policy/v2, state: elect primaries by oldest control session

A router that reconnects joins as a standby; one that kept its session
takes its routes back once available again, as on Tailscale. Both the
global and the via election use it, in place of anti-flap stickiness.

docs/ref/routes.md

@@ -345,6 +345,10 @@ availability](https://tailscale.com/docs/how-to/set-up-high-availability#subnet-
 This feature is enabled by default when at least two nodes advertise the same prefix. See the configuration options
 `node.routes.ha` in the [configuration file](configuration.md) for details.
 
+Of the routers that are online, healthy and not expired, the one connected to Headscale the longest serves a route. A
+router that reconnects becomes a standby, while a router that stays connected takes its routes back once it is approved,
+healthy or unexpired again. When none is available, the route stays with the router that served it last.
+
 ## Troubleshooting
 
 ### Enable IP forwarding

hscontrol/policy/pm.go

@@ -72,9 +72,10 @@ type PolicyManager interface {
 	// both fields are empty and the caller falls back to existing behavior.
 	ViaRoutesForPeer(viewer, peer types.NodeView) types.ViaRouteResult
 
-	// SetRouterAvailable sets the via election's runtime availability check.
+	// SetRouterSession sets the via election's runtime check: when a
+	// router's control session began and whether it may win.
 	// The callback runs under the policy lock and must remain lock-free.
-	SetRouterAvailable(available func(types.NodeID) bool)
+	SetRouterSession(session func(types.NodeID) (time.Time, bool))
 
 	// NodeCapMap returns compiled policy capabilities for the given node,
 	// or nil when none target it. The map is a defensive clone; its values

hscontrol/policy/v2/policy.go

@@ -119,7 +119,7 @@ type PolicyManager struct {
 	// called on every dispatched change, skips pm.mu when idle.
 	nodeAttrsPending atomic.Bool
 
-	routerAvailable func(types.NodeID) bool
+	routerSession func(types.NodeID) (time.Time, bool)
 }
 
 // filterAndPolicy combines the compiled filter rules with policy content for hashing.
@@ -1452,10 +1452,12 @@ func (pm *PolicyManager) NodeCanApproveRoute(node types.NodeView, route netip.Pr
 	return false
 }
 
-// SetRouterAvailable installs the availability check used by the via election.
-// Runtime status changes bypass SetNodes; nil treats every router as available.
-// The callback runs with pm.mu held and must remain a lock-free snapshot read.
-func (pm *PolicyManager) SetRouterAvailable(available func(types.NodeID) bool) {
+// SetRouterSession installs the session check used by the via election: when
+// a router's control session began and whether it may win. Runtime status
+// changes bypass SetNodes; nil treats every router as available, with one
+// session age. The callback runs with pm.mu held and must remain a lock-free
+// snapshot read.
+func (pm *PolicyManager) SetRouterSession(session func(types.NodeID) (time.Time, bool)) {
 	if pm == nil {
 		return
 	}
@@ -1463,7 +1465,7 @@ func (pm *PolicyManager) SetRouterAvailable(available func(types.NodeID) bool) {
 	pm.mu.Lock()
 	defer pm.mu.Unlock()
 
-	pm.routerAvailable = available
+	pm.routerSession = session
 }
 
 // ViaRoutesForPeer computes via grant effects for a viewer-peer pair.
@@ -1638,8 +1640,9 @@ func (pm *PolicyManager) ViaRoutesForPeer(viewer, peer types.NodeView) types.Via
 	// When neither condition is met, per-viewer via steering applies.
 	if len(result.Include) > 0 || len(result.Exclude) > 0 {
 		// Multi-router via election: when a via grant's tag matches
-		// multiple peers advertising the same prefix, only the
-		// lowest-ID available peer keeps the prefix in Include. The others
+		// multiple peers advertising the same prefix, only the available
+		// peer with the oldest control session (lower ID on a tie) keeps
+		// the prefix in Include, as the global election picks. The others
 		// move to Exclude. With no available router, the lowest ID keeps the
 		// prefix so the viewer still gets exactly one route.
 		//
@@ -1671,7 +1674,10 @@ func (pm *PolicyManager) ViaRoutesForPeer(viewer, peer types.NodeView) types.Via
 					continue
 				}
 
-				var viaPrimaryID, viaFallbackID types.NodeID
+				var (
+					viaPrimaryID, viaFallbackID types.NodeID
+					viaPrimarySince             time.Time
+				)
 
 				for _, viaTag := range grant.Via {
 					for _, node := range pm.nodes.All() {
@@ -1681,13 +1687,18 @@ func (pm *PolicyManager) ViaRoutesForPeer(viewer, peer types.NodeView) types.Via
 								viaFallbackID = node.ID()
 							}
 
-							if pm.routerAvailable != nil &&
-								!pm.routerAvailable(node.ID()) {
-								continue
+							var since time.Time
+
+							if pm.routerSession != nil {
+								var ok bool
+								if since, ok = pm.routerSession(node.ID()); !ok {
+									continue
+								}
 							}
 
-							if viaPrimaryID == 0 || node.ID() < viaPrimaryID {
-								viaPrimaryID = node.ID()
+							if viaPrimaryID == 0 || since.Before(viaPrimarySince) ||
+								(since.Equal(viaPrimarySince) && node.ID() < viaPrimaryID) {
+								viaPrimaryID, viaPrimarySince = node.ID(), since
 							}
 						}
 					}

hscontrol/policy/v2/policy_test.go

@@ -2797,20 +2797,26 @@ func TestViaRoutesForPeer(t *testing.T) {
 		pm, err := NewPolicyManager([]byte(pol), users, nodes.ViewSlice())
 		require.NoError(t, err)
 
+		// sessions maps each available router to when its control session
+		// began.
+		at := func(sec int64) time.Time { return time.Unix(sec, 0) }
+
 		for _, tt := range []struct {
-			name      string
-			available []types.NodeID
-			want      types.NodeID
+			name     string
+			sessions map[types.NodeID]time.Time
+			want     types.NodeID
 		}{
-			{name: "all available", available: []types.NodeID{2, 3, 4}, want: 2},
-			{name: "lowest ID down", available: []types.NodeID{3, 4}, want: 3},
-			{name: "only highest ID left", available: []types.NodeID{4}, want: 4},
+			{name: "all available", sessions: map[types.NodeID]time.Time{2: at(1), 3: at(1), 4: at(1)}, want: 2},
+			{name: "lowest ID down", sessions: map[types.NodeID]time.Time{3: at(1), 4: at(1)}, want: 3},
+			{name: "only highest ID left", sessions: map[types.NodeID]time.Time{4: at(1)}, want: 4},
 			{name: "whole group down", want: 2},
-			{name: "lowest ID recovers", available: []types.NodeID{2, 3, 4}, want: 2},
+			{name: "oldest session wins", sessions: map[types.NodeID]time.Time{2: at(3), 3: at(1), 4: at(2)}, want: 3},
 		} {
 			t.Run(tt.name, func(t *testing.T) {
-				pm.SetRouterAvailable(func(id types.NodeID) bool {
-					return slices.Contains(tt.available, id)
+				pm.SetRouterSession(func(id types.NodeID) (time.Time, bool) {
+					since, ok := tt.sessions[id]
+
+					return since, ok
 				})
 
 				for _, router := range nodes[1:] {

hscontrol/servertest/connect_race_test.go

@@ -37,10 +37,10 @@ func TestConnectDisconnectRace(t *testing.T) {
 	// real noise/poll path. After this, [state.NodeStore] + primaryRoutes already
 	// have the node, and [state.State.Connect] has been called once.
 	//
-	// Only c2 advertises the route. [tailcfg.NodeView.PrimaryRoutes] preserves a current
-	// primary across changes (anti-flap, see primary.go), so if both
-	// nodes were advertising, c1 (lower NodeID) would stay primary and
-	// the test could never observe the route slipping out of c2's
+	// Only c2 advertises the route. The election keeps a route on the
+	// router with the oldest control session, so if both nodes were
+	// advertising, c1 (connected first) would stay primary and the test
+	// could never observe the route slipping out of c2's
 	// [tailcfg.NodeView.PrimaryRoutes] — it would never have been there in the first place.
 	c1 := servertest.NewClient(t, srv, "race-r1", servertest.WithUser(user))
 	c2 := servertest.NewClient(t, srv, "race-r2", servertest.WithUser(user))

hscontrol/servertest/expiry_compat_test.go

@@ -36,11 +36,6 @@ var expirySkipReasons = map[string]string{
 	"expiry/C2-expire-r1/r1/paths-r2":        expiredPeerPaths,
 	"expiry/C2-expire-r1/r1/self":            selfExpiredFlags,
 
