diff --git a/src/node/handlers/dispatch.rs b/src/node/handlers/dispatch.rs
index 4f9f9d9..a0909d0 100644
--- a/src/node/handlers/dispatch.rs
+++ b/src/node/handlers/dispatch.rs
@@ -109,13 +109,29 @@ impl Node {
         }
 
         // MMP teardown log (before we drop the peer)
+        let peer_name = self.peer_aliases.get(node_addr)
+            .cloned()
+            .unwrap_or_else(|| peer.identity().short_npub());
         if let Some(mmp) = peer.mmp() {
-            let name = self.peer_aliases.get(node_addr)
-                .cloned()
-                .unwrap_or_else(|| peer.identity().short_npub());
-            Self::log_mmp_teardown(&name, mmp);
+            Self::log_mmp_teardown(&peer_name, mmp);
         }
 
+        // Remove any end-to-end session associated with this peer.
+        //
+        // Sessions are tracked separately from peers (self.sessions vs self.peers).
+        // Leaving a stale session alive after removing the peer causes:
+        //   1. check_session_mmp_reports() keeps logging stale "MMP session metrics"
+        //      with frozen counters until purge_idle_sessions() eventually fires.
+        //   2. initiate_session() finds is_established() == true on the stale entry
+        //      and silently returns Ok(()), preventing a new session from being
+        //      established even after the link layer reconnects successfully.
+        if let Some(session_entry) = self.sessions.remove(node_addr)
+            && let Some(mmp) = session_entry.mmp()
+        {
+            Self::log_session_mmp_teardown(&peer_name, mmp);
+        }
+        self.pending_tun_packets.remove(node_addr);
+
         let link_id = peer.link_id();
         let transport_id = peer.transport_id();
 
diff --git a/src/node/lifecycle.rs b/src/node/lifecycle.rs
index e92c8dd..b88746b 100644
--- a/src/node/lifecycle.rs
+++ b/src/node/lifecycle.rs
@@ -17,15 +17,29 @@ impl Node {
     /// For each peer configured with AutoConnect policy, creates a link and
     /// peer entry, then starts the Noise handshake by sending the first message.
     pub(super) async fn initiate_peer_connections(&mut self) {
-        // Build display name map from all configured peers (alias or short npub)
-        for peer_config in self.config.peers() {
-            if let Ok(identity) = PeerIdentity::from_npub(&peer_config.npub) {
-                let name = peer_config
-                    .alias
-                    .clone()
-                    .unwrap_or_else(|| identity.short_npub());
-                self.peer_aliases.insert(*identity.node_addr(), name);
-            }
+        // Build display name map from all configured peers (alias or short npub),
+        // and pre-seed the identity cache from each peer's npub so that TUN packets
+        // addressed to a configured peer can be dispatched (and trigger session
+        // initiation) immediately on startup — without waiting for the link-layer
+        // handshake to complete first.
+        let peer_identities: Vec<(PeerIdentity, Option<String>)> = self
+            .config
+            .peers()
+            .iter()
+            .filter_map(|pc| {
+                PeerIdentity::from_npub(&pc.npub)
+                    .ok()
+                    .map(|id| (id, pc.alias.clone()))
+            })
+            .collect();
+
+        for (identity, alias) in peer_identities {
+            let name = alias.unwrap_or_else(|| identity.short_npub());
+            self.peer_aliases.insert(*identity.node_addr(), name);
+            // Pre-seed identity cache. The parity may be wrong (npub is x-only)
+            // but will be corrected to the real value when the peer is promoted
+            // after a successful Noise handshake.
+            self.register_identity(*identity.node_addr(), identity.pubkey_full());
         }
 
         // Collect peer configs to avoid borrow conflicts
diff --git a/src/node/retry.rs b/src/node/retry.rs
index c5354da..9473796 100644
--- a/src/node/retry.rs
+++ b/src/node/retry.rs
@@ -129,6 +129,12 @@ impl Node {
     ///
     /// Looks up the peer in auto-connect config and checks `auto_reconnect`.
     /// If enabled, feeds the peer into the retry system with unlimited retries.
+    ///
+    /// If a retry entry already exists (e.g. from a previous failed handshake
+    /// attempt during an earlier reconnect cycle), the existing retry count is
+    /// preserved and incremented rather than reset to zero. This ensures
+    /// exponential backoff accumulates across repeated link-dead events instead
+    /// of resetting to the base interval on every peer removal.
     pub(super) fn schedule_reconnect(&mut self, node_addr: NodeAddr, now_ms: u64) {
         // Find peer in auto-connect config
         let peer_config = self
@@ -155,6 +161,24 @@ impl Node {
 
