diff --git a/cmd/cli/dns_intercept_darwin.go b/cmd/cli/dns_intercept_darwin.go
index 7b72df9c..364a29ad 100644
--- a/cmd/cli/dns_intercept_darwin.go
+++ b/cmd/cli/dns_intercept_darwin.go
@@ -41,6 +41,11 @@ const (
 	// 2s re-check misses.
 	pfAnchorRecheckDelayLong = 4 * time.Second
 
+	// pfExecFailureBackoff suppresses repeated pfctl/scutil checks briefly after
+	// macOS reports local resource exhaustion. Without this, network-change storms
+	// can turn a pf ruleset race into a fork/file-descriptor exhaustion loop.
+	pfExecFailureBackoff = 5 * time.Second
+
 	// pfVPNInterfacePrefixes lists interface name prefixes that indicate VPN/tunnel
 	// interfaces on macOS. Used to add interface-specific DNS intercept rules so that
 	// VPN software with "pass out quick on <iface>" rules cannot bypass our intercept.
@@ -1303,6 +1308,15 @@ func (p *prog) ensurePFAnchorActive() bool {
 	if p.dnsInterceptState == nil {
 		return false
 	}
+	if !p.pfEnsureRunning.CompareAndSwap(false, true) {
+		mainLog.Load().Debug().Msg("DNS intercept watchdog: check already running, skipping duplicate")
+		return false
+	}
+	defer p.pfEnsureRunning.Store(false)
+
+	if p.pfExecBackoffActive() {
+		return false
+	}
 
 	// While stabilizing (VPN connecting), suppress all restores.
 	// The stabilization loop will restore once pf settles.
@@ -1336,6 +1350,9 @@ func (p *prog) ensurePFAnchorActive() bool {
 	// Check 1: anchor references in the main ruleset.
 	natOut, err := exec.Command("pfctl", "-sn").CombinedOutput()
 	if err != nil {
+		if p.pfBackoffResourceExhaustion(err, natOut, "dump NAT rules") {
+			return false
+		}
 		mainLog.Load().Warn().Err(err).Msg("DNS intercept watchdog: could not dump NAT rules")
 		return false
 	}
@@ -1348,6 +1365,9 @@ func (p *prog) ensurePFAnchorActive() bool {
 	if !needsRestore {
 		filterOut, err := exec.Command("pfctl", "-sr").CombinedOutput()
 		if err != nil {
+			if p.pfBackoffResourceExhaustion(err, filterOut, "dump filter rules") {
+				return false
+			}
 			mainLog.Load().Warn().Err(err).Msg("DNS intercept watchdog: could not dump filter rules")
 			return false
 		}
@@ -1365,6 +1385,9 @@ func (p *prog) ensurePFAnchorActive() bool {
 	if !needsRestore {
 		anchorFilter, err := exec.Command("pfctl", "-a", pfAnchorName, "-sr").CombinedOutput()
 		if err != nil || len(strings.TrimSpace(string(anchorFilter))) == 0 {
+			if p.pfBackoffResourceExhaustion(err, anchorFilter, "dump anchor filter rules") {
+				return false
+			}
 			mainLog.Load().Warn().Msg("DNS intercept watchdog: anchor has no filter rules — content was flushed")
 			needsRestore = true
 		}
@@ -1372,6 +1395,9 @@ func (p *prog) ensurePFAnchorActive() bool {
 	if !needsRestore {
 		anchorNat, err := exec.Command("pfctl", "-a", pfAnchorName, "-sn").CombinedOutput()
 		if err != nil || len(strings.TrimSpace(string(anchorNat))) == 0 {
+			if p.pfBackoffResourceExhaustion(err, anchorNat, "dump anchor NAT rules") {
+				return false
+			}
 			mainLog.Load().Warn().Msg("DNS intercept watchdog: anchor has no rdr rules — translation was flushed (will cause packet loop on lo0)")
 			needsRestore = true
 		}
@@ -1405,6 +1431,7 @@ func (p *prog) ensurePFAnchorActive() bool {
 	// Restore: re-inject anchor references into the main ruleset.
 	mainLog.Load().Info().Msg("DNS intercept watchdog: restoring pf anchor references")
 	if err := p.ensurePFAnchorReference(); err != nil {
+		p.pfBackoffResourceExhaustion(err, nil, "restore anchor references")
 		mainLog.Load().Error().Err(err).Msg("DNS intercept watchdog: failed to restore anchor references")
 		return true
 	}
@@ -1421,6 +1448,7 @@ func (p *prog) ensurePFAnchorActive() bool {
 	if err := os.WriteFile(pfAnchorFile, []byte(rulesStr), 0644); err != nil {
 		mainLog.Load().Error().Err(err).Msg("DNS intercept watchdog: failed to write anchor file")
 	} else if out, err := exec.Command("pfctl", "-a", pfAnchorName, "-f", pfAnchorFile).CombinedOutput(); err != nil {
+		p.pfBackoffResourceExhaustion(err, out, "load rebuilt anchor")
 		mainLog.Load().Error().Err(err).Msgf("DNS intercept watchdog: failed to load rebuilt anchor (output: %s)", strings.TrimSpace(string(out)))
 	} else {
 		flushPFStates()
@@ -1458,6 +1486,49 @@ func (p *prog) scheduleDNSAfterVPNSettleRefresh(reason string, delay time.Durati
 	})
 }
 
+func (p *prog) pfExecBackoffActive() bool {
+	until := p.pfExecBackoffUntil.Load()
+	if until == 0 {
+		return false
+	}
+	remaining := time.Until(time.UnixMilli(until))
+	if remaining <= 0 {
+		p.pfExecBackoffUntil.CompareAndSwap(until, 0)
+		return false
+	}
+	mainLog.Load().Debug().Dur("remaining", remaining).Msg("DNS intercept watchdog: suppressed during pf exec backoff")
+	return true
+}
+
+func (p *prog) pfBackoffResourceExhaustion(err error, output []byte, operation string) bool {
+	if !isResourceExhaustion(err, output) {
+		return false
+	}
+	until := time.Now().Add(pfExecFailureBackoff)
+	p.pfExecBackoffUntil.Store(until.UnixMilli())
+	mainLog.Load().Warn().Err(err).Dur("backoff", pfExecFailureBackoff).Str("operation", operation).
+		Msg("DNS intercept watchdog: backing off after local exec resource exhaustion")
+	return true
+}
+
+func isResourceExhaustion(err error, output []byte) bool {
+	if err == nil && len(output) == 0 {
+		return false
+	}
+	var msg string
+	if err != nil {
+		msg = err.Error()
+	}
+	if len(output) > 0 {
+		msg += "\n" + string(output)
+	}
+	msg = strings.ToLower(msg)
+	return strings.Contains(msg, "resource temporarily unavailable") ||
+		strings.Contains(msg, "too many open files") ||
+		strings.Contains(msg, "too many processes") ||
+		strings.Contains(msg, "cannot allocate memory")
+}
+
 // pfWatchdog periodically checks that our pf anchor is still active.
 // Other programs (e.g., Windscribe desktop app, macOS configd) can replace
 // scheduleDelayedRechecks schedules delayed re-checks after a network change event.
@@ -1470,8 +1541,19 @@ func (p *prog) scheduleDNSAfterVPNSettleRefresh(reason string, delay time.Durati
 //
 // Two delays (2s and 4s) cover both fast and slow VPN teardowns.
 func (p *prog) scheduleDelayedRechecks() {
+	p.pfDelayedRecheckMu.Lock()
+	defer p.pfDelayedRecheckMu.Unlock()
+
+	for _, timer := range p.pfDelayedRecheckTimers {
+		if timer != nil {
+			timer.Stop()
+		}
+	}
+	p.pfDelayedRecheckTimers = p.pfDelayedRecheckTimers[:0]
+
 	for _, delay := range []time.Duration{pfAnchorRecheckDelay, pfAnchorRecheckDelayLong} {
-		time.AfterFunc(delay, func() {
+		delay := delay
+		timer := time.AfterFunc(delay, func() {
 			if p.dnsInterceptState == nil || p.pfStabilizing.Load() {
 				return
 			}
@@ -1485,6 +1567,7 @@ func (p *prog) scheduleDelayedRechecks() {
 				p.vpnDNS.Refresh(true)
 			}
 		})
+		p.pfDelayedRecheckTimers = append(p.pfDelayedRecheckTimers, timer)
 	}
 }
 
diff --git a/cmd/cli/dns_intercept_darwin_test.go b/cmd/cli/dns_intercept_darwin_test.go
index d0834d7f..c8d90b03 100644
--- a/cmd/cli/dns_intercept_darwin_test.go
+++ b/cmd/cli/dns_intercept_darwin_test.go
@@ -3,6 +3,7 @@
 package cli
 
 import (
+	"errors"
 	"strings"
 	"testing"
 
