diff --git a/CHANGELOG.md b/CHANGELOG.md
index 60fa077..0a084f8 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -149,6 +149,18 @@ with v0.3.x peers.
   passes 70/70. The constant and surrounding draw/redraw machinery
   are kept in place pending diagnosis of why XX cutover state cleanup
   doesn't absorb variable-interval rekeys the way IK does.
+- Receive hot path: removed two per-packet copies. New borrowed
+  `SessionDatagramRef` decoder is used in the forwarding handler so
+  local delivery and coordinate-cache warming no longer allocate or
+  copy the session payload; the owned `SessionDatagram` is materialized
+  only when re-encoding for the next hop. Owned `SessionDatagram::
+  decode` is reimplemented as `Ref::decode + into_owned`, so the two
+  decoders cannot drift. On Linux + macOS the `recvmmsg` / `recvmsg_x`
+  receive loop now moves each filled slot buffer into `ReceivedPacket`
+  via `mem::replace` instead of cloning it, and `TransportAddr` is
+  formatted directly from the `SocketAddr` without an intermediate
+  `String`. Focused decode bench: ref 1.6 ns/op vs owned 34.7 ns/op
+  (21.4x).
 
 ### Fixed
 
@@ -191,6 +203,20 @@ with v0.3.x peers.
   after the test's previous one-shot grep gave up, producing a
   pre-existing flake on next-branch CI. Success-path cost is
   unchanged — the helper returns as soon as the pattern appears.
+- Nostr-discovered NAT-traversal events (`BootstrapEvent::Established`
+  and `BootstrapEvent::Failed`) for peers that are already connected
+  or actively handshaking are now short-circuited at the
+  `poll_nostr_discovery` dispatch sites before any cooldown
+  bookkeeping or fallback retry scheduling runs. Stale `Failed` events
+  previously poisoned the per-peer failure-state cooldown of healthy
+  peers and could trigger redundant retraversal attempts via
+  `schedule_retry` / `try_peer_addresses`; stale `Established`
+  handoffs could attempt to adopt a second socket against a live
+  connection. A defense-in-depth guard was added to
+  `adopt_established_traversal` so the same invariant holds if a
+  future caller bypasses the outer dispatch check. As a side benefit,
+  narrows a cooldown-poisoning vector previously available to an
+  attacker injecting stale failure events for an active peer.
 
 ## [0.3.0] - 2026-05-11
 
diff --git a/src/cache/coord_cache.rs b/src/cache/coord_cache.rs
index 84fda86..a0c914a 100644
--- a/src/cache/coord_cache.rs
+++ b/src/cache/coord_cache.rs
@@ -205,6 +205,40 @@ impl CoordCache {
         self.entries.clear();
     }
 
+    /// Drop entries whose cached destination ancestry contains the given
+    /// `NodeAddr`.
+    ///
+    /// Used at parent-position-change sites: when our own position in the
+    /// tree changes, destinations downstream of us (whose cached coordinates
+    /// embed our previous prefix) have stale path information and must be
+    /// re-learned. Entries whose ancestry does not include `node_addr` are
+    /// unaffected by the local position change and are retained.
+    ///
+    /// Returns the count of entries removed.
+    pub fn invalidate_via_node(&mut self, node_addr: &NodeAddr) -> usize {
+        let len_before = self.entries.len();
+        self.entries
+            .retain(|_, entry| !entry.coords().contains(node_addr));
+        len_before - self.entries.len()
+    }
+
+    /// Drop entries whose cached destination `root_id` differs from
+    /// `current_root`.
+    ///
+    /// Used at root-change sites (become_root, root handover via
+    /// TreeAnnounce). `find_next_hop` returns `None` for any destination
+    /// whose root does not match the local root, so entries from a stale
+    /// root cannot route and would otherwise occupy cache slots until
+    /// TTL expiry.
+    ///
+    /// Returns the count of entries removed.
+    pub fn invalidate_other_roots(&mut self, current_root: &NodeAddr) -> usize {
+        let len_before = self.entries.len();
+        self.entries
+            .retain(|_, entry| entry.coords().root_id() == current_root);
+        len_before - self.entries.len()
+    }
+
     /// Evict one entry (expired first, then LRU).
     fn evict_one(&mut self, current_time_ms: u64) {
         // First try to evict an expired entry
@@ -507,4 +541,101 @@ mod tests {
         assert_eq!(stats.expired, 0);
         assert_eq!(stats.avg_age_ms, 0);
     }
+
+    // ===== Surgical invalidation tests =====
+
+    #[test]
+    fn test_invalidate_via_node_at_self_depth() {
+        // Entry whose own NodeAddr (depth 0) is the invalidation target.
+        let mut cache = CoordCache::new(100, 1000);
+        let target = make_node_addr(1);
+
+        cache.insert(target, make_coords(&[1, 0]), 0);
+        assert_eq!(cache.len(), 1);
+
+        let removed = cache.invalidate_via_node(&target);
+        assert_eq!(removed, 1);
+        assert_eq!(cache.len(), 0);
+    }
+
+    #[test]
+    fn test_invalidate_via_node_interior() {
+        // Entry whose ancestry contains the target in the interior of the path.
+        let mut cache = CoordCache::new(100, 1000);
+        let dest = make_node_addr(5);
+        // Path: 5 -> 3 -> 1 -> 0 (root). Target 3 appears at depth 1.
+        cache.insert(dest, make_coords(&[5, 3, 1, 0]), 0);
+
+        let removed = cache.invalidate_via_node(&make_node_addr(3));
+        assert_eq!(removed, 1);
+        assert_eq!(cache.len(), 0);
+    }
+
+    #[test]
+    fn test_invalidate_via_node_absent() {
+        // Entry whose ancestry does NOT contain the target must be retained.
+        let mut cache = CoordCache::new(100, 1000);
+        let dest = make_node_addr(5);
+        cache.insert(dest, make_coords(&[5, 3, 1, 0]), 0);
+
+        let removed = cache.invalidate_via_node(&make_node_addr(99));
+        assert_eq!(removed, 0);
+        assert_eq!(cache.len(), 1);
+        assert!(cache.contains(&dest, 0));
+    }
+
+    #[test]
+    fn test_invalidate_via_node_empty_cache() {
+        let mut cache = CoordCache::new(100, 1000);
+        let removed = cache.invalidate_via_node(&make_node_addr(1));
+        assert_eq!(removed, 0);
+        assert_eq!(cache.len(), 0);
+    }
+
+    #[test]
+    fn test_invalidate_other_roots_current_root_kept() {
+        let mut cache = CoordCache::new(100, 1000);
+        // Entries rooted at addr(0)
+        cache.insert(make_node_addr(1), make_coords(&[1, 0]), 0);
+        cache.insert(make_node_addr(2), make_coords(&[2, 0]), 0);
+
+        let removed = cache.invalidate_other_roots(&make_node_addr(0));
+        assert_eq!(removed, 0);
+        assert_eq!(cache.len(), 2);
+    }
+
+    #[test]
+    fn test_invalidate_other_roots_different_root_dropped() {
+        let mut cache = CoordCache::new(100, 1000);
+        // Three entries rooted at addr(0), one rooted at addr(9)
+        cache.insert(make_node_addr(1), make_coords(&[1, 0]), 0);
+        cache.insert(make_node_addr(2), make_coords(&[2, 0]), 0);
+        cache.insert(make_node_addr(3), make_coords(&[3, 0]), 0);
+        cache.insert(make_node_addr(4), make_coords(&[4, 9]), 0);
+
+        let removed = cache.invalidate_other_roots(&make_node_addr(0));
+        assert_eq!(removed, 1);
+        assert_eq!(cache.len(), 3);
+        assert!(!cache.contains(&make_node_addr(4), 0));
+        assert!(cache.contains(&make_node_addr(1), 0));
+    }
+
+    #[test]
+    fn test_invalidate_other_roots_all_match() {
+        let mut cache = CoordCache::new(100, 1000);
+        cache.insert(make_node_addr(1), make_coords(&[1, 0]), 0);
+        cache.insert(make_node_addr(2), make_coords(&[2, 0]), 0);
+
+        let removed = cache.invalidate_other_roots(&make_node_addr(0));
+        assert_eq!(removed, 0);
+        assert_eq!(cache.len(), 2);
+    }
+
+    #[test]
+    fn test_invalidate_other_roots_empty_cache() {
+        let mut cache = CoordCache::new(100, 1000);
+        let removed = cache.invalidate_other_roots(&make_node_addr(0));
+        assert_eq!(removed, 0);
+        assert_eq!(cache.len(), 0);
+    }
 }
diff --git a/src/control/listening.rs b/src/control/listening.rs
index 6b00ac6..2f271ab 100644
--- a/src/control/listening.rs
+++ b/src/control/listening.rs
@@ -15,7 +15,9 @@
 //! See `docs/design/fips-security.md` for the operator-side narrative
 //! that motivates the panel.
 
-use std::net::{IpAddr, Ipv6Addr};
+#[cfg(target_os = "linux")]
+use std::net::IpAddr;
+use std::net::Ipv6Addr;
 
 /// Transport protocol of a listening socket.
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
diff --git a/src/discovery/nostr/runtime.rs b/src/discovery/nostr/runtime.rs
index eea46f5..26f3b4b 100644
--- a/src/discovery/nostr/runtime.rs
+++ b/src/discovery/nostr/runtime.rs
@@ -1537,4 +1537,10 @@ impl NostrDiscovery {
         let mut cache = self.advert_cache.write().await;
         cache.insert(npub, advert);
     }
+
+    /// Queue a bootstrap event directly for lifecycle tests without live relays
+    /// or a running traversal task.
+    pub(crate) fn push_event_for_test(&self, event: BootstrapEvent) {
+        let _ = self.event_tx.send(event);
+    }
 }
diff --git a/src/node/handlers/forwarding.rs b/src/node/handlers/forwarding.rs
index 245004b..e96dfc0 100644
--- a/src/node/handlers/forwarding.rs
+++ b/src/node/handlers/forwarding.rs
@@ -12,7 +12,8 @@ use crate::node::session_wire::{
 };
 use crate::node::{Node, NodeError};
 use crate::protocol::{
-    CoordsRequired, MtuExceeded, PathBroken, SessionAck, SessionDatagram, SessionSetup,
+    CoordsRequired, MtuExceeded, PathBroken, SessionAck, SessionDatagram, SessionDatagramRef,
+    SessionSetup,
 };
 use std::time::{Duration, Instant};
 use tracing::{debug, warn};
@@ -30,7 +31,7 @@ impl Node {
     ) {
         self.stats_mut().forwarding.record_received(payload.len());
 
