Revert V4 mapped to V6 not working on Darwin/Freebsd for V6 netbird

Restore the pre-#6626 WG watcher shape
(single EnableWgWatcher reading the baseline inside the goroutine, no
PrepareInitialHandshake, no ctx-recheck before onDisconnected) to test whether
#6626 causes the missed handshake progression and subsequent reconnect loop.

Kept the WGW-DIAG logs.

This reverts commit 06839a4731.

client/internal/peer/conn.go

@@ -803,17 +803,15 @@ func (conn *Conn) isConnectedOnAllWay() (status guard.ConnStatus) {
 }
 
 func (conn *Conn) enableWgWatcherIfNeeded(enabledTime time.Time) {
-	if !conn.wgWatcher.PrepareInitialHandshake() {
-		return
+	if !conn.wgWatcher.IsEnabled() {
+		wgWatcherCtx, wgWatcherCancel := context.WithCancel(conn.ctx)
+		conn.wgWatcherCancel = wgWatcherCancel
+		conn.wgWatcherWg.Add(1)
+		go func() {
+			defer conn.wgWatcherWg.Done()
+			conn.wgWatcher.EnableWgWatcher(wgWatcherCtx, enabledTime, conn.onWGDisconnected, conn.onWGHandshakeSuccess)
+		}()
 	}
-
-	wgWatcherCtx, wgWatcherCancel := context.WithCancel(conn.ctx)
-	conn.wgWatcherCancel = wgWatcherCancel
-	conn.wgWatcherWg.Add(1)
-	go func() {
-		defer conn.wgWatcherWg.Done()
-		conn.wgWatcher.EnableWgWatcher(wgWatcherCtx, enabledTime, conn.onWGDisconnected, conn.onWGHandshakeSuccess)
-	}()
 }
 
 func (conn *Conn) disableWgWatcherIfNeeded() {
@@ -932,18 +930,22 @@ func (conn *Conn) AgentVersionString() string {
 
 func (conn *Conn) presharedKey(remoteRosenpassKey []byte) *wgtypes.Key {
 	if conn.config.RosenpassConfig.PubKey == nil {
+		conn.Log.Warnf("PSK-DIAG: rosenpass off -> static PSK (present=%v)", conn.config.WgConfig.PreSharedKey != nil)
 		return conn.config.WgConfig.PreSharedKey
 	}
 
 	if remoteRosenpassKey == nil && conn.config.RosenpassConfig.PermissiveMode {
+		conn.Log.Warnf("PSK-DIAG: rosenpass permissive + no remote RP key -> static PSK bridge (present=%v)", conn.config.WgConfig.PreSharedKey != nil)
 		return conn.config.WgConfig.PreSharedKey
 	}
 
 	// If Rosenpass has already set a PSK for this peer, return nil to prevent
 	// UpdatePeer from overwriting the Rosenpass-managed key.
 	if conn.rosenpassInitializedPresharedKeyValidator != nil && conn.rosenpassInitializedPresharedKeyValidator(conn.config.Key) {
+		conn.Log.Warnf("PSK-DIAG: rosenpass initialized -> returning nil (keep existing on-wire PSK; NOT re-set)")
 		return nil
 	}
+	conn.Log.Warnf("PSK-DIAG: rosenpass strict, not yet initialized -> seeding PSK (staticPresent=%v)", conn.config.WgConfig.PreSharedKey != nil)
 
 	// Use NetBird PSK as the seed for Rosenpass. This same PSK is passed to
 	// Rosenpass as PeerConfig.PresharedKey, ensuring the derived post-quantum

client/internal/peer/endpoint.go

@@ -38,6 +38,8 @@ func (e *EndpointUpdater) ConfigureWGEndpoint(addr *net.UDPAddr, presharedKey *w
 	e.mu.Lock()
 	defer e.mu.Unlock()
 
+	e.log.Warnf("PSK-DIAG: ConfigureWGEndpoint endpoint=%s psk_present=%v", addr, presharedKey != nil)
+
 	if e.initiator {
 		e.log.Debugf("configure up WireGuard as initiator")
 		return e.configureAsInitiator(addr, presharedKey)
@@ -51,6 +53,8 @@ func (e *EndpointUpdater) SwitchWGEndpoint(addr *net.UDPAddr, presharedKey *wgty
 	e.mu.Lock()
 	defer e.mu.Unlock()
 
+	e.log.Warnf("PSK-DIAG: SwitchWGEndpoint endpoint=%s psk_present=%v", addr, presharedKey != nil)
+
 	// prevent to run new update while cancel the previous update
 	e.waitForCloseTheDelayedUpdate()
 
@@ -69,6 +73,8 @@ func (e *EndpointUpdater) RemoveEndpointAddress() error {
 	e.mu.Lock()
 	defer e.mu.Unlock()
 