-	"expiry/C3-restore-r1/r1/routes-observer": restoredRouterTakesRoute,
-	"expiry/C3-restore-r1/r1/routes-r2":       restoredRouterTakesRoute,
-	"expiry/C3-restore-r1/r2/routes-observer": restoredRouterTakesRoute,
-	"expiry/C3-restore-r1/r2/routes-r1":       restoredRouterTakesRoute,
-
 	"expiry/C4-expire-r2/r2/online-observer": expiredPeerOffline,
 	"expiry/C4-expire-r2/r2/online-r1":       expiredPeerOffline,
 	"expiry/C4-expire-r2/r2/paths-observer":  expiredPeerPaths,
@@ -114,11 +109,10 @@ var expirySkipReasons = map[string]string{
 }
 
 const (
-	selfExpiredFlags         = "SELF_EXPIRED_FLAGS: headscale sends an expired node its own entry with MachineAuthorized false and Expired set; SaaS keeps MachineAuthorized true and leaves Expired out"
-	expiredPeerPaths         = "EXPIRED_PEER_PATHS: headscale sends peers an expired node's endpoints and home DERP; SaaS leaves both out"
-	clientExpiryRefused      = "CLIENT_EXPIRY_REFUSED: headscale refuses a client's request for an earlier key expiry; SaaS stores it, and the node expires then"
-	nameserversFromConfig    = "NAMESERVERS_FROM_CONFIG: headscale reads nameservers from its config and cannot change them while running; SaaS pushes a change to every node, an expired one too"
-	restoredRouterTakesRoute = "RESTORED_ROUTER_TAKES_ROUTE: SaaS hands a route back to a router whose key expiry is lifted, taking it from the standby that served it meanwhile; headscale keeps the standby"
+	selfExpiredFlags      = "SELF_EXPIRED_FLAGS: headscale sends an expired node its own entry with MachineAuthorized false and Expired set; SaaS keeps MachineAuthorized true and leaves Expired out"
+	expiredPeerPaths      = "EXPIRED_PEER_PATHS: headscale sends peers an expired node's endpoints and home DERP; SaaS leaves both out"
+	clientExpiryRefused   = "CLIENT_EXPIRY_REFUSED: headscale refuses a client's request for an earlier key expiry; SaaS stores it, and the node expires then"
+	nameserversFromConfig = "NAMESERVERS_FROM_CONFIG: headscale reads nameservers from its config and cannot change them while running; SaaS pushes a change to every node, an expired one too"
 )
 
 // TestNodeExpiryCompat replays key expiry against SaaS: an expired router

hscontrol/servertest/grants_test.go

@@ -1243,8 +1243,22 @@ func TestGrantViaSubnetFailover(t *testing.T) {
 				setExpiry(t, srv, router1ID, nil)
 			}
 
-			clients["client"].WaitForCondition(t, "route returns to router-1",
-				30*time.Second, onlyHolder("router-1"))
+			// router-1 keeps its control session through health and expiry
+			// changes, so it is older than router-2 and takes the route
+			// back; a reconnect opens a newer one.
+			want := "router-1"
+			if mode == "offline" {
+				want = "router-2"
+			}
+
+			clients["client"].WaitForCondition(t, "route settles on "+want,
+				30*time.Second, func(nm *netmap.NetworkMap) bool {
+					router1Online := slices.ContainsFunc(nm.Peers, func(p tailcfg.NodeView) bool {
+						return p.Hostinfo().Hostname() == "router-1" && p.Online().Get()
+					})
+
+					return router1Online && onlyHolder(want)(nm)
+				})
 			assert.Contains(t, srv.State().GetNodePrimaryRoutes(otherID), route)
 		})
 	}

hscontrol/servertest/ha_health_test.go

@@ -121,9 +121,10 @@ func TestHAHealthProbe_UnhealthyFailover(t *testing.T) {
 	assert.NotContains(t, primaries1, route, "node 1 should not be primary")
 }
 
-// TestHAHealthProbe_RecoveryNoFlap verifies that marking an unhealthy
-// node healthy again does NOT cause it to reclaim primary (stability).
-func TestHAHealthProbe_RecoveryNoFlap(t *testing.T) {
+// TestHAHealthProbe_RecoveryReturnsRoute verifies that marking an unhealthy
+// node healthy again hands it the route back: it kept the older control
+// session.
+func TestHAHealthProbe_RecoveryReturnsRoute(t *testing.T) {
 	t.Parallel()
 
 	srv := servertest.NewServer(t)
@@ -145,12 +146,12 @@ func TestHAHealthProbe_RecoveryNoFlap(t *testing.T) {
 	primaries := srv.State().GetNodePrimaryRoutes(nodeID2)
 	require.Contains(t, primaries, route, "node 2 should be primary")
 
-	// Recovery: node 1 healthy again. Node 2 should STAY primary.
+	// Recovery: node 1 healthy again, with the older session.
 	changed := srv.State().SetNodeHealth(nodeID1, true)
-	assert.False(t, changed, "recovery should not change primaries (no flap)")
+	assert.True(t, changed, "recovery should hand the route back")
 
-	primaries = srv.State().GetNodePrimaryRoutes(nodeID2)
-	assert.Contains(t, primaries, route, "node 2 should remain primary after recovery")
+	primaries = srv.State().GetNodePrimaryRoutes(nodeID1)
+	assert.Contains(t, primaries, route, "node 1 should be primary again after recovery")
 }
 
 // TestHAHealthProbe_ConnectClearsUnhealthy verifies that reconnecting
@@ -290,9 +291,8 @@ func TestHAHealthProbe_SetUnhealthyNoRoutesIsNoOp(t *testing.T) {
 // [HAHealthProber.ProbeOnce], the timeout path observes the reconnect
 // and bails out instead of installing a spurious Unhealthy bit.
 //
-// Without the guard, the primary fails over to the standby and the
-// anti-flap election preserves that choice even after the original
-// primary is fully back online.
+// Without the guard, the stale timeout marks the reconnected primary
+// unhealthy and the route fails over to the standby.
 func TestHAHealthProbe_ReconnectDuringProbeKeepsHealthy(t *testing.T) {
 	t.Parallel()
 
@@ -445,7 +445,8 @@ func TestHAHealthProbe_DeferralScopedToPrefix(t *testing.T) {
 	assert.True(t, srv.State().IsNodeHealthy(officeID1))
 	assert.True(t, srv.State().IsNodeHealthy(labID1),
 		"a deferral on another prefix must not withhold the lab node's recovery")
-	assert.Contains(t, srv.State().GetNodePrimaryRoutes(labID2), labRoute)
+	assert.Contains(t, srv.State().GetNodePrimaryRoutes(labID1), labRoute,
+		"the recovered lab node has the older session and takes its route back")
 }
 
 func TestHAHealthProbe_LastAdvertiserRecovers(t *testing.T) {

hscontrol/servertest/ha_property_test.go

@@ -27,8 +27,8 @@ type haPropClient struct {
 	name  string
 	route netip.Prefix
 	// connected mirrors the prober's view (mapBatcher.IsConnected).
-	// Tracked alongside the runtime read so anti-flap invariants can
-	// reason about the candidate set the prober actually evaluates.
+	// Tracked alongside the runtime read so the invariants can reason
+	// about the candidate set the prober actually evaluates.
 	connected bool
 	// freshSinceReconnect is true between a successful Reconnect and
 	// the prober's second observation of the new SessionEpoch — the
@@ -138,7 +138,7 @@ func readPrimaries(
 // checkHAInvariants walks every prefix → primary mapping in the live
 // snapshot and asserts the six properties documented in the test
 // header. prevPrimaries is the snapshot taken before the just-applied
-// op so anti-flap can compare moves.
+// op, the primary that keeps a prefix while no router is available.
 //
 //nolint:gocyclo // invariant checker over several independent properties
 func checkHAInvariants(
@@ -146,7 +146,6 @@ func checkHAInvariants(
 	srv *servertest.TestServer,
 	clients []*haPropClient,
 	prevPrimaries map[netip.Prefix]types.NodeID,
-	skipAntiFlap bool,
 ) {
 	rt.Helper()
 
