[management] Account-scoped ephemeral peer cleanup

Replace the per-peer linked list with a per-account map keyed by
accountID. Each entry holds only the latest disconnect timestamp we
have observed for that account and a single timer that fires the next
sweep. Sweeps query the database for the authoritative stale set,
batch the deletes through peers.Manager.DeletePeers, then drop the
account from the tracker when lastDisc + lifeTime <= now (else
re-arm at horizon + cleanupWindow).

The drop rule is the entire termination story: an account stays
tracked only while OnPeerDisconnected keeps refreshing the
timestamp. There is no internal feedback loop that can advance
lastDisc on its own, so once disconnects stop the account drops in
at most one sweep.

A timestamp beats the ref-counter alternative because the counter
drifts positive in three real situations the cleanup loop has no
signal for: peers deleted via the API while offline, peers that
reconnect within the lifetime window, and management restarts. The
timestamp design never claims to know the size of the stale set —
it only knows the latest disconnect we observed and uses that to
bound when it is safe to drop the account.

OnPeerConnected becomes a no-op. The sweep query already filters
reconnected peers at the database level (peer_status_connected =
false in the WHERE clause), so there is nothing the in-memory
tracker needs to do on reconnect. The interface method is preserved
for call-site compatibility.

LoadInitialPeers no longer runs the catch-up query synchronously.
It schedules a deferred load via time.AfterFunc at a random delay
between 8 and 10 minutes. Without the jitter, every management
replica in a fleet-wide deploy would issue the catch-up query
simultaneously. The catch-up itself is one GROUP BY against the
peers table:
```sql
  SELECT account_id, MAX(peer_status_last_seen)
  FROM peers
  WHERE ephemeral = true AND peer_status_connected = false
  GROUP BY account_id
```
For each row the tracker seeds an entry and arms a sweep at
max(now, last_seen + lifeTime) + cleanupWindow — so accounts whose
backlog is already stale get cleaned soon after the delay elapses,
and accounts that disconnected recently wait the remaining window.
OnPeerDisconnected calls that arrive during the delay window seed
the tracker live, and the catch-up query skips accounts that are
already tracked.

Stop() cancels both the deferred initial-load timer and every
per-account sweep timer, and flips a stopped flag so subsequent
OnPeerDisconnected calls are ignored. This makes restarts and test
teardown clean.

Two new store methods:
  GetStaleEphemeralPeerIDsForAccount(ctx, accountID, olderThan)
  GetEphemeralAccountsLastDisconnect(ctx)
Both are scoped, indexable queries that the existing peers table
supports without schema changes.

The pending metric is renamed from
management.ephemeral.peers.pending to
management.ephemeral.accounts.tracked to reflect the new semantics
(it now counts accounts on the cleanup list, not peers). Method
names on the metrics type are unchanged so no production call site
has to move. No new metric labels, no per-account cardinality.

The algorithm was validated against an in-memory SQLite peers
table through an 11-scenario prototype kept under proto/, including
pathological-churn and 4-hour randomized simulations. All scenarios
terminate; max observed per-account sweep rate stays bounded near
the lifeTime + cleanupWindow cadence even under sustained
disconnect churn.

Verification: go build, go vet, race-clean tests across the
ephemeral, store, and telemetry packages, plus a clean
golangci-lint pass on the touched packages.

client/cmd/status.go

@@ -43,16 +43,16 @@ func init() {
 	ipsFilterMap = make(map[string]struct{})
 	prefixNamesFilterMap = make(map[string]struct{})
 	statusCmd.PersistentFlags().BoolVarP(&detailFlag, "detail", "d", false, "display detailed status information in human-readable format")
-	statusCmd.PersistentFlags().BoolVar(&jsonFlag, "json", false, "display detailed status information in json format")
-	statusCmd.PersistentFlags().BoolVar(&yamlFlag, "yaml", false, "display detailed status information in yaml format")
-	statusCmd.PersistentFlags().BoolVar(&ipv4Flag, "ipv4", false, "display only NetBird IPv4 of this peer, e.g., --ipv4 will output 100.64.0.33")
-	statusCmd.PersistentFlags().BoolVar(&ipv6Flag, "ipv6", false, "display only NetBird IPv6 of this peer")
+	statusCmd.PersistentFlags().BoolVarP(&jsonFlag, "json", "j", false, "display detailed status information in json format")
+	statusCmd.PersistentFlags().BoolVarP(&yamlFlag, "yaml", "y", false, "display detailed status information in yaml format")
+	statusCmd.PersistentFlags().BoolVarP(&ipv4Flag, "ipv4", "4", false, "display only NetBird IPv4 of this peer, e.g., --ipv4 will output 100.64.0.33")
+	statusCmd.PersistentFlags().BoolVarP(&ipv6Flag, "ipv6", "6", false, "display only NetBird IPv6 of this peer")
 	statusCmd.MarkFlagsMutuallyExclusive("detail", "json", "yaml", "ipv4", "ipv6")
-	statusCmd.PersistentFlags().StringSliceVar(&ipsFilter, "filter-by-ips", []string{}, "filters the detailed output by a list of one or more IPs (v4 or v6), e.g., --filter-by-ips 100.64.0.100,fd00::1")
-	statusCmd.PersistentFlags().StringSliceVar(&prefixNamesFilter, "filter-by-names", []string{}, "filters the detailed output by a list of one or more peer FQDN or hostnames, e.g., --filter-by-names peer-a,peer-b.netbird.cloud")
-	statusCmd.PersistentFlags().StringVar(&statusFilter, "filter-by-status", "", "filters the detailed output by connection status(idle|connecting|connected), e.g., --filter-by-status connected")
-	statusCmd.PersistentFlags().StringVar(&connectionTypeFilter, "filter-by-connection-type", "", "filters the detailed output by connection type (P2P|Relayed), e.g., --filter-by-connection-type P2P")
-	statusCmd.PersistentFlags().StringVar(&checkFlag, "check", "", "run a health check and exit with code 0 on success, 1 on failure (live|ready|startup)")
+	statusCmd.PersistentFlags().StringSliceVarP(&ipsFilter, "filter-by-ips", "I", []string{}, "filters the detailed output by a list of one or more IPs (v4 or v6), e.g., --filter-by-ips 100.64.0.100,fd00::1")
+	statusCmd.PersistentFlags().StringSliceVarP(&prefixNamesFilter, "filter-by-names", "N", []string{}, "filters the detailed output by a list of one or more peer FQDN or hostnames, e.g., --filter-by-names peer-a,peer-b.netbird.cloud")
+	statusCmd.PersistentFlags().StringVarP(&statusFilter, "filter-by-status", "S", "", "filters the detailed output by connection status(idle|connecting|connected), e.g., --filter-by-status connected")
+	statusCmd.PersistentFlags().StringVarP(&connectionTypeFilter, "filter-by-connection-type", "T", "", "filters the detailed output by connection type (P2P|Relayed), e.g., --filter-by-connection-type P2P")
+	statusCmd.PersistentFlags().StringVarP(&checkFlag, "check", "C", "", "run a health check and exit with code 0 on success, 1 on failure (live|ready|startup)")
 }
 
 func statusFunc(cmd *cobra.Command, args []string) error {

client/embed/embed.go

@@ -336,7 +336,7 @@ func (c *Client) ListenTCP(address string) (net.Listener, error) {
 	if err != nil {
 		return nil, fmt.Errorf("split host port: %w", err)
 	}
-	listenAddr := fmt.Sprintf("%s:%s", addr, port)
+	listenAddr := net.JoinHostPort(addr.String(), port)
 
 	tcpAddr, err := net.ResolveTCPAddr("tcp", listenAddr)
 	if err != nil {
@@ -357,7 +357,7 @@ func (c *Client) ListenUDP(address string) (net.PacketConn, error) {
 	if err != nil {
 		return nil, fmt.Errorf("split host port: %w", err)
 	}
-	listenAddr := fmt.Sprintf("%s:%s", addr, port)
+	listenAddr := net.JoinHostPort(addr.String(), port)
 
 	udpAddr, err := net.ResolveUDPAddr("udp", listenAddr)
 	if err != nil {

client/installer.nsis

@@ -260,15 +260,23 @@ WriteRegStr ${REG_ROOT} "${UNINSTALL_PATH}"  "Publisher" "${COMP_NAME}"
 
 WriteRegStr ${REG_ROOT} "${UI_REG_APP_PATH}" "" "$INSTDIR\${UI_APP_EXE}"
 
-; Create autostart registry entry based on checkbox
+; Drop Run, App Paths and Uninstall entries left in the 32-bit registry view
+; or HKCU by legacy installers.
+DetailPrint "Cleaning legacy 32-bit / HKCU entries..."
+DeleteRegValue HKCU "${AUTOSTART_REG_KEY}" "${APP_NAME}"
+SetRegView 32
+DeleteRegValue HKLM "${AUTOSTART_REG_KEY}" "${APP_NAME}"
+DeleteRegKey HKLM "${REG_APP_PATH}"
+DeleteRegKey HKLM "${UI_REG_APP_PATH}"
+DeleteRegKey HKLM "${UNINSTALL_PATH}"
+SetRegView 64
+
 DetailPrint "Autostart enabled: $AutostartEnabled"
 ${If} $AutostartEnabled == "1"
     WriteRegStr HKLM "${AUTOSTART_REG_KEY}" "${APP_NAME}" '"$INSTDIR\${UI_APP_EXE}.exe"'
     DetailPrint "Added autostart registry entry: $INSTDIR\${UI_APP_EXE}.exe"
 ${Else}
     DeleteRegValue HKLM "${AUTOSTART_REG_KEY}" "${APP_NAME}"
-    ; Legacy: pre-HKLM installs wrote to HKCU; clean that up too.
-    DeleteRegValue HKCU "${AUTOSTART_REG_KEY}" "${APP_NAME}"
     DetailPrint "Autostart not enabled by user"
 ${EndIf}
 
@@ -299,11 +307,16 @@ ExecWait '"$INSTDIR\${MAIN_APP_EXE}" service uninstall'
 DetailPrint "Terminating Netbird UI process..."
 ExecWait `taskkill /im ${UI_APP_EXE}.exe /f`
 
-; Remove autostart registry entry
+; Remove autostart entries from every view a previous installer may have used.
 DetailPrint "Removing autostart registry entry if exists..."
 DeleteRegValue HKLM "${AUTOSTART_REG_KEY}" "${APP_NAME}"
-; Legacy: pre-HKLM installs wrote to HKCU; clean that up too.
 DeleteRegValue HKCU "${AUTOSTART_REG_KEY}" "${APP_NAME}"
+SetRegView 32
+DeleteRegValue HKLM "${AUTOSTART_REG_KEY}" "${APP_NAME}"
+DeleteRegKey HKLM "${REG_APP_PATH}"
+DeleteRegKey HKLM "${UI_REG_APP_PATH}"
+DeleteRegKey HKLM "${UNINSTALL_PATH}"
+SetRegView 64
 
 ; Handle data deletion based on checkbox
 DetailPrint "Checking if user requested data deletion..."

client/internal/connect.go

@@ -116,7 +116,6 @@ func (c *ConnectClient) RunOniOS(
 	fileDescriptor int32,
 	networkChangeListener listener.NetworkChangeListener,
 	dnsManager dns.IosDnsManager,
-	dnsAddresses []netip.AddrPort,
 	stateFilePath string,
 ) error {
 	// Set GC percent to 5% to reduce memory usage as iOS only allows 50MB of memory for the extension.
@@ -126,7 +125,6 @@ func (c *ConnectClient) RunOniOS(
 		FileDescriptor:        fileDescriptor,
 		NetworkChangeListener: networkChangeListener,
 		DnsManager:            dnsManager,
-		HostDNSAddresses:      dnsAddresses,
 		StateFilePath:         stateFilePath,
 	}
 	return c.run(mobileDependency, nil, "")

client/internal/debug/debug.go

@@ -45,8 +45,11 @@ netbird.out: Most recent, anonymized stdout log file of the NetBird client.
 routes.txt: Detailed system routing table in tabular format including destination, gateway, interface, metrics, and protocol information, if --system-info flag was provided.
 interfaces.txt: Anonymized network interface information, if --system-info flag was provided.
 ip_rules.txt: Detailed IP routing rules in tabular format including priority, source, destination, interfaces, table, and action information (Linux only), if --system-info flag was provided.
-iptables.txt: Anonymized iptables rules with packet counters, if --system-info flag was provided.
-nftables.txt: Anonymized nftables rules with packet counters, if --system-info flag was provided.
+iptables.txt: Anonymized iptables (IPv4) rules with packet counters, if --system-info flag was provided.
+ip6tables.txt: Anonymized ip6tables (IPv6) rules with packet counters, if --system-info flag was provided.
+ipset.txt: Anonymized ipset list output, if --system-info flag was provided.
+nftables.txt: Anonymized nftables rules with packet counters across all families (ip, ip6, inet, etc.), if --system-info flag was provided.
+sysctls.txt: Forwarding, reverse-path filter, source-validation, and conntrack accounting sysctl values that the NetBird client may read or modify, if --system-info flag was provided (Linux only).
 resolv.conf: DNS resolver configuration from /etc/resolv.conf (Unix systems only), if --system-info flag was provided.
 scutil_dns.txt: DNS configuration from scutil --dns (macOS only), if --system-info flag was provided.
 resolved_domains.txt: Anonymized resolved domain IP addresses from the status recorder.
@@ -165,22 +168,33 @@ The config.txt file contains anonymized configuration information of the NetBird
 Other non-sensitive configuration options are included without anonymization.
 
 Firewall Rules (Linux only)
-The bundle includes two separate firewall rule files:
+The bundle includes the following firewall-related files:
 
 iptables.txt:
-- Complete iptables ruleset with packet counters using 'iptables -v -n -L'
+- IPv4 iptables ruleset with packet counters using 'iptables-save' and 'iptables -v -n -L'
 - Includes all tables (filter, nat, mangle, raw, security)
 - Shows packet and byte counters for each rule
 - All IP addresses are anonymized
 - Chain names, table names, and other non-sensitive information remain unchanged
 
+ip6tables.txt:
+- IPv6 ip6tables ruleset with packet counters using 'ip6tables-save' and 'ip6tables -v -n -L'
+- Same table coverage and anonymization as iptables.txt
+- Omitted when ip6tables is not installed or no IPv6 rules are present
+
+ipset.txt:
+- Output of 'ipset list' (family-agnostic)
+- IP addresses are anonymized; set names and types remain unchanged
+
 nftables.txt:
-- Complete nftables ruleset obtained via 'nft -a list ruleset'
+- Complete nftables ruleset across all families (ip, ip6, inet, arp, bridge, netdev) via 'nft -a list ruleset'
 - Includes rule handle numbers and packet counters
-- All tables, chains, and rules are included
-- Shows packet and byte counters for each rule
-- All IP addresses are anonymized
-- Chain names, table names, and other non-sensitive information remain unchanged
+- All IP addresses are anonymized; chain/table names remain unchanged
+
+sysctls.txt:
+- Forwarding (IPv4 + IPv6, global and per-interface), reverse-path filter, source-validation, conntrack accounting, and TCP-related sysctls that netbird may read or modify
+- Per-interface keys are enumerated from /proc/sys/net/ipv{4,6}/conf
+- Interface names anonymized when --anonymize is set
 
 IP Rules (Linux only)
 The ip_rules.txt file contains detailed IP routing rule information:
@@ -412,6 +426,10 @@ func (g *BundleGenerator) addSystemInfo() {
 		log.Errorf("failed to add firewall rules to debug bundle: %v", err)
 	}
 
+	if err := g.addSysctls(); err != nil {
+		log.Errorf("failed to add sysctls to debug bundle: %v", err)
+	}
+
 	if err := g.addDNSInfo(); err != nil {
 		log.Errorf("failed to add DNS info to debug bundle: %v", err)
 	}

client/internal/debug/debug_linux.go

@@ -124,15 +124,18 @@ func getSystemdLogs(serviceName string) (string, error) {
 // addFirewallRules collects and adds firewall rules to the archive
 func (g *BundleGenerator) addFirewallRules() error {
 	log.Info("Collecting firewall rules")
-	iptablesRules, err := collectIPTablesRules()
+	g.addIPTablesRulesToBundle("iptables-save", "iptables", "iptables.txt")
+	g.addIPTablesRulesToBundle("ip6tables-save", "ip6tables", "ip6tables.txt")
+
+	ipsetOutput, err := collectIPSets()
 	if err != nil {
-		log.Warnf("Failed to collect iptables rules: %v", err)
+		log.Warnf("Failed to collect ipset information: %v", err)
 	} else {
 		if g.anonymize {
-			iptablesRules = g.anonymizer.AnonymizeString(iptablesRules)
+			ipsetOutput = g.anonymizer.AnonymizeString(ipsetOutput)
 		}
-		if err := g.addFileToZip(strings.NewReader(iptablesRules), "iptables.txt"); err != nil {
-			log.Warnf("Failed to add iptables rules to bundle: %v", err)
+		if err := g.addFileToZip(strings.NewReader(ipsetOutput), "ipset.txt"); err != nil {
+			log.Warnf("Failed to add ipset output to bundle: %v", err)
 		}
 	}
 
@@ -151,44 +154,65 @@ func (g *BundleGenerator) addFirewallRules() error {
 	return nil
 }
 
-// collectIPTablesRules collects rules using both iptables-save and verbose listing
-func collectIPTablesRules() (string, error) {
-	var builder strings.Builder
-
-	saveOutput, err := collectIPTablesSave()
+// addIPTablesRulesToBundle collects iptables/ip6tables rules and writes them to the bundle.
+func (g *BundleGenerator) addIPTablesRulesToBundle(saveBin, listBin, filename string) {
+	rules, err := collectIPTablesRules(saveBin, listBin)
 	if err != nil {
-		log.Warnf("Failed to collect iptables rules using iptables-save: %v", err)
-	} else {
-		builder.WriteString("=== iptables-save output ===\n")
+		log.Warnf("Failed to collect %s rules: %v", listBin, err)
+		return
+	}
+	if g.anonymize {
+		rules = g.anonymizer.AnonymizeString(rules)
+	}
+	if err := g.addFileToZip(strings.NewReader(rules), filename); err != nil {
+		log.Warnf("Failed to add %s rules to bundle: %v", listBin, err)
+	}
+}
+
+// collectIPTablesRules collects rules using both <saveBin> and verbose listing via <listBin>.
+// Returns an error when neither command produced any output (e.g. the binary is missing),
+// so the caller can skip writing an empty file.
+func collectIPTablesRules(saveBin, listBin string) (string, error) {
+	var builder strings.Builder
+	var collected bool
+	var firstErr error
+
+	saveOutput, err := runCommand(saveBin)
+	switch {
+	case err != nil:
+		firstErr = err
+		log.Warnf("Failed to collect %s output: %v", saveBin, err)
+	case strings.TrimSpace(saveOutput) == "":
+		log.Debugf("%s produced no output, skipping", saveBin)
+	default:
+		builder.WriteString(fmt.Sprintf("=== %s output ===\n", saveBin))
 		builder.WriteString(saveOutput)
 		builder.WriteString("\n")
+		collected = true
 	}
 
-	ipsetOutput, err := collectIPSets()
-	if err != nil {
-		log.Warnf("Failed to collect ipset information: %v", err)
-	} else {
-		builder.WriteString("=== ipset list output ===\n")
-		builder.WriteString(ipsetOutput)
-		builder.WriteString("\n")
-	}
-
-	builder.WriteString("=== iptables -v -n -L output ===\n")
+	listHeader := fmt.Sprintf("=== %s -v -n -L output ===\n", listBin)
+	builder.WriteString(listHeader)
 
 	tables := []string{"filter", "nat", "mangle", "raw", "security"}
-
 	for _, table := range tables {
-		builder.WriteString(fmt.Sprintf("*%s\n", table))
-
-		stats, err := getTableStatistics(table)
+		stats, err := runCommand(listBin, "-v", "-n", "-L", "-t", table)
 		if err != nil {
-			log.Warnf("Failed to get statistics for table %s: %v", table, err)
+			if firstErr == nil {
+				firstErr = err
+			}
+			log.Warnf("Failed to get %s statistics for table %s: %v", listBin, table, err)
 			continue
 		}
+		builder.WriteString(fmt.Sprintf("*%s\n", table))
 		builder.WriteString(stats)
 		builder.WriteString("\n")
+		collected = true
 	}
 
+	if !collected {
+		return "", fmt.Errorf("collect %s rules: %w", listBin, firstErr)
+	}
 	return builder.String(), nil
 }
 
@@ -214,34 +238,15 @@ func collectIPSets() (string, error) {
 	return ipsets, nil
 }
 
-// collectIPTablesSave uses iptables-save to get rule definitions
-func collectIPTablesSave() (string, error) {
-	cmd := exec.Command("iptables-save")
+// runCommand executes a command and returns its stdout, wrapping stderr in the error on failure.
+func runCommand(name string, args ...string) (string, error) {
+	cmd := exec.Command(name, args...)
 	var stdout, stderr bytes.Buffer
 	cmd.Stdout = &stdout
 	cmd.Stderr = &stderr
 
 	if err := cmd.Run(); err != nil {
-		return "", fmt.Errorf("execute iptables-save: %w (stderr: %s)", err, stderr.String())
-	}
-
-	rules := stdout.String()
-	if strings.TrimSpace(rules) == "" {
-		return "", fmt.Errorf("no iptables rules found")
-	}
-
-	return rules, nil
-}
-
-// getTableStatistics gets verbose statistics for an entire table using iptables command
-func getTableStatistics(table string) (string, error) {
-	cmd := exec.Command("iptables", "-v", "-n", "-L", "-t", table)
-	var stdout, stderr bytes.Buffer
-	cmd.Stdout = &stdout
-	cmd.Stderr = &stderr
-
-	if err := cmd.Run(); err != nil {
-		return "", fmt.Errorf("execute iptables -v -n -L: %w (stderr: %s)", err, stderr.String())
+		return "", fmt.Errorf("execute %s: %w (stderr: %s)", name, err, stderr.String())
 	}
 
 	return stdout.String(), nil
@@ -804,3 +809,91 @@ func formatSetKeyType(keyType nftables.SetDatatype) string {
 		return fmt.Sprintf("type-%v", keyType)
 	}
 }
+
+// addSysctls collects forwarding and netbird-managed sysctl values and writes them to the bundle.
+func (g *BundleGenerator) addSysctls() error {
+	log.Info("Collecting sysctls")
+	content := collectSysctls()
+	if g.anonymize {
+		content = g.anonymizer.AnonymizeString(content)
+	}
+	if err := g.addFileToZip(strings.NewReader(content), "sysctls.txt"); err != nil {
+		return fmt.Errorf("add sysctls to bundle: %w", err)
+	}
+	return nil
+}
+
+// collectSysctls reads every sysctl that the netbird client may modify, plus
+// global IPv4/IPv6 forwarding, and returns a formatted dump grouped by topic.
+// Per-interface values are enumerated by listing /proc/sys/net/ipv{4,6}/conf.
+func collectSysctls() string {
+	var builder strings.Builder
+
+	writeSysctlGroup(&builder, "forwarding", []string{
+		"net.ipv4.ip_forward",
+		"net.ipv6.conf.all.forwarding",
+		"net.ipv6.conf.default.forwarding",
+	})
+	writeSysctlGroup(&builder, "ipv4 per-interface forwarding", listInterfaceSysctls("ipv4", "forwarding"))
+	writeSysctlGroup(&builder, "ipv6 per-interface forwarding", listInterfaceSysctls("ipv6", "forwarding"))
+	writeSysctlGroup(&builder, "rp_filter", append(
+		[]string{"net.ipv4.conf.all.rp_filter", "net.ipv4.conf.default.rp_filter"},
+		listInterfaceSysctls("ipv4", "rp_filter")...,
+	))
+	writeSysctlGroup(&builder, "src_valid_mark", append(
+		[]string{"net.ipv4.conf.all.src_valid_mark", "net.ipv4.conf.default.src_valid_mark"},
+		listInterfaceSysctls("ipv4", "src_valid_mark")...,
+	))
+	writeSysctlGroup(&builder, "conntrack", []string{
+		"net.netfilter.nf_conntrack_acct",
+		"net.netfilter.nf_conntrack_tcp_loose",
+	})
+	writeSysctlGroup(&builder, "tcp", []string{
+		"net.ipv4.tcp_tw_reuse",
+	})
+
+	return builder.String()
+}
+
+func writeSysctlGroup(builder *strings.Builder, title string, keys []string) {
+	builder.WriteString(fmt.Sprintf("=== %s ===\n", title))
+	for _, key := range keys {
+		value, err := readSysctl(key)
+		if err != nil {
+			builder.WriteString(fmt.Sprintf("%s = <error: %v>\n", key, err))
+			continue
+		}
+		builder.WriteString(fmt.Sprintf("%s = %s\n", key, value))
+	}
+	builder.WriteString("\n")
+}
+
+// listInterfaceSysctls returns net.ipvX.conf.<iface>.<leaf> keys for every
+// interface present in /proc/sys/net/ipvX/conf, skipping "all" and "default"
+// (callers add those explicitly so they appear first).
+func listInterfaceSysctls(family, leaf string) []string {
+	dir := fmt.Sprintf("/proc/sys/net/%s/conf", family)
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		return nil
+	}
+	var keys []string
+	for _, e := range entries {
+		name := e.Name()
+		if name == "all" || name == "default" {
+			continue
+		}
+		keys = append(keys, fmt.Sprintf("net.%s.conf.%s.%s", family, name, leaf))
+	}
+	sort.Strings(keys)
+	return keys
+}
+
+func readSysctl(key string) (string, error) {
+	path := fmt.Sprintf("/proc/sys/%s", strings.ReplaceAll(key, ".", "/"))
+	value, err := os.ReadFile(path)
+	if err != nil {
+		return "", err
+	}
+	return strings.TrimSpace(string(value)), nil
+}

client/internal/debug/debug_nonlinux.go

@@ -17,3 +17,8 @@ func (g *BundleGenerator) addIPRules() error {
 	// IP rules are only supported on Linux
 	return nil
 }
+
+func (g *BundleGenerator) addSysctls() error {
+	// Sysctl collection is only supported on Linux
+	return nil
+}

client/internal/dns/host.go

@@ -16,6 +16,10 @@ type hostManager interface {
 	restoreHostDNS() error
 	supportCustomPort() bool
 	string() string
+	// getOriginalNameservers returns the OS-side resolvers used as PriorityFallback
+	// upstreams: pre-takeover snapshots on desktop, the OS-pushed list on Android,
+	// hardcoded Quad9 on iOS, nil for noop / mock.
+	getOriginalNameservers() []netip.Addr
 }
 
 type SystemDNSSettings struct {
@@ -131,3 +135,11 @@ func (n noopHostConfigurator) supportCustomPort() bool {
 func (n noopHostConfigurator) string() string {
 	return "noop"
 }
+
+func (n noopHostConfigurator) getOriginalNameservers() []netip.Addr {
+	return nil
+}
+
+func (m *mockHostConfigurator) getOriginalNameservers() []netip.Addr {
+	return nil
+}

client/internal/dns/host_android.go

@@ -1,14 +1,20 @@
 package dns
 
 import (
+	"net/netip"
+
 	"github.com/netbirdio/netbird/client/internal/statemanager"
 )
 
+// androidHostManager is a noop on the OS side (Android's VPN service handles
+// DNS for us) but tracks the OS-reported resolver list pushed via
+// OnUpdatedHostDNSServer so it can serve as the fallback nameserver source.
 type androidHostManager struct {
+	holder *hostsDNSHolder
 }
 
-func newHostManager() (*androidHostManager, error) {
-	return &androidHostManager{}, nil
+func newHostManager(holder *hostsDNSHolder) (*androidHostManager, error) {
+	return &androidHostManager{holder: holder}, nil
 }
 
 func (a androidHostManager) applyDNSConfig(HostDNSConfig, *statemanager.Manager) error {
@@ -26,3 +32,12 @@ func (a androidHostManager) supportCustomPort() bool {
 func (a androidHostManager) string() string {
 	return "none"
 }
+
+func (a androidHostManager) getOriginalNameservers() []netip.Addr {
+	hosts := a.holder.get()
+	out := make([]netip.Addr, 0, len(hosts))
+	for ap := range hosts {
+		out = append(out, ap.Addr())
+	}
+	return out
+}

client/internal/dns/host_ios.go

@@ -3,6 +3,7 @@ package dns
 import (
 	"encoding/json"
 	"fmt"
+	"net/netip"
 
 	log "github.com/sirupsen/logrus"
 
@@ -20,6 +21,14 @@ func newHostManager(dnsManager IosDnsManager) (*iosHostManager, error) {
 	}, nil
 }
 
+func (a iosHostManager) getOriginalNameservers() []netip.Addr {
+	// Quad9 v4+v6: 9.9.9.9, 2620:fe::fe.
+	return []netip.Addr{
+		netip.AddrFrom4([4]byte{9, 9, 9, 9}),
+		netip.AddrFrom16([16]byte{0x26, 0x20, 0x00, 0xfe, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0xfe}),
+	}
+}
+
 func (a iosHostManager) applyDNSConfig(config HostDNSConfig, _ *statemanager.Manager) error {
 	jsonData, err := json.Marshal(config)
 	if err != nil {

client/internal/dns/host_windows.go

@@ -7,6 +7,7 @@ import (
 	"io"
 	"net/netip"
 	"os/exec"
+	"slices"
 	"strings"
 	"syscall"
 	"time"
@@ -44,9 +45,11 @@ const (
 
 	nrptMaxDomainsPerRule = 50
 
-	interfaceConfigPath          = `SYSTEM\CurrentControlSet\Services\Tcpip\Parameters\Interfaces`
-	interfaceConfigNameServerKey = "NameServer"
-	interfaceConfigSearchListKey = "SearchList"
+	interfaceConfigPath           = `SYSTEM\CurrentControlSet\Services\Tcpip\Parameters\Interfaces`
+	interfaceConfigPathV6         = `SYSTEM\CurrentControlSet\Services\Tcpip6\Parameters\Interfaces`
+	interfaceConfigNameServerKey  = "NameServer"
+	interfaceConfigDhcpNameSrvKey = "DhcpNameServer"
+	interfaceConfigSearchListKey  = "SearchList"
 
 	// Network interface DNS registration settings
 	disableDynamicUpdateKey           = "DisableDynamicUpdate"
@@ -67,10 +70,11 @@ const (
 )
 
 type registryConfigurator struct {
-	guid           string
-	routingAll     bool
-	gpo            bool
-	nrptEntryCount int
+	guid            string
+	routingAll      bool
+	gpo             bool
+	nrptEntryCount  int
+	origNameservers []netip.Addr
 }
 
 func newHostManager(wgInterface WGIface) (*registryConfigurator, error) {
@@ -94,6 +98,17 @@ func newHostManager(wgInterface WGIface) (*registryConfigurator, error) {
 		gpo:  useGPO,
 	}
 
+	origNameservers, err := configurator.captureOriginalNameservers()
+	switch {
+	case err != nil:
+		log.Warnf("capture original nameservers from non-WG adapters: %v", err)
+	case len(origNameservers) == 0:
+		log.Warnf("no original nameservers captured from non-WG adapters; DNS fallback will be empty")
+	default:
+		log.Debugf("captured %d original nameservers from non-WG adapters: %v", len(origNameservers), origNameservers)
+	}
+	configurator.origNameservers = origNameservers
+
 	if err := configurator.configureInterface(); err != nil {
 		log.Errorf("failed to configure interface settings: %v", err)
 	}
@@ -101,6 +116,98 @@ func newHostManager(wgInterface WGIface) (*registryConfigurator, error) {
 	return configurator, nil
 }
 
+// captureOriginalNameservers reads DNS addresses from every Tcpip(6) interface
+// registry key except the WG adapter. v4 and v6 servers live in separate
+// hives (Tcpip vs Tcpip6) keyed by the same interface GUID.
+func (r *registryConfigurator) captureOriginalNameservers() ([]netip.Addr, error) {
+	seen := make(map[netip.Addr]struct{})
+	var out []netip.Addr
+	var merr *multierror.Error
+	for _, root := range []string{interfaceConfigPath, interfaceConfigPathV6} {
+		addrs, err := r.captureFromTcpipRoot(root)
+		if err != nil {
+			merr = multierror.Append(merr, fmt.Errorf("%s: %w", root, err))
+			continue
+		}
+		for _, addr := range addrs {
+			if _, dup := seen[addr]; dup {
+				continue
+			}
+			seen[addr] = struct{}{}
+			out = append(out, addr)
+		}
+	}
+	return out, nberrors.FormatErrorOrNil(merr)
+}
+
+func (r *registryConfigurator) captureFromTcpipRoot(rootPath string) ([]netip.Addr, error) {
+	root, err := registry.OpenKey(registry.LOCAL_MACHINE, rootPath, registry.READ)
+	if err != nil {
+		return nil, fmt.Errorf("open key: %w", err)
+	}
+	defer closer(root)
+
+	guids, err := root.ReadSubKeyNames(-1)
+	if err != nil {
+		return nil, fmt.Errorf("read subkeys: %w", err)
+	}
+
+	var out []netip.Addr
+	for _, guid := range guids {
+		if strings.EqualFold(guid, r.guid) {
+			continue
+		}
+		out = append(out, readInterfaceNameservers(rootPath, guid)...)
+	}
+	return out, nil
+}
+
+func readInterfaceNameservers(rootPath, guid string) []netip.Addr {
+	keyPath := rootPath + "\\" + guid
+	k, err := registry.OpenKey(registry.LOCAL_MACHINE, keyPath, registry.QUERY_VALUE)
+	if err != nil {
+		return nil
+	}
+	defer closer(k)
+
+	// Static NameServer wins over DhcpNameServer for actual resolution.
+	for _, name := range []string{interfaceConfigNameServerKey, interfaceConfigDhcpNameSrvKey} {
+		raw, _, err := k.GetStringValue(name)
+		if err != nil || raw == "" {
+			continue
+		}
+		if out := parseRegistryNameservers(raw); len(out) > 0 {
+			return out
+		}
+	}
+	return nil
+}
+
+func parseRegistryNameservers(raw string) []netip.Addr {
+	var out []netip.Addr
+	for _, field := range strings.FieldsFunc(raw, func(r rune) bool { return r == ',' || r == ' ' || r == '\t' }) {
+		addr, err := netip.ParseAddr(strings.TrimSpace(field))
+		if err != nil {
+			continue
+		}
+		addr = addr.Unmap()
+		if !addr.IsValid() || addr.IsUnspecified() {
+			continue
+		}
+		// Drop unzoned link-local: not routable without a scope id. If
+		// the user wrote "fe80::1%eth0" ParseAddr preserves the zone.
+		if addr.IsLinkLocalUnicast() && addr.Zone() == "" {
+			continue
+		}
+		out = append(out, addr)
+	}
+	return out
+}
+
+func (r *registryConfigurator) getOriginalNameservers() []netip.Addr {
+	return slices.Clone(r.origNameservers)
+}
+
 func (r *registryConfigurator) supportCustomPort() bool {
 	return false
 }

client/internal/dns/hosts_dns_holder.go

@@ -25,6 +25,7 @@ func (h *hostsDNSHolder) set(list []netip.AddrPort) {
 	h.mutex.Unlock()
 }
 
+//nolint:unused
 func (h *hostsDNSHolder) get() map[netip.AddrPort]struct{} {
 	h.mutex.RLock()
 	l := h.unprotectedDNSList

client/internal/dns/local/local.go

@@ -76,8 +76,6 @@ func (d *Resolver) ID() types.HandlerID {
 	return "local-resolver"
 }
 
-func (d *Resolver) ProbeAvailability(context.Context) {}
-
 // ServeDNS handles a DNS request
 func (d *Resolver) ServeDNS(w dns.ResponseWriter, r *dns.Msg) {
 	logger := log.WithFields(log.Fields{

client/internal/dns/mock_server.go

@@ -9,6 +9,7 @@ import (
 
 	dnsconfig "github.com/netbirdio/netbird/client/internal/dns/config"
 	nbdns "github.com/netbirdio/netbird/dns"
+	"github.com/netbirdio/netbird/route"
 	"github.com/netbirdio/netbird/shared/management/domain"
 )
 
@@ -70,10 +71,6 @@ func (m *MockServer) SearchDomains() []string {
 	return make([]string, 0)
 }
 
-// ProbeAvailability mocks implementation of ProbeAvailability from the Server interface
-func (m *MockServer) ProbeAvailability() {
-}
-
 func (m *MockServer) UpdateServerConfig(domains dnsconfig.ServerDomains) error {
 	if m.UpdateServerConfigFunc != nil {
 		return m.UpdateServerConfigFunc(domains)
@@ -85,8 +82,8 @@ func (m *MockServer) PopulateManagementDomain(mgmtURL *url.URL) error {
 	return nil
 }
 
-// SetRouteChecker mock implementation of SetRouteChecker from Server interface
-func (m *MockServer) SetRouteChecker(func(netip.Addr) bool) {
+// SetRouteSources mock implementation of SetRouteSources from Server interface
+func (m *MockServer) SetRouteSources(selected, active func() route.HAMap) {
 	// Mock implementation - no-op
 }
 

client/internal/dns/network_manager_unix.go

@@ -8,6 +8,7 @@ import (
 	"errors"
 	"fmt"
 	"net/netip"
+	"slices"
 	"strings"
 	"time"
 
@@ -32,6 +33,15 @@ const (
 	networkManagerDbusDeviceGetAppliedConnectionMethod                              = networkManagerDbusDeviceInterface + ".GetAppliedConnection"
 	networkManagerDbusDeviceReapplyMethod                                           = networkManagerDbusDeviceInterface + ".Reapply"
 	networkManagerDbusDeviceDeleteMethod                                            = networkManagerDbusDeviceInterface + ".Delete"
+	networkManagerDbusDeviceIp4ConfigProperty                                       = networkManagerDbusDeviceInterface + ".Ip4Config"
+	networkManagerDbusDeviceIp6ConfigProperty                                       = networkManagerDbusDeviceInterface + ".Ip6Config"
+	networkManagerDbusDeviceIfaceProperty                                           = networkManagerDbusDeviceInterface + ".Interface"
+	networkManagerDbusGetDevicesMethod                                              = networkManagerDest + ".GetDevices"
+	networkManagerDbusIp4ConfigInterface                                            = "org.freedesktop.NetworkManager.IP4Config"
+	networkManagerDbusIp6ConfigInterface                                            = "org.freedesktop.NetworkManager.IP6Config"
+	networkManagerDbusIp4ConfigNameserverDataProperty                               = networkManagerDbusIp4ConfigInterface + ".NameserverData"
+	networkManagerDbusIp4ConfigNameserversProperty                                  = networkManagerDbusIp4ConfigInterface + ".Nameservers"
+	networkManagerDbusIp6ConfigNameserversProperty                                  = networkManagerDbusIp6ConfigInterface + ".Nameservers"
 	networkManagerDbusDefaultBehaviorFlag              networkManagerConfigBehavior = 0
 	networkManagerDbusIPv4Key                                                       = "ipv4"
 	networkManagerDbusIPv6Key                                                       = "ipv6"
@@ -51,9 +61,10 @@ var supportedNetworkManagerVersionConstraints = []string{
 }
 
 type networkManagerDbusConfigurator struct {
-	dbusLinkObject dbus.ObjectPath
-	routingAll     bool
-	ifaceName      string
+	dbusLinkObject  dbus.ObjectPath
+	routingAll      bool
+	ifaceName       string
+	origNameservers []netip.Addr
 }
 
 // the types below are based on dbus specification, each field is mapped to a dbus type
@@ -92,10 +103,200 @@ func newNetworkManagerDbusConfigurator(wgInterface string) (*networkManagerDbusC
 
 	log.Debugf("got network manager dbus Link Object: %s from net interface %s", s, wgInterface)
 