+	e.log.Warnf("PSK-DIAG: RemoveEndpointAddress -> peer re-added with only PublicKey+AllowedIPs; on-wire PSK is dropped")
+
 	e.waitForCloseTheDelayedUpdate()
 	return e.wgConfig.WgInterface.RemoveEndpointAddress(e.wgConfig.RemoteKey)
 }

client/internal/peer/guard/guard.go

@@ -85,7 +85,11 @@ func (g *Guard) reconnectLoopWithRetry(ctx context.Context, callback func()) {
 	defer g.srWatcher.RemoveListener(srReconnectedChan)
 
 	ticker := g.initialTicker(ctx)
-	defer ticker.Stop()
+	defer func() {
+		// If backoff.Ticker.send is blocked, context.Done will not close the Ticker goroutine.
+		// We have to explicitly call Stop, even if we use backoff.WithContext.
+		ticker.Stop()
+	}()
 
 	tickerChannel := ticker.C
 

client/internal/peer/guard/guard_leak_test.go

@@ -0,0 +1,92 @@
+package guard
+
+import (
+	"context"
+	"runtime"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+
+	"github.com/netbirdio/netbird/client/internal/peer/ice"
+)
+
+func newTestGuard(status connStatusFunc) *Guard {
+	srw := NewSRWatcher(nil, nil, nil, ice.Config{})
+	return NewGuard(log.WithField("test", "guard"), status, 50*time.Millisecond, srw)
+}
+
+// countBackoffTickerGoroutines returns how many goroutines are currently sitting
+// in backoff/v4.(*Ticker).run (a ticker goroutine that has not exited).
+func countBackoffTickerGoroutines() int {
+	buf := make([]byte, 1<<25) // 32MB
+	n := runtime.Stack(buf, true)
+	return strings.Count(string(buf[:n]), "backoff/v4.(*Ticker).run")
+}
+
+// TestGuard_ReconnectTicker_NoGoroutineLeakOnShutdown reproduces a observed
+// leak: after a shutdown burst, ticker run/send goroutines stay parked
+// forever even though every reconnect loop has exited.
+func TestGuard_ReconnectTicker_NoGoroutineLeakOnShutdown(t *testing.T) {
+	before := countBackoffTickerGoroutines()
+
+	const peers = 6000
+	cancels := make([]context.CancelFunc, 0, peers)
+	var wg sync.WaitGroup
+
+	// A status check slower than the tick cadence. This models the real
+	// isConnectedOnAllWay/callback doing work: while the loop is busy in the
+	// handler, the ticker fires the next tick and parks in send(), because
+	// send() never selects on ctx.
+	slowStatus := func() ConnStatus {
+		time.Sleep(70 * time.Millisecond)
+		return ConnStatusConnected
+	}
+
+	for range peers {
+		g := newTestGuard(slowStatus)
+		ctx, cancel := context.WithCancel(context.Background())
+		cancels = append(cancels, cancel)
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			g.Start(ctx, func() {})
+		}()
+		// Force the live ticker to be a newReconnectTicker.
+		g.SetRelayedConnDisconnected()
+	}
+
+	// Let the replacement tickers get past their 800ms initial interval, so
+	// many are parked in send() waiting on the (slow) consumer when we tear
+	// everything down.
+	time.Sleep(1500 * time.Millisecond)
+
+	// Shutdown burst: cancel every peer at once, like engine teardown.
+	for _, c := range cancels {
+		c()
+	}
+
+	// Every reconnect loop must return
+	waitCh := make(chan struct{})
+	go func() { wg.Wait(); close(waitCh) }()
+	select {
+	case <-waitCh:
+	case <-time.After(30 * time.Second):
+		t.Fatal("not all reconnect loops returned after ctx cancel")
+	}
+
+	// Give any correctly-stopped ticker goroutines time to unwind.
+	for range 50 {
+		runtime.Gosched()
+		time.Sleep(10 * time.Millisecond)
+	}
+
+	leaked := countBackoffTickerGoroutines() - before
+	t.Logf("backoff Ticker.run goroutines still parked after teardown of %d peers: %d", peers, leaked)
+	if leaked > 0 {
+		t.Errorf("LEAK: %d backoff ticker goroutines parked after all reconnect loops exited "+
+			"(defer ticker.Stop() stops the initial ticker, not the live replacement)", leaked)
+	}
+}

client/internal/peer/wg_watcher.go

@@ -31,9 +31,7 @@ type WGWatcher struct {
 	stateDump     *stateDump
 
 	enabled   bool
-	muEnabled sync.Mutex
-	// initialHandshake is not thread-safe; never call PrepareInitialHandshake and EnableWgWatcher concurrently.
-	initialHandshake time.Time
+	muEnabled sync.RWMutex
 
 	resetCh chan struct{}
 }
@@ -48,38 +46,40 @@ func NewWGWatcher(log *log.Entry, wgIfaceStater WGInterfaceStater, peerKey strin
 	}
 }
 
