer rather than asking twice.
          if ! selection=$(node scripts/release-pending-publish.mjs select --event "$EVENT" --head "$SHA" ${BEFORE:+--before "$BEFORE"}); then
            echo "::error::could not select the version commit for ${SHA} (reason above). Refusing to decide whether a release is pending."
            exit 1
          fi
          jq . <<<"$selection"
          selected_version=$(jq -r '.version' <<<"$selection")
          version_commit=$(jq -r '.versionCommit' <<<"$selection")
          npm_state=$(jq -r '.npm' <<<"$selection")
          pending=$(jq -r '.pending' <<<"$selection")
          reason=$(jq -r '.reason' <<<"$selection")
          range_detail=$(jq -r '.range.detail' <<<"$selection")
          if [ "$selected_version" != "$version" ]; then
            echo "::error::the version-commit walk read @objectstack/cli@${selected_version} at ${SHA}, this step read ${version}. Two readers of one object disagree; refusing to act on either."
            exit 1
          fi
          echo "version-commit=${version_commit}" >> "$GITHUB_OUTPUT"
          echo "version commit: ${version_commit} — the landing that brought main to ${version}"

          # ── npm ───────────────────────────────────────────────────────────
          # `unknown` (npm could not be read) keeps this step's historical
          # reading, "not on npm": no backfill runs off a guess, and a pending
          # publish is a prompt a human still has to approve.
          if [ "$npm_state" != 'present' ]; then
            if [ "$pending" = 'true' ]; then
              # THE predicate (ADR-0125 D1, amended). This push carries the
              # version commit and nothing has shipped it. This job still cannot
              # publish anything — it says so and stays green; the `publish` job
              # below reads this output, and IT stops dead on the `release`
              # environment until a required reviewer approves it.
              echo "publish-pending=true" >> "$GITHUB_OUTPUT"
              echo "::notice::main carries @objectstack/cli@${version}, which is NOT on npm; this ${EVENT} queues its publish on the version commit ${version_commit}. Once the full-battery job reads the battery at or before that commit green, a deployment is queued and waits for a maintainer to approve the 'release' environment."
              {
                echo "## Release ${version} is waiting for your approval"
                echo
                echo "main (\`${SHA}\`) carries **@objectstack/cli@${version}**, which is not on npm."
                echo "Its version commit is \`${version_commit}\`."
                echo
                echo "The **Publish ${version} to npm** job below is held at the \`release\`"
                echo "environment gate. Nothing has been checked out, built or published —"
                echo "GitHub holds the whole job until a required reviewer approves it."
                echo "It starts only after **Full battery green at the version commit** finds"
                echo "the newest full-battery runs at or before \`${version_commit}\` green."
                echo
                echo "**Review this before approving:** approving publishes the tree of the"
                echo "version commit \`${version_commit}\` — the commit that brought main to"
                echo "${version}, i.e. what the Version Packages PR decided — and nothing that"
                echo "landed after it, whatever main's head is by then."
              } >> "$GITHUB_STEP_SUMMARY"

              # ── what ${version} ships without a CHANGELOG entry (#21361) ───
              # The merge queue lands the Version Packages PR on top of the
              # current main, and the PR cannot be refreshed while queued, so
              # the version commit's tree can carry landings whose changesets
              # it never consumed: their code ships in this release and its
              # CHANGELOG does not name them (17.3.0 through 17.6.0 each did).
              # This is the run that queues the publish, so the approver reads
              # the list here, before the release is cut. A warning annotation:
              # the level this job gives every finding that leaves the release
              # incomplete without stopping it (the Releases, the D4 asset, the
              # image, an unreadable range). It REPORTS and never refuses, so
              # it neither fails this step nor holds the publish job. A report
              # that cannot be measured says so in an error annotation (a
              # shallow clone is refused, never answered from a graft
              # boundary), and the publish stays queued all the same.
              if ! node scripts/release-pending-publish.mjs unconsumed --version-commit "$version_commit" --version "$version"; then
                echo "The changesets the version commit ${version_commit} did not consume were NOT MEASURED (reason above). This report never holds a release, so the publish of ${version} stays queued."
                {
                  echo
                  echo "**Changesets: NOT MEASURED.** Whether ${version}'s CHANGELOG leaves out a change that landed in"
                  echo "the version commit's tree is unknown; the reason is in this step's log."
                } >> "$GITHUB_STEP_SUMMARY"
              fi
              exit 0
            fi
            echo "publish-pending=false" >> "$GITHUB_OUTPUT"
            case "$reason" in
              version-commit-landed-earlier)
                echo "::notice::main carries @objectstack/cli@${version}, not on npm, and its version commit ${version_commit} landed before this push (${range_detail}). Its publish prompt belongs to the push that carried it; this push queues none and evicts none. If no 'Publish ${version} to npm' prompt is waiting, dispatch this workflow (the repair lane) to queue it on ${version_commit}."
                ;;
              range-unreadable)
                echo "::warning::main carries @objectstack/cli@${version}, not on npm, and this push's range cannot be read (${range_detail}), so it queues no publish rather than guess whether it carried the version commit ${version_commit}. If no 'Publish ${version} to npm' prompt is waiting, dispatch this workflow (the repair lane) to queue it."
                ;;
              *)
                echo "::error::the predicate answered '${reason}' for an unpublished version; this step has no branch for it. Refusing to guess."
                exit 1
                ;;
            esac
            {
              echo "## Release ${version} is not queued by this push"
              echo
              echo "main (\`${SHA}\`) carries **@objectstack/cli@${version}**, which is not on npm."
              echo "Its version commit is \`${version_commit}\`; ${range_detail}."
              echo "Only the push that carries the version commit queues its publish."
            } >> "$GITHUB_STEP_SUMMARY"
            exit 0
          fi
          echo "publish-pending=false" >> "$GITHUB_OUTPUT"
          echo "npm: @objectstack/cli@${version} is present."

          # Every repair below is over an ALREADY-PUBLISHED version, so none of
          # it can mint anything. This is the #4900 case — published, then died
          # before the Releases / D4 asset / image existed. "Published" means
          # the WHOLE fixed group, and that is read further down, before any
          # backfill is requested: `cli` on npm only says a publish STARTED.

          # ── GitHub Releases + the ADR-0087 D4 asset ───────────────────────
          # Two anchors, not all 69: @objectstack/cli is the fixed group's
          # canary and @objectstack/spec is both the historical failure (its
          # ~343k body hit the API's 125k limit) and D4's mount point. The
          # backfill itself is idempotent create-or-update across the whole set.
          releases_ok=true
          gh release view "@objectstack/cli@${version}"  >/dev/null 2>&1 || releases_ok=false
          gh release view "@objectstack/spec@${version}" >/dev/null 2>&1 || releases_ok=false
          if [ "$releases_ok" = true ]; then
            gh release view "@objectstack/spec@${version}" --json assets \
              --jq '.assets[].name' 2>/dev/null | grep -qx 'spec-changes.json' || releases_ok=false
          fi

          # ── runtime image ─────────────────────────────────────────────────
          # A failed probe counts as MISSING on purpose: a redundant rebuild
          # costs a few minutes, a wrongly-skipped one leaves a published npm
          # version with no image and nothing to say so.
          image_ok=false
          if token=$(curl -fsS "https://ghcr.io/token?scope=repository:${GITHUB_REPOSITORY}:pull&service=ghcr.io" 2>/dev/null) \
            && token=$(node -p 'JSON.parse(process.argv[1]).token' "$token" 2>/dev/null) \
            && curl -fsS -o /dev/null -H "Authorization: Bearer $token" \
                 -H 'Accept: application/vnd.oci.image.index.v1+json' \
                 -H 'Accept: application/vnd.docker.distribution.manifest.list.v2+json' \
                 "https://ghcr.io/v2/${GITHUB_REPOSITORY}/manifests/$version" 2>/dev/null
          then
            image_ok=true
          fi

          if [ "$releases_ok" = true ] && [ "$image_ok" = true ]; then
            echo "GitHub Releases + ADR-0087 D4 asset are present for ${version}."
            echo "ghcr: image for $version is present — nothing to backfill."
            exit 0
          fi

