diff --git a/frontend/src/components/Flow/useFlowSocket.ts b/frontend/src/components/Flow/useFlowSocket.ts
index 5519d59..c0cde04 100644
--- a/frontend/src/components/Flow/useFlowSocket.ts
+++ b/frontend/src/components/Flow/useFlowSocket.ts
@@ -25,6 +25,15 @@ const RECONNECT_MAX = 30000
  * handler instead of retrying forever.
  */
 const AUTH_ATTEMPTS = 3
+/**
+ * How long the engine holds a node down before it surfaces as an issue.
+ *
+ * ponytail: a hand copy of `HEALTH_FLOOR` in
+ * `backend/fluksio/flow/controller.py:122`, which drifts silently if that one
+ * moves — nothing breaks, the border just turns early or late. Carrying the
+ * floor in the `node_health` payload is the upgrade path.
+ */
+const HEALTH_FLOOR = 60_000
 
 type FlowEvent =
   | {
@@ -152,6 +161,20 @@ let watchers = 0
 let rejected = 0
 let client: QueryClient | null = null
 const authHandlers = new Set<() => void>()
+/** Floor re-reads still owed, per node, so a recovery can cancel its own. */
+const healthTimers = new Map<string, ReturnType<typeof setTimeout>>()
+
+/** Everything a node's health shows up in, which is two unrelated keys. */
+function invalidateHealth(flow: string | undefined) {
+  // `graph` is a sibling of the flow's own key rather than a child, so the
+  // detail invalidation alone never reaches Home's fault rings.
+  for (const queryKey of [
+    flow ? flowKeys.detail(flow) : flowKeys.all,
+    flowKeys.graph,
+  ]) {
+    client?.invalidateQueries({ queryKey })
+  }
+}
 
 function schedule() {
   if (timer) return
@@ -258,15 +281,28 @@ function connect() {
           ts: message.ts ?? Date.now() / 1000,
         })
         break
-      case "node_health":
+      case "node_health": {
         // The canvas draws node health from the flow detail's `issues`, which
         // the server derives — so the screen only moved on mount, navigation
         // or a rebuild, never when health actually flipped. The store had a
         // health map of its own and nothing ever read it.
-        client?.invalidateQueries({
-          queryKey: message.flow ? flowKeys.detail(message.flow) : flowKeys.all,
-        })
+        invalidateHealth(message.flow)
         if (message.health === "down") {
+          // This fires at the transition, while the floor is still hiding the
+          // node, and the engine publishes nothing when the floor expires — so
+          // read again ourselves at that moment, or the issue waits for an
+          // unrelated invalidation. Only the first `down` schedules it: the
+          // engine restarts its own clock on a transition, not on a connector
+          // rotating its error string while staying down.
+          if (!healthTimers.has(message.node)) {
+            healthTimers.set(
+              message.node,
+              setTimeout(() => {
+                healthTimers.delete(message.node)
+                invalidateHealth(message.flow)
+              }, HEALTH_FLOOR),
+            )
+          }
           liveStore.recordEngineEvent({
             type: message.type,
             flow: message.flow,
@@ -274,8 +310,14 @@ function connect() {
             detail: message.detail ?? "Reported itself down.",
             ts: message.ts ?? Date.now() / 1000,
           })
+        } else {
+          // Recovery clears at once server-side, so nothing is owed.
+          const owed = healthTimers.get(message.node)
+          if (owed) clearTimeout(owed)
+          healthTimers.delete(message.node)
         }
         break
+      }
       case "flow_quarantined":
       case "engine_degraded":
       case "engine_fatal":
@@ -404,6 +446,8 @@ function release(onAuthFailure?: () => void) {
     clearTimeout(timer)
     timer = null
   }
+  for (const owed of healthTimers.values()) clearTimeout(owed)
+  healthTimers.clear()
   const ws = socket
   socket = null
   ws?.close()