-// PrepareInitialHandshake reserves the watcher and reads the peer's current WireGuard
-// handshake time. It must be called before the peer is (re)configured on the WireGuard
-// interface, so the captured baseline reflects the state prior to this connection attempt
-// instead of racing with that configuration. Returns ok=false if the watcher is already
-// running, in which case EnableWgWatcher must not be called.
-func (w *WGWatcher) PrepareInitialHandshake() (ok bool) {
+// EnableWgWatcher starts the WireGuard watcher. If it is already enabled, it will return immediately and do nothing.
+// The watcher runs until ctx is cancelled. Caller is responsible for context lifecycle management.
+// NOTE: reverted to the pre-#6626 shape for bisecting the NHN issue.
+func (w *WGWatcher) EnableWgWatcher(ctx context.Context, enabledTime time.Time, onDisconnectedFn func(), onHandshakeSuccessFn func(when time.Time)) {
 	w.muEnabled.Lock()
 	if w.enabled {
 		w.muEnabled.Unlock()
-		return false
+		return
 	}
 
 	w.log.Debugf("enable WireGuard watcher")
 	w.enabled = true
 	w.muEnabled.Unlock()
 
-	handshake, _ := w.wgState()
-	w.initialHandshake = handshake
-	return true
-}
+	initialHandshake, err := w.wgState()
+	if err != nil {
+		w.log.Warnf("failed to read initial wg stats: %v", err)
+	}
+	w.log.Warnf("PSK-DIAG: watcher baseline handshake=%v (zero=%v) [pre-6626 revert]", initialHandshake, initialHandshake.IsZero())
 
-// EnableWgWatcher runs the WireGuard watcher loop using the handshake baseline captured by
-// PrepareInitialHandshake. The watcher runs until ctx is cancelled. Caller is responsible
-// for context lifecycle management.
-func (w *WGWatcher) EnableWgWatcher(ctx context.Context, enabledTime time.Time, onDisconnectedFn func(), onHandshakeSuccessFn func(when time.Time)) {
-	w.periodicHandshakeCheck(ctx, onDisconnectedFn, onHandshakeSuccessFn, enabledTime, w.initialHandshake)
+	w.periodicHandshakeCheck(ctx, onDisconnectedFn, onHandshakeSuccessFn, enabledTime, initialHandshake)
 
 	w.muEnabled.Lock()
 	w.enabled = false
 	w.muEnabled.Unlock()
 }
 
+// IsEnabled returns true if the WireGuard watcher is currently enabled
+func (w *WGWatcher) IsEnabled() bool {
+	w.muEnabled.RLock()
+	defer w.muEnabled.RUnlock()
+	return w.enabled
+}
+
 // Reset signals the watcher that the WireGuard peer has been reset and a new
 // handshake is expected. This restarts the handshake timeout from scratch.
 func (w *WGWatcher) Reset() {
@@ -92,6 +92,7 @@ func (w *WGWatcher) Reset() {
 // wgStateCheck help to check the state of the WireGuard handshake and relay connection
 func (w *WGWatcher) periodicHandshakeCheck(ctx context.Context, onDisconnectedFn func(), onHandshakeSuccessFn func(when time.Time), enabledTime time.Time, initialHandshake time.Time) {
 	w.log.Infof("WireGuard watcher started")
+	w.log.Warnf("WGW-DIAG: watcher started id=%d baseline=%v (zero=%v) firstCheckIn=%v", enabledTime.UnixNano(), initialHandshake, initialHandshake.IsZero(), wgHandshakeOvertime)
 
 	timer := time.NewTimer(wgHandshakeOvertime)
 	defer timer.Stop()
@@ -101,18 +102,17 @@ func (w *WGWatcher) periodicHandshakeCheck(ctx context.Context, onDisconnectedFn
 	for {
 		select {
 		case <-timer.C:
+			w.log.Warnf("WGW-DIAG: check fire id=%d lastHandshake=%v", enabledTime.UnixNano(), lastHandshake)
 			handshake, ok := w.handshakeCheck(lastHandshake)
 			if !ok {
-				if ctx.Err() != nil {
-					return
-				}
+				w.log.Warnf("WGW-DIAG: check failed -> firing onDisconnected (TEARDOWN, pre-6626 no ctx-recheck) id=%d", enabledTime.UnixNano())
 				onDisconnectedFn()
 				return
 			}
 			if lastHandshake.IsZero() {
 				elapsed := calcElapsed(enabledTime, *handshake)
 				w.log.Infof("first wg handshake detected within: %.2fsec, (%s)", elapsed, handshake)
-				if onHandshakeSuccessFn != nil && ctx.Err() == nil {
+				if onHandshakeSuccessFn != nil {
 					onHandshakeSuccessFn(*handshake)
 				}
 			}
@@ -123,15 +123,15 @@ func (w *WGWatcher) periodicHandshakeCheck(ctx context.Context, onDisconnectedFn
 			timer.Reset(resetTime)
 			w.stateDump.WGcheckSuccess()
 
