Why

DeviceLinkRepositoryImplTest.reconcilePrunesShortCodesNoLongerInCatalog fails intermittently in CI. It is a test-harness flake, not a product bug — the repository logic is correct; the test mixed runTest's virtual clock with Room's real dispatcher.

Root cause

reconcile() guards its network fetch with withTimeoutOrNull(5s), whose deadline is measured against the calling coroutine's clock. The test ran on runTest + UnconfinedTestDispatcher (virtual time), while Room executes its DB work on the real ioDispatcher — which runTest neither owns nor awaits. Each time the coroutine parked on a real IO thread, the test scheduler went idle and runTest fast-forwarded virtual time; under load it jumped past the 5s budget, so the fetch was reported as timed out, store() was skipped, and the cache kept the pre-prune rows [a, b] instead of [a].

This is purely a test artifact: in production the timeout runs on the wall clock against a real network call, so it never misfires.

The previously-added multiConnection = false for in-memory DBs was a separate — and on Room 3.0.0-rc01, moot — mitigation: Room already forces a single connection for in-memory databases (name == null) regardless of the pool config, so it was never what made or broke this test.

Changes

🐛 Bug Fixes

• Run DeviceLinkRepositoryImplTest on real dispatchers + runBlocking instead of UnconfinedTestDispatcher + runTest, so reconcile()'s withTimeoutOrNull is measured on the wall clock (matching production) and runTest's virtual-clock fast-forward can no longer spuriously expire it.

🧹 Chores

• Correct the configureCommon KDoc: clarify that Room 3 always serves in-memory databases from a single connection regardless of the pool config, and that this test's determinism comes from the wall clock, not from multiConnection.

Testing Performed

• The flake only surfaces under concurrent IO-pool contention (like CI), so a temporary 24-thread stress harness was used to reproduce it. On main it failed ~0.3% of runs, and the failure count exactly matched the network refresh timed out log count — confirming the timeout-misfire mechanism. With this change, ~18,000 concurrent + ~15,000 sequential runs completed with zero failures, timedOut=0. The harness was removed before this PR.
• Full baseline: ./gradlew spotlessApply spotlessCheck detekt assembleDebug test allTests — BUILD SUCCESSFUL; all 6 DeviceLinkRepositoryImplTest methods pass.