[client] Wait for signal receive watchdog to stop before reconnect

The per-stream watchReceiveStream goroutine was started fire-and-forget
and never joined. On reconnect a lingering watchdog could still flip
shared client state (receiveStalled, the disconnect notifier) on the
freshly established stream, since cancelStream only cancels its own
stream context.

Track the watchdog with a WaitGroup and wait for it to exit (after
cancelling its stream) before the operation returns, so each reconnect
starts with no stale watchdog.

client/cmd/debug.go

@@ -130,7 +130,7 @@ func debugConfigDump(cmd *cobra.Command, _ []string) error {
 
 	client := proto.NewDaemonServiceClient(conn)
 	resp, err := client.GetConfig(cmd.Context(), &proto.GetConfigRequest{
-		ProfileName: activeProf.Name,
+		ProfileName: string(activeProf.ID),
 		Username:    currUser.Username,
 	})
 	if err != nil {

client/internal/dns/mgmt/mgmt.go

@@ -51,13 +51,20 @@ type cachedRecord struct {
 }
 
 // Resolver caches critical NetBird infrastructure domains.
-// records, refreshing, mgmtDomain and serverDomains are all guarded by mutex.
+// records, refreshing, failedResolves, mgmtDomain and serverDomains are all
+// guarded by mutex.
 type Resolver struct {
 	records       map[dns.Question]*cachedRecord
 	mgmtDomain    *domain.Domain
 	serverDomains *dnsconfig.ServerDomains
 	mutex         sync.RWMutex
 
+	// failedResolves records the last failed initial resolve per domain so a
+	// domain that never resolves isn't retried on every server-domains update
+	// until refreshBackoff elapses. Entries are cleared on success and pruned
+	// to the current server-domains set.
+	failedResolves map[domain.Domain]time.Time
+
 	chain            ChainResolver
 	chainMaxPriority int
 	refreshGroup     singleflight.Group
@@ -76,9 +83,10 @@ type Resolver struct {
 // NewResolver creates a new management domains cache resolver.
 func NewResolver() *Resolver {
 	return &Resolver{
-		records:    make(map[dns.Question]*cachedRecord),
-		refreshing: make(map[dns.Question]*atomic.Bool),
-		cacheTTL:   resolveCacheTTL(),
+		records:        make(map[dns.Question]*cachedRecord),
+		refreshing:     make(map[dns.Question]*atomic.Bool),
+		failedResolves: make(map[domain.Domain]time.Time),
+		cacheTTL:       resolveCacheTTL(),
 	}
 }
 
@@ -173,7 +181,9 @@ func (m *Resolver) continueToNext(w dns.ResponseWriter, r *dns.Msg) {
 
 // AddDomain resolves a domain and stores its A/AAAA records in the cache.
 // A family that resolves NODATA (nil err, zero records) evicts any stale
-// entry for that qtype.
+// entry for that qtype. When one family hard-errors while the other succeeds,
+// the resolved family is still cached but AddDomain returns an error so the
+// caller retries the incomplete resolve rather than treating it as complete.
 func (m *Resolver) AddDomain(ctx context.Context, d domain.Domain) error {
 	dnsName := strings.ToLower(dns.Fqdn(d.PunycodeString()))
 
@@ -203,6 +213,10 @@ func (m *Resolver) AddDomain(ctx context.Context, d domain.Domain) error {
 	log.Debugf("added/updated domain=%s with %d A records and %d AAAA records",
 		d.SafeString(), len(aRecords), len(aaaaRecords))
 
+	if errA != nil || errAAAA != nil {
+		return fmt.Errorf("resolve %s: incomplete, a family failed: %w", d.SafeString(), errors.Join(errA, errAAAA))
+	}
+
 	return nil
 }
 
@@ -462,6 +476,7 @@ func (m *Resolver) RemoveDomain(d domain.Domain) error {
 	delete(m.records, qAAAA)
 	delete(m.refreshing, qA)
 	delete(m.refreshing, qAAAA)
+	delete(m.failedResolves, d)
 
 	log.Debugf("removed domain=%s from cache", d.SafeString())
 	return nil
@@ -505,6 +520,7 @@ func (m *Resolver) UpdateFromServerDomains(ctx context.Context, serverDomains dn
 		allDomains := m.extractDomainsFromServerDomains(updatedServerDomains)
 		currentDomains := m.GetCachedDomains()
 		removedDomains = m.removeStaleDomains(currentDomains, allDomains)
+		m.pruneFailedResolves(allDomains)
 	}
 
 	m.addNewDomains(ctx, newDomains)
@@ -577,13 +593,85 @@ func (m *Resolver) isManagementDomain(domain domain.Domain) bool {
 	return m.mgmtDomain != nil && domain == *m.mgmtDomain
 }
 