-	return &networkManagerDbusConfigurator{
+	c := &networkManagerDbusConfigurator{
 		dbusLinkObject: dbus.ObjectPath(s),
 		ifaceName:      wgInterface,
-	}, nil
+	}
+
+	origNameservers, err := c.captureOriginalNameservers()
+	switch {
+	case err != nil:
+		log.Warnf("capture original nameservers from NetworkManager: %v", err)
+	case len(origNameservers) == 0:
+		log.Warnf("no original nameservers captured from non-WG NetworkManager devices; DNS fallback will be empty")
+	default:
+		log.Debugf("captured %d original nameservers from non-WG NetworkManager devices: %v", len(origNameservers), origNameservers)
+	}
+	c.origNameservers = origNameservers
+	return c, nil
+}
+
+// captureOriginalNameservers reads DNS servers from every NM device's
+// IP4Config / IP6Config except our WG device.
+func (n *networkManagerDbusConfigurator) captureOriginalNameservers() ([]netip.Addr, error) {
+	devices, err := networkManagerListDevices()
+	if err != nil {
+		return nil, fmt.Errorf("list devices: %w", err)
+	}
+
+	seen := make(map[netip.Addr]struct{})
+	var out []netip.Addr
+	for _, dev := range devices {
+		if dev == n.dbusLinkObject {
+			continue
+		}
+		ifaceName := readNetworkManagerDeviceInterface(dev)
+		for _, addr := range readNetworkManagerDeviceDNS(dev) {
+			addr = addr.Unmap()
+			if !addr.IsValid() || addr.IsUnspecified() {
+				continue
+			}
+			// IP6Config.Nameservers is a byte slice without zone info;
+			// reattach the device's interface name so a captured fe80::…
+			// stays routable.
+			if addr.IsLinkLocalUnicast() && ifaceName != "" {
+				addr = addr.WithZone(ifaceName)
+			}
+			if _, dup := seen[addr]; dup {
+				continue
+			}
+			seen[addr] = struct{}{}
+			out = append(out, addr)
+		}
+	}
+	return out, nil
+}
+
+func readNetworkManagerDeviceInterface(devicePath dbus.ObjectPath) string {
+	obj, closeConn, err := getDbusObject(networkManagerDest, devicePath)
+	if err != nil {
+		return ""
+	}
+	defer closeConn()
+	v, err := obj.GetProperty(networkManagerDbusDeviceIfaceProperty)
+	if err != nil {
+		return ""
+	}
+	s, _ := v.Value().(string)
+	return s
+}
+
+func networkManagerListDevices() ([]dbus.ObjectPath, error) {
+	obj, closeConn, err := getDbusObject(networkManagerDest, networkManagerDbusObjectNode)
+	if err != nil {
+		return nil, fmt.Errorf("dbus NetworkManager: %w", err)
+	}
+	defer closeConn()
+	var devs []dbus.ObjectPath
+	if err := obj.Call(networkManagerDbusGetDevicesMethod, dbusDefaultFlag).Store(&devs); err != nil {
+		return nil, err
+	}
+	return devs, nil
+}
+
+func readNetworkManagerDeviceDNS(devicePath dbus.ObjectPath) []netip.Addr {
+	obj, closeConn, err := getDbusObject(networkManagerDest, devicePath)
+	if err != nil {
+		return nil
+	}
+	defer closeConn()
+
+	var out []netip.Addr
+	if path := readNetworkManagerConfigPath(obj, networkManagerDbusDeviceIp4ConfigProperty); path != "" {
+		out = append(out, readIPv4ConfigDNS(path)...)
+	}
+	if path := readNetworkManagerConfigPath(obj, networkManagerDbusDeviceIp6ConfigProperty); path != "" {
+		out = append(out, readIPv6ConfigDNS(path)...)
+	}
+	return out
+}
+
+func readNetworkManagerConfigPath(obj dbus.BusObject, property string) dbus.ObjectPath {
+	v, err := obj.GetProperty(property)
+	if err != nil {
+		return ""
+	}
+	path, ok := v.Value().(dbus.ObjectPath)
+	if !ok || path == "/" {
+		return ""
+	}
+	return path
+}
+
+func readIPv4ConfigDNS(path dbus.ObjectPath) []netip.Addr {
+	obj, closeConn, err := getDbusObject(networkManagerDest, path)
+	if err != nil {
+		return nil
+	}
+	defer closeConn()
+
+	// NameserverData (NM 1.13+) carries strings; older NMs only expose the
+	// legacy uint32 Nameservers property.
+	if out := readIPv4NameserverData(obj); len(out) > 0 {
+		return out
+	}
+	return readIPv4LegacyNameservers(obj)
+}
+
+func readIPv4NameserverData(obj dbus.BusObject) []netip.Addr {
+	v, err := obj.GetProperty(networkManagerDbusIp4ConfigNameserverDataProperty)
+	if err != nil {
+		return nil
+	}
+	entries, ok := v.Value().([]map[string]dbus.Variant)
+	if !ok {
+		return nil
+	}
+	var out []netip.Addr
+	for _, entry := range entries {
+		addrVar, ok := entry["address"]
+		if !ok {
+			continue
+		}
+		s, ok := addrVar.Value().(string)
+		if !ok {
+			continue
+		}
+		if a, err := netip.ParseAddr(s); err == nil {
+			out = append(out, a)
+		}
+	}
+	return out
+}
+
+func readIPv4LegacyNameservers(obj dbus.BusObject) []netip.Addr {
+	v, err := obj.GetProperty(networkManagerDbusIp4ConfigNameserversProperty)
+	if err != nil {
+		return nil
+	}
+	raw, ok := v.Value().([]uint32)
+	if !ok {
+		return nil
+	}
+	out := make([]netip.Addr, 0, len(raw))
+	for _, n := range raw {
+		var b [4]byte
+		binary.LittleEndian.PutUint32(b[:], n)
+		out = append(out, netip.AddrFrom4(b))
+	}
+	return out
+}
+
+func readIPv6ConfigDNS(path dbus.ObjectPath) []netip.Addr {
+	obj, closeConn, err := getDbusObject(networkManagerDest, path)
+	if err != nil {
+		return nil
+	}
+	defer closeConn()
+	v, err := obj.GetProperty(networkManagerDbusIp6ConfigNameserversProperty)
+	if err != nil {
+		return nil
+	}
+	raw, ok := v.Value().([][]byte)
+	if !ok {
+		return nil
+	}
+	out := make([]netip.Addr, 0, len(raw))
+	for _, b := range raw {
+		if a, ok := netip.AddrFromSlice(b); ok {
+			out = append(out, a)
+		}
+	}
+	return out
+}
+
+func (n *networkManagerDbusConfigurator) getOriginalNameservers() []netip.Addr {
+	return slices.Clone(n.origNameservers)
 }
 
 func (n *networkManagerDbusConfigurator) supportCustomPort() bool {

client/internal/dns/server_android.go

@@ -1,5 +1,5 @@
 package dns
 
 func (s *DefaultServer) initialize() (manager hostManager, err error) {
-	return newHostManager()
+	return newHostManager(s.hostsDNSHolder)
 }

client/internal/dns/server_test.go

@@ -6,7 +6,7 @@ import (
 	"net"
 	"net/netip"
 	"os"
-	"strings"
+	"runtime"
 	"testing"
 	"time"
 
@@ -15,6 +15,7 @@ import (
 	log "github.com/sirupsen/logrus"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/mock"
+	"github.com/stretchr/testify/require"
 	"golang.zx2c4.com/wireguard/tun/netstack"
 	"golang.zx2c4.com/wireguard/wgctrl/wgtypes"
 
@@ -31,8 +32,10 @@ import (
 	"github.com/netbirdio/netbird/client/internal/peer"
 	"github.com/netbirdio/netbird/client/internal/statemanager"
 	"github.com/netbirdio/netbird/client/internal/stdnet"
+	"github.com/netbirdio/netbird/client/proto"
 	nbdns "github.com/netbirdio/netbird/dns"
 	"github.com/netbirdio/netbird/formatter"
+	"github.com/netbirdio/netbird/route"
 	"github.com/netbirdio/netbird/shared/management/domain"
 )
 
@@ -101,16 +104,17 @@ func init() {
 	formatter.SetTextFormatter(log.StandardLogger())
 }
 
-func generateDummyHandler(domain string, servers []nbdns.NameServer) *upstreamResolverBase {
+func generateDummyHandler(d string, servers []nbdns.NameServer) *upstreamResolverBase {
 	var srvs []netip.AddrPort
 	for _, srv := range servers {
 		srvs = append(srvs, srv.AddrPort())
 	}
-	return &upstreamResolverBase{
-		domain:          domain,
-		upstreamServers: srvs,
-		cancel:          func() {},
+	u := &upstreamResolverBase{
+		domain: domain.Domain(d),
+		cancel: func() {},
 	}
+	u.addRace(srvs)
+	return u
 }
 
 func TestUpdateDNSServer(t *testing.T) {
@@ -653,74 +657,8 @@ func TestDNSServerStartStop(t *testing.T) {
 	}
 }
 
-func TestDNSServerUpstreamDeactivateCallback(t *testing.T) {
-	hostManager := &mockHostConfigurator{}
-	server := DefaultServer{
-		ctx:           context.Background(),
-		service:       NewServiceViaMemory(&mocWGIface{}),
-		localResolver: local.NewResolver(),
-		handlerChain:  NewHandlerChain(),
-		hostManager:   hostManager,
-		currentConfig: HostDNSConfig{
-			Domains: []DomainConfig{
-				{false, "domain0", false},
-				{false, "domain1", false},
-				{false, "domain2", false},
-			},
-		},
-		statusRecorder: peer.NewRecorder("mgm"),
-	}
-
-	var domainsUpdate string
-	hostManager.applyDNSConfigFunc = func(config HostDNSConfig, statemanager *statemanager.Manager) error {
-		domains := []string{}
-		for _, item := range config.Domains {
-			if item.Disabled {
-				continue
-			}
-			domains = append(domains, item.Domain)
-		}
-		domainsUpdate = strings.Join(domains, ",")
-		return nil
-	}
-
-	deactivate, reactivate := server.upstreamCallbacks(&nbdns.NameServerGroup{
-		Domains: []string{"domain1"},
-		NameServers: []nbdns.NameServer{
-			{IP: netip.MustParseAddr("8.8.0.0"), NSType: nbdns.UDPNameServerType, Port: 53},
-		},
-	}, nil, 0)
-
-	deactivate(nil)
-	expected := "domain0,domain2"
-	domains := []string{}
-	for _, item := range server.currentConfig.Domains {
-		if item.Disabled {
-			continue
-		}
-		domains = append(domains, item.Domain)
-	}
-	got := strings.Join(domains, ",")
-	if expected != got {
-		t.Errorf("expected domains list: %q, got %q", expected, got)
-	}
-
-	reactivate()
-	expected = "domain0,domain1,domain2"
-	domains = []string{}
-	for _, item := range server.currentConfig.Domains {
-		if item.Disabled {
-			continue
-		}
-		domains = append(domains, item.Domain)
-	}
-	got = strings.Join(domains, ",")
-	if expected != got {
-		t.Errorf("expected domains list: %q, got %q", expected, domainsUpdate)
-	}
-}
-
 func TestDNSPermanent_updateHostDNS_emptyUpstream(t *testing.T) {
+	skipUnlessAndroid(t)
 	wgIFace, err := createWgInterfaceWithBind(t)
 	if err != nil {
 		t.Fatal("failed to initialize wg interface")
@@ -748,6 +686,7 @@ func TestDNSPermanent_updateHostDNS_emptyUpstream(t *testing.T) {
 }
 
 func TestDNSPermanent_updateUpstream(t *testing.T) {
+	skipUnlessAndroid(t)
 	wgIFace, err := createWgInterfaceWithBind(t)
 	if err != nil {
 		t.Fatal("failed to initialize wg interface")
@@ -841,6 +780,7 @@ func TestDNSPermanent_updateUpstream(t *testing.T) {
 }
 
 func TestDNSPermanent_matchOnly(t *testing.T) {
+	skipUnlessAndroid(t)
 	wgIFace, err := createWgInterfaceWithBind(t)
 	if err != nil {
 		t.Fatal("failed to initialize wg interface")
@@ -913,6 +853,18 @@ func TestDNSPermanent_matchOnly(t *testing.T) {
 	}
 }
 
+// skipUnlessAndroid marks tests that exercise the mobile-permanent DNS path,
+// which only matches a real production setup on android (NewDefaultServerPermanentUpstream
+// + androidHostManager). On non-android the desktop host manager replaces it
+// during Initialize and the assertion stops making sense. Skipped here until we
+// have an android CI runner.
+func skipUnlessAndroid(t *testing.T) {
+	t.Helper()
+	if runtime.GOOS != "android" {
+		t.Skip("requires android runner; mobile-permanent path doesn't match production on this OS")
+	}
+}
+
 func createWgInterfaceWithBind(t *testing.T) (*iface.WGIface, error) {
 	t.Helper()
 	ov := os.Getenv("NB_WG_KERNEL_DISABLED")
@@ -1065,7 +1017,6 @@ type mockHandler struct {
 
 func (m *mockHandler) ServeDNS(dns.ResponseWriter, *dns.Msg) {}
 func (m *mockHandler) Stop()                                 {}
-func (m *mockHandler) ProbeAvailability(context.Context)     {}
 func (m *mockHandler) ID() types.HandlerID                   { return types.HandlerID(m.Id) }
 
 type mockService struct{}
@@ -2085,6 +2036,598 @@ func TestLocalResolverPriorityConstants(t *testing.T) {
 	assert.Equal(t, "local.example.com", localMuxUpdates[0].domain)
 }
 
+// TestBuildUpstreamHandler_MergesGroupsPerDomain verifies that multiple
+// admin-defined nameserver groups targeting the same domain collapse into a
+// single handler with each group preserved as a sequential inner list.
+func TestBuildUpstreamHandler_MergesGroupsPerDomain(t *testing.T) {
+	wgInterface := &mocWGIface{}
+	service := NewServiceViaMemory(wgInterface)
+	server := &DefaultServer{
+		ctx:           context.Background(),
+		wgInterface:   wgInterface,
+		service:       service,
+		localResolver: local.NewResolver(),
+		handlerChain:  NewHandlerChain(),
+		hostManager:   &noopHostConfigurator{},
+		dnsMuxMap:     make(registeredHandlerMap),
+	}
+
+	groups := []*nbdns.NameServerGroup{
+		{
+			NameServers: []nbdns.NameServer{
+				{IP: netip.MustParseAddr("192.0.2.1"), NSType: nbdns.UDPNameServerType, Port: 53},
+			},
+			Domains: []string{"example.com"},
+		},
+		{
+			NameServers: []nbdns.NameServer{
+				{IP: netip.MustParseAddr("192.0.2.2"), NSType: nbdns.UDPNameServerType, Port: 53},
+				{IP: netip.MustParseAddr("192.0.2.3"), NSType: nbdns.UDPNameServerType, Port: 53},
+			},
+			Domains: []string{"example.com"},
+		},
+	}
+
+	muxUpdates, err := server.buildUpstreamHandlerUpdate(groups)
+	require.NoError(t, err)
+	require.Len(t, muxUpdates, 1, "same-domain groups should merge into one handler")
+	assert.Equal(t, "example.com", muxUpdates[0].domain)
+	assert.Equal(t, PriorityUpstream, muxUpdates[0].priority)
+
+	handler := muxUpdates[0].handler.(*upstreamResolver)
+	require.Len(t, handler.upstreamServers, 2, "handler should have two groups")
+	assert.Equal(t, upstreamRace{netip.MustParseAddrPort("192.0.2.1:53")}, handler.upstreamServers[0])
+	assert.Equal(t, upstreamRace{
+		netip.MustParseAddrPort("192.0.2.2:53"),
+		netip.MustParseAddrPort("192.0.2.3:53"),
+	}, handler.upstreamServers[1])
+}
+
+// TestEvaluateNSGroupHealth covers the records-only verdict. The gate
+// (overlay route selected-but-no-active-peer) is intentionally NOT an
+// input to the evaluator anymore: the verdict drives the Enabled flag,
+// which must always reflect what we actually observed. Gate-aware event
+// suppression is tested separately in the projection test.
+//
+// Matrix per upstream: {no record, fresh Ok, fresh Fail, stale Fail,
+// stale Ok, Ok newer than Fail, Fail newer than Ok}.
+// Group verdict: any fresh-working → Healthy; any fresh-broken with no
+// fresh-working → Unhealthy; otherwise Undecided.
+func TestEvaluateNSGroupHealth(t *testing.T) {
+	now := time.Now()
+	a := netip.MustParseAddrPort("192.0.2.1:53")
+	b := netip.MustParseAddrPort("192.0.2.2:53")
+
+	recentOk := UpstreamHealth{LastOk: now.Add(-2 * time.Second)}
+	recentFail := UpstreamHealth{LastFail: now.Add(-1 * time.Second), LastErr: "timeout"}
+	staleOk := UpstreamHealth{LastOk: now.Add(-10 * time.Minute)}
+	staleFail := UpstreamHealth{LastFail: now.Add(-10 * time.Minute), LastErr: "timeout"}
+	okThenFail := UpstreamHealth{
+		LastOk:   now.Add(-10 * time.Second),
+		LastFail: now.Add(-1 * time.Second),
+		LastErr:  "timeout",
+	}
+	failThenOk := UpstreamHealth{
+		LastOk:   now.Add(-1 * time.Second),
+		LastFail: now.Add(-10 * time.Second),
+		LastErr:  "timeout",
+	}
+
+	tests := []struct {
+		name         string
+		health       map[netip.AddrPort]UpstreamHealth
+		servers      []netip.AddrPort
+		wantVerdict  nsGroupVerdict
+		wantErrSubst string
+	}{
+		{
+			name:        "no record, undecided",
+			servers:     []netip.AddrPort{a},
+			wantVerdict: nsVerdictUndecided,
+		},
+		{
+			name:        "fresh success, healthy",
+			health:      map[netip.AddrPort]UpstreamHealth{a: recentOk},
+			servers:     []netip.AddrPort{a},
+			wantVerdict: nsVerdictHealthy,
+		},
+		{
+			name:         "fresh failure, unhealthy",
+			health:       map[netip.AddrPort]UpstreamHealth{a: recentFail},
+			servers:      []netip.AddrPort{a},
+			wantVerdict:  nsVerdictUnhealthy,
+			wantErrSubst: "timeout",
+		},
+		{
+			name:        "only stale success, undecided",
+			health:      map[netip.AddrPort]UpstreamHealth{a: staleOk},
+			servers:     []netip.AddrPort{a},
+			wantVerdict: nsVerdictUndecided,
+		},
+		{
+			name:        "only stale failure, undecided",
+			health:      map[netip.AddrPort]UpstreamHealth{a: staleFail},
+			servers:     []netip.AddrPort{a},
+			wantVerdict: nsVerdictUndecided,
+		},
+		{
+			name:         "both fresh, fail newer, unhealthy",
+			health:       map[netip.AddrPort]UpstreamHealth{a: okThenFail},
+			servers:      []netip.AddrPort{a},
+			wantVerdict:  nsVerdictUnhealthy,
+			wantErrSubst: "timeout",
+		},
+		{
+			name:        "both fresh, ok newer, healthy",
+			health:      map[netip.AddrPort]UpstreamHealth{a: failThenOk},
+			servers:     []netip.AddrPort{a},
+			wantVerdict: nsVerdictHealthy,
+		},
+		{
+			name: "two upstreams, one success wins",
+			health: map[netip.AddrPort]UpstreamHealth{
+				a: recentFail,
+				b: recentOk,
+			},
+			servers:     []netip.AddrPort{a, b},
+			wantVerdict: nsVerdictHealthy,
+		},
+		{
+			name: "two upstreams, one fail one unseen, unhealthy",
+			health: map[netip.AddrPort]UpstreamHealth{
+				a: recentFail,
+			},
+			servers:      []netip.AddrPort{a, b},
+			wantVerdict:  nsVerdictUnhealthy,
+			wantErrSubst: "timeout",
+		},
+		{
+			name: "two upstreams, all recent failures, unhealthy",
+			health: map[netip.AddrPort]UpstreamHealth{
+				a: {LastFail: now.Add(-5 * time.Second), LastErr: "timeout"},
+				b: {LastFail: now.Add(-1 * time.Second), LastErr: "SERVFAIL"},
+			},
+			servers:      []netip.AddrPort{a, b},
+			wantVerdict:  nsVerdictUnhealthy,
+			wantErrSubst: "SERVFAIL",
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			verdict, err := evaluateNSGroupHealth(tc.health, tc.servers, now)
+			assert.Equal(t, tc.wantVerdict, verdict, "verdict mismatch")
+			if tc.wantErrSubst != "" {
+				require.Error(t, err)
+				assert.Contains(t, err.Error(), tc.wantErrSubst)
+			} else {
+				assert.NoError(t, err)
+			}
+		})
+	}
+}
+
+// healthStubHandler is a minimal dnsMuxMap entry that exposes a fixed
+// UpstreamHealth snapshot, letting tests drive recomputeNSGroupStates
+// without spinning up real handlers.
+type healthStubHandler struct {
+	health map[netip.AddrPort]UpstreamHealth
+}
+
+func (h *healthStubHandler) ServeDNS(dns.ResponseWriter, *dns.Msg) {}
+func (h *healthStubHandler) Stop()                                 {}
+func (h *healthStubHandler) ID() types.HandlerID                   { return "health-stub" }
+func (h *healthStubHandler) UpstreamHealth() map[netip.AddrPort]UpstreamHealth {
+	return h.health
+}
+
+// TestProjection_SteadyStateIsSilent guards against duplicate events:
+// while a group stays Unhealthy tick after tick, only the first
+// Unhealthy transition may emit. Same for staying Healthy.
+func TestProjection_SteadyStateIsSilent(t *testing.T) {
+	fx := newProjTestFixture(t)
+
+	fx.setHealth(UpstreamHealth{LastFail: time.Now(), LastErr: "timeout"})
+	fx.tick()
+	fx.expectEvent("unreachable", "first fail emits warning")
+
+	fx.setHealth(UpstreamHealth{LastFail: time.Now(), LastErr: "timeout"})
+	fx.tick()
+	fx.tick()
+	fx.expectNoEvent("staying unhealthy must not re-emit")
+
+	fx.setHealth(UpstreamHealth{LastOk: time.Now()})
+	fx.tick()
+	fx.expectEvent("recovered", "recovery on transition")
+
+	fx.tick()
+	fx.tick()
+	fx.expectNoEvent("staying healthy must not re-emit")
+}
+
+// projTestFixture is the common setup for the projection tests: a
+// single-upstream group whose route classification the test can flip by
+// assigning to selected/active. Callers drive failures/successes by
+// mutating stub.health and calling refreshHealth.
+type projTestFixture struct {
+	t        *testing.T
+	recorder *peer.Status
+	events   <-chan *proto.SystemEvent
+	server   *DefaultServer
+	stub     *healthStubHandler
+	group    *nbdns.NameServerGroup
+	srv      netip.AddrPort
+	selected route.HAMap
+	active   route.HAMap
+}
+
+func newProjTestFixture(t *testing.T) *projTestFixture {
+	t.Helper()
+	recorder := peer.NewRecorder("mgm")
+	sub := recorder.SubscribeToEvents()
+	t.Cleanup(func() { recorder.UnsubscribeFromEvents(sub) })
+
+	srv := netip.MustParseAddrPort("100.64.0.1:53")
+	fx := &projTestFixture{
+		t:        t,
+		recorder: recorder,
+		events:   sub.Events(),
+		stub:     &healthStubHandler{health: map[netip.AddrPort]UpstreamHealth{}},
+		srv:      srv,
+		group: &nbdns.NameServerGroup{
+			Domains:     []string{"example.com"},
+			NameServers: []nbdns.NameServer{{IP: srv.Addr(), NSType: nbdns.UDPNameServerType, Port: int(srv.Port())}},
+		},
+	}
+	fx.server = &DefaultServer{
+		ctx:              context.Background(),
+		wgInterface:      &mocWGIface{},
+		statusRecorder:   recorder,
+		dnsMuxMap:        make(registeredHandlerMap),
+		selectedRoutes:   func() route.HAMap { return fx.selected },
+		activeRoutes:     func() route.HAMap { return fx.active },
+		warningDelayBase: defaultWarningDelayBase,
+	}
+	fx.server.dnsMuxMap["example.com"] = handlerWrapper{domain: "example.com", handler: fx.stub, priority: PriorityUpstream}
+
+	fx.server.mux.Lock()
+	fx.server.updateNSGroupStates([]*nbdns.NameServerGroup{fx.group})
+	fx.server.mux.Unlock()
+	return fx
+}
+
+func (f *projTestFixture) setHealth(h UpstreamHealth) {
+	f.stub.health = map[netip.AddrPort]UpstreamHealth{f.srv: h}
+}
+
+func (f *projTestFixture) tick() []peer.NSGroupState {
+	f.server.refreshHealth()
+	return f.recorder.GetDNSStates()
+}
+
+func (f *projTestFixture) expectNoEvent(why string) {
+	f.t.Helper()
+	select {
+	case evt := <-f.events:
+		f.t.Fatalf("unexpected event (%s): %+v", why, evt)
+	case <-time.After(100 * time.Millisecond):
+	}
+}
+
+func (f *projTestFixture) expectEvent(substr, why string) *proto.SystemEvent {
+	f.t.Helper()
+	select {
+	case evt := <-f.events:
+		assert.Contains(f.t, evt.Message, substr, why)
+		return evt
+	case <-time.After(time.Second):
+		f.t.Fatalf("expected event (%s) with %q", why, substr)
+		return nil
+	}
+}
+
+var overlayNetForTest = netip.MustParsePrefix("100.64.0.0/16")
+var overlayMapForTest = route.HAMap{"overlay": {{Network: overlayNetForTest}}}
+
+// TestProjection_PublicFailEmitsImmediately covers rule 1: an upstream
+// that is not inside any selected route (public DNS) fires the warning
+// on the first Unhealthy tick, no grace period.
+func TestProjection_PublicFailEmitsImmediately(t *testing.T) {
+	fx := newProjTestFixture(t)
+
+	fx.setHealth(UpstreamHealth{LastFail: time.Now(), LastErr: "timeout"})
+	states := fx.tick()
+	require.Len(t, states, 1)
+	assert.False(t, states[0].Enabled)
+	fx.expectEvent("unreachable", "public DNS failure")
+}
+
+// TestProjection_OverlayConnectedFailEmitsImmediately covers rule 2:
+// the upstream is inside a selected route AND the route has a Connected
+// peer. Tunnel is up, failure is real, emit immediately.
+func TestProjection_OverlayConnectedFailEmitsImmediately(t *testing.T) {
+	fx := newProjTestFixture(t)
+	fx.selected = overlayMapForTest
+	fx.active = overlayMapForTest
+
+	fx.setHealth(UpstreamHealth{LastFail: time.Now(), LastErr: "timeout"})
+	states := fx.tick()
+	require.Len(t, states, 1)
+	assert.False(t, states[0].Enabled)
+	fx.expectEvent("unreachable", "overlay + connected failure")
+}
+
+// TestProjection_OverlayNotConnectedDelaysWarning covers rule 3: the
+// upstream is routed but no peer is Connected (Connecting/Idle/missing).
+// First tick: Unhealthy display, no warning. After the grace window
+// elapses with no recovery, the warning fires.
+func TestProjection_OverlayNotConnectedDelaysWarning(t *testing.T) {
+	grace := 50 * time.Millisecond
+	fx := newProjTestFixture(t)
+	fx.server.warningDelayBase = grace
+	fx.selected = overlayMapForTest
+	// active stays nil: routed but not connected.
+
+	fx.setHealth(UpstreamHealth{LastFail: time.Now(), LastErr: "timeout"})
+	states := fx.tick()
+	require.Len(t, states, 1)
+	assert.False(t, states[0].Enabled, "display must reflect failure even during grace window")
+	fx.expectNoEvent("first fail tick within grace window")
+
+	time.Sleep(grace + 10*time.Millisecond)
+	fx.setHealth(UpstreamHealth{LastFail: time.Now(), LastErr: "timeout"})
+	fx.tick()
+	fx.expectEvent("unreachable", "warning after grace window")
+}
+
+// TestProjection_OverlayAddrNoRouteDelaysWarning covers an upstream
+// whose address is inside the WireGuard overlay range but is not
+// covered by any selected route (peer-to-peer DNS without an explicit
+// route). Until a peer reports Connected for that address, startup
+// failures must be held just like the routed case.
+func TestProjection_OverlayAddrNoRouteDelaysWarning(t *testing.T) {
+	recorder := peer.NewRecorder("mgm")
+	sub := recorder.SubscribeToEvents()
+	t.Cleanup(func() { recorder.UnsubscribeFromEvents(sub) })
+
+	overlayPeer := netip.MustParseAddrPort("100.66.100.5:53")
+	server := &DefaultServer{
+		ctx:              context.Background(),
+		wgInterface:      &mocWGIface{},
+		statusRecorder:   recorder,
+		dnsMuxMap:        make(registeredHandlerMap),
+		selectedRoutes:   func() route.HAMap { return nil },
+		activeRoutes:     func() route.HAMap { return nil },
+		warningDelayBase: 50 * time.Millisecond,
+	}
+	group := &nbdns.NameServerGroup{
+		Domains:     []string{"example.com"},
+		NameServers: []nbdns.NameServer{{IP: overlayPeer.Addr(), NSType: nbdns.UDPNameServerType, Port: int(overlayPeer.Port())}},
+	}
+	stub := &healthStubHandler{health: map[netip.AddrPort]UpstreamHealth{
+		overlayPeer: {LastFail: time.Now(), LastErr: "timeout"},
+	}}
+	server.dnsMuxMap["example.com"] = handlerWrapper{domain: "example.com", handler: stub, priority: PriorityUpstream}
+
+	server.mux.Lock()
+	server.updateNSGroupStates([]*nbdns.NameServerGroup{group})
+	server.mux.Unlock()
+	server.refreshHealth()
+
+	select {
+	case evt := <-sub.Events():
+		t.Fatalf("unexpected event during grace window: %+v", evt)
+	case <-time.After(100 * time.Millisecond):
+	}
+
+	time.Sleep(60 * time.Millisecond)
+	stub.health = map[netip.AddrPort]UpstreamHealth{overlayPeer: {LastFail: time.Now(), LastErr: "timeout"}}
+	server.refreshHealth()
+
+	select {
+	case evt := <-sub.Events():
+		assert.Contains(t, evt.Message, "unreachable")
+	case <-time.After(time.Second):
+		t.Fatal("expected warning after grace window")
+	}
+}
+
+// TestProjection_StopClearsHealthState verifies that Stop wipes the
+// per-group projection state so a subsequent Start doesn't inherit
+// sticky flags (notably everHealthy) that would bypass the grace
+// window during the next peer handshake.
+func TestProjection_StopClearsHealthState(t *testing.T) {
+	wgIface := &mocWGIface{}
+	server := &DefaultServer{
+		ctx:               context.Background(),
+		wgInterface:       wgIface,
+		service:           NewServiceViaMemory(wgIface),
+		hostManager:       &noopHostConfigurator{},
+		extraDomains:      map[domain.Domain]int{},
+		dnsMuxMap:         make(registeredHandlerMap),
+		statusRecorder:    peer.NewRecorder("mgm"),
+		selectedRoutes:    func() route.HAMap { return nil },
+		activeRoutes:      func() route.HAMap { return nil },
+		warningDelayBase:  defaultWarningDelayBase,
+		currentConfigHash: ^uint64(0),
+	}
+	server.ctx, server.ctxCancel = context.WithCancel(context.Background())
+
+	srv := netip.MustParseAddrPort("8.8.8.8:53")
+	group := &nbdns.NameServerGroup{
+		Domains:     []string{"example.com"},
+		NameServers: []nbdns.NameServer{{IP: srv.Addr(), NSType: nbdns.UDPNameServerType, Port: int(srv.Port())}},
+	}
+	stub := &healthStubHandler{health: map[netip.AddrPort]UpstreamHealth{srv: {LastOk: time.Now()}}}
+	server.dnsMuxMap["example.com"] = handlerWrapper{domain: "example.com", handler: stub, priority: PriorityUpstream}
+
+	server.mux.Lock()
+	server.updateNSGroupStates([]*nbdns.NameServerGroup{group})
+	server.mux.Unlock()
+	server.refreshHealth()
+
+	server.healthProjectMu.Lock()
+	p, ok := server.nsGroupProj[generateGroupKey(group)]
+	server.healthProjectMu.Unlock()
+	require.True(t, ok, "projection state should exist after tick")
+	require.True(t, p.everHealthy, "tick with success must set everHealthy")
+
+	server.Stop()
+
+	server.healthProjectMu.Lock()
+	cleared := server.nsGroupProj == nil
+	server.healthProjectMu.Unlock()
+	assert.True(t, cleared, "Stop must clear nsGroupProj")
+}
+
+// TestProjection_OverlayRecoversDuringGrace covers the happy path of
+// rule 3: startup failures while the peer is handshaking, then the peer
+// comes up and a query succeeds before the grace window elapses. No
+// warning should ever have fired, and no recovery either.
+func TestProjection_OverlayRecoversDuringGrace(t *testing.T) {
+	fx := newProjTestFixture(t)
+	fx.server.warningDelayBase = 200 * time.Millisecond
+	fx.selected = overlayMapForTest
+
+	fx.setHealth(UpstreamHealth{LastFail: time.Now(), LastErr: "timeout"})
+	fx.tick()
+	fx.expectNoEvent("fail within grace, warning suppressed")
+
+	fx.active = overlayMapForTest
+	fx.setHealth(UpstreamHealth{LastOk: time.Now()})
+	states := fx.tick()
+	require.Len(t, states, 1)
+	assert.True(t, states[0].Enabled)
+	fx.expectNoEvent("recovery without prior warning must not emit")
+}
+
+// TestProjection_RecoveryOnlyAfterWarning enforces the invariant the
+// whole design leans on: recovery events only appear when a warning
+// event was actually emitted for the current streak. A Healthy verdict
+// without a prior warning is silent, so the user never sees "recovered"
+// out of thin air.
+func TestProjection_RecoveryOnlyAfterWarning(t *testing.T) {
+	fx := newProjTestFixture(t)
+
+	fx.setHealth(UpstreamHealth{LastOk: time.Now()})
+	states := fx.tick()
+	require.Len(t, states, 1)
+	assert.True(t, states[0].Enabled)
+	fx.expectNoEvent("first healthy tick should not recover anything")
+
+	fx.setHealth(UpstreamHealth{LastFail: time.Now(), LastErr: "timeout"})
+	fx.tick()
+	fx.expectEvent("unreachable", "public fail emits immediately")
+
+	fx.setHealth(UpstreamHealth{LastOk: time.Now()})
+	fx.tick()
+	fx.expectEvent("recovered", "recovery follows real warning")
+
+	fx.setHealth(UpstreamHealth{LastFail: time.Now(), LastErr: "timeout"})
+	fx.tick()
+	fx.expectEvent("unreachable", "second cycle warning")
+
+	fx.setHealth(UpstreamHealth{LastOk: time.Now()})
+	fx.tick()
+	fx.expectEvent("recovered", "second cycle recovery")
+}
+
+// TestProjection_EverHealthyOverridesDelay covers rule 4: once a group
+// has ever been Healthy, subsequent failures skip the grace window even
+// if classification says "routed + not connected". The system has
+// proved it can work, so any new failure is real.
+func TestProjection_EverHealthyOverridesDelay(t *testing.T) {
+	fx := newProjTestFixture(t)
+	// Large base so any emission must come from the everHealthy bypass, not elapsed time.
+	fx.server.warningDelayBase = time.Hour
+	fx.selected = overlayMapForTest
+	fx.active = overlayMapForTest
+
+	// Establish "ever healthy".
+	fx.setHealth(UpstreamHealth{LastOk: time.Now()})
+	fx.tick()
+	fx.expectNoEvent("first healthy tick")
+
+	// Peer drops. Query fails. Routed + not connected → normally grace,
+	// but everHealthy flag bypasses it.
+	fx.active = nil
+	fx.setHealth(UpstreamHealth{LastFail: time.Now(), LastErr: "timeout"})
+	fx.tick()
+	fx.expectEvent("unreachable", "failure after ever-healthy must be immediate")
+}
+
+// TestProjection_ReconnectBlipEmitsPair covers the explicit tradeoff
+// from the design discussion: once a group has been healthy, a brief
+// reconnect that produces a failing tick will fire warning + recovery.
+// This is by design: user-visible blips are accurate signal, not noise.
+func TestProjection_ReconnectBlipEmitsPair(t *testing.T) {
+	fx := newProjTestFixture(t)
+	fx.selected = overlayMapForTest
+	fx.active = overlayMapForTest
+
+	fx.setHealth(UpstreamHealth{LastOk: time.Now()})
+	fx.tick()
+
+	fx.setHealth(UpstreamHealth{LastFail: time.Now(), LastErr: "timeout"})
+	fx.tick()
+	fx.expectEvent("unreachable", "blip warning")
+
+	fx.setHealth(UpstreamHealth{LastOk: time.Now()})
+	fx.tick()
+	fx.expectEvent("recovered", "blip recovery")
+}
+
+// TestProjection_MixedGroupEmitsImmediately covers the multi-upstream
+// rule: a group with at least one public upstream is in the "immediate"
+// category regardless of the other upstreams' routing, because the
+// public one has no peer-startup excuse. Prevents public-DNS failures
+// from being hidden behind a routed sibling.
+func TestProjection_MixedGroupEmitsImmediately(t *testing.T) {
+	recorder := peer.NewRecorder("mgm")
+	sub := recorder.SubscribeToEvents()
+	t.Cleanup(func() { recorder.UnsubscribeFromEvents(sub) })
+	events := sub.Events()
+
+	public := netip.MustParseAddrPort("8.8.8.8:53")
+	overlay := netip.MustParseAddrPort("100.64.0.1:53")
+	overlayMap := route.HAMap{"overlay": {{Network: netip.MustParsePrefix("100.64.0.0/16")}}}
+
+	server := &DefaultServer{
+		ctx:              context.Background(),
+		statusRecorder:   recorder,
+		dnsMuxMap:        make(registeredHandlerMap),
+		selectedRoutes:   func() route.HAMap { return overlayMap },
+		activeRoutes:     func() route.HAMap { return nil },
+		warningDelayBase: time.Hour,
+	}
+	group := &nbdns.NameServerGroup{
+		Domains: []string{"example.com"},
+		NameServers: []nbdns.NameServer{
+			{IP: public.Addr(), NSType: nbdns.UDPNameServerType, Port: int(public.Port())},
+			{IP: overlay.Addr(), NSType: nbdns.UDPNameServerType, Port: int(overlay.Port())},
+		},
+	}
+	stub := &healthStubHandler{
+		health: map[netip.AddrPort]UpstreamHealth{
+			public:  {LastFail: time.Now(), LastErr: "servfail"},
+			overlay: {LastFail: time.Now(), LastErr: "timeout"},
+		},
+	}
+	server.dnsMuxMap["example.com"] = handlerWrapper{domain: "example.com", handler: stub, priority: PriorityUpstream}
+
+	server.mux.Lock()
+	server.updateNSGroupStates([]*nbdns.NameServerGroup{group})
+	server.mux.Unlock()
+	server.refreshHealth()
+
+	select {
+	case evt := <-events:
+		assert.Contains(t, evt.Message, "unreachable")
+	case <-time.After(time.Second):
+		t.Fatal("expected immediate warning because group contains a public upstream")
+	}
+}
+
 func TestDNSLoopPrevention(t *testing.T) {
 	wgInterface := &mocWGIface{}
 	service := NewServiceViaMemory(wgInterface)
@@ -2183,17 +2726,18 @@ func TestDNSLoopPrevention(t *testing.T) {
 
 			if tt.expectedHandlers > 0 {
 				handler := muxUpdates[0].handler.(*upstreamResolver)
-				assert.Len(t, handler.upstreamServers, len(tt.expectedServers))
+				flat := handler.flatUpstreams()
+				assert.Len(t, flat, len(tt.expectedServers))
 
 				if tt.shouldFilterOwnIP {
-					for _, upstream := range handler.upstreamServers {
+					for _, upstream := range flat {
 						assert.NotEqual(t, dnsServerIP, upstream.Addr())
 					}
 				}
 
 				for _, expected := range tt.expectedServers {
 					found := false
-					for _, upstream := range handler.upstreamServers {
+					for _, upstream := range flat {
 						if upstream.Addr() == expected {
 							found = true
 							break

client/internal/dns/systemd_linux.go

@@ -8,6 +8,7 @@ import (
 	"fmt"
 	"net"
 	"net/netip"
+	"slices"
 	"time"
 
 	"github.com/godbus/dbus/v5"
@@ -40,10 +41,17 @@ const (
 )
 
 type systemdDbusConfigurator struct {
-	dbusLinkObject dbus.ObjectPath
-	ifaceName      string
+	dbusLinkObject  dbus.ObjectPath
+	ifaceName       string
+	wgIndex         int
+	origNameservers []netip.Addr
 }
 
+const (
+	systemdDbusLinkDNSProperty          = systemdDbusLinkInterface + ".DNS"
+	systemdDbusLinkDefaultRouteProperty = systemdDbusLinkInterface + ".DefaultRoute"
+)
+
 // the types below are based on dbus specification, each field is mapped to a dbus type
 // see https://dbus.freedesktop.org/doc/dbus-specification.html#basic-types for more details on dbus types
 // see https://www.freedesktop.org/software/systemd/man/org.freedesktop.resolve1.html on resolve1 input types
@@ -79,10 +87,145 @@ func newSystemdDbusConfigurator(wgInterface string) (*systemdDbusConfigurator, e
 
 	log.Debugf("got dbus Link interface: %s from net interface %s and index %d", s, iface.Name, iface.Index)
 