          # ── the WHOLE fixed group, before any backfill (#20627) ──────────
          # `cli` on npm is not "the publish is complete": `changeset publish`
          # commits the 69 packages over minutes and cli is not the last. npm's
          # own `time` on 17.5.0: cli 07:58:57Z, console 08:05:24Z, spec — the
          # last — 08:09:33Z. A landing audited at 08:05:53Z read cli, took this
          # backfill, and its image build went red installing a group npm did
          # not have yet. So no backfill is requested until every package is on
          # npm, read by the ONE definition of "published" the publish job's
          # own post-publish check uses — scripts/release-verify-npm.mjs, asked
          # once (`--probe`) instead of waited on for 15 minutes. Its set is the
          # VERSION COMMIT's workspace, the tree the publish job checked out and
          # verified; this push's tree may carry a package no release of
          # ${version} contains, which would read as absent forever.
          #
          # Asked only when something IS missing, so the common path (nothing
          # to repair) pays no 69 registry reads. The cli reading above stays
          # the PUBLISH predicate on purpose: which push queues a publish is
          # ADR-0125 D1's question, and a publish that stopped partway already
          # has its repair — the `force` dispatch (D4) — not a push.
          vc_tree="${RUNNER_TEMP}/release-version-commit"
          git worktree add --quiet --detach "$vc_tree" "$version_commit"
          # The backfill steps below build from this tree, not from this
          # checkout (github.sha), so it is handed to them by name.
          echo "version-tree=${vc_tree}" >> "$GITHUB_OUTPUT"
          if ! group=$(RELEASE_VERSION="$version" node scripts/release-verify-npm.mjs --probe --root "$vc_tree"); then
            echo "::error::could not read whether the whole fixed group of ${version} is on npm (reason above). Refusing to request a backfill nobody measured."
            exit 1
          fi
          jq -c . <<<"$group"
          group_state=$(jq -r '.state' <<<"$group")
          group_checked=$(jq -r '.checked' <<<"$group")
          group_short=$(jq -r '.missing | length' <<<"$group")
          group_names=$(jq -r '[.missing[] | "\(.name)@\(.version)"] | (.[:5] | join(", ")) + (if length > 5 then ", and \(length - 5) more" else "" end)' <<<"$group")
          case "$group_state" in
            published)
              echo "npm: all ${group_checked} package(s) of the ${version} fixed group are present."
              ;;
            unpublished)
              echo "::notice::@objectstack/cli@${version} is on npm, but ${group_short} of ${group_checked} package(s) of its fixed group are not yet (${group_names}). A publish of ${version} is still shipping, or one stopped partway. Nothing is backfilled from a partly published group: the publish job creates the Releases, the ADR-0087 D4 asset and the runtime image itself, and a landing after the group is complete backfills whatever it did not. A publish that stopped partway is red in its own run, naming what is missing; dispatch this workflow with 'force' to finish it (ADR-0125 D4)."
              {
                echo "## Release ${version} is not fully on npm yet"
                echo
                echo "main (\`${SHA}\`) carries **@objectstack/cli@${version}**, which is on npm, but"
                echo "${group_short} of ${group_checked} packages of its fixed group are not: ${group_names}."
                echo "No GitHub Release, ADR-0087 D4 asset or runtime image is backfilled from a"
                echo "partly published group."
              } >> "$GITHUB_STEP_SUMMARY"
              exit 0
              ;;
            unknown)
              echo "::warning::@objectstack/cli@${version} is on npm, but npm could not be read for ${group_short} of ${group_checked} package(s) of its fixed group (${group_names}), so whether the whole group is published is unknown. Nothing is backfilled off a guess; the next landing reads again."
              exit 0
              ;;
            *)
              echo "::error::the group probe answered '${group_state}'; this step has no branch for it. Refusing to guess."
              exit 1
              ;;
          esac

          # ── the publish in flight, before any backfill ────────────────────
          # npm settles BEFORE the publish job reaches its own "Create GitHub
          # Releases" step, and before that run's `docker` job even exists (it
          # needs `publish`, so GitHub creates it when the publish job
          # completes). A landing audited in that window found the group
          # complete and wrote beside the publishing run. On 17.6.0 the publish
          # of run 36955885276 wrote its Releases 03:03:47Z -> 03:05:42Z, this
          # backfill in run 36958423332 started writing them at 03:04:37Z, and
          # five tags got two Release objects each; and both runs built and
          # pushed the 17.6.0 image, 03:05:49Z -> 03:10:02Z beside 03:05:48Z ->
          # 03:09:18Z. The writers are in different runs, so no `needs:` can
          # order them; the publishing run creates its Releases, D4 asset and
          # image, and a landing after it finishes backfills whatever it did
          # not. scripts/release-pending-publish.mjs `in-flight` (its
          # --self-test, run by lint.yml, holds the rule) answers both in one
          # read: `.state` for the publish job, `.docker.state` for that run's
          # image build, in flight from the approval until its docker job
          # completes. Anything it cannot read answers `in-flight`, so nothing
          # is backfilled off a guess. It only ever leaves `releases-missing`
          # and `image-missing` unset: this step stays green.
          flight_asked=true
          flight=$(GITHUB_TOKEN="$GH_TOKEN" node scripts/release-pending-publish.mjs in-flight --version "$version" --version-commit "$version_commit" --head "$SHA" --workflow release.yml) || flight_asked=false
          if [ "$flight_asked" = true ]; then
            jq -c . <<<"$flight"
          fi
          if [ "$releases_ok" = true ]; then
            echo "GitHub Releases + ADR-0087 D4 asset are present for ${version}."
          elif [ "$flight_asked" != true ]; then
            echo "::warning::${version}'s GitHub Releases or ADR-0087 D4 asset are incomplete, and whether its publish is still in flight could not be asked (reason above). Nothing is backfilled off a guess; the next landing reads again."
          elif [ "$(jq -r '.state' <<<"$flight")" != 'clear' ]; then
            if [ "$(jq -r '.reason' <<<"$flight")" = 'publish-in-flight' ]; then
              echo "::notice::${version}'s GitHub Releases or ADR-0087 D4 asset are incomplete, but its publish is still in flight ($(jq -r '.detail' <<<"$flight")). That run creates them; nothing is backfilled beside it. A landing after it finishes backfills whatever it did not."
            else
              echo "::warning::${version}'s GitHub Releases or ADR-0087 D4 asset are incomplete, and whether its publish is still in flight could not be read ($(jq -r '.detail' <<<"$flight")). Nothing is backfilled off a guess; the next landing reads again."
            fi
          else
            echo "::warning::${version} is on npm but its GitHub Releases or the ADR-0087 D4 asset are incomplete (#4900) — backfilling."
            echo "releases-missing=true" >> "$GITHUB_OUTPUT"
          fi
          if [ "$image_ok" = true ]; then
            echo "ghcr: image for $version is present."
          elif [ "$flight_asked" != true ]; then
            echo "::warning::No ghcr image for ${version} (or the registry could not be probed), and whether its publishing run is still building it could not be asked (reason above). No Docker job is requested off a guess; the next landing reads again."
          elif [ "$(jq -r '.docker.state' <<<"$flight")" != 'clear' ]; then
            if [ "$(jq -r '.docker.reason' <<<"$flight")" = 'docker-in-flight' ]; then
              echo "::notice::No ghcr image for ${version} yet, but its publishing run has not finished building it ($(jq -r '.docker.detail' <<<"$flight")). That run pushes the image; no Docker job is requested beside it. A landing after it finishes requests one if the image is still missing."
            else
              echo "::warning::No ghcr image for ${version} (or the registry could not be probed), and whether its publishing run is still building it could not be read ($(jq -r '.docker.detail' <<<"$flight")). No Docker job is requested off a guess; the next landing reads again."
            fi
          else
            echo "::warning::No ghcr image for $version (or the registry could not be probed) — requesting the Docker job."
            echo "image-missing=true" >> "$GITHUB_OUTPUT"
          fi

      # Everything below is skipped on the overwhelmingly common path (nothing to
      # repair), which is why the install is here rather than at the top of the job.
      #
      # ── Every backfill builds from the VERSION COMMIT's tree ─────────────
      # The publish job checks out the version commit and runs that commit's
      # own scripts over that commit's own tree (ADR-0125 D1 as amended
      # 2026-09-29). A repair describes the same release, so the steps below
      # run in the worktree the audit made of the version commit (its
      # `version-tree` output), with that commit's lockfile and tooling, and
      # never in this checkout. This checkout is `github.sha`, the head of
      # whichever push is being audited. A landing after the version commit
      # that touched packages/spec/src would otherwise make the D4 asset
      # describe a tree npm did not ship, and a Release created here would
      # point its CHANGELOG link, and any tag it has to create, at that landing.
      #
      # The checkout itself is NOT swapped: the audit above reads `github.sha`
      # and its tripwire reads this workspace, and both stay as they are. Each
      # step re-asserts that its tree is the version commit and refuses
      # otherwise, because an empty `version-tree` would leave the step in
      # this checkout, which is the defect, with nothing said.
      #
      # Pinned by battery 13 of `node scripts/release-verify-npm.mjs
      # --self-test` (lint.yml runs it): these steps' own text runs after the
      # audit's, and a landing that moved packages/spec/src is the fixture.
      - name: Setup pnpm
        if: steps.audit.outputs.releases-missing == 'true'
        uses: ./.github/actions/setup-pnpm