-// addNewDomains resolves and caches all domains from the update
+// addNewDomains resolves and caches domains that are not yet in the cache,
+// running the lookups concurrently. Domains already cached are skipped and left
+// to the stale-while-revalidate refresh path, so a sync never re-resolves them
+// synchronously: once NetBird owns the OS resolver the resolve runs through the
+// handler chain and would otherwise dial the managed upstreams under the engine
+// sync lock on every update.
 func (m *Resolver) addNewDomains(ctx context.Context, newDomains domain.List) {
+	var wg sync.WaitGroup
+	seen := make(map[domain.Domain]struct{}, len(newDomains))
 	for _, newDomain := range newDomains {
-		if err := m.AddDomain(ctx, newDomain); err != nil {
-			log.Warnf("failed to add/update domain=%s: %v", newDomain.SafeString(), err)
-		} else {
-			log.Debugf("added/updated management cache domain=%s", newDomain.SafeString())
+		if _, dup := seen[newDomain]; dup {
+			continue
+		}
+		seen[newDomain] = struct{}{}
+
+		if !m.needsResolve(newDomain) {
+			continue
+		}
+
+		wg.Add(1)
+		go func(d domain.Domain) {
+			defer wg.Done()
+			if err := m.AddDomain(ctx, d); err != nil {
+				m.markResolveFailed(d)
+				log.Warnf("failed to add/update domain=%s: %v", d.SafeString(), err)
+				return
+			}
+			m.clearResolveFailed(d)
+			log.Debugf("added/updated management cache domain=%s", d.SafeString())
+		}(newDomain)
+	}
+	wg.Wait()
+}
+
+// needsResolve reports whether d should be resolved now. A recent failed or
+// incomplete resolve gates retries on the backoff even when one family is
+// already cached, so a transiently-failed family is retried instead of being
+// treated as fully resolved. Otherwise a domain with any cached record is left
+// to the stale-while-revalidate refresh path.
+func (m *Resolver) needsResolve(d domain.Domain) bool {
+	dnsName := strings.ToLower(dns.Fqdn(d.PunycodeString()))
+
+	m.mutex.RLock()
+	defer m.mutex.RUnlock()
+
+	if failedAt, ok := m.failedResolves[d]; ok {
+		return time.Since(failedAt) >= refreshBackoff
+	}
+
+	for _, qtype := range []uint16{dns.TypeA, dns.TypeAAAA} {
+		q := dns.Question{Name: dnsName, Qtype: qtype, Qclass: dns.ClassINET}
+		if _, ok := m.records[q]; ok {
+			return false
+		}
+	}
+	return true
+}
+
+func (m *Resolver) markResolveFailed(d domain.Domain) {
+	m.mutex.Lock()
+	m.failedResolves[d] = time.Now()
+	m.mutex.Unlock()
+}
+
+func (m *Resolver) clearResolveFailed(d domain.Domain) {
+	m.mutex.Lock()
+	delete(m.failedResolves, d)
+	m.mutex.Unlock()
+}
+
+// pruneFailedResolves drops failure markers for domains no longer present in
+// the server-domains set, keeping the map bounded to the current set (a
+// failed-only domain has no cached record, so RemoveDomain never sees it).
+func (m *Resolver) pruneFailedResolves(domains domain.List) {
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+	for d := range m.failedResolves {
+		if !slices.Contains(domains, d) {
+			delete(m.failedResolves, d)
 		}
 	}
 }

client/internal/dns/mgmt/mgmt_refresh_test.go

@@ -21,6 +21,7 @@ type fakeChain struct {
 	mu       sync.Mutex
 	calls    map[string]int
 	answers  map[string][]dns.RR
+	qErr     map[string]error
 	err      error
 	hasRoot  bool
 	onLookup func()
@@ -30,6 +31,7 @@ func newFakeChain() *fakeChain {
 	return &fakeChain{
 		calls:   map[string]int{},
 		answers: map[string][]dns.RR{},
+		qErr:    map[string]error{},
 		hasRoot: true,
 	}
 }
@@ -47,6 +49,9 @@ func (f *fakeChain) ResolveInternal(ctx context.Context, msg *dns.Msg, maxPriori
 	f.calls[key]++
 	answers := f.answers[key]
 	err := f.err
+	if err == nil {
+		err = f.qErr[key]
+	}
 	onLookup := f.onLookup
 	f.mu.Unlock()
 