-	return &systemdDbusConfigurator{
+	c := &systemdDbusConfigurator{
 		dbusLinkObject: dbus.ObjectPath(s),
 		ifaceName:      wgInterface,
-	}, nil
+		wgIndex:        iface.Index,
+	}
+
+	origNameservers, err := c.captureOriginalNameservers()
+	switch {
+	case err != nil:
+		log.Warnf("capture original nameservers from systemd-resolved: %v", err)
+	case len(origNameservers) == 0:
+		log.Warnf("no original nameservers captured from systemd-resolved default-route links; DNS fallback will be empty")
+	default:
+		log.Debugf("captured %d original nameservers from systemd-resolved default-route links: %v", len(origNameservers), origNameservers)
+	}
+	c.origNameservers = origNameservers
+	return c, nil
+}
+
+// captureOriginalNameservers reads per-link DNS from systemd-resolved for
+// every default-route link except our own WG link. Non-default-route links
+// (VPNs, docker bridges) are skipped because their upstreams wouldn't
+// actually serve host queries.
+func (s *systemdDbusConfigurator) captureOriginalNameservers() ([]netip.Addr, error) {
+	ifaces, err := net.Interfaces()
+	if err != nil {
+		return nil, fmt.Errorf("list interfaces: %w", err)
+	}
+
+	seen := make(map[netip.Addr]struct{})
+	var out []netip.Addr
+	for _, iface := range ifaces {
+		if !s.isCandidateLink(iface) {
+			continue
+		}
+		linkPath, err := getSystemdLinkPath(iface.Index)
+		if err != nil || !isSystemdLinkDefaultRoute(linkPath) {
+			continue
+		}
+		for _, addr := range readSystemdLinkDNS(linkPath) {
+			addr = normalizeSystemdAddr(addr, iface.Name)
+			if !addr.IsValid() {
+				continue
+			}
+			if _, dup := seen[addr]; dup {
+				continue
+			}
+			seen[addr] = struct{}{}
+			out = append(out, addr)
+		}
+	}
+	return out, nil
+}
+
+func (s *systemdDbusConfigurator) isCandidateLink(iface net.Interface) bool {
+	if iface.Index == s.wgIndex {
+		return false
+	}
+	if iface.Flags&net.FlagLoopback != 0 || iface.Flags&net.FlagUp == 0 {
+		return false
+	}
+	return true
+}
+
+// normalizeSystemdAddr unmaps v4-mapped-v6, drops unspecified, and reattaches
+// the link's iface name as zone for link-local v6 (Link.DNS strips it).
+// Returns the zero Addr to signal "skip this entry".
+func normalizeSystemdAddr(addr netip.Addr, ifaceName string) netip.Addr {
+	addr = addr.Unmap()
+	if !addr.IsValid() || addr.IsUnspecified() {
+		return netip.Addr{}
+	}
+	if addr.IsLinkLocalUnicast() {
+		return addr.WithZone(ifaceName)
+	}
+	return addr
+}
+
+func getSystemdLinkPath(ifIndex int) (dbus.ObjectPath, error) {
+	obj, closeConn, err := getDbusObject(systemdResolvedDest, systemdDbusObjectNode)
+	if err != nil {
+		return "", fmt.Errorf("dbus resolve1: %w", err)
+	}
+	defer closeConn()
+	var p string
+	if err := obj.Call(systemdDbusGetLinkMethod, dbusDefaultFlag, int32(ifIndex)).Store(&p); err != nil {
+		return "", err
+	}
+	return dbus.ObjectPath(p), nil
+}
+
+func isSystemdLinkDefaultRoute(linkPath dbus.ObjectPath) bool {
+	obj, closeConn, err := getDbusObject(systemdResolvedDest, linkPath)
+	if err != nil {
+		return false
+	}
+	defer closeConn()
+	v, err := obj.GetProperty(systemdDbusLinkDefaultRouteProperty)
+	if err != nil {
+		return false
+	}
+	b, ok := v.Value().(bool)
+	return ok && b
+}
+
+func readSystemdLinkDNS(linkPath dbus.ObjectPath) []netip.Addr {
+	obj, closeConn, err := getDbusObject(systemdResolvedDest, linkPath)
+	if err != nil {
+		return nil
+	}
+	defer closeConn()
+	v, err := obj.GetProperty(systemdDbusLinkDNSProperty)
+	if err != nil {
+		return nil
+	}
+	entries, ok := v.Value().([][]any)
+	if !ok {
+		return nil
+	}
+	var out []netip.Addr
+	for _, entry := range entries {
+		if len(entry) < 2 {
+			continue
+		}
+		raw, ok := entry[1].([]byte)
+		if !ok {
+			continue
+		}
+		addr, ok := netip.AddrFromSlice(raw)
+		if !ok {
+			continue
+		}
+		out = append(out, addr)
+	}
+	return out
+}
+
+func (s *systemdDbusConfigurator) getOriginalNameservers() []netip.Addr {
+	return slices.Clone(s.origNameservers)
 }
 
 func (s *systemdDbusConfigurator) supportCustomPort() bool {

client/internal/dns/upstream.go

@@ -1,3 +1,32 @@
+// Package dns implements the client-side DNS stack: listener/service on the
+// peer's tunnel address, handler chain that routes questions by domain and
+// priority, and upstream resolvers that forward what remains to configured
+// nameservers.
+//
+// # Upstream resolution and the race model
+//
+// When two or more nameserver groups target the same domain, DefaultServer
+// merges them into one upstream handler whose state is:
+//
+//	upstreamResolverBase
+//	  └── upstreamServers []upstreamRace   // one entry per source NS group
+//	        └── []netip.AddrPort           // primary, fallback, ...
+//
+// Each source nameserver group contributes one upstreamRace. Within a race
+// upstreams are tried in order: the next is used only on failure (timeout,
+// SERVFAIL, REFUSED, no response). NXDOMAIN is a valid answer and stops
+// the walk. When more than one race exists, ServeDNS fans out one
+// goroutine per race and returns the first valid answer, cancelling the
+// rest. A handler with a single race skips the fan-out.
+//
+// # Health projection
+//
+// Query outcomes are recorded per-upstream in UpstreamHealth. The server
+// periodically merges these snapshots across handlers and projects them
+// into peer.NSGroupState. There is no active probing: a group is marked
+// unhealthy only when every seen upstream has a recent failure and none
+// has a recent success. Healthy→unhealthy fires a single
+// SystemEvent_WARNING; steady-state refreshes do not duplicate it.
 package dns
 
 import (
@@ -11,11 +40,8 @@ import (
 	"slices"
 	"strings"
 	"sync"
-	"sync/atomic"
 	"time"
 
-	"github.com/cenkalti/backoff/v4"
-	"github.com/hashicorp/go-multierror"
 	"github.com/miekg/dns"
 	log "github.com/sirupsen/logrus"
 	"golang.zx2c4.com/wireguard/tun/netstack"
@@ -25,7 +51,8 @@ import (
 	"github.com/netbirdio/netbird/client/internal/dns/resutil"
 	"github.com/netbirdio/netbird/client/internal/dns/types"
 	"github.com/netbirdio/netbird/client/internal/peer"
-	"github.com/netbirdio/netbird/client/proto"
+	"github.com/netbirdio/netbird/route"
+	"github.com/netbirdio/netbird/shared/management/domain"
 )
 
 var currentMTU uint16 = iface.DefaultMTU
@@ -67,15 +94,17 @@ const (
 	// Set longer than UpstreamTimeout to ensure context timeout takes precedence
 	ClientTimeout = 5 * time.Second
 
-	reactivatePeriod = 30 * time.Second
-	probeTimeout     = 2 * time.Second
-
 	// ipv6HeaderSize + udpHeaderSize, used to derive the maximum DNS UDP
 	// payload from the tunnel MTU.
 	ipUDPHeaderSize = 60 + 8
-)
 
-const testRecord = "com."
+	// raceMaxTotalTimeout caps the combined time spent walking all upstreams
+	// within one race, so a slow primary can't eat the whole race budget.
+	raceMaxTotalTimeout = 5 * time.Second
+	// raceMinPerUpstreamTimeout is the floor applied when dividing
+	// raceMaxTotalTimeout across upstreams within a race.
+	raceMinPerUpstreamTimeout = 2 * time.Second
+)
 
 const (
 	protoUDP = "udp"
@@ -84,6 +113,69 @@ const (
 
 type dnsProtocolKey struct{}
 
+type upstreamProtocolKey struct{}
+
+// upstreamProtocolResult holds the protocol used for the upstream exchange.
+// Stored as a pointer in context so the exchange function can set it.
+type upstreamProtocolResult struct {
+	protocol string
+}
+
+type upstreamClient interface {
+	exchange(ctx context.Context, upstream string, r *dns.Msg) (*dns.Msg, time.Duration, error)
+}
+
+type UpstreamResolver interface {
+	serveDNS(r *dns.Msg) (*dns.Msg, time.Duration, error)
+	upstreamExchange(upstream string, r *dns.Msg) (*dns.Msg, time.Duration, error)
+}
+
+// upstreamRace is an ordered list of upstreams derived from one configured
+// nameserver group. Order matters: the first upstream is tried first, the
+// second only on failure, and so on. Multiple upstreamRace values coexist
+// inside one resolver when overlapping nameserver groups target the same
+// domain; those races run in parallel and the first valid answer wins.
+type upstreamRace []netip.AddrPort
+
+// UpstreamHealth is the last query-path outcome for a single upstream,
+// consumed by nameserver-group status projection.
+type UpstreamHealth struct {
+	LastOk   time.Time
+	LastFail time.Time
+	LastErr  string
+}
+
+type upstreamResolverBase struct {
+	ctx             context.Context
+	cancel          context.CancelFunc
+	upstreamClient  upstreamClient
+	upstreamServers []upstreamRace
+	domain          domain.Domain
+	upstreamTimeout time.Duration
+
+	healthMu sync.RWMutex
+	health   map[netip.AddrPort]*UpstreamHealth
+
+	statusRecorder *peer.Status
+	// selectedRoutes returns the current set of client routes the admin
+	// has enabled. Called lazily from the query hot path when an upstream
+	// might need a tunnel-bound client (iOS) and from health projection.
+	selectedRoutes func() route.HAMap
+}
+
+type upstreamFailure struct {
+	upstream netip.AddrPort
+	reason   string
+}
+
+type raceResult struct {
+	msg      *dns.Msg
+	upstream netip.AddrPort
+	protocol string
+	ede      string
+	failures []upstreamFailure
+}
+
 // contextWithDNSProtocol stores the inbound DNS protocol ("udp" or "tcp") in context.
 func contextWithDNSProtocol(ctx context.Context, network string) context.Context {
 	return context.WithValue(ctx, dnsProtocolKey{}, network)
@@ -100,16 +192,8 @@ func dnsProtocolFromContext(ctx context.Context) string {
 	return ""
 }
 
-type upstreamProtocolKey struct{}
-
-// upstreamProtocolResult holds the protocol used for the upstream exchange.
-// Stored as a pointer in context so the exchange function can set it.
-type upstreamProtocolResult struct {
-	protocol string
-}
-
-// contextWithupstreamProtocolResult stores a mutable result holder in the context.
-func contextWithupstreamProtocolResult(ctx context.Context) (context.Context, *upstreamProtocolResult) {
+// contextWithUpstreamProtocolResult stores a mutable result holder in the context.
+func contextWithUpstreamProtocolResult(ctx context.Context) (context.Context, *upstreamProtocolResult) {
 	r := &upstreamProtocolResult{}
 	return context.WithValue(ctx, upstreamProtocolKey{}, r), r
 }
@@ -124,67 +208,37 @@ func setUpstreamProtocol(ctx context.Context, protocol string) {
 	}
 }
 
-type upstreamClient interface {
-	exchange(ctx context.Context, upstream string, r *dns.Msg) (*dns.Msg, time.Duration, error)
-}
-
-type UpstreamResolver interface {
-	serveDNS(r *dns.Msg) (*dns.Msg, time.Duration, error)
-	upstreamExchange(upstream string, r *dns.Msg) (*dns.Msg, time.Duration, error)
-}
-
-type upstreamResolverBase struct {
-	ctx              context.Context
-	cancel           context.CancelFunc
-	upstreamClient   upstreamClient
-	upstreamServers  []netip.AddrPort
-	domain           string
-	disabled         bool
-	successCount     atomic.Int32
-	mutex            sync.Mutex
-	reactivatePeriod time.Duration
-	upstreamTimeout  time.Duration
-	wg               sync.WaitGroup
-
-	deactivate     func(error)
-	reactivate     func()
-	statusRecorder *peer.Status
-	routeMatch     func(netip.Addr) bool
-}
-
-type upstreamFailure struct {
-	upstream netip.AddrPort
-	reason   string
-}
-
-func newUpstreamResolverBase(ctx context.Context, statusRecorder *peer.Status, domain string) *upstreamResolverBase {
+func newUpstreamResolverBase(ctx context.Context, statusRecorder *peer.Status, d domain.Domain) *upstreamResolverBase {
 	ctx, cancel := context.WithCancel(ctx)
 
 	return &upstreamResolverBase{
-		ctx:              ctx,
-		cancel:           cancel,
-		domain:           domain,
-		upstreamTimeout:  UpstreamTimeout,
-		reactivatePeriod: reactivatePeriod,
-		statusRecorder:   statusRecorder,
+		ctx:             ctx,
+		cancel:          cancel,
+		domain:          d,
+		upstreamTimeout: UpstreamTimeout,
+		statusRecorder:  statusRecorder,
 	}
 }
 
 // String returns a string representation of the upstream resolver
 func (u *upstreamResolverBase) String() string {
-	return fmt.Sprintf("Upstream %s", u.upstreamServers)
+	return fmt.Sprintf("Upstream %s", u.flatUpstreams())
 }
 
-// ID returns the unique handler ID
+// ID returns the unique handler ID. Race groupings and within-race
+// ordering are both part of the identity: [[A,B]] and [[A],[B]] query
+// the same servers but with different semantics (serial fallback vs
+// parallel race), so their handlers must not collide.
 func (u *upstreamResolverBase) ID() types.HandlerID {
-	servers := slices.Clone(u.upstreamServers)
-	slices.SortFunc(servers, func(a, b netip.AddrPort) int { return a.Compare(b) })
-
 	hash := sha256.New()
-	hash.Write([]byte(u.domain + ":"))
-	for _, s := range servers {
-		hash.Write([]byte(s.String()))
-		hash.Write([]byte("|"))
+	hash.Write([]byte(u.domain.PunycodeString() + ":"))
+	for _, race := range u.upstreamServers {
+		hash.Write([]byte("["))
+		for _, s := range race {
+			hash.Write([]byte(s.String()))
+			hash.Write([]byte("|"))
+		}
+		hash.Write([]byte("]"))
 	}
 	return types.HandlerID("upstream-" + hex.EncodeToString(hash.Sum(nil)[:8]))
 }
@@ -194,13 +248,31 @@ func (u *upstreamResolverBase) MatchSubdomains() bool {
 }
 
 func (u *upstreamResolverBase) Stop() {
-	log.Debugf("stopping serving DNS for upstreams %s", u.upstreamServers)
+	log.Debugf("stopping serving DNS for upstreams %s", u.flatUpstreams())
 	u.cancel()
+}
 
-	u.mutex.Lock()
-	u.wg.Wait()
-	u.mutex.Unlock()
+// flatUpstreams is for logging and ID hashing only, not for dispatch.
+func (u *upstreamResolverBase) flatUpstreams() []netip.AddrPort {
+	var out []netip.AddrPort
+	for _, g := range u.upstreamServers {
+		out = append(out, g...)
+	}
+	return out
+}
 
+// setSelectedRoutes swaps the accessor used to classify overlay-routed
+// upstreams. Called when route sources are wired after the handler was
+// built (permanent / iOS constructors).
+func (u *upstreamResolverBase) setSelectedRoutes(selected func() route.HAMap) {
+	u.selectedRoutes = selected
+}
+
+func (u *upstreamResolverBase) addRace(servers []netip.AddrPort) {
+	if len(servers) == 0 {
+		return
+	}
+	u.upstreamServers = append(u.upstreamServers, slices.Clone(servers))
 }
 
 // ServeDNS handles a DNS request
@@ -242,82 +314,201 @@ func (u *upstreamResolverBase) prepareRequest(r *dns.Msg) {
 }
 
 func (u *upstreamResolverBase) tryUpstreamServers(ctx context.Context, w dns.ResponseWriter, r *dns.Msg, logger *log.Entry) (bool, []upstreamFailure) {
-	timeout := u.upstreamTimeout
-	if len(u.upstreamServers) > 1 {
-		maxTotal := 5 * time.Second
-		minPerUpstream := 2 * time.Second
-		scaledTimeout := maxTotal / time.Duration(len(u.upstreamServers))
-		if scaledTimeout > minPerUpstream {
-			timeout = scaledTimeout
-		} else {
-			timeout = minPerUpstream
-		}
+	groups := u.upstreamServers
+	switch len(groups) {
+	case 0:
+		return false, nil
+	case 1:
+		return u.tryOnlyRace(ctx, w, r, groups[0], logger)
+	default:
+		return u.raceAll(ctx, w, r, groups, logger)
+	}
+}
+
+func (u *upstreamResolverBase) tryOnlyRace(ctx context.Context, w dns.ResponseWriter, r *dns.Msg, group upstreamRace, logger *log.Entry) (bool, []upstreamFailure) {
+	res := u.tryRace(ctx, r, group)
+	if res.msg == nil {
+		return false, res.failures
+	}
+	if res.ede != "" {
+		resutil.SetMeta(w, "ede", res.ede)
+	}
+	u.writeSuccessResponse(w, res.msg, res.upstream, r.Question[0].Name, res.protocol, logger)
+	return true, res.failures
+}
+
+// raceAll runs one worker per group in parallel, taking the first valid
+// answer and cancelling the rest.
+func (u *upstreamResolverBase) raceAll(ctx context.Context, w dns.ResponseWriter, r *dns.Msg, groups []upstreamRace, logger *log.Entry) (bool, []upstreamFailure) {
+	raceCtx, cancel := context.WithCancel(ctx)
+	defer cancel()
+
+	// Buffer sized to len(groups) so workers never block on send, even
+	// after the coordinator has returned.
+	results := make(chan raceResult, len(groups))
+	for _, g := range groups {
+		// tryRace clones the request per attempt, so workers never share
+		// a *dns.Msg and concurrent EDNS0 mutations can't race.
+		go func(g upstreamRace) {
+			results <- u.tryRace(raceCtx, r, g)
+		}(g)
 	}
 
 	var failures []upstreamFailure
-	for _, upstream := range u.upstreamServers {
-		if failure := u.queryUpstream(ctx, w, r, upstream, timeout, logger); failure != nil {
-			failures = append(failures, *failure)
-		} else {
-			return true, failures
+	for range groups {
+		select {
+		case res := <-results:
+			failures = append(failures, res.failures...)
+			if res.msg != nil {
+				if res.ede != "" {
+					resutil.SetMeta(w, "ede", res.ede)
+				}
+				u.writeSuccessResponse(w, res.msg, res.upstream, r.Question[0].Name, res.protocol, logger)
+				return true, failures
+			}
+		case <-ctx.Done():
+			return false, failures
 		}
 	}
 	return false, failures
 }
 
-// queryUpstream queries a single upstream server. Returns nil on success, or failure info to try next upstream.
-func (u *upstreamResolverBase) queryUpstream(parentCtx context.Context, w dns.ResponseWriter, r *dns.Msg, upstream netip.AddrPort, timeout time.Duration, logger *log.Entry) *upstreamFailure {
-	var rm *dns.Msg
-	var t time.Duration
-	var err error
+func (u *upstreamResolverBase) tryRace(ctx context.Context, r *dns.Msg, group upstreamRace) raceResult {
+	timeout := u.upstreamTimeout
+	if len(group) > 1 {
+		// Cap the whole walk at raceMaxTotalTimeout: per-upstream timeouts
+		// still honor raceMinPerUpstreamTimeout as a floor for correctness
+		// on slow links, but the outer context ensures the combined walk
+		// cannot exceed the cap regardless of group size.
+		timeout = max(raceMaxTotalTimeout/time.Duration(len(group)), raceMinPerUpstreamTimeout)
+		var cancel context.CancelFunc
+		ctx, cancel = context.WithTimeout(ctx, raceMaxTotalTimeout)
+		defer cancel()
+	}
+
+	var failures []upstreamFailure
+	for _, upstream := range group {
+		if ctx.Err() != nil {
+			return raceResult{failures: failures}
+		}
+		// Clone the request per attempt: the exchange path mutates EDNS0
+		// options in-place, so reusing the same *dns.Msg across sequential
+		// upstreams would carry those mutations (e.g. a reduced UDP size)
+		// into the next attempt.
+		res, failure := u.queryUpstream(ctx, r.Copy(), upstream, timeout)
+		if failure != nil {
+			failures = append(failures, *failure)
+			continue
+		}
+		res.failures = failures
+		return res
+	}
+	return raceResult{failures: failures}
+}
+
+func (u *upstreamResolverBase) queryUpstream(parentCtx context.Context, r *dns.Msg, upstream netip.AddrPort, timeout time.Duration) (raceResult, *upstreamFailure) {
+	ctx, cancel := context.WithTimeout(parentCtx, timeout)
+	defer cancel()
+	ctx, upstreamProto := contextWithUpstreamProtocolResult(ctx)
 
 	// Advertise EDNS0 so the upstream may include Extended DNS Errors
 	// (RFC 8914) in failure responses; we use those to short-circuit
 	// failover for definitive answers like DNSSEC validation failures.
-	// Operate on a copy so the inbound request is unchanged: a client that
-	// did not advertise EDNS0 must not see an OPT in the response.
+	// The caller already passed a per-attempt copy, so we can mutate r
+	// directly; hadEdns reflects the original client request's state and
+	// controls whether we strip the OPT from the response.
 	hadEdns := r.IsEdns0() != nil
-	reqUp := r
 	if !hadEdns {
-		reqUp = r.Copy()
-		reqUp.SetEdns0(upstreamUDPSize(), false)
+		r.SetEdns0(upstreamUDPSize(), false)
 	}
 
-	var startTime time.Time
-	var upstreamProto *upstreamProtocolResult
-	func() {
-		ctx, cancel := context.WithTimeout(parentCtx, timeout)
-		defer cancel()
-		ctx, upstreamProto = contextWithupstreamProtocolResult(ctx)
-		startTime = time.Now()
-		rm, t, err = u.upstreamClient.exchange(ctx, upstream.String(), reqUp)
-	}()
+	startTime := time.Now()
+	rm, _, err := u.upstreamClient.exchange(ctx, upstream.String(), r)
 
 	if err != nil {
-		return u.handleUpstreamError(err, upstream, startTime)
+		// A parent cancellation (e.g., another race won and the coordinator
+		// cancelled the losers) is not an upstream failure. Check both the
+		// error chain and the parent context: a transport may surface the
+		// cancellation as a read/deadline error rather than context.Canceled.
+		if errors.Is(err, context.Canceled) || errors.Is(parentCtx.Err(), context.Canceled) {
+			return raceResult{}, &upstreamFailure{upstream: upstream, reason: "canceled"}
+		}
+		failure := u.handleUpstreamError(err, upstream, startTime)
+		u.markUpstreamFail(upstream, failure.reason)
+		return raceResult{}, failure
 	}
 
 	if rm == nil || !rm.Response {
-		return &upstreamFailure{upstream: upstream, reason: "no response"}
+		u.markUpstreamFail(upstream, "no response")
+		return raceResult{}, &upstreamFailure{upstream: upstream, reason: "no response"}
+	}
+
+	proto := ""
+	if upstreamProto != nil {
+		proto = upstreamProto.protocol
 	}
 
 	if rm.Rcode == dns.RcodeServerFailure || rm.Rcode == dns.RcodeRefused {
 		if code, ok := nonRetryableEDE(rm); ok {
-			resutil.SetMeta(w, "ede", edeName(code))
 			if !hadEdns {
 				stripOPT(rm)
 			}
-			u.writeSuccessResponse(w, rm, upstream, r.Question[0].Name, t, upstreamProto, logger)
-			return nil
+			u.markUpstreamOk(upstream)
+			return raceResult{msg: rm, upstream: upstream, protocol: proto, ede: edeName(code)}, nil
 		}
-		return &upstreamFailure{upstream: upstream, reason: dns.RcodeToString[rm.Rcode]}
+		reason := dns.RcodeToString[rm.Rcode]
+		u.markUpstreamFail(upstream, reason)
+		return raceResult{}, &upstreamFailure{upstream: upstream, reason: reason}
 	}
 
 	if !hadEdns {
 		stripOPT(rm)
 	}
-	u.writeSuccessResponse(w, rm, upstream, r.Question[0].Name, t, upstreamProto, logger)
-	return nil
+
+	u.markUpstreamOk(upstream)
+	return raceResult{msg: rm, upstream: upstream, protocol: proto}, nil
+}
+
+// healthEntry returns the mutable health record for addr, lazily creating
+// the map and the entry. Caller must hold u.healthMu.
+func (u *upstreamResolverBase) healthEntry(addr netip.AddrPort) *UpstreamHealth {
+	if u.health == nil {
+		u.health = make(map[netip.AddrPort]*UpstreamHealth)
+	}
+	h := u.health[addr]
+	if h == nil {
+		h = &UpstreamHealth{}
+		u.health[addr] = h
+	}
+	return h
+}
+
+func (u *upstreamResolverBase) markUpstreamOk(addr netip.AddrPort) {
+	u.healthMu.Lock()
+	defer u.healthMu.Unlock()
+	h := u.healthEntry(addr)
+	h.LastOk = time.Now()
+	h.LastFail = time.Time{}
+	h.LastErr = ""
+}
+
+func (u *upstreamResolverBase) markUpstreamFail(addr netip.AddrPort, reason string) {
+	u.healthMu.Lock()
+	defer u.healthMu.Unlock()
+	h := u.healthEntry(addr)
+	h.LastFail = time.Now()
+	h.LastErr = reason
+}
+
+// UpstreamHealth returns a snapshot of per-upstream query outcomes.
+func (u *upstreamResolverBase) UpstreamHealth() map[netip.AddrPort]UpstreamHealth {
+	u.healthMu.RLock()
+	defer u.healthMu.RUnlock()
+	out := make(map[netip.AddrPort]UpstreamHealth, len(u.health))
+	for k, v := range u.health {
+		out[k] = *v
+	}
+	return out
 }
 
 // upstreamUDPSize returns the EDNS0 UDP buffer size we advertise to upstreams,
@@ -358,12 +549,23 @@ func (u *upstreamResolverBase) handleUpstreamError(err error, upstream netip.Add
 	return &upstreamFailure{upstream: upstream, reason: reason}
 }
 
-func (u *upstreamResolverBase) writeSuccessResponse(w dns.ResponseWriter, rm *dns.Msg, upstream netip.AddrPort, domain string, t time.Duration, upstreamProto *upstreamProtocolResult, logger *log.Entry) bool {
-	u.successCount.Add(1)
+func (u *upstreamResolverBase) debugUpstreamTimeout(upstream netip.AddrPort) string {
+	if u.statusRecorder == nil {
+		return ""
+	}
 
+	peerInfo := findPeerForIP(upstream.Addr(), u.statusRecorder)
+	if peerInfo == nil {
+		return ""
+	}
+
+	return fmt.Sprintf("(routes through NetBird peer %s)", FormatPeerStatus(peerInfo))
+}
+
+func (u *upstreamResolverBase) writeSuccessResponse(w dns.ResponseWriter, rm *dns.Msg, upstream netip.AddrPort, domain string, proto string, logger *log.Entry) {
 	resutil.SetMeta(w, "upstream", upstream.String())
-	if upstreamProto != nil && upstreamProto.protocol != "" {
-		resutil.SetMeta(w, "upstream_protocol", upstreamProto.protocol)
+	if proto != "" {
+		resutil.SetMeta(w, "upstream_protocol", proto)
 	}
 
 	// Clear Zero bit from external responses to prevent upstream servers from
@@ -372,14 +574,11 @@ func (u *upstreamResolverBase) writeSuccessResponse(w dns.ResponseWriter, rm *dn
 
 	if err := w.WriteMsg(rm); err != nil {
 		logger.Errorf("failed to write DNS response for question domain=%s: %s", domain, err)
-		return true
 	}
-
-	return true
 }
 
 func (u *upstreamResolverBase) logUpstreamFailures(domain string, failures []upstreamFailure, succeeded bool, logger *log.Entry) {
-	totalUpstreams := len(u.upstreamServers)
+	totalUpstreams := len(u.flatUpstreams())
 	failedCount := len(failures)
 	failureSummary := formatFailures(failures)
 
@@ -434,119 +633,6 @@ func edeName(code uint16) string {
 	return fmt.Sprintf("EDE %d", code)
 }
 
-// ProbeAvailability tests all upstream servers simultaneously and
-// disables the resolver if none work
-func (u *upstreamResolverBase) ProbeAvailability(ctx context.Context) {
-	u.mutex.Lock()
-	defer u.mutex.Unlock()
-
-	// avoid probe if upstreams could resolve at least one query
-	if u.successCount.Load() > 0 {
-		return
-	}
-
-	var success bool
-	var mu sync.Mutex
-	var wg sync.WaitGroup
-
-	var errs *multierror.Error
-	for _, upstream := range u.upstreamServers {
-		wg.Add(1)
-		go func(upstream netip.AddrPort) {
-			defer wg.Done()
-			err := u.testNameserver(u.ctx, ctx, upstream, 500*time.Millisecond)
-			if err != nil {
-				mu.Lock()
-				errs = multierror.Append(errs, err)
-				mu.Unlock()
-				log.Warnf("probing upstream nameserver %s: %s", upstream, err)
-				return
-			}
-
-			mu.Lock()
-			success = true
-			mu.Unlock()
-		}(upstream)
-	}
-
-	wg.Wait()
-
-	select {
-	case <-ctx.Done():
-		return
-	case <-u.ctx.Done():
-		return
-	default:
-	}
-
-	// didn't find a working upstream server, let's disable and try later
-	if !success {
-		u.disable(errs.ErrorOrNil())
-
-		if u.statusRecorder == nil {
-			return
-		}
-
-		u.statusRecorder.PublishEvent(
-			proto.SystemEvent_WARNING,
-			proto.SystemEvent_DNS,
-			"All upstream servers failed (probe failed)",
-			"Unable to reach one or more DNS servers. This might affect your ability to connect to some services.",
-			map[string]string{"upstreams": u.upstreamServersString()},
-		)
-	}
-}
-
-// waitUntilResponse retries, in an exponential interval, querying the upstream servers until it gets a positive response
-func (u *upstreamResolverBase) waitUntilResponse() {
-	exponentialBackOff := &backoff.ExponentialBackOff{
-		InitialInterval:     500 * time.Millisecond,
-		RandomizationFactor: 0.5,
-		Multiplier:          1.1,
-		MaxInterval:         u.reactivatePeriod,
-		MaxElapsedTime:      0,
-		Stop:                backoff.Stop,
-		Clock:               backoff.SystemClock,
-	}
-
-	operation := func() error {
-		select {
-		case <-u.ctx.Done():
-			return backoff.Permanent(fmt.Errorf("exiting upstream retry loop for upstreams %s: parent context has been canceled", u.upstreamServersString()))
-		default:
-		}
-
-		for _, upstream := range u.upstreamServers {
-			if err := u.testNameserver(u.ctx, nil, upstream, probeTimeout); err != nil {
-				log.Tracef("upstream check for %s: %s", upstream, err)
-			} else {
-				// at least one upstream server is available, stop probing
-				return nil
-			}
-		}
-
-		log.Tracef("checking connectivity with upstreams %s failed. Retrying in %s", u.upstreamServersString(), exponentialBackOff.NextBackOff())
-		return fmt.Errorf("upstream check call error")
-	}
-
-	err := backoff.Retry(operation, backoff.WithContext(exponentialBackOff, u.ctx))
-	if err != nil {
-		if errors.Is(err, context.Canceled) {
-			log.Debugf("upstream retry loop exited for upstreams %s", u.upstreamServersString())
-		} else {
-			log.Warnf("upstream retry loop exited for upstreams %s: %v", u.upstreamServersString(), err)
-		}
-		return
-	}
-
-	log.Infof("upstreams %s are responsive again. Adding them back to system", u.upstreamServersString())
-	u.successCount.Add(1)
-	u.reactivate()
-	u.mutex.Lock()
-	u.disabled = false
-	u.mutex.Unlock()
-}
-
 // isTimeout returns true if the given error is a network timeout error.
 //
 // Copied from k8s.io/apimachinery/pkg/util/net.IsTimeout
@@ -558,45 +644,6 @@ func isTimeout(err error) bool {
 	return false
 }
 
-func (u *upstreamResolverBase) disable(err error) {
-	if u.disabled {
-		return
-	}
-
-	log.Warnf("Upstream resolving is Disabled for %v", reactivatePeriod)
-	u.successCount.Store(0)
-	u.deactivate(err)
-	u.disabled = true
-	u.wg.Add(1)
-	go func() {
-		defer u.wg.Done()
-		u.waitUntilResponse()
-	}()
-}
-
-func (u *upstreamResolverBase) upstreamServersString() string {
-	var servers []string
-	for _, server := range u.upstreamServers {
-		servers = append(servers, server.String())
-	}
-	return strings.Join(servers, ", ")
-}
-
-func (u *upstreamResolverBase) testNameserver(baseCtx context.Context, externalCtx context.Context, server netip.AddrPort, timeout time.Duration) error {
-	mergedCtx, cancel := context.WithTimeout(baseCtx, timeout)
-	defer cancel()
-
-	if externalCtx != nil {
-		stop2 := context.AfterFunc(externalCtx, cancel)
-		defer stop2()
-	}
-
-	r := new(dns.Msg).SetQuestion(testRecord, dns.TypeSOA)
-
-	_, _, err := u.upstreamClient.exchange(mergedCtx, server.String(), r)
-	return err
-}
-
 // clientUDPMaxSize returns the maximum UDP response size the client accepts.
 func clientUDPMaxSize(r *dns.Msg) int {
 	if opt := r.IsEdns0(); opt != nil {
@@ -608,13 +655,10 @@ func clientUDPMaxSize(r *dns.Msg) int {
 // ExchangeWithFallback exchanges a DNS message with the upstream server.
 // It first tries to use UDP, and if it is truncated, it falls back to TCP.
 // If the inbound request came over TCP (via context), it skips the UDP attempt.
-// If the passed context is nil, this will use Exchange instead of ExchangeContext.
 func ExchangeWithFallback(ctx context.Context, client *dns.Client, r *dns.Msg, upstream string) (*dns.Msg, time.Duration, error) {
 	// If the request came in over TCP, go straight to TCP upstream.
 	if dnsProtocolFromContext(ctx) == protoTCP {
-		tcpClient := *client
-		tcpClient.Net = protoTCP
-		rm, t, err := tcpClient.ExchangeContext(ctx, r, upstream)
+		rm, t, err := toTCPClient(client).ExchangeContext(ctx, r, upstream)
 		if err != nil {
 			return nil, t, fmt.Errorf("with tcp: %w", err)
 		}
@@ -634,18 +678,7 @@ func ExchangeWithFallback(ctx context.Context, client *dns.Client, r *dns.Msg, u
 		opt.SetUDPSize(maxUDPPayload)
 	}
 
-	var (
-		rm  *dns.Msg
-		t   time.Duration
-		err error
-	)
-
-	if ctx == nil {
-		rm, t, err = client.Exchange(r, upstream)
-	} else {
-		rm, t, err = client.ExchangeContext(ctx, r, upstream)
-	}
-
+	rm, t, err := client.ExchangeContext(ctx, r, upstream)
 	if err != nil {
 		return nil, t, fmt.Errorf("with udp: %w", err)
 	}
@@ -659,15 +692,7 @@ func ExchangeWithFallback(ctx context.Context, client *dns.Client, r *dns.Msg, u
 	// data than the client's buffer, we could truncate locally and skip
 	// the TCP retry.
 
-	tcpClient := *client
-	tcpClient.Net = protoTCP
-
-	if ctx == nil {
-		rm, t, err = tcpClient.Exchange(r, upstream)
-	} else {
-		rm, t, err = tcpClient.ExchangeContext(ctx, r, upstream)
-	}
-
+	rm, t, err = toTCPClient(client).ExchangeContext(ctx, r, upstream)
 	if err != nil {
 		return nil, t, fmt.Errorf("with tcp: %w", err)
 	}
@@ -681,6 +706,25 @@ func ExchangeWithFallback(ctx context.Context, client *dns.Client, r *dns.Msg, u
 	return rm, t, nil
 }
 
+// toTCPClient returns a copy of c configured for TCP. If c's Dialer has a
+// *net.UDPAddr bound as LocalAddr (iOS does this to keep the source IP on
+// the tunnel interface), it is converted to the equivalent *net.TCPAddr
+// so net.Dialer doesn't reject the TCP dial with "mismatched local
+// address type".
+func toTCPClient(c *dns.Client) *dns.Client {
+	tcp := *c
+	tcp.Net = protoTCP
+	if tcp.Dialer == nil {
+		return &tcp
+	}
+	d := *tcp.Dialer
+	if ua, ok := d.LocalAddr.(*net.UDPAddr); ok {
+		d.LocalAddr = &net.TCPAddr{IP: ua.IP, Port: ua.Port, Zone: ua.Zone}
+	}
+	tcp.Dialer = &d
+	return &tcp
+}
+
 // ExchangeWithNetstack performs a DNS exchange using netstack for dialing.
 // This is needed when netstack is enabled to reach peer IPs through the tunnel.
 func ExchangeWithNetstack(ctx context.Context, nsNet *netstack.Net, r *dns.Msg, upstream string) (*dns.Msg, error) {
@@ -822,15 +866,36 @@ func findPeerForIP(ip netip.Addr, statusRecorder *peer.Status) *peer.State {
 	return bestMatch
 }
 
-func (u *upstreamResolverBase) debugUpstreamTimeout(upstream netip.AddrPort) string {
-	if u.statusRecorder == nil {
-		return ""
+// haMapRouteCount returns the total number of routes across all HA
+// groups in the map. route.HAMap is keyed by HAUniqueID with slices of
+// routes per key, so len(hm) is the number of HA groups, not routes.
+func haMapRouteCount(hm route.HAMap) int {
+	total := 0
+	for _, routes := range hm {
+		total += len(routes)
 	}
-
-	peerInfo := findPeerForIP(upstream.Addr(), u.statusRecorder)
-	if peerInfo == nil {
-		return ""
-	}
-
-	return fmt.Sprintf("(routes through NetBird peer %s)", FormatPeerStatus(peerInfo))
+	return total
+}
+
+// haMapContains checks whether ip is covered by any concrete prefix in
+// the HA map. haveDynamic is reported separately: dynamic (domain-based)
+// routes carry a placeholder Network that can't be prefix-checked, so we
+// can't know at this point whether ip is reached through one. Callers
+// decide how to interpret the unknown: health projection treats it as
+// "possibly routed" to avoid emitting false-positive warnings during
+// startup, while iOS dial selection requires a concrete match before
+// binding to the tunnel.
+func haMapContains(hm route.HAMap, ip netip.Addr) (matched, haveDynamic bool) {
+	for _, routes := range hm {
+		for _, r := range routes {
+			if r.IsDynamic() {
+				haveDynamic = true
+				continue
+			}
+			if r.Network.Contains(ip) {
+				return true, haveDynamic
+			}
+		}
+	}
+	return false, haveDynamic
 }

client/internal/dns/upstream_android.go

@@ -11,6 +11,7 @@ import (
 
 	"github.com/netbirdio/netbird/client/internal/peer"
 	nbnet "github.com/netbirdio/netbird/client/net"
+	"github.com/netbirdio/netbird/shared/management/domain"
 )
 
 type upstreamResolver struct {
@@ -26,9 +27,9 @@ func newUpstreamResolver(
 	_ WGIface,
 	statusRecorder *peer.Status,
 	hostsDNSHolder *hostsDNSHolder,
-	domain string,
+	d domain.Domain,
 ) (*upstreamResolver, error) {
-	upstreamResolverBase := newUpstreamResolverBase(ctx, statusRecorder, domain)
+	upstreamResolverBase := newUpstreamResolverBase(ctx, statusRecorder, d)
 	c := &upstreamResolver{
 		upstreamResolverBase: upstreamResolverBase,
 		hostsDNSHolder:       hostsDNSHolder,

client/internal/dns/upstream_general.go

@@ -12,6 +12,7 @@ import (
 	"golang.zx2c4.com/wireguard/tun/netstack"
 
 	"github.com/netbirdio/netbird/client/internal/peer"
+	"github.com/netbirdio/netbird/shared/management/domain"
 )
 
 type upstreamResolver struct {
@@ -24,9 +25,9 @@ func newUpstreamResolver(
 	wgIface WGIface,
 	statusRecorder *peer.Status,
 	_ *hostsDNSHolder,
-	domain string,
+	d domain.Domain,
 ) (*upstreamResolver, error) {
-	upstreamResolverBase := newUpstreamResolverBase(ctx, statusRecorder, domain)
+	upstreamResolverBase := newUpstreamResolverBase(ctx, statusRecorder, d)
 	nonIOS := &upstreamResolver{
 		upstreamResolverBase: upstreamResolverBase,
 		nsNet:                wgIface.GetNet(),

client/internal/dns/upstream_ios.go

@@ -15,6 +15,7 @@ import (
 	"golang.org/x/sys/unix"
 
 	"github.com/netbirdio/netbird/client/internal/peer"
+	"github.com/netbirdio/netbird/shared/management/domain"
 )
 
 type upstreamResolverIOS struct {
@@ -27,9 +28,9 @@ func newUpstreamResolver(
 	wgIface WGIface,
 	statusRecorder *peer.Status,
 	_ *hostsDNSHolder,
-	domain string,
+	d domain.Domain,
 ) (*upstreamResolverIOS, error) {
-	upstreamResolverBase := newUpstreamResolverBase(ctx, statusRecorder, domain)
+	upstreamResolverBase := newUpstreamResolverBase(ctx, statusRecorder, d)
 
 	ios := &upstreamResolverIOS{
 		upstreamResolverBase: upstreamResolverBase,
@@ -62,9 +63,16 @@ func (u *upstreamResolverIOS) exchange(ctx context.Context, upstream string, r *
 		upstreamIP = upstreamIP.Unmap()
 	}
 	addr := u.wgIface.Address()
+	var routed bool
+	if u.selectedRoutes != nil {
+		// Only a concrete prefix match binds to the tunnel: dialing
+		// through a private client for an upstream we can't prove is
+		// routed would break public resolvers.
+		routed, _ = haMapContains(u.selectedRoutes(), upstreamIP)
+	}
 	needsPrivate := addr.Network.Contains(upstreamIP) ||
 		addr.IPv6Net.Contains(upstreamIP) ||
-		(u.routeMatch != nil && u.routeMatch(upstreamIP))
+		routed
 	if needsPrivate {
 		log.Debugf("using private client to query %s via upstream %s", r.Question[0].Name, upstream)
 		client, err = GetClientPrivate(u.wgIface, upstreamIP, timeout)
@@ -73,8 +81,7 @@ func (u *upstreamResolverIOS) exchange(ctx context.Context, upstream string, r *
 		}
 	}
 
-	// Cannot use client.ExchangeContext because it overwrites our Dialer
-	return ExchangeWithFallback(nil, client, r, upstream)
+	return ExchangeWithFallback(ctx, client, r, upstream)
 }
 
 // GetClientPrivate returns a new DNS client bound to the local IP of the Netbird interface.

client/internal/dns/upstream_test.go

@@ -6,6 +6,7 @@ import (
 	"net"
 	"net/netip"
 	"strings"
+	"sync/atomic"
 	"testing"
 	"time"
 
