fix(tests): make the event buses register synchronously and kill the App-target flake

[Adron]

Summary

Closes #82. The intermittent, unattributable "1 failure" in the App target had two causes, and only one of them was a test-hygiene problem.

The flake is now reproduced, named, and fixed — with the machinery to name the next one.

1. The real race was in production code

All four feature event buses registered their subscriber inside a Task:

return AsyncStream { continuation in
    Task { await self.storage.register(id: id, continuation: continuation) }
}

So events() returned a stream that was not yet in the subscriber table. A caller that subscribed and immediately wrote could miss its own event, and no number of Task.yield()s closed the window — the wait was on actor scheduling, not on cooperative yields.

Two tests papered over this with a fixed 10 ms sleep commented "give the subscription a beat to register", then awaited a 1-second expectation. When the beat was not enough, the result was a 1 s timeout — exactly the "one unnamed failure that does not reproduce" signature in the issue.

This was not only a test problem. In the running app, a view that subscribes in .task { } and immediately triggers a write can drop its own event — the unread badge that occasionally does not move.

The four identical private actors are now one EventBusStorage<Event> that registers synchronously under a Mutex, inside the AsyncStream build closure (which AsyncStream invokes synchronously during init). post broadcasts synchronously too, and subscriberCount lets a test assert a subscription is live rather than wait for one.

2. Fixed wall-clock waits, and the yield-count waits the issue's table missed

AppTests/Support/AsyncSettle.swift carries one idiom:

Helper	For
settle(until:)	poll for a post-condition; XCTFail at the caller's line when it expires
settleQuiet(for:)	the deliberately short bounded wait, for assertions that something did not happen
settledValue(of:)	wait for a counter to stop moving before taking a baseline

settledValue exists for test_givenRunningPoll_whenStopped_thenNoFurtherUpdatesLand, which sampled the poll count the instant after stopPolling(): a poll already in flight still records, so the old version failed a correctly-cancelled loop.

CurrentUserStoreTests already had a private settle(until:) and TagCompletionViewModelTests already had the G20 fix; both now point at the shared one.

The issue's nine-file table was incomplete — it counted Task.sleep, and the actual flake spent Task.yield() instead. Swept for that class too; see below.

3. Name the test first — and it paid immediately

• scripts/gate.sh — the whole E2E gate in one command, -resultBundlePath per App-target run, and an xcresulttool pass that prints the failing test identifiers. scripts/gate.sh app 20 is flake-hunt mode.
• .github/workflows/ci.yml — same capture, a ::error:: annotation naming the failing tests, and the bundle uploaded as an artefact on failure.
• The gate's Decision 0003 grep is anchored to column 0. The unanchored form in the checklist matches the prose in file headers that say a file does not import the kit, and reports a violation on a compliant tree.

The very first run with capture named a failure (a defect in one of the new tests: _ = bus.events() drops the stream, whose deinit correctly unregisters — now asserted in both directions).

The flake, reproduced and named

$ xcodebuild test … -test-iterations 20 -run-tests-until-failure
DocumentEditorViewModelTests/test_givenBodyEdit_whenDebounceElapses_thenCallsUpdate()
    DocumentEditorViewModelTests.swift:59: XCTAssertFalse failed

waitForSaveCompletion yielded eight times and then fell through silently, so a save that had not finished failed the caller's assertion rather than reporting that the wait expired. The same command now runs 20/20 clean.

The other instance of the class was FollowRequestRowViewModelTests, whose five yields were waiting for the bus registration that is now synchronous — replaced by an assertion that the subscription is live.

Deliberately not in this PR

The four sleeps in InterlinedPersistence's outbox tests are a different animal: they exist to force distinct Date() timestamps, because outboxEntries() sorts by enqueuedAt alone, which is not a total order. That is a FIFO-ordering defect in a document-sync queue, not a test race, and fixing it means a SwiftData schema change. Filed as #84 rather than smuggled into a test-hygiene PR.

FileLogTests:73 and SwiftDataDocumentStoreTests are untouched for the opposite reason: their sleep calls are polling ticks inside correct wait loops.

Verification

• xcodebuild build → ** BUILD SUCCEEDED **
• xcodebuild test (App) → Executed 975 tests, with 0 failures · ** TEST SUCCEEDED ** (was 968)
• -test-iterations 20 -run-tests-until-failure → 20/20 clean, exit 0 (previously stopped at iteration 3)
• swift test InterlinedDomain → Executed 1012 tests, with 0 failures
• swift test InterlinedPersistence → Executed 140 tests, with 0 failures
• swift test InterlinedKit → Executed 483 tests, with 6 failures — ⚠️ all six are the env-gated live ContractTests, failing with "Too many attempts. Please try again later." The account was rate-limited by unrelated API recon running in the same session; they are not regressions and pass once the limit clears.
• Decision 0003 (anchored) → zero hits

New tests: EventBusStorageTests — 7 cases, each written so it would fail, not merely flake, against the old asynchronous registration.

Acceptance mapping

Criterion	Status
The gate names the failing test when it fails	✅ scripts/gate.sh + CI annotation + uploaded bundle
No App-target test depends on a fixed wall-clock sleep	✅ zero Task.sleep in AppTests/ outside the polling ticks in AsyncSettle
20 consecutive clean App-target runs	✅ 20/20

🤖 Generated with Claude Code

https://claude.ai/code/session_016gSWb3scYobtxLJioV1qF9