      - name: Install dependencies (the version commit's tree)
        if: steps.audit.outputs.releases-missing == 'true'
        env:
          VERSION_COMMIT: ${{ steps.audit.outputs.version-commit }}
          VERSION_TREE: ${{ steps.audit.outputs.version-tree }}
        run: |
          if [ -z "$VERSION_TREE" ] || [ "$(git -C "$VERSION_TREE" rev-parse HEAD 2>/dev/null)" != "$VERSION_COMMIT" ]; then
            echo "::error::the version commit's tree ('${VERSION_TREE}') is not at the version commit '${VERSION_COMMIT}'. Refusing to backfill from any other tree."
            exit 1
          fi
          cd "$VERSION_TREE"
          pnpm install --frozen-lockfile

      - name: Backfill GitHub Releases (bodies truncated to the API limit)
        if: steps.audit.outputs.releases-missing == 'true'
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          # No publish happened in this run, so there is no publishedPackages
          # JSON. RELEASE_VERSION drives the whole publishable workspace — the
          # Changesets `fixed` group releases every public package at one
          # version, which check-changeset-fixed.mjs gates.
          RELEASE_VERSION: ${{ steps.audit.outputs.version }}
          VERSION_COMMIT: ${{ steps.audit.outputs.version-commit }}
          VERSION_TREE: ${{ steps.audit.outputs.version-tree }}
        # The script reads every package and CHANGELOG.md from the tree it
        # lives in, and takes `GITHUB_SHA` as each Release's `target_commitish`
        # and as the ref of its CHANGELOG permalink. Both must be the version
        # commit. GitHub ignores `target_commitish` when the tag already exists,
        # but a publish that died before its tag push leaves no tag, and the
        # Release then CREATES it at `target_commitish`. So the version
        # commit's own script runs in its own tree, with GITHUB_SHA set for that
        # one process only.
        run: |
          if [ -z "$VERSION_TREE" ] || [ "$(git -C "$VERSION_TREE" rev-parse HEAD 2>/dev/null)" != "$VERSION_COMMIT" ]; then
            echo "::error::the version commit's tree ('${VERSION_TREE}') is not at the version commit '${VERSION_COMMIT}'. Refusing to backfill from any other tree."
            exit 1
          fi
          cd "$VERSION_TREE"
          GITHUB_SHA="$VERSION_COMMIT" node scripts/release-github-releases.mjs

      - name: Backfill spec-changes.json on the GitHub Release (ADR-0087 D4)
        # Ordering is load-bearing: `gh release upload` needs the Release the
        # step above creates.
        #
        # `--prepare` regenerates the manifest against the previously published
        # tarball exactly as the publish job does: in the version commit's
        # tree, with that commit's generator. So the asset this repair uploads
        # carries the same per-release section the npm artifact does — and, from
        # the version commit that taught `--prepare` the upgrade guide on, the
        # same guide. It cannot
        # repair the npm artifact itself — that tarball is immutable — and this
        # lane never publishes one.
        if: steps.audit.outputs.releases-missing == 'true'
        env:
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          RELEASE_VERSION: ${{ steps.audit.outputs.version }}
          VERSION_COMMIT: ${{ steps.audit.outputs.version-commit }}
          VERSION_TREE: ${{ steps.audit.outputs.version-tree }}
        run: |
          if [ -z "$VERSION_TREE" ] || [ "$(git -C "$VERSION_TREE" rev-parse HEAD 2>/dev/null)" != "$VERSION_COMMIT" ]; then
            echo "::error::the version commit's tree ('${VERSION_TREE}') is not at the version commit '${VERSION_COMMIT}'. Refusing to backfill from any other tree."
            exit 1
          fi
          cd "$VERSION_TREE"
          bash scripts/release-spec-changes.sh --prepare
          bash scripts/release-spec-changes.sh --attach

  # ══════════════════════════════════════════════════════════════════════════
  # PUSH LANE 3 — stale approval prompts. Cancels; never publishes, never
  # approves, never touches a job that has started.
  # ══════════════════════════════════════════════════════════════════════════
  stale-prompts:
    name: Cancel publish prompts for versions already on npm
    # ADR-0125 D1 as amended 2026-09-29: a prompt for a version that is already
    # published does not stay waiting.
    #
    # WHY IT MUST BE CANCELLED, NOT MERELY REPORTED — measured on the runs, not
    # assumed: the publish job that holds `release-publish-<ref>` waits at the
    # `release` environment WHILE HOLDING the group, and every later publish
    # job pends behind it. So a prompt left waiting after its version shipped
    # is not just a lookalike, it HIDES the next real one: 17.4.0's run
    # 34308599522 waited 20 days, was the only visible prompt when 17.5.0 was
    # queued, and was approved by mistake; 17.5.0 then left run 36539819278
    # waiting at 08:11Z, after cli@17.5.0 reached npm at 07:58:57Z.
    #
    # What is cancelled, exactly (scripts/release-pending-publish.mjs `sweep`,
    # pinned by its --self-test): a WAITING run of this workflow, not this run,
    # started by `push`, whose `Publish VERSION to npm (awaiting approval)` job
    # is waiting at the `release` environment, with VERSION present on npm.
    #   - A dispatch run is never touched: it is a human's own act, and the D4
    #     `force` repair legitimately waits for a version whose CLI is on npm.
    #   - A job GitHub holds at an environment has run no step, so this never
    #     interrupts a publish; the one it cancels could publish nothing — and
    #     if an approval lands in the second before the cancel, `publish`'s own
    #     guard refuses a push-lane release that is already on npm.
    # Rejecting instead was measured and is not available: the review endpoint
    # answers only a REQUIRED REVIEWER of the environment, which the workflow
    # token is not (and must not become).
    #
    # Same two events as `release-integrity`, for the same reason: bookkeeping
    # events get no release machinery. Not `needs`-ed by `publish` and needing
    # nothing itself: a failed read here must be loud (a red job on main) but
    # must never be able to hold up a release, and the order does not matter —
    # a publish job queued behind a stale holder takes the group the moment
    # the holder is cancelled.
    if: >-
      github.event_name == 'push' ||
      (github.event_name == 'workflow_dispatch' && !inputs.refresh_version_pr)
    runs-on: ubuntu-latest
    # Run wall of the 376 push runs among the 400 most recent (2026-10-22..27), an upper
    # bound for this job: p99 4.6 min, max 6.9; its own job reading on 18.1.0 (run
    # 40078805767) is 1.0 min.
    timeout-minutes: 10
    permissions:
      # `actions: write` is the cancel (POST .../runs/{id}/cancel); it also
      # covers the reads (the waiting-runs list, the runs at each version
      # commit's push, each named run, its jobs and its pending deployments).
      # `contents: read` is the checkout.
      actions: write
      contents: read
    steps:
      # `fetch-depth: 0`: a push-lane prompt can only hang on the push that
      # carried a version commit, and the sweep finds those on main's
      # first-parent history; it refuses a shallow clone.
      - name: Checkout repository
        uses: actions/checkout@v7
        with:
          fetch-depth: 0

      - name: Setup Node.js
        uses: actions/setup-node@v7
        with:
          node-version-file: .nvmrc

      - name: Cancel waiting publish prompts whose version is already on npm
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: node scripts/release-pending-publish.mjs sweep --workflow release.yml

  # ══════════════════════════════════════════════════════════════════════════
  # PUSH LANE 4 — the full battery, read before the approval exists (#22998).
  # Reads; never publishes, never approves, never cancels.
  # ══════════════════════════════════════════════════════════════════════════
  full-battery:
    name: Full battery green at the version commit
    # A PR or queue build carries ci.yml's FAST SET since #22998 (a queue
    # build adds the three dogfood shards since #30461); `dogfood-verify` and
    # the temporal leg run only on main's push, 4-hourly schedule and
    # dispatch runs. So "tested before a release" is read HERE, as a gate: the
    # newest completed `schedule` or `push` runs of `CI` and `Lint & Type
    # Check` whose head is the version commit or a first-parent ancestor of it
    # must both have concluded `success` (a `cancelled` run carries no
    # verdict). The selection rule and its pins live in
    # scripts/ci/scheduled-full-run.mjs (HALF 3), whose --self-test lint.yml
    # runs; this job only calls it.
    #
    # `publish` needs this job and its `if:` opens with `success()`, so a
    # refusal here skips `publish` before any deployment is queued: no
    # approval prompt appears for a version whose battery is red. Recovery is
    # the ordinary one — fix or re-run the red run at or before the version
    # commit, then the repair dispatch re-audits and queues the publish.
    #
    # Same predicate as `publish`, verbatim, so the read happens exactly when
    # a publish would be queued and on no other event; a skipped gate reads as
    # a skipped `needs` to `publish`, whose `success()` is then false — the
    # same fail-closed direction. `!inputs.snapshot_next` likewise: a snapshot
    # dispatch skips the audit, and this gate says so in its own guard.
    needs: [release-integrity]
    if: >-
      success() &&
      (needs.release-integrity.outputs.publish-pending == 'true' ||
       (github.event_name == 'workflow_dispatch' && inputs.force &&
        !inputs.refresh_version_pr && !inputs.snapshot_next))
    runs-on: ubuntu-latest
    timeout-minutes: 10
    permissions:
      # `actions: read` is the run listings of ci.yml and lint.yml; `contents:
      # read` is the checkout.
      actions: read
      contents: read
    steps:
      # `fetch-depth: 0`: "at or before the version commit" is a first-parent
      # walk from it, and the script refuses a shallow clone rather than read a
      # window that stops at the graft.
      - name: Checkout repository
        uses: actions/checkout@v7
        with:
          fetch-depth: 0