@@ -141,3 +142,47 @@ func TestPFAddressFamily(t *testing.T) {
 		}
 	}
 }
+
+func TestIsResourceExhaustion(t *testing.T) {
+	tests := []struct {
+		name   string
+		err    error
+		output []byte
+		want   bool
+	}{
+		{
+			name: "exec start failure",
+			err:  errors.New("fork/exec /sbin/pfctl: resource temporarily unavailable"),
+			want: true,
+		},
+		{
+			name:   "fd exhaustion from stderr output",
+			err:    errors.New("exit status 1"),
+			output: []byte("pfctl: Pipe: Too many open files"),
+			want:   true,
+		},
+		{
+			name: "process exhaustion from wrapped restore error",
+			err:  errors.New("failed to dump running filter rules: exit status 1 (output: too many processes)"),
+			want: true,
+		},
+		{
+			name:   "ordinary pf syntax failure",
+			err:    errors.New("exit status 1"),
+			output: []byte("pfctl: syntax error"),
+			want:   false,
+		},
+		{
+			name: "nil error and empty output",
+			want: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := isResourceExhaustion(tt.err, tt.output); got != tt.want {
+				t.Fatalf("isResourceExhaustion() = %v, want %v", got, tt.want)
+			}
+		})
+	}
+}
diff --git a/cmd/cli/dns_intercept_windows.go b/cmd/cli/dns_intercept_windows.go
index a79fa0f6..5c3ecb34 100644
--- a/cmd/cli/dns_intercept_windows.go
+++ b/cmd/cli/dns_intercept_windows.go
@@ -278,6 +278,9 @@ type wfpState struct {
 	loopbackProtectActive bool
 	// loopbackPermitIDs stores the filter IDs for the loopback protect permits.
 	loopbackPermitIDs []uint64
+	// nrptRecoveryLimiter prevents repeated Windows policy/Dnscache signaling
+	// when another agent keeps putting NRPT back into a broken state.
+	nrptRecoveryLimiter nrptRecoveryLimiter
 }
 
 // Lazy-loaded WFP DLL procedures.
@@ -343,9 +346,10 @@ const (
 //   - GP path: SOFTWARE\Policies\...\DnsPolicyConfig (Group Policy)
 //   - Local path: SYSTEM\CurrentControlSet\...\DnsPolicyConfig (service store)
 //
-// If ANY rules exist in the GP path (from IT policy, VPN, MDM, etc.), DNS Client
-// enters "GP mode" and ignores ALL local-path rules entirely. Conversely, if the
-// GP path is empty/absent, DNS Client reads from the local path only.
+// If the GP path contains real rules (from IT policy, VPN, MDM, etc.), DNS
+// Client enters "GP mode" and ignores ALL local-path rules entirely. An empty GP
+// parent key is worse: it still puts DNS Client in GP mode, but contributes no
+// usable rule, so our local catch-all is hidden until that empty parent is gone.
 //
 // Strategy (matching Tailscale's approach):
 //   - Always write to the local path (baseline for non-domain machines).
@@ -395,18 +399,20 @@ func otherGPRulesExist() bool {
 	return false
 }
 
-// cleanGPPath removes our CtrldCatchAll rule from the GP path and deletes
-// the GP DnsPolicyConfig parent key if no other rules remain. Removing the
-// empty GP key is critical: its mere existence forces DNS Client into "GP mode"
-// where local-path rules are ignored.
-func cleanGPPath() {
+// cleanGPPath removes only ctrld's GP-path rule and deletes the GP parent when
+// no rules remain. The return value tells callers whether the parent key was
+// actually deleted, which means DNS Client should be signaled once.
+//
+// Do not leave an empty GP parent behind: Windows treats the parent key itself
+// as the policy store boundary, so an empty key can still hide local-path rules.
+func cleanGPPath() bool {
 	// Delete our specific rule.
 	registry.DeleteKey(registry.LOCAL_MACHINE, nrptBaseKey+`\`+nrptRuleName)
 
 	// If the GP parent key is now empty, delete it entirely to exit "GP mode".
 	k, err := registry.OpenKey(registry.LOCAL_MACHINE, nrptBaseKey, registry.ENUMERATE_SUB_KEYS)
 	if err != nil {
-		return // Key doesn't exist — clean state.
+		return false // Key doesn't exist — clean state.
 	}
 	names, err := k.ReadSubKeyNames(-1)
 	k.Close()
@@ -414,12 +420,14 @@ func cleanGPPath() {
 		if len(names) > 0 {
 			mainLog.Load().Debug().Strs("remaining", names).Msg("DNS intercept: GP path has other rules, leaving parent key")
 		}
-		return
+		return false
 	}
 	// Empty — delete it to exit "GP mode".
 	if err := registry.DeleteKey(registry.LOCAL_MACHINE, nrptBaseKey); err == nil {
 		mainLog.Load().Info().Msg("DNS intercept: deleted empty GP DnsPolicyConfig key (exits GP mode)")
+		return true
 	}
+	return false
 }
 
 // writeNRPTRule writes a single NRPT catch-all rule at the given registry keyPath.
@@ -542,10 +550,11 @@ func refreshNRPTPolicy() {
 // Group Policy refresh so NRPT changes take effect immediately.
 // Uses DnsFlushResolverCache from dnsapi.dll + RefreshPolicyEx from userenv.dll.
 func flushDNSCache() {
-	// Step 1: Refresh GP so DNS Client loads the new NRPT rules from registry.
 	refreshNRPTPolicy()
+	flushDNSCacheOnly()
+}
 
-	// Step 2: Flush the DNS cache so stale entries from pre-NRPT resolution are cleared.
+func flushDNSCacheOnly() {
 	if err := dnsapiDLL.Load(); err == nil {
 		if err := procDnsFlushResolverCache.Find(); err == nil {
 			ret, _, _ := procDnsFlushResolverCache.Call()
@@ -555,7 +564,6 @@ func flushDNSCache() {
 			}
 		}
 	}
-	// Fallback: use ipconfig /flushdns.
 	if out, err := exec.Command("ipconfig", "/flushdns").CombinedOutput(); err != nil {
 		mainLog.Load().Debug().Msgf("DNS intercept: ipconfig /flushdns failed: %v: %s", err, string(out))
 	} else {
@@ -563,6 +571,12 @@ func flushDNSCache() {
 	}
 }
 
+func signalNRPTChange() {
+	refreshNRPTPolicy()
+	sendParamChange()
+	flushDNSCacheOnly()
+}
+
 // startDNSIntercept activates WFP-based DNS interception on Windows.
 // It creates a WFP sublayer and adds filters that block all outbound DNS (port 53)
 // traffic except to localhost (127.0.0.1/::1), ensuring all DNS queries must go
@@ -598,21 +612,19 @@ func (p *prog) startDNSIntercept() error {
 	mainLog.Load().Info().Msgf("DNS intercept: initializing (mode: %s)", interceptMode)
 	logNRPTParentKeyState("pre-write")
 
-	// Two-phase empty parent key recovery: if the GP DnsPolicyConfig key exists
-	// but is empty, DNS Client has cached a "no rules" state and won't accept
-	// new rules even after they're written. Delete the empty key and signal DNS
-	// Client to reset before writing our rule.
-	// Two-phase recovery handles its own 2s signaling burst internally.
-	cleanEmptyNRPTParent()
+	// Empty parent key recovery: if the GP DnsPolicyConfig key exists but is
+	// empty, DNS Client enters GP mode and hides local rules. Delete empty
+	// parents first, then send one change signal so DNS Client drops stale state.
+	if cleanEmptyNRPTParent() {
+		signalNRPTChange()
+	}
 
 	if err := addNRPTCatchAllRule(listenerIP); err != nil {
 		return fmt.Errorf("dns intercept: failed to add NRPT catch-all rule: %w", err)
 	}
 	logNRPTParentKeyState("post-write")
 	state.nrptActive = true
-	refreshNRPTPolicy()
-	sendParamChange()
-	flushDNSCache()
+	signalNRPTChange()
 	mainLog.Load().Info().Msgf("DNS intercept: NRPT catch-all rule active — all DNS queries directed to %s", listenerIP)
 
