ci: skip puppeteer postinstall download to avoid Chrome race on Node 24 (#912)

## What

Set `PUPPETEER_SKIP_DOWNLOAD=1` on the `pnpm install` step in the `Build
& Test` matrix, so puppeteer's postinstall doesn't try to download
Chrome. The dedicated `Install Chrome for puppeteer` step (already in
the workflow) keeps doing the download synchronously and reliably.

## Why

Node 24 jobs have been failing intermittently with:

```
BrowserLaunchError: Failed to launch browser
Caused by: Could not find Chrome (ver. 146.0.7680.153).
```

Step timing shows the `Install Chrome for puppeteer` step has been
silently no-opping on Node 24 — exiting in ~1 s with no output, vs ~4 s
with a `chrome@146.0.7680.153 …` line on a healthy run.

### Root cause

Puppeteer's postinstall hook
([`install.mjs`](https://github.com/puppeteer/puppeteer/blob/puppeteer-v24.40.0/packages/puppeteer/install.mjs))
is **fire-and-forget**:

```js
const {downloadBrowsers} = await importInstaller();
downloadBrowsers();   // ← not awaited
```

Inside, `downloadBrowsers()` kicks off downloads for chrome,
chrome-headless-shell, and firefox in parallel. When `pnpm install`
returns before they finish, the process is reaped mid-download and
`~/.cache/puppeteer/chrome/linux-146.0.7680.153/` is left
half-populated.

The next step then runs `pnpm exec puppeteer browsers install chrome`,
which hands off to `@puppeteer/browsers`. That library's "already
installed" check is purely "does this directory exist with the right
buildId?" — so it sees the half-written directory, assumes Chrome is
installed, and exits silently. Tests then explode at runtime because the
binary isn't actually there.

That explains everything we observed:
- **Only Node 24** — probably just timing; Node 24's faster startup
loses the race more often.
- **Intermittent** — it's a literal race between `pnpm install`
returning and the download completing.
- **Retries sometimes work** — when the download happens to finish in
time, no orphaned directory, all green.
- **Has been a slow-burn issue** — but the Renovate/TS-v6 lockfile bumps
that invalidate the pnpm-store cache made the race lose more often.

### Fix

`PUPPETEER_SKIP_DOWNLOAD=1` on the install step makes puppeteer's
postinstall log "Skipping downloading browsers as instructed" and return
immediately — no racy background download, no orphaned cache directory.
The explicit step that follows then performs the download synchronously
and reliably.

Credit to @barjin for [suggesting this
approach](#912 (comment))
— it's much cleaner than the `rm -rf ~/.cache/puppeteer` workaround in
the first commit, which only mopped up after the race instead of
preventing it.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: B4nan

.github/workflows/check.yaml

@@ -41,6 +41,16 @@ jobs:
 
             - name: Install pnpm and dependencies
               uses: apify/actions/pnpm-install@v1.1.2
+              env:
+                  # Puppeteer's postinstall kicks off the Chrome download in a
+                  # fire-and-forget Promise (downloadBrowsers() is not awaited
+                  # in puppeteer's install.mjs). When pnpm install returns
+                  # before the download finishes, ~/.cache/puppeteer is left
+                  # half-populated, and the subsequent `browsers install`
+                  # silently no-ops because the directory already exists.
+                  # Skip the postinstall download entirely and let the explicit
+                  # step below do it synchronously.
+                  PUPPETEER_SKIP_DOWNLOAD: 1
 
             - name: Install Chrome for puppeteer
               run: pnpm exec puppeteer browsers install chrome