@@ -75,6 +80,12 @@ func (f *fakeChain) setAnswer(name string, qtype uint16, ip string) {
 	}
 }
 
+func (f *fakeChain) setErr(name string, qtype uint16, err error) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.qErr[name+"|"+dns.TypeToString[qtype]] = err
+}
+
 func (f *fakeChain) callCount(name string, qtype uint16) int {
 	f.mu.Lock()
 	defer f.mu.Unlock()

client/internal/dns/mgmt/mgmt_resolve_test.go

@@ -0,0 +1,183 @@
+package mgmt
+
+import (
+	"context"
+	"errors"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/miekg/dns"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	dnsconfig "github.com/netbirdio/netbird/client/internal/dns/config"
+	"github.com/netbirdio/netbird/shared/management/domain"
+)
+
+// A domain already in the cache must not be re-resolved on a subsequent server
+// domains update; it is left to the stale-while-revalidate refresh path.
+func TestResolver_UpdateFromServerDomains_SkipsCached(t *testing.T) {
+	r := NewResolver()
+	chain := newFakeChain()
+	chain.setAnswer("signal.example.com.", dns.TypeA, "10.0.0.2")
+	r.SetChainResolver(chain, 50)
+
+	sd := dnsconfig.ServerDomains{Signal: domain.Domain("signal.example.com")}
+
+	_, err := r.UpdateFromServerDomains(context.Background(), sd)
+	require.NoError(t, err)
+	require.Equal(t, 1, chain.callCount("signal.example.com.", dns.TypeA),
+		"first update must resolve the domain")
+
+	_, err = r.UpdateFromServerDomains(context.Background(), sd)
+	require.NoError(t, err)
+	assert.Equal(t, 1, chain.callCount("signal.example.com.", dns.TypeA),
+		"cached domain must not be re-resolved on a subsequent update")
+}
+
+// New domains in a single update must resolve concurrently rather than serially.
+func TestResolver_AddNewDomains_ResolvesConcurrently(t *testing.T) {
+	r := NewResolver()
+	chain := newFakeChain()
+
+	var inflight, maxInflight atomic.Int32
+	chain.onLookup = func() {
+		n := inflight.Add(1)
+		for {
+			old := maxInflight.Load()
+			if n <= old || maxInflight.CompareAndSwap(old, n) {
+				break
+			}
+		}
+		time.Sleep(50 * time.Millisecond)
+		inflight.Add(-1)
+	}
+
+	relays := []domain.Domain{"a.example.com", "b.example.com", "c.example.com", "d.example.com"}
+	for _, d := range relays {
+		chain.setAnswer(dns.Fqdn(string(d)), dns.TypeA, "10.0.0.2")
+	}
+	r.SetChainResolver(chain, 50)
+
+	start := time.Now()
+	_, err := r.UpdateFromServerDomains(context.Background(), dnsconfig.ServerDomains{Relay: relays})
+	require.NoError(t, err)
+	elapsed := time.Since(start)
+
+	assert.GreaterOrEqual(t, int(maxInflight.Load()), 2, "domains must resolve concurrently")
+	// Serial resolution of 4 domains would take at least 4*50ms; concurrent is far less.
+	assert.Less(t, elapsed, 300*time.Millisecond, "resolution should not be serial")
+}
+
+// A domain that fails to resolve must not be retried on every update; the
+// failure backoff suppresses re-resolution until it expires.
+func TestResolver_UpdateFromServerDomains_BacksOffFailures(t *testing.T) {
+	r := NewResolver()
+	chain := newFakeChain()
+	chain.err = errors.New("resolve boom")
+	r.SetChainResolver(chain, 50)
+
+	sd := dnsconfig.ServerDomains{Signal: domain.Domain("signal.example.com")}
+
+	_, err := r.UpdateFromServerDomains(context.Background(), sd)
+	require.NoError(t, err)
+	require.Equal(t, 1, chain.callCount("signal.example.com.", dns.TypeA),
+		"first update must attempt the resolve")
+
+	_, err = r.UpdateFromServerDomains(context.Background(), sd)
+	require.NoError(t, err)
+	assert.Equal(t, 1, chain.callCount("signal.example.com.", dns.TypeA),
+		"failed resolve must back off and not retry on the next update")
+}
+
+// A domain listed under more than one server-domain type (e.g. STUN and TURN on
+// the same host) must be resolved once per update, not once per occurrence.