 	// Step 2: In hard mode, also set up WFP filters to block non-local DNS.
@@ -1555,7 +1567,7 @@ func (p *prog) scheduleDelayedRechecks() {
 					mainLog.Load().Error().Err(err).Msg("DNS intercept: failed to re-add NRPT catch-all rule")
 					state.nrptActive = false
 				} else {
-					flushDNSCache()
+					signalNRPTChange()
 					mainLog.Load().Info().Msg("DNS intercept: NRPT catch-all rule restored")
 				}
 			}
@@ -1593,14 +1605,22 @@ func (p *prog) nrptHealthMonitor(state *wfpState) {
 
 			// Step 1: Check registry key exists.
 			if !nrptCatchAllRuleExists() {
+				now := time.Now()
+				if ok, wait := state.nrptRecoveryLimiter.allow(now, p.cfg); !ok {
+					if state.nrptRecoveryLimiter.shouldLogSkip(now) {
+						mainLog.Load().Warn().Dur("remaining", wait).
+							Msg("DNS intercept: NRPT rule restore suppressed after repeated recovery flows")
+					}
+					continue
+				}
 				mainLog.Load().Warn().Msg("DNS intercept: NRPT health check — catch-all rule missing, restoring")
 				if err := addNRPTCatchAllRule(state.listenerIP); err != nil {
 					mainLog.Load().Error().Err(err).Msg("DNS intercept: failed to restore NRPT catch-all rule")
 					state.nrptActive = false
 					continue
 				}
-				refreshNRPTPolicy()
-				flushDNSCache()
+				signalNRPTChange()
+				state.nrptRecoveryLimiter.recordRecoveryFlow(time.Now(), p.cfg)
 				mainLog.Load().Info().Msg("DNS intercept: NRPT catch-all rule restored by health monitor")
 				// After restoring, verify it's actually working.
 				go p.nrptProbeAndHeal()
@@ -1612,6 +1632,8 @@ func (p *prog) nrptHealthMonitor(state *wfpState) {
 			if !p.probeNRPT() {
 				mainLog.Load().Warn().Msg("DNS intercept: NRPT health check — rule present but probe failed, running heal cycle")
 				go p.nrptProbeAndHeal()
+			} else {
+				state.nrptRecoveryLimiter.recordStableSuccess()
 			}
 
 			// Step 3: In hard mode, also verify WFP sublayer.
@@ -1710,43 +1732,39 @@ func sendParamChange() {
 }
 
 // cleanEmptyNRPTParent removes empty NRPT parent keys that block activation.
-// An empty DnsPolicyConfig key (exists but no subkeys) causes DNS Client to
-// cache "no rules" and ignore subsequently-added rules.
+// Empty GP and local parents have different failure shapes:
+//   - empty GP parent: DNS Client is in GP mode and ignores local-path rules;
+//   - empty local parent: DNS Client can cache an empty local policy store.
 //
-// Also cleans the GP path entirely if it has no non-ctrld rules, since the GP
-// path's existence forces DNS Client into "GP mode" where it ignores the local
-// service store path.
+// This helper only changes registry state. The caller sends the single
+// RefreshPolicyEx/paramchange/flush signal after it knows cleanup occurred.
 //
-// Returns true if cleanup was performed (caller should add a delay).
+// Returns true if cleanup was performed (caller should signal DNS Client).
 func cleanEmptyNRPTParent() bool {
-	cleaned := false
-
 	// Always clean the GP path — its existence blocks local path activation.
-	cleanGPPath()
+	cleaned := cleanGPPath()
 
 	// Clean empty local/direct path parent key.
-	k, err := registry.OpenKey(registry.LOCAL_MACHINE, nrptDirectKey, registry.ENUMERATE_SUB_KEYS)
-	if err != nil {
-		return false
-	}
-	names, err := k.ReadSubKeyNames(-1)
-	k.Close()
-	if err != nil || len(names) > 0 {
-		return false
+	if !nrptParentKeyEmpty(nrptDirectKey) {
+		return cleaned
 	}
 
 	mainLog.Load().Warn().Msg("DNS intercept: found empty NRPT local parent key (blocks activation) — removing")
 	if err := registry.DeleteKey(registry.LOCAL_MACHINE, nrptDirectKey); err != nil {
 		mainLog.Load().Warn().Err(err).Msg("DNS intercept: failed to delete empty NRPT local parent key")
-		return false
+		return cleaned
 	}
-	cleaned = true
+	return true
+}
 
-	// Signal DNS Client to process the deletion and reset its internal cache.
-	mainLog.Load().Info().Msg("DNS intercept: empty NRPT parent key removed — signaling DNS Client")
-	sendParamChange()
-	flushDNSCache()
-	return cleaned
+func nrptParentKeyEmpty(keyPath string) bool {
+	k, err := registry.OpenKey(registry.LOCAL_MACHINE, keyPath, registry.ENUMERATE_SUB_KEYS)
+	if err != nil {
+		return false
+	}
+	names, err := k.ReadSubKeyNames(-1)
+	k.Close()
+	return err == nil && len(names) == 0
 }
 
 // logNRPTParentKeyState logs the state of both NRPT registry paths for diagnostics.
@@ -1783,18 +1801,39 @@ func logNRPTParentKeyState(context string) {
 // nrptProbeAndHeal runs the NRPT probe with retries and escalating remediation.
 // Called asynchronously after startup and from the health monitor.
 //
-// Retry sequence (each attempt: GP refresh + paramchange + flush → sleep → probe):
-//  1. Immediate probe
-//  2. GP refresh + paramchange + flush → 1s → probe
-//  3. GP refresh + paramchange + flush → 2s → probe
-//  4. GP refresh + paramchange + flush → 4s → probe
+// Retry sequence:
+//  1. Immediate probe.
+//  2. If the GP parent is empty, clean it immediately, signal once, then probe.
+//     This is intentionally before the normal retry loop: policy refresh and
+//     Dnscache paramchange cannot make local rules visible while GP mode is
+//     selected by an empty GP parent.
+//  3. Otherwise, signal DNS Client with increasing backoff between probes.
 func (p *prog) nrptProbeAndHeal() {
+	state, _ := p.dnsInterceptState.(*wfpState)
+	if state != nil {
+		now := time.Now()
+		if ok, wait := state.nrptRecoveryLimiter.allow(now, p.cfg); !ok {
+			if state.nrptRecoveryLimiter.shouldLogSkip(now) {
+				mainLog.Load().Warn().Dur("remaining", wait).
+					Msg("DNS intercept: NRPT recovery suppressed after repeated failed recovery flows")
+			}
+			return
+		}
+	}
+
 	if !nrptProbeRunning.CompareAndSwap(false, true) {
 		mainLog.Load().Debug().Msg("DNS intercept: NRPT probe already running, skipping")
 		return
 	}
 	defer nrptProbeRunning.Store(false)
 
+	remediated := false
+	defer func() {
+		if remediated && state != nil {
+			state.nrptRecoveryLimiter.recordRecoveryFlow(time.Now(), p.cfg)
+		}
+	}()
+
 	mainLog.Load().Info().Msg("DNS intercept: starting NRPT verification probe sequence")
 
 	// Log parent key state for diagnostics.
@@ -1805,17 +1844,37 @@ func (p *prog) nrptProbeAndHeal() {
 		mainLog.Load().Info().Msg("DNS intercept: NRPT verified working")
 		return
 	}
+	remediated = true
+
+	// If the GP parent exists but is empty, do not burn retries on Windows
+	// signaling. Those retries create SIEM noise but cannot succeed because DNS
+	// Client is still reading the empty GP store instead of the populated local
+	// store. Delete the blocker, send one notification, then re-probe.
+	if nrptParentKeyEmpty(nrptBaseKey) {
+		mainLog.Load().Warn().Msg("DNS intercept: NRPT probe failed with empty GP parent — cleaning before retry signaling")
+		if cleanEmptyNRPTParent() {
+			signalNRPTChange()
+			time.Sleep(1 * time.Second)
+			logNRPTParentKeyState("empty-gp-after-clean")
+			if p.probeNRPT() {
+				mainLog.Load().Info().Msg("DNS intercept: NRPT verified working after empty GP parent cleanup")
+				return
+			}
+		}
+		if nrptParentKeyEmpty(nrptBaseKey) {
+			mainLog.Load().Warn().Msg("DNS intercept: empty GP NRPT parent still present after cleanup; skipping redundant policy refresh retries")
+			return
+		}
+	}
 