-        let mut datagram = match SessionDatagram::decode(payload) {
+        let datagram_ref = match SessionDatagramRef::decode(payload) {
             Ok(dg) => dg,
             Err(e) => {
                 self.stats_mut()
@@ -41,35 +42,41 @@ impl Node {
             }
         };
 
-        // TTL enforcement: decrement and drop if exhausted
-        if !datagram.decrement_ttl() {
+        // TTL enforcement: decrement for forwarding and drop only if the
+        // received datagram was already exhausted.
+        if datagram_ref.ttl == 0 {
             self.stats_mut()
                 .forwarding
                 .record_ttl_exhausted(payload.len());
             debug!(
-                src = %datagram.src_addr,
-                dest = %datagram.dest_addr,
+                src = %datagram_ref.src_addr,
+                dest = %datagram_ref.dest_addr,
                 "SessionDatagram TTL exhausted, dropping"
             );
             return;
         }
+        let forwarded_ttl = datagram_ref.ttl - 1;
 
         // Coordinate cache warming from plaintext session-layer headers
-        self.try_warm_coord_cache(&datagram);
+        self.try_warm_coord_cache_ref(&datagram_ref);
 
-        // Local delivery: dispatch to session layer handlers
-        if datagram.dest_addr == *self.node_addr() {
+        // Local delivery: dispatch to session layer handlers without
+        // materializing an owned SessionDatagram payload Vec.
+        if datagram_ref.dest_addr == *self.node_addr() {
             self.stats_mut().forwarding.record_delivered(payload.len());
             self.handle_session_payload(
-                &datagram.src_addr,
-                &datagram.payload,
-                datagram.path_mtu,
+                &datagram_ref.src_addr,
+                datagram_ref.payload,
+                datagram_ref.path_mtu,
                 incoming_ce,
             )
             .await;
             return;
         }
 
+        let mut datagram = datagram_ref.into_owned();
+        datagram.ttl = forwarded_ttl;
+
         // Find next hop toward destination
         let next_hop_addr = match self.find_next_hop(&datagram.dest_addr) {
             Some(peer) => *peer.node_addr(),
@@ -153,8 +160,8 @@ impl Node {
     ///
     /// Decode failures are logged and silently ignored — they don't block
     /// forwarding.
-    fn try_warm_coord_cache(&mut self, datagram: &SessionDatagram) {
-        let prefix = match FspCommonPrefix::parse(&datagram.payload) {
+    fn try_warm_coord_cache_ref(&mut self, datagram: &SessionDatagramRef<'_>) {
+        let prefix = match FspCommonPrefix::parse(datagram.payload) {
             Some(p) => p,
             None => return,
         };
diff --git a/src/node/handlers/mmp.rs b/src/node/handlers/mmp.rs
index ce28da4..d58ab40 100644
--- a/src/node/handlers/mmp.rs
+++ b/src/node/handlers/mmp.rs
@@ -155,7 +155,9 @@ impl Node {
                     warn!(error = %e, "Failed to sign declaration after first-RTT parent eval");
                     return;
                 }
-                self.coord_cache.clear();
+                // Surgical invalidation — see CoordCache::invalidate_via_node doc.
+                self.coord_cache
+                    .invalidate_via_node(self.identity.node_addr());
                 self.reset_discovery_backoff();
                 self.stats_mut().tree.parent_switched += 1;
                 self.stats_mut().tree.parent_switches += 1;
@@ -180,7 +182,9 @@ impl Node {
                     warn!(error = %e, "Failed to sign self-root declaration after first-RTT");
                     return;
                 }
-                self.coord_cache.clear();
+                // Surgical invalidation — see CoordCache::invalidate_other_roots doc.
+                self.coord_cache
+                    .invalidate_other_roots(self.identity.node_addr());
                 self.reset_discovery_backoff();
                 self.stats_mut().tree.parent_switched += 1;
                 self.stats_mut().tree.parent_switches += 1;
diff --git a/src/node/lifecycle.rs b/src/node/lifecycle.rs
index 729cad7..fe7b0ee 100644
--- a/src/node/lifecycle.rs
+++ b/src/node/lifecycle.rs
@@ -122,12 +122,7 @@ impl Node {
         }
 
         // Check if connection already in progress to this peer
-        let already_connecting = self.connections.values().any(|conn| {
-            conn.expected_identity()
-                .map(|id| id.node_addr() == &peer_node_addr)
-                .unwrap_or(false)
-        });
-        if already_connecting {
+        if self.is_connecting_to_peer(&peer_node_addr) {
             debug!(
                 npub = %peer_config.npub,
                 "Connection already in progress, skipping"
@@ -139,6 +134,14 @@ impl Node {
             .await
     }
 
+    fn is_connecting_to_peer(&self, peer_node_addr: &NodeAddr) -> bool {
+        self.connections.values().any(|conn| {
+            conn.expected_identity()
+                .map(|id| id.node_addr() == peer_node_addr)
+                .unwrap_or(false)
+        })
+    }
+
     /// Initiate a connection to a peer on a specific transport and address.
     ///
     /// For connectionless transports (UDP, Ethernet): allocates a link, starts
@@ -451,6 +454,23 @@ impl Node {
             match event {
                 BootstrapEvent::Established { traversal } => {
                     let peer_npub = traversal.peer_npub.clone();
+                    if let Ok(peer_identity) = PeerIdentity::from_npub(&peer_npub) {
+                        let peer_addr = *peer_identity.node_addr();
+                        if self.peers.contains_key(&peer_addr) {
+                            debug!(
+                                peer_npub = %peer_npub,
+                                "Ignoring established NAT traversal for already-connected peer"
+                            );
+                            continue;
+                        }
+                        if self.is_connecting_to_peer(&peer_addr) {
+                            debug!(
+                                peer_npub = %peer_npub,
+                                "Ignoring established NAT traversal while peer handshake is already in progress"
+                            );
+                            continue;
+                        }
+                    }
                     match self.adopt_established_traversal(traversal).await {
                         Ok(_) => {
                             info!(peer_npub = %peer_npub, "Adopted NAT traversal socket");
@@ -467,6 +487,28 @@ impl Node {
                     peer_config,
                     reason,
                 } => {
+                    let peer_identity = match PeerIdentity::from_npub(&peer_config.npub) {
+                        Ok(identity) => identity,
+                        Err(_) => continue,
+                    };
+                    let node_addr = *peer_identity.node_addr();
+                    if self.peers.contains_key(&node_addr) {
+                        debug!(
+                            npub = %peer_config.npub,
+                            error = %reason,
+                            "Ignoring failed NAT traversal for already-connected peer"
+                        );
+                        continue;
+                    }
+                    if self.is_connecting_to_peer(&node_addr) {
+                        debug!(
+                            npub = %peer_config.npub,
+                            error = %reason,
+                            "Ignoring failed NAT traversal while peer handshake is already in progress"
+                        );
+                        continue;
+                    }
+
                     let now_ms = Self::now_ms();
                     let decision = bootstrap.record_traversal_failure(&peer_config.npub, now_ms);
                     if decision.should_warn {
@@ -517,11 +559,6 @@ impl Node {
                         });
                     }
 
-                    let peer_identity = match PeerIdentity::from_npub(&peer_config.npub) {
-                        Ok(identity) => identity,
-                        Err(_) => continue,
-                    };
-
                     if self
                         .try_peer_addresses(&peer_config, peer_identity, false)
                         .await
@@ -530,7 +567,6 @@ impl Node {
                         continue;
                     }
 
-                    let node_addr = *peer_identity.node_addr();
                     self.schedule_retry(node_addr, now_ms);
                     if let Some(cooldown_until_ms) = decision.cooldown_until_ms
                         && let Some(state) = self.retry_pending.get_mut(&node_addr)
@@ -1737,6 +1773,22 @@ impl Node {
         peer_identity: PeerIdentity,
         allow_bootstrap_nat: bool,
     ) -> Result<(), NodeError> {
+        let peer_node_addr = *peer_identity.node_addr();
+        if self.peers.contains_key(&peer_node_addr) {
+            debug!(
+                npub = %peer_config.npub,
+                "Peer already exists, skipping address attempts"
+            );
+            return Ok(());
+        }
+        if self.is_connecting_to_peer(&peer_node_addr) {
+            debug!(
+                npub = %peer_config.npub,
+                "Connection already in progress, skipping address attempts"
+            );
+            return Ok(());
+        }
+
         // Static-first dialing: avoid delaying configured address attempts on
         // advert fetch/network latency.
         let static_addresses = self.static_peer_addresses(peer_config);
@@ -1880,6 +1932,20 @@ impl Node {
             }
         })?;
         let peer_node_addr = *peer_identity.node_addr();
+        if self.peers.contains_key(&peer_node_addr) {
+            debug!(
+                peer_npub = %traversal.peer_npub,
+                "Ignoring NAT traversal handoff for already-connected peer"
+            );
+            return Err(NodeError::PeerAlreadyExists(peer_node_addr));
+        }
+        if self.is_connecting_to_peer(&peer_node_addr) {
+            debug!(
+                peer_npub = %traversal.peer_npub,
+                "Ignoring NAT traversal handoff while peer handshake is already in progress"
+            );
+            return Err(NodeError::PeerAlreadyExists(peer_node_addr));
+        }
 
         self.peer_aliases
             .insert(peer_node_addr, peer_identity.short_npub());
diff --git a/src/node/tests/bootstrap.rs b/src/node/tests/bootstrap.rs
index a051cea..43b66df 100644
--- a/src/node/tests/bootstrap.rs
+++ b/src/node/tests/bootstrap.rs
@@ -133,6 +133,51 @@ async fn test_failed_adopted_traversal_cleans_up_transport() {
     );
 }
 