         let base_interval_ms = self.config.node.retry.base_interval_secs * 1000;
         let max_backoff_ms = self.config.node.retry.max_backoff_secs * 1000;
+        let peer_name = self.peer_display_name(&node_addr);
+
+        // If we already have accumulated backoff from previous failed attempts,
+        // preserve and bump it rather than resetting to zero. This prevents the
+        // exponential backoff from being discarded on each link-dead cycle.
+        if let Some(state) = self.retry_pending.get_mut(&node_addr) {
+            state.reconnect = true;
+            state.retry_count += 1;
+            let delay = state.backoff_ms(base_interval_ms, max_backoff_ms);
+            state.retry_after_ms = now_ms + delay;
+            info!(
+                peer = %peer_name,
+                retry = state.retry_count,
+                delay_secs = delay / 1000,
+                "Scheduling auto-reconnect after link-dead removal (backoff preserved)"
+            );
+            return;
+        }
 
         let mut state = RetryState::new(pc);
         state.reconnect = true;
@@ -162,7 +186,7 @@ impl Node {
         state.retry_after_ms = now_ms + delay;
 
         info!(
-            peer = %self.peer_display_name(&node_addr),
+            peer = %peer_name,
             delay_secs = delay / 1000,
             "Scheduling auto-reconnect after link-dead removal"
         );
diff --git a/src/node/tests/disconnect.rs b/src/node/tests/disconnect.rs
index a5ad34f..567543d 100644
--- a/src/node/tests/disconnect.rs
+++ b/src/node/tests/disconnect.rs
@@ -228,6 +228,94 @@ async fn test_disconnect_chain_partition() {
     cleanup_nodes(&mut nodes).await;
 }
 
+/// Removing a peer via disconnect must also remove the associated end-to-end session.
+///
+/// Regression test for issue #5: `remove_active_peer` previously left the
+/// `SessionEntry` alive in `self.sessions` after evicting the peer from
+/// `self.peers`. This caused:
+///   1. Stale "MMP session metrics" logs with frozen counters until
+///      `purge_idle_sessions` eventually fired (up to idle_timeout_secs later).
+///   2. `initiate_session` silently returning `Ok(())` on the stale Established
+///      entry's guard check, preventing a new session from being created even
+///      after the link layer reconnected successfully.
+#[tokio::test]
+async fn test_disconnect_clears_session() {
+    use crate::identity::Identity;
+    use crate::node::session::{EndToEndState, SessionEntry};
+    use crate::noise::HandshakeState;
+
+    // Two-node topology: 0 -- 1.
+    let edges = vec![(0, 1)];
+    let mut nodes = run_tree_test(2, &edges, false).await;
+    verify_tree_convergence(&nodes);
+
+    let node0_addr = *nodes[0].node.node_addr();
+    let node1_addr = *nodes[1].node.node_addr();
+
+    // Inject a synthetic Established session entry into node 1's session table
+    // to simulate the state after a completed XK handshake with node 0.
+    let remote_identity = Identity::generate();
+    {
+        let our_identity = nodes[1].node.identity().clone();
+
+        let mut initiator = HandshakeState::new_initiator(
+            our_identity.keypair(),
+            remote_identity.pubkey_full(),
+        );
+        let mut responder = HandshakeState::new_responder(remote_identity.keypair());
+        let mut init_epoch = [0u8; 8];
+        rand::Rng::fill_bytes(&mut rand::rng(), &mut init_epoch);
+        initiator.set_local_epoch(init_epoch);
+        let mut resp_epoch = [0u8; 8];
+        rand::Rng::fill_bytes(&mut rand::rng(), &mut resp_epoch);
+        responder.set_local_epoch(resp_epoch);
+        let msg1 = initiator.write_message_1().unwrap();
+        responder.read_message_1(&msg1).unwrap();
+        let msg2 = responder.write_message_2().unwrap();
+        initiator.read_message_2(&msg2).unwrap();
+        let session = initiator.into_session().unwrap();
+
+        let entry = SessionEntry::new(
+            node0_addr,
+            remote_identity.pubkey_full(),
+            EndToEndState::Established(session),
+            1_000,
+            true,
+        );
+        nodes[1].node.sessions.insert(node0_addr, entry);
+    }
+
+    assert_eq!(nodes[1].node.session_count(), 1, "Session should exist before disconnect");
+    assert_eq!(nodes[1].node.peer_count(), 1, "Peer should exist before disconnect");
+
+    // Node 0 sends Disconnect to node 1.
+    let disconnect = crate::protocol::Disconnect::new(DisconnectReason::Shutdown);
+    nodes[0]
+        .node
+        .send_encrypted_link_message(&node1_addr, &disconnect.encode())
+        .await
+        .expect("Failed to send disconnect");
+
+    tokio::time::sleep(Duration::from_millis(50)).await;
+    process_available_packets(&mut nodes).await;
+
+    // Peer must be gone.
+    assert_eq!(
+        nodes[1].node.peer_count(), 0,
+        "Peer should be removed after disconnect"
+    );
+
+    // Session must also be gone — core regression check for issue #5.
+    // Before the fix, session_count() would still be 1 here because
+    // remove_active_peer didn't remove self.sessions[node0_addr].
+    assert_eq!(
+        nodes[1].node.session_count(), 0,
+        "Session must be cleaned up when peer is removed (regression: issue #5)"
+    );
+
+    cleanup_nodes(&mut nodes).await;
+}
+
 /// Verify that different disconnect reasons are handled correctly.
 ///
 /// Sends each reason code and verifies the peer is removed regardless.
diff --git a/src/node/tests/unit.rs b/src/node/tests/unit.rs
index ea8d5a4..98eeaaf 100644
--- a/src/node/tests/unit.rs
+++ b/src/node/tests/unit.rs
@@ -702,6 +702,90 @@ fn test_schedule_retry_skips_connected_peer() {
     );
 }
 