-	// Attempts 2-4: GP refresh + paramchange + flush with increasing backoff
+	// Attempts 2-4: signal DNS Client with increasing backoff between probes.
 	delays := []time.Duration{1 * time.Second, 2 * time.Second, 4 * time.Second}
 	for i, delay := range delays {
 		attempt := i + 2
 		mainLog.Load().Info().Int("attempt", attempt).Dur("delay", delay).
-			Msg("DNS intercept: NRPT probe failed, retrying with GP refresh + paramchange")
+			Msg("DNS intercept: NRPT probe failed, retrying with policy refresh + paramchange")
 		logNRPTParentKeyState(fmt.Sprintf("probe-attempt-%d", attempt))
-		refreshNRPTPolicy()
-		sendParamChange()
-		flushDNSCache()
+		signalNRPTChange()
 		time.Sleep(delay)
 		if p.probeNRPT() {
 			mainLog.Load().Info().Int("attempt", attempt).
@@ -1829,7 +1888,7 @@ func (p *prog) nrptProbeAndHeal() {
 	// signal DNS Client to forget it, wait, then re-add and signal again.
 	mainLog.Load().Warn().Msg("DNS intercept: all probes failed — attempting two-phase NRPT recovery (delete → signal → re-add)")
 	listenerIP := "127.0.0.1"
-	if state, ok := p.dnsInterceptState.(*wfpState); ok {
+	if state != nil {
 		listenerIP = state.listenerIP
 	}
 
@@ -1837,9 +1896,7 @@ func (p *prog) nrptProbeAndHeal() {
 	_ = removeNRPTCatchAllRule()
 	// If parent key is now empty after removing our rule, delete it too.
 	cleanEmptyNRPTParent()
-	refreshNRPTPolicy()
-	sendParamChange()
-	flushDNSCache()
+	signalNRPTChange()
 	logNRPTParentKeyState("nuclear-after-delete")
 
 	// Wait for DNS Client to process the deletion.
@@ -1850,9 +1907,7 @@ func (p *prog) nrptProbeAndHeal() {
 		mainLog.Load().Error().Err(err).Msg("DNS intercept: failed to re-add NRPT after nuclear recovery")
 		return
 	}
-	refreshNRPTPolicy()
-	sendParamChange()
-	flushDNSCache()
+	signalNRPTChange()
 	logNRPTParentKeyState("nuclear-after-readd")
 
 	// Final probe after recovery.
diff --git a/cmd/cli/dns_proxy.go b/cmd/cli/dns_proxy.go
index fe37f609..4c08083a 100644
--- a/cmd/cli/dns_proxy.go
+++ b/cmd/cli/dns_proxy.go
@@ -1733,6 +1733,15 @@ func (p *prog) checkUpstreamOnce(upstream string, uc *ctrld.UpstreamConfig) erro
 			mainLog.Load().Debug().Err(err).Msgf("Upstream %s check failed after %v (WFP loopback protect active)", upstream, duration)
 			return errOsHealthcheckSuppressed
 		}
+		// A no-route/network-unreachable failure means the endpoint's address
+		// family is available locally but unroutable (e.g. an IPv6 DoH endpoint
+		// while IPv6 is up but has no route). These repeat until the route
+		// returns and are handled by bounded backoff in the recovery loop, so
+		// keep them at debug to avoid sustained error-log spam.
+		if ctrldnet.IsUnreachable(err) {
+			mainLog.Load().Debug().Err(err).Msgf("Upstream %s check failed after %v (network unreachable)", upstream, duration)
+			return err
+		}
 		mainLog.Load().Error().Err(err).Msgf("Upstream %s check failed after %v", upstream, duration)
 		return err
 	}
@@ -1937,6 +1946,9 @@ func (p *prog) handleRecovery(reason RecoveryReason) {
 // waitForUpstreamRecovery checks the provided upstreams concurrently until one recovers.
 // It returns the name of the recovered upstream or an error if the check times out.
 func (p *prog) waitForUpstreamRecovery(ctx context.Context, upstreams map[string]*ctrld.UpstreamConfig) (string, error) {
+	recoveryCtx, cancel := context.WithCancel(ctx)
+	defer cancel()
+
 	recoveredCh := make(chan string, 1)
 	var wg sync.WaitGroup
 
@@ -1948,9 +1960,10 @@ func (p *prog) waitForUpstreamRecovery(ctx context.Context, upstreams map[string
 			defer wg.Done()
 			mainLog.Load().Debug().Msgf("Starting recovery check loop for upstream: %s", name)
 			attempts := 0
+			unreachableStreak := 0
 			for {
 				select {
-				case <-ctx.Done():
+				case <-recoveryCtx.Done():
 					mainLog.Load().Debug().Msgf("Context canceled for upstream %s", name)
 					return
 				default:
@@ -1962,13 +1975,30 @@ func (p *prog) waitForUpstreamRecovery(ctx context.Context, upstreams map[string
 						select {
 						case recoveredCh <- name:
 							mainLog.Load().Debug().Msgf("Sent recovery notification for upstream %s", name)
+							cancel()
 						default:
 							mainLog.Load().Debug().Msg("Recovery channel full, another upstream already recovered")
 						}
 						return
 					}
-					mainLog.Load().Debug().Msgf("Upstream %s check failed, sleeping before retry", name)
-					time.Sleep(checkUpstreamBackoffSleep)
+					// Back off the retry cadence for an unroutable endpoint so a
+					// host with IPv6 up but no route to the IPv6 DoH endpoint does
+					// not re-bootstrap/re-check every checkUpstreamBackoffSleep and
+					// spam the log. The backoff is bounded (checkUpstreamUnreachableBackoffMax)
+					// so the endpoint is still re-probed and recovers when the route
+					// returns; any other failure resets to the base cadence.
+					sleep := checkUpstreamBackoffSleep
+					if ctrldnet.IsUnreachable(err) {
+						unreachableStreak++
+						sleep = unreachableRecoveryBackoff(unreachableStreak)
+						mainLog.Load().Debug().Msgf("Upstream %s unreachable (streak %d), backing off %s before retry", name, unreachableStreak, sleep)
+					} else {
+						unreachableStreak = 0
+						mainLog.Load().Debug().Msgf("Upstream %s check failed, sleeping before retry", name)
+					}
+					if !sleepWithContext(recoveryCtx, sleep) {
+						return
+					}
 
 					// if this is the upstreamOS and it's the 3rd attempt (or multiple of 3),
 					// we should try to reinit the OS resolver to ensure we can recover
@@ -1996,6 +2026,17 @@ func (p *prog) waitForUpstreamRecovery(ctx context.Context, upstreams map[string
 	return recovered, nil
 }
 
