diff --git a/v2/crates/wifi-densepose-sensing-server/tests/mqtt_integration.rs b/v2/crates/wifi-densepose-sensing-server/tests/mqtt_integration.rs
index 50975617..4edb8a7c 100644
--- a/v2/crates/wifi-densepose-sensing-server/tests/mqtt_integration.rs
+++ b/v2/crates/wifi-densepose-sensing-server/tests/mqtt_integration.rs
@@ -274,38 +274,40 @@ async fn state_messages_published_on_snapshot_broadcast() {
     let (tx, rx) = broadcast::channel::<VitalsSnapshot>(32);
     let _handle = spawn(cfg, builder, rx);
 
-    // Wait long enough for the publisher to:
-    //   (a) connect to the broker (rumqttc connects on first publish)
-    //   (b) complete the initial 21+ QoS-1 discovery publishes
-    //   (c) reach its select! and begin draining state_rx
-    // 3s is well past what we measured locally for the full ramp-up
-    // (median ~800ms on a fast loopback; doubled for CI safety).
-    tokio::time::sleep(Duration::from_secs(3)).await;
+    // Iter 46 — instead of front-loading 6 snapshots and hoping the
+    // publisher's startup beats them, drive snapshots in a background
+    // task THROUGHOUT the capture window. CI runners can be slow to
+    // boot the publisher (mosquitto sidecar + cold cargo cache + slow
+    // QoS-1 discovery publishes), so a "publisher must be ready by
+    // t=3s" assumption is fragile. Steady-state ON/OFF traffic for the
+    // full 14 s window guarantees both states appear in the capture
+    // even if the first 3-5 s of publishes are missed.
+    let tx_bg = tx.clone();
+    let drive = tokio::spawn(async move {
+        // Brief warm-up before first publish.
+        tokio::time::sleep(Duration::from_secs(1)).await;
+        for i in 0..40 {
+            let _ = tx_bg.send(VitalsSnapshot {
+                node_id: "inttest3".into(),
+                timestamp_ms: 1779_512_400_000 + (i as i64) * 300,
+                presence: i % 2 == 0,
+                fall_detected: false,
+                motion: if i % 2 == 0 { 0.40 } else { 0.02 },
+                motion_energy: 800.0,
+                presence_score: if i % 2 == 0 { 0.95 } else { 0.10 },
+                breathing_rate_bpm: Some(14.0),
+                heartrate_bpm: Some(72.0),
+                n_persons: if i % 2 == 0 { 1 } else { 0 },
+                rssi_dbm: Some(-48.0),
+                vital_confidence: 0.9,
+            });
+            tokio::time::sleep(Duration::from_millis(300)).await;
+        }
+    });
 
-    // Fire snapshots repeatedly so a single dropped broadcast doesn't
-    // tank the test. Each tx.send is fanout to ALL receivers, so the
-    // publisher receives every one.
-    for i in 0..6 {
-        let _ = tx.send(VitalsSnapshot {
-            node_id: "inttest3".into(),
-            timestamp_ms: 1779_512_400_000 + (i as i64) * 100,
-            presence: i % 2 == 0,
-            fall_detected: false,
-            motion: if i % 2 == 0 { 0.40 } else { 0.02 },
-            motion_energy: 800.0,
-            presence_score: if i % 2 == 0 { 0.95 } else { 0.10 },
-            breathing_rate_bpm: Some(14.0),
-            heartrate_bpm: Some(72.0),
-            n_persons: if i % 2 == 0 { 1 } else { 0 },
-            rssi_dbm: Some(-48.0),
-            vital_confidence: 0.9,
-        });
-        tokio::time::sleep(Duration::from_millis(200)).await;
-    }
-
-    // Capture window — generous so we don't race the publisher's
-    // change-detection on the presence binary_sensor.
-    let msgs = collect_published(&mut sub_loop, Duration::from_secs(8)).await;
+    // 14 s window covers warm-up + 12 s of steady-state ON/OFF traffic.
+    let msgs = collect_published(&mut sub_loop, Duration::from_secs(14)).await;
+    drive.abort();
     let _ = sub.disconnect().await;
 
     // Diagnostic: dump every captured topic so we can see what (if