The end-to-end CI shards fail when a runner is slow to install the browser or ffmpeg #838

What happens

An end-to-end shard fails before a single test runs when its hosted runner is slow to install the browser's system packages or ffmpeg. The tests are fine; the install steps are not.

The first e2e-full run on big-release (run 39827053220, commit 6c3cf58) shows it. Shard 2 finished in 7 minutes. Shards 1 and 3 were cancelled by the 30 minute job limit:

The job log shows where the time went: apt fetched 20.9 MB from azure.archive.ubuntu.com at 29.2 kB/s ("Fetched 20.9 MB in 11min 56s"), with several Ign: lines where apt gave up on a package and fetched it again. The Chromium download itself, from Playwright's CDN, took 9 seconds. The runner image was ubuntu-24.04, version 20261012.334.1.

The same stall hit other runs the same week:

• CI run 39823430200 (PR ci(e2e): scoped e2e on staging PRs, full suite after each merge to big-release #836): shard 1 spent 15.5 minutes in the Playwright install, then was cancelled in the ffmpeg install.
• CI run 39792892907 (PR fix(vocal-split): say when Lead + Backing could not be made #834): shard 3 sat in the Playwright install for 2 hours on its first attempt.
• CI run 39827062156: shard 3, 11.5 minutes in the Playwright install, then cancelled in the ffmpeg install.

Every job also carries a notice that the ubuntu-latest label moves to Ubuntu 26 from October 19, 2026 (actions/runner-images#14748), so the environment under the suite can change between two runs of the same commit.

Who it affects

Anyone with a pull request open, and the full-suite run on every merge to big-release. A cancelled shard reads as a red PR, needs a manual re-run, and hides whether the change itself is sound.

What should happen

• A slow or hung install fails within minutes, says which step and why, and is retried, instead of burning the whole 30 minutes.
• Most runs do not depend on the Ubuntu mirror at all for things that rarely change.
• The environment under the suite does not change because a runner label moved.
• When a stall does happen, the run's summary shows which setup step took the time, without reading logs.

Constraints

• The e2e backend needs an ffmpeg with libmp3lame, aac, flac and libvorbis; the ogg export specs need Vorbis.
• No unverified binary may be downloaded or run: a download is pinned by checksum.
• Job names frontend-e2e (1) to (3), the spec selector and the scoped and full cadence stay as they are.
• No change to specs, app code, uv.lock or package files.
• Hosted runners only; the self-hosted runners are not involved here.