+func sleepWithContext(ctx context.Context, d time.Duration) bool {
+	timer := time.NewTimer(d)
+	defer timer.Stop()
+	select {
+	case <-timer.C:
+		return true
+	case <-ctx.Done():
+		return false
+	}
+}
+
 // buildRecoveryUpstreams constructs the map of upstream configurations to test.
 // For OS failures we supply the manual OS resolver upstream configuration.
 // For network change or regular failure we use the upstreams defined in p.cfg (ignoring OS).
diff --git a/cmd/cli/nrpt_recovery_limiter_windows.go b/cmd/cli/nrpt_recovery_limiter_windows.go
new file mode 100644
index 00000000..64b6526a
--- /dev/null
+++ b/cmd/cli/nrpt_recovery_limiter_windows.go
@@ -0,0 +1,101 @@
+//go:build windows
+
+package cli
+
+import (
+	"sync"
+	"time"
+
+	"github.com/Control-D-Inc/ctrld"
+)
+
+const (
+	// Default to current behavior: keep recovering indefinitely unless configured.
+	defaultNRPTRecoveryMaxAttempts = 0
+	defaultNRPTRecoveryCooldown    = 30 * time.Minute
+
+	// Require more than one good health tick before clearing the circuit. A probe can
+	// pass briefly after delete/re-add even when another agent recreates broken NRPT state.
+	nrptRecoveryStableSuccessesToReset = 2
+)
+
+type nrptRecoveryLimiter struct {
+	mu              sync.Mutex
+	attempts        int
+	stableSuccesses int
+	cooldownUntil   time.Time
+	lastSkipLog     time.Time
+}
+
+func nrptRecoveryMaxAttempts(cfg *ctrld.Config) int {
+	if cfg != nil && cfg.Service.NRPTRecoveryMaxAttempts != nil {
+		return *cfg.Service.NRPTRecoveryMaxAttempts
+	}
+	return defaultNRPTRecoveryMaxAttempts
+}
+
+func nrptRecoveryCooldown(cfg *ctrld.Config) time.Duration {
+	if cfg != nil && cfg.Service.NRPTRecoveryCooldown != nil {
+		return *cfg.Service.NRPTRecoveryCooldown
+	}
+	return defaultNRPTRecoveryCooldown
+}
+
+func (l *nrptRecoveryLimiter) allow(now time.Time, cfg *ctrld.Config) (bool, time.Duration) {
+	maxAttempts := nrptRecoveryMaxAttempts(cfg)
+	if maxAttempts <= 0 {
+		return true, 0
+	}
+
+	l.mu.Lock()
+	defer l.mu.Unlock()
+
+	if now.Before(l.cooldownUntil) {
+		return false, l.cooldownUntil.Sub(now)
+	}
+	return true, 0
+}
+
+func (l *nrptRecoveryLimiter) recordRecoveryFlow(now time.Time, cfg *ctrld.Config) {
+	maxAttempts := nrptRecoveryMaxAttempts(cfg)
+	if maxAttempts <= 0 {
+		return
+	}
+
+	cooldown := nrptRecoveryCooldown(cfg)
+	if cooldown <= 0 {
+		cooldown = defaultNRPTRecoveryCooldown
+	}
+
+	l.mu.Lock()
+	defer l.mu.Unlock()
+
+	l.stableSuccesses = 0
+	l.attempts++
+	if l.attempts >= maxAttempts {
+		l.cooldownUntil = now.Add(cooldown)
+	}
+}
+
+func (l *nrptRecoveryLimiter) recordStableSuccess() {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+
+	l.stableSuccesses++
+	if l.stableSuccesses >= nrptRecoveryStableSuccessesToReset {
+		l.attempts = 0
+		l.cooldownUntil = time.Time{}
+		l.lastSkipLog = time.Time{}
+	}
+}
+
+func (l *nrptRecoveryLimiter) shouldLogSkip(now time.Time) bool {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+
+	if l.lastSkipLog.IsZero() || now.Sub(l.lastSkipLog) >= 5*time.Minute {
+		l.lastSkipLog = now
+		return true
+	}
+	return false
+}
diff --git a/cmd/cli/nrpt_recovery_limiter_windows_test.go b/cmd/cli/nrpt_recovery_limiter_windows_test.go
new file mode 100644
index 00000000..d0974a54
--- /dev/null
+++ b/cmd/cli/nrpt_recovery_limiter_windows_test.go
@@ -0,0 +1,74 @@
+//go:build windows
+
+package cli
+
+import (
+	"testing"
+	"time"
+
+	"github.com/Control-D-Inc/ctrld"
+)
+
+func TestNRPTRecoveryLimiterCooldownAndStableReset(t *testing.T) {
+	maxAttempts := 2
+	cooldown := 10 * time.Minute
+	cfg := &ctrld.Config{}
+	cfg.Service.NRPTRecoveryMaxAttempts = &maxAttempts
+	cfg.Service.NRPTRecoveryCooldown = &cooldown
+
+	limiter := &nrptRecoveryLimiter{}
+	now := time.Unix(100, 0)
+
+	if ok, wait := limiter.allow(now, cfg); !ok || wait != 0 {
+		t.Fatalf("initial allow = %v, %v; want true, 0", ok, wait)
+	}
+
+	limiter.recordRecoveryFlow(now, cfg)
+	if ok, wait := limiter.allow(now.Add(time.Second), cfg); !ok || wait != 0 {
+		t.Fatalf("allow after first flow = %v, %v; want true, 0", ok, wait)
+	}
+
+	limiter.recordRecoveryFlow(now.Add(2*time.Second), cfg)
+	if ok, wait := limiter.allow(now.Add(3*time.Second), cfg); ok || wait <= 0 {
+		t.Fatalf("allow after max flows = %v, %v; want false, positive wait", ok, wait)
+	}
+
+	limiter.recordStableSuccess()
+	if ok, _ := limiter.allow(now.Add(4*time.Second), cfg); ok {
+		t.Fatal("one stable success cleared cooldown; want cooldown to remain")
+	}
+
+	limiter.recordStableSuccess()
+	if ok, wait := limiter.allow(now.Add(5*time.Second), cfg); !ok || wait != 0 {
+		t.Fatalf("allow after stable reset = %v, %v; want true, 0", ok, wait)
+	}
+}
+
+func TestNRPTRecoveryLimiterDefaultIsUnlimited(t *testing.T) {
+	cfg := &ctrld.Config{}
+	limiter := &nrptRecoveryLimiter{}
+	now := time.Unix(100, 0)
+
+	for i := 0; i < 10; i++ {
+		limiter.recordRecoveryFlow(now.Add(time.Duration(i)*time.Second), cfg)
+	}
+	if ok, wait := limiter.allow(now.Add(time.Hour), cfg); !ok || wait != 0 {
+		t.Fatalf("default allow after recovery flows = %v, %v; want true, 0", ok, wait)
+	}
+}
+
+func TestNRPTRecoveryLimiterUnlimited(t *testing.T) {
+	maxAttempts := 0
+	cfg := &ctrld.Config{}
+	cfg.Service.NRPTRecoveryMaxAttempts = &maxAttempts
+
+	limiter := &nrptRecoveryLimiter{}
+	now := time.Unix(100, 0)
+
+	for i := 0; i < 10; i++ {
+		limiter.recordRecoveryFlow(now.Add(time.Duration(i)*time.Second), cfg)
+	}
+	if ok, wait := limiter.allow(now.Add(time.Hour), cfg); !ok || wait != 0 {
+		t.Fatalf("unlimited allow = %v, %v; want true, 0", ok, wait)
+	}
+}
diff --git a/cmd/cli/prog.go b/cmd/cli/prog.go
index 5c5b2759..5da24906 100644
--- a/cmd/cli/prog.go
+++ b/cmd/cli/prog.go
@@ -34,6 +34,7 @@ import (
 	"github.com/Control-D-Inc/ctrld/internal/clientinfo"
 	"github.com/Control-D-Inc/ctrld/internal/controld"
 	"github.com/Control-D-Inc/ctrld/internal/dnscache"
+	ctrldnet "github.com/Control-D-Inc/ctrld/internal/net"
 	"github.com/Control-D-Inc/ctrld/internal/router"
 	"github.com/Control-D-Inc/ctrld/internal/router/dnsmasq"
 )
@@ -188,6 +189,21 @@ type prog struct {
 	// interception with exponential backoff and auto-heals if broken.
 	pfMonitorRunning atomic.Bool //lint:ignore U1000 used on darwin
 
