test(updates): macOS auto-update E2E test pipeline with real install + relaunch

[charlesvien]

Problem

The electron-builder + electron-updater migration gave us a real auto-update path (download, install, relaunch), but nothing tests it end to end.

Changes

Two tiny test-gated hooks in the main process, both inert unless POSTHOG_E2E_UPDATE_FEED is set (no production behavior change):

• electron-updater.ts points the updater at a local feed via setFeedURL. disableDifferentialDownload, autoDownload, autoInstallOnAppQuit and isSupported are untouched.
• main/index.ts exposes globalThis.__e2eUpdates (check/download/install/status) wired to the real UpdatesService.

Harness:

• scripts/dev-update/serve.mjs: dependency-free, range-capable static feed server.
• scripts/dev-update/build-pair.sh: builds a signed 2.0.0 feed and a runnable signed 1.0.0 app (SKIP_NOTARIZE=1).
• tests/e2e/fixtures/update.ts: ditto run-copy so a retry restarts from 1.0.0, plist version read, pkill helpers.
• tests/e2e/tests/update.spec.ts: phase 1 drives download then install via the main-process hook; phase 2 waits for the on-disk bundle to swap, kills the auto-relaunched instance, then fresh-launches and asserts app.getVersion() is 2.0.0.

CI: a dedicated code-update-e2e.yml, macOS only, reusing the existing build-macos steps and Apple signing secrets. It is not on the PR gate (nightly + workflow_dispatch); a temporary push trigger on this branch lets it run on the PR and is meant to be removed before merge.

The swap is signature-gated (Squirrel.Mac matches the designated requirement), which is why this runs in CI where the signing cert is present rather than as a local unit test. Notarization is skipped on purpose: it is a Gatekeeper concern, not what the in-place update verifies.

How did you test this?

Ran locally:

• pnpm --filter code typecheck passes with the hooks.
• Biome clean on all new and changed files.
• node --check serve.mjs and bash -n build-pair.sh pass.
• The e2e tsconfig typechecks the two new e2e files (the only errors are pre-existing in smoke.spec.ts).

Not run locally: the full signed build plus end-to-end swap (two signed mac builds, needs the Developer ID cert). That runs in the new workflow on this branch via the temporary push trigger, which is the real validation.

Automatic notifications

• Publish to changelog?
• Alert Sales and Marketing teams?

[github-actions]

React Doctor found no issues in the changed files. 🎉

_{Reviewed by React Doctor for commit 460340d.}

[greptile-apps]

Prompt To Fix All With AI

Fix the following 3 code review issues. Work through them one at a time, proposing concise fixes.
---### Issue 1 of 3
.github/workflows/code-update-e2e.yml:10-14
**Temporary push trigger not removed before merge**
The PR description explicitly states this push trigger "is meant to be removed before merge." It's still present. If the source branch (`test/macos-auto-update-e2e`) continues to exist in the repo after the merge, any future push to it will fire this workflow — which runs two full signed macOS builds and exercises the real install. That's expensive (and confusing) for what should be a nightly-only job.
### Issue 2 of 3
apps/code/tests/e2e/tests/update.spec.ts:63-130
**`app` and `updated` not closed in finally**
If `pollStatus` for check/download times out (phase 1 fails), the `app` Electron process is left running — only `feed.kill()` is in the `finally` block. Similarly, if the `expect(version).toBe(NEW_VERSION)` assertion at line 126 fails, `updated` is leaked. Playwright does not automatically close `electron.launch()` instances on test failure. Adding `app.close().catch(() => {})` / `updated.close().catch(() => {})` to the `finally` block (tracked via `let app` / `let updated` in outer scope) would prevent stale processes from interfering with a retry.
### Issue 3 of 3
apps/code/scripts/dev-update/serve.mjs:50-52
**Suffix-range requests (`bytes=-N`) compute wrong offsets**
When the range header is `bytes=-500` (last 500 bytes), the regex captures `match[1] = ""` and `match[2] = "500"`. Because `""` is falsy, `start` is set to `0` instead of `stat.size - 500`, so the response sends the wrong slice of the file. The electron-updater differential/blockmap code can emit suffix ranges during verification. In practice `electron-updater` on this path likely uses open-ended ranges (`bytes=0-`), but a defensive fix would be: `const start = match?.[1] ? Number(match[1]) : match?.[2] ? stat.size - Number(match[2]) : 0`.

_{Reviews (1): Last reviewed commit: "add macos auto-update e2e harness and ci" | Re-trigger Greptile}

[charlesvien]

• test(updates): macOS auto-update E2E test pipeline with real install + relaunch #2828 👈 (View in Graphite)
• feat(updates): update modal, auto-download setting, changelog #2818
• chore(code): upgrade Electron 42, React 19.2, Vite 7 and deps #2817
• refactor(code): migrate build and dev to electron-vite #2811
• refactor(code): migrate from electron forge to electron-builder #2795
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[stamphog]

Gates denied because this PR adds a new CI/CD workflow file (.github/workflows/code-update-e2e.yml), which is on the infra/CI deny-list and requires human review. Additionally, there are unresolved bot comments flagging a push trigger that was explicitly noted for removal before merge but wasn't removed, and a functional bug in the range-request handling in serve.mjs.