      - name: Setup Node.js
        uses: actions/setup-node@v7
        with:
          node-version-file: .nvmrc

      - name: Read the full battery at or before the version commit
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          VERSION_COMMIT: ${{ needs.release-integrity.outputs.version-commit }}
        run: node scripts/ci/scheduled-full-run.mjs --release-gate --release-commit "$VERSION_COMMIT"

  # ══════════════════════════════════════════════════════════════════════════
  # HUMAN LANE — the ONLY job in this repository that publishes a STABLE
  # version (the `next` snapshot lane at the end of this file is the other
  # publish, held at the same environment).
  # ══════════════════════════════════════════════════════════════════════════
  publish:
    # The job NAME is the approval screen (ADR-0125 D2). GitHub shows the job
    # name and the environment on the review prompt, so computing the name from
    # the audited version is what replaces the typed one: the maintainer
    # confirms a version they are SHOWN, read from the object database at
    # github.sha, rather than one they recall. #10146 is why that matters — the
    # version in that failure's log was @object-ui/console@17.5.0, the vendored
    # objectui build, while the repo was publishing 17.1.0.
    name: Publish ${{ needs.release-integrity.outputs.cli-version }} to npm (awaiting approval)
    # `full-battery` (#22998): the publish lane refuses, before the approval
    # prompt exists, a version whose full battery at or before the version
    # commit is not green. `success()` below is what makes that refusal bind.
    needs: [release-integrity, full-battery]
    # ⛔ THE barrier is now the environment, not the trigger (ADR-0125 D3).
    #
    #   An environment with NO protection rules passes AUTOMATICALLY and
    #   SILENTLY, and a run that passed an unprotected gate is indistinguishable
    #   in the log from one a human approved. Under this trigger that is not a
    #   degraded gate, it is NO gate: the push lane would publish end to end
    #   with nobody deciding, which is exactly the rc.3 / rc.4 incident — 69
    #   packages, tags, Releases and a runtime image, twice in one week, no
    #   human in the trigger chain.
    #
    #   Settings → Environments → release → Required reviewers was confirmed
    #   configured by the maintainer on 2026-08-20, and ADR-0125 is conditional
    #   on it staying that way. ⚠️ No file in this repo can assert it — not this
    #   YAML, not a CI gate, not the ADR. It is checked by a human opening the
    #   settings page. If those reviewers are ever removed, revert the `push`
    #   trigger to `workflow_dispatch` in the SAME change; do not leave this
    #   running.
    environment: release
    # `publish-pending` is the whole trigger predicate: on a push, true exactly
    # when that push carries the version commit and the version is absent from
    # npm; on the repair dispatch, when the version is absent from npm (ADR-0125
    # D1 as amended 2026-09-29). Unset (an audit that died before deciding)
    # compares false, so this fails CLOSED. `force` is dispatch-only and exists
    # for the partial-publish repair D4 describes.
    #
    # `success() && (...)`, with the parentheses load-bearing: `&&` binds tighter
    # than `||`, so `success() && A || B` would let the force branch publish on
    # top of a FAILED audit — and the audit is what computes the version this
    # job's name, guard and tag all read. If the audit dies, nothing publishes.
    #
    # `!inputs.refresh_version_pr` on the force branch is #11233's half, and it
    # is written even though it is currently redundant. A refresh dispatch skips
    # `release-integrity`, and a SKIPPED `needs` job makes `success()` false — so
    # the belt already holds (the `docker` job below documents that same GitHub
    # behaviour from the other direction). Redundant is not the same as
    # unnecessary: what makes the force branch safe would then be a fact about a
    # DIFFERENT job's `if:`, discoverable only by reading it, and the next person
    # to touch either guard gets no warning. This is the one job in the
    # repository that publishes a STABLE version; its guard states its own
    # preconditions.
    #
    # `!inputs.snapshot_next` is the snapshot lane's half (ADR-0125, amendment
    # of 2026-11-05), redundant in exactly the same way and written for the
    # same reason: `force` never reaches the snapshot lane, and the snapshot
    # input never reaches this job.
    if: >-
      success() &&
      (needs.release-integrity.outputs.publish-pending == 'true' ||
       (github.event_name == 'workflow_dispatch' && inputs.force &&
        !inputs.refresh_version_pr && !inputs.snapshot_next))
    runs-on: ubuntu-latest
    # ⛔ A timeout here cancels mid-`changeset publish`, the half-published state the
    # concurrency note below exists to prevent, so this bound is set from the steps' own
    # ceilings, not from the one reading: 18.1.0's job (run 40078805767) took 32.6 min,
    # the publish step 15.5 of it. Ceilings: build + Console build about 15 min cold,
    # `release-verify-npm.mjs` up to 35 min after `changeset publish`, the dist-tags
    # step's own 15 min; about 80 min in all, plus half again. The approval wait at the
    # `release` environment is not job time.
    timeout-minutes: 120
    # One publish at a time per ref, and never cancelled — a run cancelled
    # mid-`changeset publish` is the state that leaves a fixed group half on npm.
    #
    # ⚠️ ADR-0125 D5's eviction applies to THIS group too, and did, until the
    # 2026-09-29 amendment of D1. Measured on the runs: the job holding the
    # group waits at the `release` environment WHILE HOLDING it; every later
    # publish job pends behind it and evicts the pending one before it. While
    # the predicate was "absent from npm", every landing queued one, so the
    # prompt moved under the maintainer and whichever was pending last shipped
    # ITS head. Now only the push carrying the version commit queues a publish,
    # so the group holds one job per version and nothing evicts it; and a job
    # left holding the group for a version already on npm is cancelled by
    # `stale-prompts` above, never by `cancel-in-progress`.
    concurrency:
      group: release-publish-${{ github.ref }}
      cancel-in-progress: false
    permissions:
      contents: write
    outputs:
      published: ${{ steps.publish.outputs.published }}
      cli-version: ${{ steps.guards.outputs.version }}
    steps:
      # THE VERSION COMMIT, not `github.sha` (ADR-0125 D1 as amended
      # 2026-09-29). `github.sha` is the head of whichever push queued this run;
      # the version commit is the landing that brought main to the version the
      # approval screen names — the tree the Version Packages PR decided to
      # release. On 17.5.0 the two were eight PRs apart, one of them breaking.
      #
      # `fetch-depth: 0` is load-bearing for the guard below: "the version
      # commit is on main" is an ancestry test against the freshly fetched
      # `origin/main`, and on a shallow clone a non-ancestor answer is not
      # evidence of anything.
      - name: Checkout repository
        uses: actions/checkout@v7
        with:
          ref: ${{ needs.release-integrity.outputs.version-commit }}
          fetch-depth: 0

      - name: Setup Node.js
        uses: actions/setup-node@v7
        with:
          # .nvmrc's pin (check-node-version holds every workflow to it). It
          # must also satisfy the downstream hotcrm smoke below, which clones
          # hotcrm's latest release and installs it: pnpm install aborts with
          # ERR_PNPM_UNSUPPORTED_ENGINE outside that release's engines.node
          # (v6.0.0, the latest on 2026-10-26: `^22.11 || ^24 || >=26`).
          node-version-file: .nvmrc

      # ──────────────────────────────────────────────────────────────────────
      # The publish lane may only ever ship a commit that is ALREADY on main.
      # That is the other half of #6170's title: rc.3 and rc.4 tagged commits
      # that lived only on `changeset-release/main`, so main kept stale versions
      # and every later release recomputed an npm-occupied number. Nothing here
      # runs `changeset version`; this job publishes what the ref carries, or it
      # fails.
      # ──────────────────────────────────────────────────────────────────────
      - name: Guard the approved release (branch, version commit, and the tree matches it)
        id: guards
        env:
          # What release-integrity read out of the object database at
          # github.sha, and what the approval screen named. Read through env,
          # never interpolated into the shell.
          AUDITED: ${{ needs.release-integrity.outputs.cli-version }}
          # The commit the checkout above was asked for. Empty would make
          # `ref:` fall back to github.sha — the very defect — so it is refused.
          VERSION_COMMIT: ${{ needs.release-integrity.outputs.version-commit }}
          # 'true' only on the dispatch-only D4 repair, whose whole point is a
          # version whose CLI is already on npm.
          FORCE: ${{ github.event_name == 'workflow_dispatch' && inputs.force }}
        run: |
          if [ "${GITHUB_REF}" != "refs/heads/main" ]; then
            echo "::error::the publish lane may only run on main (got ${GITHUB_REF}). Publishing from any other ref would tag and ship code that never landed."
            exit 1
          fi