+	// pfEnsureRunning ensures only one pf anchor validation/restoration runs at a time.
+	// Network-change callbacks, delayed rechecks, and the periodic watchdog can all
+	// converge during macOS interface churn; concurrent pfctl/scutil exec storms can
+	// exhaust process/file limits and make the outage worse.
+	pfEnsureRunning atomic.Bool //lint:ignore U1000 used on darwin
+
+	// pfExecBackoffUntil suppresses pf anchor validation after pfctl/scutil execs
+	// fail due host resource exhaustion (fork unavailable, too many open files).
+	pfExecBackoffUntil atomic.Int64 //lint:ignore U1000 used on darwin
+
+	// pfDelayedRecheckTimers coalesces delayed DNS-intercept rechecks after noisy
+	// network changes. Protected by pfDelayedRecheckMu.
+	pfDelayedRecheckMu     sync.Mutex    //lint:ignore U1000 used on darwin
+	pfDelayedRecheckTimers []*time.Timer //lint:ignore U1000 used on darwin
+
 	// pfProbeExpected holds the domain name of a pending pf interception probe.
 	// When non-empty, the DNS handler checks incoming queries against this value
 	// and signals pfProbeCh if matched. The probe verifies that pf's rdr rules
@@ -1328,13 +1344,14 @@ func errAddrInUse(err error) bool {
 
 var _ = errAddrInUse
 
+// The unreachable winsock errnos (ENETUNREACH/EHOSTUNREACH) are matched via
+// ctrldnet.IsUnreachable, which owns their definitions.
+//
 // https://learn.microsoft.com/en-us/windows/win32/winsock/windows-sockets-error-codes-2
 var (
 	windowsECONNREFUSED = syscall.Errno(10061)
-	windowsENETUNREACH  = syscall.Errno(10051)
 	windowsEINVAL       = syscall.Errno(10022)
 	windowsEADDRINUSE   = syscall.Errno(10048)
-	windowsEHOSTUNREACH = syscall.Errno(10065)
 )
 
 func errUrlNetworkError(err error) bool {
@@ -1351,14 +1368,14 @@ func errNetworkError(err error) bool {
 		if opErr.Temporary() {
 			return true
 		}
+		if ctrldnet.IsUnreachable(err) {
+			return true
+		}
 		switch {
 		case errors.Is(opErr.Err, syscall.ECONNREFUSED),
 			errors.Is(opErr.Err, syscall.EINVAL),
-			errors.Is(opErr.Err, syscall.ENETUNREACH),
-			errors.Is(opErr.Err, windowsENETUNREACH),
 			errors.Is(opErr.Err, windowsEINVAL),
-			errors.Is(opErr.Err, windowsECONNREFUSED),
-			errors.Is(opErr.Err, windowsEHOSTUNREACH):
+			errors.Is(opErr.Err, windowsECONNREFUSED):
 			return true
 		}
 	}
diff --git a/cmd/cli/prog_test.go b/cmd/cli/prog_test.go
index 6831f5ae..562a8e03 100644
--- a/cmd/cli/prog_test.go
+++ b/cmd/cli/prog_test.go
@@ -1,7 +1,11 @@
 package cli
 
 import (
+	"context"
+	"net"
+	"net/url"
 	"runtime"
+	"syscall"
 	"testing"
 	"time"
 
@@ -12,6 +16,32 @@ import (
 	"github.com/Control-D-Inc/ctrld"
 )
 
+func TestErrNetworkErrorTreatsNoRouteAsNetworkError(t *testing.T) {
+	err := &net.OpError{Op: "dial", Net: "tcp", Err: syscall.EHOSTUNREACH}
+	assert.True(t, errNetworkError(err))
+	assert.True(t, errUrlNetworkError(&url.Error{Op: "Get", URL: "https://dns.controld.com", Err: err}))
+}
+
+func TestSleepWithContext(t *testing.T) {
+	assert.True(t, sleepWithContext(context.Background(), time.Millisecond))
+
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+
+	start := time.Now()
+	assert.False(t, sleepWithContext(ctx, time.Minute))
+	assert.Less(t, time.Since(start), 100*time.Millisecond)
+}
+
+func TestUnreachableRecoveryBackoff(t *testing.T) {
+	// Streak starts at the base cadence and doubles each attempt, capped at the max.
+	assert.Equal(t, checkUpstreamBackoffSleep, unreachableRecoveryBackoff(0))
+	assert.Equal(t, checkUpstreamBackoffSleep, unreachableRecoveryBackoff(1))
+	assert.Equal(t, 2*checkUpstreamBackoffSleep, unreachableRecoveryBackoff(2))
+	assert.Equal(t, 4*checkUpstreamBackoffSleep, unreachableRecoveryBackoff(3))
+	assert.Equal(t, checkUpstreamUnreachableBackoffMax, unreachableRecoveryBackoff(100))
+}
+
 func Test_prog_dnsWatchdogEnabled(t *testing.T) {
 	p := &prog{cfg: &ctrld.Config{}}
 
diff --git a/cmd/cli/upstream_monitor.go b/cmd/cli/upstream_monitor.go
index 6e19e38a..3eeb24b5 100644
--- a/cmd/cli/upstream_monitor.go
+++ b/cmd/cli/upstream_monitor.go
@@ -12,8 +12,27 @@ const (
 	maxFailureRequest = 50
 	// checkUpstreamBackoffSleep is the time interval between each upstream checks.
 	checkUpstreamBackoffSleep = 2 * time.Second
+	// checkUpstreamUnreachableBackoffMax caps the recovery retry interval for an
+	// endpoint that keeps failing with a network-unreachable error. It bounds
+	// the backoff so an unroutable endpoint is still re-probed periodically and
+	// recovers once the route returns.
+	checkUpstreamUnreachableBackoffMax = 60 * time.Second
 )
 
+// unreachableRecoveryBackoff returns the retry interval for the given streak of
+// consecutive network-unreachable failures. It starts at checkUpstreamBackoffSleep
+// and doubles each attempt, capped at checkUpstreamUnreachableBackoffMax.
+func unreachableRecoveryBackoff(streak int) time.Duration {
+	d := checkUpstreamBackoffSleep
+	for i := 1; i < streak; i++ {
+		d *= 2
+		if d >= checkUpstreamUnreachableBackoffMax {
+			return checkUpstreamUnreachableBackoffMax
+		}
+	}
+	return d
+}
+
 // upstreamMonitor performs monitoring upstreams health.
 type upstreamMonitor struct {
 	cfg *ctrld.Config
diff --git a/cmd/cli/vpn_dns.go b/cmd/cli/vpn_dns.go
index f6f8c20c..115ed80a 100644
--- a/cmd/cli/vpn_dns.go
+++ b/cmd/cli/vpn_dns.go
@@ -6,6 +6,7 @@ import (
 	"runtime"
 	"strings"
 	"sync"
+	"sync/atomic"
 
 	"github.com/rs/zerolog"
 	"tailscale.com/net/netmon"
@@ -50,6 +51,9 @@ type vpnDNSManager struct {
 	// discoverVPNDNS is injected for tests so Refresh does not depend on the
 	// runner host's real VPN/virtual adapter state.
 	discoverVPNDNS func(context.Context) []ctrld.VPNDNSConfig
+	// refreshRunning keeps noisy network-change storms from running overlapping
+	// scutil/networksetup VPN DNS discovery work.
+	refreshRunning atomic.Bool
 	// Called when VPN DNS server list changes, to update intercept exemptions.
 	onServersChanged vpnDNSExemptFunc
 }
@@ -69,6 +73,11 @@ func newVPNDNSManager(exemptFunc vpnDNSExemptFunc) *vpnDNSManager {
 // Called on network change events.
 func (m *vpnDNSManager) Refresh(guardAgainstNoNameservers bool) {
 	logger := mainLog.Load()
+	if !m.refreshRunning.CompareAndSwap(false, true) {
+		logger.Debug().Msg("VPN DNS refresh already running, skipping duplicate")
+		return
+	}
+	defer m.refreshRunning.Store(false)
 
 	logger.Debug().Msg("Refreshing VPN DNS configurations")
 	discoverVPNDNS := m.discoverVPNDNS
diff --git a/cmd/cli/vpn_dns_test.go b/cmd/cli/vpn_dns_test.go
index 6a6d520d..846dc8ef 100644
--- a/cmd/cli/vpn_dns_test.go
+++ b/cmd/cli/vpn_dns_test.go
@@ -2,6 +2,8 @@ package cli
 
 import (
 	"context"
+	"sync"
+	"sync/atomic"
 	"testing"
 
 	"github.com/Control-D-Inc/ctrld"
@@ -14,6 +16,36 @@ func withVPNDNSSettlingEnabled(t *testing.T) {
 	t.Cleanup(func() { vpnDNSSettlingEnabled = old })
 }
 