+#[tokio::test]
+async fn test_adopted_traversal_skips_already_connected_peer() {
+    let mut node = make_node();
+    let (packet_tx, packet_rx) = packet_channel(64);
+    node.packet_tx = Some(packet_tx);
+    node.packet_rx = Some(packet_rx);
+    node.state = NodeState::Running;
+
+    let transport_id = TransportId::new(1);
+    let link_id = LinkId::new(1);
+    let (conn, peer_identity) = make_completed_connection(&mut node, link_id, transport_id, 1_000);
+    let peer_node_addr = *peer_identity.node_addr();
+    node.add_connection(conn).unwrap();
+    node.promote_connection(link_id, peer_identity, 2_000)
+        .unwrap();
+
+    let link_count = node.link_count();
+    let transport_count = node.transport_count();
+
+    let adopted_socket = std::net::UdpSocket::bind("127.0.0.1:0").unwrap();
+    let handoff = EstablishedTraversal::new(
+        "sess-stale",
+        peer_identity.npub(),
+        "127.0.0.1:9".parse().unwrap(),
+        adopted_socket,
+    )
+    .with_transport_name("nostr-stale");
+
+    let result = node.adopt_established_traversal(handoff).await;
+    assert!(
+        matches!(result, Err(NodeError::PeerAlreadyExists(addr)) if addr == peer_node_addr),
+        "stale traversal handoff should be ignored once the peer is already active"
+    );
+    assert_eq!(
+        node.link_count(),
+        link_count,
+        "ignored traversal must not create a duplicate link"
+    );
+    assert_eq!(
+        node.transport_count(),
+        transport_count,
+        "ignored traversal must not leak an adopted transport"
+    );
+}
+
 #[tokio::test]
 async fn test_third_peer_can_handshake_via_adopted_transport_socket() {
     let mut node_a = make_node(); // Existing traversal peer (Alice)
diff --git a/src/node/tests/unit.rs b/src/node/tests/unit.rs
index 3f46e54..d5a8080 100644
--- a/src/node/tests/unit.rs
+++ b/src/node/tests/unit.rs
@@ -1,7 +1,9 @@
 use super::*;
+use crate::discovery::nostr::{BootstrapEvent, NostrDiscovery};
 use crate::peer::PromotionResult;
 use crate::transport::udp::UdpTransport;
 use crate::transport::{TransportHandle, packet_channel};
+use std::sync::Arc;
 
 #[test]
 fn test_node_creation() {
@@ -782,6 +784,135 @@ fn test_schedule_retry_skips_connected_peer() {
     );
 }
 
+#[tokio::test]
+async fn test_try_peer_addresses_skips_connected_peer() {
+    let mut node = make_node();
+    let transport_id = TransportId::new(1);
+    let link_id = LinkId::new(1);
+    let (conn, peer_identity) = make_completed_connection(&mut node, link_id, transport_id, 1000);
+    let peer_config = crate::config::PeerConfig::new(peer_identity.npub(), "udp", "127.0.0.1:9");
+
+    node.add_connection(conn).unwrap();
+    node.promote_connection(link_id, peer_identity, 2000)
+        .unwrap();
+    let link_count = node.link_count();
+    let connection_count = node.connection_count();
+
+    node.try_peer_addresses(&peer_config, peer_identity, true)
+        .await
+        .unwrap();
+
+    assert_eq!(
+        node.link_count(),
+        link_count,
+        "stale retry/traversal fallback must not create a duplicate link"
+    );
+    assert_eq!(
+        node.connection_count(),
+        connection_count,
+        "stale retry/traversal fallback must not create a duplicate handshake"
+    );
+}
+
+#[tokio::test]
+async fn test_try_peer_addresses_skips_connecting_peer() {
+    let mut node = make_node();
+    let peer_identity = make_peer_identity();
+    let peer_config = crate::config::PeerConfig::new(peer_identity.npub(), "udp", "127.0.0.1:9");
+    let pending = PeerConnection::outbound(LinkId::new(1), peer_identity, 1000);
+    node.add_connection(pending).unwrap();
+
+    node.try_peer_addresses(&peer_config, peer_identity, true)
+        .await
+        .unwrap();
+
+    assert_eq!(
+        node.connection_count(),
+        1,
+        "stale retry/traversal fallback must not start a second handshake"
+    );
+    assert_eq!(
+        node.link_count(),
+        0,
+        "stale retry/traversal fallback must not allocate a link while a handshake is pending"
+    );
+}
+
+#[tokio::test]
+async fn test_nostr_traversal_failure_skips_connected_peer() {
+    let mut node = make_node();
+    let transport_id = TransportId::new(1);
+    let link_id = LinkId::new(1);
+    let (conn, peer_identity) = make_completed_connection(&mut node, link_id, transport_id, 1000);
+    node.add_connection(conn).unwrap();
+    node.promote_connection(link_id, peer_identity, 2000)
+        .unwrap();
+
+    let bootstrap = Arc::new(NostrDiscovery::new_for_test());
+    bootstrap.push_event_for_test(BootstrapEvent::Failed {
+        peer_config: crate::config::PeerConfig::new(peer_identity.npub(), "udp", "127.0.0.1:9"),
+        reason: "stale traversal failure".to_string(),
+    });
+    node.nostr_discovery = Some(bootstrap.clone());
+
+    node.poll_nostr_discovery().await;
+
+    assert!(
+        bootstrap.failure_state_snapshot().is_empty(),
+        "stale failures for connected peers must not affect traversal cooldown"
+    );
+    assert!(
+        node.retry_pending.is_empty(),
+        "stale failures for connected peers must not enqueue reconnect attempts"
+    );
+}
+
+#[tokio::test]
+async fn test_nostr_traversal_established_skips_connected_peer() {
+    use crate::discovery::EstablishedTraversal;
+    use std::net::UdpSocket;
+
+    let mut node = make_node();
+    let transport_id = TransportId::new(1);
+    let link_id = LinkId::new(1);
+    let (conn, peer_identity) = make_completed_connection(&mut node, link_id, transport_id, 1000);
+    node.add_connection(conn).unwrap();
+    node.promote_connection(link_id, peer_identity, 2000)
+        .unwrap();
+    let link_count = node.link_count();
+    let connection_count = node.connection_count();
+
+    let bootstrap = Arc::new(NostrDiscovery::new_for_test());
+    let socket = UdpSocket::bind("127.0.0.1:0").expect("bind local UDP socket");
+    let remote_addr = "127.0.0.1:9999".parse().expect("parse remote addr");
+    bootstrap.push_event_for_test(BootstrapEvent::Established {
+        traversal: EstablishedTraversal::new(
+            "test-session",
+            peer_identity.npub(),
+            remote_addr,
+            socket,
+        ),
+    });
+    node.nostr_discovery = Some(bootstrap.clone());
+
+    node.poll_nostr_discovery().await;
+
+    assert_eq!(
+        node.link_count(),
+        link_count,
+        "stale established handoff must not allocate a new link"
+    );
+    assert_eq!(
+        node.connection_count(),
+        connection_count,
+        "stale established handoff must not start a new handshake"
+    );
+    assert!(
+        node.retry_pending.is_empty(),
+        "stale established handoff must not enqueue a reconnect"
+    );
+}
+
 #[tokio::test]
 async fn test_process_pending_retries_drops_expired_entries() {
     let mut node = make_node();
diff --git a/src/node/tree.rs b/src/node/tree.rs
index 4e166b1..e560384 100644
--- a/src/node/tree.rs
+++ b/src/node/tree.rs
@@ -255,7 +255,9 @@ impl Node {
                 warn!(error = %e, "Failed to sign declaration after parent switch");
                 return;
             }
-            self.coord_cache.clear();
+            // Surgical invalidation — see CoordCache::invalidate_via_node doc.
+            self.coord_cache
+                .invalidate_via_node(self.identity.node_addr());
             self.reset_discovery_backoff();
 
             self.stats_mut().tree.parent_switched += 1;
@@ -266,7 +268,7 @@ impl Node {
                 new_seq = new_seq,
                 new_root = %self.tree_state.root(),
                 depth = self.tree_state.my_coords().depth(),
-                "Parent switched, flushed coord cache, announcing to all peers"
+                "Parent switched, invalidated downstream coord cache entries, announcing to all peers"
             );
             if flap_dampened {
                 self.stats_mut().tree.flap_dampened += 1;
@@ -286,7 +288,9 @@ impl Node {
                 warn!(error = %e, "Failed to sign self-root declaration");
                 return;
             }
-            self.coord_cache.clear();
+            // Surgical invalidation — see CoordCache::invalidate_other_roots doc.
+            self.coord_cache
+                .invalidate_other_roots(self.identity.node_addr());
             self.reset_discovery_backoff();
             self.stats_mut().tree.parent_switched += 1;
             self.stats_mut().tree.parent_switches += 1;
@@ -320,7 +324,12 @@ impl Node {
                         warn!(error = %e, "Failed to sign declaration after loop detection");
                         return;
                     }
-                    self.coord_cache.clear();
+                    // handle_parent_lost may promote to root OR find new parent;
+                    // cover both invalidation classes.
+                    self.coord_cache
+                        .invalidate_via_node(self.identity.node_addr());
+                    self.coord_cache
+                        .invalidate_other_roots(self.tree_state.root());
                     self.reset_discovery_backoff();
                     self.send_tree_announce_to_all().await;
                 }
@@ -355,7 +364,9 @@ impl Node {
                 warn!(error = %e, "Failed to sign declaration after parent update");
                 return;
             }
-            self.coord_cache.clear();
+            // Surgical invalidation — see CoordCache::invalidate_via_node doc.
+            self.coord_cache
+                .invalidate_via_node(self.identity.node_addr());
             self.reset_discovery_backoff();
 
             let new_addrs: Vec<NodeAddr> =
@@ -445,7 +456,9 @@ impl Node {
                 warn!(error = %e, "Failed to sign declaration after periodic parent re-eval");
                 return;
             }
-            self.coord_cache.clear();
+            // Surgical invalidation — see CoordCache::invalidate_via_node doc.
+            self.coord_cache
+                .invalidate_via_node(self.identity.node_addr());
             self.reset_discovery_backoff();
 
             self.stats_mut().tree.parent_switched += 1;
@@ -474,7 +487,9 @@ impl Node {
                 warn!(error = %e, "Failed to sign self-root declaration in periodic reeval");
                 return;
             }
-            self.coord_cache.clear();
+            // Surgical invalidation — see CoordCache::invalidate_other_roots doc.
+            self.coord_cache
+                .invalidate_other_roots(self.identity.node_addr());
             self.reset_discovery_backoff();
             self.stats_mut().tree.parent_switched += 1;
             self.stats_mut().tree.parent_switches += 1;
diff --git a/src/protocol/link.rs b/src/protocol/link.rs
index f1c7bd2..e8044f3 100644
--- a/src/protocol/link.rs
+++ b/src/protocol/link.rs
@@ -307,6 +307,19 @@ pub struct SessionDatagram {
     pub payload: Vec<u8>,
 }
 
+/// Borrowed view of a session datagram payload.
+///
+/// This avoids allocating and copying the inner payload when the caller only
+/// needs to inspect or locally deliver it.
+#[derive(Clone, Copy, Debug)]
+pub struct SessionDatagramRef<'a> {
+    pub src_addr: NodeAddr,
+    pub dest_addr: NodeAddr,
+    pub ttl: u8,
+    pub path_mtu: u16,
+    pub payload: &'a [u8],
+}
+
 /// SessionDatagram fixed header size: msg_type(1) + ttl(1) + path_mtu(2) + src_addr(16) + dest_addr(16).
 pub const SESSION_DATAGRAM_HEADER_SIZE: usize = 36;
 