@@ -73,7 +74,7 @@ func TestUpstreamResolver_ServeDNS(t *testing.T) {
 					servers = append(servers, netip.AddrPortFrom(addrPort.Addr().Unmap(), addrPort.Port()))
 				}
 			}
-			resolver.upstreamServers = servers
+			resolver.addRace(servers)
 			resolver.upstreamTimeout = testCase.timeout
 			if testCase.cancelCTX {
 				cancel()
@@ -132,20 +133,10 @@ func (m *mockNetstackProvider) GetInterfaceGUIDString() (string, error) {
 	return "", nil
 }
 
-type mockUpstreamResolver struct {
-	r   *dns.Msg
-	rtt time.Duration
-	err error
-}
-
-// exchange mock implementation of exchange from upstreamResolver
-func (c mockUpstreamResolver) exchange(_ context.Context, _ string, _ *dns.Msg) (*dns.Msg, time.Duration, error) {
-	return c.r, c.rtt, c.err
-}
-
 type mockUpstreamResponse struct {
-	msg *dns.Msg
-	err error
+	msg   *dns.Msg
+	err   error
+	delay time.Duration
 }
 
 type mockUpstreamResolverPerServer struct {
@@ -153,63 +144,19 @@ type mockUpstreamResolverPerServer struct {
 	rtt       time.Duration
 }
 
-func (c mockUpstreamResolverPerServer) exchange(_ context.Context, upstream string, _ *dns.Msg) (*dns.Msg, time.Duration, error) {
-	if r, ok := c.responses[upstream]; ok {
-		return r.msg, c.rtt, r.err
+func (c mockUpstreamResolverPerServer) exchange(ctx context.Context, upstream string, _ *dns.Msg) (*dns.Msg, time.Duration, error) {
+	r, ok := c.responses[upstream]
+	if !ok {
+		return nil, c.rtt, fmt.Errorf("no mock response for %s", upstream)
 	}
-	return nil, c.rtt, fmt.Errorf("no mock response for %s", upstream)
-}
-
-func TestUpstreamResolver_DeactivationReactivation(t *testing.T) {
-	mockClient := &mockUpstreamResolver{
-		err: dns.ErrTime,
-		r:   new(dns.Msg),
-		rtt: time.Millisecond,
-	}
-
-	resolver := &upstreamResolverBase{
-		ctx:              context.TODO(),
-		upstreamClient:   mockClient,
-		upstreamTimeout:  UpstreamTimeout,
-		reactivatePeriod: time.Microsecond * 100,
-	}
-	addrPort, _ := netip.ParseAddrPort("0.0.0.0:1") // Use valid port for parsing, test will still fail on connection
-	resolver.upstreamServers = []netip.AddrPort{netip.AddrPortFrom(addrPort.Addr().Unmap(), addrPort.Port())}
-
-	failed := false
-	resolver.deactivate = func(error) {
-		failed = true
-		// After deactivation, make the mock client work again
-		mockClient.err = nil
-	}
-
-	reactivated := false
-	resolver.reactivate = func() {
-		reactivated = true
-	}
-
-	resolver.ProbeAvailability(context.TODO())
-
-	if !failed {
-		t.Errorf("expected that resolving was deactivated")
-		return
-	}
-
-	if !resolver.disabled {
-		t.Errorf("resolver should be Disabled")
-		return
-	}
-
-	time.Sleep(time.Millisecond * 200)
-
-	if !reactivated {
-		t.Errorf("expected that resolving was reactivated")
-		return
-	}
-
-	if resolver.disabled {
-		t.Errorf("should be enabled")
+	if r.delay > 0 {
+		select {
+		case <-time.After(r.delay):
+		case <-ctx.Done():
+			return nil, c.rtt, ctx.Err()
+		}
 	}
+	return r.msg, c.rtt, r.err
 }
 
 func TestUpstreamResolver_Failover(t *testing.T) {
@@ -339,9 +286,9 @@ func TestUpstreamResolver_Failover(t *testing.T) {
 			resolver := &upstreamResolverBase{
 				ctx:             ctx,
 				upstreamClient:  trackingClient,
-				upstreamServers: []netip.AddrPort{upstream1, upstream2},
 				upstreamTimeout: UpstreamTimeout,
 			}
+			resolver.addRace([]netip.AddrPort{upstream1, upstream2})
 
 			var responseMSG *dns.Msg
 			responseWriter := &test.MockResponseWriter{
@@ -421,9 +368,9 @@ func TestUpstreamResolver_SingleUpstreamFailure(t *testing.T) {
 	resolver := &upstreamResolverBase{
 		ctx:             ctx,
 		upstreamClient:  mockClient,
-		upstreamServers: []netip.AddrPort{upstream},
 		upstreamTimeout: UpstreamTimeout,
 	}
+	resolver.addRace([]netip.AddrPort{upstream})
 
 	var responseMSG *dns.Msg
 	responseWriter := &test.MockResponseWriter{
@@ -440,6 +387,136 @@ func TestUpstreamResolver_SingleUpstreamFailure(t *testing.T) {
 	assert.Equal(t, dns.RcodeServerFailure, responseMSG.Rcode, "single upstream SERVFAIL should return SERVFAIL")
 }
 
+// TestUpstreamResolver_RaceAcrossGroups covers two nameserver groups
+// configured for the same domain, with one broken group. The merge+race
+// path should answer as fast as the working group and not pay the timeout
+// of the broken one on every query.
+func TestUpstreamResolver_RaceAcrossGroups(t *testing.T) {
+	broken := netip.MustParseAddrPort("192.0.2.1:53")
+	working := netip.MustParseAddrPort("192.0.2.2:53")
+	successAnswer := "192.0.2.100"
+	timeoutErr := &net.OpError{Op: "read", Err: fmt.Errorf("i/o timeout")}
+
+	mockClient := &mockUpstreamResolverPerServer{
+		responses: map[string]mockUpstreamResponse{
+			// Force the broken upstream to only unblock via timeout /
+			// cancellation so the assertion below can't pass if races
+			// were run serially.
+			broken.String():  {err: timeoutErr, delay: 500 * time.Millisecond},
+			working.String(): {msg: buildMockResponse(dns.RcodeSuccess, successAnswer)},
+		},
+		rtt: time.Millisecond,
+	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	resolver := &upstreamResolverBase{
+		ctx:             ctx,
+		upstreamClient:  mockClient,
+		upstreamTimeout: 250 * time.Millisecond,
+	}
+	resolver.addRace([]netip.AddrPort{broken})
+	resolver.addRace([]netip.AddrPort{working})
+
+	var responseMSG *dns.Msg
+	responseWriter := &test.MockResponseWriter{
+		WriteMsgFunc: func(m *dns.Msg) error {
+			responseMSG = m
+			return nil
+		},
+	}
+
+	inputMSG := new(dns.Msg).SetQuestion("example.com.", dns.TypeA)
+	start := time.Now()
+	resolver.ServeDNS(responseWriter, inputMSG)
+	elapsed := time.Since(start)
+
+	require.NotNil(t, responseMSG, "should write a response")
+	assert.Equal(t, dns.RcodeSuccess, responseMSG.Rcode)
+	require.NotEmpty(t, responseMSG.Answer)
+	assert.Contains(t, responseMSG.Answer[0].String(), successAnswer)
+	// Working group answers in a single RTT; the broken group's
+	// timeout (100ms) must not block the response.
+	assert.Less(t, elapsed, 100*time.Millisecond, "race must not wait for broken group's timeout")
+}
+
+// TestUpstreamResolver_AllGroupsFail checks that when every group fails the
+// resolver returns SERVFAIL rather than leaking a partial response.
+func TestUpstreamResolver_AllGroupsFail(t *testing.T) {
+	a := netip.MustParseAddrPort("192.0.2.1:53")
+	b := netip.MustParseAddrPort("192.0.2.2:53")
+
+	mockClient := &mockUpstreamResolverPerServer{
+		responses: map[string]mockUpstreamResponse{
+			a.String(): {msg: buildMockResponse(dns.RcodeServerFailure, "")},
+			b.String(): {msg: buildMockResponse(dns.RcodeServerFailure, "")},
+		},
+		rtt: time.Millisecond,
+	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	resolver := &upstreamResolverBase{
+		ctx:             ctx,
+		upstreamClient:  mockClient,
+		upstreamTimeout: UpstreamTimeout,
+	}
+	resolver.addRace([]netip.AddrPort{a})
+	resolver.addRace([]netip.AddrPort{b})
+
+	var responseMSG *dns.Msg
+	responseWriter := &test.MockResponseWriter{
+		WriteMsgFunc: func(m *dns.Msg) error {
+			responseMSG = m
+			return nil
+		},
+	}
+
+	resolver.ServeDNS(responseWriter, new(dns.Msg).SetQuestion("example.com.", dns.TypeA))
+	require.NotNil(t, responseMSG)
+	assert.Equal(t, dns.RcodeServerFailure, responseMSG.Rcode)
+}
+
+// TestUpstreamResolver_HealthTracking verifies that query-path results are
+// recorded into per-upstream health, which is what projects back to
+// NSGroupState for status reporting.
+func TestUpstreamResolver_HealthTracking(t *testing.T) {
+	ok := netip.MustParseAddrPort("192.0.2.10:53")
+	bad := netip.MustParseAddrPort("192.0.2.11:53")
+
+	mockClient := &mockUpstreamResolverPerServer{
+		responses: map[string]mockUpstreamResponse{
+			ok.String():  {msg: buildMockResponse(dns.RcodeSuccess, "192.0.2.100")},
+			bad.String(): {msg: buildMockResponse(dns.RcodeServerFailure, "")},
+		},
+		rtt: time.Millisecond,
+	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	resolver := &upstreamResolverBase{
+		ctx:             ctx,
+		upstreamClient:  mockClient,
+		upstreamTimeout: UpstreamTimeout,
+	}
+	resolver.addRace([]netip.AddrPort{ok, bad})
+
+	responseWriter := &test.MockResponseWriter{WriteMsgFunc: func(m *dns.Msg) error { return nil }}
+	resolver.ServeDNS(responseWriter, new(dns.Msg).SetQuestion("example.com.", dns.TypeA))
+
+	health := resolver.UpstreamHealth()
+	require.Contains(t, health, ok)
+	assert.False(t, health[ok].LastOk.IsZero(), "ok upstream should have LastOk set")
+	assert.Empty(t, health[ok].LastErr)
+
+	// bad upstream was never tried because ok answered first; its health
+	// should remain unset.
+	assert.NotContains(t, health, bad, "sibling upstream should not be queried when primary answers")
+}
+
 func TestFormatFailures(t *testing.T) {
 	testCases := []struct {
 		name     string
@@ -665,10 +742,10 @@ func TestExchangeWithFallback_EDNS0Capped(t *testing.T) {
 	// Verify that a client EDNS0 larger than our MTU-derived limit gets
 	// capped in the outgoing request so the upstream doesn't send a
 	// response larger than our read buffer.
-	var receivedUDPSize uint16
+	var receivedUDPSize atomic.Uint32
 	udpHandler := dns.HandlerFunc(func(w dns.ResponseWriter, r *dns.Msg) {
 		if opt := r.IsEdns0(); opt != nil {
-			receivedUDPSize = opt.UDPSize()
+			receivedUDPSize.Store(uint32(opt.UDPSize()))
 		}
 		m := new(dns.Msg)
 		m.SetReply(r)
@@ -699,7 +776,7 @@ func TestExchangeWithFallback_EDNS0Capped(t *testing.T) {
 	require.NotNil(t, rm)
 
 	expectedMax := uint16(currentMTU - ipUDPHeaderSize)
-	assert.Equal(t, expectedMax, receivedUDPSize,
+	assert.Equal(t, expectedMax, uint16(receivedUDPSize.Load()),
 		"upstream should see capped EDNS0, not the client's 4096")
 }
 
@@ -874,7 +951,7 @@ func TestUpstreamResolver_NonRetryableEDEShortCircuits(t *testing.T) {
 	resolver := &upstreamResolverBase{
 		ctx:             ctx,
 		upstreamClient:  tracking,
-		upstreamServers: []netip.AddrPort{upstream1, upstream2},
+		upstreamServers: []upstreamRace{{upstream1, upstream2}},
 		upstreamTimeout: UpstreamTimeout,
 	}
 

client/internal/engine.go

@@ -512,16 +512,7 @@ func (e *Engine) Start(netbirdConfig *mgmProto.NetbirdConfig, mgmtURL *url.URL)
 
 	e.routeManager.SetRouteChangeListener(e.mobileDep.NetworkChangeListener)
 
-	e.dnsServer.SetRouteChecker(func(ip netip.Addr) bool {
-		for _, routes := range e.routeManager.GetSelectedClientRoutes() {
-			for _, r := range routes {
-				if r.Network.Contains(ip) {
-					return true
-				}
-			}
-		}
-		return false
-	})
+	e.dnsServer.SetRouteSources(e.routeManager.GetSelectedClientRoutes, e.routeManager.GetActiveClientRoutes)
 
 	if err = e.wgInterfaceCreate(); err != nil {
 		log.Errorf("failed creating tunnel interface %s: [%s]", e.config.WgIfaceName, err.Error())
@@ -1386,9 +1377,6 @@ func (e *Engine) updateNetworkMap(networkMap *mgmProto.NetworkMap) error {
 
 	e.networkSerial = serial
 
-	// Test received (upstream) servers for availability right away instead of upon usage.
-	// If no server of a server group responds this will disable the respective handler and retry later.
-	go e.dnsServer.ProbeAvailability()
 	return nil
 }
 
@@ -1932,7 +1920,7 @@ func (e *Engine) newDnsServer(dnsConfig *nbdns.Config) (dns.Server, error) {
 		return dnsServer, nil
 
 	case "ios":
-		dnsServer := dns.NewDefaultServerIos(e.ctx, e.wgInterface, e.mobileDep.DnsManager, e.mobileDep.HostDNSAddresses, e.statusRecorder, e.config.DisableDNS)
+		dnsServer := dns.NewDefaultServerIos(e.ctx, e.wgInterface, e.mobileDep.DnsManager, e.statusRecorder, e.config.DisableDNS)
 		return dnsServer, nil
 
 	default:

client/internal/routemanager/manager.go

@@ -53,6 +53,7 @@ type Manager interface {
 	GetRouteSelector() *routeselector.RouteSelector
 	GetClientRoutes() route.HAMap
 	GetSelectedClientRoutes() route.HAMap
+	GetActiveClientRoutes() route.HAMap
 	GetClientRoutesWithNetID() map[route.NetID][]*route.Route
 	SetRouteChangeListener(listener listener.NetworkChangeListener)
 	InitialRouteRange() []string
@@ -485,6 +486,39 @@ func (m *DefaultManager) GetSelectedClientRoutes() route.HAMap {
 	return m.routeSelector.FilterSelectedExitNodes(maps.Clone(m.clientRoutes))
 }
 
+// GetActiveClientRoutes returns the subset of selected client routes
+// that are currently reachable: the route's peer is Connected and is
+// the one actively carrying the route (not just an HA sibling).
+func (m *DefaultManager) GetActiveClientRoutes() route.HAMap {
+	m.mux.Lock()
+	selected := m.routeSelector.FilterSelectedExitNodes(maps.Clone(m.clientRoutes))
+	recorder := m.statusRecorder
+	m.mux.Unlock()
+
+	if recorder == nil {
+		return selected
+	}
+
+	out := make(route.HAMap, len(selected))
+	for id, routes := range selected {
+		for _, r := range routes {
+			st, err := recorder.GetPeer(r.Peer)
+			if err != nil {
+				continue
+			}
+			if st.ConnStatus != peer.StatusConnected {
+				continue
+			}
+			if _, hasRoute := st.GetRoutes()[r.Network.String()]; !hasRoute {
+				continue
+			}
+			out[id] = routes
+			break
+		}
+	}
+	return out
+}
+
 // GetClientRoutesWithNetID returns the current routes from the route map, but the keys consist of the network ID only
 func (m *DefaultManager) GetClientRoutesWithNetID() map[route.NetID][]*route.Route {
 	m.mux.Lock()
@@ -704,7 +738,10 @@ func (m *DefaultManager) collectExitNodeInfo(clientRoutes route.HAMap) exitNodeI
 }
 
 func (m *DefaultManager) isExitNodeRoute(routes []*route.Route) bool {
-	return len(routes) > 0 && routes[0].Network.String() == vars.ExitNodeCIDR
+	if len(routes) == 0 {
+		return false
+	}
+	return route.IsV4DefaultRoute(routes[0].Network) || route.IsV6DefaultRoute(routes[0].Network)
 }
 
 func (m *DefaultManager) categorizeUserSelection(netID route.NetID, info *exitNodeInfo) {

client/internal/routemanager/mock.go

@@ -19,6 +19,7 @@ type MockManager struct {
 	GetRouteSelectorFunc         func() *routeselector.RouteSelector
 	GetClientRoutesFunc          func() route.HAMap
 	GetSelectedClientRoutesFunc  func() route.HAMap
+	GetActiveClientRoutesFunc    func() route.HAMap
 	GetClientRoutesWithNetIDFunc func() map[route.NetID][]*route.Route
 	StopFunc                     func(manager *statemanager.Manager)
 }
@@ -78,6 +79,14 @@ func (m *MockManager) GetSelectedClientRoutes() route.HAMap {
 	return nil
 }
 
+// GetActiveClientRoutes mock implementation of GetActiveClientRoutes from the Manager interface
+func (m *MockManager) GetActiveClientRoutes() route.HAMap {
+	if m.GetActiveClientRoutesFunc != nil {
+		return m.GetActiveClientRoutesFunc()
+	}
+	return nil
+}
+
 // GetClientRoutesWithNetID mock implementation of GetClientRoutesWithNetID from Manager interface
 func (m *MockManager) GetClientRoutesWithNetID() map[route.NetID][]*route.Route {
 	if m.GetClientRoutesWithNetIDFunc != nil {

client/internal/routeselector/routeselector.go

@@ -4,6 +4,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"slices"
+	"strings"
 	"sync"
 
 	"github.com/hashicorp/go-multierror"
@@ -12,10 +13,6 @@ import (
 	"github.com/netbirdio/netbird/route"
 )
 
-const (
-	exitNodeCIDR = "0.0.0.0/0"
-)
-
 type RouteSelector struct {
 	mu               sync.RWMutex
 	deselectedRoutes map[route.NetID]struct{}
@@ -124,13 +121,7 @@ func (rs *RouteSelector) IsSelected(routeID route.NetID) bool {
 	rs.mu.RLock()
 	defer rs.mu.RUnlock()
 
-	if rs.deselectAll {
-		return false
-	}
-
-	_, deselected := rs.deselectedRoutes[routeID]
-	isSelected := !deselected
-	return isSelected
+	return rs.isSelectedLocked(routeID)
 }
 
 // FilterSelected removes unselected routes from the provided map.
@@ -144,23 +135,22 @@ func (rs *RouteSelector) FilterSelected(routes route.HAMap) route.HAMap {
 
 	filtered := route.HAMap{}
 	for id, rt := range routes {
-		netID := id.NetID()
-		_, deselected := rs.deselectedRoutes[netID]
-		if !deselected {
+		if !rs.isDeselectedLocked(id.NetID()) {
 			filtered[id] = rt
 		}
 	}
 	return filtered
 }
 
-// HasUserSelectionForRoute returns true if the user has explicitly selected or deselected this specific route
+// HasUserSelectionForRoute returns true if the user has explicitly selected or deselected this route.
+// Intended for exit-node code paths: a v6 exit-node pair (e.g. "MyExit-v6") with no explicit state of
+// its own inherits its v4 base's state, so legacy persisted selections that predate v6 pairing
+// transparently apply to the synthesized v6 entry.
 func (rs *RouteSelector) HasUserSelectionForRoute(routeID route.NetID) bool {
 	rs.mu.RLock()
 	defer rs.mu.RUnlock()
 
-	_, selected := rs.selectedRoutes[routeID]
-	_, deselected := rs.deselectedRoutes[routeID]
-	return selected || deselected
+	return rs.hasUserSelectionForRouteLocked(rs.effectiveNetID(routeID))
 }
 
 func (rs *RouteSelector) FilterSelectedExitNodes(routes route.HAMap) route.HAMap {
@@ -174,7 +164,7 @@ func (rs *RouteSelector) FilterSelectedExitNodes(routes route.HAMap) route.HAMap
 	filtered := make(route.HAMap, len(routes))
 	for id, rt := range routes {
 		netID := id.NetID()
-		if rs.isDeselected(netID) {
+		if rs.isDeselectedLocked(netID) {
 			continue
 		}
 
@@ -189,13 +179,48 @@ func (rs *RouteSelector) FilterSelectedExitNodes(routes route.HAMap) route.HAMap
 	return filtered
 }
 
-func (rs *RouteSelector) isDeselected(netID route.NetID) bool {
+// effectiveNetID returns the v4 base for a "-v6" exit pair entry that has no explicit
+// state of its own, so selections made on the v4 entry govern the v6 entry automatically.
+// Only call this from exit-node-specific code paths: applying it to a non-exit "-v6" route
+// would make it inherit unrelated v4 state. Must be called with rs.mu held.
+func (rs *RouteSelector) effectiveNetID(id route.NetID) route.NetID {
+	name := string(id)
+	if !strings.HasSuffix(name, route.V6ExitSuffix) {
+		return id
+	}
+	if _, ok := rs.selectedRoutes[id]; ok {
+		return id
+	}
+	if _, ok := rs.deselectedRoutes[id]; ok {
+		return id
+	}
+	return route.NetID(strings.TrimSuffix(name, route.V6ExitSuffix))
+}
+
+func (rs *RouteSelector) isSelectedLocked(routeID route.NetID) bool {
+	if rs.deselectAll {
+		return false
+	}
+	_, deselected := rs.deselectedRoutes[routeID]
+	return !deselected
+}
+
+func (rs *RouteSelector) isDeselectedLocked(netID route.NetID) bool {
+	if rs.deselectAll {
+		return true
+	}
 	_, deselected := rs.deselectedRoutes[netID]
-	return deselected || rs.deselectAll
+	return deselected
+}
+
+func (rs *RouteSelector) hasUserSelectionForRouteLocked(routeID route.NetID) bool {
+	_, selected := rs.selectedRoutes[routeID]
+	_, deselected := rs.deselectedRoutes[routeID]
+	return selected || deselected
 }
 
 func isExitNode(rt []*route.Route) bool {
-	return len(rt) > 0 && rt[0].Network.String() == exitNodeCIDR
+	return len(rt) > 0 && (route.IsV4DefaultRoute(rt[0].Network) || route.IsV6DefaultRoute(rt[0].Network))
 }
 
 func (rs *RouteSelector) applyExitNodeFilter(
@@ -204,26 +229,23 @@ func (rs *RouteSelector) applyExitNodeFilter(
 	rt []*route.Route,
 	out route.HAMap,
 ) {
-
-	if rs.hasUserSelections() {
-		// user made explicit selects/deselects
-		if rs.IsSelected(netID) {
+	// Exit-node path: apply the v4/v6 pair mirror so a deselect on the v4 base also
+	// drops the synthesized v6 entry that lacks its own explicit state.
+	effective := rs.effectiveNetID(netID)
+	if rs.hasUserSelectionForRouteLocked(effective) {
+		if rs.isSelectedLocked(effective) {
 			out[id] = rt
 		}
 		return
 	}
 
-	// no explicit selections: only include routes marked !SkipAutoApply (=AutoApply)
+	// no explicit selection for this route: defer to management's SkipAutoApply flag
 	sel := collectSelected(rt)
 	if len(sel) > 0 {
 		out[id] = sel
 	}
 }
 
-func (rs *RouteSelector) hasUserSelections() bool {
-	return len(rs.selectedRoutes) > 0 || len(rs.deselectedRoutes) > 0
-}
-
 func collectSelected(rt []*route.Route) []*route.Route {
 	var sel []*route.Route
 	for _, r := range rt {

client/internal/routeselector/routeselector_test.go

@@ -330,6 +330,137 @@ func TestRouteSelector_FilterSelectedExitNodes(t *testing.T) {
 	assert.Len(t, filtered, 0) // No routes should be selected
 }
 
+// TestRouteSelector_V6ExitPairInherits covers the v4/v6 exit-node pair selection
+// mirror. The mirror is scoped to exit-node code paths: HasUserSelectionForRoute
+// and FilterSelectedExitNodes resolve a "-v6" entry without explicit state to its
+// v4 base, so legacy persisted selections that predate v6 pairing transparently
+// apply to the synthesized v6 entry. General lookups (IsSelected, FilterSelected)
+// stay literal so unrelated routes named "*-v6" don't inherit unrelated state.
+func TestRouteSelector_V6ExitPairInherits(t *testing.T) {
+	all := []route.NetID{"exit1", "exit1-v6", "exit2", "exit2-v6", "corp", "corp-v6"}
+
+	t.Run("HasUserSelectionForRoute mirrors deselected v4 base", func(t *testing.T) {
+		rs := routeselector.NewRouteSelector()
+		require.NoError(t, rs.DeselectRoutes([]route.NetID{"exit1"}, all))
+
+		assert.True(t, rs.HasUserSelectionForRoute("exit1-v6"), "v6 pair sees v4 base's user selection")
+
+		// unrelated v6 with no v4 base touched is unaffected
+		assert.False(t, rs.HasUserSelectionForRoute("exit2-v6"))
+	})
+
+	t.Run("IsSelected stays literal for non-exit lookups", func(t *testing.T) {
+		rs := routeselector.NewRouteSelector()
+		require.NoError(t, rs.DeselectRoutes([]route.NetID{"corp"}, all))
+
+		// A non-exit route literally named "corp-v6" must not inherit "corp"'s state
+		// via the mirror; the mirror only applies in exit-node code paths.
+		assert.False(t, rs.IsSelected("corp"))
+		assert.True(t, rs.IsSelected("corp-v6"), "non-exit *-v6 routes must not inherit unrelated v4 state")
+	})
+
+	t.Run("explicit v6 state overrides v4 base in filter", func(t *testing.T) {
+		rs := routeselector.NewRouteSelector()
+		require.NoError(t, rs.DeselectRoutes([]route.NetID{"exit1"}, all))
+		require.NoError(t, rs.SelectRoutes([]route.NetID{"exit1-v6"}, true, all))
+
+		v4Route := &route.Route{NetID: "exit1", Network: netip.MustParsePrefix("0.0.0.0/0")}
+		v6Route := &route.Route{NetID: "exit1-v6", Network: netip.MustParsePrefix("::/0")}
+		routes := route.HAMap{
+			"exit1|0.0.0.0/0": {v4Route},
+			"exit1-v6|::/0":   {v6Route},
+		}
+
+		filtered := rs.FilterSelectedExitNodes(routes)
+		assert.NotContains(t, filtered, route.HAUniqueID("exit1|0.0.0.0/0"))
+		assert.Contains(t, filtered, route.HAUniqueID("exit1-v6|::/0"), "explicit v6 select wins over v4 base")
+	})
+
+	t.Run("non-v6-suffix routes unaffected", func(t *testing.T) {
+		rs := routeselector.NewRouteSelector()
+		require.NoError(t, rs.DeselectRoutes([]route.NetID{"exit1"}, all))
+
+		// A route literally named "exit1-something" must not pair-resolve.
+		assert.False(t, rs.HasUserSelectionForRoute("exit1-something"))
+	})
+
+	t.Run("filter v6 paired with deselected v4 base", func(t *testing.T) {
+		rs := routeselector.NewRouteSelector()
+		require.NoError(t, rs.DeselectRoutes([]route.NetID{"exit1"}, all))
+
+		v4Route := &route.Route{NetID: "exit1", Network: netip.MustParsePrefix("0.0.0.0/0")}
+		v6Route := &route.Route{NetID: "exit1-v6", Network: netip.MustParsePrefix("::/0")}
+		routes := route.HAMap{
+			"exit1|0.0.0.0/0": {v4Route},
+			"exit1-v6|::/0":   {v6Route},
+		}
+
+		filtered := rs.FilterSelectedExitNodes(routes)
+		assert.Empty(t, filtered, "deselecting v4 base must also drop the v6 pair")
+	})
+
+	t.Run("non-exit *-v6 routes pass through FilterSelectedExitNodes", func(t *testing.T) {
+		rs := routeselector.NewRouteSelector()
+		require.NoError(t, rs.DeselectRoutes([]route.NetID{"corp"}, all))
+
+		// A non-default-route entry named "corp-v6" is not an exit node and
+		// must not be skipped because its v4 base "corp" is deselected.
+		corpV6 := &route.Route{NetID: "corp-v6", Network: netip.MustParsePrefix("10.0.0.0/8")}
+		routes := route.HAMap{
+			"corp-v6|10.0.0.0/8": {corpV6},
+		}
+
+		filtered := rs.FilterSelectedExitNodes(routes)
+		assert.Contains(t, filtered, route.HAUniqueID("corp-v6|10.0.0.0/8"),
+			"non-exit *-v6 routes must not inherit unrelated v4 state in FilterSelectedExitNodes")
+	})
+}
+
+// TestRouteSelector_SkipAutoApplyPerRoute verifies that management's
+// SkipAutoApply flag governs each untouched route independently, even when
+// the user has explicit selections on other routes.
+func TestRouteSelector_SkipAutoApplyPerRoute(t *testing.T) {
+	autoApplied := &route.Route{
+		NetID:         "Auto",
+		Network:       netip.MustParsePrefix("0.0.0.0/0"),
+		SkipAutoApply: false,
+	}
+	skipApply := &route.Route{
+		NetID:         "Skip",
+		Network:       netip.MustParsePrefix("0.0.0.0/0"),
+		SkipAutoApply: true,
+	}
+	routes := route.HAMap{
+		"Auto|0.0.0.0/0": {autoApplied},
+		"Skip|0.0.0.0/0": {skipApply},
+	}
+
+	rs := routeselector.NewRouteSelector()
+	// User makes an unrelated explicit selection elsewhere.
+	require.NoError(t, rs.DeselectRoutes([]route.NetID{"Unrelated"}, []route.NetID{"Auto", "Skip", "Unrelated"}))
+
+	filtered := rs.FilterSelectedExitNodes(routes)
+	assert.Contains(t, filtered, route.HAUniqueID("Auto|0.0.0.0/0"), "AutoApply route should be included")
+	assert.NotContains(t, filtered, route.HAUniqueID("Skip|0.0.0.0/0"), "SkipAutoApply route should be excluded without explicit user selection")
+}
+
+// TestRouteSelector_V6ExitIsExitNode verifies that ::/0 routes are recognized
+// as exit nodes by the selector's filter path.
+func TestRouteSelector_V6ExitIsExitNode(t *testing.T) {
+	v6Exit := &route.Route{
+		NetID:         "V6Only",
+		Network:       netip.MustParsePrefix("::/0"),
+		SkipAutoApply: true,
+	}
+	routes := route.HAMap{
+		"V6Only|::/0": {v6Exit},
+	}
+
+	rs := routeselector.NewRouteSelector()
+	filtered := rs.FilterSelectedExitNodes(routes)
+	assert.Empty(t, filtered, "::/0 should be treated as an exit node and respect SkipAutoApply")
+}
+
 func TestRouteSelector_NewRoutesBehavior(t *testing.T) {
 	initialRoutes := []route.NetID{"route1", "route2", "route3"}
 	newRoutes := []route.NetID{"route1", "route2", "route3", "route4", "route5"}

client/ios/NetBirdSDK/client.go

@@ -162,11 +162,7 @@ func (c *Client) Run(fd int32, interfaceName string, envList *EnvList) error {
 	cfg.WgIface = interfaceName
 
 	c.connectClient = internal.NewConnectClient(ctx, cfg, c.recorder)
-	hostDNS := []netip.AddrPort{
-		netip.MustParseAddrPort("9.9.9.9:53"),
-		netip.MustParseAddrPort("149.112.112.112:53"),
-	}
-	return c.connectClient.RunOniOS(fd, c.networkChangeListener, c.dnsManager, hostDNS, c.stateFile)
+	return c.connectClient.RunOniOS(fd, c.networkChangeListener, c.dnsManager, c.stateFile)
 }
 
 // Stop the internal client and free the resources

client/netbird.wxs

@@ -64,6 +64,13 @@
 					<RegistryValue Name="InstalledByMSI" Type="integer" Value="1" KeyPath="yes" />
 				</RegistryKey>
 			</Component>
+			<!-- Drop the HKCU Run\Netbird value written by legacy NSIS installers. -->
+			<Component Id="NetbirdLegacyHKCUCleanup" Guid="*">
+				<RegistryValue Root="HKCU" Key="Software\NetBird GmbH\Installer"
+					Name="LegacyHKCUCleanup" Type="integer" Value="1" KeyPath="yes" />
+				<RemoveRegistryValue Root="HKCU"
+					Key="Software\Microsoft\Windows\CurrentVersion\Run" Name="Netbird" />
+			</Component>
 		</StandardDirectory>
 
 		<StandardDirectory Id="CommonAppDataFolder">
@@ -76,10 +83,28 @@
 			</Directory>
 		</StandardDirectory>
 
+		<!-- Drop Run, App Paths and Uninstall entries written by legacy NSIS
+		     installers into the 32-bit registry view (HKLM\Software\Wow6432Node). -->
+		<Component Id="NetbirdLegacyWow6432Cleanup" Directory="NetbirdInstallDir"
+			Guid="bda5d628-16bd-4086-b2c1-5099d8d51763" Bitness="always32">
+			<RegistryValue Root="HKLM" Key="Software\NetBird GmbH\Installer"
+				Name="LegacyWow6432Cleanup" Type="integer" Value="1" KeyPath="yes" />
+			<RemoveRegistryValue Root="HKLM"
+				Key="Software\Microsoft\Windows\CurrentVersion\Run" Name="Netbird" />
+			<RemoveRegistryKey Action="removeOnInstall" Root="HKLM"
+				Key="Software\Microsoft\Windows\CurrentVersion\App Paths\Netbird" />
+			<RemoveRegistryKey Action="removeOnInstall" Root="HKLM"
+				Key="Software\Microsoft\Windows\CurrentVersion\App Paths\Netbird-ui" />
+			<RemoveRegistryKey Action="removeOnInstall" Root="HKLM"
+				Key="Software\Microsoft\Windows\CurrentVersion\Uninstall\Netbird" />
+		</Component>
+
 		<ComponentGroup Id="NetbirdFilesComponent">
 			<ComponentRef Id="NetbirdFiles" />
 			<ComponentRef Id="NetbirdAumidRegistry" />
 			<ComponentRef Id="NetbirdAutoStart" />
+			<ComponentRef Id="NetbirdLegacyHKCUCleanup" />
+			<ComponentRef Id="NetbirdLegacyWow6432Cleanup" />
 		</ComponentGroup>
 
 		<util:CloseApplication Id="CloseNetBird" CloseMessage="no" Target="netbird.exe" RebootPrompt="no" />

client/ui/network.go

@@ -193,7 +193,15 @@ func getOverlappingNetworks(routes []*proto.Network) []*proto.Network {
 }
 
 func isDefaultRoute(routeRange string) bool {
-	return routeRange == "0.0.0.0/0" || routeRange == "::/0"
+	// routeRange is the merged display string from the daemon, e.g. "0.0.0.0/0",
+	// "::/0", or "0.0.0.0/0, ::/0" when a v4 exit node has a paired v6 entry.
+	for _, part := range strings.Split(routeRange, ",") {
+		switch strings.TrimSpace(part) {
+		case "0.0.0.0/0", "::/0":
+			return true
+		}
+	}
+	return false
 }
 
 func getExitNodeNetworks(routes []*proto.Network) []*proto.Network {

combined/cmd/root.go

@@ -67,10 +67,6 @@ func init() {
 	rootCmd.AddCommand(newTokenCommands())
 }
 
-func RootCmd() *cobra.Command {
-	return rootCmd
-}
-
 func Execute() error {
 	return rootCmd.Execute()
 }
@@ -172,7 +168,7 @@ func initializeConfig() error {
 // serverInstances holds all server instances created during startup.
 type serverInstances struct {
 	relaySrv      *relayServer.Server
-	mgmtSrv       mgmtServer.Server
+	mgmtSrv       *mgmtServer.BaseServer
 	signalSrv     *signalServer.Server
 	healthcheck   *healthcheck.Server
 	stunServer    *stun.Server
@@ -328,24 +324,19 @@ func setupServerHooks(servers *serverInstances, cfg *CombinedConfig) {
 		return
 	}
 
-	if s, ok := servers.mgmtSrv.GetContainer(mgmtServer.ContainerKeyBaseServer); ok {
-		if baseServer, ok := s.(*mgmtServer.BaseServer); ok {
-			baseServer.AfterInit(func(s *mgmtServer.BaseServer) {
-				grpcSrv := s.GRPCServer()
+	servers.mgmtSrv.AfterInit(func(s *mgmtServer.BaseServer) {
+		grpcSrv := s.GRPCServer()
 
-				if servers.signalSrv != nil {
-					proto.RegisterSignalExchangeServer(grpcSrv, servers.signalSrv)
-					log.Infof("Signal server registered on port %s", cfg.Server.ListenAddress)
-				}
-
-				s.SetHandlerFunc(createCombinedHandler(grpcSrv, s.APIHandler(), servers.relaySrv, servers.metricsServer.Meter, cfg))
-				if servers.relaySrv != nil {
-					log.Infof("Relay WebSocket handler added (path: /relay)")
-				}
-			})
+		if servers.signalSrv != nil {
+			proto.RegisterSignalExchangeServer(grpcSrv, servers.signalSrv)
+			log.Infof("Signal server registered on port %s", cfg.Server.ListenAddress)
 		}
-	}
 
+		s.SetHandlerFunc(createCombinedHandler(grpcSrv, s.APIHandler(), servers.relaySrv, servers.metricsServer.Meter, cfg))
+		if servers.relaySrv != nil {
+			log.Infof("Relay WebSocket handler added (path: /relay)")
+		}
+	})
 }
 
 func startServers(wg *sync.WaitGroup, srv *relayServer.Server, httpHealthcheck *healthcheck.Server, stunServer *stun.Server, metricsServer *sharedMetrics.Metrics) {
@@ -355,32 +346,38 @@ func startServers(wg *sync.WaitGroup, srv *relayServer.Server, httpHealthcheck *
 		log.Infof("Relay WebSocket multiplexed on management port (no separate relay listener)")
 	}
 
-	wg.Go(func() {
+	wg.Add(1)
+	go func() {
+		defer wg.Done()
 		log.Infof("running metrics server: %s%s", metricsServer.Addr, metricsServer.Endpoint)
 		if err := metricsServer.ListenAndServe(); !errors.Is(err, http.ErrServerClosed) {
 			log.Fatalf("failed to start metrics server: %v", err)
 		}
-	})
+	}()
 
-	wg.Go(func() {
+	wg.Add(1)
+	go func() {
+		defer wg.Done()
 		if err := httpHealthcheck.ListenAndServe(); !errors.Is(err, http.ErrServerClosed) {
 			log.Fatalf("failed to start healthcheck server: %v", err)
 		}
-	})
+	}()
 
 	if stunServer != nil {
-		wg.Go(func() {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
 			if err := stunServer.Listen(); err != nil {
 				if errors.Is(err, stun.ErrServerClosed) {
 					return
 				}
 				log.Errorf("STUN server error: %v", err)
 			}
-		})
+		}()
 	}
 }
 
-func shutdownServers(ctx context.Context, srv *relayServer.Server, httpHealthcheck *healthcheck.Server, stunServer *stun.Server, mgmtSrv mgmtServer.Server, metricsServer *sharedMetrics.Metrics) error {
+func shutdownServers(ctx context.Context, srv *relayServer.Server, httpHealthcheck *healthcheck.Server, stunServer *stun.Server, mgmtSrv *mgmtServer.BaseServer, metricsServer *sharedMetrics.Metrics) error {
 	var errs error
 
 	if err := httpHealthcheck.Shutdown(ctx); err != nil {
@@ -494,7 +491,7 @@ func handleTLSConfig(cfg *CombinedConfig) (*tls.Config, bool, error) {
 	return nil, false, nil
 }
 
-func createManagementServer(cfg *CombinedConfig, mgmtConfig *nbconfig.Config) (mgmtServer.Server, error) {
+func createManagementServer(cfg *CombinedConfig, mgmtConfig *nbconfig.Config) (*mgmtServer.BaseServer, error) {
 	mgmt := cfg.Management
 
 	// Extract port from listen address
@@ -505,7 +502,7 @@ func createManagementServer(cfg *CombinedConfig, mgmtConfig *nbconfig.Config) (m
 	}
 	mgmtPort, _ := strconv.Atoi(portStr)
 
-	mgmtSrv := newServer(
+	mgmtSrv := mgmtServer.NewServer(
 		&mgmtServer.Config{
 			NbConfig:                mgmtConfig,
 			DNSDomain:               "",

combined/cmd/server.go

@@ -1,13 +0,0 @@
-package cmd
-
-import (
-	mgmtServer "github.com/netbirdio/netbird/management/internals/server"
-)
-
-var newServer = func(cfg *mgmtServer.Config) mgmtServer.Server {
-	return mgmtServer.NewServer(cfg)
-}
-
-func SetNewServer(fn func(*mgmtServer.Config) mgmtServer.Server) {
-	newServer = fn
-}

management/internals/modules/peers/ephemeral/manager/ephemeral.go

@@ -2,6 +2,7 @@ package manager
 
 import (
 	"context"
+	"math/rand"
 	"sync"
 	"time"
 