          # ── the version commit (ADR-0125 D1 as amended 2026-09-29) ────────
          # Every assertion here fails closed: the checkout must be the commit
          # the audit selected, that commit must already be on main, and it must
          # be the commit that CHANGED the version — not merely one carrying it.
          if ! printf '%s' "$VERSION_COMMIT" | grep -Eqx '[0-9a-f]{40}'; then
            echo "::error::release-integrity named no version commit ('${VERSION_COMMIT}'). An empty ref would check out github.sha — the head of whichever push queued this run — so refusing."
            exit 1
          fi
          head=$(git rev-parse HEAD)
          if [ "$head" != "$VERSION_COMMIT" ]; then
            echo "::error::the workspace is at ${head}, not the version commit ${VERSION_COMMIT}. Refusing to publish a tree nobody selected."
            exit 1
          fi
          if [ "$(git rev-parse --is-shallow-repository)" != 'false' ]; then
            echo "::error::shallow clone: an ancestry answer here would not be evidence. The checkout above must keep fetch-depth: 0."
            exit 1
          fi
          if ! git merge-base --is-ancestor "$VERSION_COMMIT" refs/remotes/origin/main; then
            echo "::error::the version commit ${VERSION_COMMIT} is not on main as fetched just now. Publishing it would tag and ship code that never landed."
            exit 1
          fi
          if ! git merge-base --is-ancestor "$VERSION_COMMIT" "$GITHUB_SHA"; then
            echo "::error::the version commit ${VERSION_COMMIT} is not an ancestor of ${GITHUB_SHA}, the head it was selected from."
            exit 1
          fi

          # R2's tripwire, kept (ADR-0125 D1), now against the version commit.
          # The typed version is gone, so this is the thing standing between an
          # approval and a publish of something main does not carry. It is the
          # exact assertion the 2026-08-03 recover-publish step lacked when it
          # shipped rc.3 off a re-versioned workspace: read the version from the
          # OBJECT DATABASE at the commit being published and refuse if the
          # checked-out tree disagrees.
          committed=$(git show "${VERSION_COMMIT}:packages/cli/package.json" | jq -r '.version')
          declared=$(jq -r '.version' packages/cli/package.json)
          if [ -z "$committed" ] || [ "$committed" = "null" ]; then
            echo "::error::could not read @objectstack/cli version at ${VERSION_COMMIT}"
            exit 1
          fi
          if [ "$declared" != "$committed" ]; then
            echo "::error::workspace carries @objectstack/cli@${declared} but ${VERSION_COMMIT} carries ${committed} — something re-versioned this workspace (#6170). Refusing to publish a version main does not have."
            exit 1
          fi
          parent=$(git show "${VERSION_COMMIT}^1:packages/cli/package.json" 2>/dev/null | jq -r '.version' || true)
          if [ "$parent" = "$committed" ]; then
            echo "::error::${VERSION_COMMIT} carries @objectstack/cli@${committed} but so does its parent, so it is not the commit that changed the version. Refusing to publish from anything but the version commit."
            exit 1
          fi

          # The approval was given against the audited number. If the tree has
          # moved since, the human approved a different release than the one
          # about to ship — refuse rather than ship the surprise.
          if [ -n "$AUDITED" ] && [ "$AUDITED" != "$committed" ]; then
            echo "::error::the release approved was @objectstack/cli@${AUDITED} but ${VERSION_COMMIT} carries ${committed}. Refusing to publish a version nobody approved."
            exit 1
          fi

          # ── a stale prompt, approved anyway ───────────────────────────────
          # A prompt can outlive its release: queued while the version was
          # absent, approved after some other run shipped it. The 17.4.0 prompt
          # of run 34308599522 was approved that way, in place of 17.5.0's.
          # Approving one must publish nothing and say so. `force` is exempt —
          # a partial publish leaves the CLI canary on npm by definition. An
          # unreadable registry is not evidence of staleness: `changeset
          # publish` skips an already-published version by itself.
          #
          # The same three-way reading as classifyNpmView() in
          # scripts/release-pending-publish.mjs, INLINED on purpose: this guard
          # runs before the checked-out tree is trusted, so it executes nothing
          # from that tree — git, jq and npm only, as the rest of it does. (The
          # version commit's tree may also predate the script.)
          if [ "$FORCE" != 'true' ]; then
            if npm_out=$(npm view "@objectstack/cli@${committed}" version 2>"${RUNNER_TEMP}/npm-view.err"); then
              if [ "$(printf '%s' "$npm_out" | tr -d '[:space:]')" = "$committed" ]; then
                echo "::error::@objectstack/cli@${committed} is already on npm, so this approval prompt was stale when it was approved — nothing is published from it. If a newer release is due, its own 'Publish VERSION to npm' prompt is the one to approve; if ${committed} is only partly published, dispatch this workflow with 'force' (ADR-0125 D4)."
                exit 1
              fi
            elif ! grep -q 'E404' "${RUNNER_TEMP}/npm-view.err"; then
              cat "${RUNNER_TEMP}/npm-view.err"
              echo "::warning::npm could not be read to confirm @objectstack/cli@${committed} is still unpublished; continuing, since changeset publish skips any version the registry already has."
            fi
          fi

          echo "version=$committed" >> "$GITHUB_OUTPUT"
          echo "Publishing @objectstack/cli@${committed} from its version commit ${VERSION_COMMIT} (approved on the 'release' environment; this run was queued by ${GITHUB_SHA})."
          {
            echo "## Publishing ${committed}"
            echo
            echo "- version commit: \`${VERSION_COMMIT}\` — the tree that is built, tagged and published"
            echo "- queued by: \`${GITHUB_REF}\` @ \`${GITHUB_SHA}\` (\`${GITHUB_EVENT_NAME}\`)"
            echo "- started by: \`${GITHUB_ACTOR}\`"
            echo "- authorised by: the \`release\` environment approval on this run —"
            echo "  see the run's deployment history for the reviewer and timestamp"
          } >> "$GITHUB_STEP_SUMMARY"

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

      # Mostly a CONSUMER now, not a seeder: this job used to run on every main
      # push and warmed the cache for everyone; it now runs only when a human
      # publishes. lint.yml's "Save Turbo cache (main only)" is the seeder. The
      # key is namespaced by `github.job`, which changed from `release` to
      # `publish` — the first release after this PR builds cold once. Keyed on
      # the VERSION COMMIT, the tree actually built, not on `github.sha`
      # (ADR-0125 D1 as amended 2026-09-29); turbo re-hashes its inputs either
      # way, so the key names the entry and decides nothing about correctness.
      #
      # ⚠️ From the checkout on, the steps are this FILE at `github.sha` while
      # the scripts they run are the VERSION COMMIT's. On the version push the
      # two are the same commit or a merge-queue batch apart; on a repair
      # dispatch they can be further apart. A step naming a script the version
      # commit predates fails on the missing file, which is the right failure:
      # a release is built by the version commit's own tooling.
      - name: Setup Turbo cache
        uses: actions/cache@v6
        with:
          path: .turbo/cache
          key: ${{ runner.os }}-turbo-${{ github.job }}-${{ github.ref_name }}-${{ needs.release-integrity.outputs.version-commit }}
          restore-keys: |
            ${{ runner.os }}-turbo-${{ github.job }}-${{ github.ref_name }}-
            ${{ runner.os }}-turbo-${{ github.job }}-

      - name: Verify Changesets "fixed" group covers every public package
        run: pnpm check:changeset-fixed

      # ──────────────────────────────────────────────────────────────────────
      # ⛔ TOMBSTONE — no objectui pin is read, built or checked here, and none
      # is re-added (objectstack#25711, D1 ruling 「同一个版本号组」).
      #
      # The Console is a workspace package now: `ui/apps/console`
      # (@object-ui/console) builds against this tree's @objectstack/* through
      # `workspace:^`, and `@objectstack/console#build:dist` (root turbo.json)
      # copies that dist with the tracked sdui.manifest.json; `pnpm run release`
      # runs it (`pnpm run console:build`) after the build below. The
      # pin-currency gate, the vendored-dist cache keyed on `.objectui-sha`,
      # its SHA stamp check and
      # its spec-injection check are gone with the pin: each answered "was the
      # Console built from the objectui commit, and against the spec, this
      # release names", and both are true by construction when the Console
      # source and the spec are one tree at one commit. The single-zod canary
      # survives inside packages/console/build-dist.sh, on the bytes it ships.
      # ──────────────────────────────────────────────────────────────────────