-			w.log.Debugf("WireGuard watcher reset timer: %v", resetTime)
+			w.log.Warnf("WGW-DIAG: check ok id=%d handshake=%v nextCheckIn=%v", enabledTime.UnixNano(), handshake, resetTime)
 		case <-w.resetCh:
-			w.log.Infof("WireGuard watcher received peer reset, restarting handshake timeout")
+			w.log.Warnf("WGW-DIAG: peer reset received, restarting timeout id=%d", enabledTime.UnixNano())
 			lastHandshake = time.Time{}
 			enabledTime = time.Now()
 			timer.Stop()
 			timer.Reset(wgHandshakeOvertime)
 		case <-ctx.Done():
-			w.log.Infof("WireGuard watcher stopped")
+			w.log.Warnf("WGW-DIAG: watcher stopped (ctx done) id=%d", enabledTime.UnixNano())
 			return
 		}
 	}

client/internal/peer/wg_watcher_test.go

@@ -7,7 +7,6 @@ import (
 	"time"
 
 	log "github.com/sirupsen/logrus"
-	"github.com/stretchr/testify/require"
 
 	"github.com/netbirdio/netbird/client/iface/configurer"
 )
@@ -35,9 +34,6 @@ func TestWGWatcher_EnableWgWatcher(t *testing.T) {
 	ctx, cancel := context.WithCancel(context.Background())
 	defer cancel()
 
-	ok := watcher.PrepareInitialHandshake()
-	require.True(t, ok, "watcher should not be enabled yet")
-
 	onDisconnected := make(chan struct{}, 1)
 	go watcher.EnableWgWatcher(ctx, time.Now(), func() {
 		mlog.Infof("onDisconnectedFn")
@@ -66,9 +62,6 @@ func TestWGWatcher_ReEnable(t *testing.T) {
 	watcher := NewWGWatcher(mlog, mocWgIface, "", newStateDump("peer", mlog, &Status{}))
 
 	ctx, cancel := context.WithCancel(context.Background())
-	ok := watcher.PrepareInitialHandshake()
-	require.True(t, ok, "watcher should not be enabled yet")
-
 	wg := &sync.WaitGroup{}
 	wg.Add(1)
 	go func() {
@@ -83,9 +76,6 @@ func TestWGWatcher_ReEnable(t *testing.T) {
 	ctx, cancel = context.WithCancel(context.Background())
 	defer cancel()
 
-	ok = watcher.PrepareInitialHandshake()
-	require.True(t, ok, "watcher should be re-enabled after the previous run stopped")
-
 	onDisconnected := make(chan struct{}, 1)
 	go watcher.EnableWgWatcher(ctx, time.Now(), func() {
 		onDisconnected <- struct{}{}

client/internal/rosenpass/manager.go

@@ -82,9 +82,15 @@ func (m *Manager) GetPubKey() []byte {
 	return m.spk
 }
 
-// GetAddress returns the address of the Rosenpass server
+// GetAddress returns the address of the Rosenpass server.
+//
+// The server binds v4-only (0.0.0.0). Rosenpass reaches peers over their
+// WireGuard overlay IP, which is always IPv4, so a v4 socket suffices. This
+// avoids the AF_INET6 -> IPv4 send rejection (EDESTADDRREQ) that a [::]
+// (dual-stack) socket hits on macOS/BSD when sending to a 4-byte IPv4
+// destination.
 func (m *Manager) GetAddress() *net.UDPAddr {
-	return &net.UDPAddr{Port: m.port}
+	return &net.UDPAddr{IP: net.IPv4zero, Port: m.port}
 }
 
 // addPeer adds a new peer to the Rosenpass server
@@ -109,20 +115,14 @@ func (m *Manager) addPeer(rosenpassPubKey []byte, rosenpassAddr string, wireGuar
 		if err != nil {
 			return fmt.Errorf("failed to parse rosenpass address: %w", err)
 		}
+		// Resolve as udp4: our Rosenpass server binds v4-only (see GetAddress)
+		// and the peer WireGuard overlay IP is always IPv4. This keeps the
+		// destination a 4-byte IPv4 address that matches our v4 listening
+		// socket, avoiding the AF_INET6 -> IPv4 send rejection on macOS/BSD.
 		peerAddr := net.JoinHostPort(wireGuardIP, strPort)
-		if pcfg.Endpoint, err = net.ResolveUDPAddr("udp", peerAddr); err != nil {
+		if pcfg.Endpoint, err = net.ResolveUDPAddr("udp4", peerAddr); err != nil {
 			return fmt.Errorf("failed to resolve peer endpoint address: %w", err)
 		}
-		// Our local Rosenpass UDP server binds on the IPv6 wildcard ([::]) — see
-		// GetAddress(). The remote peer's endpoint (pcfg.Endpoint) is the destination
-		// our server will sendto when initiating handshakes. ResolveUDPAddr returns a
-		// 4-byte IPv4 for IPv4 hosts, which the kernel rejects (EDESTADDRREQ) when
-		// sent from an AF_INET6 socket. Normalize the remote endpoint to IPv4-mapped
-		// IPv6 so its address family matches our listening socket.
-		// TODO: maybe bind the Rosenpass UDP server to the peer wg IP addr
-		if v4 := pcfg.Endpoint.IP.To4(); v4 != nil {
-			pcfg.Endpoint.IP = v4.To16()
-		}
 	}
 	peerID, err := m.server.AddPeer(pcfg)
 	if err != nil {
@@ -280,13 +280,15 @@ func (m *Manager) OnConnected(remoteWireGuardKey string, remoteRosenpassPubKey [
 	}
 