@@ -246,31 +245,17 @@ func checkHAInvariants(
 			)
 		}
 
-		// Invariant 5 (anti-flap): if the previous snapshot already had
-		// a primary for this prefix AND that primary is still a healthy,
-		// online, advertising candidate, the new snapshot must keep it.
-		// Skipped on settle reads where we re-check without an op in
-		// between — the prober may run in a deferred batch and the
-		// election may legitimately move once on its own timeline.
-		if !skipAntiFlap {
-			if prev, hadPrev := prevPrimaries[prefix]; hadPrev && prev != owner {
-				prevNV, exists := st.GetNodeByID(prev)
-				if exists && prevNV.Valid() {
-					prevOnline, prevKnown := prevNV.IsOnline().GetOk()
-					prevApproved := prevNV.AllApprovedRoutes()
-
-					stillCandidate := prevKnown && prevOnline &&
-						slices.Contains(prevApproved, prefix)
-					stillHealthy := st.IsNodeHealthy(prev)
-
-					if stillCandidate && stillHealthy {
-						rt.Fatalf(
-							"prefix %s flapped primary %d -> %d "+
-								"while previous primary was still a healthy candidate",
-							prefix, prev, owner,
-						)
-					}
-				}
+		// Invariant 5 (oldest session): no healthy online candidate
+		// opened its control session before the primary's. A router
+		// that reconnects joins as a standby; one that keeps its
+		// session wins again once healthy.
+		for _, c := range advertisersByPrefix[prefix] {
+			cv, _ := st.GetNodeByID(c)
+			if st.IsNodeHealthy(c) && cv.SessionStart().Before(nv.SessionStart()) {
+				rt.Fatalf(
+					"prefix %s primary %d holds it over %d, whose session is older",
+					prefix, owner, c,
+				)
 			}
 		}
 	}
@@ -487,17 +472,6 @@ func TestHAProberProperty(t *testing.T) {
 
 			prev := snapshotPrimaries(rt, srv, clients)
 
-			// Anti-flap is meaningful only across operations that do
-			// not themselves mutate the candidate set or per-node
-			// health. Connect/Disconnect both mutate IsOnline (via
-			// state.Connect, or via the asynchronous grace-period
-			// state.Disconnect), and the resulting election move is
-			// expected, not a flap. ProbeOnce, by contrast, must keep
-			// the primary on the previous owner whenever that owner
-			// is still a candidate — this is what the reconnect-
-			// during-probe and dual-disconnect guards preserve.
-			isProberOp := op == 2
-
 			switch op {
 			case 0: // ClientDisconnect
 				idx := idxGen.Draw(rt, fmt.Sprintf("disc_idx_%d", step))
@@ -506,8 +480,8 @@ func TestHAProberProperty(t *testing.T) {
 				// Refuse to disconnect the last connected node. The
 				// prober's HANodes gate requires ≥2 online candidates
 				// per prefix; with only one node left, no probe runs
-				// and the prober-driven invariants (anti-flap,
-				// reconnect-defer) become unfalsifiable. Keeping at
+				// and the prober-driven invariants (health
+				// failover, reconnect-defer) become unfalsifiable. Keeping at
 				// least two candidates is enough to exercise every
 				// failure shape we care about.
 				connectedCount := 0
@@ -615,15 +589,13 @@ func TestHAProberProperty(t *testing.T) {
 				// to wait on; the no-op step still verifies that two
 				// consecutive reads (one snapshot, the same snapshot)
 				// stay consistent. This is invariant 3: stable
-				// conditions = stable primary. We skip the anti-flap
-				// comparison so a primary that legitimately changed
-				// in the prior op is not re-flagged here.
-				checkHAInvariants(rt, srv, clients, prev, true)
+				// conditions = stable primary.
+				checkHAInvariants(rt, srv, clients, prev)
 
 				continue
 			}
 
-			checkHAInvariants(rt, srv, clients, prev, !isProberOp)
+			checkHAInvariants(rt, srv, clients, prev)
 		}
 	})
 }

hscontrol/servertest/via_ha_compat_test.go

@@ -221,18 +221,7 @@ func runViaHACompat(t *testing.T, c *testcapture.Capture) {
 // viaHAFailoverSkipReasons lists checks where headscale differs from SaaS,
 // keyed <test_id>/<step>/<viewer>/<check>. A listed check that matches
 // fails, so a fix must delete its entry.
-var viaHAFailoverSkipReasons = map[string]string{
-	"via-ha-failover/V2/client-a/route-owner":  viaLowestIDAfterRestart,
-	"via-ha-failover/V10/client-a/route-owner": viaLowestIDAfterRestart,
-
-	"via-ha-failover/V12/client-a/route-owner": readvertisedRouterTakesRoute,
-	"via-ha-failover/V12/client-b/route-owner": readvertisedRouterTakesRoute,
-}
-
-const (
-	viaLowestIDAfterRestart      = "VIA_LOWEST_ID_AFTER_RESTART: headscale serves a via route from the group's lowest-ID router once it reconnects; SaaS keeps the router that served while it was down"
-	readvertisedRouterTakesRoute = "READVERTISED_ROUTER_TAKES_ROUTE: SaaS hands a route back to an HA router that withdrew and re-advertised it, taking it from the router that served it meanwhile; headscale keeps that router"
-)
+var viaHAFailoverSkipReasons = map[string]string{}
 
 // viaHAFailoverViewers are the clients whose views are compared: one reaches
 // the route only through the via group, the other through the global

hscontrol/state/node_store.go

@@ -169,8 +169,8 @@ type Snapshot struct {
 	allNodes    []types.NodeView
 
 	// routes maps each prefix to its current primary advertiser. The
-	// previous assignment is carried over when still valid so the
-	// primary does not flap on every unrelated batch.
+	// next election reads it to keep a prefix where it was while no
+	// router is available.
 	routes         map[netip.Prefix]types.NodeID
 	isPrimaryRoute map[types.NodeID]bool
 }
@@ -217,7 +217,8 @@ func updateChanges(pre, post *types.Node) (bool, bool) {
 
 	return false, pre.Online() != post.Online() ||
 		!preView.Expiry().GetOr(time.Time{}).Equal(postView.Expiry().GetOr(time.Time{})) ||
-		pre.Unhealthy != post.Unhealthy
+		pre.Unhealthy != post.Unhealthy ||
+		!pre.SessionStart.Equal(post.SessionStart)
 }
 
 // PutNode adds or updates a node in the store.
@@ -727,8 +728,8 @@ func snapshotFromNodes(
 	if reuseRoutes {
 		routes, isPrimaryRoute = prev.routes, prev.isPrimaryRoute
 	} else {
-		// Carrying the previous assignment forward lets a still-valid
-		// primary survive unrelated batches.
+		// The previous assignment keeps a prefix mapped while no router
+		// is available.
 		var prevRoutes map[netip.Prefix]types.NodeID
 		if prev != nil {
 			prevRoutes = prev.routes
@@ -818,8 +819,9 @@ func advertisers(
 
 // electPrimaryRoutes picks the primary router for each non-exit prefix any
 // node approves. A router is available when it is online, unexpired and
-// healthy. The previous primary is preserved while it is available
-// (anti-flap); otherwise the lowest-NodeID available router wins.
+// healthy. The available router with the oldest control session wins, the
+// lower NodeID on a tie: a router that reconnects joins as a standby, and
+// one that keeps its session takes its routes back once available again.
 //
 // With no router available the prefix stays mapped, so peers keep the route
 // and use it once its router returns: the previous primary keeps it while it
@@ -843,22 +845,24 @@ func electPrimaryRoutes(
 
 	routes := make(map[netip.Prefix]types.NodeID, len(approvers))
 	for prefix, ids := range approvers {
+		var oldest types.NodeID
+
+		// ids ascend, so a tie keeps the lower ID.
+		for _, id := range ids {
+			if available(id) && (oldest == 0 ||
+				nodes[id].SessionStart.Before(nodes[oldest].SessionStart)) {
+				oldest = id
+			}
+		}
+
+		if oldest != 0 {
+			routes[prefix] = oldest
+
+			continue
+		}
+
 		cur, hasPrev := prev[prefix]
-		hasPrev = hasPrev && slices.Contains(ids, cur)
-
-		if hasPrev && available(cur) {
-			routes[prefix] = cur
-
-			continue
-		}
-
-		if i := slices.IndexFunc(ids, available); i >= 0 {
-			routes[prefix] = ids[i]
-
-			continue
-		}
-
-		if hasPrev {
+		if hasPrev && slices.Contains(ids, cur) {
 			routes[prefix] = cur
 
 			continue
@@ -1095,12 +1099,12 @@ func (s *NodeStore) IsNodeHealthy(id types.NodeID) bool {
 	return !n.Unhealthy
 }
 