      - name: Build
        run: pnpm run build

      - name: Downstream backward-compat smoke (live hotcrm)
        # Pre-publish check (#2035): the about-to-publish @objectstack/spec
        # should not break a real third-party consumer pinned to a published
        # release. The deterministic in-repo floor is
        # @objectstack/downstream-contract, which blocks; this is the live
        # ceiling, and it is ADVISORY: it runs and reports on every publish, and
        # a failure does NOT block. `BLOCKING` below is the whole switch.
        #
        # How it got here. #3600 made it advisory for the rc pre-mode window and
        # promised it would re-arm at `changeset pre exit`. It did (#8643,
        # 2026-08-14), against a hand-pinned HOTCRM_REF that was still a 14.7-era
        # hotcrm, so the re-armed gate stood between the 17.0.0 GA publish and a
        # migration nobody had done yet: the deadlock #3600 named, one event
        # later. It went back to advisory, and the hand pin then sat four majors
        # behind hotcrm, so every publish emitted a warning that was certain to
        # appear and meant nothing.
        #
        # The posture now is a ruling, not a pending to-do (objectstack#25770,
        # D10 = A, 2026-10-25): the step stays advisory, and its ref is no
        # longer pinned by hand. It resolves at run time to hotcrm's LATEST
        # PUBLISHED RELEASE (`releases/latest` -> `tag_name`), so the check
        # always runs against the consumer as it ships today. ⛔ Never hard-code
        # a tag here again. The read is fail-soft: an unresolvable tag (network
        # error, non-200, empty or malformed `tag_name`) is a `::notice::` and a
        # skip, never a failure. Making this blocking is a new-gate decision for
        # the maintainer, not a switch left for someone to flip.
        env:
          # Read-only use: the releases API for a public repo. Authenticated so
          # the call is not subject to the anonymous per-IP rate limit shared
          # across hosted runners.
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          # '1' = a failure fails the publish. '0' = reported only.
          BLOCKING: '0'
        run: |
          api='https://api.github.com/repos/objectstack-ai/hotcrm/releases/latest'
          auth=()
          if [ -n "${GH_TOKEN:-}" ]; then
            auth=(-H "Authorization: Bearer ${GH_TOKEN}")
          fi
          body="$(mktemp)"
          code="$(curl -sS --max-time 30 --retry 2 -o "$body" -w '%{http_code}' \
            -H 'Accept: application/vnd.github+json' "${auth[@]}" "$api")" || true
          tag=''
          if [ "$code" = '200' ]; then
            tag="$(jq -r '.tag_name // empty' "$body" 2>/dev/null || true)"
          fi
          rm -f "$body"
          # The token's only use is the read above. hotcrm's install below runs
          # third-party lifecycle scripts, which must not inherit it.
          unset GH_TOKEN auth
          if ! printf '%s' "$tag" | grep -Eqx '[A-Za-z0-9][A-Za-z0-9._/-]*'; then
            echo "::notice::could not resolve hotcrm's latest published release from ${api} (HTTP ${code:-none}, no usable tag_name) — skipping the advisory downstream smoke. The publish continues."
            exit 0
          fi
          export HOTCRM_REF="$tag"
          echo "Resolved hotcrm's latest published release: ${HOTCRM_REF}"

          if bash scripts/downstream-smoke.sh; then
            echo "::notice::hotcrm@${HOTCRM_REF} is compatible with the about-to-publish @objectstack/spec."
            exit 0
          fi

          if [ "$BLOCKING" = '1' ]; then
            echo "::error::hotcrm@${HOTCRM_REF} is incompatible with the about-to-publish @objectstack/spec — refusing to publish a release that breaks a real downstream consumer."
            exit 1
          fi

          echo "::warning::hotcrm@${HOTCRM_REF} (hotcrm's latest release) is incompatible with the about-to-publish @objectstack/spec. ADVISORY ONLY — the publish continues; the in-repo @objectstack/downstream-contract floor is the blocking check."

      # ──────────────────────────────────────────────────────────────────────
      # ADR-0087 D4 — the per-release section, and the gate that proves it.
      #
      # Both steps run BEFORE `changeset publish`, and that ordering is the
      # whole design. The section has to be in the tree when the tarball is
      # packed (a consumer's tooling reads `node_modules`, not a Release page),
      # and a delta that disagrees with the artifacts must stop the release
      # while stopping it is still free — after `changeset publish` the tarball
      # is immutable and the only remaining repair is another version.
      #
      # `--no-git-checks` is what `changeset publish` passes to `pnpm publish`
      # (its own source), so the working-tree edits `--prepare` makes do not
      # block the publish.
      #
      # Since the #22449 B′ ruling this is also where BOTH registry projections
      # are generated for the tarball — spec-changes.json's per-major records and
      # aggregate, and packages/spec/protocol-upgrade-guide.md — and `--verify`
      # refuses a packed copy that differs from a fresh generation. No committed
      # copy is read: the pull-request check generates the same projections in
      # memory and renders their diff, and compares nothing.
      # ──────────────────────────────────────────────────────────────────────
      - name: Generate the per-release spec-changes section (ADR-0087 D4)
        env:
          RELEASE_VERSION: ${{ steps.guards.outputs.version }}
        run: bash scripts/release-spec-changes.sh --prepare

      - name: Correctness gate — the delta must match both tarballs
        # ⛔ A failure here fails the release, by design (#17080): a wrong change
        # file is worse than none, because a consumer stops looking once it has
        # one. The gate names the disagreeing exports and the direction of each
        # disagreement, so a held release arrives with its own diagnosis.
        env:
          RELEASE_VERSION: ${{ steps.guards.outputs.version }}
        run: bash scripts/release-spec-changes.sh --verify

      # ──────────────────────────────────────────────────────────────────────
      # The publish itself. `pnpm run release` = build (both scopes; the turbo
      # cache makes it the build above) + console:build (the Console dist into
      # packages/console) + scripts/release-publish.sh, which runs, in order:
      #   1. `node ui/scripts/check-published-dist-tooling.mjs --no-build`, which
      #      refuses the publish when a published @object-ui package carries
      #      tooling material (a test, a mock, a `*.tsbuildinfo`) or has no build
      #      output. `--no-build` because the build above is what gets packed,
      #      so the scan reads exactly those artifacts;
      #   2. `changeset publish`, with npm's own per-package answers recorded for
      #      scripts/release-verify-npm.mjs, which this step runs next;
      #   3. ONE atomic `git push origin --tags` (#2191: the action's concurrent
      #      per-tag pushes raced GitHub's ref backend and lost ~half the tags).
      #
      # changesets/action is NOT used here, and that is deliberate: handed a
      # workspace with pending changesets it would take the VERSION path and mint
      # a commit. `changeset publish` can only ever publish the versions the
      # checked-out package.json files already declare — the versions the guard
      # step above proved main carries.
      # ──────────────────────────────────────────────────────────────────────
      - name: Publish to npm + push version tags
        id: publish
        env:
          NPM_TOKEN: ${{ secrets.NPM_TOKEN }}
          VERSION: ${{ steps.guards.outputs.version }}
        run: |
          printf '//registry.npmjs.org/:_authToken=%s\n' "$NPM_TOKEN" >> "$HOME/.npmrc"
          git config user.name 'github-actions[bot]'
          git config user.email '41898282+github-actions[bot]@users.noreply.github.com'

          pnpm run release

          # `changeset publish` skips versions already on the registry, so a
          # re-dispatch over a partially-published release is a repair, not a
          # duplicate. What is NOT optional is that EVERY package is on npm when
          # this step ends.
          #
          # ⚠️ This checks what happens AFTER a publish, never WHETHER one runs.
          # The gating above it — `environment: release`, the `publish-pending`
          # predicate and its parenthesisation, `force`'s dispatch-only
          # semantics — is the "ONLY A HUMAN PUBLISHES" invariant and is not
          # this check's business.
          #
          # It used to be one `npm view` of @objectstack/cli, and #15321 is the
          # two ways that failed on the 17.3.0 release, which had succeeded
          # completely:
          #
          #   ① It raced the registry. npm committed cli@17.3.0 at 10:53:25.108
          #      and this read at 10:53:32 — seven seconds later — and got an
          #      absence. The write path and the CDN-fronted read path are
          #      eventually consistent; a single shot immediately after a
          #      69-package burst bets the release on how fast reads settle.
          #      scripts/release-verify-npm.mjs retries with bounded backoff for
          #      15 minutes, derived from that release's own npm `time` field:
          #      the last package (@objectstack/runtime) committed 7m07s after
          #      the first read.
          #
          #   ② It looked at 1 package of 69. Had a package genuinely failed to
          #      publish, this step could not have seen it — and the false red
          #      on `cli` would have MASKED it. The verifier now derives the
          #      whole publishable set from the workspace (never a transcribed
          #      list) and names, on failure, exactly which packages are absent,
          #      so the `force` repair dispatch has something to act on.
          #
          # ⛔ It still fails CLOSED. The retry absorbs LATENCY, never absence:
          # a package still missing when the budget is spent exits non-zero, and
          # so does a registry that could not be read at all. No `|| true`, no
          # downgrade to a warning — the value of this step is that it reds.
          # `published=true` below is written only on its success.
          #
          # `$VERSION` is handed over as RELEASE_VERSION — the spelling the two
          # steps below already use for the same value — so the verifier refuses
          # outright if the workspace it derives targets from disagrees with the
          # version this run was approved for.
          if ! RELEASE_VERSION="$VERSION" node scripts/release-verify-npm.mjs; then
            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: ${{ steps.guards.outputs.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. Uploads both projections: spec-changes.json and
        # protocol-upgrade-guide.md.
        if: ${{ !cancelled() && steps.publish.outputs.published == 'true' }}
        env:
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          RELEASE_VERSION: ${{ steps.guards.outputs.version }}
        # Uploads the files the two steps above generated and verified — the
        # Release assets and the npm artifact are the same bytes by construction,
        # not by two runs agreeing.
        run: bash scripts/release-spec-changes.sh --attach

