web-testing

First reproduce the flake with `--repeat-each=20 --fully-parallel`, then classify it as timing, isolation, environment, or infrastructure. Fix timing with retrying web-first assertions, isolation with per-worker state or `test.lock`, and add `--fail-on-flaky-tests` so retries cannot hide future flakes.

Create a setup project matched by `testMatch` that logs in once and calls `context.storageState({ path })`, then have real projects declare `dependencies: ['setup']` and `use: { storageState }`. Note that storageState captures cookies and localStorage only, not sessionStorage.

A fixed wait runs the assertion at a wall-clock offset, so it fails when the app is slower than the guess and stops catching regressions when the app is faster. Replace it with a retrying assertion like `expect(locator).toHaveText(...)`, which polls until the condition holds or times out.

Only with `fullyParallel: true`. Without it, `--shard` distributes whole files, so shards can come back wildly uneven, including empty shards that finish green and prove nothing. Enable fullyParallel first, then shard, and merge per-shard blob reports with `merge-reports`.

The browser cache key has no version component, so it restores binaries built for an older `@playwright/test` and `playwright install` treats the cache as a hit. Put the resolved Playwright version in the cache key, or use the version-tagged official container image.

Reserve end-to-end tests for critical journeys that cross tiers like sign-in or checkout. Unit logic, API contracts, edge-case permutations, load testing, and native mobile UI automation belong to other tools and should not be covered by a browser suite.