Pin calendar version when changesets bumps a major

[andrewmcgov]

Background

The Version Packages (2026-10-rc) PR (#4603) landed with the version 2027.0.0-rc.1 instead of 2026.10.0-rc.1.

Root cause

We deliberately mark API-version cuts as major changesets so the generated CHANGELOG reads "Major Changes". But changeset version feeds that bump type straight into semver.inc(), and on a calendar version a semver major bump always rolls the year:

semver.inc('2026.10.0-rc.1', 'major')  ->  '2027.0.0'

node-semver has a "pre-major" shortcut that would keep 1.0.0-rc.0 at 1.0.0, but it only fires when minor and patch are both 0. Our minor is the month, so it never fires:

inc('1.0.0-rc.0',     'major') = 1.0.0      <- shortcut fires
inc('2026.10.0-rc.0', 'major') = 2027.0.0   <- minor=10, real major bump

This is not new. The same thing happened on 2026-07-rc (#4290) — the bot's commit produced 2027.0.0-rc.1 from 2026.7.0-rc.0, and it only reached npm as 2026.7.0-rc.1 because it was hand-corrected in three follow-up commits ("lint fix", "update to correct versioning", "update extension tester example test to new api version"). It will recur on every RC branch that receives a major changeset.

Note that only major is hazardous on an RC branch. minor and patch are no-ops there, because node-semver leaves minor alone when patch is 0 and a prerelease is present — which is why rc.2 released cleanly off a minor changeset.

Left unfixed, the real risk isn't the red CI (that's loud and obvious) — it's that 2027.0.0-rc.1 reaching npm would outrank every future 2026.x release on the rc dist-tag permanently, and npm versions can't be reused.

Solution

In this repo YYYY.MM is chosen by a human at RC-cut time. Changeset bump types are changelog semantics only and must never move it.

So scripts/changeset-version.mjs wraps changeset version: run it as normal, then re-pin YYYY.MM and keep whatever patch/prerelease counter changesets calculated.

// Calendar segment untouched (normal patch bump) — nothing to do.
if (next.major === prev.major && next.minor === prev.minor) return computed;

const pre = next.prerelease.length ? `-${next.prerelease.join('.')}` : '';
let pinned = `${prev.major}.${prev.minor}.${prev.patch}${pre}`;

// Stable branch: no prerelease counter to advance, so move the patch instead.
if (!semver.gt(pinned, before)) {
  pinned = `${prev.major}.${prev.minor}.${prev.patch + 1}${pre}`;
}

It then swaps the stale version across every packages/*/package.json and CHANGELOG.md — covering package versions, internal dependency ranges, and changelog headings in one pass — and finishes with prettier --write .changeset/pre.json, which fixes the second CI failure permanently.

It's wired in as the version input of changesets/action, which defaults to changeset version:

uses: changesets/action@e0145edc7d9d8679003495b11f87bd8ef63c0cba # v1.5.3
with:
  title: Version Packages (${{ github.ref_name }})
  version: yarn version:calendar
  publish: yarn run deploy ...

semver is added as a root devDependency, pinned to ^7.5.3 so it reuses the resolution already in yarn.lock (no lockfile change, no duplicate copy in node_modules).

How the release flow works now

Nothing changes about how you release. Each release is still two pushes to the version branch:

1. Merge a PR containing .changeset/foo.md
        │
        ▼  push to 2026-10-rc
2. deploy.yml → changesets job
        │
   ┌────┴──────────────────────────────────────────────┐
   │ unconsumed changesets exist?                      │
   │   YES → runs `version` input                      │
   │         = yarn version:calendar   ◄── NEW         │
   │         opens/updates "Version Packages" PR       │
   │   NO  → runs `publish` input                      │
   │         = changeset publish → npm                 │
   └───────────────────────────────────────────────────┘
        │
3. Merge the "Version Packages" PR
        │
        ▼  push to 2026-10-rc
4. deploy.yml runs again → nothing unconsumed → publish → npm

The script runs only at step 2, on the runner. It never runs on the publish leg, and never locally unless you invoke yarn version:calendar yourself. In pre mode "unconsumed" means not yet listed in pre.json's changesets array, which is why merging the Version Packages PR doesn't re-trigger versioning even though the .md files remain on the branch.

Deliberately unaffected: /snapit, deploy:unstable and deploy:internal all produce 0.0.0-* snapshot versions where calendar pinning would be meaningless, so they bypass the wrapper.

Scenario	changesets computes	after fix
New major changeset mid-RC	2027.0.0-rc.3	2026.10.0-rc.3
Normal patch changeset mid-RC	2026.10.0-rc.3	untouched, no-op
major on a stable branch	2027.0.0	2026.10.1 (no collision with 2026.10.0)

Full CI gate green locally: yarn build, yarn type-check, eslint . --max-warnings 0 (on a pristine tree), prettier --check, tester typecheck + 75 tests, all 4 example suites, yarn docs:admin.