+func TestVPNDNSRefreshSkipsConcurrentDuplicate(t *testing.T) {
+	m := newVPNDNSManager(nil)
+	started := make(chan struct{})
+	release := make(chan struct{})
+	done := make(chan struct{})
+	var once sync.Once
+	var calls atomic.Int32
+
+	m.discoverVPNDNS = func(context.Context) []ctrld.VPNDNSConfig {
+		calls.Add(1)
+		once.Do(func() { close(started) })
+		<-release
+		return nil
+	}
+
+	go func() {
+		defer close(done)
+		m.Refresh(true)
+	}()
+
+	<-started
+	m.Refresh(true)
+	close(release)
+	<-done
+
+	if calls.Load() != 1 {
+		t.Fatalf("expected overlapping refresh to be skipped, got %d discovery calls", calls.Load())
+	}
+}
+
 func TestVPNDNSRefreshRetainsStateForOneGuardedEmptyDiscovery(t *testing.T) {
 	withVPNDNSSettlingEnabled(t)
 	var gotExemptions []vpnDNSExemption
diff --git a/config.go b/config.go
index 1fc82e2f..fd94675e 100644
--- a/config.go
+++ b/config.go
@@ -241,6 +241,8 @@ type ServiceConfig struct {
 	ForceRefetchWaitTime    *int           `mapstructure:"force_refetch_wait_time" toml:"force_refetch_wait_time,omitempty"`
 	LeakOnUpstreamFailure   *bool          `mapstructure:"leak_on_upstream_failure" toml:"leak_on_upstream_failure,omitempty"`
 	InterceptMode           string         `mapstructure:"intercept_mode" toml:"intercept_mode,omitempty" validate:"omitempty,oneof=off dns hard"`
+	NRPTRecoveryMaxAttempts *int           `mapstructure:"nrpt_recovery_max_attempts" toml:"nrpt_recovery_max_attempts,omitempty" validate:"omitempty,gte=0"`
+	NRPTRecoveryCooldown    *time.Duration `mapstructure:"nrpt_recovery_cooldown" toml:"nrpt_recovery_cooldown,omitempty"`
 	Daemon                  bool           `mapstructure:"-" toml:"-"`
 	AllocateIP              bool           `mapstructure:"-" toml:"-"`
 }
diff --git a/docs/config.md b/docs/config.md
index 99e98c9c..8f21a4c2 100644
--- a/docs/config.md
+++ b/docs/config.md
@@ -295,6 +295,22 @@ If a remote upstream fails to resolve a query or is unreachable, `ctrld` will fo
 - Required: no
 - Default: true on Windows, MacOS and non-router Linux.
 
+### nrpt_recovery_max_attempts
+Windows DNS intercept mode uses NRPT health probes and recovery when Windows stops routing queries to the local `ctrld` listener. This limits how many consecutive recovery flows can run before `ctrld` enters a cooldown and stops making policy/Dnscache changes.
+
+Set to `0` to disable this circuit breaker and keep retrying indefinitely.
+
+- Type: integer
+- Required: no
+- Default: 0 (unlimited, current behavior)
+
+### nrpt_recovery_cooldown
+Cooldown duration after `nrpt_recovery_max_attempts` consecutive Windows NRPT recovery flows. During cooldown, `ctrld` logs the suppressed recovery and avoids additional `RefreshPolicyEx`, Dnscache `paramchange`, and DNS cache flush calls.
+
+- Type: time duration string
+- Required: no
+- Default: 30m
+
 ## Upstream
 The `[upstream]` section specifies the DNS upstream servers that `ctrld` will forward DNS requests to.
 
diff --git a/docs/wfp-dns-intercept.md b/docs/wfp-dns-intercept.md
index 2e7ece27..527d20d6 100644
--- a/docs/wfp-dns-intercept.md
+++ b/docs/wfp-dns-intercept.md
@@ -91,6 +91,20 @@ ctrld uses an adaptive strategy (matching [Tailscale's approach](https://github.
    the empty GP parent key. This ensures DNS Client stays in "local mode" where
    the local-path rule activates immediately via `paramchange`.
 
+### Reproducing the Empty GP Parent Case
+
+This is a production code reference, so the temporary repro script is not kept in
+the repository. For MR !942 review, the test script and exact before/after steps
+are posted in the MR discussion. The scenario to compare is:
+
+1. Run the same approved PowerShell repro script against a pre-fix build and this
+   branch with the same ctrld config.
+2. Create an empty GP NRPT parent key while ctrld is running in DNS intercept mode.
+3. Confirm pre-fix logs can spend policy refresh/paramchange retries while the GP
+   parent remains empty.
+4. Confirm post-fix logs clean the empty GP parent, send one NRPT-change signal,
+   and re-probe before normal retries.
+
 ### VPN Coexistence
 
 NRPT uses most-specific-match. VPN NRPT rules for specific domains (e.g.,
diff --git a/internal/net/net.go b/internal/net/net.go
index f4b55860..677e90dc 100644
--- a/internal/net/net.go
+++ b/internal/net/net.go
@@ -157,6 +157,101 @@ type parallelDialerResult struct {
 	err  error
 }
 
+const (
+	// unreachableBackoffBase is the initial suppression window applied to a
+	// dial address after it returns a network-unreachable error (e.g.
+	// "connect: no route to host"). The window grows exponentially up to
+	// unreachableBackoffMax on repeated failures, and is cleared as soon as
+	// the address dials successfully.
+	unreachableBackoffBase = 5 * time.Second
+	// unreachableBackoffMax caps the suppression window so an address is
+	// always re-probed within a bounded interval, preserving recovery when
+	// the route comes back.
+	unreachableBackoffMax = 60 * time.Second
+)
+
+// Windows winsock codes for the unreachable errnos. A failing connect on
+// Windows surfaces these raw WSA codes (WSAENETUNREACH/WSAEHOSTUNREACH), whereas
+// syscall.ENETUNREACH/EHOSTUNREACH are Go's portable "invented" values
+// (APPLICATION_ERROR + iota) that never equal them. Matching these explicitly is
+// therefore required for the classifier to detect unreachable errors on Windows;
+// errors.Is against the syscall.* constants alone would not.
+//
+// https://learn.microsoft.com/en-us/windows/win32/winsock/windows-sockets-error-codes-2
+var (
+	windowsENETUNREACH  = syscall.Errno(10051)
+	windowsEHOSTUNREACH = syscall.Errno(10065)
+)
+
+// IsUnreachable reports whether err indicates the destination network or host
+// has no route (ENETUNREACH/EHOSTUNREACH). These are the errors produced when
+// an endpoint's address family is available locally but unroutable, e.g. an
+// IPv6 DoH endpoint while the host has IPv6 but no route to it.
+func IsUnreachable(err error) bool {
+	if err == nil {
+		return false
+	}
+	var opErr *net.OpError
+	if errors.As(err, &opErr) {
+		return errors.Is(opErr.Err, syscall.ENETUNREACH) ||
+			errors.Is(opErr.Err, syscall.EHOSTUNREACH) ||
+			errors.Is(opErr.Err, windowsENETUNREACH) ||
+			errors.Is(opErr.Err, windowsEHOSTUNREACH)
+	}
+	return false
+}
+
+type unreachableEntry struct {
+	until   time.Time
+	backoff time.Duration
+}
+
+// unreachableTracker records dial addresses that recently failed with a
+// network-unreachable error so ParallelDialer can temporarily stop hammering
+// them. This prevents an unroutable endpoint from generating a sustained dial
+// /health-check storm, while still re-probing each address once its bounded
+// backoff window expires so genuine recovery is never permanently blocked.
+type unreachableTracker struct {
+	mu      sync.Mutex
+	entries map[string]unreachableEntry
+}
+
+var unreachable = &unreachableTracker{entries: make(map[string]unreachableEntry)}
+
+// suppressed reports whether addr is currently within its unreachable backoff
+// window and should be skipped.
+func (t *unreachableTracker) suppressed(addr string, now time.Time) bool {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	e, ok := t.entries[addr]
+	return ok && now.Before(e.until)
+}
+
+// markUnreachable extends the suppression window for addr using a bounded
+// exponential backoff.
+func (t *unreachableTracker) markUnreachable(addr string, now time.Time) {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	e := t.entries[addr]
+	if e.backoff == 0 {
+		e.backoff = unreachableBackoffBase
+	} else {
+		e.backoff *= 2
+		if e.backoff > unreachableBackoffMax {
+			e.backoff = unreachableBackoffMax
+		}
+	}
+	e.until = now.Add(e.backoff)
+	t.entries[addr] = e
+}
+
+// markReachable clears any suppression for addr after a successful dial.
+func (t *unreachableTracker) markReachable(addr string) {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	delete(t.entries, addr)
+}
+
 type ParallelDialer struct {
 	net.Dialer
 }
@@ -165,26 +260,63 @@ func (d *ParallelDialer) DialContext(ctx context.Context, network string, addrs
 	if len(addrs) == 0 {
 		return nil, errors.New("empty addresses")
 	}
+
+	// Skip addresses that recently returned a network-unreachable error so an
+	// unroutable endpoint (e.g. an IPv6 DoH address while the host has IPv6
+	// but no route to it) does not generate a sustained dial storm. Suppression
+	// is bounded: once an address's backoff window expires it is re-probed, so
+	// genuine recovery is preserved.
+	now := time.Now()
+	live := make([]string, 0, len(addrs))
+	var suppressed int
+	for _, addr := range addrs {
+		if unreachable.suppressed(addr, now) {
+			suppressed++
+			continue
+		}
+		live = append(live, addr)
+	}
+	if len(live) == 0 {
+		// Every candidate is within its unreachable backoff window. Fail fast
+		// and quietly instead of re-dialing known-unroutable addresses; the
+		// windows expire and re-probe shortly, so recovery still happens.
+		logger.Debug().Msgf("skipping %d unreachable address(es), all in backoff", suppressed)
+		// TODO: the errno here is hardcoded to EHOSTUNREACH, but the actual
+		// failure that triggered suppression may have been ENETUNREACH. This is
+		// harmless today (IsUnreachable treats both the same and nothing else
+		// inspects the errno), but if these errors are ever recorded/reported we
+		// should retain the real error in unreachableEntry and surface it here.
+		return nil, &net.OpError{Op: "dial", Net: network, Err: syscall.EHOSTUNREACH}
+	}
+	if suppressed > 0 {
+		logger.Debug().Msgf("skipping %d unreachable address(es) in backoff", suppressed)
+	}
+
 	ctx, cancel := context.WithCancel(ctx)
 	defer cancel()
 
