Drive RepositoriesFeature delays with an injected clock

[onevcat]

Problem

RepositoriesFeatureTests/worktreeInfoEventRepositoryPullRequestRefreshQueuesWhileAvailabilityUnknown
fails intermittently in CI with:

Must handle 2 received actions before sending an action.
Unhandled actions: [refreshGithubIntegrationAvailability, githubIntegrationAvailabilityUpdated(false)]

It passes locally and in isolation (10/10), so it read as noise — but it is a real
non-deterministic-clock defect, not flakiness.

Root cause

The GitHub-integration recovery loop sleeps on a real ContinuousClock() and re-emits
refreshGithubIntegrationAvailability every githubIntegrationRecoveryInterval (15s). A TestStore
cannot control a real clock, so under the full parallel CI run a test's window between starting the
recovery loop (on githubIntegrationAvailabilityUpdated(false)) and cancelling it (on
setGithubIntegrationEnabled(false)) can exceed 15s of wall-clock — the real clock then fires an
extra refreshGithubIntegrationAvailability → githubIntegrationAvailabilityUpdated pair right as
the test sends its next action, exactly the two unhandled actions above.

Fix

Inject @Dependency(\.continuousClock) and sleep on it in the recovery loop. The two tests that
reach the recovery path now provide a TestClock that is never advanced, so the loop suspends and
never fires during the test — deterministic regardless of CI load, aligning with the repo rule
"in unit tests, never use Task.sleep; use TestClock."

The one-shot delayed PR refresh (2s) and toast auto-dismiss (3s) keep their original real clock:
they are not the flake, and injecting them would force every test embedding RepositoriesFeature
that triggers a toast or delayed refresh (e.g. AppFeature*Tests) to supply a clock. Scoping the
change to the recovery loop keeps the blast radius to exactly the two recovery-path tests.

Verification

make check, make test (zero failures), make build-app. RepositoriesFeatureTests: 251 passed;
the three AppFeature* toast tests that a broad injection had broken pass unchanged. The
previously-flaky test is now deterministic (the recovery effect cannot fire without a clock advance).

https://claude.ai/code/session_01AQ4X2bx8DnU74wwThhV9D8