test: replace the remaining wall-clock bounds with time limits and ordering

[MarcoDotIO]

Follow-up to #22. With ~3,150 Swift Testing tests running in parallel on the macOS runner, the cooperative pool saturates and runs stall for 5–8 s (CI run 36644330107). Upper bounds on elapsed time fail spuriously under that load. This PR removes the bounds that #22 left and adds one-minute time limits. Each test still fails, instead of hanging, when the code it guards regresses.

Stacked on #22. This branch starts from test/event-driven-test-waits, so until #22 merges the diff also shows its two commits. The three commits below are this PR's.

Changes

Test	Old bound	Now
GatewayDeviceAuthOffActorTests.issuedTokenPersistenceNeverBlocksTheChannelActor	< 3 s	Ordering. The lock is released only after the actor answers, so the token can reach disk only if the actor answered while the write was pending. A write on the actor fails with SQLITE_BUSY after 30 s and the token never lands. A time limit alone would miss this, because 30 s < 1 min. A new DEBUG hook, _test_setDeviceTokenPersistenceStartedHandler, replaces the 200 ms sleep that waited for hello-ok to reach the write.
FileTransferNodeCommandsTests.file fetch refuses a FIFO promptly	< 2 s	Time limit. A blocking open(2) is a syscall, so the cancellation handler opens the FIFO's write end once to release it.
AsyncTimeoutRaceTests.operationThatIgnoresCancellationStillTimesOut	< 5 s	Time limit. The operation parks on a gate that the test opens on cancellation, instead of a never-resumed continuation that a loser-joining race would wait on forever.
ChatLinkPreviewTests.total deadline can fire before the session starts	< 1 s	Time limit. URLSession's own timeouts move to 3600 s, so only the fetcher's 0 s deadline can end the fetch. The fetch is awaited through AsyncTimeout.withTimeout(seconds: 0) because a deadline lost before start leaves it unresumed even on cancellation.
gatewayCoreWaitUntil (56 call sites, 8 suites)	10 s / 15 s	Polls until the condition holds or the test is cancelled. Every calling suite gets .timeLimit(.minutes(1)). On cancellation it records GatewayCoreWaitTimeout as an issue before throwing. Swift Testing drops errors thrown after a time-limit cancellation, and the issue says which wait hung.
ChatViewModelSessionActionTests wait helpers	15 s	Same approach. waitForForkStart awaits the gate stream directly, and the suite gets a time limit.

Verification

• swift test on macOS: 3,150 tests in 341 suites passed. Scripts/lint-swift.sh: 0 violations.
• Pool starvation (scratch copies that spawn activeProcessorCount * 2 blockers of Task.detached(priority: Task.currentPriority) { usleep(6_000_000) } at the critical point): all four new tests pass after about 12 s. The old AsyncTimeout and link-preview versions fail their bounds.
• Regressions (scratch source edits, reverted). Each new test fails on the time limit and the run ends at about 60 s without hanging:
  • device-auth write moved back onto the actor. Instrumented, the actor answered after 30.35 s and the token was never stored.
  • O_NONBLOCK removed from the file-fetch open.
  • AsyncTimeout rewritten as a task group that joins the loser.
  • link preview: the deadline never aborts; separately, an abort that lands before start is lost.
  • gatewayCoreWaitUntil { false } and eventually { false } end on the time limit. The gateway wait reports Timeout waiting for: <label>.

Remaining

waitUntil in TestAsyncHelpers.swift (15 s default, about 760 call sites) still uses a wall-clock deadline. It was out of scope here.

🤖 Generated with Claude Code