@@ -363,6 +376,14 @@ impl SessionDatagram {
 
     /// Decode from link-layer payload (after msg_type byte has been consumed).
     pub fn decode(payload: &[u8]) -> Result<Self, ProtocolError> {
+        let view = SessionDatagramRef::decode(payload)?;
+        Ok(view.into_owned())
+    }
+}
+
+impl<'a> SessionDatagramRef<'a> {
+    /// Decode a borrowed view from link-layer payload after the msg_type byte.
+    pub fn decode(payload: &'a [u8]) -> Result<Self, ProtocolError> {
         // ttl(1) + path_mtu(2) + src_addr(16) + dest_addr(16) = 35
         if payload.len() < 35 {
             return Err(ProtocolError::MessageTooShort {
@@ -376,16 +397,26 @@ impl SessionDatagram {
         src_bytes.copy_from_slice(&payload[3..19]);
         let mut dest_bytes = [0u8; 16];
         dest_bytes.copy_from_slice(&payload[19..35]);
-        let inner_payload = payload[35..].to_vec();
 
         Ok(Self {
             src_addr: NodeAddr::from_bytes(src_bytes),
             dest_addr: NodeAddr::from_bytes(dest_bytes),
             ttl,
             path_mtu,
-            payload: inner_payload,
+            payload: &payload[35..],
         })
     }
+
+    /// Materialize an owned datagram for forwarding/re-encoding paths.
+    pub fn into_owned(self) -> SessionDatagram {
+        SessionDatagram {
+            src_addr: self.src_addr,
+            dest_addr: self.dest_addr,
+            ttl: self.ttl,
+            path_mtu: self.path_mtu,
+            payload: self.payload.to_vec(),
+        }
+    }
 }
 
 // Legacy type alias for compatibility during transition
@@ -561,6 +592,75 @@ mod tests {
         assert_eq!(decoded.payload, payload);
     }
 
+    #[test]
+    fn test_session_datagram_ref_decode_borrows_payload() {
+        let src = make_node_addr(0xAA);
+        let dest = make_node_addr(0xBB);
+        let payload = vec![0x10, 0x00, 0x05, 0x00, 1, 2, 3, 4, 5];
+        let dg = SessionDatagram::new(src, dest, payload.clone())
+            .with_ttl(32)
+            .with_path_mtu(1400);
+
+        let encoded = dg.encode();
+        let decoded = SessionDatagramRef::decode(&encoded[1..]).unwrap();
+
+        assert_eq!(decoded.src_addr, src);
+        assert_eq!(decoded.dest_addr, dest);
+        assert_eq!(decoded.ttl, 32);
+        assert_eq!(decoded.path_mtu, 1400);
+        assert_eq!(decoded.payload, payload.as_slice());
+        assert_eq!(
+            decoded.payload.as_ptr(),
+            encoded[SESSION_DATAGRAM_HEADER_SIZE..].as_ptr()
+        );
+    }
+
+    #[test]
+    #[ignore = "performance benchmark; run explicitly with --ignored --nocapture"]
+    fn bench_session_datagram_decode_owned_vs_ref() {
+        use std::hint::black_box;
+        use std::time::Instant;
+
+        const ITERS: usize = 300_000;
+
+        let src = make_node_addr(0xAA);
+        let dest = make_node_addr(0xBB);
+        let payload = vec![0x5A; 1200];
+        let datagram = SessionDatagram::new(src, dest, payload)
+            .with_ttl(32)
+            .with_path_mtu(1400);
+        let encoded = datagram.encode();
+        let link_payload = &encoded[1..];
+
+        let ref_start = Instant::now();
+        let mut ref_bytes = 0usize;
+        for _ in 0..ITERS {
+            let decoded = SessionDatagramRef::decode(black_box(link_payload)).unwrap();
+            ref_bytes = ref_bytes.wrapping_add(decoded.payload.len());
+            black_box(decoded);
+        }
+        let ref_elapsed = ref_start.elapsed();
+
+        let owned_start = Instant::now();
+        let mut owned_bytes = 0usize;
+        for _ in 0..ITERS {
+            let decoded = SessionDatagram::decode(black_box(link_payload)).unwrap();
+            owned_bytes = owned_bytes.wrapping_add(decoded.payload.len());
+            black_box(decoded);
+        }
+        let owned_elapsed = owned_start.elapsed();
+
+        assert_eq!(ref_bytes, owned_bytes);
+        println!(
+            "SessionDatagram decode: ref={:.1} ns/op owned={:.1} ns/op speedup={:.2}x iters={} payload_bytes={}",
+            ref_elapsed.as_secs_f64() * 1_000_000_000.0 / ITERS as f64,
+            owned_elapsed.as_secs_f64() * 1_000_000_000.0 / ITERS as f64,
+            owned_elapsed.as_secs_f64() / ref_elapsed.as_secs_f64(),
+            ITERS,
+            link_payload.len() - 35
+        );
+    }
+
     #[test]
     fn test_session_datagram_empty_payload() {
         let dg = SessionDatagram::new(make_node_addr(1), make_node_addr(2), Vec::new());
diff --git a/src/protocol/mod.rs b/src/protocol/mod.rs
index 60756ee..da5b746 100644
--- a/src/protocol/mod.rs
+++ b/src/protocol/mod.rs
@@ -34,7 +34,7 @@ pub use error::ProtocolError;
 pub use filter::{FilterAnnounce, FilterNack};
 pub use link::{
     Disconnect, DisconnectReason, HandshakeMessageType, LinkMessageType,
-    SESSION_DATAGRAM_HEADER_SIZE, SessionDatagram,
+    SESSION_DATAGRAM_HEADER_SIZE, SessionDatagram, SessionDatagramRef,
 };
 pub use negotiation::{
     FMP_FEAT_PROFILE_MASK, FMP_FEAT_PROVIDES_RR, FMP_FEAT_PROVIDES_SR, FMP_FEAT_WANTS_RR,
diff --git a/src/transport/mod.rs b/src/transport/mod.rs
index 41d0f8d..3fa2c14 100644
--- a/src/transport/mod.rs
+++ b/src/transport/mod.rs
@@ -399,6 +399,15 @@ impl TransportAddr {
         Self(s.as_bytes().to_vec())
     }
 
+    /// Create a UDP/TCP transport address directly from a socket address.
+    pub fn from_socket_addr(addr: std::net::SocketAddr) -> Self {
+        use std::io::Write;
+
+        let mut buf = Vec::with_capacity(56);
+        write!(&mut buf, "{addr}").expect("Vec<u8>::write_fmt is infallible");
+        Self(buf)
+    }
+
     /// Get the raw bytes.
     pub fn as_bytes(&self) -> &[u8] {
         &self.0
@@ -1311,6 +1320,15 @@ mod tests {
         assert_eq!(addr2.as_str(), Some("hello"));
     }
 
+    #[test]
+    fn test_transport_addr_from_socket_addr() {
+        let addr = TransportAddr::from_socket_addr("127.0.0.1:2121".parse().unwrap());
+        assert_eq!(addr.as_str(), Some("127.0.0.1:2121"));
+
+        let addr = TransportAddr::from_socket_addr("[::1]:2121".parse().unwrap());
+        assert_eq!(addr.as_str(), Some("[::1]:2121"));
+    }
+
     #[test]
     fn test_link_stats_basic() {
         let mut stats = LinkStats::new();
diff --git a/src/transport/udp/mod.rs b/src/transport/udp/mod.rs
index 2530def..7fddcbd 100644
--- a/src/transport/udp/mod.rs
+++ b/src/transport/udp/mod.rs
@@ -413,10 +413,10 @@ impl Drop for UdpTransport {
 
 /// UDP receive loop - runs as a spawned task.
 ///
-/// On Linux, drains the kernel UDP queue in 32-packet bursts via `recvmmsg`
-/// to amortise the per-syscall + per-task-wakeup overhead. macOS / Windows
-/// fall through to single-packet `recv_from`. Either way every datagram
-/// is forwarded to `packet_tx` in arrival order.
+/// Drains the kernel UDP queue in 32-packet bursts via `recvmmsg` (Linux) or
+/// `recvmsg_x` (macOS) to amortise the per-syscall + per-task-wakeup overhead.
+/// Other unix targets and Windows fall through to single-packet `recv_from`.
+/// Either way every datagram is forwarded to `packet_tx` in arrival order.
 async fn udp_receive_loop(
     socket: AsyncUdpSocket,
     transport_id: TransportId,
@@ -426,11 +426,13 @@ async fn udp_receive_loop(
 ) {
     debug!(transport_id = %transport_id, "UDP receive loop starting");
 
-    #[cfg(target_os = "linux")]
+    #[cfg(any(target_os = "linux", target_os = "macos"))]
     {
         const BATCH: usize = 32;
         let buf_size = mtu as usize + 100;
-        // One contiguous backing alloc; slice it for recvmmsg.
+        // One Vec per recvmmsg / recvmsg_x slot. When a packet lands, move the
+        // filled buffer directly into ReceivedPacket and install a fresh empty
+        // buffer for the next syscall, avoiding a per-packet memcpy.
         let mut backing: Vec<Vec<u8>> = (0..BATCH).map(|_| vec![0u8; buf_size]).collect();
         let mut addrs: [Option<std::net::SocketAddr>; BATCH] = std::array::from_fn(|_| None);
         let mut lens: [usize; BATCH] = [0; BATCH];
@@ -454,8 +456,7 @@ async fn udp_receive_loop(
                         };
                         stats.record_recv(len);
 
-                        let buf = &backing[i][..len];
-                        if is_punch_packet(buf) {
+                        if is_punch_packet(&backing[i][..len]) {
                             trace!(
                                 transport_id = %transport_id,
                                 remote_addr = %remote_addr,
@@ -465,8 +466,9 @@ async fn udp_receive_loop(
                             continue;
                         }
 
-                        let data = buf.to_vec();
-                        let addr = TransportAddr::from_string(&remote_addr.to_string());
+                        let mut data = std::mem::replace(&mut backing[i], vec![0u8; buf_size]);
+                        data.truncate(len);
+                        let addr = TransportAddr::from_socket_addr(remote_addr);
                         let packet = ReceivedPacket::new(transport_id, addr, data);
 
                         trace!(
@@ -497,7 +499,7 @@ async fn udp_receive_loop(
         }
     }
 
-    #[cfg(not(target_os = "linux"))]
+    #[cfg(not(any(target_os = "linux", target_os = "macos")))]
     {
         let mut buf = vec![0u8; mtu as usize + 100];
 