 	done := make(chan struct{})
 	defer close(done)
-	ch := make(chan *parallelDialerResult, len(addrs))
+	ch := make(chan *parallelDialerResult, len(live))
 	var wg sync.WaitGroup
-	wg.Add(len(addrs))
+	wg.Add(len(live))
 	go func() {
 		wg.Wait()
 		close(ch)
 	}()
 
-	for _, addr := range addrs {
+	for _, addr := range live {
 		go func(addr string) {
 			defer wg.Done()
 			logger.Debug().Msgf("dialing to %s", addr)
 			conn, err := d.Dialer.DialContext(ctx, network, addr)
 			if err != nil {
 				logger.Debug().Msgf("failed to dial %s: %v", addr, err)
+				if IsUnreachable(err) {
+					unreachable.markUnreachable(addr, time.Now())
+				}
+			} else {
+				unreachable.markReachable(addr)
 			}
 			select {
 			case ch <- &parallelDialerResult{conn: conn, err: err}:
diff --git a/internal/net/net_test.go b/internal/net/net_test.go
index 7df3e09c..38187c5c 100644
--- a/internal/net/net_test.go
+++ b/internal/net/net_test.go
@@ -2,10 +2,77 @@ package net
 
 import (
 	"context"
+	"net"
+	"syscall"
 	"testing"
 	"time"
 )
 
+func TestIsUnreachable(t *testing.T) {
+	tests := []struct {
+		name string
+		err  error
+		want bool
+	}{
+		{"nil", nil, false},
+		{"enetunreach", &net.OpError{Op: "dial", Net: "tcp", Err: syscall.ENETUNREACH}, true},
+		{"ehostunreach", &net.OpError{Op: "dial", Net: "tcp", Err: syscall.EHOSTUNREACH}, true},
+		{"windows enetunreach", &net.OpError{Op: "dial", Net: "tcp", Err: windowsENETUNREACH}, true},
+		{"windows ehostunreach", &net.OpError{Op: "dial", Net: "tcp", Err: windowsEHOSTUNREACH}, true},
+		{"connection refused", &net.OpError{Op: "dial", Net: "tcp", Err: syscall.ECONNREFUSED}, false},
+		{"not an opError", syscall.ENETUNREACH, false},
+	}
+	for _, tc := range tests {
+		tc := tc
+		t.Run(tc.name, func(t *testing.T) {
+			if got := IsUnreachable(tc.err); got != tc.want {
+				t.Errorf("IsUnreachable(%v) = %v, want %v", tc.err, got, tc.want)
+			}
+		})
+	}
+}
+
+func TestUnreachableTracker(t *testing.T) {
+	tr := &unreachableTracker{entries: make(map[string]unreachableEntry)}
+	const addr = "[2606:1a40::22]:443"
+	now := time.Unix(0, 0)
+
+	// Not suppressed before any failure.
+	if tr.suppressed(addr, now) {
+		t.Fatal("addr suppressed before any failure")
+	}
+
+	// First failure suppresses for the base window.
+	tr.markUnreachable(addr, now)
+	if !tr.suppressed(addr, now.Add(unreachableBackoffBase-time.Millisecond)) {
+		t.Fatal("addr not suppressed within base backoff window")
+	}
+	if tr.suppressed(addr, now.Add(unreachableBackoffBase)) {
+		t.Fatal("addr still suppressed at end of base backoff window")
+	}
+
+	// Backoff grows exponentially and is capped at the max.
+	tr.markUnreachable(addr, now)
+	if got := tr.entries[addr].backoff; got != 2*unreachableBackoffBase {
+		t.Fatalf("backoff after second failure = %v, want %v", got, 2*unreachableBackoffBase)
+	}
+	for i := 0; i < 10; i++ {
+		tr.markUnreachable(addr, now)
+	}
+	if got := tr.entries[addr].backoff; got != unreachableBackoffMax {
+		t.Fatalf("backoff not capped: got %v, want %v", got, unreachableBackoffMax)
+	}
+
+	// A successful dial clears suppression entirely.
+	tr.markReachable(addr)
+	if tr.suppressed(addr, now) {
+		t.Fatal("addr still suppressed after markReachable")
+	}
+	if _, ok := tr.entries[addr]; ok {
+		t.Fatal("entry not removed after markReachable")
+	}
+}
+
 func TestProbeStackTimeout(t *testing.T) {
 	done := make(chan struct{})
 	started := make(chan struct{})
diff --git a/resolver.go b/resolver.go
index fbd2ad61..2284d04a 100644
--- a/resolver.go
+++ b/resolver.go
@@ -45,8 +45,12 @@ const (
 
 const controldPublicDns = "76.76.2.0"
 
+const maxConcurrentOSResolverExchanges = 128
+
 var controldPublicDnsWithPort = net.JoinHostPort(controldPublicDns, "53")
 
+var osResolverExchangeSem = make(chan struct{}, maxConcurrentOSResolverExchanges)
+
 var localResolver Resolver
 
 func init() {
@@ -136,14 +140,15 @@ func availableNameservers() []string {
 // It's the caller's responsibility to ensure the system DNS is in a clean state before
 // calling this function.
 func InitializeOsResolver(guardAgainstNoNameservers bool) []string {
+	resolverMutex.Lock()
+	defer resolverMutex.Unlock()
+
 	nameservers := availableNameservers()
 	// if no nameservers, return empty slice so we dont remove all nameservers
 	if len(nameservers) == 0 && guardAgainstNoNameservers {
 		return []string{}
 	}
 	ns := initializeOsResolver(nameservers)
-	resolverMutex.Lock()
-	defer resolverMutex.Unlock()
 	or = newResolverWithNameserver(ns)
 	return ns
 }
@@ -466,6 +471,13 @@ func (o *osResolver) resolve(ctx context.Context, msg *dns.Msg) (*dns.Msg, error
 		for _, server := range servers {
 			go func(server string) {
 				defer wg.Done()
+				release, ok := acquireOSResolverExchangeSlot(ctx)
+				if !ok {
+					ch <- &osResolverResult{err: ctx.Err(), server: server, lan: isLan}
+					return
+				}
+				defer release()
+
 				var answer *dns.Msg
 				var err error
 				var localOSResolverIP net.IP
@@ -576,6 +588,15 @@ func (o *osResolver) resolve(ctx context.Context, msg *dns.Msg) (*dns.Msg, error
 	return nil, errors.Join(errs...)
 }
 
+func acquireOSResolverExchangeSlot(ctx context.Context) (func(), bool) {
+	select {
+	case osResolverExchangeSem <- struct{}{}:
+		return func() { <-osResolverExchangeSem }, true
+	case <-ctx.Done():
+		return nil, false
+	}
+}
+
 func (o *osResolver) removeCache(key string) {
 	o.cache.Delete(key)
 }