[Ekwav]

CI ran the whole suite three times (about 27 minutes for a passing job), so one flaky spec failed unrelated PRs such as #1567. This fixes the flaky specs at their cause, without retries, and repeats only the specs a change touches.

Flaky specs

• reduced-motion: this was an app bug. OfflineBanner only listened for offline events, so a connection lost before hydration was never shown. It now also reads navigator.onLine on mount, and the test goes offline before the page loads.
• search: a flag set by an intercept was checked once; it is now checked with a retrying should.
• flip-tracking-recovery: failed during the first days of each month, because the calendar shows the month of today − 7 days and today's date wasn't on it. The test now opens the next month in that case.
• dom-translation: waits for the held history request before checking the translated text.
• bazaar-snapshot-recovery: rewritten on cy.clock (new cypress/support/clock.ts). The old retry test finished in 0.9 s, so it never actually exercised the 3 s retry; now it does.

Speed: the premium polling and slot-assignment tests run on a faked clock instead of real waits. Three tests were merged into two, with the same coverage; the suite has 280 tests. One pass takes about 4:35 instead of 6:47.

Gate: scripts/verify-cypress.mjs first runs the changed specs, and specs importing a changed support module, two extra times. Then it runs the suite once. The base is the merge commit's first parent for PRs and github.event.before for pushes; manual runs get the full pass only. A passing job is expected to take about 9–13 minutes.

Verified locally: a production test build passed the new gate 3 times in a row (each run: the burn-in specs twice, then 280/280). The five formerly flaky specs passed 10/10 each in fresh processes.

This replaces #1563 and #1564.

🤖 Generated with Claude Code