@@ -955,4 +957,49 @@ mod tests {
         t1.stop_async().await.unwrap();
         t2.stop_async().await.unwrap();
     }
+
+    /// Burst more than one datagram into the kernel queue before yielding to
+    /// the receive loop, then assert all are delivered in arrival order. On
+    /// Linux/macOS this exercises the recvmmsg / recvmsg_x batching path
+    /// (multiple datagrams reaped per syscall); on other unix targets it
+    /// degrades to N single-packet recvmsg calls and still must pass.
+    #[tokio::test]
+    async fn test_burst_recv_batch() {
+        let (tx1, _rx1) = packet_channel(100);
+        let (tx2, mut rx2) = packet_channel(100);
+
+        let mut t1 = UdpTransport::new(TransportId::new(1), None, make_config(0), tx1);
+        let mut t2 = UdpTransport::new(TransportId::new(2), None, make_config(0), tx2);
+
+        t1.start_async().await.unwrap();
+        t2.start_async().await.unwrap();
+
+        let addr2 = TransportAddr::from_string(&t2.local_addr().unwrap().to_string());
+
+        // Fire BURST datagrams back-to-back. Each carries its index in the
+        // first 4 bytes so we can verify per-datagram boundaries (recvmsg_x
+        // must not coalesce them).
+        const BURST: u32 = 10;
+        for i in 0..BURST {
+            let mut payload = vec![0u8; 32];
+            payload[..4].copy_from_slice(&i.to_be_bytes());
+            payload[4..].fill(b'x');
+            t1.send_async(&addr2, &payload).await.unwrap();
+        }
+
+        // Drain. Order must match send order (UDP loopback is in-order, and
+        // recvmmsg/recvmsg_x preserve it across the batch).
+        for expected in 0..BURST {
+            let packet = timeout(Duration::from_secs(1), rx2.recv())
+                .await
+                .expect("timeout draining burst")
+                .expect("channel closed");
+            assert_eq!(packet.data.len(), 32);
+            let got = u32::from_be_bytes(packet.data[..4].try_into().unwrap());
+            assert_eq!(got, expected, "datagram out of order");
+        }
+
+        t1.stop_async().await.unwrap();
+        t2.stop_async().await.unwrap();
+    }
 }
diff --git a/src/transport/udp/socket.rs b/src/transport/udp/socket.rs
index cef256f..d8a7c9e 100644
--- a/src/transport/udp/socket.rs
+++ b/src/transport/udp/socket.rs
@@ -29,13 +29,41 @@ mod platform {
     use std::os::unix::io::{AsRawFd, RawFd};
     use tokio::io::unix::AsyncFd;
 
-    /// Maximum number of datagrams a single recvmmsg / sendmmsg syscall
-    /// will pull from / push to the kernel. Tuned to amortise syscall +
-    /// per-task-wakeup overhead across a useful burst without blowing
-    /// the stack (each slot owns an mmsghdr + sockaddr_storage + iovec).
-    #[cfg(target_os = "linux")]
+    /// Maximum number of datagrams a single recvmmsg / recvmsg_x / sendmmsg
+    /// syscall will pull from / push to the kernel. Tuned to amortise syscall +
+    /// per-task-wakeup overhead across a useful burst without blowing the
+    /// stack (each slot owns an mmsghdr/msghdr_x + sockaddr_storage + iovec).
+    #[cfg(any(target_os = "linux", target_os = "macos"))]
     const BATCH_SIZE: usize = 32;
 
+    /// Darwin-private `msghdr_x` for the `recvmsg_x` / `sendmsg_x` syscalls.
+    /// Layout matches `bsd/sys/socket_private.h` in xnu — same as `msghdr` plus
+    /// a trailing `msg_datalen` (per-message bytes-received output, in lieu of
+    /// the `msg_len` field that `mmsghdr` uses on Linux).
+    #[cfg(target_os = "macos")]
+    #[repr(C)]
+    #[allow(non_camel_case_types)]
+    struct msghdr_x {
+        msg_name: *mut libc::c_void,
+        msg_namelen: libc::socklen_t,
+        msg_iov: *mut libc::iovec,
+        msg_iovlen: libc::c_int,
+        msg_control: *mut libc::c_void,
+        msg_controllen: libc::socklen_t,
+        msg_flags: libc::c_int,
+        msg_datalen: usize,
+    }
+
+    #[cfg(target_os = "macos")]
+    unsafe extern "C" {
+        fn recvmsg_x(
+            s: libc::c_int,
+            msgp: *const msghdr_x,
+            cnt: libc::c_uint,
+            flags: libc::c_int,
+        ) -> isize;
+    }
+
     /// Wrapper around a `socket2::Socket` providing sync send/recv with
     /// `SO_RXQ_OVFL` ancillary data parsing.
     pub struct UdpRawSocket {
@@ -240,10 +268,10 @@ mod platform {
         /// value is a cumulative counter since socket creation; it is 0 if
         /// `SO_RXQ_OVFL` is not supported.
         ///
-        /// On Linux the production receive path uses `recv_batch` (recvmmsg);
-        /// this single-packet variant remains for non-Linux unix targets and
-        /// for the local `tests` module.
-        #[cfg_attr(target_os = "linux", allow(dead_code))]
+        /// The production receive path on Linux/macOS uses `recv_batch`
+        /// (recvmmsg / recvmsg_x); this single-packet variant remains for
+        /// other unix targets and for the local `tests` module.
+        #[cfg_attr(any(target_os = "linux", target_os = "macos"), allow(dead_code))]
         pub fn recv_from(&self, buf: &mut [u8]) -> std::io::Result<(usize, SocketAddr, u32)> {
             let fd = self.inner.as_raw_fd();
 
@@ -391,6 +419,56 @@ mod platform {
             Ok((count, drops))
         }
 
+        /// Receive up to `BATCH_SIZE` datagrams in a single `recvmsg_x` syscall
+        /// (macOS). Same `(count, drops)` contract as the Linux `recv_batch`,
+        /// except `drops` is always 0 — Darwin has no `SO_RXQ_OVFL` equivalent.
+        ///
+        /// `recvmsg_x` is a Darwin-private syscall (not in the public SDK) but
+        /// is shipped in production xnu and is used by quinn-udp for the same
+        /// per-syscall-amortisation reason as our Linux `recvmmsg` path.
+        #[cfg(target_os = "macos")]
+        pub fn recv_batch(
+            &self,
+            bufs: &mut [&mut [u8]],
+            addrs: &mut [Option<SocketAddr>],
+            lens: &mut [usize],
+        ) -> std::io::Result<(usize, u32)> {
+            let n = bufs.len().min(addrs.len()).min(lens.len()).min(BATCH_SIZE);
+            if n == 0 {
+                return Ok((0, 0));
+            }
+            let fd = self.inner.as_raw_fd();
+
+            let mut iovs: [libc::iovec; BATCH_SIZE] = unsafe { std::mem::zeroed() };
+            let mut storages: [libc::sockaddr_storage; BATCH_SIZE] = unsafe { std::mem::zeroed() };
+            let mut msgs: [msghdr_x; BATCH_SIZE] = unsafe { std::mem::zeroed() };
+
+            for i in 0..n {
+                iovs[i].iov_base = bufs[i].as_mut_ptr() as *mut libc::c_void;
+                iovs[i].iov_len = bufs[i].len();
+                msgs[i].msg_name = &mut storages[i] as *mut _ as *mut libc::c_void;
+                msgs[i].msg_namelen =
+                    std::mem::size_of::<libc::sockaddr_storage>() as libc::socklen_t;
+                msgs[i].msg_iov = &mut iovs[i];
+                msgs[i].msg_iovlen = 1;
+                // No cmsg consumption — leave msg_control null. (msg_controllen
+                // is documented as not overwritten by macOS recvmsg_x; zeroed
+                // init keeps it sane.)
+            }
+
+            let r = unsafe { recvmsg_x(fd, msgs.as_ptr(), n as libc::c_uint, 0) };
+            if r < 0 {
+                return Err(std::io::Error::last_os_error());
+            }
+            let count = r as usize;
+            for i in 0..count {
+                lens[i] = msgs[i].msg_datalen;
+                addrs[i] = sockaddr_to_socket_addr(&storages[i]).ok();
+            }
+
+            Ok((count, 0))
+        }
+
         /// Wrap this socket in a tokio `AsyncFd` for async I/O.
         pub fn into_async(self) -> Result<AsyncUdpSocket, TransportError> {
             let async_fd = AsyncFd::new(self)
@@ -440,11 +518,11 @@ mod platform {
 
         /// Receive a payload, source address, and kernel drop counter.
         ///
-        /// Returns `(bytes_read, source_addr, kernel_drops)`. On Linux the
-        /// production receive path uses `recv_batch`; this single-packet
-        /// variant remains for non-Linux unix targets and for the local
-        /// `tests` module.
-        #[cfg_attr(target_os = "linux", allow(dead_code))]
+        /// Returns `(bytes_read, source_addr, kernel_drops)`. On Linux/macOS
+        /// the production receive path uses `recv_batch`; this single-packet
+        /// variant remains for other unix targets and for the local `tests`
+        /// module.
+        #[cfg_attr(any(target_os = "linux", target_os = "macos"), allow(dead_code))]
         pub async fn recv_from(
             &self,
             buf: &mut [u8],
@@ -465,9 +543,10 @@ mod platform {
         }
 