-// IsRouterAvailable reports whether id may win a primary election: online,
-// unexpired and healthy.
-func (s *NodeStore) IsRouterAvailable(id types.NodeID) bool {
+// RouterSession reports when id's control session began and whether id may
+// win a primary election: online, unexpired and healthy.
+func (s *NodeStore) RouterSession(id types.NodeID) (time.Time, bool) {
 	n, ok := s.data.Load().nodesByID[id]
 
-	return ok && servesRoutes(&n) && !n.Unhealthy
+	return n.SessionStart, ok && servesRoutes(&n) && !n.Unhealthy
 }
 
 // PrimaryRoutes returns the snapshot's prefix→primary map. The map is

hscontrol/state/node_store_test.go

@@ -1967,6 +1967,7 @@ func TestUpdateChangesReportsRelationFields(t *testing.T) {
 		{name: "online", mutate: func(n *types.Node) { n.IsOnline = new(true) }, wantElection: true},
 		{name: "unhealthy", mutate: func(n *types.Node) { n.Unhealthy = true }, wantElection: true},
 		{name: "expiry", mutate: func(n *types.Node) { n.Expiry = new(time.Now().Add(-time.Minute)) }, wantElection: true},
+		{name: "session start", mutate: func(n *types.Node) { n.SessionStart = time.Now() }, wantElection: true},
 		{name: "endpoint", mutate: func(n *types.Node) { n.Endpoints = []netip.AddrPort{netip.MustParseAddrPort("192.0.2.1:1")} }},
 		{name: "lastseen", mutate: func(n *types.Node) { n.LastSeen = new(time.Now()) }},
 	}

hscontrol/state/primaries_property_test.go

@@ -6,6 +6,7 @@ import (
 	"net/netip"
 	"slices"
 	"testing"
+	"time"
 
 	"github.com/juanfont/headscale/hscontrol/types"
 	"pgregory.net/rapid"
@@ -23,7 +24,10 @@ import (
 // is the one inside snapshotFromNodes; the model's job is to predict
 // the same answer from a separate, deliberately direct implementation.
 type primariesModel struct {
+	// connected mirrors IsOnline; since holds the step at which the
+	// node's current session opened.
 	connected map[types.NodeID]bool
+	since     map[types.NodeID]int
 	// announced mirrors Hostinfo.RoutableIPs — what the client says
 	// it can route. The election does not look at this directly; it
 	// is used to recompute the effective set whenever ApprovedRoutes
@@ -40,16 +44,17 @@ type primariesModel struct {
 	prefixes  map[types.NodeID][]netip.Prefix
 	unhealthy map[types.NodeID]bool
 
-	// primary[p] is the current primary for prefix p. The
-	// implementation preserves the current primary across changes to
-	// avoid flapping, so the model has to track this across
-	// operations rather than recompute a fresh choice each time.
+	// primary[p] is the current primary for prefix p. With no router
+	// available the implementation keeps the previous primary, so the
+	// model has to track this across operations rather than recompute
+	// a fresh choice each time.
 	primary map[netip.Prefix]types.NodeID
 }
 
 func newPrimariesModel() *primariesModel {
 	return &primariesModel{
 		connected: map[types.NodeID]bool{},
+		since:     map[types.NodeID]int{},
 		announced: map[types.NodeID][]netip.Prefix{},
 		approved:  map[types.NodeID][]netip.Prefix{},
 		prefixes:  map[types.NodeID][]netip.Prefix{},
@@ -120,6 +125,15 @@ func (m *primariesModel) approversByPrefix() map[netip.Prefix][]types.NodeID {
 	return out
 }
 
+// connect mirrors a session opening on node n at step.
+func (m *primariesModel) connect(n types.NodeID, step int) {
+	if !m.connected[n] {
+		m.since[n] = step
+	}
+
+	m.connected[n] = true
+}
+
 // available reports whether n may win an election: connected and healthy.
 func (m *primariesModel) available(n types.NodeID) bool {
 	return m.connected[n] && !m.unhealthy[n]
@@ -138,19 +152,25 @@ func (m *primariesModel) updatePrimaries() {
 	}
 
 	for p, nodes := range approvers {
+		// The oldest session wins; nodes ascend, so a tie keeps the
+		// lower ID.
+		var oldest types.NodeID
+
+		for _, n := range nodes {
+			if m.available(n) && (oldest == 0 || m.since[n] < m.since[oldest]) {
+				oldest = n
+			}
+		}
+
+		if oldest != 0 {
+			m.primary[p] = oldest
+
+			continue
+		}
+
 		cur, hasPrev := m.primary[p]
 		hasPrev = hasPrev && slices.Contains(nodes, cur)
 
-		if hasPrev && m.available(cur) {
-			continue
-		}
-
-		if i := slices.IndexFunc(nodes, m.available); i >= 0 {
-			m.primary[p] = nodes[i]
-
-			continue
-		}
-
 		// No router available: the previous primary keeps the prefix,
 		// else the lowest-ID approver the prober has not declared
 		// unreachable; with neither the prefix is unmapped.
@@ -189,8 +209,8 @@ func samePrefixSet(a, b []netip.Prefix) bool {
 
 // checkPrimariesProperties asserts every rule we expect of the snapshot's
 // primaries map given the model. prevSnapshotPrimaries is the snapshot's
-// PrimaryRoutes() reading taken before the just-applied op, used to
-// catch flap regressions that move primary off a still-eligible owner.
+// PrimaryRoutes() reading taken before the just-applied op, the primary
+// that may stay unhealthy while no router is available.
 func checkPrimariesProperties(
 	rt *rapid.T,
 	ns *NodeStore,
@@ -327,21 +347,15 @@ func checkPrimariesProperties(
 			)
 		}
 
-		// Anti-flap: if the previous snapshot already had a primary
-		// for this prefix AND that primary is still a healthy
-		// candidate after the op, the election must keep it. Flapping
-		// the primary for an unrelated reason violates the contract
-		// the integration tests rely on (HA failover only moves on
-		// loss of candidacy or health).
-		if prev, hadPrev := prevSnapshotPrimaries[prefix]; hadPrev {
-			stillCandidate := slices.Contains(candidates, prev)
-			stillHealthy := !m.unhealthy[prev]
-
-			if stillCandidate && stillHealthy && owner != prev {
+		// Oldest session: no available router opened its control
+		// session before the primary's. A router that reconnects joins
+		// as a standby; one that keeps its session wins again.
+		for _, c := range approversByPrefix[prefix] {
+			cv, _ := ns.GetNode(c)
+			if m.available(c) && cv.SessionStart().Before(nv.SessionStart()) {
 				rt.Fatalf(
-					"prefix %s flapped primary %d -> %d "+
-						"while previous primary was still a healthy candidate",
-					prefix, prev, owner,
+					"prefix %s primary %d holds it over %d, whose session is older",
+					prefix, owner, c,
 				)
 			}
 		}
@@ -447,10 +461,20 @@ func TestPrimaryRoutesProperty(t *testing.T) {
 		for step := range opCount {
 			op := rapid.IntRange(0, 8).Draw(rt, fmt.Sprintf("op_%d", step))
 
