The babysitting business kit has a browser test file that drives the app the way a sitter would: load the example, add a lead, book a sit, record a payment, reload and check it all stayed. It also checks the keyboard tabs, three guide pages and the offline service worker. Seven tests, 4.9 seconds, the bulk of the kit's 6.6-second suite.
The browser and the little file server were already shared. What the spec reporter showed was that the tests ran one after another, even though each opens its own browser context, with its own storage and service worker. They have nothing to wait for from each other.
Node's test runner runs tests in sequence unless you say otherwise. One line says otherwise: put them inside describe(..., { concurrency: true }). Now the file takes as long as its longest test plus start-up: 2.9 seconds, median of five runs, against 4.9 before. All 333 tests in the kit still pass, and a test we broke on purpose still fails the run.
It is two seconds. It matters because agent sessions run this suite whenever they touch the kit, and the fix had no cost: the tests were already isolated, only the runner didn't know. The measurement and the next candidates are in docs/history/ci-perf-2026-11-16-babysitting-e2e-concurrent.md.