+func TestResolver_AddNewDomains_DedupesDuplicateDomains(t *testing.T) {
+	r := NewResolver()
+	chain := newFakeChain()
+	chain.setAnswer("dup.example.com.", dns.TypeA, "10.0.0.9")
+	r.SetChainResolver(chain, 50)
+
+	sd := dnsconfig.ServerDomains{
+		Stuns: []domain.Domain{"dup.example.com"},
+		Turns: []domain.Domain{"dup.example.com"},
+	}
+
+	_, err := r.UpdateFromServerDomains(context.Background(), sd)
+	require.NoError(t, err)
+	assert.Equal(t, 1, chain.callCount("dup.example.com.", dns.TypeA),
+		"a domain appearing under multiple server-domain types must resolve once")
+}
+
+// A failure marker must be dropped once its domain leaves the server-domains set
+// so the map stays bounded to the current set.
+func TestResolver_UpdateFromServerDomains_PrunesFailedResolves(t *testing.T) {
+	r := NewResolver()
+	chain := newFakeChain()
+	chain.err = errors.New("resolve boom")
+	r.SetChainResolver(chain, 50)
+
+	_, err := r.UpdateFromServerDomains(context.Background(), dnsconfig.ServerDomains{Signal: domain.Domain("gone.example.com")})
+	require.NoError(t, err)
+	r.mutex.RLock()
+	_, marked := r.failedResolves[domain.Domain("gone.example.com")]
+	r.mutex.RUnlock()
+	require.True(t, marked, "failed resolve must be recorded")
+
+	_, err = r.UpdateFromServerDomains(context.Background(), dnsconfig.ServerDomains{Signal: domain.Domain("other.example.com")})
+	require.NoError(t, err)
+	r.mutex.RLock()
+	_, stillMarked := r.failedResolves[domain.Domain("gone.example.com")]
+	r.mutex.RUnlock()
+	assert.False(t, stillMarked, "failure marker for a domain no longer in the set must be pruned")
+}
+
+// When one family hard-errors while the other resolves, the domain is cached
+// for the working family but recorded as incomplete so the failed family is
+// retried under backoff instead of being treated as fully resolved forever.
+func TestResolver_AddNewDomains_RetriesPartialFamilyFailure(t *testing.T) {
+	d := domain.Domain("relay.example.com")
+	r := NewResolver()
+	chain := newFakeChain()
+	chain.setAnswer("relay.example.com.", dns.TypeA, "10.0.0.2")
+	chain.setErr("relay.example.com.", dns.TypeAAAA, errors.New("servfail"))
+	r.SetChainResolver(chain, 50)
+
+	_, err := r.UpdateFromServerDomains(context.Background(), dnsconfig.ServerDomains{Relay: []domain.Domain{d}})
+	require.NoError(t, err)
+
+	r.mutex.RLock()
+	_, aCached := r.records[dns.Question{Name: "relay.example.com.", Qtype: dns.TypeA, Qclass: dns.ClassINET}]
+	_, marked := r.failedResolves[d]
+	r.mutex.RUnlock()
+	require.True(t, aCached, "the working family must still be cached")
+	require.True(t, marked, "a partial failure must be recorded so the failed family is retried")
+
+	assert.False(t, r.needsResolve(d), "within the backoff window the domain is not retried")
+
+	r.mutex.Lock()
+	r.failedResolves[d] = time.Now().Add(-2 * refreshBackoff)
+	r.mutex.Unlock()
+	assert.True(t, r.needsResolve(d), "after the backoff elapses the domain is retried to pick up the missing family")
+}
+
+// A family that returns NODATA (legitimately absent, e.g. an IPv4-only host) is
+// not a failure: the domain must not be marked for retry, otherwise it would be
+// re-resolved on every sync.
+func TestResolver_AddNewDomains_NodataIsNotFailure(t *testing.T) {
+	d := domain.Domain("v4only.example.com")
+	r := NewResolver()
+	chain := newFakeChain()
+	chain.setAnswer("v4only.example.com.", dns.TypeA, "10.0.0.2")
+	r.SetChainResolver(chain, 50)
+
+	_, err := r.UpdateFromServerDomains(context.Background(), dnsconfig.ServerDomains{Relay: []domain.Domain{d}})
+	require.NoError(t, err)
+
+	r.mutex.RLock()
+	_, marked := r.failedResolves[d]
+	r.mutex.RUnlock()
+	assert.False(t, marked, "a NODATA family must not be recorded as a failure")
+	assert.False(t, r.needsResolve(d), "an IPv4-only host must not be re-resolved on later syncs")
+}