 	rpKeyHash := hashRosenpassKey(remoteRosenpassPubKey)
-	log.Debugf("received remote rosenpass key %s, my key %s", rpKeyHash, m.rpKeyHash)
+	initiator := bytes.Compare(m.spk, remoteRosenpassPubKey) == 1
+	log.Warnf("PSK-DIAG: remote peer %s advertises rosenpass (key %s, addr %s); starting RP negotiation as initiator=%v", remoteWireGuardKey, rpKeyHash, remoteRosenpassAddr, initiator)
 
 	err := m.addPeer(remoteRosenpassPubKey, remoteRosenpassAddr, wireGuardIP, remoteWireGuardKey)
 	if err != nil {
 		log.Errorf("failed to add rosenpass peer: %s", err)
 		return
 	}
+	log.Warnf("PSK-DIAG: rosenpass peer added for %s (RP negotiation started; watch for 'set PSK on WG' to confirm completion)", remoteWireGuardKey)
 }
 
 // IsPresharedKeyInitialized returns true if Rosenpass has completed a handshake

client/internal/rosenpass/netbird_handler.go

@@ -100,6 +100,7 @@ func (h *NetbirdHandler) outputKey(_ rp.KeyOutputReason, pid rp.PeerID, psk rp.K
 		log.Errorf("Failed to apply rosenpass key: %v", err)
 		return
 	}
+	log.Warnf("PSK-DIAG: rosenpass set PSK on WG for peer %s (updateOnly=%v)", peerKey, isInitialized)
 
 	// Mark peer as isInitialized after the successful first rotation
 	if !isInitialized {

client/internal/routemanager/client/client.go

@@ -231,11 +231,15 @@ func (w *Watcher) getBestRouteFromStatuses(routePeerStatuses map[route.ID]router
 	switch {
 	case chosen == "":
 		var peers []string
-		for _, r := range w.routes {
-			peers = append(peers, r.Peer)
+		for id, r := range w.routes {
+			st := "unknown(no status)"
+			if ps, ok := routePeerStatuses[id]; ok {
+				st = fmt.Sprintf("%s relayed=%v lat=%v", ps.status, ps.relayed, ps.latency)
+			}
+			peers = append(peers, fmt.Sprintf("%s{%s}", r.Peer, st))
 		}
 
-		log.Infof("network [%v] has not been assigned a routing peer as no peers from the list %s are currently available", w.handler, peers)
+		log.Warnf("ROUTE-DIAG: network [%v] no routing peer available yet (all candidates connecting/absent); candidates: %s", w.handler, peers)
 	case chosen != currID:
 		// we compare the current score + 10ms to the chosen score to avoid flapping between routes
 		if currScore != 0 && currScore+0.01 > chosenScore {
@@ -247,7 +251,7 @@ func (w *Watcher) getBestRouteFromStatuses(routePeerStatuses map[route.ID]router
 		if rt := w.routes[chosen]; rt != nil {
 			p = rt.Peer
 		}
-		log.Infof("New chosen route is %s with peer %s with score %f for network [%v]", chosen, p, chosenScore, w.handler)
+		log.Warnf("ROUTE-DIAG: new chosen route %s peer %s score %f for network [%v] (routing peer now usable)", chosen, p, chosenScore, w.handler)
 	}
 