+			// connect opens a session, as State.Connect does, when the
+			// node is offline.
+			connect := func(n *types.Node) {
+				if !n.Online() {
+					n.SessionStart = time.Unix(int64(step), 0)
+				}
+
+				n.IsOnline = new(true)
+			}
+
 			// Snapshot primaries before applying the op so the
-			// anti-flap invariant has a stable reference. Reading
-			// after the model has already changed would compare a
-			// stale snapshot to a moved model.
+			// healthy-preference check knows the previous primary.
+			// Reading after the model has already changed would compare
+			// a stale snapshot to a moved model.
 			prevPrimaries := snapshotPrimariesCopy(ns)
 
 			switch op {
@@ -459,7 +483,7 @@ func TestPrimaryRoutesProperty(t *testing.T) {
 				prefs := prefixSubsetGen.Draw(rt, fmt.Sprintf("prefs_%d", step))
 
 				ns.UpdateNode(id, func(n *types.Node) {
-					n.IsOnline = new(true)
+					connect(n)
 					n.Unhealthy = false
 					n.Hostinfo = &tailcfg.Hostinfo{RoutableIPs: prefs}
 					n.ApprovedRoutes = prefs
@@ -468,12 +492,7 @@ func TestPrimaryRoutesProperty(t *testing.T) {
 				m.announced[id] = prefs
 				m.approved[id] = prefs
 				m.recomputeEffective(id)
-
-				if len(prefs) == 0 {
-					delete(m.connected, id)
-				} else {
-					m.connected[id] = true
-				}
+				m.connect(id, step)
 
 				delete(m.unhealthy, id)
 
@@ -503,7 +522,7 @@ func TestPrimaryRoutesProperty(t *testing.T) {
 				prefs := prefixSubsetGen.Draw(rt, fmt.Sprintf("prefs_%d", step))
 
 				ns.UpdateNode(id, func(n *types.Node) {
-					n.IsOnline = new(true)
+					connect(n)
 					n.Hostinfo = &tailcfg.Hostinfo{RoutableIPs: prefs}
 					n.ApprovedRoutes = prefs
 				})
@@ -511,12 +530,7 @@ func TestPrimaryRoutesProperty(t *testing.T) {
 				m.announced[id] = prefs
 				m.approved[id] = prefs
 				m.recomputeEffective(id)
-
-				if len(prefs) == 0 {
-					delete(m.connected, id)
-				} else {
-					m.connected[id] = true
-				}
+				m.connect(id, step)
 
 			case 5: // BatchProbeResults — atomic health flips per cycle.
 				// Mirrors State.BatchSetNodeHealth: per-node

hscontrol/state/primaries_test.go

@@ -24,6 +24,24 @@ func mp(prefix string) netip.Prefix {
 type primariesFixture struct {
 	t  *testing.T
 	ns *NodeStore
+
+	// clock orders the control sessions advertise and approveRoutes open.
+	clock time.Time
+}
+
+// session returns an update that opens a control session, as State.Connect
+// does, when the node is offline: each opens later than the one before.
+func (f *primariesFixture) session() func(*types.Node) {
+	f.clock = f.clock.Add(time.Second)
+	start := f.clock
+
+	return func(n *types.Node) {
+		if !n.Online() {
+			n.SessionStart = start
+		}
+
+		n.IsOnline = new(true)
+	}
 }
 
 func newPrimariesFixture(t *testing.T, ids ...types.NodeID) *primariesFixture {
@@ -45,8 +63,10 @@ func newPrimariesFixture(t *testing.T, ids ...types.NodeID) *primariesFixture {
 // prefs argument leaves the node online but advertising nothing.
 func (f *primariesFixture) advertise(id types.NodeID, prefs ...netip.Prefix) {
 	f.t.Helper()
+
+	connect := f.session()
 	f.ns.UpdateNode(id, func(n *types.Node) {
-		n.IsOnline = new(true)
+		connect(n)
 		n.Unhealthy = false
 		n.Hostinfo = &tailcfg.Hostinfo{RoutableIPs: prefs}
 		n.ApprovedRoutes = prefs
@@ -58,8 +78,10 @@ func (f *primariesFixture) advertise(id types.NodeID, prefs ...netip.Prefix) {
 // Unhealthy.
 func (f *primariesFixture) approveRoutes(id types.NodeID, prefs ...netip.Prefix) {
 	f.t.Helper()
+
+	connect := f.session()
 	f.ns.UpdateNode(id, func(n *types.Node) {
-		n.IsOnline = new(true)
+		connect(n)
 		n.Hostinfo = &tailcfg.Hostinfo{RoutableIPs: prefs}
 		n.ApprovedRoutes = prefs
 	})
@@ -147,10 +169,10 @@ func TestPrimaries_OverlappingRoutesLowerIDWins(t *testing.T) {
 	f.requireNodeRoutes(2)
 }
 
-func TestPrimaries_AntiFlapPreservesCurrentPrimary(t *testing.T) {
+func TestPrimaries_ReconnectJoinsAsStandby(t *testing.T) {
 	// A primary that disappears (advertiser leaves the set) should
-	// trigger failover. When the original primary returns, the new
-	// primary keeps the assignment — anti-flap.
+	// trigger failover. When the original primary returns on a new
+	// session, the new primary's session is older and keeps the route.
 	f := newPrimariesFixture(t, 1, 2)
 	f.advertise(1, mp("192.168.1.0/24"))
 	f.advertise(2, mp("192.168.1.0/24"))
@@ -276,15 +298,75 @@ func TestPrimaries_UnhealthyTriggersFailover(t *testing.T) {
 	f.requirePrimary(mp("192.168.1.0/24"), 2)
 }
 
-func TestPrimaries_RecoveryFromUnhealthyNoFlap(t *testing.T) {
-	f := newPrimariesFixture(t, 1, 2)
-	f.advertise(1, mp("192.168.1.0/24"))
-	f.advertise(2, mp("192.168.1.0/24"))
-	f.unhealthy(1)
-	f.requirePrimary(mp("192.168.1.0/24"), 2)
+func TestPrimaries_OldestSessionWins(t *testing.T) {
+	prefix := mp("192.168.1.0/24")
+	withdraw := func(f *primariesFixture, id types.NodeID) {
+		f.ns.UpdateNode(id, func(n *types.Node) { n.Hostinfo = &tailcfg.Hostinfo{} })
+	}
 
-	f.healthy(1)
-	f.requirePrimary(mp("192.168.1.0/24"), 2)
+	tests := []struct {
+		name  string
+		setup func(*primariesFixture)
+		want  types.NodeID
+	}{
+		{
+			name: "older session over lower ID",
+			setup: func(f *primariesFixture) {
+				f.advertise(2, prefix)
+				f.advertise(1, prefix)
+			},
+			want: 2,
+		},
+		{
+			name: "healthy again takes the route back",
+			setup: func(f *primariesFixture) {
+				f.advertise(1, prefix)
+				f.advertise(2, prefix)
+				f.unhealthy(1)
+				f.healthy(1)
+			},
+			want: 1,
+		},
+		{
+			name: "approved again takes the route back",
+			setup: func(f *primariesFixture) {
+				f.advertise(1, prefix)
+				f.advertise(2, prefix)
+				f.approveRoutes(1)
+				f.approveRoutes(1, prefix)
+			},
+			want: 1,
+		},
+		{
+			name: "advertised again takes the route back",
+			setup: func(f *primariesFixture) {
+				f.advertise(1, prefix)
+				f.advertise(2, prefix)
+				withdraw(f, 1)
+				f.requirePrimary(prefix, 2)
+				f.approveRoutes(1, prefix)
+			},
+			want: 1,
+		},
+		{
+			name: "reconnected router stays standby",
+			setup: func(f *primariesFixture) {
+				f.advertise(1, prefix)
+				f.advertise(2, prefix)
+				f.disconnect(1)
+				f.advertise(1, prefix)
+			},
+			want: 2,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			f := newPrimariesFixture(t, 1, 2)
+			tt.setup(f)
+			f.requirePrimary(prefix, tt.want)
+		})
+	}
 }
 
 func TestPrimaries_AllUnhealthyKeepsAPrimary(t *testing.T) {
@@ -332,15 +414,19 @@ func TestPrimaries_ExpiredRouterNotElected(t *testing.T) {
 	f.advertise(1, prefix)
 	f.advertise(2, prefix)
 	f.requirePrimary(prefix, 1)
-	require.True(t, f.ns.IsRouterAvailable(1))
+	_, ok := f.ns.RouterSession(1)
+	require.True(t, ok)
 
 	f.ns.UpdateNode(1, func(n *types.Node) {
 		n.Expiry = new(time.Now().Add(-time.Minute))
 	})
 
 	f.requirePrimary(prefix, 2)
-	require.False(t, f.ns.IsRouterAvailable(1))
-	require.True(t, f.ns.IsRouterAvailable(2))
+	_, ok = f.ns.RouterSession(1)
+	require.False(t, ok)
+
+	_, ok = f.ns.RouterSession(2)
+	require.True(t, ok)
 }
 
 func TestPrimaries_ExitRouteNotElected(t *testing.T) {

hscontrol/state/state.go

@@ -281,7 +281,7 @@ func NewState(cfg *types.Config) (*State, error) {
 	batchTimeout := cmp.Or(cfg.Tuning.NodeStoreBatchTimeout, defaultNodeStoreBatchTimeout)
 
 	nodeStore := NewNodeStore(nodes, policyPeersFunc(polMan), batchSize, batchTimeout)
-	polMan.SetRouterAvailable(nodeStore.IsRouterAvailable)
+	polMan.SetRouterSession(nodeStore.RouterSession)
 	nodeStore.Start()
 