management/server/account_test.go

@@ -1916,6 +1916,117 @@ func TestDefaultAccountManager_MarkPeerConnected_PeerLoginExpiration(t *testing.
 	}
 }
 
+func TestDefaultAccountManager_MarkPeerDisconnected_SchedulesInactivityExpiration(t *testing.T) {
+	manager, _, err := createManager(t)
+	require.NoError(t, err, "unable to create account manager")
+
+	accountID, err := manager.GetAccountIDByUserID(context.Background(), auth.UserAuth{UserId: userID})
+	require.NoError(t, err, "unable to create an account")
+
+	key, err := wgtypes.GenerateKey()
+	require.NoError(t, err, "unable to generate WireGuard key")
+	peerPubKey := key.PublicKey().String()
+
+	_, _, _, _, err = manager.AddPeer(context.Background(), "", "", userID, &nbpeer.Peer{
+		Key:                         peerPubKey,
+		Meta:                        nbpeer.PeerSystemMeta{Hostname: "test-peer"},
+		InactivityExpirationEnabled: true,
+	}, false)
+	require.NoError(t, err, "unable to add peer")
+
+	_, err = manager.UpdateAccountSettings(context.Background(), accountID, userID, &types.Settings{
+		PeerLoginExpiration:             time.Hour,
+		PeerLoginExpirationEnabled:      true,
+		PeerInactivityExpiration:        time.Hour,
+		PeerInactivityExpirationEnabled: true,
+		Extra:                           &types.ExtraSettings{},
+	})
+	require.NoError(t, err, "expecting to update account settings successfully but got error")
+
+	// Establish a session so the matching-token disconnect is actually applied.
+	streamStartTime := time.Now().UTC()
+	err = manager.MarkPeerConnected(context.Background(), peerPubKey, accountID, streamStartTime.UnixNano(), nil)
+	require.NoError(t, err, "unable to mark peer connected")
+
+	// Install the mock only now, so the assertion observes the disconnect, not
+	// the earlier connect.
+	scheduled := make(chan struct{}, 1)
+	manager.peerInactivityExpiry = &MockScheduler{
+		CancelFunc: func(ctx context.Context, IDs []string) {},
+		ScheduleFunc: func(ctx context.Context, in time.Duration, ID string, job func() (nextRunIn time.Duration, reschedule bool)) {
+			select {
+			case scheduled <- struct{}{}:
+			default:
+			}
+		},
+	}
+
+	err = manager.MarkPeerDisconnected(context.Background(), peerPubKey, accountID, streamStartTime.UnixNano())
+	require.NoError(t, err, "unable to mark peer disconnected")
+
+	select {
+	case <-scheduled:
+		// expected: disconnect re-armed the inactivity expiry timer
+	case <-time.After(time.Second):
+		t.Fatal("expected inactivity expiration to be rescheduled when an eligible peer disconnects")
+	}
+}
+
+func TestDefaultAccountManager_MarkPeerDisconnected_SkipsInactivityExpirationWhenDisabled(t *testing.T) {
+	manager, _, err := createManager(t)
+	require.NoError(t, err, "unable to create account manager")
+
+	accountID, err := manager.GetAccountIDByUserID(context.Background(), auth.UserAuth{UserId: userID})
+	require.NoError(t, err, "unable to create an account")
+
+	key, err := wgtypes.GenerateKey()
+	require.NoError(t, err, "unable to generate WireGuard key")
+	peerPubKey := key.PublicKey().String()
+
+	_, _, _, _, err = manager.AddPeer(context.Background(), "", "", userID, &nbpeer.Peer{
+		Key:                         peerPubKey,
+		Meta:                        nbpeer.PeerSystemMeta{Hostname: "test-peer"},
+		InactivityExpirationEnabled: true,
+	}, false)
+	require.NoError(t, err, "unable to add peer")
+
+	// Peer is eligible (SSO + inactivity enabled) but the account-level setting
+	// stays disabled, so disconnect must not schedule anything.
+	_, err = manager.UpdateAccountSettings(context.Background(), accountID, userID, &types.Settings{
+		PeerLoginExpiration:             time.Hour,
+		PeerLoginExpirationEnabled:      true,
+		PeerInactivityExpiration:        time.Hour,
+		PeerInactivityExpirationEnabled: false,
+		Extra:                           &types.ExtraSettings{},
+	})
+	require.NoError(t, err, "expecting to update account settings successfully but got error")
+
+	streamStartTime := time.Now().UTC()
+	err = manager.MarkPeerConnected(context.Background(), peerPubKey, accountID, streamStartTime.UnixNano(), nil)
+	require.NoError(t, err, "unable to mark peer connected")
+
+	scheduled := make(chan struct{}, 1)
+	manager.peerInactivityExpiry = &MockScheduler{
+		CancelFunc: func(ctx context.Context, IDs []string) {},
+		ScheduleFunc: func(ctx context.Context, in time.Duration, ID string, job func() (nextRunIn time.Duration, reschedule bool)) {
+			select {
+			case scheduled <- struct{}{}:
+			default:
+			}
+		},
+	}
+
+	err = manager.MarkPeerDisconnected(context.Background(), peerPubKey, accountID, streamStartTime.UnixNano())
+	require.NoError(t, err, "unable to mark peer disconnected")
+
+	select {
+	case <-scheduled:
+		t.Fatal("inactivity expiration must not be scheduled while the account-level setting is disabled")
+	case <-time.After(200 * time.Millisecond):
+		// expected: nothing scheduled
+	}
+}
+
 func TestDefaultAccountManager_OnPeerDisconnected_LastSeenCheck(t *testing.T) {
 	manager, _, err := createManager(t)
 	require.NoError(t, err, "unable to create account manager")