 	return chosen, chosenStatus

client/internal/routemanager/exit_node_selection_test.go

@@ -0,0 +1,191 @@
+package routemanager
+
+import (
+	"net/netip"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/client/internal/routeselector"
+	"github.com/netbirdio/netbird/route"
+)
+
+func newExitNodeTestManager() *DefaultManager {
+	return &DefaultManager{routeSelector: routeselector.NewRouteSelector()}
+}
+
+func exitRoute(netID, peer string, skipAutoApply bool) *route.Route {
+	return &route.Route{
+		NetID:         route.NetID(netID),
+		Network:       netip.MustParsePrefix("0.0.0.0/0"),
+		Peer:          peer,
+		SkipAutoApply: skipAutoApply,
+	}
+}
+
+func TestPickPreferredExitNode(t *testing.T) {
+	tests := []struct {
+		name string
+		info exitNodeInfo
+		want route.NetID
+	}{
+		{
+			name: "persisted user selection wins over management",
+			info: exitNodeInfo{
+				allIDs:               []route.NetID{"a", "b", "c"},
+				userSelected:         []route.NetID{"b"},
+				selectedByManagement: []route.NetID{"a"},
+			},
+			want: "b",
+		},
+		{
+			name: "multiple user-selected self-heal to deterministic min",
+			info: exitNodeInfo{
+				allIDs:       []route.NetID{"a", "b", "c"},
+				userSelected: []route.NetID{"c", "a"},
+			},
+			want: "a",
+		},
+		{
+			name: "explicit opt-out keeps none",
+			info: exitNodeInfo{
+				allIDs:         []route.NetID{"a", "b"},
+				userDeselected: []route.NetID{"a", "b"},
+			},
+			want: "",
+		},
+		{
+			name: "fresh defaults to management auto-apply pick",
+			info: exitNodeInfo{
+				allIDs:               []route.NetID{"a", "b", "c"},
+				selectedByManagement: []route.NetID{"b"},
+			},
+			want: "b",
+		},
+		{
+			name: "no user pick and no management auto-apply selects none",
+			info: exitNodeInfo{
+				allIDs: []route.NetID{"c", "a", "b"},
+			},
+			want: "",
+		},
+		{
+			name: "user-deselect does not block a management auto-apply sibling",
+			info: exitNodeInfo{
+				allIDs:               []route.NetID{"a", "b"},
+				userDeselected:       []route.NetID{"a"},
+				selectedByManagement: []route.NetID{"b"},
+			},
+			want: "b",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			assert.Equal(t, tt.want, pickPreferredExitNode(tt.info), "preferred exit node")
+		})
+	}
+}
+
+func TestEnforceSingleExitNode(t *testing.T) {
+	m := newExitNodeTestManager()
+	all := []route.NetID{"a", "b", "c"}
+
+	m.enforceSingleExitNode("b", all)
+	assert.False(t, m.routeSelector.IsSelected("a"), "a should be deselected")
+	assert.True(t, m.routeSelector.IsSelected("b"), "b should be the only selected exit node")
+	assert.False(t, m.routeSelector.IsSelected("c"), "c should be deselected")
+
+	// Switching the preferred node moves the single selection.
+	m.enforceSingleExitNode("c", all)
+	assert.False(t, m.routeSelector.IsSelected("a"), "a stays deselected")
+	assert.False(t, m.routeSelector.IsSelected("b"), "b should now be deselected")
+	assert.True(t, m.routeSelector.IsSelected("c"), "c should now be selected")
+
+	// Empty preferred turns every exit node off.
+	m.enforceSingleExitNode("", all)
+	for _, id := range all {
+		assert.False(t, m.routeSelector.IsSelected(id), "no exit node should be selected")
+	}
+}
+
+func TestEnforceSingleExitNode_RespectsDeselectAll(t *testing.T) {
+	m := newExitNodeTestManager()
+	m.routeSelector.DeselectAllRoutes()
+
+	m.enforceSingleExitNode("b", []route.NetID{"a", "b"})
+
+	assert.True(t, m.routeSelector.IsDeselectAll(), "global deselect-all must stay in effect")
+	assert.False(t, m.routeSelector.IsSelected("b"), "no exit node should be forced on while deselect-all is set")
+}
+
+func TestUpdateRouteSelectorFromManagement_FreshSelectsOne(t *testing.T) {
+	m := newExitNodeTestManager()
+	routes := route.HAMap{
+		"exitA|0.0.0.0/0":    {exitRoute("exitA", "p1", false)},
+		"exitB|0.0.0.0/0":    {exitRoute("exitB", "p2", false)},
+		"lan|192.168.1.0/24": {{NetID: "lan", Network: netip.MustParsePrefix("192.168.1.0/24"), Peer: "p3"}},
+		"exitC|0.0.0.0/0":    {exitRoute("exitC", "p4", false)},
+	}
+
+	m.updateRouteSelectorFromManagement(routes)
+
+	// Exactly one exit node (the deterministic first) is selected.
+	assert.True(t, m.routeSelector.IsSelected("exitA"), "exitA is the deterministic default")
+	assert.False(t, m.routeSelector.IsSelected("exitB"), "exitB must not also be selected")
+	assert.False(t, m.routeSelector.IsSelected("exitC"), "exitC must not also be selected")
+	// Non-exit routes are left at their default-on state.
+	assert.True(t, m.routeSelector.IsSelected("lan"), "non-exit route selection is untouched")
+}
+
+func TestUpdateRouteSelectorFromManagement_HonorsPersistedPick(t *testing.T) {
+	m := newExitNodeTestManager()
+	routes := route.HAMap{
+		"exitA|0.0.0.0/0": {exitRoute("exitA", "p1", false)},
+		"exitB|0.0.0.0/0": {exitRoute("exitB", "p2", false)},
+	}
+	all := []route.NetID{"exitA", "exitB"}
+
+	// Simulate the state the runtime select path leaves behind: exactly one
+	// exit node explicitly selected, its sibling deselected.
+	require.NoError(t, m.routeSelector.SelectRoutes([]route.NetID{"exitB"}, true, all))
+	require.NoError(t, m.routeSelector.DeselectRoutes([]route.NetID{"exitA"}, all))
+
+	m.updateRouteSelectorFromManagement(routes)
+
+	assert.True(t, m.routeSelector.IsSelected("exitB"), "persisted pick must stay selected")
+	assert.False(t, m.routeSelector.IsSelected("exitA"), "the other exit node stays deselected")
+}
+
+func TestUpdateRouteSelectorFromManagement_OptOutKeepsNone(t *testing.T) {
+	m := newExitNodeTestManager()
+	routes := route.HAMap{
+		"exitA|0.0.0.0/0": {exitRoute("exitA", "p1", false)},
+		"exitB|0.0.0.0/0": {exitRoute("exitB", "p2", false)},
+	}
+	all := []route.NetID{"exitA", "exitB"}
+
+	// User deselected exit nodes and selected none.
+	require.NoError(t, m.routeSelector.DeselectRoutes(all, all))
+
+	m.updateRouteSelectorFromManagement(routes)
+
+	assert.False(t, m.routeSelector.IsSelected("exitA"), "opt-out keeps exitA off")
+	assert.False(t, m.routeSelector.IsSelected("exitB"), "opt-out keeps exitB off")
+}
+
+func TestUpdateRouteSelectorFromManagement_NoAutoApplySelectsNone(t *testing.T) {
+	m := newExitNodeTestManager()
+	// SkipAutoApply=true: management offers the exit nodes but doesn't request
+	// auto-activation, so none should be selected until the user picks one.
+	routes := route.HAMap{
+		"exitA|0.0.0.0/0": {exitRoute("exitA", "p1", true)},
+		"exitB|0.0.0.0/0": {exitRoute("exitB", "p2", true)},
+	}
+
+	m.updateRouteSelectorFromManagement(routes)
+
+	assert.False(t, m.routeSelector.IsSelected("exitA"), "no auto-apply keeps exitA off")
+	assert.False(t, m.routeSelector.IsSelected("exitB"), "no auto-apply keeps exitB off")
+}