 	// Last: NewLRU starts the TTL sweeper goroutine, which only Close stops.
@@ -702,6 +702,13 @@ func (s *State) Connect(id types.NodeID) ([]change.Change, uint64) {
 	node, ok := s.nodeStore.UpdateNode(id, func(n *types.Node) {
 		n.SessionEpoch++
 		epoch = n.SessionEpoch
+
+		// A reconnect within poll.go's offline grace keeps the age: peers
+		// never saw the node leave, so a blip does not move its routes.
+		if n.ActiveSessions == 0 {
+			n.SessionStart = time.Now()
+		}
+
 		n.ActiveSessions++
 		n.IsOnline = new(n.ShouldBeOnline())
 		n.Unhealthy = false

integration/route_test.go

@@ -312,6 +312,8 @@ func TestHASubnetRouterFailover(t *testing.T) {
 
 	client := allClients[3]
 
+	orderRouterSessions(t, headscale, subRouter1, subRouter2, subRouter3)
+
 	t.Logf("%s (%s) picked as client", client.Hostname(), client.MustID())
 	t.Logf("=== Initial Route Advertisement - Setting up HA configuration with 3 routers ===")
 	t.Logf("[%s] Starting test section", time.Now().Format(TimestampFormat))
@@ -1028,7 +1030,7 @@ func TestHASubnetRouterFailover(t *testing.T) {
 	t.Logf("[%s] Starting test section", time.Now().Format(TimestampFormat))
 	t.Logf("  Current state: Router 1 STANDBY, Router 2 STANDBY, Router 3 PRIMARY")
 	t.Logf("  Action: Disabling route approval on router 3 (route still advertised but not approved)")
-	t.Logf("  Expected: Router 1 (%s) should become new PRIMARY (lowest ID with approved route)", subRouter1.Hostname())
+	t.Logf("  Expected: Router 1 (%s) should become new PRIMARY (came back before router 2, so its session is older)", subRouter1.Hostname())
 	t.Logf("  Expected: Router 2 (%s) remains STANDBY", subRouter2.Hostname())
 	t.Logf("  Expected: Router 3 (%s) goes to advertised-only state (no longer serving)", subRouter3.Hostname())
 	_, err = headscale.ApproveRoutes(mustParseID(MustFindNode(subRouter3.Hostname(), nodes).Id), []netip.Prefix{})
@@ -1195,14 +1197,15 @@ func TestHASubnetRouterFailover(t *testing.T) {
 
 	checkFailureAndPrintRoutes(t, client)
 
-	// enable the route of subnet router 1, no change expected
+	// Re-enable the route of subnet router 1. Its control session is older
+	// than router 2's (it came back first), so it takes the route back.
 	t.Logf("=== ROUTE RE-ENABLE TEST: Re-approving route on router 1 (%s) ===", subRouter1.Hostname())
 	t.Logf("[%s] Starting test section", time.Now().Format(TimestampFormat))
 	t.Logf("  Current state: Router 1 advertised-only, Router 2 PRIMARY, Router 3 advertised-only")
 	t.Logf("  Action: Re-enabling route approval on router 1")
-	t.Logf("  Expected: Router 2 (%s) remains PRIMARY (stability - no unnecessary flapping)", subRouter2.Hostname())
-	t.Logf("  Expected: Router 1 (%s) becomes STANDBY (approved but not primary)", subRouter1.Hostname())
-	t.Logf("  Expected: HA fully restored with Router 2 PRIMARY and Router 1 STANDBY")
+	t.Logf("  Expected: Router 1 (%s) becomes PRIMARY (older control session)", subRouter1.Hostname())
+	t.Logf("  Expected: Router 2 (%s) becomes STANDBY", subRouter2.Hostname())
+	t.Logf("  Expected: HA fully restored with Router 1 PRIMARY and Router 2 STANDBY")
 
 	r1Node := MustFindNode(subRouter1.Hostname(), nodes)
 	_, err = headscale.ApproveRoutes(
@@ -1216,10 +1219,10 @@ func TestHASubnetRouterFailover(t *testing.T) {
 		assert.NoError(c, err)
 		assert.Len(c, nodes, 6)
 
-		requireNodeRouteCountWithCollect(c, MustFindNode(subRouter1.Hostname(), nodes), 1, 1, 0)
-		requireNodeRouteCountWithCollect(c, MustFindNode(subRouter2.Hostname(), nodes), 1, 1, 1)
+		requireNodeRouteCountWithCollect(c, MustFindNode(subRouter1.Hostname(), nodes), 1, 1, 1)
+		requireNodeRouteCountWithCollect(c, MustFindNode(subRouter2.Hostname(), nodes), 1, 1, 0)
 		requireNodeRouteCountWithCollect(c, MustFindNode(subRouter3.Hostname(), nodes), 1, 0, 0)
-	}, propagationTime, 200*time.Millisecond, "Re-enable verification: Router 1 approved as STANDBY, Router 2 remains PRIMARY (no flapping), full HA restored")
+	}, propagationTime, 200*time.Millisecond, "Re-enable verification: Router 1 PRIMARY, Router 2 STANDBY, full HA restored")
 
 	// Verify that the route is announced from subnet router 1
 	assert.EventuallyWithT(t, func(c *assert.CollectT) {
@@ -1238,33 +1241,33 @@ func TestHASubnetRouterFailover(t *testing.T) {
 			return
 		}
 
-		assert.Nil(c, srs1PeerStatus.PrimaryRoutes)
-		assert.NotNil(c, srs2PeerStatus.PrimaryRoutes)
+		assert.NotNil(c, srs1PeerStatus.PrimaryRoutes)
+		assert.Nil(c, srs2PeerStatus.PrimaryRoutes)
 		assert.Nil(c, srs3PeerStatus.PrimaryRoutes)
 
-		if srs2PeerStatus.PrimaryRoutes != nil {
+		if srs1PeerStatus.PrimaryRoutes != nil {
 			assert.Contains(c,
-				srs2PeerStatus.PrimaryRoutes.AsSlice(),
+				srs1PeerStatus.PrimaryRoutes.AsSlice(),
 				pref,
 			)
 		}
-	}, propagationTime, 200*time.Millisecond, "Verifying Router 2 remains PRIMARY after Router 1 route re-enabled")
+	}, propagationTime, 200*time.Millisecond, "Verifying Router 1 is PRIMARY after its route is re-enabled")
 
 	assert.EventuallyWithT(t, func(c *assert.CollectT) {
-		assertCurlDockerHostname(c, client, weburl, "Verifying client can reach webservice through router 2 after route re-enable")
-	}, propagationTime, 200*time.Millisecond, "Verifying client can reach webservice through router 2 after route re-enable")
+		assertCurlDockerHostname(c, client, weburl, "Verifying client can reach webservice through router 1 after route re-enable")
+	}, propagationTime, 200*time.Millisecond, "Verifying client can reach webservice through router 1 after route re-enable")
 
 	assert.EventuallyWithT(t, func(c *assert.CollectT) {
 		tr, err := client.Traceroute(webip)
 		assert.NoError(c, err)
 
