diff --git a/fleetd/src/test/java/dev/ltms/fleet/FleetdAssemblyFleetAppTest.java b/fleetd/src/test/java/dev/ltms/fleet/FleetdAssemblyFleetAppTest.java
index 36ca40d..52e49ba 100644
--- a/fleetd/src/test/java/dev/ltms/fleet/FleetdAssemblyFleetAppTest.java
+++ b/fleetd/src/test/java/dev/ltms/fleet/FleetdAssemblyFleetAppTest.java
@@ -8,6 +8,7 @@ import dev.ltms.fleet.herdr.HerdrClient;
 import io.javalin.Javalin;
 import org.junit.jupiter.api.AfterEach;
 import org.junit.jupiter.api.Test;
+import org.junit.jupiter.api.Timeout;
 import org.junit.jupiter.api.io.TempDir;
 
 import java.net.URI;
@@ -71,7 +72,23 @@ import static org.junit.jupiter.api.Assertions.assertEquals;
  * below is what this class relies on for CB-185's {@code FleetApp} half; {@code
  * FleetAppTwoDaemonTest} remains the full behavioural proof that {@code FleetApp} itself merges
  * {@code /sessions} correctly once handed two clients.
+ *
+ * <p><strong>fleetd #629 follow-up.</strong> The fix below (see {@link TwoHerdrResourcePorts})
+ * makes {@link #healthzGoesRedWhenTheLeadDaemonIsDownEvenThoughTheMemberIsUp}'s fake {@code
+ * nanoClock()} frozen unless {@code herdrPollWait()} itself advances it. That is a sharper pin
+ * than an assertion — if a future edit to {@code FleetdAssembly} ever bypasses {@code
+ * ports.herdrPollWait()} again (e.g. reverting to a hardcoded {@code Thread.sleep}), the clock
+ * never advances, {@code Fleetd#awaitHerdr}'s deadline is never reached, and this test hangs
+ * forever instead of failing — proven by deliberately reintroducing that exact regression while
+ * fixing this ticket. {@code @Timeout} turns that silent hang into a bounded, named test failure:
+ * {@code SEPARATE_THREAD} so JUnit's timeout governor can actually interrupt a thread stuck in a
+ * real {@code Thread.sleep} loop (the default {@code SAME_THREAD} mode cannot — it only measures
+ * elapsed time after the test method returns on its own, which never happens here). 10 seconds is
+ * roughly 150x the real passing times measured here (~0.06s), so a slow CI machine has no reason
+ * to flake, and it is still 3x faster than discovering the regression by burning a CI job's whole
+ * wall-clock budget.
  */
+@Timeout(value = 10, unit = TimeUnit.SECONDS, threadMode = Timeout.ThreadMode.SEPARATE_THREAD)
 class FleetdAssemblyFleetAppTest {
 
     private static final class TwoHerdrResourcePorts implements ResourcePorts {