@@ -11,240 +12,344 @@ import (
 	"github.com/netbirdio/netbird/management/internals/modules/peers/ephemeral"
 	"github.com/netbirdio/netbird/management/server/activity"
 	nbpeer "github.com/netbirdio/netbird/management/server/peer"
-
 	"github.com/netbirdio/netbird/management/server/store"
+	"github.com/netbirdio/netbird/management/server/telemetry"
 )
 
 const (
-	// cleanupWindow is the time window to wait after nearest peer deadline to start the cleanup procedure.
+	// cleanupWindow is the small grace period added on top of the
+	// staleness horizon before a sweep fires. It absorbs minor clock
+	// skew between the management server and the database and avoids
+	// firing a sweep right at the boundary where last_seen could still
+	// be one tick under the threshold.
 	cleanupWindow = 1 * time.Minute
+
+	// initialLoadMinDelay and initialLoadMaxDelay bracket the random
+	// delay applied before the post-restart catch-up query runs. Spread
+	// across replicas this prevents a thundering herd of catch-up
+	// queries hitting the database simultaneously after a deploy.
+	initialLoadMinDelay = 8 * time.Minute
+	initialLoadMaxDelay = 10 * time.Minute
 )
 
 var (
 	timeNow = time.Now
 )
 
-type ephemeralPeer struct {
-	id        string
-	accountID string
-	deadline  time.Time
-	next      *ephemeralPeer
+// accountEntry is the per-account state held by the cleanup tracker.
+// We don't track which peers are pending — the sweep query gets the
+// authoritative list straight from the database every time. We only
+// need to know the latest disconnect we've observed for this account
+// (so we can decide when it's safe to drop the entry) and the timer
+// that will fire the next sweep.
+type accountEntry struct {
+	lastDisconnectedAt time.Time
+	timer              *time.Timer
 }
 
-// todo: consider to remove peer from ephemeral list when the peer has been deleted via API. If we do not do it
-// in worst case we will get invalid error message in this manager.
-
-// EphemeralManager keep a list of ephemeral peers. After EphemeralLifeTime inactivity the peer will be deleted
-// automatically. Inactivity means the peer disconnected from the Management server.
+// EphemeralManager tracks accounts that may have ephemeral peers in
+// need of cleanup and runs a per-account sweep at the appropriate
+// time. State is in-memory and account-scoped: a sweep deletes any
+// ephemeral peer in the account that has been disconnected for at
+// least lifeTime, then either drops the account from the tracker
+// (when no recent disconnects have arrived) or re-arms the timer.
 type EphemeralManager struct {
 	store        store.Store
 	peersManager peers.Manager
 
-	headPeer  *ephemeralPeer
-	tailPeer  *ephemeralPeer
-	peersLock sync.Mutex
-	timer     *time.Timer
+	accountsLock sync.Mutex
+	accounts     map[string]*accountEntry
+
+	// initialLoadTimer is the one-shot timer used to defer the
+	// post-restart catch-up query; held so Stop() can cancel it.
+	initialLoadTimer *time.Timer
+	// stopped is flipped by Stop() so any timer that fires after
+	// teardown becomes a no-op instead of touching a half-dismantled
+	// store.
+	stopped bool
 
 	lifeTime      time.Duration
 	cleanupWindow time.Duration
+
+	// initialLoadDelay returns the wall-clock delay to wait before
+	// running the post-restart catch-up query. Pluggable so tests can
+	// fire the load immediately.
+	initialLoadDelay func() time.Duration
+
+	// bgCtx is the long-lived context captured at LoadInitialPeers
+	// time. Timer-driven sweeps use it because they fire long after
+	// the original gRPC handler ctx that produced an OnPeerDisconnected
+	// call has been cancelled.
+	bgCtx context.Context
+
+	// metrics is nil-safe; methods on telemetry.EphemeralPeersMetrics
+	// no-op when the receiver is nil so deployments without an app
+	// metrics provider work unchanged.
+	metrics *telemetry.EphemeralPeersMetrics
 }
 
 // NewEphemeralManager instantiate new EphemeralManager
 func NewEphemeralManager(store store.Store, peersManager peers.Manager) *EphemeralManager {
 	return &EphemeralManager{
-		store:        store,
-		peersManager: peersManager,
-
-		lifeTime:      ephemeral.EphemeralLifeTime,
-		cleanupWindow: cleanupWindow,
+		store:            store,
+		peersManager:     peersManager,
+		accounts:         make(map[string]*accountEntry),
+		lifeTime:         ephemeral.EphemeralLifeTime,
+		cleanupWindow:    cleanupWindow,
+		initialLoadDelay: defaultInitialLoadDelay,
 	}
 }
 
-// LoadInitialPeers load from the database the ephemeral type of peers and schedule a cleanup procedure to the head
-// of the linked list (to the most deprecated peer). At the end of cleanup it schedules the next cleanup to the new
-// head.
+// SetMetrics attaches a metrics collector. Pass nil to detach.
+func (e *EphemeralManager) SetMetrics(m *telemetry.EphemeralPeersMetrics) {
+	e.accountsLock.Lock()
+	e.metrics = m
+	e.accountsLock.Unlock()
+}
+
+// LoadInitialPeers schedules the post-restart catch-up query for a
+// random moment 8-10 minutes from now. Returns immediately. The
+// catch-up populates the per-account tracker from the database so any
+// peers that disconnected before the restart still get cleaned up.
+//
+// The random delay is critical: without it, every management replica
+// hitting the same Postgres instance after a deploy would issue the
+// catch-up query simultaneously.
 func (e *EphemeralManager) LoadInitialPeers(ctx context.Context) {
-	e.peersLock.Lock()
-	defer e.peersLock.Unlock()
-
-	e.loadEphemeralPeers(ctx)
-	if e.headPeer != nil {
-		e.timer = time.AfterFunc(e.lifeTime, func() {
-			e.cleanup(ctx)
-		})
-	}
-}
-
-// Stop timer
-func (e *EphemeralManager) Stop() {
-	e.peersLock.Lock()
-	defer e.peersLock.Unlock()
-
-	if e.timer != nil {
-		e.timer.Stop()
-	}
-}
-
-// OnPeerConnected remove the peer from the linked list of ephemeral peers. Because it has been called when the peer
-// is active the manager will not delete it while it is active.
-func (e *EphemeralManager) OnPeerConnected(ctx context.Context, peer *nbpeer.Peer) {
-	if !peer.Ephemeral {
+	e.accountsLock.Lock()
+	defer e.accountsLock.Unlock()
+	if e.stopped {
 		return
 	}
 
-	log.WithContext(ctx).Tracef("remove peer from ephemeral list: %s", peer.ID)
+	e.bgCtx = ctx
 
-	e.peersLock.Lock()
-	defer e.peersLock.Unlock()
-
-	e.removePeer(peer.ID)
-
-	// stop the unnecessary timer
-	if e.headPeer == nil && e.timer != nil {
-		e.timer.Stop()
-		e.timer = nil
-	}
+	delay := e.initialLoadDelay()
+	log.WithContext(ctx).Infof("ephemeral peer initial load scheduled in %s", delay)
+	e.initialLoadTimer = time.AfterFunc(delay, func() {
+		e.loadInitialAccounts(e.bgCtx)
+	})
 }
 
-// OnPeerDisconnected add the peer to the linked list of ephemeral peers. Because of the peer
-// is inactive it will be deleted after the EphemeralLifeTime period.
+// Stop cancels the deferred initial load and any per-account timers.
+func (e *EphemeralManager) Stop() {
+	e.accountsLock.Lock()
+	defer e.accountsLock.Unlock()
+
+	e.stopped = true
+	if e.initialLoadTimer != nil {
+		e.initialLoadTimer.Stop()
+		e.initialLoadTimer = nil
+	}
+	for _, entry := range e.accounts {
+		if entry.timer != nil {
+			entry.timer.Stop()
+		}
+	}
+	e.accounts = make(map[string]*accountEntry)
+}
+
+// OnPeerConnected is a no-op in the account-scoped design. The sweep
+// query filters out connected peers at the database level, so we don't
+// need an explicit "remove from list" signal when a peer reconnects.
+// Kept on the interface to preserve the existing call sites.
+func (e *EphemeralManager) OnPeerConnected(_ context.Context, _ *nbpeer.Peer) {
+}
+
+// OnPeerDisconnected registers a disconnect for the peer's account and
+// arms a sweep if one isn't already scheduled. Non-ephemeral peers are
+// ignored.
 func (e *EphemeralManager) OnPeerDisconnected(ctx context.Context, peer *nbpeer.Peer) {
 	if !peer.Ephemeral {
 		return
 	}
 
-	log.WithContext(ctx).Tracef("add peer to ephemeral list: %s", peer.ID)
-
-	e.peersLock.Lock()
-	defer e.peersLock.Unlock()
-
-	if e.isPeerOnList(peer.ID) {
-		return
-	}
-
-	e.addPeer(peer.AccountID, peer.ID, e.newDeadLine())
-	if e.timer == nil {
-		delay := e.headPeer.deadline.Sub(timeNow()) + e.cleanupWindow
-		if delay < 0 {
-			delay = 0
-		}
-		e.timer = time.AfterFunc(delay, func() {
-			e.cleanup(ctx)
-		})
-	}
-}
-
-func (e *EphemeralManager) loadEphemeralPeers(ctx context.Context) {
-	peers, err := e.store.GetAllEphemeralPeers(ctx, store.LockingStrengthNone)
-	if err != nil {
-		log.WithContext(ctx).Debugf("failed to load ephemeral peers: %s", err)
-		return
-	}
-
-	t := e.newDeadLine()
-	for _, p := range peers {
-		e.addPeer(p.AccountID, p.ID, t)
-	}
-
-	log.WithContext(ctx).Debugf("loaded ephemeral peer(s): %d", len(peers))
-}
-
-func (e *EphemeralManager) cleanup(ctx context.Context) {
-	log.Tracef("on ephemeral cleanup")
-	deletePeers := make(map[string]*ephemeralPeer)
-
-	e.peersLock.Lock()
 	now := timeNow()
-	for p := e.headPeer; p != nil; p = p.next {
-		if now.Before(p.deadline) {
-			break
-		}
 
-		deletePeers[p.id] = p
-		e.headPeer = p.next
-		if p.next == nil {
-			e.tailPeer = nil
-		}
+	e.accountsLock.Lock()
+	defer e.accountsLock.Unlock()
+	if e.stopped {
+		return
 	}
 
-	if e.headPeer != nil {
-		delay := e.headPeer.deadline.Sub(timeNow()) + e.cleanupWindow
-		if delay < 0 {
-			delay = 0
-		}
-		e.timer = time.AfterFunc(delay, func() {
-			e.cleanup(ctx)
+	entry, existed := e.accounts[peer.AccountID]
+	if !existed {
+		entry = &accountEntry{}
+		e.accounts[peer.AccountID] = entry
+		e.metrics.IncPending()
+	}
+	entry.lastDisconnectedAt = now
+
+	if entry.timer == nil {
+		delay := e.lifeTime + e.cleanupWindow
+		log.WithContext(ctx).Tracef("ephemeral: scheduling sweep for account %s in %s", peer.AccountID, delay)
+		accountID := peer.AccountID
+		entry.timer = time.AfterFunc(delay, func() {
+			e.sweep(e.bgCtxOrFallback(ctx), accountID)
 		})
-	} else {
-		e.timer = nil
-	}
-
-	e.peersLock.Unlock()
-
-	peerIDsPerAccount := make(map[string][]string)
-	for id, p := range deletePeers {
-		peerIDsPerAccount[p.accountID] = append(peerIDsPerAccount[p.accountID], id)
-	}
-
-	for accountID, peerIDs := range peerIDsPerAccount {
-		log.WithContext(ctx).Tracef("cleanup: deleting %d ephemeral peers for account %s", len(peerIDs), accountID)
-		err := e.peersManager.DeletePeers(ctx, accountID, peerIDs, activity.SystemInitiator, true)
-		if err != nil {
-			log.WithContext(ctx).Errorf("failed to delete ephemeral peers: %s", err)
-		}
 	}
 }
 
-func (e *EphemeralManager) addPeer(accountID string, peerID string, deadline time.Time) {
-	ep := &ephemeralPeer{
-		id:        peerID,
-		accountID: accountID,
-		deadline:  deadline,
+// bgCtxOrFallback returns the long-lived background context captured at
+// LoadInitialPeers time, falling back to the supplied ctx when the
+// manager hasn't been started through LoadInitialPeers (e.g. in tests
+// that drive the manager directly). Must be called with the lock held
+// or before the timer is armed.
+func (e *EphemeralManager) bgCtxOrFallback(ctx context.Context) context.Context {
+	if e.bgCtx != nil {
+		return e.bgCtx
 	}
-
-	if e.headPeer == nil {
-		e.headPeer = ep
-	}
-	if e.tailPeer != nil {
-		e.tailPeer.next = ep
-	}
-	e.tailPeer = ep
+	return ctx
 }
 
-func (e *EphemeralManager) removePeer(id string) {
-	if e.headPeer == nil {
+// loadInitialAccounts runs the post-restart catch-up query and seeds
+// the tracker with one entry per account that has at least one
+// disconnected ephemeral peer.
+func (e *EphemeralManager) loadInitialAccounts(ctx context.Context) {
+	accounts, err := e.store.GetEphemeralAccountsLastDisconnect(ctx)
+	if err != nil {
+		log.WithContext(ctx).Errorf("failed to load ephemeral accounts on startup: %v", err)
 		return
 	}
 
-	if e.headPeer.id == id {
-		e.headPeer = e.headPeer.next
-		if e.tailPeer.id == id {
-			e.tailPeer = nil
+	now := timeNow()
+	added := 0
+
+	e.accountsLock.Lock()
+	defer e.accountsLock.Unlock()
+	if e.stopped {
+		return
+	}
+
+	for accountID, lastDisc := range accounts {
+		// If we already learned about this account via an
+		// OnPeerDisconnected that arrived during the random delay
+		// window, prefer the live timestamp.
+		if _, alreadyTracked := e.accounts[accountID]; alreadyTracked {
+			continue
 		}
+
+		entry := &accountEntry{lastDisconnectedAt: lastDisc}
+		horizon := lastDisc.Add(e.lifeTime)
+
+		var delay time.Duration
+		if horizon.After(now) {
+			delay = horizon.Sub(now) + e.cleanupWindow
+		} else {
+			// Already past the staleness window — sweep right away
+			// (one cleanupWindow later, to keep startup load smooth
+			// when many accounts qualify at once).
+			delay = e.cleanupWindow
+		}
+		idForClosure := accountID
+		entry.timer = time.AfterFunc(delay, func() {
+			e.sweep(ctx, idForClosure)
+		})
+		e.accounts[accountID] = entry
+		added++
+	}
+
+	e.metrics.AddPending(int64(added))
+	log.WithContext(ctx).Debugf("ephemeral: loaded %d account(s) for cleanup tracking", added)
+}
+
+// sweep runs the cleanup pass for a single account. It queries the
+// database for disconnected ephemeral peers that have crossed the
+// staleness window, deletes them via peers.Manager, and then decides
+// whether to drop the account from the tracker or re-arm the timer.
+func (e *EphemeralManager) sweep(ctx context.Context, accountID string) {
+	now := timeNow()
+
+	e.accountsLock.Lock()
+	entry, ok := e.accounts[accountID]
+	if !ok || e.stopped {
+		e.accountsLock.Unlock()
+		return
+	}
+	lastDisc := entry.lastDisconnectedAt
+	entry.timer = nil
+	e.accountsLock.Unlock()
+
+	threshold := now.Add(-e.lifeTime)
+	stalePeerIDs, err := e.store.GetStaleEphemeralPeerIDsForAccount(ctx, accountID, threshold)
+	if err != nil {
+		log.WithContext(ctx).Errorf("ephemeral: failed to query stale peers for account %s: %v", accountID, err)
+		e.metrics.CountCleanupError()
+		e.rearm(ctx, accountID, e.cleanupWindow)
 		return
 	}
 
-	for p := e.headPeer; p.next != nil; p = p.next {
-		if p.next.id == id {
-			// if we remove the last element from the chain then set the last-1 as tail
-			if e.tailPeer.id == id {
-				e.tailPeer = p
-			}
-			p.next = p.next.next
+	if len(stalePeerIDs) > 0 {
+		log.WithContext(ctx).Tracef("ephemeral: deleting %d peer(s) for account %s", len(stalePeerIDs), accountID)
+		if err := e.peersManager.DeletePeers(ctx, accountID, stalePeerIDs, activity.SystemInitiator, true); err != nil {
+			log.WithContext(ctx).Errorf("ephemeral: failed to delete peers for account %s: %v", accountID, err)
+			e.metrics.CountCleanupError()
+			e.rearm(ctx, accountID, e.cleanupWindow)
 			return
 		}
+		e.metrics.CountCleanupRun()
+		e.metrics.CountPeersCleaned(int64(len(stalePeerIDs)))
 	}
+
+	e.accountsLock.Lock()
+	defer e.accountsLock.Unlock()
+	if e.stopped {
+		return
+	}
+	entry, ok = e.accounts[accountID]
+	if !ok {
+		return
+	}
+
+	// Drop rule: if every disconnect we've observed has now crossed
+	// the staleness window, the sweep we just ran saw everything that
+	// could possibly need cleaning. Dropping is safe — a future
+	// disconnect will recreate the entry. The check uses the latest
+	// lastDisc, which may have advanced (concurrently with the sweep
+	// itself) due to a new OnPeerDisconnected, in which case we
+	// correctly re-arm.
+	horizon := entry.lastDisconnectedAt.Add(e.lifeTime)
+	if !horizon.After(now) {
+		delete(e.accounts, accountID)
+		e.metrics.DecPending(1)
+		log.WithContext(ctx).Tracef("ephemeral: dropping account %s (lastDisc=%s, horizon=%s, now=%s)",
+			accountID, lastDisc, horizon, now)
+		return
+	}
+
+	delay := horizon.Sub(now) + e.cleanupWindow
+	idForClosure := accountID
+	entry.timer = time.AfterFunc(delay, func() {
+		e.sweep(ctx, idForClosure)
+	})
 }
 
-func (e *EphemeralManager) isPeerOnList(id string) bool {
-	for p := e.headPeer; p != nil; p = p.next {
-		if p.id == id {
-			return true
-		}
+// rearm reschedules a sweep `delay` from now. Used after a recoverable
+// error in the sweep path so the account doesn't get stuck.
+func (e *EphemeralManager) rearm(ctx context.Context, accountID string, delay time.Duration) {
+	e.accountsLock.Lock()
+	defer e.accountsLock.Unlock()
+	if e.stopped {
+		return
 	}
-	return false
+	entry, ok := e.accounts[accountID]
+	if !ok {
+		return
+	}
+	idForClosure := accountID
+	entry.timer = time.AfterFunc(delay, func() {
+		e.sweep(ctx, idForClosure)
+	})
 }
 
-func (e *EphemeralManager) newDeadLine() time.Time {
-	return timeNow().Add(e.lifeTime)
+// defaultInitialLoadDelay returns a random duration in
+// [initialLoadMinDelay, initialLoadMaxDelay). Process-wide
+// math/rand is acceptable here — the delay is purely a smoothing
+// jitter, not a security primitive.
+func defaultInitialLoadDelay() time.Duration {
+	span := int64(initialLoadMaxDelay - initialLoadMinDelay)
+	if span <= 0 {
+		return initialLoadMinDelay
+	}
+	return initialLoadMinDelay + time.Duration(rand.Int63n(span))
 }

management/internals/modules/peers/ephemeral/manager/ephemeral_test.go

@@ -2,299 +2,544 @@ package manager
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"sync"
+	"sync/atomic"
 	"testing"
 	"time"
 
 	"github.com/golang/mock/gomock"
 	log "github.com/sirupsen/logrus"
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 
 	nbdns "github.com/netbirdio/netbird/dns"
 	"github.com/netbirdio/netbird/management/internals/modules/peers"
 	"github.com/netbirdio/netbird/management/internals/modules/peers/ephemeral"
-	nbAccount "github.com/netbirdio/netbird/management/server/account"
 	nbpeer "github.com/netbirdio/netbird/management/server/peer"
 	"github.com/netbirdio/netbird/management/server/store"
 	"github.com/netbirdio/netbird/management/server/types"
 	"github.com/netbirdio/netbird/route"
 )
 
+// MockStore is a thin in-memory stand-in that implements only the two
+// methods the EphemeralManager uses. It honors the account / ephemeral
+// / connected / lastSeen attributes of each peer so the cleanup logic
+// can be exercised end-to-end without bringing up sqlite or Postgres.
 type MockStore struct {
 	store.Store
+	mu      sync.Mutex
 	account *types.Account
 }
 
-func (s *MockStore) GetAllEphemeralPeers(_ context.Context, _ store.LockingStrength) ([]*nbpeer.Peer, error) {
-	var peers []*nbpeer.Peer
-	for _, v := range s.account.Peers {
-		if v.Ephemeral {
-			peers = append(peers, v)
+func (s *MockStore) GetStaleEphemeralPeerIDsForAccount(_ context.Context, accountID string, olderThan time.Time) ([]string, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.account == nil || s.account.Id != accountID {
+		return nil, nil
+	}
+	var ids []string
+	for _, p := range s.account.Peers {
+		if !p.Ephemeral {
+			continue
+		}
+		if p.Status == nil || p.Status.Connected {
+			continue
+		}
+		if p.Status.LastSeen.Before(olderThan) {
+			ids = append(ids, p.ID)
 		}
 	}
-	return peers, nil
+	return ids, nil
 }
 
-type MockAccountManager struct {
-	mu sync.Mutex
-	nbAccount.Manager
-	store             *MockStore
-	deletePeerCalls   int
-	bufferUpdateCalls map[string]int
-	wg                *sync.WaitGroup
-}
-
-func (a *MockAccountManager) DeletePeer(_ context.Context, accountID, peerID, userID string) error {
-	a.mu.Lock()
-	defer a.mu.Unlock()
-	a.deletePeerCalls++
-	delete(a.store.account.Peers, peerID)
-	if a.wg != nil {
-		a.wg.Done()
+func (s *MockStore) GetEphemeralAccountsLastDisconnect(_ context.Context) (map[string]time.Time, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	out := map[string]time.Time{}
+	if s.account == nil {
+		return out, nil
 	}
-	return nil
-}
-
-func (a *MockAccountManager) GetDeletePeerCalls() int {
-	a.mu.Lock()
-	defer a.mu.Unlock()
-	return a.deletePeerCalls
-}
-
-func (a *MockAccountManager) BufferUpdateAccountPeers(ctx context.Context, accountID string, reason types.UpdateReason) {
-	a.mu.Lock()
-	defer a.mu.Unlock()
-	if a.bufferUpdateCalls == nil {
-		a.bufferUpdateCalls = make(map[string]int)
+	var latest time.Time
+	hasAny := false
+	for _, p := range s.account.Peers {
+		if !p.Ephemeral || p.Status == nil || p.Status.Connected {
+			continue
+		}
+		if !hasAny || p.Status.LastSeen.After(latest) {
+			latest = p.Status.LastSeen
+			hasAny = true
+		}
 	}
-	a.bufferUpdateCalls[accountID]++
-}
-
-func (a *MockAccountManager) GetBufferUpdateCalls(accountID string) int {
-	a.mu.Lock()
-	defer a.mu.Unlock()
-	if a.bufferUpdateCalls == nil {
-		return 0
+	if hasAny {
+		out[s.account.Id] = latest
 	}
-	return a.bufferUpdateCalls[accountID]
+	return out, nil
 }
 
-func (a *MockAccountManager) GetStore() store.Store {
-	return a.store
-}
-
-func TestNewManager(t *testing.T) {
-	t.Cleanup(func() {
-		timeNow = time.Now
-	})
-	startTime := time.Now()
+// withFakeClock pins timeNow to a settable value for the duration of t.
+// Returns a getter and a setter so subtests can advance virtual time.
+func withFakeClock(t *testing.T, start time.Time) (get func() time.Time, set func(time.Time)) {
+	t.Helper()
+	var mu sync.Mutex
+	now := start
 	timeNow = func() time.Time {
-		return startTime
+		mu.Lock()
+		defer mu.Unlock()
+		return now
 	}
+	t.Cleanup(func() { timeNow = time.Now })
 
-	store := &MockStore{}
-	ctrl := gomock.NewController(t)
-	peersManager := peers.NewMockManager(ctrl)
-
-	numberOfPeers := 5
-	numberOfEphemeralPeers := 3
-	seedPeers(store, numberOfPeers, numberOfEphemeralPeers)
-
-	// Expect DeletePeers to be called for ephemeral peers
-	peersManager.EXPECT().
-		DeletePeers(gomock.Any(), gomock.Any(), gomock.Any(), gomock.Any(), true).
-		DoAndReturn(func(ctx context.Context, accountID string, peerIDs []string, userID string, checkConnected bool) error {
-			for _, peerID := range peerIDs {
-				delete(store.account.Peers, peerID)
-			}
-			return nil
-		}).
-		AnyTimes()
-
-	mgr := NewEphemeralManager(store, peersManager)
-	mgr.loadEphemeralPeers(context.Background())
-	startTime = startTime.Add(ephemeral.EphemeralLifeTime + 1)
-	mgr.cleanup(context.Background())
-
-	if len(store.account.Peers) != numberOfPeers {
-		t.Errorf("failed to cleanup ephemeral peers, expected: %d, result: %d", numberOfPeers, len(store.account.Peers))
-	}
+	return func() time.Time {
+			mu.Lock()
+			defer mu.Unlock()
+			return now
+		}, func(v time.Time) {
+			mu.Lock()
+			defer mu.Unlock()
+			now = v
+		}
 }
 
-func TestNewManagerPeerConnected(t *testing.T) {
-	t.Cleanup(func() {
-		timeNow = time.Now
-	})
-	startTime := time.Now()
-	timeNow = func() time.Time {
-		return startTime
-	}
-
-	store := &MockStore{}
-	ctrl := gomock.NewController(t)
-	peersManager := peers.NewMockManager(ctrl)
-
-	numberOfPeers := 5
-	numberOfEphemeralPeers := 3
-	seedPeers(store, numberOfPeers, numberOfEphemeralPeers)
-
-	// Expect DeletePeers to be called for ephemeral peers (except the connected one)
-	peersManager.EXPECT().
-		DeletePeers(gomock.Any(), gomock.Any(), gomock.Any(), gomock.Any(), true).
-		DoAndReturn(func(ctx context.Context, accountID string, peerIDs []string, userID string, checkConnected bool) error {
-			for _, peerID := range peerIDs {
-				delete(store.account.Peers, peerID)
-			}
-			return nil
-		}).
-		AnyTimes()
-
-	mgr := NewEphemeralManager(store, peersManager)
-	mgr.loadEphemeralPeers(context.Background())
-	mgr.OnPeerConnected(context.Background(), store.account.Peers["ephemeral_peer_0"])
-
-	startTime = startTime.Add(ephemeral.EphemeralLifeTime + 1)
-	mgr.cleanup(context.Background())
-
-	expected := numberOfPeers + 1
-	if len(store.account.Peers) != expected {
-		t.Errorf("failed to cleanup ephemeral peers, expected: %d, result: %d", expected, len(store.account.Peers))
-	}
+// newManagerForTest builds a manager with short timers and no random
+// initial-load delay so tests run instantly.
+func newManagerForTest(t *testing.T, st store.Store, peersMgr peers.Manager) *EphemeralManager {
+	t.Helper()
+	mgr := NewEphemeralManager(st, peersMgr)
+	mgr.lifeTime = 100 * time.Millisecond
+	mgr.cleanupWindow = 10 * time.Millisecond
+	mgr.initialLoadDelay = func() time.Duration { return 0 }
+	t.Cleanup(mgr.Stop)
+	return mgr
 }
 
-func TestNewManagerPeerDisconnected(t *testing.T) {
-	t.Cleanup(func() {
-		timeNow = time.Now
-	})
-	startTime := time.Now()
-	timeNow = func() time.Time {
-		return startTime
-	}
+// TestOnPeerDisconnected_RegistersAndSweeps drives the OnPeerDisconnected
+// path with a fake clock: a single ephemeral peer disconnects, we
+// advance past the staleness window, and the sweep deletes it.
+func TestOnPeerDisconnected_RegistersAndSweeps(t *testing.T) {
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
 
-	store := &MockStore{}
-	ctrl := gomock.NewController(t)
-	peersManager := peers.NewMockManager(ctrl)
-
-	numberOfPeers := 5
-	numberOfEphemeralPeers := 3
-	seedPeers(store, numberOfPeers, numberOfEphemeralPeers)
-
-	// Expect DeletePeers to be called for the one disconnected peer
-	peersManager.EXPECT().
-		DeletePeers(gomock.Any(), gomock.Any(), gomock.Any(), gomock.Any(), true).
-		DoAndReturn(func(ctx context.Context, accountID string, peerIDs []string, userID string, checkConnected bool) error {
-			for _, peerID := range peerIDs {
-				delete(store.account.Peers, peerID)
-			}
-			return nil
-		}).
-		AnyTimes()
-
-	mgr := NewEphemeralManager(store, peersManager)
-	mgr.loadEphemeralPeers(context.Background())
-	for _, v := range store.account.Peers {
-		mgr.OnPeerConnected(context.Background(), v)
-
-	}
-	mgr.OnPeerDisconnected(context.Background(), store.account.Peers["ephemeral_peer_0"])
-
-	startTime = startTime.Add(ephemeral.EphemeralLifeTime + 1)
-	mgr.cleanup(context.Background())
-
-	expected := numberOfPeers + numberOfEphemeralPeers - 1
-	if len(store.account.Peers) != expected {
-		t.Errorf("failed to cleanup ephemeral peers, expected: %d, result: %d", expected, len(store.account.Peers))
-	}
-}
-
-func TestCleanupSchedulingBehaviorIsBatched(t *testing.T) {
-	const (
-		ephemeralPeers    = 10
-		testLifeTime      = 1 * time.Second
-		testCleanupWindow = 100 * time.Millisecond
-	)
-
-	t.Cleanup(func() {
-		timeNow = time.Now
-	})
-	startTime := time.Now()
-	timeNow = func() time.Time {
-		return startTime
-	}
-
-	mockStore := &MockStore{}
-	account := newAccountWithId(context.Background(), "account", "", "", false)
-	mockStore.account = account
-
-	wg := &sync.WaitGroup{}
-	wg.Add(ephemeralPeers)
-	mockAM := &MockAccountManager{
-		store: mockStore,
-		wg:    wg,
-	}
+	getNow, setNow := withFakeClock(t, time.Now())
 
 	ctrl := gomock.NewController(t)
-	peersManager := peers.NewMockManager(ctrl)
+	peersMgr := peers.NewMockManager(ctrl)
 
-	// Set up expectation that DeletePeers will be called once with all peer IDs
-	peersManager.EXPECT().
-		DeletePeers(gomock.Any(), account.Id, gomock.Any(), gomock.Any(), true).
-		DoAndReturn(func(ctx context.Context, accountID string, peerIDs []string, userID string, checkConnected bool) error {
-			// Simulate the actual deletion behavior
-			for _, peerID := range peerIDs {
-				err := mockAM.DeletePeer(ctx, accountID, peerID, userID)
-				if err != nil {
-					return err
-				}
+	var deletedMu sync.Mutex
+	var deleted []string
+	var deleteCalls atomic.Int32
+	peersMgr.EXPECT().
+		DeletePeers(gomock.Any(), "acc-1", gomock.Any(), gomock.Any(), true).
+		DoAndReturn(func(_ context.Context, accountID string, peerIDs []string, _ string, _ bool) error {
+			deleteCalls.Add(1)
+			mockStore.mu.Lock()
+			for _, id := range peerIDs {
+				delete(mockStore.account.Peers, id)
 			}
-			mockAM.BufferUpdateAccountPeers(ctx, accountID, types.UpdateReason{})
+			mockStore.mu.Unlock()
+			deletedMu.Lock()
+			deleted = append(deleted, peerIDs...)
+			deletedMu.Unlock()
 			return nil
-		}).
-		Times(1)
+		}).AnyTimes()
 
-	mgr := NewEphemeralManager(mockStore, peersManager)
-	mgr.lifeTime = testLifeTime
-	mgr.cleanupWindow = testCleanupWindow
+	mgr := newManagerForTest(t, mockStore, peersMgr)
 
-	// Add peers and disconnect them at slightly different times (within cleanup window)
-	for i := range ephemeralPeers {
-		p := &nbpeer.Peer{ID: fmt.Sprintf("peer-%d", i), AccountID: account.Id, Ephemeral: true}
-		mockStore.account.Peers[p.ID] = p
+	// One ephemeral peer that disconnected "now".
+	now := getNow()
+	p := &nbpeer.Peer{
+		ID:        "p1",
+		AccountID: "acc-1",
+		Ephemeral: true,
+		Status:    &nbpeer.PeerStatus{Connected: false, LastSeen: now},
+	}
+	mockStore.account.Peers[p.ID] = p
+	mgr.OnPeerDisconnected(context.Background(), p)
+
+	// Advance past lifeTime + cleanupWindow so the timer-driven sweep fires.
+	setNow(now.Add(mgr.lifeTime + 5*mgr.cleanupWindow))
+	require.Eventually(t, func() bool { return deleteCalls.Load() >= 1 }, 2*time.Second, 5*time.Millisecond,
+		"sweep should fire and delete the stale peer")
+
+	deletedMu.Lock()
+	deletedCopy := append([]string(nil), deleted...)
+	deletedMu.Unlock()
+	require.Equal(t, []string{"p1"}, deletedCopy, "only the one ephemeral peer should be deleted")
+}
+
+// TestOnPeerDisconnected_NonEphemeralIgnored: a non-ephemeral disconnect
+// must not register the account or arm any timer.
+func TestOnPeerDisconnected_NonEphemeralIgnored(t *testing.T) {
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
+	withFakeClock(t, time.Now())
+
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+	// No DeletePeers expectation — must not be called.
+
+	mgr := newManagerForTest(t, mockStore, peersMgr)
+	mgr.OnPeerDisconnected(context.Background(), &nbpeer.Peer{
+		ID:        "p1",
+		AccountID: "acc-1",
+		Ephemeral: false,
+		Status:    &nbpeer.PeerStatus{Connected: false, LastSeen: timeNow()},
+	})
+
+	mgr.accountsLock.Lock()
+	require.Empty(t, mgr.accounts, "non-ephemeral disconnect must not register an account")
+	mgr.accountsLock.Unlock()
+}
+
+// TestSweep_DropsAccountWhenIdle: after a sweep cleans the stale peers,
+// if no more disconnects have arrived the account must be dropped from
+// the in-memory tracker.
+func TestSweep_DropsAccountWhenIdle(t *testing.T) {
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
+	getNow, setNow := withFakeClock(t, time.Now())
+
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+	peersMgr.EXPECT().
+		DeletePeers(gomock.Any(), "acc-1", gomock.Any(), gomock.Any(), true).
+		DoAndReturn(func(_ context.Context, _ string, peerIDs []string, _ string, _ bool) error {
+			mockStore.mu.Lock()
+			for _, id := range peerIDs {
+				delete(mockStore.account.Peers, id)
+			}
+			mockStore.mu.Unlock()
+			return nil
+		}).AnyTimes()
+
+	mgr := newManagerForTest(t, mockStore, peersMgr)
+
+	now := getNow()
+	p := &nbpeer.Peer{ID: "p1", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: now}}
+	mockStore.account.Peers[p.ID] = p
+	mgr.OnPeerDisconnected(context.Background(), p)
+
+	setNow(now.Add(mgr.lifeTime + 5*mgr.cleanupWindow))
+
+	require.Eventually(t, func() bool {
+		mgr.accountsLock.Lock()
+		defer mgr.accountsLock.Unlock()
+		return len(mgr.accounts) == 0
+	}, 2*time.Second, 5*time.Millisecond, "account should be dropped after sweep with no new disconnects")
+}
+
+// TestSweep_ReArmsWhenNewDisconnectArrived: simulate the race where a
+// fresh disconnect arrives just before the sweep fires. The sweep must
+// observe the updated lastDisc and re-arm rather than drop.
+func TestSweep_ReArmsWhenNewDisconnectArrived(t *testing.T) {
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
+	getNow, setNow := withFakeClock(t, time.Now())
+
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+	peersMgr.EXPECT().
+		DeletePeers(gomock.Any(), "acc-1", gomock.Any(), gomock.Any(), true).
+		DoAndReturn(func(_ context.Context, _ string, peerIDs []string, _ string, _ bool) error {
+			mockStore.mu.Lock()
+			for _, id := range peerIDs {
+				delete(mockStore.account.Peers, id)
+			}
+			mockStore.mu.Unlock()
+			return nil
+		}).AnyTimes()
+
+	mgr := newManagerForTest(t, mockStore, peersMgr)
+
+	now := getNow()
+	p1 := &nbpeer.Peer{ID: "p1", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: now}}
+	mockStore.account.Peers[p1.ID] = p1
+	mgr.OnPeerDisconnected(context.Background(), p1)
+
+	// Advance most of the way toward the first sweep, then introduce
+	// a fresh disconnect that resets lastDisc.
+	setNow(now.Add(mgr.lifeTime - 10*time.Millisecond))
+	p2 := &nbpeer.Peer{ID: "p2", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: getNow()}}
+	mockStore.account.Peers[p2.ID] = p2
+	mgr.OnPeerDisconnected(context.Background(), p2)
+
+	// Push past p1's staleness so the first sweep runs and cleans p1
+	// but observes p2 already on the account entry. It must re-arm.
+	setNow(now.Add(mgr.lifeTime + 5*mgr.cleanupWindow))
+
+	require.Eventually(t, func() bool {
+		mockStore.mu.Lock()
+		defer mockStore.mu.Unlock()
+		_, gone := mockStore.account.Peers["p1"]
+		return !gone
+	}, 2*time.Second, 5*time.Millisecond, "p1 should be cleaned at the first sweep")
+
+	// The account should still be tracked because p2 is younger than lifeTime
+	// from the sweep's vantage point at this moment.
+	mgr.accountsLock.Lock()
+	_, stillTracked := mgr.accounts["acc-1"]
+	mgr.accountsLock.Unlock()
+	require.True(t, stillTracked, "account should remain tracked because p2's disconnect kept it active")
+
+	// Push past p2's staleness; second sweep cleans p2 and drops the account.
+	setNow(getNow().Add(mgr.lifeTime + 5*mgr.cleanupWindow))
+	require.Eventually(t, func() bool {
+		mgr.accountsLock.Lock()
+		defer mgr.accountsLock.Unlock()
+		return len(mgr.accounts) == 0
+	}, 2*time.Second, 5*time.Millisecond, "account should drop after the final sweep")
+}
+
+// TestSweep_BatchesPeersPerAccount: many ephemeral peers disconnect on
+// the same account; a single sweep must delete them all in one
+// DeletePeers call.
+func TestSweep_BatchesPeersPerAccount(t *testing.T) {
+	const ephemeralPeers = 8
+
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
+	getNow, setNow := withFakeClock(t, time.Now())
+
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+
+	deleteBatches := make(chan []string, 4)
+	peersMgr.EXPECT().
+		DeletePeers(gomock.Any(), "acc-1", gomock.Any(), gomock.Any(), true).
+		DoAndReturn(func(_ context.Context, _ string, peerIDs []string, _ string, _ bool) error {
+			cp := append([]string(nil), peerIDs...)
+			mockStore.mu.Lock()
+			for _, id := range peerIDs {
+				delete(mockStore.account.Peers, id)
+			}
+			mockStore.mu.Unlock()
+			deleteBatches <- cp
+			return nil
+		}).Times(1)
+
+	mgr := newManagerForTest(t, mockStore, peersMgr)
+
+	now := getNow()
+	for i := 0; i < ephemeralPeers; i++ {
+		id := fmt.Sprintf("p-%d", i)
+		// Stagger by a fraction of cleanupWindow so they all fall on
+		// the same sweep tick.
+		when := now.Add(time.Duration(i) * time.Millisecond)
+		p := &nbpeer.Peer{ID: id, AccountID: "acc-1", Ephemeral: true,
+			Status: &nbpeer.PeerStatus{Connected: false, LastSeen: when}}
+		mockStore.account.Peers[id] = p
 		mgr.OnPeerDisconnected(context.Background(), p)
-		startTime = startTime.Add(testCleanupWindow / (ephemeralPeers * 2))
 	}
 
-	// Advance time past the lifetime to trigger cleanup
-	startTime = startTime.Add(testLifeTime + testCleanupWindow)
+	setNow(now.Add(mgr.lifeTime + 5*mgr.cleanupWindow))
 
-	// Wait for all deletions to complete
-	wg.Wait()
-
-	assert.Len(t, mockStore.account.Peers, 0, "all ephemeral peers should be cleaned up after the lifetime")
-	assert.Equal(t, 1, mockAM.GetBufferUpdateCalls(account.Id), "buffer update should be called once")
-	assert.Equal(t, ephemeralPeers, mockAM.GetDeletePeerCalls(), "should have deleted all peers")
+	select {
+	case batch := <-deleteBatches:
+		require.Len(t, batch, ephemeralPeers, "all peers should be deleted in a single batch")
+	case <-time.After(2 * time.Second):
+		t.Fatal("expected one batched DeletePeers call")
+	}
 }
 
-func seedPeers(store *MockStore, numberOfPeers int, numberOfEphemeralPeers int) {
-	store.account = newAccountWithId(context.Background(), "my account", "", "", false)
+// TestLoadInitialAccounts_SeedsFromStore exercises the post-restart
+// catch-up path: pre-populate the store, point the manager at it, and
+// confirm both already-stale and not-yet-stale peers get cleaned at
+// their proper times.
+func TestLoadInitialAccounts_SeedsFromStore(t *testing.T) {
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
+	getNow, setNow := withFakeClock(t, time.Now())
 
-	for i := 0; i < numberOfPeers; i++ {
-		peerId := fmt.Sprintf("peer_%d", i)
-		p := &nbpeer.Peer{
-			ID:        peerId,
-			Ephemeral: false,
-		}
-		store.account.Peers[p.ID] = p
+	now := getNow()
+	// p-stale: already past the staleness window when load runs.
+	mockStore.account.Peers["p-stale"] = &nbpeer.Peer{
+		ID: "p-stale", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: now.Add(-time.Hour)},
+	}
+	// p-fresh: disconnected but not yet stale.
+	mockStore.account.Peers["p-fresh"] = &nbpeer.Peer{
+		ID: "p-fresh", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: now},
 	}
 