-		ip, err := subRouter2.IPv4()
-		if !assert.NoError(c, err, "failed to get IPv4 for subRouter2") {
+		ip, err := subRouter1.IPv4()
+		if !assert.NoError(c, err, "failed to get IPv4 for subRouter1") {
 			return
 		}
 
 		assertTracerouteViaIPWithCollect(c, tr, ip)
-	}, propagationTime, 200*time.Millisecond, "Verifying traceroute still goes through router 2 after route re-enable")
+	}, propagationTime, 200*time.Millisecond, "Verifying traceroute goes through router 1 after route re-enable")
 
 	// Validate primary routes table state after router 1 re-approval
 	validatePrimaryRoutes(t, headscale, &types.DebugRoutes{
@@ -1274,19 +1277,20 @@ func TestHASubnetRouterFailover(t *testing.T) {
 			// Router 3 route is still not approved
 		},
 		PrimaryRoutes: map[string]types.NodeID{
-			pref.String(): types.NodeID(mustParseID(MustFindNode(subRouter2.Hostname(), nodes).Id)),
+			pref.String(): types.NodeID(mustParseID(MustFindNode(subRouter1.Hostname(), nodes).Id)),
 		},
-	}, "Router 2 should remain primary after router 1 re-approval")
+	}, "Router 1 should be primary after router 1 re-approval")
 
 	checkFailureAndPrintRoutes(t, client)
 
-	// Enable route on node 3, we now have all routes re-enabled
+	// Enable route on node 3, we now have all routes re-enabled. Router 3
+	// never reconnected, so its session is the oldest and it takes the route.
 	t.Logf("=== ROUTE RE-ENABLE TEST: Re-approving route on router 3 (%s) - Full HA Restoration ===", subRouter3.Hostname())
 	t.Logf("[%s] Starting test section", time.Now().Format(TimestampFormat))
-	t.Logf("  Current state: Router 1 STANDBY, Router 2 PRIMARY, Router 3 advertised-only")
+	t.Logf("  Current state: Router 1 PRIMARY, Router 2 STANDBY, Router 3 advertised-only")
 	t.Logf("  Action: Re-enabling route approval on router 3")
-	t.Logf("  Expected: Router 2 (%s) remains PRIMARY (stability preferred)", subRouter2.Hostname())
-	t.Logf("  Expected: Routers 1 & 3 are both STANDBY")
+	t.Logf("  Expected: Router 3 (%s) becomes PRIMARY (oldest control session)", subRouter3.Hostname())
+	t.Logf("  Expected: Routers 1 & 2 are both STANDBY")
 	t.Logf("  Expected: Full HA restored with all 3 routers available")
 
 	r3Node := MustFindNode(subRouter3.Hostname(), nodes)
@@ -1301,11 +1305,11 @@ func TestHASubnetRouterFailover(t *testing.T) {
 		assert.NoError(c, err)
 		assert.Len(c, nodes, 6)
 		require.GreaterOrEqual(t, len(nodes), 3, "need at least 3 nodes to avoid panic")
-		// After router 3 re-approval: Router 2 remains PRIMARY, Routers 1&3 are STANDBY
+		// After router 3 re-approval: Router 3 is PRIMARY, Routers 1&2 are STANDBY
 		// SubnetRoutes should only show routes for PRIMARY node (actively serving)
 		requireNodeRouteCountWithCollect(c, nodes[0], 1, 1, 0) // Router 1: STANDBY (available, approved, but not serving)
-		requireNodeRouteCountWithCollect(c, nodes[1], 1, 1, 1) // Router 2: PRIMARY (available, approved, and serving)
-		requireNodeRouteCountWithCollect(c, nodes[2], 1, 1, 0) // Router 3: STANDBY (available, approved, but not serving)
+		requireNodeRouteCountWithCollect(c, nodes[1], 1, 1, 0) // Router 2: STANDBY (available, approved, but not serving)
+		requireNodeRouteCountWithCollect(c, nodes[2], 1, 1, 1) // Router 3: PRIMARY (available, approved, and serving)
 	}, propagationTime, 200*time.Millisecond, "Waiting for route state after router 3 re-approval")
 
 	// Validate primary routes table state after router 3 re-approval
@@ -1316,9 +1320,9 @@ func TestHASubnetRouterFailover(t *testing.T) {
 			types.NodeID(mustParseID(MustFindNode(subRouter3.Hostname(), nodes).Id)): {pref},
 		},
 		PrimaryRoutes: map[string]types.NodeID{
-			pref.String(): types.NodeID(mustParseID(MustFindNode(subRouter2.Hostname(), nodes).Id)),
+			pref.String(): types.NodeID(mustParseID(MustFindNode(subRouter3.Hostname(), nodes).Id)),
 		},
-	}, "Router 2 should remain primary after router 3 re-approval")
+	}, "Router 3 should be primary after router 3 re-approval")
 
 	checkFailureAndPrintRoutes(t, client)
 }
@@ -2167,6 +2171,38 @@ func MustFindNode(hostname string, nodes []*clientv1.Node) *clientv1.Node {
 	panic("node not found")
 }
 
+// orderRouterSessions gives each router after the first a new control
+// session, one after another, so the oldest session, which the primary
+// election prefers, follows the order given. Clients log in concurrently,
+// which leaves that order to chance.
+func orderRouterSessions(t *testing.T, headscale ControlServer, routers ...TailscaleClient) {
+	t.Helper()
+
+	waitOnline := func(r TailscaleClient, online bool) {
+		t.Helper()
+		require.EventuallyWithT(t, func(c *assert.CollectT) {
+			nodes, err := headscale.ListNodes()
+			assert.NoError(c, err)
+			assert.Equal(c, online, MustFindNode(r.Hostname(), nodes).Online)
+		}, integrationutil.HAConvergeTimeout, 500*time.Millisecond,
+			"%s should have online=%t", r.Hostname(), online)
+	}
+
+	for _, r := range routers[1:] {
+		require.NoError(t, r.Down())
+	}
+
+	// A session ends only once the server's reconnect grace has passed.
+	for _, r := range routers[1:] {
+		waitOnline(r, false)
+	}
+
+	for _, r := range routers[1:] {
+		require.NoError(t, r.Up())
+		waitOnline(r, true)
+	}
+}
+
 // TestAutoApproveMultiNetwork tests auto approving of routes
 // by setting up two networks where network1 has three subnet
 // routers:
@@ -3692,6 +3728,8 @@ func TestHASubnetRouterPingFailover(t *testing.T) {
 	subRouter2 := allClients[1]
 	client := allClients[2]
 
+	orderRouterSessions(t, headscale, subRouter1, subRouter2)
+
 	t.Logf("Router 1: %s, Router 2: %s, Client: %s",
 		subRouter1.Hostname(), subRouter2.Hostname(), client.Hostname())
 
@@ -3828,32 +3866,31 @@ func TestHASubnetRouterPingFailover(t *testing.T) {
 	_, _, err = subRouter1.Execute(iptablesDel)
 	require.NoError(t, err, "failed to remove iptables rule")
 
-	// Wait for the prober to detect recovery.
+	// Wait for the prober to detect recovery. Router 1 kept its control
+	// session, the older one, so it takes the route back.
 	assert.EventuallyWithT(t, func(c *assert.CollectT) {
 		pr, err := headscale.PrimaryRoutes()
 		assert.NoError(c, err)
 
-		// Router 1 should be healthy again but NOT primary (no flapping).
 		assert.Equal(c, map[string]types.NodeID{
-			pref.String(): nodeID2,
-		}, pr.PrimaryRoutes, "router 2 should remain primary (no flapping)")
+			pref.String(): nodeID1,
+		}, pr.PrimaryRoutes, "router 1 should be primary again")
 
 		assert.Empty(c, pr.UnhealthyNodes,
 			"no nodes should be unhealthy after recovery")
-	}, propagationTime, 1*time.Second, "waiting for recovery without flapping")
+	}, propagationTime, 1*time.Second, "waiting for recovery")
 
-	// Traffic should still go through router 2 (stability).
 	assert.EventuallyWithT(t, func(c *assert.CollectT) {
 		tr, err := client.Traceroute(webip)
 		assert.NoError(c, err)
 
-		ip, err := subRouter2.IPv4()
+		ip, err := subRouter1.IPv4()
 		if !assert.NoError(c, err) {
 			return
 		}
 
 		assertTracerouteViaIPWithCollect(c, tr, ip)
-	}, propagationTime, 200*time.Millisecond, "traceroute should still go through router 2 after recovery")
+	}, propagationTime, 200*time.Millisecond, "traceroute should go through router 1 after recovery")
 }
 