      # #25581: a stable release moves `latest` and nothing else, so the `next`
      # and `rc` tags a pre-release cycle left behind keep resolving to versions
      # older than it (`npx create-objectstack@next` → 18.0.0-next.5 after
      # 18.0.0 shipped). This moves every pre-release tag that the release has
      # overtaken up to it; a tag AHEAD of it (a cycle in progress) and a tag on
      # a stable version are left alone. A pre-release publish is a no-op. Rule,
      # refusals and the one-time cleanup command: the script's header.
      #
      # ⚠️ After the publish, and outside its gating: this step mints nothing
      # and moves no tag to a version that is not already `latest`. Its
      # credential is the `$HOME/.npmrc` the publish step wrote — no secret is
      # handed to it here.
      #
      # Non-blocking by design: npm already has the whole group, so a tag left
      # stale warns (`::warning::` per package, step marked failed) and never
      # fails the job — the runtime image below `needs` this job's success.
      - name: Move stale pre-release dist-tags up to latest
        if: ${{ !cancelled() && steps.publish.outputs.published == 'true' }}
        continue-on-error: true
        timeout-minutes: 15
        env:
          RELEASE_VERSION: ${{ steps.guards.outputs.version }}
        run: node scripts/release-dist-tags.mjs --apply

  # ══════════════════════════════════════════════════════════════════════════
  # Runtime image — fed by either lane. Building an image for a version that is
  # already on npm cannot mint anything, so the push lane may request it.
  # ══════════════════════════════════════════════════════════════════════════
  docker:
    name: Docker image
    needs: [release-integrity, publish]
    # Publish the official runtime image (ghcr.io/objectstack-ai/objectstack).
    # Called as a reusable workflow so the same build can be re-run manually via
    # workflow_dispatch (e.g. base-image CVE rebuilds) — see docker-publish.yml.
    #
    # `!cancelled()` rather than the default implicit success(): at most one of
    # the two upstream jobs runs on any given event, so the other is always
    # SKIPPED — under the implicit success() this job would then never run at
    # all. It also survives a publish job that reached npm and then died
    # (#4900). The outputs are the gate; the jobs' statuses are not.
    #
    # On #11233's bookkeeping events (`schedule`, or a dispatch carrying
    # `refresh_version_pr`) NEITHER upstream job runs, which is why "at most"
    # replaced "exactly". Nothing else here changes: both outputs are then unset,
    # unset compares false against 'true', and this job stays skipped — the
    # outputs were already the gate, so a third event needed no new condition.
    if: ${{ !cancelled() && (needs.release-integrity.outputs.published == 'true' || needs.publish.outputs.published == 'true') }}
    permissions:
      contents: read
      packages: write
    uses: ./.github/workflows/docker-publish.yml
    with:
      version: ${{ needs.publish.outputs.cli-version || needs.release-integrity.outputs.cli-version }}

  # ══════════════════════════════════════════════════════════════════════════
  # SNAPSHOT LANE — a maintainer-approved `next` snapshot of main (ADR-0125,
  # amendment of 2026-11-05; ruling record 6663894403, #29466 letter B).
  # Dispatch-only, and exclusive: it starts only on a dispatch carrying
  # `snapshot_next`, and that dispatch runs none of the stable lane's jobs.
  # ══════════════════════════════════════════════════════════════════════════
  #
  # What it does, and what it never does:
  #   - snapshots the head of the newest green full-battery run on main, at
  #     distance zero (`scheduled-full-run.mjs --snapshot-commit`, confirmed by
  #     `--release-gate --require-distance-zero`);
  #   - runs the bare `changeset version --snapshot next` in its own checkout and
  #     commits none of it — the bumped manifests, the consumed changesets and
  #     the rendered CHANGELOG entries stay in the runner (`contents: read`,
  #     `persist-credentials: false`: nothing here can push);
  #   - publishes with `changeset publish --tag next --no-git-tag`, after the
  #     guard (`scripts/ci/snapshot-next.mjs`) proved every package goes out
  #     under `next` alone and no member of the fixed group would debut;
  #   - ⛔ pushes no tag, cuts no GitHub Release, uploads no ADR-0087 D4 asset,
  #     builds no runtime image, never touches the Version Packages PR, runs
  #     none of the stable post-publish acts and never moves `latest`.
  #
  # Every refusal — not on main, no distance-zero green commit, a release-
  # candidate cycle (`.changeset/pre.json` `mode: pre`), a member with no version
  # on npm, a dry run that fails — lands in the PREFLIGHT job, which carries no
  # environment, so no approval prompt ever appears for a snapshot that would be
  # refused. The publish job re-reads the cycle and the debut rule after the
  # approval, because an approval can come a day later.
  #
  # ⛔ D3 holds here exactly as for `publish`: if `release` ever has no required
  # reviewers, this lane is void, and it is removed in the same change that
  # reverts `publish` to a dispatch trigger (ADR-0125's amendment says why).
  snapshot-next-preflight:
    name: Snapshot next preflight (battery, RC cycle, debut, dry run)
    if: >-
      github.event_name == 'workflow_dispatch' && inputs.snapshot_next &&
      !inputs.refresh_version_pr
    runs-on: ubuntu-latest
    # Install, the version step and a pack of the whole fixed group; the pack
    # is the long pole.
    timeout-minutes: 40
    permissions:
      # `actions: read` is the run listings of ci.yml and lint.yml;
      # `contents: read` is the checkout.
      actions: read
      contents: read
    outputs:
      commit: ${{ steps.commit.outputs.commit }}
      commit-short: ${{ steps.commit.outputs.commit-short }}
    steps:
      # `fetch-depth: 0`: the commit is found on main's first-parent history,
      # and the reader refuses a shallow clone.
      - name: Checkout repository
        uses: actions/checkout@v7
        with:
          fetch-depth: 0
          persist-credentials: false

      - name: Setup Node.js
        uses: actions/setup-node@v7
        with:
          node-version-file: .nvmrc

      # The newest commit at which both full-battery workflows (CI, Lint & Type
      # Check) have their deciding run ON it, green — then the release gate,
      # asked at that commit with distance zero required, must agree. Both read
      # through `--json`, so no prose line is parsed. Runs off this checkout
      # with no install: scheduled-full-run.mjs imports node builtins only.
      - name: Select the commit (the newest with the full battery green on it)
        id: commit
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: |
          set -euo pipefail
          if [ "${GITHUB_REF}" != "refs/heads/main" ]; then
            echo "::error::the snapshot lane cuts from main only (got ${GITHUB_REF})."
            exit 1
          fi
          node scripts/ci/scheduled-full-run.mjs --snapshot-commit --from "${GITHUB_SHA}" --json > "${RUNNER_TEMP}/snapshot-commit.json"
          commit=$(jq -r '.commit // empty' "${RUNNER_TEMP}/snapshot-commit.json")
          if ! printf '%s' "$commit" | grep -Eqx '[0-9a-f]{40}'; then
            echo "::error::the selection named no commit ('${commit}'). Refusing."
            exit 1
          fi
          node scripts/ci/scheduled-full-run.mjs --release-gate --release-commit "$commit" --require-distance-zero --json > "${RUNNER_TEMP}/snapshot-gate.json"
          if ! jq -e --arg c "$commit" '.ok == true and .distanceZero == true and .requireDistanceZero == true and .releaseCommit == $c' "${RUNNER_TEMP}/snapshot-gate.json" > /dev/null; then
            jq . "${RUNNER_TEMP}/snapshot-gate.json"
            echo "::error::the release gate does not read the full battery green ON ${commit} (record above). Refusing."
            exit 1
          fi
          git checkout --quiet --detach "$commit"
          echo "commit=${commit}" >> "$GITHUB_OUTPUT"
          echo "commit-short=${commit:0:12}" >> "$GITHUB_OUTPUT"
          {
            echo "## \`next\` snapshot of main \`${commit}\`"
            echo
            echo "- dispatched at main \`${GITHUB_SHA}\` by \`${GITHUB_ACTOR}\`; the snapshot is cut $(jq -r '.offset' "${RUNNER_TEMP}/snapshot-commit.json") first-parent commit(s) behind it"
            jq -r '.workflows[] | "- \(.name): run \(.runId) (\(.event)) concluded \(.conclusion) ON the commit (distance \(.distance))"' "${RUNNER_TEMP}/snapshot-gate.json"
          } >> "$GITHUB_STEP_SUMMARY"