management/server/peer.go

@@ -188,6 +188,15 @@ func (am *DefaultAccountManager) MarkPeerDisconnected(ctx context.Context, peerP
 		}
 	}
 
+	if peer.AddedWithSSOLogin() && peer.InactivityExpirationEnabled {
+		settings, err := am.Store.GetAccountSettings(ctx, store.LockingStrengthNone, accountID)
+		if err != nil {
+			log.WithContext(ctx).Warnf("failed getting account settings to schedule inactivity expiration for peer %s: %v", peer.ID, err)
+		} else if settings.PeerInactivityExpirationEnabled {
+			am.checkAndSchedulePeerInactivityExpiration(ctx, accountID)
+		}
+	}
+
 	return nil
 }
 

shared/signal/client/grpc.go

@@ -78,6 +78,14 @@ type GrpcClient struct {
 	// transport-alive but no longer delivering messages. It is the source of
 	// truth IsHealthy reads, and is cleared once any frame is received again.
 	receiveStalled atomic.Bool
+	// receiveHandoffBlocked is set while the receive loop is parked handing a
+	// message to a busy decryption worker. The loop stops calling Recv (and
+	// markReceived) in that window, so the stream looks silent though it is
+	// healthy. The watchdog reads this to avoid misreading self-inflicted
+	// receive backpressure as a dead stream: reconnecting cannot help, since the
+	// new stream feeds the same worker, and only triggers a reconnect storm.
+	receiveHandoffBlocked atomic.Bool
+	watchdogWg            sync.WaitGroup
 }
 