client/internal/routemanager/manager.go

@@ -701,7 +701,13 @@ func resolveURLsToIPs(urls []string) []net.IP {
 	return ips
 }
 
-// updateRouteSelectorFromManagement updates the route selector based on the isSelected status from the management server
+// updateRouteSelectorFromManagement reconciles exit-node selection on every
+// network map: it keeps at most one exit node selected — the user's persisted
+// pick, else whatever management marks for auto-apply (SkipAutoApply=false),
+// else none. We never auto-activate an exit node the map doesn't request; it
+// stays off until the user picks it. Exit nodes are mutually exclusive, but the
+// RouteSelector stores routes with default-on semantics, so without this every
+// available exit node would report selected at once.
 func (m *DefaultManager) updateRouteSelectorFromManagement(clientRoutes route.HAMap) {
 	m.mirrorV6ExitPairSelections(clientRoutes)
 
@@ -712,13 +718,14 @@ func (m *DefaultManager) updateRouteSelectorFromManagement(clientRoutes route.HA
 		return
 	}
 
-	exitNodeInfo := m.collectExitNodeInfo(clientRoutes)
-	if len(exitNodeInfo.allIDs) == 0 {
+	info := m.collectExitNodeInfo(clientRoutes)
+	if len(info.allIDs) == 0 {
 		return
 	}
 
-	m.updateExitNodeSelections(exitNodeInfo)
-	m.logExitNodeUpdate(exitNodeInfo)
+	preferred := pickPreferredExitNode(info)
+	m.enforceSingleExitNode(preferred, info.allIDs)
+	m.logExitNodeUpdate(info, preferred)
 }
 
 // mirrorV6ExitPairSelections keeps every synthesized "-v6" exit route's selection
@@ -746,6 +753,10 @@ type exitNodeInfo struct {
 	userDeselected       []route.NetID
 }
 
+// collectExitNodeInfo categorises the available exit nodes by their persisted
+// selection state. It keys on the base (v4) NetID and skips the synthesized
+// "-v6" partner, which inherits its base's selection through the RouteSelector
+// — counting it separately would double-count the pair.
 func (m *DefaultManager) collectExitNodeInfo(clientRoutes route.HAMap) exitNodeInfo {
 	var info exitNodeInfo
 
@@ -755,6 +766,9 @@ func (m *DefaultManager) collectExitNodeInfo(clientRoutes route.HAMap) exitNodeI
 		}
 
 		netID := haID.NetID()
+		if strings.HasSuffix(string(netID), route.V6ExitSuffix) {
+			continue
+		}
 		info.allIDs = append(info.allIDs, netID)
 
 		if m.routeSelector.HasUserSelectionForRoute(netID) {
@@ -791,45 +805,52 @@ func (m *DefaultManager) checkManagementSelection(routes []*route.Route, netID r
 	}
 }
 
-func (m *DefaultManager) updateExitNodeSelections(info exitNodeInfo) {
-	routesToDeselect := m.getRoutesToDeselect(info.allIDs)
-	m.deselectExitNodes(routesToDeselect)
-	m.selectExitNodesByManagement(info.selectedByManagement, info.allIDs)
+// pickPreferredExitNode chooses the single exit node to keep selected. In order:
+//   - a persisted user selection wins (deterministic if several survive from
+//     legacy state, so the set self-heals down to one);
+//   - otherwise activate only what management marks for auto-apply
+//     (SkipAutoApply=false); the lexicographically first if it marks several.
+//
+// Returns "" when neither holds — we never force an arbitrary exit node on. A
+// route the map doesn't auto-apply stays off until the user selects it.
+// info.userDeselected is informational only: an explicit deselect simply keeps
+// that route out of both lists above, so it can't be picked.
+func pickPreferredExitNode(info exitNodeInfo) route.NetID {
+	if len(info.userSelected) > 0 {
+		return minNetID(info.userSelected)
+	}
+	if len(info.selectedByManagement) > 0 {
+		return minNetID(info.selectedByManagement)
+	}
+	return ""
 }
 