 // TestHASubnetRouterFailoverBothOffline reproduces issue #3203:
@@ -3932,6 +3969,8 @@ func TestHASubnetRouterFailoverBothOffline(t *testing.T) {
 	subRouter2 := allClients[1]
 	client := allClients[2]
 
+	orderRouterSessions(t, headscale, subRouter1, subRouter2)
+
 	t.Logf("Router 1: %s, Router 2: %s, Client: %s",
 		subRouter1.Hostname(), subRouter2.Hostname(), client.Hostname())
 
@@ -3970,7 +4009,7 @@ func TestHASubnetRouterFailoverBothOffline(t *testing.T) {
 	nodeID1 := types.NodeID(mustParseID(MustFindNode(subRouter1.Hostname(), nodes).Id))
 	nodeID2 := types.NodeID(mustParseID(MustFindNode(subRouter2.Hostname(), nodes).Id))
 
-	// Sanity: r1 starts as primary (lower NodeID).
+	// Sanity: r1 starts as primary (older session).
 	assert.EventuallyWithT(t, func(c *assert.CollectT) {
 		pr, err := headscale.PrimaryRoutes()
 		assert.NoError(c, err)
@@ -4147,6 +4186,8 @@ func TestHASubnetRouterFailoverBothOfflineCablePull(t *testing.T) {
 	subRouter2 := allClients[1]
 	client := allClients[2]
 
+	orderRouterSessions(t, headscale, subRouter1, subRouter2)
+
 	for _, r := range []TailscaleClient{subRouter1, subRouter2} {
 		_, _, err = r.Execute([]string{
 			"tailscale", "set", "--advertise-routes=" + pref.String(),
@@ -4291,7 +4332,7 @@ func TestHASubnetRouterFailoverBothOfflineCablePull(t *testing.T) {
 // disconnect leaves the long-poll TCP half-open at the peer.
 //
 // Phases:
-//  1. r1 starts as primary (lowest NodeID).
+//  1. r1 starts as primary (older session).
 //  2. r1 alone fails and recovers — failover to r2, then traffic
 //     resumes when r1 returns.
 //  3. r2 alone fails and recovers — failover, then traffic resumes.
@@ -4299,9 +4340,10 @@ func TestHASubnetRouterFailoverBothOfflineCablePull(t *testing.T) {
 //     4a. r1 down → r2 promoted.
 //     4b. r2 down → primary must NOT flap to offline r1.
 //     4c. r2 up   → r2 primary again, traffic resumes.
+//     4d. r1 up   → the older session is primary.
 //  5. Simultaneous dual failure.
-//     5a. r1 + r2 down → primary must NOT flap to offline r1.
-//     5b. both up → primary stays r2, traffic resumes.
+//     5a. r1 + r2 down → primary must not move.
+//     5b. both up → the older session is primary, traffic resumes.
 //
 // The no-flap assertions in 4b and 5a are the regression barriers
 // for #3203. Phases 2/3 are functional checks (failover works,
@@ -4371,6 +4413,8 @@ func TestHASubnetRouterFailoverDockerDisconnect(t *testing.T) {
 	subRouter2 := allClients[1]
 	client := allClients[2]
 
+	orderRouterSessions(t, headscale, subRouter1, subRouter2)
+
 	for _, r := range []TailscaleClient{subRouter1, subRouter2} {
 		_, _, err = r.Execute([]string{
 			"tailscale", "set", "--advertise-routes=" + pref.String(),
@@ -4427,8 +4471,7 @@ func TestHASubnetRouterFailoverDockerDisconnect(t *testing.T) {
 	}
 
 	// requirePrimaryStable asserts primary == want for the entire
-	// window. Catches transient flaps and verifies anti-flap on
-	// prev-primary return.
+	// window. Catches transient flaps.
 	requirePrimaryStable := func(want types.NodeID, window time.Duration, msg string) {
 		t.Helper()
 		require.Never(t, func() bool {
@@ -4443,8 +4486,47 @@ func TestHASubnetRouterFailoverDockerDisconnect(t *testing.T) {
 		}, window, 1*time.Second, msg)
 	}
 
+	// requireOlderSessionPrimary blocks until both routers are online and
+	// healthy and the one with the older control session is primary, and
+	// returns it. Whether a cable pull ends a session depends on when the
+	// server notices the dead connection, so the test reads the order.
+	requireOlderSessionPrimary := func(msg string) types.NodeID {
+		t.Helper()
+
+		var want types.NodeID
+
+		assert.EventuallyWithT(t, func(c *assert.CollectT) {
+			store, err := headscale.DebugNodeStore()
+			if !assert.NoError(c, err) {
+				return
+			}
+
+			r1, r2 := store[nodeID1], store[nodeID2]
+			for _, n := range []*types.Node{&r1, &r2} {
+				assert.True(c, n.Online(), "%s: node %d online", msg, n.ID)
+				assert.False(c, n.Unhealthy, "%s: node %d healthy", msg, n.ID)
+			}
+
+			want = nodeID1
+			if r2.SessionStart.Before(r1.SessionStart) {
+				want = nodeID2
+			}
+
+			pr, err := headscale.PrimaryRoutes()
+			if !assert.NoError(c, err) {
+				return
+			}
+
+			assert.Equal(c, map[string]types.NodeID{
+				pref.String(): want,
+			}, pr.PrimaryRoutes, msg)
+		}, propagationTime, 1*time.Second, msg)
+
+		return want
+	}
+
 	// ============================================================
-	// Phase 1: initial state — r1 (lowest NodeID) is primary.
+	// Phase 1: initial state — r1 (older session) is primary.
 	// ============================================================
 	t.Log("=== Phase 1: initial state — r1 should be primary. ===")
 	requirePrimary(nodeID1, "phase 1: r1 primary at start")
@@ -4535,28 +4617,31 @@ func TestHASubnetRouterFailoverDockerDisconnect(t *testing.T) {
 	requirePrimary(nodeID2, "phase 4c: r2 primary after reconnect")
 	requireTrafficWorks("phase 4c: client reaches webservice via r2 after recovery")
 
-	t.Log("=== Phase 4d: reconnect r1, traffic should still flow. ===")
+	t.Log("=== Phase 4d: reconnect r1, the older session should be primary. ===")
 	require.NoError(t, subRouter1.ReconnectToNetwork(usernet1),
 		"phase 4d: docker reconnect r1")
+
+	prev := requireOlderSessionPrimary("phase 4d: older session primary after r1 returns")
+
 	requireTrafficWorks("phase 4d: client still reaches webservice")
 
 	// ============================================================
 	// Phase 5: simultaneous dual failure (whole-segment outage).
-	// prev going in is r2 — primary must not flap to offline r1.
+	// The primary going in must keep the route while both are down.
 	// ============================================================
 	t.Log("=== Phase 5a: cable-pull r1 and r2 simultaneously. ===")
 	require.NoError(t, subRouter1.DisconnectFromNetwork(usernet1),
 		"phase 5a: docker disconnect r1")
 	require.NoError(t, subRouter2.DisconnectFromNetwork(usernet1),
 		"phase 5a: docker disconnect r2")
-	requirePrimaryStable(nodeID2, flapWindow,
-		"phase 5a: primary must not flap to offline r1 (issue #3203)")
+	requirePrimaryStable(prev, flapWindow,
+		"phase 5a: primary must not move while both routers are down (issue #3203)")
 
-	t.Log("=== Phase 5b: reconnect both, r2 should remain primary. ===")
+	t.Log("=== Phase 5b: reconnect both, the older session should be primary. ===")
 	require.NoError(t, subRouter1.ReconnectToNetwork(usernet1),
 		"phase 5b: docker reconnect r1")
 	require.NoError(t, subRouter2.ReconnectToNetwork(usernet1),
 		"phase 5b: docker reconnect r2")
-	requirePrimary(nodeID2, "phase 5b: r2 primary after both reconnect")
-	requireTrafficWorks("phase 5b: client reaches webservice via r2")
+	requireOlderSessionPrimary("phase 5b: older session primary after both reconnect")
+	requireTrafficWorks("phase 5b: client reaches webservice")
 }