 // NewClient creates a new Signal client
@@ -193,10 +201,18 @@ func (c *GrpcClient) Receive(ctx context.Context, msgHandler func(msg *proto.Mes
 		// Guard the receive direction: the transport can stay healthy while the
 		// server stops delivering messages. The watchdog reconnects via cancelStream.
 		c.markReceived()
-		go c.watchReceiveStream(streamCtx, cancelStream)
+		c.watchdogWg.Add(1)
+		go func() {
+			defer c.watchdogWg.Done()
+			c.watchReceiveStream(streamCtx, cancelStream)
+		}()
 
 		// start receiving messages from the Signal stream (from other peers through signal)
 		err = c.receive(stream)
+
+		cancelStream()
+		c.watchdogWg.Wait()
+
 		if err != nil {
 			// Check the parent context, not streamCtx: a watchdog-triggered
 			// cancelStream must reconnect, only a parent cancel is shutdown.
@@ -439,6 +455,16 @@ func (c *GrpcClient) idleSinceReceive() time.Duration {
 	return time.Since(time.Unix(0, c.lastReceived.Load()))
 }
 
+// receiveAlive reports whether the receive stream shows liveness: it delivered a
+// frame within the inactivity threshold, or the receive loop is currently parked
+// handing a message to a busy decryption worker. In the latter case the loop has
+// stopped calling Recv, so the stream looks silent while being healthy, and
+// reconnecting would not help, so the watchdog must treat it as alive.
+func (c *GrpcClient) receiveAlive() bool {
+	return c.idleSinceReceive() < receiveInactivityThreshold ||
+		c.receiveHandoffBlocked.Load()
+}
+
 // watchReceiveStream guards against a receive stream that is transport-alive but
 // no longer delivering messages. While the stream is idle past
 // receiveInactivityThreshold it sends a self-addressed probe that the Signal
@@ -455,7 +481,7 @@ func (c *GrpcClient) watchReceiveStream(ctx context.Context, cancelStream contex
 		case <-ctx.Done():
 			return
 		case <-ticker.C:
-			if c.idleSinceReceive() < receiveInactivityThreshold {
+			if c.receiveAlive() {
 				probeSentAt = time.Time{}
 				continue
 			}
@@ -517,9 +543,14 @@ func (c *GrpcClient) receive(stream proto.SignalExchange_ConnectStreamClient) er
 			continue
 		}
 
+		// The handoff blocks while the worker is busy, which parks this loop and
+		// stops Recv. Flag it so the watchdog does not read the resulting silence
+		// as a dead stream.
+		c.receiveHandoffBlocked.Store(true)
 		if err := c.decryptionWorker.AddMsg(c.ctx, msg); err != nil {
 			log.Errorf("failed to add message to decryption worker: %v", err)
 		}
+		c.receiveHandoffBlocked.Store(false)
 	}
 }
 

shared/signal/client/watchdog_test.go

@@ -82,3 +82,27 @@ func TestReceiveProbeRoundTrips(t *testing.T) {
 		t.Fatal("self-addressed heartbeat did not round-trip back through the signal server")
 	}
 }
+
+// TestReceiveAliveTreatsHandoffBlockAsLiveness reproduces the false positive
+// where a busy decryption worker parks the receive loop on the worker handoff,
+// so Recv (and markReceived) stops firing even though the stream is healthy.
+// With the receive stream silent past the inactivity threshold but the loop
+// blocked on handoff, the watchdog must consider the stream alive rather than
+// tear it down (reconnecting feeds the same worker and would not help).
+func TestReceiveAliveTreatsHandoffBlockAsLiveness(t *testing.T) {
+	c := &GrpcClient{}
+
+	// Receive stream silent and the loop not blocked on handoff: genuinely stalled.
+	c.lastReceived.Store(time.Now().Add(-2 * receiveInactivityThreshold).UnixNano())
+	require.False(t, c.receiveAlive(), "silent stream with the receive loop idle must be treated as stalled")
+
+	// Receive stream silent but the loop is parked handing a message to a busy
+	// worker: self-inflicted backpressure, not a dead stream, must not tear down.
+	c.receiveHandoffBlocked.Store(true)
+	require.True(t, c.receiveAlive(), "a receive loop blocked on worker handoff must keep the stream alive")
+
+	// Handoff drained, loop back to reading, a frame just arrived: alive via the receive path.
+	c.receiveHandoffBlocked.Store(false)
+	c.markReceived()
+	require.True(t, c.receiveAlive(), "a freshly received frame must keep the stream alive")
+}