-	for i := 0; i < numberOfEphemeralPeers; i++ {
-		peerId := fmt.Sprintf("ephemeral_peer_%d", i)
-		p := &nbpeer.Peer{
-			ID:        peerId,
-			Ephemeral: true,
-		}
-		store.account.Peers[p.ID] = p
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+	peersMgr.EXPECT().
+		DeletePeers(gomock.Any(), "acc-1", gomock.Any(), gomock.Any(), true).
+		DoAndReturn(func(_ context.Context, _ string, peerIDs []string, _ string, _ bool) error {
+			mockStore.mu.Lock()
+			for _, id := range peerIDs {
+				delete(mockStore.account.Peers, id)
+			}
+			mockStore.mu.Unlock()
+			return nil
+		}).AnyTimes()
+
+	mgr := newManagerForTest(t, mockStore, peersMgr)
+	// Drive loadInitialAccounts directly with the fake-clock-aware now.
+	mgr.loadInitialAccounts(context.Background())
+
+	// First sweep should fire shortly (cleanupWindow) for the stale peer.
+	setNow(now.Add(5 * mgr.cleanupWindow))
+	require.Eventually(t, func() bool {
+		mockStore.mu.Lock()
+		defer mockStore.mu.Unlock()
+		_, gone := mockStore.account.Peers["p-stale"]
+		return !gone
+	}, 2*time.Second, 5*time.Millisecond, "p-stale should be deleted on the first sweep")
+
+	// p-fresh is not yet stale; advance past its window.
+	setNow(now.Add(mgr.lifeTime + 5*mgr.cleanupWindow))
+	require.Eventually(t, func() bool {
+		mockStore.mu.Lock()
+		defer mockStore.mu.Unlock()
+		_, gone := mockStore.account.Peers["p-fresh"]
+		return !gone
+	}, 2*time.Second, 5*time.Millisecond, "p-fresh should be deleted once it crosses the staleness window")
+}
+
+// TestStop_CancelsPendingWork verifies that Stop() cancels both the
+// deferred initial load and per-account sweep timers and that
+// subsequent OnPeerDisconnected calls are ignored.
+func TestStop_CancelsPendingWork(t *testing.T) {
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
+	withFakeClock(t, time.Now())
+
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+	// DeletePeers must NOT be called after Stop.
+
+	mgr := NewEphemeralManager(mockStore, peersMgr)
+	mgr.lifeTime = 100 * time.Millisecond
+	mgr.cleanupWindow = 10 * time.Millisecond
+	// Use a long delay so the initial-load timer is still pending.
+	mgr.initialLoadDelay = func() time.Duration { return time.Hour }
+
+	mgr.LoadInitialPeers(context.Background())
+	mgr.OnPeerDisconnected(context.Background(), &nbpeer.Peer{
+		ID: "p1", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: timeNow()},
+	})
+
+	mgr.accountsLock.Lock()
+	require.NotNil(t, mgr.initialLoadTimer, "initial-load timer should be armed")
+	require.Len(t, mgr.accounts, 1, "account should be tracked after disconnect")
+	mgr.accountsLock.Unlock()
+
+	mgr.Stop()
+
+	mgr.accountsLock.Lock()
+	require.Empty(t, mgr.accounts, "Stop should clear tracked accounts")
+	require.True(t, mgr.stopped, "stopped flag must be set")
+	mgr.accountsLock.Unlock()
+
+	// Post-stop disconnect must be ignored.
+	mgr.OnPeerDisconnected(context.Background(), &nbpeer.Peer{
+		ID: "p2", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: timeNow()},
+	})
+	mgr.accountsLock.Lock()
+	require.Empty(t, mgr.accounts, "disconnects after Stop must be ignored")
+	mgr.accountsLock.Unlock()
+}
+
+// TestOnPeerConnected_IsNoop: the OnPeerConnected hook is preserved on
+// the interface but does nothing in the per-account model — the sweep
+// query filters connected peers at the DB level.
+func TestOnPeerConnected_IsNoop(t *testing.T) {
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
+	withFakeClock(t, time.Now())
+
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+
+	mgr := newManagerForTest(t, mockStore, peersMgr)
+	mgr.OnPeerDisconnected(context.Background(), &nbpeer.Peer{
+		ID: "p1", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: timeNow()},
+	})
+	mgr.accountsLock.Lock()
+	require.Len(t, mgr.accounts, 1, "disconnect should track the account")
+	mgr.accountsLock.Unlock()
+
+	mgr.OnPeerConnected(context.Background(), &nbpeer.Peer{ID: "p1", AccountID: "acc-1", Ephemeral: true})
+	mgr.accountsLock.Lock()
+	require.Len(t, mgr.accounts, 1, "OnPeerConnected must be a no-op")
+	mgr.accountsLock.Unlock()
+}
+
+// TestSweep_StoreErrorReArms: if the stale-peer query fails, the
+// account must remain tracked and a follow-up sweep gets scheduled.
+func TestSweep_StoreErrorReArms(t *testing.T) {
+	mockStore := &erroringStore{
+		MockStore: MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)},
+	}
+	getNow, setNow := withFakeClock(t, time.Now())
+
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+
+	mgr := newManagerForTest(t, mockStore, peersMgr)
+
+	p := &nbpeer.Peer{ID: "p1", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: getNow()}}
+	mockStore.account.Peers[p.ID] = p
+	mgr.OnPeerDisconnected(context.Background(), p)
+
+	mockStore.fail.Store(true)
+	setNow(getNow().Add(mgr.lifeTime + 5*mgr.cleanupWindow))
+
+	// Wait until the failing sweep has run at least once.
+	require.Eventually(t, func() bool { return mockStore.failedCalls.Load() >= 1 },
+		2*time.Second, 5*time.Millisecond, "expected at least one failing sweep")
+
+	mgr.accountsLock.Lock()
+	_, stillTracked := mgr.accounts["acc-1"]
+	mgr.accountsLock.Unlock()
+	require.True(t, stillTracked, "account must remain tracked after a sweep error")
+
+	// Recover and ensure the rearmed sweep cleans up.
+	peersMgr.EXPECT().
+		DeletePeers(gomock.Any(), "acc-1", gomock.Any(), gomock.Any(), true).
+		DoAndReturn(func(_ context.Context, _ string, peerIDs []string, _ string, _ bool) error {
+			mockStore.mu.Lock()
+			for _, id := range peerIDs {
+				delete(mockStore.account.Peers, id)
+			}
+			mockStore.mu.Unlock()
+			return nil
+		}).AnyTimes()
+	mockStore.fail.Store(false)
+
+	require.Eventually(t, func() bool {
+		mockStore.mu.Lock()
+		defer mockStore.mu.Unlock()
+		_, gone := mockStore.account.Peers["p1"]
+		return !gone
+	}, 2*time.Second, 5*time.Millisecond, "rearmed sweep should clean up after the store recovers")
+}
+
+// erroringStore is a MockStore that can be flipped into a failing mode
+// to exercise the sweep's error-rearm path.
+type erroringStore struct {
+	MockStore
+	fail        atomic.Bool
+	failedCalls atomic.Int32
+}
+
+func (s *erroringStore) GetStaleEphemeralPeerIDsForAccount(ctx context.Context, accountID string, olderThan time.Time) ([]string, error) {
+	if s.fail.Load() {
+		s.failedCalls.Add(1)
+		return nil, errors.New("synthetic store error")
+	}
+	return s.MockStore.GetStaleEphemeralPeerIDsForAccount(ctx, accountID, olderThan)
+}
+
+// TestDefaultInitialLoadDelay confirms the jitter falls inside the
+// documented [8m, 10m) range — sanity check for the production timer.
+func TestDefaultInitialLoadDelay(t *testing.T) {
+	for i := 0; i < 1000; i++ {
+		d := defaultInitialLoadDelay()
+		assert.GreaterOrEqual(t, d, initialLoadMinDelay)
+		assert.Less(t, d, initialLoadMaxDelay)
 	}
 }
 
@@ -351,3 +596,7 @@ func newAccountWithId(ctx context.Context, accountID, userID, domain string, dis
 	}
 	return acc
 }
+
+// silence the import "github.com/netbirdio/netbird/management/internals/modules/peers/ephemeral"
+// (still needed indirectly for ephemeral.EphemeralLifeTime in production paths).
+var _ = ephemeral.EphemeralLifeTime

management/internals/modules/reverseproxy/domain/manager/manager.go

@@ -304,10 +304,27 @@ func (m Manager) getClusterAllowList(ctx context.Context, accountID string) ([]s
 	if err != nil {
 		return nil, fmt.Errorf("get BYOP cluster addresses: %w", err)
 	}
-	if len(byopAddresses) > 0 {
-		return byopAddresses, nil
+	publicAddresses, err := m.proxyManager.GetActiveClusterAddresses(ctx)
+	if err != nil {
+		return nil, fmt.Errorf("get public cluster addresses: %w", err)
 	}
-	return m.proxyManager.GetActiveClusterAddresses(ctx)
+	seen := make(map[string]struct{}, len(byopAddresses)+len(publicAddresses))
+	merged := make([]string, 0, len(byopAddresses)+len(publicAddresses))
+	for _, addr := range byopAddresses {
+		if _, ok := seen[addr]; ok {
+			continue
+		}
+		seen[addr] = struct{}{}
+		merged = append(merged, addr)
+	}
+	for _, addr := range publicAddresses {
+		if _, ok := seen[addr]; ok {
+			continue
+		}
+		seen[addr] = struct{}{}
+		merged = append(merged, addr)
+	}
+	return merged, nil
 }
 
 func extractClusterFromCustomDomains(serviceDomain string, customDomains []*domain.Domain) (string, bool) {

management/internals/modules/reverseproxy/domain/manager/manager_test.go

@@ -40,22 +40,37 @@ func (m *mockProxyManager) ClusterSupportsCrowdSec(_ context.Context, _ string)
 	return nil
 }
 
-func TestGetClusterAllowList_BYOPProxy(t *testing.T) {
+func TestGetClusterAllowList_BYOPMergedWithPublic(t *testing.T) {
 	pm := &mockProxyManager{
 		getActiveClusterAddressesForAccountFunc: func(_ context.Context, accID string) ([]string, error) {
 			assert.Equal(t, "acc-123", accID)
 			return []string{"byop.example.com"}, nil
 		},
 		getActiveClusterAddressesFunc: func(_ context.Context) ([]string, error) {
-			t.Fatal("should not call GetActiveClusterAddresses when BYOP addresses exist")
-			return nil, nil
+			return []string{"eu.proxy.netbird.io"}, nil
 		},
 	}
 
 	mgr := Manager{proxyManager: pm}
 	result, err := mgr.getClusterAllowList(context.Background(), "acc-123")
 	require.NoError(t, err)
-	assert.Equal(t, []string{"byop.example.com"}, result)
+	assert.Equal(t, []string{"byop.example.com", "eu.proxy.netbird.io"}, result)
+}
+
+func TestGetClusterAllowList_DeduplicatesBYOPAndPublic(t *testing.T) {
+	pm := &mockProxyManager{
+		getActiveClusterAddressesForAccountFunc: func(_ context.Context, _ string) ([]string, error) {
+			return []string{"shared.example.com", "byop.example.com"}, nil
+		},
+		getActiveClusterAddressesFunc: func(_ context.Context) ([]string, error) {
+			return []string{"shared.example.com", "eu.proxy.netbird.io"}, nil
+		},
+	}
+
+	mgr := Manager{proxyManager: pm}
+	result, err := mgr.getClusterAllowList(context.Background(), "acc-123")
+	require.NoError(t, err)
+	assert.Equal(t, []string{"shared.example.com", "byop.example.com", "eu.proxy.netbird.io"}, result)
 }
 
 func TestGetClusterAllowList_NoBYOP_FallbackToShared(t *testing.T) {
@@ -79,10 +94,6 @@ func TestGetClusterAllowList_BYOPError_ReturnsError(t *testing.T) {
 		getActiveClusterAddressesForAccountFunc: func(_ context.Context, _ string) ([]string, error) {
 			return nil, errors.New("db error")
 		},
-		getActiveClusterAddressesFunc: func(_ context.Context) ([]string, error) {
-			t.Fatal("should not call GetActiveClusterAddresses when BYOP lookup fails")
-			return nil, nil
-		},
 	}
 
 	mgr := Manager{proxyManager: pm}
@@ -92,6 +103,23 @@ func TestGetClusterAllowList_BYOPError_ReturnsError(t *testing.T) {
 	assert.Contains(t, err.Error(), "BYOP cluster addresses")
 }
 
+func TestGetClusterAllowList_PublicError_ReturnsError(t *testing.T) {
+	pm := &mockProxyManager{
+		getActiveClusterAddressesForAccountFunc: func(_ context.Context, _ string) ([]string, error) {
+			return []string{"byop.example.com"}, nil
+		},
+		getActiveClusterAddressesFunc: func(_ context.Context) ([]string, error) {
+			return nil, errors.New("db error")
+		},
+	}
+
+	mgr := Manager{proxyManager: pm}
+	result, err := mgr.getClusterAllowList(context.Background(), "acc-123")
+	require.Error(t, err)
+	assert.Nil(t, result)
+	assert.Contains(t, err.Error(), "public cluster addresses")
+}
+
 func TestGetClusterAllowList_BYOPEmptySlice_FallbackToShared(t *testing.T) {
 	pm := &mockProxyManager{
 		getActiveClusterAddressesForAccountFunc: func(_ context.Context, _ string) ([]string, error) {
@@ -108,3 +136,19 @@ func TestGetClusterAllowList_BYOPEmptySlice_FallbackToShared(t *testing.T) {
 	assert.Equal(t, []string{"eu.proxy.netbird.io"}, result)
 }
 
+func TestGetClusterAllowList_PublicEmpty_BYOPOnly(t *testing.T) {
+	pm := &mockProxyManager{
+		getActiveClusterAddressesForAccountFunc: func(_ context.Context, _ string) ([]string, error) {
+			return []string{"byop.example.com"}, nil
+		},
+		getActiveClusterAddressesFunc: func(_ context.Context) ([]string, error) {
+			return nil, nil
+		},
+	}
+
+	mgr := Manager{proxyManager: pm}
+	result, err := mgr.getClusterAllowList(context.Background(), "acc-123")
+	require.NoError(t, err)
+	assert.Equal(t, []string{"byop.example.com"}, result)
+}
+

management/internals/modules/reverseproxy/service/manager/manager.go

@@ -306,6 +306,10 @@ func (m *Manager) validateSubdomainRequirement(ctx context.Context, domain, clus
 func (m *Manager) persistNewService(ctx context.Context, accountID string, svc *service.Service) error {
 	customPorts := m.clusterCustomPorts(ctx, svc)
 
+	if err := validateTargetReferences(ctx, m.store, accountID, svc.Targets); err != nil {
+		return err
+	}
+
 	return m.store.ExecuteInTransaction(ctx, func(transaction store.Store) error {
 		if svc.Domain != "" {
 			if err := m.checkDomainAvailable(ctx, transaction, svc.Domain, ""); err != nil {
@@ -321,10 +325,6 @@ func (m *Manager) persistNewService(ctx context.Context, accountID string, svc *
 			return err
 		}
 
-		if err := validateTargetReferences(ctx, transaction, accountID, svc.Targets); err != nil {
-			return err
-		}
-
 		if err := transaction.CreateService(ctx, svc); err != nil {
 			return fmt.Errorf("create service: %w", err)
 		}
@@ -435,6 +435,10 @@ func (m *Manager) assignPort(ctx context.Context, tx store.Store, cluster string
 func (m *Manager) persistNewEphemeralService(ctx context.Context, accountID, peerID string, svc *service.Service) error {
 	customPorts := m.clusterCustomPorts(ctx, svc)
 
+	if err := validateTargetReferences(ctx, m.store, accountID, svc.Targets); err != nil {
+		return err
+	}
+
 	return m.store.ExecuteInTransaction(ctx, func(transaction store.Store) error {
 		if err := m.validateEphemeralPreconditions(ctx, transaction, accountID, peerID, svc); err != nil {
 			return err
@@ -448,10 +452,6 @@ func (m *Manager) persistNewEphemeralService(ctx context.Context, accountID, pee
 			return err
 		}
 
-		if err := validateTargetReferences(ctx, transaction, accountID, svc.Targets); err != nil {
-			return err
-		}
-
 		if err := transaction.CreateService(ctx, svc); err != nil {
 			return fmt.Errorf("create service: %w", err)
 		}
@@ -552,10 +552,22 @@ func (m *Manager) persistServiceUpdate(ctx context.Context, accountID string, se
 	svcForCaps.ProxyCluster = effectiveCluster
 	customPorts := m.clusterCustomPorts(ctx, &svcForCaps)
 
+	if err := validateTargetReferences(ctx, m.store, accountID, service.Targets); err != nil {
+		return nil, err
+	}
+
+	// Validate subdomain requirement *before* the transaction: the underlying
+	// capability lookup talks to the main DB pool, and SQLite's single-connection
+	// pool would self-deadlock if this ran while the tx already held the only
+	// connection.
+	if err := m.validateSubdomainRequirement(ctx, service.Domain, effectiveCluster); err != nil {
+		return nil, err
+	}
+
 	var updateInfo serviceUpdateInfo
 
 	err = m.store.ExecuteInTransaction(ctx, func(transaction store.Store) error {
-		return m.executeServiceUpdate(ctx, transaction, accountID, service, &updateInfo, customPorts)
+		return m.executeServiceUpdate(ctx, transaction, accountID, service, &updateInfo, customPorts, effectiveCluster)
 	})
 
 	return &updateInfo, err
@@ -585,7 +597,7 @@ func (m *Manager) resolveEffectiveCluster(ctx context.Context, accountID string,
 	return existing.ProxyCluster, nil
 }
 
-func (m *Manager) executeServiceUpdate(ctx context.Context, transaction store.Store, accountID string, service *service.Service, updateInfo *serviceUpdateInfo, customPorts *bool) error {
+func (m *Manager) executeServiceUpdate(ctx context.Context, transaction store.Store, accountID string, service *service.Service, updateInfo *serviceUpdateInfo, customPorts *bool, effectiveCluster string) error {
 	existingService, err := transaction.GetServiceByID(ctx, store.LockingStrengthUpdate, accountID, service.ID)
 	if err != nil {
 		return err
@@ -603,17 +615,13 @@ func (m *Manager) executeServiceUpdate(ctx context.Context, transaction store.St
 	updateInfo.domainChanged = existingService.Domain != service.Domain
 
 	if updateInfo.domainChanged {
-		if err := m.handleDomainChange(ctx, transaction, accountID, service); err != nil {
+		if err := m.handleDomainChange(ctx, transaction, service, effectiveCluster); err != nil {
 			return err
 		}
 	} else {
 		service.ProxyCluster = existingService.ProxyCluster
 	}
 
-	if err := m.validateSubdomainRequirement(ctx, service.Domain, service.ProxyCluster); err != nil {
-		return err
-	}
-
 	m.preserveExistingAuthSecrets(service, existingService)
 	if err := validateHeaderAuthValues(service.Auth.HeaderAuths); err != nil {
 		return err
@@ -628,9 +636,6 @@ func (m *Manager) executeServiceUpdate(ctx context.Context, transaction store.St
 	if err := m.checkPortConflict(ctx, transaction, service); err != nil {
 		return err
 	}
-	if err := validateTargetReferences(ctx, transaction, accountID, service.Targets); err != nil {
-		return err
-	}
 	if err := transaction.UpdateService(ctx, service); err != nil {
 		return fmt.Errorf("update service: %w", err)
 	}
@@ -638,20 +643,18 @@ func (m *Manager) executeServiceUpdate(ctx context.Context, transaction store.St
 	return nil
 }
 
-func (m *Manager) handleDomainChange(ctx context.Context, transaction store.Store, accountID string, svc *service.Service) error {
+// handleDomainChange validates the new domain is free inside the transaction
+// and applies the pre-resolved cluster (computed outside the tx by
+// resolveEffectiveCluster). It must NOT call clusterDeriver here: that talks
+// to the main DB pool and would self-deadlock under SQLite (max_open_conns=1)
+// because the transaction already holds the only connection.
+func (m *Manager) handleDomainChange(ctx context.Context, transaction store.Store, svc *service.Service, effectiveCluster string) error {
 	if err := m.checkDomainAvailable(ctx, transaction, svc.Domain, svc.ID); err != nil {
 		return err
 	}
-
-	if m.clusterDeriver != nil {
-		newCluster, err := m.clusterDeriver.DeriveClusterFromDomain(ctx, accountID, svc.Domain)
-		if err != nil {
-			log.WithError(err).Warnf("could not derive cluster from domain %s", svc.Domain)
-		} else {
-			svc.ProxyCluster = newCluster
-		}
+	if effectiveCluster != "" {
+		svc.ProxyCluster = effectiveCluster
 	}
-
 	return nil
 }
 

management/internals/modules/reverseproxy/service/service.go

@@ -381,13 +381,14 @@ func (s *Service) buildPathMappings() []*proto.PathMapping {
 		}
 
 		// HTTP/HTTPS: build full URL
+		hostNoBrackets := strings.TrimSuffix(strings.TrimPrefix(target.Host, "["), "]")
 		targetURL := url.URL{
 			Scheme: target.Protocol,
-			Host:   target.Host,
+			Host:   bracketIPv6Host(hostNoBrackets),
 			Path:   "/",
 		}
 		if target.Port > 0 && !isDefaultPort(target.Protocol, target.Port) {
-			targetURL.Host = net.JoinHostPort(targetURL.Host, strconv.FormatUint(uint64(target.Port), 10))
+			targetURL.Host = net.JoinHostPort(hostNoBrackets, strconv.FormatUint(uint64(target.Port), 10))
 		}
 
 		path := "/"
@@ -405,6 +406,19 @@ func (s *Service) buildPathMappings() []*proto.PathMapping {
 	return pathMappings
 }
 
+// bracketIPv6Host wraps host in square brackets when it is an IPv6 literal, as
+// required for the Host field of net/url.URL (RFC 3986 §3.2.2). v4-mapped IPv6
+// addresses are bracketed too since their textual form contains colons.
+func bracketIPv6Host(host string) string {
+	if strings.HasPrefix(host, "[") {
+		return host
+	}
+	if addr, err := netip.ParseAddr(host); err == nil && addr.Is6() {
+		return "[" + host + "]"
+	}
+	return host
+}
+
 func operationToProtoType(op Operation) proto.ProxyMappingUpdateType {
 	switch op {
 	case Create:

management/internals/modules/reverseproxy/service/service_test.go

@@ -351,6 +351,83 @@ func TestToProtoMapping_PortInTargetURL(t *testing.T) {
 			port:       80,
 			wantTarget: "https://10.0.0.1:80/",
 		},
+		{
+			name:       "domain host without port is unchanged",
+			protocol:   "http",
+			host:       "example.com",
+			port:       0,
+			wantTarget: "http://example.com/",
+		},
+		{
+			name:       "domain host with non-default port is unchanged",
+			protocol:   "http",
+			host:       "example.com",
+			port:       8080,
+			wantTarget: "http://example.com:8080/",
+		},
+		{
+			name:       "ipv6 host without port is bracketed",
+			protocol:   "http",
+			host:       "fb00:cafe:1::3",
+			port:       0,
+			wantTarget: "http://[fb00:cafe:1::3]/",
+		},
+		{
+			name:       "ipv6 host with default port omits port and brackets host",
+			protocol:   "http",
+			host:       "fb00:cafe:1::3",
+			port:       80,
+			wantTarget: "http://[fb00:cafe:1::3]/",
+		},
+		{
+			name:       "ipv6 host with non-default port is bracketed",
+			protocol:   "http",
+			host:       "fb00:cafe:1::3",
+			port:       8080,
+			wantTarget: "http://[fb00:cafe:1::3]:8080/",
+		},
+		{
+			name:       "ipv6 loopback without port is bracketed",
+			protocol:   "http",
+			host:       "::1",
+			port:       0,
+			wantTarget: "http://[::1]/",
+		},
+		{
+			name:       "ipv6 host with 5-digit port is bracketed",
+			protocol:   "http",
+			host:       "fb00:cafe::1",
+			port:       18080,
+			wantTarget: "http://[fb00:cafe::1]:18080/",
+		},
+		{
+			name:       "pre-bracketed ipv6 without port stays single-bracketed",
+			protocol:   "http",
+			host:       "[fb00:cafe::1]",
+			port:       0,
+			wantTarget: "http://[fb00:cafe::1]/",
+		},
+		{
+			name:       "pre-bracketed ipv6 with port is not double-bracketed",
+			protocol:   "http",
+			host:       "[fb00:cafe::1]",
+			port:       8080,
+			wantTarget: "http://[fb00:cafe::1]:8080/",
+		},
+		{
+			name:       "v4-mapped ipv6 host without port is bracketed",
+			protocol:   "http",
+			host:       "::ffff:10.0.0.1",
+			port:       0,
+			wantTarget: "http://[::ffff:10.0.0.1]/",
+		},
+		{
+			name:       "full-form 8-group ipv6 without port is bracketed",
+			protocol:   "http",
+			host:       "fb00:cafe:1:0:0:0:0:3",
+			port:       0,
+			wantTarget: "http://[fb00:cafe:1:0:0:0:0:3]/",
+		},
 	}
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {

management/internals/server/controllers.go

@@ -112,7 +112,11 @@ func (s *BaseServer) AuthManager() auth.Manager {
 
 func (s *BaseServer) EphemeralManager() ephemeral.Manager {
 	return Create(s, func() ephemeral.Manager {
-		return manager.NewEphemeralManager(s.Store(), s.PeersManager())
+		em := manager.NewEphemeralManager(s.Store(), s.PeersManager())
+		if metrics := s.Metrics(); metrics != nil {
+			em.SetMetrics(metrics.EphemeralPeersMetrics())
+		}
+		return em
 	})
 }
 

management/internals/server/server.go

@@ -34,8 +34,6 @@ const (
 	ManagementLegacyPort = 33073
 	// DefaultSelfHostedDomain is the default domain used for self-hosted fresh installs.
 	DefaultSelfHostedDomain = "netbird.selfhosted"
-
-	ContainerKeyBaseServer = "baseServer"
 )
 
 type Server interface {
@@ -93,7 +91,7 @@ type Config struct {
 
 // NewServer initializes and configures a new Server instance
 func NewServer(cfg *Config) *BaseServer {
-	s := &BaseServer{
+	return &BaseServer{
 		Config:                      cfg.NbConfig,
 		container:                   make(map[string]any),
 		dnsDomain:                   cfg.DNSDomain,
@@ -106,9 +104,6 @@ func NewServer(cfg *Config) *BaseServer {
 		mgmtMetricsPort:             cfg.MgmtMetricsPort,
 		autoResolveDomains:          cfg.AutoResolveDomains,
 	}
-	s.container[ContainerKeyBaseServer] = s
-
-	return s
 }
 
 func (s *BaseServer) AfterInit(fn func(s *BaseServer)) {

management/internals/shared/grpc/proxy.go

@@ -394,6 +394,13 @@ func (s *ProxyServiceServer) sendSnapshot(ctx context.Context, conn *proxyConnec
 		if end > len(mappings) {
 			end = len(mappings)
 		}
+		for _, m := range mappings[i:end] {
+			token, err := s.tokenStore.GenerateToken(m.AccountId, m.Id, s.proxyTokenTTL())
+			if err != nil {
+				return fmt.Errorf("generate auth token for service %s: %w", m.Id, err)
+			}
+			m.AuthToken = token
+		}
 		if err := conn.stream.Send(&proto.GetMappingUpdateResponse{
 			Mapping:             mappings[i:end],
 			InitialSyncComplete: end == len(mappings),
@@ -425,18 +432,14 @@ func (s *ProxyServiceServer) snapshotServiceMappings(ctx context.Context, conn *
 		return nil, fmt.Errorf("get services from store: %w", err)
 	}
 
+	oidcCfg := s.GetOIDCValidationConfig()
 	var mappings []*proto.ProxyMapping
 	for _, service := range services {
 		if !service.Enabled || service.ProxyCluster == "" || service.ProxyCluster != conn.address {
 			continue
 		}
 
-		token, err := s.tokenStore.GenerateToken(service.AccountID, service.ID, s.proxyTokenTTL())
-		if err != nil {
-			return nil, fmt.Errorf("generate auth token for service %s: %w", service.ID, err)
-		}
-
-		m := service.ToProtoMapping(rpservice.Create, token, s.GetOIDCValidationConfig())
+		m := service.ToProtoMapping(rpservice.Create, "", oidcCfg)
 		if !proxyAcceptsMapping(conn, m) {
 			continue
 		}

management/internals/shared/grpc/proxy_snapshot_test.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"fmt"
 	"testing"
+	"time"
 
 	"github.com/golang/mock/gomock"
 	"github.com/stretchr/testify/assert"
@@ -172,3 +173,55 @@ func TestSendSnapshot_EmptySnapshot(t *testing.T) {
 	assert.Empty(t, stream.messages[0].Mapping)
 	assert.True(t, stream.messages[0].InitialSyncComplete)
 }
+
+type hookingStream struct {
+	grpc.ServerStream
+	onSend func(*proto.GetMappingUpdateResponse)
+}
+
+func (s *hookingStream) Send(m *proto.GetMappingUpdateResponse) error {
+	if s.onSend != nil {
+		s.onSend(m)
+	}
+	return nil
+}
+
+func (s *hookingStream) Context() context.Context     { return context.Background() }
+func (s *hookingStream) SetHeader(metadata.MD) error  { return nil }
+func (s *hookingStream) SendHeader(metadata.MD) error { return nil }
+func (s *hookingStream) SetTrailer(metadata.MD)       {}
+func (s *hookingStream) SendMsg(any) error            { return nil }
+func (s *hookingStream) RecvMsg(any) error            { return nil }
+
+func TestSendSnapshot_TokensRemainValidUnderSlowSend(t *testing.T) {
+	const cluster = "cluster.example.com"
+	const batchSize = 2
+	const totalServices = 6
+	const ttl = 100 * time.Millisecond
+	const sendDelay = 200 * time.Millisecond
+
+	ctrl := gomock.NewController(t)
+	mgr := rpservice.NewMockManager(ctrl)
+	mgr.EXPECT().GetGlobalServices(gomock.Any()).Return(makeServices(totalServices, cluster), nil)
+
+	s := newSnapshotTestServer(t, batchSize)
+	s.serviceManager = mgr
+	s.tokenTTL = ttl
+
+	var validateErrs []error
+	stream := &hookingStream{
+		onSend: func(resp *proto.GetMappingUpdateResponse) {
+			for _, m := range resp.Mapping {
+				if err := s.tokenStore.ValidateAndConsume(m.AuthToken, m.AccountId, m.Id); err != nil {
+					validateErrs = append(validateErrs, fmt.Errorf("svc %s: %w", m.Id, err))
+				}
+			}
+			time.Sleep(sendDelay)
+		},
+	}
+	conn := &proxyConnection{proxyID: "proxy-a", address: cluster, stream: stream}
+
+	require.NoError(t, s.sendSnapshot(context.Background(), conn))
+	require.Empty(t, validateErrs,
+		"tokens must remain valid even when batches are sent slowly: lazy per-batch generation guarantees freshness")
+}

management/internals/shared/grpc/server.go

@@ -522,10 +522,11 @@ func (s *Server) sendJob(ctx context.Context, peerKey wgtypes.Key, job *job.Even
 }
 
 func (s *Server) cancelPeerRoutines(ctx context.Context, accountID string, peer *nbpeer.Peer, streamStartTime time.Time) {
-	unlock := s.acquirePeerLockByUID(ctx, peer.Key)
+	uncanceledCTX := context.WithoutCancel(ctx)
+	unlock := s.acquirePeerLockByUID(uncanceledCTX, peer.Key)
 	defer unlock()
 
-	s.cancelPeerRoutinesWithoutLock(ctx, accountID, peer, streamStartTime)
+	s.cancelPeerRoutinesWithoutLock(uncanceledCTX, accountID, peer, streamStartTime)
 }
 
 func (s *Server) cancelPeerRoutinesWithoutLock(ctx context.Context, accountID string, peer *nbpeer.Peer, streamStartTime time.Time) {

management/internals/shared/grpc/validate_session_test.go

@@ -326,17 +326,25 @@ func (m *testValidateSessionServiceManager) GetActiveClusters(_ context.Context,
 	return nil, nil
 }
 
+func (m *testValidateSessionServiceManager) DeleteAccountCluster(_ context.Context, _, _, _ string) error {
+	return nil
+}
+
 type testValidateSessionProxyManager struct{}
 
-func (m *testValidateSessionProxyManager) Connect(_ context.Context, _, _, _ string, _ *string, _ *proxy.Capabilities) error {
+func (m *testValidateSessionProxyManager) Connect(_ context.Context, _, _, _, _ string, _ *string, _ *proxy.Capabilities) (*proxy.Proxy, error) {
+	return nil, nil
+}
+
+func (m *testValidateSessionProxyManager) Disconnect(_ context.Context, _, _ string) error {
 	return nil
 }
 
-func (m *testValidateSessionProxyManager) Disconnect(_ context.Context, _ string) error {
+func (m *testValidateSessionProxyManager) Heartbeat(_ context.Context, _ *proxy.Proxy) error {
 	return nil
 }
 
-func (m *testValidateSessionProxyManager) Heartbeat(_ context.Context, _, _, _ string) error {
+func (m *testValidateSessionProxyManager) DeleteAccountCluster(_ context.Context, _, _ string) error {
 	return nil
 }
 

management/server/account.go

@@ -291,10 +291,15 @@ func (am *DefaultAccountManager) UpdateAccountSettings(ctx context.Context, acco
 		return nil, status.NewPermissionDeniedError()
 	}
 
+	// Canonicalize the incoming range so a caller-supplied prefix with host bits
+	// (e.g. 100.64.1.1/16) compares equal to the masked form stored on network.Net.
+	newSettings.NetworkRange = newSettings.NetworkRange.Masked()
+
 	var oldSettings *types.Settings
 	var updateAccountPeers bool
 	var groupChangesAffectPeers bool
 	var reloadReverseProxy bool
+	var effectiveOldNetworkRange netip.Prefix
 
 	err = am.Store.ExecuteInTransaction(ctx, func(transaction store.Store) error {
 		var groupsUpdated bool
@@ -308,6 +313,16 @@ func (am *DefaultAccountManager) UpdateAccountSettings(ctx context.Context, acco
 			return err
 		}
 
+		// No lock: the transaction already holds Settings(Update), and network.Net is
+		// only mutated by reallocateAccountPeerIPs, which is reachable only through
+		// this same code path. A Share lock here would extend an unnecessary row lock
+		// and complicate ordering against updatePeerIPv6InTransaction.
+		network, err := transaction.GetAccountNetwork(ctx, store.LockingStrengthNone, accountID)
+		if err != nil {
+			return fmt.Errorf("get account network: %w", err)
+		}
+		effectiveOldNetworkRange = prefixFromIPNet(network.Net)
+
 		if oldSettings.Extra != nil && newSettings.Extra != nil &&
 			oldSettings.Extra.PeerApprovalEnabled && !newSettings.Extra.PeerApprovalEnabled {
 			approvedCount, err := transaction.ApproveAccountPeers(ctx, accountID)
@@ -321,7 +336,7 @@ func (am *DefaultAccountManager) UpdateAccountSettings(ctx context.Context, acco
 			}
 		}
 
-		if oldSettings.NetworkRange != newSettings.NetworkRange {
+		if newSettings.NetworkRange.IsValid() && newSettings.NetworkRange != effectiveOldNetworkRange {
 			if err = am.reallocateAccountPeerIPs(ctx, transaction, accountID, newSettings.NetworkRange); err != nil {
 				return err
 			}
@@ -396,9 +411,9 @@ func (am *DefaultAccountManager) UpdateAccountSettings(ctx context.Context, acco
 		}
 		am.StoreEvent(ctx, userID, accountID, accountID, activity.AccountDNSDomainUpdated, eventMeta)
 	}
-	if oldSettings.NetworkRange != newSettings.NetworkRange {
+	if newSettings.NetworkRange.IsValid() && newSettings.NetworkRange != effectiveOldNetworkRange {
 		eventMeta := map[string]any{
-			"old_network_range": oldSettings.NetworkRange.String(),
+			"old_network_range": effectiveOldNetworkRange.String(),
 			"new_network_range": newSettings.NetworkRange.String(),
 		}
 		am.StoreEvent(ctx, userID, accountID, accountID, activity.AccountNetworkRangeUpdated, eventMeta)
@@ -443,6 +458,22 @@ func ipv6SettingsChanged(old, updated *types.Settings) bool {
 	return !slices.Equal(oldGroups, newGroups)
 }
 
+// prefixFromIPNet returns the overlay prefix actually allocated on the account
+// network, or an invalid prefix if none is set. Settings.NetworkRange is a
+// user-facing override that is empty on legacy accounts, so the effective
+// range must be read from network.Net to compare against an incoming update.
+func prefixFromIPNet(ipNet net.IPNet) netip.Prefix {
+	if ipNet.IP == nil {
+		return netip.Prefix{}
+	}
+	addr, ok := netip.AddrFromSlice(ipNet.IP)
+	if !ok {
+		return netip.Prefix{}
+	}
+	ones, _ := ipNet.Mask.Size()
+	return netip.PrefixFrom(addr.Unmap(), ones)
+}
+
 func (am *DefaultAccountManager) validateSettingsUpdate(ctx context.Context, transaction store.Store, newSettings, oldSettings *types.Settings, userID, accountID string) error {
 	halfYearLimit := 180 * 24 * time.Hour
 	if newSettings.PeerLoginExpiration > halfYearLimit {
@@ -1837,35 +1868,32 @@ func domainIsUpToDate(domain string, domainCategory string, userAuth auth.UserAu
 	return domainCategory == types.PrivateCategory || userAuth.DomainCategory != types.PrivateCategory || domain != userAuth.Domain
 }
 
+// SyncAndMarkPeer is the per-Sync entry point: it refreshes the peer's
+// network map and then marks the peer connected with a session token
+// derived from syncTime (the moment the gRPC stream opened). Any
+// concurrent stream that started earlier loses the optimistic-lock race
+// in MarkPeerConnected and bails without writing.
 func (am *DefaultAccountManager) SyncAndMarkPeer(ctx context.Context, accountID string, peerPubKey string, meta nbpeer.PeerSystemMeta, realIP net.IP, syncTime time.Time) (*nbpeer.Peer, *types.NetworkMap, []*posture.Checks, int64, error) {
 	peer, netMap, postureChecks, dnsfwdPort, err := am.SyncPeer(ctx, types.PeerSync{WireGuardPubKey: peerPubKey, Meta: meta}, accountID)
 	if err != nil {
 		return nil, nil, nil, 0, fmt.Errorf("error syncing peer: %w", err)
 	}
 
-	err = am.MarkPeerConnected(ctx, peerPubKey, true, realIP, accountID, syncTime)
-	if err != nil {
+	if err := am.MarkPeerConnected(ctx, peerPubKey, realIP, accountID, syncTime.UnixNano()); err != nil {
 		log.WithContext(ctx).Warnf("failed marking peer as connected %s %v", peerPubKey, err)
 	}
 
 	return peer, netMap, postureChecks, dnsfwdPort, nil
 }
 
+// OnPeerDisconnected is invoked when a sync stream ends. It marks the
+// peer disconnected only when the stored SessionStartedAt matches the
+// nanosecond token derived from streamStartTime — i.e. only when this
+// is the stream that currently owns the peer's session. A mismatch
+// means a newer stream has already replaced us, so the disconnect is
+// dropped.
 func (am *DefaultAccountManager) OnPeerDisconnected(ctx context.Context, accountID string, peerPubKey string, streamStartTime time.Time) error {
-	peer, err := am.Store.GetPeerByPeerPubKey(ctx, store.LockingStrengthNone, peerPubKey)
-	if err != nil {
-		log.WithContext(ctx).Warnf("failed to get peer %s for disconnect check: %v", peerPubKey, err)
-		return nil
-	}
-
-	if peer.Status.LastSeen.After(streamStartTime) {
-		log.WithContext(ctx).Tracef("peer %s has newer activity (lastSeen=%s > streamStart=%s), skipping disconnect",
-			peerPubKey, peer.Status.LastSeen.Format(time.RFC3339), streamStartTime.Format(time.RFC3339))
-		return nil
-	}
-
-	err = am.MarkPeerConnected(ctx, peerPubKey, false, nil, accountID, time.Now().UTC())
-	if err != nil {
+	if err := am.MarkPeerDisconnected(ctx, peerPubKey, accountID, streamStartTime.UnixNano()); err != nil {
 		log.WithContext(ctx).Warnf("failed marking peer as disconnected %s %v", peerPubKey, err)
 	}
 	return nil
@@ -2487,6 +2515,18 @@ func (am *DefaultAccountManager) buildIPv6AllowedPeers(ctx context.Context, tran
 			allowedPeers[peerID] = struct{}{}
 		}
 	}
+
+	// Embedded proxy peers sit outside regular group membership but must
+	// participate in any v6-enabled overlay to reach v6-only peers.
+	peers, err := transaction.GetAccountPeers(ctx, store.LockingStrengthNone, accountID, "", "")
+	if err != nil {
+		return nil, fmt.Errorf("get peers: %w", err)
+	}
+	for _, p := range peers {
+		if p.ProxyMeta.Embedded {
+			allowedPeers[p.ID] = struct{}{}
+		}
+	}
 	return allowedPeers, nil
 }
 

management/server/account/manager.go

@@ -61,7 +61,8 @@ type Manager interface {
 	GetUserFromUserAuth(ctx context.Context, userAuth auth.UserAuth) (*types.User, error)
 	ListUsers(ctx context.Context, accountID string) ([]*types.User, error)
 	GetPeers(ctx context.Context, accountID, userID, nameFilter, ipFilter string) ([]*nbpeer.Peer, error)
-	MarkPeerConnected(ctx context.Context, peerKey string, connected bool, realIP net.IP, accountID string, syncTime time.Time) error
+	MarkPeerConnected(ctx context.Context, peerKey string, realIP net.IP, accountID string, sessionStartedAt int64) error
+	MarkPeerDisconnected(ctx context.Context, peerKey string, accountID string, sessionStartedAt int64) error
 	DeletePeer(ctx context.Context, accountID, peerID, userID string) error
 	UpdatePeer(ctx context.Context, accountID, userID string, p *nbpeer.Peer) (*nbpeer.Peer, error)
 	UpdatePeerIP(ctx context.Context, accountID, userID, peerID string, newIP netip.Addr) error

management/server/account/manager_mock.go

@@ -1305,17 +1305,31 @@ func (mr *MockManagerMockRecorder) LoginPeer(ctx, login interface{}) *gomock.Cal
 }
 