         /// Drain up to `BATCH_SIZE` datagrams from the kernel via
-        /// `recvmmsg` (Linux). Returns `(count, kernel_drops)`; same
-        /// buffer / addr / len contract as `UdpRawSocket::recv_batch`.
-        #[cfg(target_os = "linux")]
+        /// `recvmmsg` (Linux) or `recvmsg_x` (macOS). Returns
+        /// `(count, kernel_drops)`; same buffer / addr / len contract as
+        /// `UdpRawSocket::recv_batch`. `kernel_drops` is always 0 on macOS.
+        #[cfg(any(target_os = "linux", target_os = "macos"))]
         pub async fn recv_batch(
             &self,
             bufs: &mut [&mut [u8]],
@@ -739,4 +818,167 @@ mod tests {
         assert_eq!(&buf[..n], payload);
         assert_eq!(src, addr1);
     }
+
+    /// Microbench: compare per-packet `recv_from` (single recvmsg syscall +
+    /// task wakeup per datagram — the macOS pre-recvmsg_x baseline) vs
+    /// `recv_batch` (the new recvmsg_x path, up to 32 datagrams per syscall).
+    /// Both modes run back-to-back in this binary on loopback so the only
+    /// thing that varies is the receive-syscall strategy. Sender is a tight
+    /// `socket.send_to()` loop in a separate task; receiver counts datagrams
+    /// drained over a fixed wall-clock window per mode.
+    ///
+    /// Run with:
+    ///   cargo test --release -p fips --lib transport::udp::socket::tests::bench_udp_recv_amortization -- --ignored --nocapture
+    ///
+    /// Sender runs on a dedicated *blocking* OS thread (std::net::UdpSocket
+    /// in default blocking mode) so it always saturates the kernel rx queue
+    /// regardless of how the tokio receiver schedules. That's the scenario
+    /// where recvmmsg / recvmsg_x is meant to win: the receiver wakes up to
+    /// find N packets already buffered, and one syscall reaps the burst.
+    #[cfg(any(target_os = "linux", target_os = "macos"))]
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    #[ignore]
+    async fn bench_udp_recv_amortization() {
+        use std::sync::Arc;
+        use std::sync::atomic::{AtomicBool, Ordering};
+        use std::time::{Duration, Instant};
+
+        const RECV_BUF: usize = 4 * 1024 * 1024;
+        const SEND_BUF: usize = 1024 * 1024;
+        const PAYLOAD_LEN: usize = 100;
+        const WINDOW: Duration = Duration::from_secs(3);
+        const WARMUP: Duration = Duration::from_millis(500);
+
+        async fn run_mode(
+            label: &str,
+            batched: bool,
+            sender_threads: usize,
+        ) -> (u64, u64, Duration) {
+            let rx_sock = UdpRawSocket::open("127.0.0.1:0".parse().unwrap(), RECV_BUF, SEND_BUF)
+                .expect("rx bind");
+            let rx_addr = rx_sock.local_addr();
+            let rx = rx_sock.into_async().expect("rx into_async");
+
+            // Senders: N dedicated blocking std threads. More threads → deeper
+            // kernel rx queue → larger amortization opportunity for recv_batch.
+            // ENOBUFS / EAGAIN just yield and retry; we want saturation, not
+            // perfect accounting. Sent count is best-effort.
+            let stop = Arc::new(AtomicBool::new(false));
+            let mut sender_handles = Vec::with_capacity(sender_threads);
+            for _ in 0..sender_threads {
+                let stop_tx = stop.clone();
+                sender_handles.push(std::thread::spawn(move || {
+                    let sock = std::net::UdpSocket::bind("127.0.0.1:0").expect("tx bind");
+                    sock.connect(rx_addr).expect("tx connect");
+                    let payload = vec![0xABu8; PAYLOAD_LEN];
+                    let mut sent: u64 = 0;
+                    while !stop_tx.load(Ordering::Relaxed) {
+                        match sock.send(&payload) {
+                            Ok(_) => sent += 1,
+                            Err(_) => std::thread::yield_now(),
+                        }
+                    }
+                    sent
+                }));
+            }
+
+            // Warm-up: let the sender thread reach steady state and the
+            // kernel rx queue start filling.
+            tokio::time::sleep(WARMUP).await;
+
+            let start = Instant::now();
+            let deadline = start + WINDOW;
+            let mut recv_count: u64 = 0;
+            let mut last_drops: u32 = 0;
+
+            if batched {
+                const BATCH: usize = 32;
+                let mut backing: Vec<Vec<u8>> =
+                    (0..BATCH).map(|_| vec![0u8; PAYLOAD_LEN + 64]).collect();
+                let mut addrs: [Option<SocketAddr>; BATCH] = std::array::from_fn(|_| None);
+                let mut lens: [usize; BATCH] = [0; BATCH];
+                let mut batch_sum: u64 = 0;
+                let mut batch_calls: u64 = 0;
+
+                while Instant::now() < deadline {
+                    let mut bufs: [&mut [u8]; BATCH] = {
+                        let mut iter = backing.iter_mut();
+                        std::array::from_fn(|_| iter.next().unwrap().as_mut_slice())
+                    };
+                    match rx.recv_batch(&mut bufs, &mut addrs, &mut lens).await {
+                        Ok((n, drops)) => {
+                            recv_count += n as u64;
+                            batch_sum += n as u64;
+                            batch_calls += 1;
+                            last_drops = drops;
+                        }
+                        Err(_) => break,
+                    }
+                }
+                let avg_batch = if batch_calls > 0 {
+                    batch_sum as f64 / batch_calls as f64
+                } else {
+                    0.0
+                };
+                eprintln!(
+                    "[{:>10}] avg_batch_per_call={:.2} ({} calls)",
+                    label, avg_batch, batch_calls
+                );
+            } else {
+                let mut buf = vec![0u8; PAYLOAD_LEN + 64];
+                while Instant::now() < deadline {
+                    match rx.recv_from(&mut buf).await {
+                        Ok((_n, _src, drops)) => {
+                            recv_count += 1;
+                            last_drops = drops;
+                        }
+                        Err(_) => break,
+                    }
+                }
+            }
+            let elapsed = start.elapsed();
+
+            stop.store(true, Ordering::Relaxed);
+            drop(rx);
+            let sent: u64 = sender_handles
+                .into_iter()
+                .map(|h| h.join().unwrap_or(0))
+                .sum();
+
+            let pps = (recv_count as f64) / elapsed.as_secs_f64();
+            let mbps =
+                (recv_count as f64) * (PAYLOAD_LEN as f64) * 8.0 / 1e6 / elapsed.as_secs_f64();
+            eprintln!(
+                "[{:>10}] recv={:>10} sent={:>10} elapsed={:?} pps={:>12.0} mbps={:>7.1} kdrops={}",
+                label, recv_count, sent, elapsed, pps, mbps, last_drops
+            );
+            (recv_count, sent, elapsed)
+        }
+
+        eprintln!("--- udp recv amortization bench ---");
+        eprintln!(
+            "payload={}B window={:?} warmup={:?} runtime=multi_thread(2)",
+            PAYLOAD_LEN, WINDOW, WARMUP
+        );
+
+        // Sweep sender concurrency. Each level shows how the win scales as
+        // the rx queue gets deeper (more amortization opportunity).
+        for senders in [1usize, 2, 4, 8] {
+            eprintln!("\n=== sender_threads = {} ===", senders);
+            let (b_recv, _, b_el) = run_mode(" recv_from", false, senders).await;
+            let (x_recv, _, x_el) = run_mode("recv_batch", true, senders).await;
+            let (x_recv2, _, x_el2) = run_mode("recv_batch", true, senders).await;
+            let (b_recv2, _, b_el2) = run_mode(" recv_from", false, senders).await;
+
+            let baseline_pps =
+                (b_recv as f64 / b_el.as_secs_f64() + b_recv2 as f64 / b_el2.as_secs_f64()) / 2.0;
+            let batched_pps =
+                (x_recv as f64 / x_el.as_secs_f64() + x_recv2 as f64 / x_el2.as_secs_f64()) / 2.0;
+            let speedup = batched_pps / baseline_pps;
+            eprintln!(
+                "--- senders={}: baseline={:.0} pps  batched={:.0} pps  speedup={:.2}x ---",
+                senders, baseline_pps, batched_pps, speedup
+            );
+        }
+    }
 }
diff --git a/testing/boringtun/.gitignore b/testing/boringtun/.gitignore
new file mode 100644
index 0000000..9ab870d
--- /dev/null
+++ b/testing/boringtun/.gitignore
@@ -0,0 +1 @@
+generated/
diff --git a/testing/boringtun/Dockerfile b/testing/boringtun/Dockerfile
new file mode 100644
index 0000000..0b44c21
--- /dev/null
+++ b/testing/boringtun/Dockerfile
@@ -0,0 +1,28 @@
+# Minimal boringtun-cli image for throughput benchmarking against FIPS.
+# Userspace WireGuard (boringtun) + wireguard-tools + iperf3 in one
+# image. Run two containers on a docker bridge, configure a WG peer
+# between them, then iperf3 client→server across the tunnel.
+
+FROM rust:1-bookworm AS build
+# 0.6.0 (crates.io HEAD) is from 2022 and turns in ~1.7 Mbps in
+# this harness — clearly broken. Pull the boringtun-cli binary from
+# the cloudflare HEAD instead; it gets the post-0.6 perf fixes and
+# the multi-threaded recv loop.
+RUN git clone --depth 1 https://github.com/cloudflare/boringtun /usr/src/boringtun \
+    && cargo install --path /usr/src/boringtun/boringtun-cli
+
+FROM debian:bookworm-slim
+ENV DEBIAN_FRONTEND=noninteractive
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends \
+        iperf3 \
+        iproute2 \
+        wireguard-tools \
+        iputils-ping \
+        netcat-openbsd \
+        ca-certificates \
+    && rm -rf /var/lib/apt/lists/*
+COPY --from=build /usr/local/cargo/bin/boringtun-cli /usr/local/bin/boringtun-cli
+COPY entrypoint.sh /entrypoint.sh
+RUN chmod +x /entrypoint.sh
+ENTRYPOINT ["/entrypoint.sh"]
diff --git a/testing/boringtun/README.md b/testing/boringtun/README.md
new file mode 100644
index 0000000..aeb1be6
--- /dev/null
+++ b/testing/boringtun/README.md
@@ -0,0 +1,29 @@
+# BoringTun Throughput Baseline
+
+This harness runs two userspace WireGuard peers with Cloudflare BoringTun and
+measures single-stream TCP throughput with `iperf3`. It is intended as a simple
+baseline for comparing FIPS tunnel throughput against another userspace tunnel.
+
+```bash
+docker build -t boringtun-test:latest testing/boringtun
+testing/boringtun/scripts/generate-keys.sh
+docker compose -f testing/boringtun/docker-compose.yml up -d
+testing/boringtun/scripts/bench.sh
+docker compose -f testing/boringtun/docker-compose.yml down
+```
+
+The generated WireGuard key material is written under
+`testing/boringtun/generated/` and is ignored by git.
+
+For FIPS-to-FIPS revision comparisons, use the static topology comparison
+script:
+
+```bash
+testing/static/scripts/iperf-compare-refs.sh origin/master HEAD mesh
+```
+
+That script builds each ref into a separate `fips-test:*` image, runs the
+same static `iperf3` topology against both images, and prints a bandwidth
+summary for each path. Override `DURATION`, `PARALLEL`, `SETTLE_SECONDS`, or
+`IPERF_TIMEOUT` in the environment when needed. Set `RUNS=3` or similar to
+repeat each ref and print aggregate results.
diff --git a/testing/boringtun/docker-compose.yml b/testing/boringtun/docker-compose.yml
new file mode 100644
index 0000000..81c1220
--- /dev/null
+++ b/testing/boringtun/docker-compose.yml
@@ -0,0 +1,46 @@
+# Minimal 2-node boringtun setup for iperf3 throughput comparison.
+# Both containers join the same docker bridge; each runs a boringtun
+# userspace WireGuard tunnel and they peer with each other over UDP.
+# Inner WG IPs: 10.99.0.1/24 (alice) <-> 10.99.0.2/24 (bob).
+
+networks:
+  bt-net:
+    driver: bridge
+    ipam:
+      config:
+        - subnet: 172.99.0.0/24
+
+x-boringtun-common: &boringtun-common
+  image: boringtun-test:latest
+  cap_add:
+    - NET_ADMIN
+  devices:
+    - /dev/net/tun:/dev/net/tun
+  sysctls:
+    - net.ipv4.conf.all.forwarding=1
+  restart: "no"
+  env_file:
+    - ./generated/peers.env
+
+services:
+  alice:
+    <<: *boringtun-common
+    container_name: bt-alice
+    hostname: alice
+    environment:
+      - ROLE=alice
+      - PEER_HOST=bob
+    networks:
+      bt-net:
+        ipv4_address: 172.99.0.10
+
+  bob:
+    <<: *boringtun-common
+    container_name: bt-bob
+    hostname: bob
+    environment:
+      - ROLE=bob
+      - PEER_HOST=alice
+    networks:
+      bt-net:
+        ipv4_address: 172.99.0.11
diff --git a/testing/boringtun/entrypoint.sh b/testing/boringtun/entrypoint.sh
new file mode 100755
index 0000000..a440125
--- /dev/null
+++ b/testing/boringtun/entrypoint.sh
@@ -0,0 +1,72 @@
+#!/bin/bash
+# Bring up a boringtun-userspace WireGuard tunnel and then sleep
+# indefinitely so the benchmark scripts can drive it via docker exec.
+#
+# Required env:
+#   ROLE      "alice" or "bob"
+#   ALICE_WG_IP  WG inner IP for alice (e.g. 10.99.0.1/24)
+#   BOB_WG_IP    WG inner IP for bob (e.g. 10.99.0.2/24)
+#   ALICE_PUB   alice's WG public key
+#   BOB_PUB     bob's WG public key
+#   ALICE_PRIV  alice's WG private key
+#   BOB_PRIV    bob's WG private key
+#   PEER_HOST   the other container's hostname on the docker bridge
+#   PEER_PORT   the other container's WG UDP port (default 51820)
+#
+# boringtun-cli runs in foreground (--foreground) so wg-quick-style
+# configuration is done manually via `ip`, `wg`, and `wg set`.
+
+set -e
+
+PORT="${PEER_PORT:-51820}"
+case "$ROLE" in
+    alice)
+        OUR_IP="$ALICE_WG_IP"
+        OUR_PRIV="$ALICE_PRIV"
+        PEER_PUB="$BOB_PUB"
+        ;;
+    bob)
+        OUR_IP="$BOB_WG_IP"
+        OUR_PRIV="$BOB_PRIV"
+        PEER_PUB="$ALICE_PUB"
+        ;;
+    *)
+        echo "ROLE must be alice or bob, got '$ROLE'" >&2
+        exit 1
+        ;;
+esac
+
+echo "[$ROLE] starting boringtun-cli foreground on wg0"
+# `--foreground` keeps the userspace driver in the container
+# foreground. boringtun-cli sets WG_TUN_NAME_FILE to /tmp/wg0.name
+# when --foreground; we just hardcode the device name.
+boringtun-cli --foreground --disable-drop-privileges wg0 &
+BORINGTUN_PID=$!
+
+# wait for the tun device to appear
+for i in $(seq 1 50); do
+    if ip link show wg0 >/dev/null 2>&1; then
+        break
+    fi
+    sleep 0.1
+done
+
+echo "[$ROLE] configuring wg0 with $OUR_IP listening on $PORT"
+PRIV_FILE=$(mktemp)
+chmod 600 "$PRIV_FILE"
+printf '%s' "$OUR_PRIV" >"$PRIV_FILE"
+wg set wg0 private-key "$PRIV_FILE" listen-port "$PORT"
+rm -f "$PRIV_FILE"
+
+ip address add "$OUR_IP" dev wg0
+ip link set up dev wg0
+
+echo "[$ROLE] adding peer pubkey, endpoint $PEER_HOST:$PORT"
+wg set wg0 peer "$PEER_PUB" allowed-ips 10.99.0.0/24 endpoint "$PEER_HOST:$PORT" persistent-keepalive 25
+
+echo "[$ROLE] ready"
+wg show wg0
+
+# Park here so the container stays up; we'll run iperf3 etc via
+# docker exec.
+wait $BORINGTUN_PID
diff --git a/testing/boringtun/scripts/bench.sh b/testing/boringtun/scripts/bench.sh
new file mode 100755
index 0000000..f8acbb9
--- /dev/null
+++ b/testing/boringtun/scripts/bench.sh
@@ -0,0 +1,25 @@
+#!/bin/bash
+# End-to-end iperf3 bandwidth test between two boringtun containers.
+# Output mirrors testing/static/scripts/iperf-test.sh so the FIPS
+# numbers are directly comparable.
+set -euo pipefail
+
+DURATION="${DURATION:-10}"
+PARALLEL="${PARALLEL:-1}"
+
+echo "=== boringtun iperf3 throughput (single TCP stream, ${DURATION}s) ==="
+
+# Run iperf3 server on alice (background), client on bob.
+docker exec -d bt-alice iperf3 -s -1 -B 10.99.0.1 -p 5201
+sleep 1
+
+# wait for tun handshake to settle (boringtun + WG keepalive)
+sleep 2
+
+# Client: bob → alice over WG (10.99.0.1)
+OUT=$(docker exec bt-bob iperf3 -c 10.99.0.1 -p 5201 -t "$DURATION" -P "$PARALLEL" -J)
+
+# Pull SUM bps.
+MBPS=$(echo "$OUT" | python3 -c "import json,sys; d=json.load(sys.stdin); print(f\"{d['end']['sum_received']['bits_per_second'] / 1_000_000:.2f}\")")
+
+echo "boringtun bob -> alice : ${MBPS} Mbits/sec"
diff --git a/testing/boringtun/scripts/generate-keys.sh b/testing/boringtun/scripts/generate-keys.sh
new file mode 100755
index 0000000..0c0c7f5
--- /dev/null
+++ b/testing/boringtun/scripts/generate-keys.sh
@@ -0,0 +1,40 @@
+#!/bin/bash
+# Generate WG keypairs for alice and bob and write the
+# generated/peers.env file the docker-compose.yml reads.
+# Idempotent: skips if file already exists.
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+OUT_DIR="$SCRIPT_DIR/../generated"
+ENV_FILE="$OUT_DIR/peers.env"
+
+if [ -f "$ENV_FILE" ]; then
+    echo "$ENV_FILE already exists; reusing keys"
+    exit 0
+fi
+
+mkdir -p "$OUT_DIR"
+
+# Use the built boringtun-test image to generate keys — host may
+# not have wireguard-tools installed. The `pubkey` step needs `-i`
+# so stdin is piped through; `genkey` doesn't need it but matching
+# flags keeps the two calls symmetric.
+GEN_CMD="docker run --rm --entrypoint wg boringtun-test:latest"
+PUB_CMD="docker run --rm -i --entrypoint wg boringtun-test:latest"
+
+ALICE_PRIV=$($GEN_CMD genkey)
+ALICE_PUB=$(printf '%s' "$ALICE_PRIV" | $PUB_CMD pubkey)
+BOB_PRIV=$($GEN_CMD genkey)
+BOB_PUB=$(printf '%s' "$BOB_PRIV" | $PUB_CMD pubkey)
+
+cat >"$ENV_FILE" <<EOF
+ALICE_PRIV=$ALICE_PRIV
+ALICE_PUB=$ALICE_PUB
+BOB_PRIV=$BOB_PRIV
+BOB_PUB=$BOB_PUB
+ALICE_WG_IP=10.99.0.1/24
+BOB_WG_IP=10.99.0.2/24
+PEER_PORT=51820
+EOF
+chmod 600 "$ENV_FILE"
+echo "wrote $ENV_FILE"
diff --git a/testing/scripts/build.sh b/testing/scripts/build.sh
index b72e47a..35781ea 100755
--- a/testing/scripts/build.sh
+++ b/testing/scripts/build.sh
@@ -40,6 +40,8 @@ done
 UNAME_S=$(uname -s)
 CARGO_TARGET="x86_64-unknown-linux-musl"
 