-func (m *DefaultManager) getRoutesToDeselect(allIDs []route.NetID) []route.NetID {
-	var routesToDeselect []route.NetID
-	for _, netID := range allIDs {
-		if !m.routeSelector.HasUserSelectionForRoute(netID) {
-			routesToDeselect = append(routesToDeselect, netID)
+// enforceSingleExitNode makes preferred the only selected exit node: every other
+// available exit node is deselected and preferred (if any) is selected, without
+// disturbing non-exit route selections. The whole reconciliation runs under a
+// single RouteSelector lock (SetExclusiveExitNode) so a concurrent deselect-all
+// cannot interleave and get undone; a global deselect-all is left untouched so
+// the user's "all off" stays in effect.
+func (m *DefaultManager) enforceSingleExitNode(preferred route.NetID, allIDs []route.NetID) {
+	m.routeSelector.SetExclusiveExitNode(preferred, allIDs)
+}
+
+func (m *DefaultManager) logExitNodeUpdate(info exitNodeInfo, preferred route.NetID) {
+	log.Debugf("Exit node selection: %d available, preferred=%q (%d user-selected, %d user-deselected, %d management-selected)",
+		len(info.allIDs), preferred, len(info.userSelected), len(info.userDeselected), len(info.selectedByManagement))
+}
+
+// minNetID returns the lexicographically smallest NetID, for a deterministic
+// default pick that stays stable across restarts.
+func minNetID(ids []route.NetID) route.NetID {
+	if len(ids) == 0 {
+		return ""
+	}
+	best := ids[0]
+	for _, id := range ids[1:] {
+		if id < best {
+			best = id
 		}
 	}
-	return routesToDeselect
-}
-
-func (m *DefaultManager) deselectExitNodes(routesToDeselect []route.NetID) {
-	if len(routesToDeselect) == 0 {
-		return
-	}
-
-	err := m.routeSelector.DeselectRoutes(routesToDeselect, routesToDeselect)
-	if err != nil {
-		log.Warnf("Failed to deselect exit nodes: %v", err)
-	}
-}
-
-func (m *DefaultManager) selectExitNodesByManagement(selectedByManagement []route.NetID, allIDs []route.NetID) {
-	if len(selectedByManagement) == 0 {
-		return
-	}
-
-	err := m.routeSelector.SelectRoutes(selectedByManagement, true, allIDs)
-	if err != nil {
-		log.Warnf("Failed to select exit nodes: %v", err)
-	}
-}
-
-func (m *DefaultManager) logExitNodeUpdate(info exitNodeInfo) {
-	log.Debugf("Updated route selector: %d exit nodes available, %d selected by management, %d user-selected, %d user-deselected",
-		len(info.allIDs), len(info.selectedByManagement), len(info.userSelected), len(info.userDeselected))
+	return best
 }

client/internal/routeselector/routeselector.go

@@ -115,7 +115,38 @@ func (rs *RouteSelector) DeselectAllRoutes() {
 	clear(rs.selectedRoutes)
 }
 
-// IsDeselectAll reports whether the user has explicitly deselected all routes.
+// SetExclusiveExitNode atomically makes preferred the only selected exit node
+// among exitIDs: every other ID in exitIDs is deselected and preferred (when
+// non-empty) is selected, all under a single lock. Holding the lock across the
+// whole reconciliation prevents a concurrent DeselectAllRoutes from interleaving
+// between the deselect and select steps and being silently undone. A global
+// deselect-all is left untouched so the user's "all off" stays in effect;
+// non-exit routes are never referenced, so their selection is preserved.
+func (rs *RouteSelector) SetExclusiveExitNode(preferred route.NetID, exitIDs []route.NetID) {
+	rs.mu.Lock()
+	defer rs.mu.Unlock()
+
+	if rs.deselectAll {
+		return
+	}
+
+	for _, id := range exitIDs {
+		if id == preferred {
+			continue
+		}
+		rs.deselectedRoutes[id] = struct{}{}
+		delete(rs.selectedRoutes, id)
+	}
+
+	if preferred != "" {
+		delete(rs.deselectedRoutes, preferred)
+		rs.selectedRoutes[preferred] = struct{}{}
+	}
+}
+
+// IsDeselectAll reports whether the global "deselect all" flag is set, i.e. the
+// user explicitly disabled every route. Callers enforcing per-route invariants
+// (e.g. single exit node) should leave the selection untouched when it is.
 func (rs *RouteSelector) IsDeselectAll() bool {
 	rs.mu.RLock()
 	defer rs.mu.RUnlock()