Story tests flake: a random story file fails with "Cannot connect to the iframe"

New data point from PR #59: on the retry, all 164 stories passed, then timeout 360 killed the process (exit 124). So part of the "hang" is vitest not exiting after a green run, probably because the browser provider or a Storybook/Vite server handle is left open, not a test that never finishes.

Things to try:

• teardownTimeout
• closing the Playwright browser in a global teardown
• running with --pool=threads
• checking for a still-open dev server

Also note: CI runners aren't the only constrained resource. The agent host is a production k8s node, so local test:stories runs should be rare.

#64 is merged. It runs story files one at a time, sets a root teardownTimeout, allows 10 minutes per attempt and retries once.

Running files one at a time does not fix the flake. On #64's own CI, the first attempt still failed one file (EnvBadge: "Failed to fetch dynamically imported module"), then the retry passed. An attempt takes about a minute.

The files that fail are usually ones added recently (EnvBadge, Pipeline, Toggle, Vote, Skeleton). That points at Vite optimising dependencies when a file's module graph is first requested, rather than at running files in parallel.

Next things to try:

• Pre-bundle the design-system dependencies with optimizeDeps.include.
• Run vite optimize (a warmup pass) before the story run.
• Pin @vitest/browser and Storybook versions in case it's a regression.

The retry keeps CI usable in the meantime.

More data:

• Local vs CI. Locally (CI=true, nice'd, a production-node host), all 50 story files pass on the issue-detail branch. On CI the same branch fails.
• Where it hangs. The first attempt froze on the next file right after Colour.stories until the 10-minute timeout. Vite logged [optimizer] bundling dependencies... but never scanning.
• The retry. It failed StatusIcon.stories with "Failed to fetch dynamically imported module" plus "Cannot connect to the iframe", and stopped after 35 of 50 files on an unhandled error.

So this is a slow-runner race in Vite dependency optimisation, or in the browser iframe. It isn't a test bug, and running files serially doesn't help.

Interim merge policy: a story-test CI failure counts as non-blocking only when all of these hold:

1. every test that ran passed;
2. the failure is the iframe, dynamic-import or timeout symptom;
3. the PR's stories pass locally.

Next fix to try: pre-warm the optimiser cache before the run (run a vitest browser pass on one tiny story, then the full run), or pin optimizeDeps.include to the exact dependency list Vite reports after a local run.