+TARGET_ROOT="${CARGO_TARGET_DIR:-$PROJECT_ROOT/target}"
+
 if [ "$UNAME_S" = "Darwin" ]; then
     echo "Detected macOS host — using cross-compilation for Linux..."
 
@@ -57,12 +59,12 @@ if [ "$UNAME_S" = "Darwin" ]; then
     echo "Building FIPS for Linux (release) using cargo-zigbuild..."
     cargo zigbuild --release --target "$CARGO_TARGET" --manifest-path="$PROJECT_ROOT/Cargo.toml" "${CARGO_BUILD_ARGS[@]}"
 
-    TARGET_DIR="$PROJECT_ROOT/target/$CARGO_TARGET/release"
+    TARGET_DIR="$TARGET_ROOT/$CARGO_TARGET/release"
 else
     echo "Building FIPS (release)..."
     cargo build --release --manifest-path="$PROJECT_ROOT/Cargo.toml" "${CARGO_BUILD_ARGS[@]}"
 
-    TARGET_DIR="$PROJECT_ROOT/target/release"
+    TARGET_DIR="$TARGET_ROOT/release"
 fi
 
 echo "Copying binaries to $DOCKER_DIR/"
diff --git a/testing/static/README.md b/testing/static/README.md
index 8d22a08..d922035 100644
--- a/testing/static/README.md
+++ b/testing/static/README.md
@@ -270,6 +270,19 @@ Runs iperf3 with:
 - Parallel streams: 8 (`-P 8`)
 - Protocol: TCP over IPv6
 
+For before/after measurements across commits or branches:
+
+```bash
+./testing/static/scripts/iperf-compare-refs.sh origin/master HEAD mesh
+```
+
+The comparison script builds each ref into a separate Docker image, runs the
+same topology and `iperf3` settings for both images, and prints a bandwidth
+summary. Override `DURATION`, `PARALLEL`, `SETTLE_SECONDS`, `IPERF_TIMEOUT`,
+or `RUNS` in the environment when needed. `RUNS` is the total number of
+measurements per ref; for example, `RUNS=3` runs each ref three times and
+prints both per-run and aggregate tables.
+
 ## Network Impairment
 
