The suite was intermittently failing, with different tests each time and runs stretching past two minutes. Two separate causes, neither of them the behaviour under test. State leaked between tests. The editor persists edits to disk, so any test that typed into a note permanently changed the vault for every test after it — which made results depend on execution order and read as timing flakiness. Chasing it as a timing problem produced nonsense: the same scenario appeared to fail at 0ms, then only at 120ms, then not at all, because each probe was measuring accumulated file corruption rather than typing speed. `resetWorkspace` now restores the fixture notes as well as clearing the persisted tab layout. Enter was pressed inside CodeMirror's `interactionDelay`. `acceptCompletion` deliberately refuses a completion whose popup opened less than 75ms ago, so a fast typist cannot accept something they never saw. Synthetic input is quick enough to land inside that window; a human is not. The test now waits past it. Specs also share a single Electron instance rather than launching per test. On macOS a quitting Electron regularly gets stuck mid-exit — `ps` shows `UNEs`, reparented to launchd, where not even SIGKILL reaches it — and those hung processes accumulate until launches slow enough to time out. Four launches per suite instead of fourteen keeps that rare; it is a mitigation, not a cure. Three consecutive full runs: 14/14 in ~8s, down from 2.4m with failures. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>