 // MarkPeerConnected mocks base method.
-func (m *MockManager) MarkPeerConnected(ctx context.Context, peerKey string, connected bool, realIP net.IP, accountID string, syncTime time.Time) error {
+func (m *MockManager) MarkPeerConnected(ctx context.Context, peerKey string, realIP net.IP, accountID string, sessionStartedAt int64) error {
 	m.ctrl.T.Helper()
-	ret := m.ctrl.Call(m, "MarkPeerConnected", ctx, peerKey, connected, realIP, accountID, syncTime)
+	ret := m.ctrl.Call(m, "MarkPeerConnected", ctx, peerKey, realIP, accountID, sessionStartedAt)
 	ret0, _ := ret[0].(error)
 	return ret0
 }
 
 // MarkPeerConnected indicates an expected call of MarkPeerConnected.
-func (mr *MockManagerMockRecorder) MarkPeerConnected(ctx, peerKey, connected, realIP, accountID, syncTime interface{}) *gomock.Call {
+func (mr *MockManagerMockRecorder) MarkPeerConnected(ctx, peerKey, realIP, accountID, sessionStartedAt interface{}) *gomock.Call {
 	mr.mock.ctrl.T.Helper()
-	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "MarkPeerConnected", reflect.TypeOf((*MockManager)(nil).MarkPeerConnected), ctx, peerKey, connected, realIP, accountID, syncTime)
+	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "MarkPeerConnected", reflect.TypeOf((*MockManager)(nil).MarkPeerConnected), ctx, peerKey, realIP, accountID, sessionStartedAt)
+}
+
+// MarkPeerDisconnected mocks base method.
+func (m *MockManager) MarkPeerDisconnected(ctx context.Context, peerKey string, accountID string, sessionStartedAt int64) error {
+	m.ctrl.T.Helper()
+	ret := m.ctrl.Call(m, "MarkPeerDisconnected", ctx, peerKey, accountID, sessionStartedAt)
+	ret0, _ := ret[0].(error)
+	return ret0
+}
+
+// MarkPeerDisconnected indicates an expected call of MarkPeerDisconnected.
+func (mr *MockManagerMockRecorder) MarkPeerDisconnected(ctx, peerKey, accountID, sessionStartedAt interface{}) *gomock.Call {
+	mr.mock.ctrl.T.Helper()
+	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "MarkPeerDisconnected", reflect.TypeOf((*MockManager)(nil).MarkPeerDisconnected), ctx, peerKey, accountID, sessionStartedAt)
 }
 
 // OnPeerDisconnected mocks base method.

management/server/account_test.go

@@ -1813,7 +1813,7 @@ func TestDefaultAccountManager_UpdatePeer_PeerLoginExpiration(t *testing.T) {
 	accountID, err := manager.GetAccountIDByUserID(context.Background(), auth.UserAuth{UserId: userID})
 	require.NoError(t, err, "unable to get the account")
 
-	err = manager.MarkPeerConnected(context.Background(), key.PublicKey().String(), true, nil, accountID, time.Now().UTC())
+	err = manager.MarkPeerConnected(context.Background(), key.PublicKey().String(), nil, accountID, time.Now().UTC().UnixNano())
 	require.NoError(t, err, "unable to mark peer connected")
 
 	_, err = manager.UpdateAccountSettings(context.Background(), accountID, userID, &types.Settings{
@@ -1884,7 +1884,7 @@ func TestDefaultAccountManager_MarkPeerConnected_PeerLoginExpiration(t *testing.
 	require.NoError(t, err, "unable to get the account")
 
 	// when we mark peer as connected, the peer login expiration routine should trigger
-	err = manager.MarkPeerConnected(context.Background(), key.PublicKey().String(), true, nil, accountID, time.Now().UTC())
+	err = manager.MarkPeerConnected(context.Background(), key.PublicKey().String(), nil, accountID, time.Now().UTC().UnixNano())
 	require.NoError(t, err, "unable to mark peer connected")
 
 	failed := waitTimeout(wg, time.Second)
@@ -1910,15 +1910,16 @@ func TestDefaultAccountManager_OnPeerDisconnected_LastSeenCheck(t *testing.T) {
 	}, false)
 	require.NoError(t, err, "unable to add peer")
 
-	t.Run("disconnect peer when streamStartTime is after LastSeen", func(t *testing.T) {
-		err = manager.MarkPeerConnected(context.Background(), peerPubKey, true, nil, accountID, time.Now().UTC())
+	t.Run("disconnect peer when session token matches", func(t *testing.T) {
+		streamStartTime := time.Now().UTC()
+		err = manager.MarkPeerConnected(context.Background(), peerPubKey, nil, accountID, streamStartTime.UnixNano())
 		require.NoError(t, err, "unable to mark peer connected")
 
 		peer, err := manager.Store.GetPeerByPeerPubKey(context.Background(), store.LockingStrengthNone, peerPubKey)
 		require.NoError(t, err, "unable to get peer")
 		require.True(t, peer.Status.Connected, "peer should be connected")
-
-		streamStartTime := time.Now().UTC()
+		require.Equal(t, streamStartTime.UnixNano(), peer.Status.SessionStartedAt,
+			"SessionStartedAt should equal the token we passed in")
 
 		err = manager.OnPeerDisconnected(context.Background(), accountID, peerPubKey, streamStartTime)
 		require.NoError(t, err)
@@ -1926,49 +1927,127 @@ func TestDefaultAccountManager_OnPeerDisconnected_LastSeenCheck(t *testing.T) {
 		peer, err = manager.Store.GetPeerByPeerPubKey(context.Background(), store.LockingStrengthNone, peerPubKey)
 		require.NoError(t, err)
 		require.False(t, peer.Status.Connected, "peer should be disconnected")
+		require.Equal(t, int64(0), peer.Status.SessionStartedAt, "SessionStartedAt should be reset to 0")
 	})
 
-	t.Run("skip disconnect when LastSeen is after streamStartTime (zombie stream protection)", func(t *testing.T) {
-		err = manager.MarkPeerConnected(context.Background(), peerPubKey, true, nil, accountID, time.Now().UTC())
+	t.Run("skip disconnect when stored session is newer (zombie stream protection)", func(t *testing.T) {
+		// Newer stream wins on connect (sets SessionStartedAt = now ns).
+		streamStartTime := time.Now().UTC()
+		err = manager.MarkPeerConnected(context.Background(), peerPubKey, nil, accountID, streamStartTime.UnixNano())
 		require.NoError(t, err, "unable to mark peer connected")
 
 		peer, err := manager.Store.GetPeerByPeerPubKey(context.Background(), store.LockingStrengthNone, peerPubKey)
 		require.NoError(t, err)
 		require.True(t, peer.Status.Connected, "peer should be connected")
 
-		streamStartTime := peer.Status.LastSeen.Add(-1 * time.Hour)
+		// Older stream tries to mark disconnect with its own (older) session token —
+		// fencing kicks in and the write is dropped.
+		staleStreamStartTime := streamStartTime.Add(-1 * time.Hour)
 
-		err = manager.OnPeerDisconnected(context.Background(), accountID, peerPubKey, streamStartTime)
+		err = manager.OnPeerDisconnected(context.Background(), accountID, peerPubKey, staleStreamStartTime)
 		require.NoError(t, err)
 
 		peer, err = manager.Store.GetPeerByPeerPubKey(context.Background(), store.LockingStrengthNone, peerPubKey)
 		require.NoError(t, err)
 		require.True(t, peer.Status.Connected,
-			"peer should remain connected because LastSeen > streamStartTime (zombie stream protection)")
+			"peer should remain connected because the stored session is newer than the disconnect token")
+		require.Equal(t, streamStartTime.UnixNano(), peer.Status.SessionStartedAt,
+			"SessionStartedAt should still hold the winning stream's token")
 	})
 
-	t.Run("skip stale connect when peer already has newer LastSeen (blocked goroutine protection)", func(t *testing.T) {
+	t.Run("skip stale connect when stored session is newer (blocked goroutine protection)", func(t *testing.T) {
 		node2SyncTime := time.Now().UTC()
-		err = manager.MarkPeerConnected(context.Background(), peerPubKey, true, nil, accountID, node2SyncTime)
+		err = manager.MarkPeerConnected(context.Background(), peerPubKey, nil, accountID, node2SyncTime.UnixNano())
 		require.NoError(t, err, "node 2 should connect peer")
 
 		peer, err := manager.Store.GetPeerByPeerPubKey(context.Background(), store.LockingStrengthNone, peerPubKey)
 		require.NoError(t, err)
 		require.True(t, peer.Status.Connected, "peer should be connected")
-		require.Equal(t, node2SyncTime.Unix(), peer.Status.LastSeen.Unix(), "LastSeen should be node2SyncTime")
+		require.Equal(t, node2SyncTime.UnixNano(), peer.Status.SessionStartedAt,
+			"SessionStartedAt should equal node2SyncTime token")
 
 		node1StaleSyncTime := node2SyncTime.Add(-1 * time.Minute)
-		err = manager.MarkPeerConnected(context.Background(), peerPubKey, true, nil, accountID, node1StaleSyncTime)
+		err = manager.MarkPeerConnected(context.Background(), peerPubKey, nil, accountID, node1StaleSyncTime.UnixNano())
 		require.NoError(t, err, "stale connect should not return error")
 
 		peer, err = manager.Store.GetPeerByPeerPubKey(context.Background(), store.LockingStrengthNone, peerPubKey)
 		require.NoError(t, err)
 		require.True(t, peer.Status.Connected, "peer should still be connected")
-		require.Equal(t, node2SyncTime.Unix(), peer.Status.LastSeen.Unix(),
-			"LastSeen should NOT be overwritten by stale syncTime from blocked goroutine")
+		require.Equal(t, node2SyncTime.UnixNano(), peer.Status.SessionStartedAt,
+			"SessionStartedAt should NOT be overwritten by stale token from blocked goroutine")
 	})
 }
 
+// TestDefaultAccountManager_MarkPeerConnected_ConcurrentRace exercises the
+// fencing protocol under contention: many goroutines race to mark the
+// same peer connected with distinct session tokens at the same time.
+// The contract is that the highest token always wins and is what remains
+// in the store, regardless of execution order.
+func TestDefaultAccountManager_MarkPeerConnected_ConcurrentRace(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 get 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: "race-peer"},
+	}, false)
+	require.NoError(t, err, "unable to add peer")
+
+	const workers = 16
+	base := time.Now().UTC().UnixNano()
+	tokens := make([]int64, workers)
+	for i := range tokens {
+		// Spread tokens by 1ms so the comparison is unambiguous; the
+		// largest is index workers-1.
+		tokens[i] = base + int64(i)*int64(time.Millisecond)
+	}
+	expected := tokens[workers-1]
+
+	var ready sync.WaitGroup
+	ready.Add(workers)
+	var start sync.WaitGroup
+	start.Add(1)
+	var done sync.WaitGroup
+	done.Add(workers)
+
+	// require.* calls t.FailNow which is documented as unsafe from
+	// non-test goroutines (it calls runtime.Goexit on the wrong stack and
+	// races with the WaitGroup). Collect errors here and assert from the
+	// main goroutine after done.Wait().
+	errs := make(chan error, workers)
+
+	for i := 0; i < workers; i++ {
+		token := tokens[i]
+		go func() {
+			defer done.Done()
+			ready.Done()
+			start.Wait()
+			errs <- manager.MarkPeerConnected(context.Background(), peerPubKey, nil, accountID, token)
+		}()
+	}
+
+	ready.Wait()
+	start.Done()
+	done.Wait()
+	close(errs)
+	for err := range errs {
+		require.NoError(t, err, "MarkPeerConnected must not error under contention")
+	}
+
+	peer, err := manager.Store.GetPeerByPeerPubKey(context.Background(), store.LockingStrengthNone, peerPubKey)
+	require.NoError(t, err)
+	require.True(t, peer.Status.Connected, "peer should be connected after the race")
+	require.Equal(t, expected, peer.Status.SessionStartedAt,
+		"the largest token must win regardless of execution order")
+}
+
 func TestDefaultAccountManager_UpdateAccountSettings_PeerLoginExpiration(t *testing.T) {
 	manager, _, err := createManager(t)
 	require.NoError(t, err, "unable to create account manager")
@@ -1991,7 +2070,7 @@ func TestDefaultAccountManager_UpdateAccountSettings_PeerLoginExpiration(t *test
 	account, err := manager.Store.GetAccount(context.Background(), accountID)
 	require.NoError(t, err, "unable to get the account")
 
-	err = manager.MarkPeerConnected(context.Background(), key.PublicKey().String(), true, nil, accountID, time.Now().UTC())
+	err = manager.MarkPeerConnected(context.Background(), key.PublicKey().String(), nil, accountID, time.Now().UTC().UnixNano())
 	require.NoError(t, err, "unable to mark peer connected")
 
 	wg := &sync.WaitGroup{}
@@ -3970,6 +4049,96 @@ func TestDefaultAccountManager_UpdateAccountSettings_NetworkRangeChange(t *testi
 	}
 }
 
+// TestDefaultAccountManager_UpdateAccountSettings_NetworkRangePreserved guards against
+// peer IP reallocation when a settings update carries the network range that is already
+// in use. Legacy accounts have Settings.NetworkRange unset in the DB while network.Net
+// holds the actual allocated overlay; the dashboard backfills the GET response from
+// network.Net and echoes the value back on PUT, so the diff must be against the
+// effective range to avoid renumbering every peer on an unrelated settings change.
+func TestDefaultAccountManager_UpdateAccountSettings_NetworkRangePreserved(t *testing.T) {
+	manager, _, account, peer1, peer2, peer3 := setupNetworkMapTest(t)
+	ctx := context.Background()
+
+	settings, err := manager.Store.GetAccountSettings(ctx, store.LockingStrengthNone, account.Id)
+	require.NoError(t, err)
+	require.False(t, settings.NetworkRange.IsValid(), "precondition: new accounts leave Settings.NetworkRange unset")
+
+	network, err := manager.Store.GetAccountNetwork(ctx, store.LockingStrengthNone, account.Id)
+	require.NoError(t, err)
+	require.NotNil(t, network.Net.IP, "precondition: network.Net should be allocated")
+	addr, ok := netip.AddrFromSlice(network.Net.IP)
+	require.True(t, ok)
+	ones, _ := network.Net.Mask.Size()
+	effective := netip.PrefixFrom(addr.Unmap(), ones)
+	require.True(t, effective.IsValid())
+
+	before := map[string]netip.Addr{peer1.ID: peer1.IP, peer2.ID: peer2.IP, peer3.ID: peer3.IP}
+
+	// Round-trip the effective range as if the dashboard echoed back the GET-backfilled value.
+	_, err = manager.UpdateAccountSettings(ctx, account.Id, userID, &types.Settings{
+		PeerLoginExpirationEnabled: true,
+		PeerLoginExpiration:        types.DefaultPeerLoginExpiration,
+		NetworkRange:               effective,
+		Extra:                      &types.ExtraSettings{},
+	})
+	require.NoError(t, err)
+
+	peers, err := manager.Store.GetAccountPeers(ctx, store.LockingStrengthNone, account.Id, "", "")
+	require.NoError(t, err)
+	require.Len(t, peers, len(before))
+	for _, p := range peers {
+		assert.Equal(t, before[p.ID], p.IP, "peer %s IP should not change when range matches effective", p.ID)
+	}
+
+	// Carrying the same range with host bits set must also be a no-op once canonicalized.
+	hostBitsForm := netip.PrefixFrom(peer1.IP, ones)
+	require.NotEqual(t, effective, hostBitsForm, "precondition: host-bit form should differ before masking")
+	_, err = manager.UpdateAccountSettings(ctx, account.Id, userID, &types.Settings{
+		PeerLoginExpirationEnabled: true,
+		PeerLoginExpiration:        types.DefaultPeerLoginExpiration,
+		NetworkRange:               hostBitsForm,
+		Extra:                      &types.ExtraSettings{},
+	})
+	require.NoError(t, err)
+
+	peers, err = manager.Store.GetAccountPeers(ctx, store.LockingStrengthNone, account.Id, "", "")
+	require.NoError(t, err)
+	for _, p := range peers {
+		assert.Equal(t, before[p.ID], p.IP, "peer %s IP should not change for host-bit-set equivalent range", p.ID)
+	}
+
+	// Omitting NetworkRange (invalid prefix) must also be a no-op.
+	_, err = manager.UpdateAccountSettings(ctx, account.Id, userID, &types.Settings{
+		PeerLoginExpirationEnabled: true,
+		PeerLoginExpiration:        types.DefaultPeerLoginExpiration,
+		Extra:                      &types.ExtraSettings{},
+	})
+	require.NoError(t, err)
+
+	peers, err = manager.Store.GetAccountPeers(ctx, store.LockingStrengthNone, account.Id, "", "")
+	require.NoError(t, err)
+	for _, p := range peers {
+		assert.Equal(t, before[p.ID], p.IP, "peer %s IP should not change when NetworkRange omitted", p.ID)
+	}
+
+	// Sanity: an actually different range still triggers reallocation.
+	newRange := netip.MustParsePrefix("100.99.0.0/16")
+	_, err = manager.UpdateAccountSettings(ctx, account.Id, userID, &types.Settings{
+		PeerLoginExpirationEnabled: true,
+		PeerLoginExpiration:        types.DefaultPeerLoginExpiration,
+		NetworkRange:               newRange,
+		Extra:                      &types.ExtraSettings{},
+	})
+	require.NoError(t, err)
+
+	peers, err = manager.Store.GetAccountPeers(ctx, store.LockingStrengthNone, account.Id, "", "")
+	require.NoError(t, err)
+	for _, p := range peers {
+		assert.True(t, newRange.Contains(p.IP), "peer %s should be in new range %s, got %s", p.ID, newRange, p.IP)
+		assert.NotEqual(t, before[p.ID], p.IP, "peer %s IP should change on real range update", p.ID)
+	}
+}
+
 func TestDefaultAccountManager_UpdateAccountSettings_IPv6EnabledGroups(t *testing.T) {
 	manager, _, account, peer1, peer2, peer3 := setupNetworkMapTest(t)
 	ctx := context.Background()

management/server/idp/embedded.go

@@ -6,7 +6,9 @@ import (
 	"errors"
 	"fmt"
 	"net/http"
+	"net/url"
 	"os"
+	"path"
 	"strings"
 
 	"github.com/dexidp/dex/storage"
@@ -138,10 +140,13 @@ func (c *EmbeddedIdPConfig) ToYAMLConfig() (*dex.YAMLConfig, error) {
 		return nil, fmt.Errorf("invalid IdP storage config: %w", err)
 	}
 
-	// Build CLI redirect URIs including the device callback (both relative and absolute)
+	// Build CLI redirect URIs including the device callback. Dex uses the issuer-relative
+	// path (for example, /oauth2/device/callback) when completing the device flow, so
+	// include it explicitly in addition to the legacy bare path and absolute URL.
 	cliRedirectURIs := c.CLIRedirectURIs
 	cliRedirectURIs = append(cliRedirectURIs, "/device/callback")
-	cliRedirectURIs = append(cliRedirectURIs, c.Issuer+"/device/callback")
+	cliRedirectURIs = append(cliRedirectURIs, issuerRelativeDeviceCallback(c.Issuer))
+	cliRedirectURIs = append(cliRedirectURIs, strings.TrimSuffix(c.Issuer, "/")+"/device/callback")
 
 	// Build dashboard redirect URIs including the OAuth callback for proxy authentication
 	dashboardRedirectURIs := c.DashboardRedirectURIs
@@ -154,6 +159,10 @@ func (c *EmbeddedIdPConfig) ToYAMLConfig() (*dex.YAMLConfig, error) {
 	// MGMT api and the dashboard, adding baseURL means less configuration for the instance admin
 	dashboardPostLogoutRedirectURIs = append(dashboardPostLogoutRedirectURIs, baseURL)
 
+	redirectURIs := make([]string, 0)
+	redirectURIs = append(redirectURIs, cliRedirectURIs...)
+	redirectURIs = append(redirectURIs, dashboardRedirectURIs...)
+
 	cfg := &dex.YAMLConfig{
 		Issuer: c.Issuer,
 		Storage: dex.Storage{
@@ -179,14 +188,14 @@ func (c *EmbeddedIdPConfig) ToYAMLConfig() (*dex.YAMLConfig, error) {
 				ID:                     staticClientDashboard,
 				Name:                   "NetBird Dashboard",
 				Public:                 true,
-				RedirectURIs:           dashboardRedirectURIs,
+				RedirectURIs:           redirectURIs,
 				PostLogoutRedirectURIs: sanitizePostLogoutRedirectURIs(dashboardPostLogoutRedirectURIs),
 			},
 			{
 				ID:           staticClientCLI,
 				Name:         "NetBird CLI",
 				Public:       true,
-				RedirectURIs: cliRedirectURIs,
+				RedirectURIs: redirectURIs,
 			},
 		},
 		StaticConnectors: c.StaticConnectors,
@@ -217,6 +226,14 @@ func (c *EmbeddedIdPConfig) ToYAMLConfig() (*dex.YAMLConfig, error) {
 	return cfg, nil
 }
 
+func issuerRelativeDeviceCallback(issuer string) string {
+	u, err := url.Parse(issuer)
+	if err != nil || u.Path == "" {
+		return "/device/callback"
+	}
+	return path.Join(u.Path, "/device/callback")
+}
+
 // Due to how the frontend generates the logout, sometimes it appends a trailing slash
 // and because Dex only allows exact matches, we need to make sure we always have both
 // versions of each provided uri
@@ -299,7 +316,7 @@ func resolveSessionCookieEncryptionKey(configuredKey string) (string, error) {
 		}
 	}
 
-	return "", fmt.Errorf("invalid embedded IdP session cookie encryption key: %s (or sessionCookieEncryptionKey) must be 16, 24, or 32 bytes as a raw string or base64-encoded to one of those lengths; got %d raw bytes", sessionCookieEncryptionKeyEnv, len([]byte(key)))
+	return "", fmt.Errorf("invalid embedded IdP session cookie encryption key:%s (or sessionCookieEncryptionKey) must be 16, 24, or 32 bytes as a raw string or base64-encoded to one of those lengths; got %d raw bytes", sessionCookieEncryptionKeyEnv, len([]byte(key)))
 }
 
 func validSessionCookieEncryptionKeyLength(length int) bool {

management/server/idp/embedded_test.go

@@ -314,6 +314,34 @@ func TestEmbeddedIdPManager_UpdateUserPassword(t *testing.T) {
 	})
 }
 
+func TestEmbeddedIdPConfig_ToYAMLConfig_IncludesDeviceCallbackRedirectURI(t *testing.T) {
+	config := &EmbeddedIdPConfig{
+		Enabled: true,
+		Issuer:  "https://example.com/oauth2",
+		Storage: EmbeddedStorageConfig{
+			Type: "sqlite3",
+			Config: EmbeddedStorageTypeConfig{
+				File: filepath.Join(t.TempDir(), "dex.db"),
+			},
+		},
+	}
+
+	yamlConfig, err := config.ToYAMLConfig()
+	require.NoError(t, err)
+
+	var cliRedirectURIs []string
+	for _, client := range yamlConfig.StaticClients {
+		if client.ID == staticClientCLI {
+			cliRedirectURIs = client.RedirectURIs
+			break
+		}
+	}
+	require.NotEmpty(t, cliRedirectURIs)
+	assert.Contains(t, cliRedirectURIs, "/device/callback")
+	assert.Contains(t, cliRedirectURIs, "/oauth2/device/callback")
+	assert.Contains(t, cliRedirectURIs, "https://example.com/oauth2/device/callback")
+}
+
 func TestEmbeddedIdPConfig_ToYAMLConfig_SessionCookieEncryptionKey(t *testing.T) {
 	t.Setenv(sessionCookieEncryptionKeyEnv, "")
 

management/server/mock_server/account_mock.go

@@ -38,7 +38,8 @@ type MockAccountManager struct {
 	GetUserFromUserAuthFunc               func(ctx context.Context, userAuth auth.UserAuth) (*types.User, error)
 	ListUsersFunc                         func(ctx context.Context, accountID string) ([]*types.User, error)
 	GetPeersFunc                          func(ctx context.Context, accountID, userID, nameFilter, ipFilter string) ([]*nbpeer.Peer, error)
-	MarkPeerConnectedFunc                 func(ctx context.Context, peerKey string, connected bool, realIP net.IP, syncTime time.Time) error
+	MarkPeerConnectedFunc                 func(ctx context.Context, peerKey string, realIP net.IP, accountID string, sessionStartedAt int64) error
+	MarkPeerDisconnectedFunc              func(ctx context.Context, peerKey string, accountID string, sessionStartedAt int64) error
 	SyncAndMarkPeerFunc                   func(ctx context.Context, accountID string, peerPubKey string, meta nbpeer.PeerSystemMeta, realIP net.IP, syncTime time.Time) (*nbpeer.Peer, *types.NetworkMap, []*posture.Checks, int64, error)
 	DeletePeerFunc                        func(ctx context.Context, accountID, peerKey, userID string) error
 	GetNetworkMapFunc                     func(ctx context.Context, peerKey string) (*types.NetworkMap, error)
@@ -227,7 +228,14 @@ func (am *MockAccountManager) SyncAndMarkPeer(ctx context.Context, accountID str
 	return nil, nil, nil, 0, status.Errorf(codes.Unimplemented, "method MarkPeerConnected is not implemented")
 }
 
-func (am *MockAccountManager) OnPeerDisconnected(_ context.Context, accountID string, peerPubKey string, streamStartTime time.Time) error {
+func (am *MockAccountManager) OnPeerDisconnected(ctx context.Context, accountID string, peerPubKey string, streamStartTime time.Time) error {
+	// Mirror DefaultAccountManager.OnPeerDisconnected: drive the fencing
+	// hook so tests that inject MarkPeerDisconnectedFunc actually observe
+	// disconnect events. Falls through to nil when no hook is set, which
+	// is the original behaviour.
+	if am.MarkPeerDisconnectedFunc != nil {
+		return am.MarkPeerDisconnectedFunc(ctx, peerPubKey, accountID, streamStartTime.UnixNano())
+	}
 	return nil
 }
 
@@ -328,13 +336,21 @@ func (am *MockAccountManager) GetAccountIDByUserID(ctx context.Context, userAuth
 }
 
 // MarkPeerConnected mock implementation of MarkPeerConnected from server.AccountManager interface
-func (am *MockAccountManager) MarkPeerConnected(ctx context.Context, peerKey string, connected bool, realIP net.IP, accountID string, syncTime time.Time) error {
+func (am *MockAccountManager) MarkPeerConnected(ctx context.Context, peerKey string, realIP net.IP, accountID string, sessionStartedAt int64) error {
 	if am.MarkPeerConnectedFunc != nil {
-		return am.MarkPeerConnectedFunc(ctx, peerKey, connected, realIP, syncTime)
+		return am.MarkPeerConnectedFunc(ctx, peerKey, realIP, accountID, sessionStartedAt)
 	}
 	return status.Errorf(codes.Unimplemented, "method MarkPeerConnected is not implemented")
 }
 
+// MarkPeerDisconnected mock implementation of MarkPeerDisconnected from server.AccountManager interface
+func (am *MockAccountManager) MarkPeerDisconnected(ctx context.Context, peerKey string, accountID string, sessionStartedAt int64) error {
+	if am.MarkPeerDisconnectedFunc != nil {
+		return am.MarkPeerDisconnectedFunc(ctx, peerKey, accountID, sessionStartedAt)
+	}
+	return status.Errorf(codes.Unimplemented, "method MarkPeerDisconnected is not implemented")
+}
+
 // DeleteAccount mock implementation of DeleteAccount from server.AccountManager interface
 func (am *MockAccountManager) DeleteAccount(ctx context.Context, accountID, userID string) error {
 	if am.DeleteAccountFunc != nil {

management/server/peer.go

@@ -16,7 +16,6 @@ import (
 	"golang.org/x/exp/maps"
 
 	nbdns "github.com/netbirdio/netbird/dns"
-	"github.com/netbirdio/netbird/management/server/geolocation"
 	"github.com/netbirdio/netbird/management/server/idp"
 	routerTypes "github.com/netbirdio/netbird/management/server/networks/routers/types"
 	"github.com/netbirdio/netbird/management/server/permissions/modules"
@@ -29,6 +28,7 @@ import (
 
 	"github.com/netbirdio/netbird/management/server/activity"
 	nbpeer "github.com/netbirdio/netbird/management/server/peer"
+	"github.com/netbirdio/netbird/management/server/telemetry"
 	"github.com/netbirdio/netbird/shared/management/status"
 )
 
@@ -63,56 +63,64 @@ func (am *DefaultAccountManager) GetPeers(ctx context.Context, accountID, userID
 	return am.Store.GetUserPeers(ctx, store.LockingStrengthNone, accountID, userID)
 }
 
-// MarkPeerConnected marks peer as connected (true) or disconnected (false)
-// syncTime is used as the LastSeen timestamp and for stale request detection
-func (am *DefaultAccountManager) MarkPeerConnected(ctx context.Context, peerPubKey string, connected bool, realIP net.IP, accountID string, syncTime time.Time) error {
-	var peer *nbpeer.Peer
-	var settings *types.Settings
-	var expired bool
-	var err error
-	var skipped bool
+// MarkPeerConnected marks a peer as connected with optimistic-locked
+// fencing on PeerStatus.SessionStartedAt. The sessionStartedAt argument
+// is the start time of the gRPC sync stream that owns this update,
+// expressed as Unix nanoseconds — only the call whose token is greater
+// than what's stored wins. LastSeen is written by the database itself;
+// we never pass it down.
+//
+// Disconnects use MarkPeerDisconnected and require the session to match
+// exactly; see PeerStatus.SessionStartedAt for the protocol.
+func (am *DefaultAccountManager) MarkPeerConnected(ctx context.Context, peerPubKey string, realIP net.IP, accountID string, sessionStartedAt int64) error {
+	start := time.Now()
+	defer func() {
+		am.metrics.AccountManagerMetrics().RecordPeerStatusUpdateDuration(telemetry.PeerStatusConnect, time.Since(start))
+	}()
 
-	err = am.Store.ExecuteInTransaction(ctx, func(transaction store.Store) error {
-		peer, err = transaction.GetPeerByPeerPubKey(ctx, store.LockingStrengthUpdate, peerPubKey)
-		if err != nil {
-			return err
+	peer, err := am.Store.GetPeerByPeerPubKey(ctx, store.LockingStrengthNone, peerPubKey)
+	if err != nil {
+		outcome := telemetry.PeerStatusError
+		if s, ok := status.FromError(err); ok && s.Type() == status.NotFound {
+			outcome = telemetry.PeerStatusPeerNotFound
 		}
-
-		if connected && !syncTime.After(peer.Status.LastSeen) {
-			log.WithContext(ctx).Tracef("peer %s has newer activity (lastSeen=%s >= syncTime=%s), skipping connect",
-				peer.ID, peer.Status.LastSeen.Format(time.RFC3339), syncTime.Format(time.RFC3339))
-			skipped = true
-			return nil
-		}
-
-		expired, err = updatePeerStatusAndLocation(ctx, am.geo, transaction, peer, connected, realIP, accountID, syncTime)
+		am.metrics.AccountManagerMetrics().CountPeerStatusUpdate(telemetry.PeerStatusConnect, outcome)
 		return err
-	})
-	if skipped {
+	}
+
+	updated, err := am.Store.MarkPeerConnectedIfNewerSession(ctx, accountID, peer.ID, sessionStartedAt)
+	if err != nil {
+		am.metrics.AccountManagerMetrics().CountPeerStatusUpdate(telemetry.PeerStatusConnect, telemetry.PeerStatusError)
+		return err
+	}
+	if !updated {
+		am.metrics.AccountManagerMetrics().CountPeerStatusUpdate(telemetry.PeerStatusConnect, telemetry.PeerStatusStale)
+		log.WithContext(ctx).Tracef("peer %s already has a newer session in store, skipping connect", peer.ID)
 		return nil
 	}
-	if err != nil {
-		return err
+	am.metrics.AccountManagerMetrics().CountPeerStatusUpdate(telemetry.PeerStatusConnect, telemetry.PeerStatusApplied)
+
+	if am.geo != nil && realIP != nil {
+		am.updatePeerLocationIfChanged(ctx, accountID, peer, realIP)
 	}
 
+	expired := peer.Status != nil && peer.Status.LoginExpired
+
 	if peer.AddedWithSSOLogin() {
-		settings, err = am.Store.GetAccountSettings(ctx, store.LockingStrengthNone, accountID)
+		settings, err := am.Store.GetAccountSettings(ctx, store.LockingStrengthNone, accountID)
 		if err != nil {
 			return err
 		}
-
 		if peer.LoginExpirationEnabled && settings.PeerLoginExpirationEnabled {
 			am.schedulePeerLoginExpiration(ctx, accountID)
 		}
-
 		if peer.InactivityExpirationEnabled && settings.PeerInactivityExpirationEnabled {
 			am.checkAndSchedulePeerInactivityExpiration(ctx, accountID)
 		}
 	}
 
 	if expired {
-		err = am.networkMapController.OnPeersUpdated(ctx, accountID, []string{peer.ID})
-		if err != nil {
+		if err = am.networkMapController.OnPeersUpdated(ctx, accountID, []string{peer.ID}); err != nil {
 			return fmt.Errorf("notify network map controller of peer update: %w", err)
 		}
 	}
@@ -120,41 +128,60 @@ func (am *DefaultAccountManager) MarkPeerConnected(ctx context.Context, peerPubK
 	return nil
 }
 
-func updatePeerStatusAndLocation(ctx context.Context, geo geolocation.Geolocation, transaction store.Store, peer *nbpeer.Peer, connected bool, realIP net.IP, accountID string, syncTime time.Time) (bool, error) {
-	oldStatus := peer.Status.Copy()
-	newStatus := oldStatus
-	newStatus.LastSeen = syncTime
-	newStatus.Connected = connected
-	// whenever peer got connected that means that it logged in successfully
-	if newStatus.Connected {
-		newStatus.LoginExpired = false
-	}
-	peer.Status = newStatus
+// MarkPeerDisconnected marks a peer as disconnected, but only when the
+// stored session token matches the one passed in. A mismatch means a
+// newer stream has already taken ownership of the peer — disconnects from
+// the older stream are ignored. LastSeen is written by the database.
+func (am *DefaultAccountManager) MarkPeerDisconnected(ctx context.Context, peerPubKey string, accountID string, sessionStartedAt int64) error {
+	start := time.Now()
+	defer func() {
+		am.metrics.AccountManagerMetrics().RecordPeerStatusUpdateDuration(telemetry.PeerStatusDisconnect, time.Since(start))
+	}()
 
-	if geo != nil && realIP != nil {
-		location, err := geo.Lookup(realIP)
-		if err != nil {
-			log.WithContext(ctx).Warnf("failed to get location for peer %s realip: [%s]: %v", peer.ID, realIP.String(), err)
-		} else {
-			peer.Location.ConnectionIP = realIP
-			peer.Location.CountryCode = location.Country.ISOCode
-			peer.Location.CityName = location.City.Names.En
-			peer.Location.GeoNameID = location.City.GeonameID
-			err = transaction.SavePeerLocation(ctx, accountID, peer)
-			if err != nil {
-				log.WithContext(ctx).Warnf("could not store location for peer %s: %s", peer.ID, err)
-			}
-		}
-	}
-
-	log.WithContext(ctx).Debugf("saving peer status for peer %s is connected: %t", peer.ID, connected)
-
-	err := transaction.SavePeerStatus(ctx, accountID, peer.ID, *newStatus)
+	peer, err := am.Store.GetPeerByPeerPubKey(ctx, store.LockingStrengthNone, peerPubKey)
 	if err != nil {
-		return false, err
+		outcome := telemetry.PeerStatusError
+		if s, ok := status.FromError(err); ok && s.Type() == status.NotFound {
+			outcome = telemetry.PeerStatusPeerNotFound
+		}
+		am.metrics.AccountManagerMetrics().CountPeerStatusUpdate(telemetry.PeerStatusDisconnect, outcome)
+		return err
 	}
 
-	return oldStatus.LoginExpired, nil
+	updated, err := am.Store.MarkPeerDisconnectedIfSameSession(ctx, accountID, peer.ID, sessionStartedAt)
+	if err != nil {
+		am.metrics.AccountManagerMetrics().CountPeerStatusUpdate(telemetry.PeerStatusDisconnect, telemetry.PeerStatusError)
+		return err
+	}
+	if !updated {
+		am.metrics.AccountManagerMetrics().CountPeerStatusUpdate(telemetry.PeerStatusDisconnect, telemetry.PeerStatusStale)
+		log.WithContext(ctx).Tracef("peer %s session token mismatch on disconnect (token=%d), skipping",
+			peer.ID, sessionStartedAt)
+		return nil
+	}
+	am.metrics.AccountManagerMetrics().CountPeerStatusUpdate(telemetry.PeerStatusDisconnect, telemetry.PeerStatusApplied)
+	return nil
+}
+
+// updatePeerLocationIfChanged refreshes the geolocation on a separate
+// row update, only when the connection IP actually changed. Geo lookups
+// are expensive so we skip same-IP reconnects.
+func (am *DefaultAccountManager) updatePeerLocationIfChanged(ctx context.Context, accountID string, peer *nbpeer.Peer, realIP net.IP) {
+	if peer.Location.ConnectionIP != nil && peer.Location.ConnectionIP.Equal(realIP) {
+		return
+	}
+	location, err := am.geo.Lookup(realIP)
+	if err != nil {
+		log.WithContext(ctx).Warnf("failed to get location for peer %s realip: [%s]: %v", peer.ID, realIP.String(), err)
+		return
+	}
+	peer.Location.ConnectionIP = realIP
+	peer.Location.CountryCode = location.Country.ISOCode
+	peer.Location.CityName = location.City.Names.En
+	peer.Location.GeoNameID = location.City.GeonameID
+	if err := am.Store.SavePeerLocation(ctx, accountID, peer); err != nil {
+		log.WithContext(ctx).Warnf("could not store location for peer %s: %s", peer.ID, err)
+	}
 }
 
 // UpdatePeer updates peer. Only Peer.Name, Peer.SSHEnabled, Peer.LoginExpirationEnabled and Peer.InactivityExpirationEnabled can be updated.
@@ -762,16 +789,19 @@ func (am *DefaultAccountManager) AddPeer(ctx context.Context, accountID, setupKe
 		newPeer.IP = freeIP
 
 		if len(settings.IPv6EnabledGroups) > 0 && network.NetV6.IP != nil {
-			var allGroupID string
-			if !peer.ProxyMeta.Embedded {
-				allGroup, err := am.Store.GetGroupByName(ctx, store.LockingStrengthNone, accountID, "All")
-				if err != nil {
-					log.WithContext(ctx).Debugf("get All group for IPv6 allocation: %v", err)
-				} else {
+			// Embedded proxy peers are not group members but participate in any
+			// IPv6-enabled overlay so reverse-proxy traffic reaches v6-only peers.
+			allocate := peer.ProxyMeta.Embedded
+			if !allocate {
+				var allGroupID string
+				if allGroup, err := am.Store.GetGroupByName(ctx, store.LockingStrengthNone, accountID, types.GroupAllName); err == nil {
 					allGroupID = allGroup.ID
+				} else {
+					log.WithContext(ctx).Debugf("get All group for IPv6 allocation: %v", err)
 				}
+				allocate = peerWillHaveIPv6(settings, peerAddConfig.GroupsToAdd, allGroupID)
 			}
-			if peerWillHaveIPv6(settings, peerAddConfig.GroupsToAdd, allGroupID) {
+			if allocate {
 				v6Prefix, err := netip.ParsePrefix(network.NetV6.String())
 				if err != nil {
 					return nil, nil, nil, fmt.Errorf("parse IPv6 prefix: %w", err)

management/server/peer/peer.go

@@ -74,8 +74,19 @@ type ProxyMeta struct {
 }
 
 type PeerStatus struct { //nolint:revive
-	// LastSeen is the last time peer was connected to the management service
+	// LastSeen is the last time the peer status was updated (i.e. the last
+	// time we observed the peer being alive on a sync stream). Written by
+	// the database (CURRENT_TIMESTAMP) — callers do not supply it.
 	LastSeen time.Time
+	// SessionStartedAt records when the currently-active sync stream began,
+	// stored as Unix nanoseconds. It acts as the optimistic-locking token
+	// for status updates: a stream is only allowed to mutate the peer's
+	// status when its own token strictly exceeds the stored token (when connecting)
+	// or matches it exactly (for disconnects). Zero means "no
+	// active session". Integer nanoseconds are used so equality is
+	// precision-safe across drivers, and so the predicates compose to a
+	// single bigint comparison.
+	SessionStartedAt int64
 	// Connected indicates whether peer is connected to the management service or not
 	Connected bool
 	// LoginExpired
@@ -375,10 +386,14 @@ func (p *Peer) EventMeta(dnsDomain string) map[string]any {
 	return meta
 }
 