+/// Test that schedule_reconnect preserves accumulated backoff across link-dead cycles.
+///
+/// Regression test for issue #5: previously `schedule_reconnect` always created a
+/// fresh `RetryState` with `retry_count=0`, discarding any backoff accumulated by
+/// prior failed handshake attempts. On repeated link-dead evictions the node would
+/// restart exponential backoff from the base interval every time instead of
+/// continuing to back off.
+#[test]
+fn test_schedule_reconnect_preserves_backoff() {
+    let peer_identity = Identity::generate();
+    let peer_npub = peer_identity.npub();
+    let peer_node_addr = *PeerIdentity::from_npub(&peer_npub).unwrap().node_addr();
+
+    let mut config = Config::new();
+    config.peers.push(crate::config::PeerConfig::new(
+        peer_npub,
+        "udp",
+        "10.0.0.2:2121",
+    ));
+
+    let mut node = Node::new(config).unwrap();
+
+    // Simulate two stale handshake timeouts incrementing the retry count.
+    node.schedule_retry(peer_node_addr, 1_000);  // count=1, delay=10s
+    node.schedule_retry(peer_node_addr, 11_000); // count=2, delay=20s
+    {
+        let state = node.retry_pending.get(&peer_node_addr).unwrap();
+        assert_eq!(state.retry_count, 2, "Two failures should yield count=2");
+    }
+
+    // Now simulate a link-dead removal triggering schedule_reconnect.
+    // The existing retry entry (count=2) should be preserved and bumped to 3,
+    // NOT reset to 0 as it was before the fix.
+    node.schedule_reconnect(peer_node_addr, 31_000);
+
+    let state = node.retry_pending.get(&peer_node_addr).unwrap();
+    assert!(
+        state.reconnect,
+        "Entry should be marked as reconnect"
+    );
+    assert_eq!(
+        state.retry_count, 3,
+        "schedule_reconnect should increment existing count (was 2), not reset to 0 (regression: issue #5)"
+    );
+
+    // With count=3, backoff should be 5s * 2^3 = 40s.
+    let base_ms = node.config.node.retry.base_interval_secs * 1000;
+    let max_ms = node.config.node.retry.max_backoff_secs * 1000;
+    let expected_delay = state.backoff_ms(base_ms, max_ms);
+    assert_eq!(
+        state.retry_after_ms, 31_000 + expected_delay,
+        "retry_after_ms should reflect count=3 backoff"
+    );
+}
+
+/// Test that schedule_reconnect on a fresh peer (no prior retry entry) starts at count=0.
+#[test]
+fn test_schedule_reconnect_fresh_state() {
+    let peer_identity = Identity::generate();
+    let peer_npub = peer_identity.npub();
+    let peer_node_addr = *PeerIdentity::from_npub(&peer_npub).unwrap().node_addr();
+
+    let mut config = Config::new();
+    config.peers.push(crate::config::PeerConfig::new(
+        peer_npub,
+        "udp",
+        "10.0.0.2:2121",
+    ));
+
+    let mut node = Node::new(config).unwrap();
+
+    // No prior retry entry — first reconnect should use base delay.
+    node.schedule_reconnect(peer_node_addr, 1_000);
+
+    let state = node.retry_pending.get(&peer_node_addr).unwrap();
+    assert!(state.reconnect, "Entry should be marked as reconnect");
+    assert_eq!(state.retry_count, 0, "Fresh reconnect should start at count=0");
+    // Base delay: 5s * 2^0 = 5s
+    let base_ms = node.config.node.retry.base_interval_secs * 1000;
+    let max_ms = node.config.node.retry.max_backoff_secs * 1000;
+    let expected_delay = state.backoff_ms(base_ms, max_ms);
+    assert_eq!(state.retry_after_ms, 1_000 + expected_delay);
+}
+
 /// Test that promote_connection clears retry_pending.
 #[test]
 fn test_promote_clears_retry_pending() {