fix: do not fatally abort install on nm-online connectivity check

pkg/config/cos.go

@@ -588,14 +588,28 @@ func UpdateManagementInterfaceConfig(mgmtInterface Network, dnsNameServers []str
 			logrus.Error(err, string(output))
 			return err
 		}
-		// This next command waits up to 30 seconds to ensure there's
-		// a connection.  Without this, it's possible that a slow DHCP
-		// server won't return in time, and the installer will subsequently
-		// fail the check for a default route.
-		output, err = exec.Command("nm-online", "-x").CombinedOutput()
+		// Wait (up to the timeout) for NetworkManager to finish (re)starting the
+		// management network before continuing.
+		//
+		// Use "-s" (wait for startup to complete) instead of "-x": "-x" exits
+		// immediately and defeats the intended wait, and startup (interfaces
+		// finished activating) is the right signal rather than *global*
+		// connectivity, which an isolated management network may never reach.
+		//
+		// The 45s timeout gives a slow-rate LACP bond headroom to converge:
+		// lacp_rate=slow emits an LACPDU only every 30s, so a single missed
+		// interval can cost ~30s before the bond syncs.  Because "-s" returns as
+		// soon as startup completes, this budget only costs time on a genuinely
+		// slow or degraded link.
+		//
+		// The check is best-effort and must NOT be fatal.  On the
+		// already-installed first-boot path it runs before the RancherD config is
+		// written, so aborting the install here on a briefly-degraded link (e.g. a
+		// bond still re-converging) leaves the node with no role until manual
+		// repair.  See harvester/harvester#10885.
+		output, err = exec.Command("nm-online", "-s", "-t", "45").CombinedOutput()
 		if err != nil {
-			logrus.Error(err, string(output))
-			return err
+			logrus.Warnf("nm-online did not confirm the management network within the timeout; continuing anyway: %v: %s", err, string(output))
 		}
 	}