-// Copy PeerStatus
+// Copy PeerStatus. SessionStartedAt must be propagated so clone-based
+// callers (Peer.Copy, MarkLoginExpired, UpdateLastLogin) don't silently
+// reset the fencing token to zero — that would let any subsequent
+// SavePeerStatus write reopen the optimistic-lock window.
 func (p *PeerStatus) Copy() *PeerStatus {
 	return &PeerStatus{
 		LastSeen:         p.LastSeen,
+		SessionStartedAt: p.SessionStartedAt,
 		Connected:        p.Connected,
 		LoginExpired:     p.LoginExpired,
 		RequiresApproval: p.RequiresApproval,

management/server/store/sql_store.go

@@ -8,6 +8,7 @@ import (
 	"fmt"
 	"net"
 	"net/netip"
+	"net/url"
 	"os"
 	"path/filepath"
 	"runtime"
@@ -497,8 +498,9 @@ func (s *SqlStore) SavePeerStatus(ctx context.Context, accountID, peerID string,
 	peerCopy.Status = &peerStatus
 
 	fieldsToUpdate := []string{
-		"peer_status_last_seen", "peer_status_connected",
-		"peer_status_login_expired", "peer_status_required_approval",
+		"peer_status_last_seen", "peer_status_session_started_at",
+		"peer_status_connected", "peer_status_login_expired",
+		"peer_status_requires_approval",
 	}
 	result := s.db.Model(&nbpeer.Peer{}).
 		Select(fieldsToUpdate).
@@ -515,6 +517,69 @@ func (s *SqlStore) SavePeerStatus(ctx context.Context, accountID, peerID string,
 	return nil
 }
 
+// MarkPeerConnectedIfNewerSession is an atomic optimistic-locked update.
+// The peer is marked connected with the given session token only when
+// the stored SessionStartedAt is strictly smaller than the incoming
+// one — equivalently, when no newer stream has already taken ownership.
+// The sentinel zero (set on peer creation or after a disconnect) counts
+// as the smallest possible token. This is the write half of the
+// fencing protocol described on PeerStatus.SessionStartedAt.
+//
+// The post-write side effects in the caller — geo lookup,
+// schedulePeerLoginExpiration, checkAndSchedulePeerInactivityExpiration,
+// OnPeersUpdated — all run AFTER this method returns and are deliberately
+// outside the database write so they cannot extend the row-lock window.
+//
+// LastSeen is set to the database's clock (CURRENT_TIMESTAMP) at the
+// moment the row is written. The caller never supplies LastSeen because
+// the value would otherwise drift under lock contention — a Go-side
+// time.Now() taken before the write can land minutes later than the
+// actual UPDATE under load, which previously caused real ordering bugs.
+func (s *SqlStore) MarkPeerConnectedIfNewerSession(ctx context.Context, accountID, peerID string, newSessionStartedAt int64) (bool, error) {
+	result := s.db.WithContext(ctx).
+		Model(&nbpeer.Peer{}).
+		Where(accountAndIDQueryCondition, accountID, peerID).
+		Where("peer_status_session_started_at < ?", newSessionStartedAt).
+		Updates(map[string]any{
+			"peer_status_connected":          true,
+			"peer_status_last_seen":          gorm.Expr("CURRENT_TIMESTAMP"),
+			"peer_status_session_started_at": newSessionStartedAt,
+			"peer_status_login_expired":      false,
+		})
+	if result.Error != nil {
+		return false, status.Errorf(status.Internal, "mark peer connected: %v", result.Error)
+	}
+	return result.RowsAffected > 0, nil
+}
+
+// MarkPeerDisconnectedIfSameSession is an atomic optimistic-locked update.
+// The peer is marked disconnected only when the stored SessionStartedAt
+// matches the incoming token — meaning the stream that owns the current
+// session is the one ending. If a newer stream has already replaced the
+// session, the update is skipped. LastSeen is set to CURRENT_TIMESTAMP at
+// write time; see MarkPeerConnectedIfNewerSession for the rationale.
+//
+// A zero sessionStartedAt is rejected at the call site; the underlying
+// WHERE on equality would otherwise match every never-connected peer.
+func (s *SqlStore) MarkPeerDisconnectedIfSameSession(ctx context.Context, accountID, peerID string, sessionStartedAt int64) (bool, error) {
+	if sessionStartedAt == 0 {
+		return false, nil
+	}
+	result := s.db.WithContext(ctx).
+		Model(&nbpeer.Peer{}).
+		Where(accountAndIDQueryCondition, accountID, peerID).
+		Where("peer_status_session_started_at = ?", sessionStartedAt).
+		Updates(map[string]any{
+			"peer_status_connected":          false,
+			"peer_status_last_seen":          gorm.Expr("CURRENT_TIMESTAMP"),
+			"peer_status_session_started_at": int64(0),
+		})
+	if result.Error != nil {
+		return false, status.Errorf(status.Internal, "mark peer disconnected: %v", result.Error)
+	}
+	return result.RowsAffected > 0, nil
+}
+
 func (s *SqlStore) SavePeerLocation(ctx context.Context, accountID string, peerWithLocation *nbpeer.Peer) error {
 	// To maintain data integrity, we create a copy of the peer's location to prevent unintended updates to other fields.
 	var peerCopy nbpeer.Peer
@@ -1722,9 +1787,10 @@ func (s *SqlStore) getPeers(ctx context.Context, accountID string) ([]nbpeer.Pee
 	inactivity_expiration_enabled, last_login, created_at, ephemeral, extra_dns_labels, allow_extra_dns_labels, meta_hostname,
 	meta_go_os, meta_kernel, meta_core, meta_platform, meta_os, meta_os_version, meta_wt_version, meta_ui_version,
 	meta_kernel_version, meta_network_addresses, meta_system_serial_number, meta_system_product_name, meta_system_manufacturer,
-	meta_environment, meta_flags, meta_files, meta_capabilities, peer_status_last_seen, peer_status_connected, peer_status_login_expired,
-	peer_status_requires_approval, location_connection_ip, location_country_code, location_city_name,
-	location_geo_name_id, proxy_meta_embedded, proxy_meta_cluster, ipv6 FROM peers WHERE account_id = $1`
+	meta_environment, meta_flags, meta_files, meta_capabilities, peer_status_last_seen, peer_status_session_started_at,
+	peer_status_connected, peer_status_login_expired, peer_status_requires_approval, location_connection_ip,
+	location_country_code, location_city_name, location_geo_name_id, proxy_meta_embedded, proxy_meta_cluster, ipv6
+	FROM peers WHERE account_id = $1`
 	rows, err := s.pool.Query(ctx, query, accountID)
 	if err != nil {
 		return nil, err
@@ -1737,6 +1803,7 @@ func (s *SqlStore) getPeers(ctx context.Context, accountID string) ([]nbpeer.Pee
 			lastLogin, createdAt                                                                            sql.NullTime
 			sshEnabled, loginExpirationEnabled, inactivityExpirationEnabled, ephemeral, allowExtraDNSLabels sql.NullBool
 			peerStatusLastSeen                                                                              sql.NullTime
+			peerStatusSessionStartedAt                                                                      sql.NullInt64
 			peerStatusConnected, peerStatusLoginExpired, peerStatusRequiresApproval, proxyEmbedded          sql.NullBool
 			ip, extraDNS, netAddr, env, flags, files, capabilities, connIP, ipv6                            []byte
 			metaHostname, metaGoOS, metaKernel, metaCore, metaPlatform                                      sql.NullString
@@ -1751,8 +1818,9 @@ func (s *SqlStore) getPeers(ctx context.Context, accountID string) ([]nbpeer.Pee
 			&allowExtraDNSLabels, &metaHostname, &metaGoOS, &metaKernel, &metaCore, &metaPlatform,
 			&metaOS, &metaOSVersion, &metaWtVersion, &metaUIVersion, &metaKernelVersion, &netAddr,
 			&metaSystemSerialNumber, &metaSystemProductName, &metaSystemManufacturer, &env, &flags, &files, &capabilities,
-			&peerStatusLastSeen, &peerStatusConnected, &peerStatusLoginExpired, &peerStatusRequiresApproval, &connIP,
-			&locationCountryCode, &locationCityName, &locationGeoNameID, &proxyEmbedded, &proxyCluster, &ipv6)
+			&peerStatusLastSeen, &peerStatusSessionStartedAt, &peerStatusConnected, &peerStatusLoginExpired,
+			&peerStatusRequiresApproval, &connIP, &locationCountryCode, &locationCityName, &locationGeoNameID,
+			&proxyEmbedded, &proxyCluster, &ipv6)
 
 		if err == nil {
 			if lastLogin.Valid {
@@ -1779,6 +1847,9 @@ func (s *SqlStore) getPeers(ctx context.Context, accountID string) ([]nbpeer.Pee
 			if peerStatusLastSeen.Valid {
 				p.Status.LastSeen = peerStatusLastSeen.Time
 			}
+			if peerStatusSessionStartedAt.Valid {
+				p.Status.SessionStartedAt = peerStatusSessionStartedAt.Int64
+			}
 			if peerStatusConnected.Valid {
 				p.Status.Connected = peerStatusConnected.Bool
 			}
@@ -2794,12 +2865,27 @@ func NewSqliteStore(ctx context.Context, dataDir string, metrics telemetry.AppMe
 		connStr = filepath.Join(dataDir, filePath)
 	}
 
-	// Append query parameters: user-provided take precedence, otherwise default to cache=shared on non-Windows
-	if hasQuery {
-		connStr += "?" + query
-	} else if runtime.GOOS != "windows" {
+	// Compose query parameters. User-provided ?_busy_timeout (or its mattn alias
+	// ?_timeout) overrides our default; otherwise inject 30s so SQLite waits at
+	// most that long on a lock instead of blocking the only Go-side connection.
+	// mattn/go-sqlite3 applies PRAGMA from the DSN on every fresh connection, so
+	// the value survives ConnMaxIdleTime/ConnMaxLifetime recycling. cache=shared
+	// stays the default on non-Windows for the same reason as before.
+	parsed, _ := url.ParseQuery(query)
+	var defaults []string
+	if parsed.Get("_busy_timeout") == "" && parsed.Get("_timeout") == "" {
+		defaults = append(defaults, "_busy_timeout=30000")
+	}
+	if !hasQuery && runtime.GOOS != "windows" {
 		// To avoid `The process cannot access the file because it is being used by another process` on Windows
-		connStr += "?cache=shared"
+		defaults = append(defaults, "cache=shared")
+	}
+	parts := defaults
+	if hasQuery {
+		parts = append(parts, query)
+	}
+	if len(parts) > 0 {
+		connStr += "?" + strings.Join(parts, "&")
 	}
 
 	db, err := gorm.Open(sqlite.Open(connStr), getGormConfig())
@@ -3377,6 +3463,49 @@ func (s *SqlStore) GetAllEphemeralPeers(ctx context.Context, lockStrength Lockin
 	return allEphemeralPeers, nil
 }
 
+// GetStaleEphemeralPeerIDsForAccount returns IDs of disconnected
+// ephemeral peers in the given account whose last_seen is strictly
+// older than olderThan.
+func (s *SqlStore) GetStaleEphemeralPeerIDsForAccount(ctx context.Context, accountID string, olderThan time.Time) ([]string, error) {
+	var ids []string
+	err := s.db.WithContext(ctx).
+		Model(&nbpeer.Peer{}).
+		Where("account_id = ? AND ephemeral = ? AND peer_status_connected = ? AND peer_status_last_seen < ?",
+			accountID, true, false, olderThan).
+		Pluck("id", &ids).Error
+	if err != nil {
+		log.WithContext(ctx).Errorf("failed to query stale ephemeral peers for account %s: %v", accountID, err)
+		return nil, status.Errorf(status.Internal, "query stale ephemeral peers")
+	}
+	return ids, nil
+}
+
+// GetEphemeralAccountsLastDisconnect returns the latest peer_status_last_seen
+// per account across disconnected ephemeral peers. Returns one entry per
+// account that has at least one such peer.
+func (s *SqlStore) GetEphemeralAccountsLastDisconnect(ctx context.Context) (map[string]time.Time, error) {
+	type row struct {
+		AccountID string
+		LastSeen  time.Time
+	}
+	var rows []row
+	err := s.db.WithContext(ctx).
+		Model(&nbpeer.Peer{}).
+		Select("account_id, MAX(peer_status_last_seen) AS last_seen").
+		Where("ephemeral = ? AND peer_status_connected = ?", true, false).
+		Group("account_id").
+		Scan(&rows).Error
+	if err != nil {
+		log.WithContext(ctx).Errorf("failed to load ephemeral-account last disconnect map: %v", err)
+		return nil, status.Errorf(status.Internal, "load ephemeral accounts")
+	}
+	out := make(map[string]time.Time, len(rows))
+	for _, r := range rows {
+		out[r.AccountID] = r.LastSeen
+	}
+	return out, nil
+}
+
 // DeletePeer removes a peer from the store.
 func (s *SqlStore) DeletePeer(ctx context.Context, accountID string, peerID string) error {
 	result := s.db.Delete(&nbpeer.Peer{}, accountAndIDQueryCondition, accountID, peerID)
@@ -3402,7 +3531,7 @@ func (s *SqlStore) IncrementNetworkSerial(ctx context.Context, accountId string)
 }
 
 func (s *SqlStore) ExecuteInTransaction(ctx context.Context, operation func(store Store) error) error {
-	timeoutCtx, cancel := context.WithTimeout(context.Background(), s.transactionTimeout)
+	timeoutCtx, cancel := context.WithTimeout(ctx, s.transactionTimeout)
 	defer cancel()
 
 	startTime := time.Now()

management/server/store/sql_store_test.go

@@ -4592,3 +4592,55 @@ func TestSqlStore_DeleteZoneDNSRecords(t *testing.T) {
 	require.NoError(t, err)
 	assert.Equal(t, 0, len(remainingRecords))
 }
+
+// TestNewSqliteStore_BusyTimeoutApplied opens a fresh SQLite store and verifies
+// that the _busy_timeout DSN parameter took effect at the driver level. Without
+// this, lock contention on the single SQLite connection waits indefinitely on
+// the Go side and can be hidden behind the 5-minute transactionTimeout.
+func TestNewSqliteStore_BusyTimeoutApplied(t *testing.T) {
+	dir := t.TempDir()
+	store, err := NewSqliteStore(context.Background(), dir, nil, true)
+	require.NoError(t, err)
+	t.Cleanup(func() {
+		_ = store.Close(context.Background())
+	})
+
+	sqlDB, err := store.db.DB()
+	require.NoError(t, err)
+	row := sqlDB.QueryRow("PRAGMA busy_timeout")
+	var busyTimeout int
+	require.NoError(t, row.Scan(&busyTimeout))
+	assert.Equal(t, 30000, busyTimeout, "SQLite busy_timeout must be set via DSN so it survives connection recycling")
+}
+
+// TestNewSqliteStore_BusyTimeoutRespectsUserOverride confirms that an operator
+// passing _busy_timeout or its mattn alias _timeout via NB_STORE_ENGINE_SQLITE_FILE
+// wins over our 30s default. This guards the DSN merge logic in NewSqliteStore.
+func TestNewSqliteStore_BusyTimeoutRespectsUserOverride(t *testing.T) {
+	cases := []struct {
+		name     string
+		envFile  string
+		expected int
+	}{
+		{name: "explicit _busy_timeout wins", envFile: "store.db?_busy_timeout=5000", expected: 5000},
+		{name: "alias _timeout wins", envFile: "store.db?_timeout=7000", expected: 7000},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Setenv("NB_STORE_ENGINE_SQLITE_FILE", tc.envFile)
+			dir := t.TempDir()
+			store, err := NewSqliteStore(context.Background(), dir, nil, true)
+			require.NoError(t, err)
+			t.Cleanup(func() {
+				_ = store.Close(context.Background())
+			})
+
+			sqlDB, err := store.db.DB()
+			require.NoError(t, err)
+			row := sqlDB.QueryRow("PRAGMA busy_timeout")
+			var busyTimeout int
+			require.NoError(t, row.Scan(&busyTimeout))
+			assert.Equal(t, tc.expected, busyTimeout)
+		})
+	}
+}

management/server/store/store.go

@@ -165,8 +165,32 @@ type Store interface {
 	GetAccountPeersWithExpiration(ctx context.Context, lockStrength LockingStrength, accountID string) ([]*nbpeer.Peer, error)
 	GetAccountPeersWithInactivity(ctx context.Context, lockStrength LockingStrength, accountID string) ([]*nbpeer.Peer, error)
 	GetAllEphemeralPeers(ctx context.Context, lockStrength LockingStrength) ([]*nbpeer.Peer, error)
+	// GetStaleEphemeralPeerIDsForAccount returns the IDs of disconnected
+	// ephemeral peers whose last_seen is strictly older than olderThan,
+	// scoped to a single account. Used by the per-account cleanup sweep.
+	GetStaleEphemeralPeerIDsForAccount(ctx context.Context, accountID string, olderThan time.Time) ([]string, error)
+	// GetEphemeralAccountsLastDisconnect returns, for every account that
+	// has at least one disconnected ephemeral peer, the most recent
+	// last_seen across that account's disconnected ephemeral peers. Used
+	// to reconstruct the per-account cleanup tracker after a restart.
+	GetEphemeralAccountsLastDisconnect(ctx context.Context) (map[string]time.Time, error)
 	SavePeer(ctx context.Context, accountID string, peer *nbpeer.Peer) error
 	SavePeerStatus(ctx context.Context, accountID, peerID string, status nbpeer.PeerStatus) error
+	// MarkPeerConnectedIfNewerSession sets the peer to connected with the
+	// given session token, but only when the stored SessionStartedAt is
+	// strictly less than newSessionStartedAt (the sentinel zero counts as
+	// "older"). LastSeen is recorded by the database at the moment the
+	// row is updated — never by the caller — so it always reflects the
+	// real write time even under lock contention.
+	// Returns true when the update happened, false when this stream lost
+	// the race against a newer session.
+	MarkPeerConnectedIfNewerSession(ctx context.Context, accountID, peerID string, newSessionStartedAt int64) (bool, error)
+	// MarkPeerDisconnectedIfSameSession sets the peer to disconnected and
+	// resets SessionStartedAt to zero, but only when the stored
+	// SessionStartedAt equals the given sessionStartedAt. LastSeen is
+	// recorded by the database. Returns true when the update happened,
+	// false when a newer session has taken over.
+	MarkPeerDisconnectedIfSameSession(ctx context.Context, accountID, peerID string, sessionStartedAt int64) (bool, error)
 	SavePeerLocation(ctx context.Context, accountID string, peer *nbpeer.Peer) error
 	ApproveAccountPeers(ctx context.Context, accountID string) (int, error)
 	DeletePeer(ctx context.Context, accountID string, peerID string) error

management/server/store/store_mock.go

@@ -1376,6 +1376,36 @@ func (mr *MockStoreMockRecorder) GetAllEphemeralPeers(ctx, lockStrength interfac
 	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "GetAllEphemeralPeers", reflect.TypeOf((*MockStore)(nil).GetAllEphemeralPeers), ctx, lockStrength)
 }
 
+// GetStaleEphemeralPeerIDsForAccount mocks base method.
+func (m *MockStore) GetStaleEphemeralPeerIDsForAccount(ctx context.Context, accountID string, olderThan time.Time) ([]string, error) {
+	m.ctrl.T.Helper()
+	ret := m.ctrl.Call(m, "GetStaleEphemeralPeerIDsForAccount", ctx, accountID, olderThan)
+	ret0, _ := ret[0].([]string)
+	ret1, _ := ret[1].(error)
+	return ret0, ret1
+}
+
+// GetStaleEphemeralPeerIDsForAccount indicates an expected call of GetStaleEphemeralPeerIDsForAccount.
+func (mr *MockStoreMockRecorder) GetStaleEphemeralPeerIDsForAccount(ctx, accountID, olderThan interface{}) *gomock.Call {
+	mr.mock.ctrl.T.Helper()
+	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "GetStaleEphemeralPeerIDsForAccount", reflect.TypeOf((*MockStore)(nil).GetStaleEphemeralPeerIDsForAccount), ctx, accountID, olderThan)
+}
+
+// GetEphemeralAccountsLastDisconnect mocks base method.
+func (m *MockStore) GetEphemeralAccountsLastDisconnect(ctx context.Context) (map[string]time.Time, error) {
+	m.ctrl.T.Helper()
+	ret := m.ctrl.Call(m, "GetEphemeralAccountsLastDisconnect", ctx)
+	ret0, _ := ret[0].(map[string]time.Time)
+	ret1, _ := ret[1].(error)
+	return ret0, ret1
+}
+
+// GetEphemeralAccountsLastDisconnect indicates an expected call of GetEphemeralAccountsLastDisconnect.
+func (mr *MockStoreMockRecorder) GetEphemeralAccountsLastDisconnect(ctx interface{}) *gomock.Call {
+	mr.mock.ctrl.T.Helper()
+	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "GetEphemeralAccountsLastDisconnect", reflect.TypeOf((*MockStore)(nil).GetEphemeralAccountsLastDisconnect), ctx)
+}
+
 // GetAllProxyAccessTokens mocks base method.
 func (m *MockStore) GetAllProxyAccessTokens(ctx context.Context, lockStrength LockingStrength) ([]*types2.ProxyAccessToken, error) {
 	m.ctrl.T.Helper()
@@ -2878,6 +2908,36 @@ func (mr *MockStoreMockRecorder) SavePeerStatus(ctx, accountID, peerID, status i
 	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "SavePeerStatus", reflect.TypeOf((*MockStore)(nil).SavePeerStatus), ctx, accountID, peerID, status)
 }
 
+// MarkPeerConnectedIfNewerSession mocks base method.
+func (m *MockStore) MarkPeerConnectedIfNewerSession(ctx context.Context, accountID, peerID string, newSessionStartedAt int64) (bool, error) {
+	m.ctrl.T.Helper()
+	ret := m.ctrl.Call(m, "MarkPeerConnectedIfNewerSession", ctx, accountID, peerID, newSessionStartedAt)
+	ret0, _ := ret[0].(bool)
+	ret1, _ := ret[1].(error)
+	return ret0, ret1
+}
+
+// MarkPeerConnectedIfNewerSession indicates an expected call of MarkPeerConnectedIfNewerSession.
+func (mr *MockStoreMockRecorder) MarkPeerConnectedIfNewerSession(ctx, accountID, peerID, newSessionStartedAt interface{}) *gomock.Call {
+	mr.mock.ctrl.T.Helper()
+	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "MarkPeerConnectedIfNewerSession", reflect.TypeOf((*MockStore)(nil).MarkPeerConnectedIfNewerSession), ctx, accountID, peerID, newSessionStartedAt)
+}
+
+// MarkPeerDisconnectedIfSameSession mocks base method.
+func (m *MockStore) MarkPeerDisconnectedIfSameSession(ctx context.Context, accountID, peerID string, sessionStartedAt int64) (bool, error) {
+	m.ctrl.T.Helper()
+	ret := m.ctrl.Call(m, "MarkPeerDisconnectedIfSameSession", ctx, accountID, peerID, sessionStartedAt)
+	ret0, _ := ret[0].(bool)
+	ret1, _ := ret[1].(error)
+	return ret0, ret1
+}
+
+// MarkPeerDisconnectedIfSameSession indicates an expected call of MarkPeerDisconnectedIfSameSession.
+func (mr *MockStoreMockRecorder) MarkPeerDisconnectedIfSameSession(ctx, accountID, peerID, sessionStartedAt interface{}) *gomock.Call {
+	mr.mock.ctrl.T.Helper()
+	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "MarkPeerDisconnectedIfSameSession", reflect.TypeOf((*MockStore)(nil).MarkPeerDisconnectedIfSameSession), ctx, accountID, peerID, sessionStartedAt)
+}
+
 // SavePolicy mocks base method.
 func (m *MockStore) SavePolicy(ctx context.Context, policy *types2.Policy) error {
 	m.ctrl.T.Helper()

management/server/telemetry/accountmanager_metrics.go

@@ -16,6 +16,8 @@ type AccountManagerMetrics struct {
 	getPeerNetworkMapDurationMs  metric.Float64Histogram
 	networkMapObjectCount        metric.Int64Histogram
 	peerMetaUpdateCount          metric.Int64Counter
+	peerStatusUpdateCounter      metric.Int64Counter
+	peerStatusUpdateDurationMs   metric.Float64Histogram
 }
 
 // NewAccountManagerMetrics creates an instance of AccountManagerMetrics
@@ -64,6 +66,24 @@ func NewAccountManagerMetrics(ctx context.Context, meter metric.Meter) (*Account
 		return nil, err
 	}
 
+	// peerStatusUpdateCounter records every attempt to mark a peer as connected or disconnected
+	peerStatusUpdateCounter, err := meter.Int64Counter("management.account.peer.status.update.counter",
+		metric.WithUnit("1"),
+		metric.WithDescription("Number of peer status update attempts, labeled by operation (connect|disconnect) and outcome (applied|stale|error|peer_not_found)"))
+	if err != nil {
+		return nil, err
+	}
+
+	peerStatusUpdateDurationMs, err := meter.Float64Histogram("management.account.peer.status.update.duration.ms",
+		metric.WithUnit("milliseconds"),
+		metric.WithExplicitBucketBoundaries(
+			1, 5, 15, 25, 50, 100, 250, 500, 1000, 2000, 5000,
+		),
+		metric.WithDescription("Duration of a peer status update (fence UPDATE + post-write side effects), labeled by operation"))
+	if err != nil {
+		return nil, err
+	}
+
 	return &AccountManagerMetrics{
 		ctx:                          ctx,
 		getPeerNetworkMapDurationMs:  getPeerNetworkMapDurationMs,
@@ -71,10 +91,35 @@ func NewAccountManagerMetrics(ctx context.Context, meter metric.Meter) (*Account
 		updateAccountPeersCounter:    updateAccountPeersCounter,
 		networkMapObjectCount:        networkMapObjectCount,
 		peerMetaUpdateCount:          peerMetaUpdateCount,
+		peerStatusUpdateCounter:      peerStatusUpdateCounter,
+		peerStatusUpdateDurationMs:   peerStatusUpdateDurationMs,
 	}, nil
 
 }
 
+// PeerStatusOperation labels the kind of fence-locked peer status write.
+type PeerStatusOperation string
+
+// PeerStatusOutcome labels how a fence-locked peer status write resolved.
+type PeerStatusOutcome string
+
+const (
+	PeerStatusConnect    PeerStatusOperation = "connect"
+	PeerStatusDisconnect PeerStatusOperation = "disconnect"
+
+	// PeerStatusApplied — the fence WHERE matched and the UPDATE landed.
+	PeerStatusApplied PeerStatusOutcome = "applied"
+	// PeerStatusStale — the fence WHERE rejected the write because a
+	// newer session has already taken ownership (connect: stored token
+	// >= incoming; disconnect: stored token != incoming).
+	PeerStatusStale PeerStatusOutcome = "stale"
+	// PeerStatusError — the store returned a non-NotFound error.
+	PeerStatusError PeerStatusOutcome = "error"
+	// PeerStatusPeerNotFound — the peer lookup failed (the peer was
+	// deleted between the gRPC sync handshake and the status write).
+	PeerStatusPeerNotFound PeerStatusOutcome = "peer_not_found"
+)
+
 // CountUpdateAccountPeersDuration counts the duration of updating account peers
 func (metrics *AccountManagerMetrics) CountUpdateAccountPeersDuration(duration time.Duration) {
 	metrics.updateAccountPeersDurationMs.Record(metrics.ctx, float64(duration.Nanoseconds())/1e6)
@@ -104,3 +149,23 @@ func (metrics *AccountManagerMetrics) CountUpdateAccountPeersTriggered(resource,
 func (metrics *AccountManagerMetrics) CountPeerMetUpdate() {
 	metrics.peerMetaUpdateCount.Add(metrics.ctx, 1)
 }
+
+// CountPeerStatusUpdate increments the connect/disconnect counter,
+// labeled by operation and outcome. Both labels are bounded enums.
+func (metrics *AccountManagerMetrics) CountPeerStatusUpdate(op PeerStatusOperation, outcome PeerStatusOutcome) {
+	metrics.peerStatusUpdateCounter.Add(metrics.ctx, 1,
+		metric.WithAttributes(
+			attribute.String("operation", string(op)),
+			attribute.String("outcome", string(outcome)),
+		),
+	)
+}
+
+// RecordPeerStatusUpdateDuration records the wall-clock time spent
+// running a peer status update (including post-write side effects),
+// labeled by operation.
+func (metrics *AccountManagerMetrics) RecordPeerStatusUpdateDuration(op PeerStatusOperation, d time.Duration) {
+	metrics.peerStatusUpdateDurationMs.Record(metrics.ctx, float64(d.Nanoseconds())/1e6,
+		metric.WithAttributes(attribute.String("operation", string(op))),
+	)
+}

management/server/telemetry/app_metrics.go

@@ -29,6 +29,7 @@ type MockAppMetrics struct {
 	StoreMetricsFunc             func() *StoreMetrics
 	UpdateChannelMetricsFunc     func() *UpdateChannelMetrics
 	AddAccountManagerMetricsFunc func() *AccountManagerMetrics
+	EphemeralPeersMetricsFunc    func() *EphemeralPeersMetrics
 }
 
 // GetMeter mocks the GetMeter function of the AppMetrics interface
@@ -103,6 +104,14 @@ func (mock *MockAppMetrics) AccountManagerMetrics() *AccountManagerMetrics {
 	return nil
 }
 
+// EphemeralPeersMetrics mocks the MockAppMetrics function of the EphemeralPeersMetrics interface
+func (mock *MockAppMetrics) EphemeralPeersMetrics() *EphemeralPeersMetrics {
+	if mock.EphemeralPeersMetricsFunc != nil {
+		return mock.EphemeralPeersMetricsFunc()
+	}
+	return nil
+}
+
 // AppMetrics is metrics interface
 type AppMetrics interface {
 	GetMeter() metric2.Meter
@@ -114,6 +123,7 @@ type AppMetrics interface {
 	StoreMetrics() *StoreMetrics
 	UpdateChannelMetrics() *UpdateChannelMetrics
 	AccountManagerMetrics() *AccountManagerMetrics
+	EphemeralPeersMetrics() *EphemeralPeersMetrics
 }
 
 // defaultAppMetrics are core application metrics based on OpenTelemetry https://opentelemetry.io/
@@ -129,6 +139,7 @@ type defaultAppMetrics struct {
 	storeMetrics          *StoreMetrics
 	updateChannelMetrics  *UpdateChannelMetrics
 	accountManagerMetrics *AccountManagerMetrics
+	ephemeralMetrics      *EphemeralPeersMetrics
 }
 
 // IDPMetrics returns metrics for the idp package
@@ -161,6 +172,11 @@ func (appMetrics *defaultAppMetrics) AccountManagerMetrics() *AccountManagerMetr
 	return appMetrics.accountManagerMetrics
 }
 
+// EphemeralPeersMetrics returns metrics for the ephemeral peer cleanup loop
+func (appMetrics *defaultAppMetrics) EphemeralPeersMetrics() *EphemeralPeersMetrics {
+	return appMetrics.ephemeralMetrics
+}
+
 // Close stop application metrics HTTP handler and closes listener.
 func (appMetrics *defaultAppMetrics) Close() error {
 	if appMetrics.listener == nil {
@@ -245,6 +261,11 @@ func NewDefaultAppMetrics(ctx context.Context) (AppMetrics, error) {
 		return nil, fmt.Errorf("failed to initialize account manager metrics: %w", err)
 	}
 
+	ephemeralMetrics, err := NewEphemeralPeersMetrics(ctx, meter)
+	if err != nil {
+		return nil, fmt.Errorf("failed to initialize ephemeral peers metrics: %w", err)
+	}
+
 	return &defaultAppMetrics{
 		Meter:                 meter,
 		ctx:                   ctx,
@@ -254,6 +275,7 @@ func NewDefaultAppMetrics(ctx context.Context) (AppMetrics, error) {
 		storeMetrics:          storeMetrics,
 		updateChannelMetrics:  updateChannelMetrics,
 		accountManagerMetrics: accountManagerMetrics,
+		ephemeralMetrics:      ephemeralMetrics,
 	}, nil
 }
 
@@ -290,6 +312,11 @@ func NewAppMetricsWithMeter(ctx context.Context, meter metric2.Meter) (AppMetric
 		return nil, fmt.Errorf("failed to initialize account manager metrics: %w", err)
 	}
 
+	ephemeralMetrics, err := NewEphemeralPeersMetrics(ctx, meter)
+	if err != nil {
+		return nil, fmt.Errorf("failed to initialize ephemeral peers metrics: %w", err)
+	}
+
 	return &defaultAppMetrics{
 		Meter:                 meter,
 		ctx:                   ctx,
@@ -300,5 +327,6 @@ func NewAppMetricsWithMeter(ctx context.Context, meter metric2.Meter) (AppMetric
 		storeMetrics:          storeMetrics,
 		updateChannelMetrics:  updateChannelMetrics,
 		accountManagerMetrics: accountManagerMetrics,
+		ephemeralMetrics:      ephemeralMetrics,
 	}, nil
 }

management/server/telemetry/ephemeral_metrics.go

@@ -0,0 +1,115 @@
+package telemetry
+
+import (
+	"context"
+
+	"go.opentelemetry.io/otel/metric"
+)
+
+// EphemeralPeersMetrics tracks the ephemeral peer cleanup pipeline: how
+// many accounts are currently being tracked for cleanup, how many sweep
+// runs deleted at least one peer, how many peers have been removed, and
+// how many delete batches failed.
+type EphemeralPeersMetrics struct {
+	ctx context.Context
+
+	pending      metric.Int64UpDownCounter
+	cleanupRuns  metric.Int64Counter
+	peersCleaned metric.Int64Counter
+	errors       metric.Int64Counter
+}
+
+// NewEphemeralPeersMetrics constructs the ephemeral cleanup counters.
+func NewEphemeralPeersMetrics(ctx context.Context, meter metric.Meter) (*EphemeralPeersMetrics, error) {
+	pending, err := meter.Int64UpDownCounter("management.ephemeral.accounts.tracked",
+		metric.WithUnit("1"),
+		metric.WithDescription("Number of accounts currently tracked for ephemeral peer cleanup"))
+	if err != nil {
+		return nil, err
+	}
+
+	cleanupRuns, err := meter.Int64Counter("management.ephemeral.cleanup.runs.counter",
+		metric.WithUnit("1"),
+		metric.WithDescription("Number of ephemeral cleanup sweeps that deleted at least one peer"))
+	if err != nil {
+		return nil, err
+	}
+
+	peersCleaned, err := meter.Int64Counter("management.ephemeral.peers.cleaned.counter",
+		metric.WithUnit("1"),
+		metric.WithDescription("Total number of ephemeral peers deleted by the cleanup loop"))
+	if err != nil {
+		return nil, err
+	}
+
+	errors, err := meter.Int64Counter("management.ephemeral.cleanup.errors.counter",
+		metric.WithUnit("1"),
+		metric.WithDescription("Number of ephemeral cleanup batches (per account) that failed to delete"))
+	if err != nil {
+		return nil, err
+	}
+
+	return &EphemeralPeersMetrics{
+		ctx:          ctx,
+		pending:      pending,
+		cleanupRuns:  cleanupRuns,
+		peersCleaned: peersCleaned,
+		errors:       errors,
+	}, nil
+}
+
+// All methods are nil-receiver safe so callers that haven't wired metrics
+// (tests, self-hosted with metrics off) can invoke them unconditionally.
+
+// IncPending bumps the tracked-accounts gauge when a new account
+// becomes eligible for ephemeral cleanup tracking.
+func (m *EphemeralPeersMetrics) IncPending() {
+	if m == nil {
+		return
+	}
+	m.pending.Add(m.ctx, 1)
+}
+
+// AddPending bumps the tracked-accounts gauge by n — used at startup
+// when the catch-up query seeds the tracker.
+func (m *EphemeralPeersMetrics) AddPending(n int64) {
+	if m == nil || n <= 0 {
+		return
+	}
+	m.pending.Add(m.ctx, n)
+}
+
+// DecPending decreases the tracked-accounts gauge when an account is
+// dropped from the tracker (no more disconnects to chase).
+func (m *EphemeralPeersMetrics) DecPending(n int64) {
+	if m == nil || n <= 0 {
+		return
+	}
+	m.pending.Add(m.ctx, -n)
+}
+
+// CountCleanupRun records one cleanup pass that processed >0 peers. Idle
+// ticks (nothing to do) deliberately don't increment so the rate
+// reflects useful work.
+func (m *EphemeralPeersMetrics) CountCleanupRun() {
+	if m == nil {
+		return
+	}
+	m.cleanupRuns.Add(m.ctx, 1)
+}
+
+// CountPeersCleaned records the number of peers a single tick deleted.
+func (m *EphemeralPeersMetrics) CountPeersCleaned(n int64) {
+	if m == nil || n <= 0 {
+		return
+	}
+	m.peersCleaned.Add(m.ctx, n)
+}
+
+// CountCleanupError records a failed delete batch.
+func (m *EphemeralPeersMetrics) CountCleanupError() {
+	if m == nil {
+		return
+	}
+	m.errors.Add(m.ctx, 1)
+}

management/server/types/account.go

@@ -598,28 +598,21 @@ func (a *Account) GetPeerGroups(peerID string) LookupMap {
 	return groupList
 }
 
-// PeerIPv6Allowed reports whether the given peer is in any of the account's IPv6 enabled groups.
+// PeerIPv6Allowed reports whether the given peer participates in the IPv6 overlay.
 // Returns false if IPv6 is disabled or no groups are configured.
 func (a *Account) PeerIPv6Allowed(peerID string) bool {
-	if len(a.Settings.IPv6EnabledGroups) == 0 {
-		return false
-	}
-
-	for _, groupID := range a.Settings.IPv6EnabledGroups {
-		group, ok := a.Groups[groupID]
-		if !ok {
-			continue
-		}
-		if slices.Contains(group.Peers, peerID) {
-			return true
-		}
-	}
-	return false
+	_, ok := a.peerIPv6AllowedSet()[peerID]
+	return ok
 }
 
-// peerIPv6AllowedSet returns a set of peer IDs that belong to any IPv6-enabled group.
+// peerIPv6AllowedSet returns the set of peer IDs that participate in the IPv6 overlay:
+// members of any IPv6-enabled group, plus every embedded proxy peer (which sit outside
+// regular group membership but must reach v6-enabled peers).
 func (a *Account) peerIPv6AllowedSet() map[string]struct{} {
 	result := make(map[string]struct{})
+	if len(a.Settings.IPv6EnabledGroups) == 0 {
+		return result
+	}
 	for _, groupID := range a.Settings.IPv6EnabledGroups {
 		group, ok := a.Groups[groupID]
 		if !ok {
@@ -629,6 +622,11 @@ func (a *Account) peerIPv6AllowedSet() map[string]struct{} {
 			result[peerID] = struct{}{}
 		}
 	}
+	for id, p := range a.Peers {
+		if p != nil && p.ProxyMeta.Embedded {
+			result[id] = struct{}{}
+		}
+	}
 	return result
 }
 

management/server/types/group.go

@@ -92,9 +92,12 @@ func (g *Group) HasPeers() bool {
 	return len(g.Peers) > 0
 }
 
+// GroupAllName is the reserved name of the default group that contains every peer in an account.
+const GroupAllName = "All"
+
 // IsGroupAll checks if the group is a default "All" group.
 func (g *Group) IsGroupAll() bool {
-	return g.Name == "All"
+	return g.Name == GroupAllName
 }
 
 // AddPeer adds peerID to Peers if not present, returning true if added.

management/server/types/ipv6_groups_test.go

@@ -232,3 +232,33 @@ func TestIPv6RecalculationOnGroupChange(t *testing.T) {
 		assert.True(t, account.PeerIPv6Allowed("peer3"), "peer3 now in infra")
 	})
 }
+
+func TestPeerIPv6AllowedEmbeddedProxy(t *testing.T) {
+	account := &Account{
+		Peers: map[string]*nbpeer.Peer{
+			"peer1": {ID: "peer1"},
+			"proxy": {ID: "proxy", ProxyMeta: nbpeer.ProxyMeta{Embedded: true, Cluster: "netbird.test"}},
+		},
+		Groups: map[string]*Group{
+			"group-devs": {ID: "group-devs", Peers: []string{"peer1"}},
+		},
+		Settings: &Settings{},
+	}
+
+	t.Run("embedded proxy allowed when any v6 group exists, without group membership", func(t *testing.T) {
+		account.Settings.IPv6EnabledGroups = []string{"group-devs"}
+		assert.True(t, account.PeerIPv6Allowed("proxy"), "embedded proxy participates in v6 overlay")
+		assert.True(t, account.PeerIPv6Allowed("peer1"), "regular peer in enabled group still allowed")
+	})
+
+	t.Run("embedded proxy denied when no v6 group enabled", func(t *testing.T) {
+		account.Settings.IPv6EnabledGroups = nil
+		assert.False(t, account.PeerIPv6Allowed("proxy"), "v6 disabled account-wide denies embedded proxies too")
+	})
+
+	t.Run("non-embedded peer outside any enabled group is not pulled in", func(t *testing.T) {
+		account.Settings.IPv6EnabledGroups = []string{"group-devs"}
+		account.Peers["lonely"] = &nbpeer.Peer{ID: "lonely"}
+		assert.False(t, account.PeerIPv6Allowed("lonely"), "embedded-proxy bypass must not leak to regular peers")
+	})
+}