      # From here on the tree is the selected commit, with its own tooling.
      - name: Setup workspace
        uses: ./.github/actions/setup-workspace

      - name: Verify Changesets "fixed" group covers every public package
        run: pnpm check:changeset-fixed

      # The RC cycle, the fixed group and the debut rule, read at the selected
      # commit. The cycle is read a second time at main's tip as this job's
      # checkout fetched it, before the detach (the detach moves HEAD, never
      # `refs/remotes/origin/main`): the selected commit is older than main's
      # tip, and a cycle entered since would own `next`. A refusal here fails
      # the job, so `snapshot-next` is skipped and no deployment is queued.
      - name: Refuse during an RC cycle, or before every group member has debuted
        run: node scripts/ci/snapshot-next.mjs preflight --main-ref refs/remotes/origin/main

      # The dry run: the bare version step, the guard over changesets' publish
      # plan and over every packed manifest (pnpm substitutes `workspace:*` at
      # pack time), then the tree put back. Publishes nothing; no token is read.
      - name: Dry run (version, plan, packed manifests, tag)
        run: node scripts/ci/snapshot-next.mjs dry-run --main-ref refs/remotes/origin/main

  snapshot-next:
    # The approval screen (ADR-0125 D2, and the amendment's rule for this lane):
    # it says this is a `next` snapshot and names the main commit it is cut
    # from. ⛔ It must never read like the stable prompt, which
    # scripts/release-pending-publish.mjs reads as `Publish <one token> to npm
    # (awaiting approval)` (PUBLISH_JOB_NAME) — pinned by snapshot-next.mjs's
    # self-test.
    name: Snapshot next of main ${{ needs.snapshot-next-preflight.outputs.commit-short }} (awaiting approval)
    needs: [snapshot-next-preflight]
    # Held whole until a required reviewer approves (D2, D3): no checkout, no
    # build, no version step, no publish before the click.
    environment: release
    if: >-
      success() &&
      github.event_name == 'workflow_dispatch' && inputs.snapshot_next &&
      !inputs.refresh_version_pr
    runs-on: ubuntu-latest
    # Build + Console build about 15 min cold, the guard's pack a few minutes,
    # `changeset publish` about 15, the registry read-back up to 35, then the
    # dist-tag read. The approval wait is not job time.
    timeout-minutes: 120
    # Its OWN group (ADR-0125 D5 as the amendment reads it): a snapshot waiting
    # at the environment holds this group, never `release-publish-<ref>`, so it
    # neither hides nor evicts a stable prompt, and none evicts it. Never
    # cancelled in progress: a cancel mid-publish leaves the group half on npm.
    concurrency:
      group: release-snapshot-next-${{ github.ref }}
      cancel-in-progress: false
    permissions:
      # Read only: this job pushes no tag and cuts no Release, so it holds no
      # credential that could.
      contents: read
    steps:
      - name: Checkout the selected commit
        uses: actions/checkout@v7
        with:
          ref: ${{ needs.snapshot-next-preflight.outputs.commit }}
          fetch-depth: 0
          persist-credentials: false

      - name: Setup Node.js
        uses: actions/setup-node@v7
        with:
          node-version-file: .nvmrc

      # Before anything from the tree is trusted: the checkout is the commit the
      # preflight selected, and that commit is on main.
      - name: Guard the approved snapshot (branch, commit)
        env:
          COMMIT: ${{ needs.snapshot-next-preflight.outputs.commit }}
        run: |
          set -euo pipefail
          if [ "${GITHUB_REF}" != "refs/heads/main" ]; then
            echo "::error::the snapshot lane may only run on main (got ${GITHUB_REF})."
            exit 1
          fi
          if ! printf '%s' "$COMMIT" | grep -Eqx '[0-9a-f]{40}'; then
            echo "::error::the preflight named no commit ('${COMMIT}'). An empty ref would check out github.sha, which the battery has not passed. Refusing."
            exit 1
          fi
          if [ "$(git rev-parse HEAD)" != "$COMMIT" ]; then
            echo "::error::the workspace is at $(git rev-parse HEAD), not the selected commit ${COMMIT}. Refusing."
            exit 1
          fi
          if [ "$(git rev-parse --is-shallow-repository)" != 'false' ]; then
            echo "::error::shallow clone: an ancestry answer here would not be evidence."
            exit 1
          fi
          if ! git merge-base --is-ancestor "$COMMIT" refs/remotes/origin/main; then
            echo "::error::${COMMIT} is not on main as fetched just now. Refusing."
            exit 1
          fi
          {
            echo "## Publishing a \`next\` snapshot of \`${COMMIT}\`"
            echo
            echo "- started by: \`${GITHUB_ACTOR}\`; authorised by the \`release\` environment approval on this run"
            echo "- dist-tag: \`next\` alone; no tag, no GitHub Release, \`latest\` untouched"
          } >> "$GITHUB_STEP_SUMMARY"

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

      - name: Verify Changesets "fixed" group covers every public package
        run: pnpm check:changeset-fixed

      # Re-read after the approval, which can come a day after the preflight:
      # the cycle and the debut rule again, and every member's dist-tags as they
      # stand now — the record `verify` compares `latest` against. The cycle is
      # read in the selected commit's tree AND at the `refs/remotes/origin/main`
      # this job's checkout fetched just now (the ancestry guard above reads the
      # same ref), so a cycle entered during the approval wait refuses here.
      - name: Refuse during an RC cycle, or before every group member has debuted (after the approval)
        run: node scripts/ci/snapshot-next.mjs preflight --main-ref refs/remotes/origin/main --record "${RUNNER_TEMP}/snapshot-next-before.json"

      # ⛔ The BARE command, never `pnpm run version`: that chain aborts at
      # sync-protocol-version.mjs on a major below 1, and the snapshot is
      # `0.0.0-next-<datetime>`. Nothing it writes is committed.
      - name: Version the workspace as a next snapshot (never committed)
        id: version
        run: |
          set -euo pipefail
          pnpm exec changeset version --snapshot next
          version=$(jq -r '.version' packages/cli/package.json)
          if ! printf '%s' "$version" | grep -Eqx '0\.0\.0-next-[0-9]{14}'; then
            echo "::error::the version step left @objectstack/cli at '${version}', not a next snapshot. Refusing."
            exit 1
          fi
          echo "version=${version}" >> "$GITHUB_OUTPUT"
          echo "- version: \`${version}\`" >> "$GITHUB_STEP_SUMMARY"

      - name: Build
        run: pnpm run build

      - name: Build the Console dist
        run: pnpm run console:build

      # The two ADR-0087 registry projections the spec tarball ships, with NO
      # per-release section — the pre-release half of the script, reading
      # nothing from npm — then the gate that the packed copies match a fresh
      # generation.
      - name: Generate and verify the spec projections (no per-release section)
        env:
          RELEASE_VERSION: ${{ steps.version.outputs.version }}
        run: |
          set -euo pipefail
          bash scripts/release-spec-changes.sh --generate
          bash scripts/release-spec-changes.sh --verify

      # THE GUARD: every package publishes under `next` alone, at one snapshot
      # version of the whole group, and every packed manifest resolves inside
      # the snapshot. Fails the job before anything reaches the registry.
      - name: Guard the publish (dist-tag, versions, plan, packed manifests)
        run: node scripts/ci/snapshot-next.mjs guard --pack-dir "${RUNNER_TEMP}/snapshot-next-pack"

      # The publish: scripts/release-snapshot-publish.sh runs the stable lane's
      # published-dist tooling refusal, then the ruling's command,
      # `changeset publish --tag next --no-git-tag`, off the workspace's own
      # binary (never through `pnpm run`, whose npm_config_globalconfig breaks
      # the nested publish — scripts/release-publish.sh says why). No tag is
      # created, and this job holds no credential that could push one.
      - name: Publish to npm under next (no git tag)
        env:
          NPM_TOKEN: ${{ secrets.NPM_TOKEN }}
        run: |
          set -euo pipefail
          printf '//registry.npmjs.org/:_authToken=%s\n' "$NPM_TOKEN" >> "$HOME/.npmrc"
          bash scripts/release-snapshot-publish.sh

      # Every package of the group at the snapshot version on npm, with the
      # stable lane's own bounded read-back (it absorbs latency, never absence).
      - name: Verify every package reached npm
        env:
          RELEASE_VERSION: ${{ steps.version.outputs.version }}
        run: node scripts/release-verify-npm.mjs

      # `next` names the snapshot on every member, and no `latest` moved to it.
      # A `latest` on the snapshot is an incident for the maintainer — never
      # repaired with a counter-publish (Prime Directive 15).
      - name: Verify the dist-tags (next on the snapshot, latest unmoved)
        env:
          VERSION: ${{ steps.version.outputs.version }}
        run: node scripts/ci/snapshot-next.mjs verify --record "${RUNNER_TEMP}/snapshot-next-before.json" --version "$VERSION"