 The `netem.sh` script simulates adverse network conditions using `tc`/`netem`
diff --git a/testing/static/docker-compose.yml b/testing/static/docker-compose.yml
index 3cdb9ac..410fe61 100644
--- a/testing/static/docker-compose.yml
+++ b/testing/static/docker-compose.yml
@@ -13,7 +13,7 @@ networks:
         - subnet: fd02::/64
 
 x-fips-common: &fips-common
-  image: fips-test:latest
+  image: ${FIPS_TEST_IMAGE:-fips-test:latest}
   cap_add:
     - NET_ADMIN
   devices:
diff --git a/testing/static/scripts/iperf-compare-refs.sh b/testing/static/scripts/iperf-compare-refs.sh
new file mode 100755
index 0000000..ce6ec66
--- /dev/null
+++ b/testing/static/scripts/iperf-compare-refs.sh
@@ -0,0 +1,202 @@
+#!/bin/bash
+# Build two git refs into separate FIPS test images and run the same static
+# iperf3 topology against each one. This keeps performance PR evidence
+# repeatable: compare origin/master with a candidate branch using identical
+# configs, topology, duration, and parallelism.
+#
+# Usage:
+#   ./testing/static/scripts/iperf-compare-refs.sh <base-ref> <candidate-ref> [mesh|chain]
+#
+# Environment:
+#   DURATION=10        iperf3 duration passed through to iperf-test.sh
+#   PARALLEL=8         iperf3 parallel streams passed through to iperf-test.sh
+#   SETTLE_SECONDS=3   topology startup delay passed through to iperf-test.sh
+#   IPERF_TIMEOUT      per-path timeout, defaults to DURATION + 30
+#   RUNS=1             total measurement runs per ref
+set -euo pipefail
+
+if [ "$#" -lt 2 ] || [ "$#" -gt 3 ]; then
+    echo "Usage: $0 <base-ref> <candidate-ref> [mesh|chain]" >&2
+    exit 2
+fi
+
+BASE_REF="$1"
+CANDIDATE_REF="$2"
+PROFILE="${3:-mesh}"
+RUNS="${RUNS:-1}"
+
+if ! [[ "$RUNS" =~ ^[1-9][0-9]*$ ]]; then
+    echo "RUNS must be a positive integer" >&2
+    exit 2
+fi
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PROJECT_ROOT="$(cd "$SCRIPT_DIR/../../.." && pwd)"
+COMPOSE_FILE="$PROJECT_ROOT/testing/static/docker-compose.yml"
+TMP_DIR="$(mktemp -d "${TMPDIR:-/tmp}/fips-iperf-compare.XXXXXX")"
+WORKTREES=()
+FAILED_RUNS=0
+
+cleanup() {
+    docker compose -f "$COMPOSE_FILE" --profile "$PROFILE" down >/dev/null 2>&1 || true
+    for wt in "${WORKTREES[@]}"; do
+        git -C "$PROJECT_ROOT" worktree remove --force "$wt" >/dev/null 2>&1 || true
+    done
+    rm -rf "$TMP_DIR"
+}
+trap cleanup EXIT
+
+slug_ref() {
+    echo "$1" \
+        | tr '[:upper:]' '[:lower:]' \
+        | sed 's/[^a-z0-9_.-]/-/g; s/^-*//; s/-*$//' \
+        | cut -c1-48
+}
+
+image_tag() {
+    local label="$1"
+    local ref="$2"
+    local slug
+    slug="$(slug_ref "$ref")"
+    [ -n "$slug" ] || slug="ref"
+    printf 'fips-test:compare-%s-%s\n' "$label" "$slug"
+}
+
+build_ref_image() {
+    local label="$1"
+    local ref="$2"
+    local tag="$3"
+
+    local wt="$TMP_DIR/$label"
+    local target_dir="$PROJECT_ROOT/target/iperf-compare-$label-$(slug_ref "$ref")"
+
+    echo ""
+    echo "=== Building $label: $ref -> $tag ==="
+    git -C "$PROJECT_ROOT" worktree add --detach "$wt" "$ref"
+    WORKTREES+=("$wt")
+
+    mkdir -p "$target_dir"
+    ln -s "$target_dir" "$wt/target"
+    (
+        cd "$wt"
+        CARGO_TARGET_DIR="$wt/target" ./testing/scripts/build.sh
+    )
+    docker tag fips-test:latest "$tag"
+}
+
+run_profile() {
+    local label="$1"
+    local image="$2"
+    local run="$3"
+    local log="$TMP_DIR/$label-run-$run.log"
+    local duration="${DURATION:-10}"
+    local parallel="${PARALLEL:-8}"
+    local settle_seconds="${SETTLE_SECONDS:-3}"
+    local iperf_timeout="${IPERF_TIMEOUT:-$((duration + 30))}"
+
+    echo ""
+    echo "=== Running $label run $run/$RUNS with $image ==="
+    FIPS_TEST_IMAGE="$image" docker compose -f "$COMPOSE_FILE" --profile "$PROFILE" up -d --force-recreate
+    if DURATION="$duration" PARALLEL="$parallel" \
+        SETTLE_SECONDS="$settle_seconds" IPERF_TIMEOUT="$iperf_timeout" \
+        "$SCRIPT_DIR/iperf-test.sh" "$PROFILE" | tee "$log"; then
+        :
+    else
+        local status="$?"
+        echo "=== $label exited with status $status ===" | tee -a "$log"
+        FAILED_RUNS=1
+    fi
+    docker compose -f "$COMPOSE_FILE" --profile "$PROFILE" down
+}
+
+print_summary() {
+    local label="$1"
+    local run="$2"
+    local log="$TMP_DIR/$label-run-$run.log"
+
+    [ -f "$log" ] || return 0
+
+    awk -v label="$label" -v run="$run" '
+        /^=== .* ===$/ && $0 !~ /FIPS iperf3 Bandwidth Test/ && $0 !~ /Results:/ {
+            test=$0
+            sub(/^=== /, "", test)
+            sub(/ ===$/, "", test)
+        }
+        /^Bandwidth:/ {
+            print label "\t" run "\t" test "\t" $2 " " $3
+        }
+    ' "$log"
+}
+
+print_aggregate() {
+    local summary="$1"
+
+    awk -F '\t' '
+        NR == 1 { next }
+
+        function to_mbps(value, unit) {
+            if (unit ~ /^Gbits/) return value * 1000
+            if (unit ~ /^Mbits/) return value
+            if (unit ~ /^Kbits/) return value / 1000
+            return value
+        }
+
+        {
+            split($4, parts, " ")
+            mbps = to_mbps(parts[1] + 0, parts[2])
+            key = $1 SUBSEP $3
+            if (!(key in seen)) {
+                seen[key] = 1
+                order[++order_len] = key
+                refs[key] = $1
+                tests[key] = $3
+                min[key] = mbps
+                max[key] = mbps
+            }
+            count[key]++
+            sum[key] += mbps
+            if (mbps < min[key]) min[key] = mbps
+            if (mbps > max[key]) max[key] = mbps
+        }
+
+        END {
+            print "ref\ttest\truns\tavg_mbps\tmin_mbps\tmax_mbps"
+            for (i = 1; i <= order_len; i++) {
+                key = order[i]
+                printf "%s\t%s\t%d\t%.0f\t%.0f\t%.0f\n",
+                    refs[key], tests[key], count[key],
+                    sum[key] / count[key], min[key], max[key]
+            }
+        }
+    ' "$summary"
+}
+
+"$SCRIPT_DIR/generate-configs.sh" "$PROFILE"
+
+BASE_IMAGE="$(image_tag base "$BASE_REF")"
+CANDIDATE_IMAGE="$(image_tag candidate "$CANDIDATE_REF")"
+
+build_ref_image base "$BASE_REF" "$BASE_IMAGE"
+build_ref_image candidate "$CANDIDATE_REF" "$CANDIDATE_IMAGE"
+
+for run in $(seq 1 "$RUNS"); do
+    run_profile base "$BASE_IMAGE" "$run"
+    run_profile candidate "$CANDIDATE_IMAGE" "$run"
+done
+
+echo ""
+echo "=== Summary ==="
+SUMMARY_FILE="$TMP_DIR/summary.tsv"
+{
+    printf 'ref\trun\ttest\tbandwidth\n'
+    for run in $(seq 1 "$RUNS"); do
+        print_summary base "$run"
+        print_summary candidate "$run"
+    done
+} | tee "$SUMMARY_FILE"
+
+echo ""
+echo "=== Aggregate ==="
+print_aggregate "$SUMMARY_FILE"
+
+exit "$FAILED_RUNS"
diff --git a/testing/static/scripts/iperf-test.sh b/testing/static/scripts/iperf-test.sh
index 74d1767..450db5e 100755
--- a/testing/static/scripts/iperf-test.sh
+++ b/testing/static/scripts/iperf-test.sh
@@ -18,8 +18,10 @@ if [ "$2" = "--live" ] || [ "$1" = "--live" ]; then
     [ "$1" = "--live" ] && PROFILE="mesh"
 fi
 
-DURATION=10
-PARALLEL=8
+DURATION="${DURATION:-10}"
+PARALLEL="${PARALLEL:-8}"
+SETTLE_SECONDS="${SETTLE_SECONDS:-3}"
+IPERF_TIMEOUT="${IPERF_TIMEOUT:-$((DURATION + 30))}"
 PASSED=0
 FAILED=0
 
@@ -47,7 +49,7 @@ iperf_test() {
     if [ "$LIVE_OUTPUT" = true ]; then
         # Show live output
         echo "Running iperf3 test (live output):"
-        if docker exec "fips-$client_node" iperf3 -c "${dest_npub}.fips" -t "$DURATION" -P "$PARALLEL"; then
+        if docker exec "fips-$client_node" timeout "$IPERF_TIMEOUT" iperf3 -c "${dest_npub}.fips" -t "$DURATION" -P "$PARALLEL"; then
             PASSED=$((PASSED + 1))
         else
             echo "FAIL"
@@ -57,7 +59,7 @@ iperf_test() {
         # Capture and summarize output
         echo -n "Running iperf3 test... "
         local output
-        if output=$(docker exec "fips-$client_node" iperf3 -c "${dest_npub}.fips" -t "$DURATION" -P "$PARALLEL" 2>&1); then
+        if output=$(docker exec "fips-$client_node" timeout "$IPERF_TIMEOUT" iperf3 -c "${dest_npub}.fips" -t "$DURATION" -P "$PARALLEL" 2>&1); then
             # Check if we got valid results
             if echo "$output" | grep -q "sender"; then
                 # Extract and display results (get SUM line for aggregate bandwidth)
@@ -83,8 +85,8 @@ echo "=== FIPS iperf3 Bandwidth Test ($PROFILE topology) ==="
 echo ""
 
 # Wait for nodes to converge
-echo "Waiting 3s for mesh convergence..."
-sleep 3
+echo "Waiting ${SETTLE_SECONDS}s for mesh convergence..."
+sleep "$SETTLE_SECONDS"
 
 if [ "$PROFILE" = "mesh" ] || [ "$PROFILE" = "mesh-public" ]; then
     # Test key paths in mesh topology
@@ -120,4 +122,4 @@ fi
 
 echo ""
 echo "=== Results: $PASSED passed, $FAILED failed ==="
-[ "$FAILED" -eq 0 ] && exit 0 || exit 1
\ No newline at end of file
+[ "$FAILED" -eq 0 ] && exit 0 || exit 1