diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 6ebfb3a6..68f5d798 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -84,6 +84,11 @@ jobs:
       # a matrix suite.
       - name: Run convergence-gate unit tests
         run: bash testing/lib/wait-converge-test.sh
+      # Runs package-linux.yml's own version derivation on a release tag, a
+      # candidate tag and a branch. Not a matrix suite either, so it is kept in
+      # step with ci-local.sh's run_deb_version by hand.
+      - name: Check the Debian version derived for tags and branches
+        run: bash testing/check-deb-version.sh
 
   fmt:
     name: Format check
@@ -636,6 +641,36 @@ jobs:
         timeout-minutes: 5
         run: bash testing/openwrt/maintainer-scripts-test.sh
 
+# ─────────────────────────────────────────────────────────────────────────────
+# Job 2f – systemd tarball upgrade scenarios
+#
+# Runs the tarball's install.sh twice in a systemd container, the second time
+# as an upgrade, with fips, fips-dns and fips-gateway in a known state, and
+# checks that the units running before the upgrade are running after it. The
+# binaries are stubs, so like the OpenWrt job it needs no FIPS build and no
+# shared test image, and has a job of its own.
+#
+# The leg keeps `suite:` because testing/check-ci-parity.sh matches it against
+# TARBALL_INSTALL_SUITES in ci-local.sh; the step below does not read it.
+# ─────────────────────────────────────────────────────────────────────────────
+  tarball-install:
+    name: systemd tarball (${{ matrix.suite }})
+    runs-on: ubuntu-latest
+    if: ${{ !inputs.skip_integration }}
+
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - suite: tarball-install
+
+    steps:
+      - uses: actions/checkout@d23441a48e516b6c34aea4fa41551a30e30af803 # v6
+
+      - name: Run the systemd tarball upgrade scenarios
+        timeout-minutes: 10
+        run: bash testing/tarball-install/test.sh
+
 # ─────────────────────────────────────────────────────────────────────────────
 # Job 3 – Integration tests (static mesh + chaos simulation)
 #
diff --git a/.github/workflows/package-linux.yml b/.github/workflows/package-linux.yml
index 7a7ff22d..d9f20e17 100644
--- a/.github/workflows/package-linux.yml
+++ b/.github/workflows/package-linux.yml
@@ -18,6 +18,7 @@ jobs:
     runs-on: ubuntu-latest
     outputs:
       linux_package_version: ${{ steps.linux_version.outputs.linux_package_version }}
+      deb_package_version: ${{ steps.linux_version.outputs.deb_package_version }}
     steps:
       - uses: actions/checkout@d23441a48e516b6c34aea4fa41551a30e30af803 # v6
         with:
@@ -42,7 +43,18 @@ jobs:
             VERSION="${BASE_VERSION}+${BRANCH}.${HEIGHT}.${HASH}"
           fi
 
+          # dpkg reads X.Y.Z-rcN as revision rcN of X.Y.Z and sorts it above
+          # the release; X.Y.Z~rcN sorts below it. git refuses '~' in a ref
+          # name, so the tag carries '-' and this maps it, for the .deb only.
+          # testing/check-deb-version.sh runs this step's text.
+          DEB_VERSION="$VERSION"
+          if [[ "$GITHUB_REF" == refs/tags/* ]] \
+            && [[ "$VERSION" =~ ^([0-9]+\.[0-9]+\.[0-9]+)-((alpha|beta|pre|rc)[0-9]*)$ ]]; then
+            DEB_VERSION="${BASH_REMATCH[1]}~${BASH_REMATCH[2]}"
+          fi
+
           echo "linux_package_version=${VERSION}" >> "$GITHUB_OUTPUT"
+          echo "deb_package_version=${DEB_VERSION}" >> "$GITHUB_OUTPUT"
 
   build:
     name: Build Linux artifacts (${{ matrix.artifact_arch }})
@@ -128,7 +140,7 @@ jobs:
           : ${GITHUB_OUTPUT:=/tmp/github_output}
 
           packaging/debian/build-deb-container.sh \
-            --version "${{ needs.determine-versioning.outputs.linux_package_version }}" \
+            --version "${{ needs.determine-versioning.outputs.deb_package_version }}" \
             --output-dir deploy \
             --image-archive "$RUNNER_TEMP/deb-builder-image.tar" \
             | tee /tmp/build-deb-container.log
@@ -149,6 +161,19 @@ jobs:
               ;;
           esac
 
+          # A candidate's package Version carries '~' (see determine-versioning),
+          # and cargo-deb puts it in the file name. GitHub renames a release
+          # asset whose name has special characters, which would leave
+          # checksums-linux.txt naming a file the release does not have. The
+          # file takes the tag's '-' instead; the Version inside is unchanged.
+          case "$DEB_FILE" in
+            *~*)
+              RENAMED="$(dirname "$DEB_FILE")/$(basename "$DEB_FILE" | tr '~' '-')"
+              mv "$DEB_FILE" "$RENAMED"
+              DEB_FILE="$RENAMED"
+              ;;
+          esac
+
           # Record it relative to the checkout: upload-artifact derives the
           # archive layout from the common ancestor of its paths, and an
           # absolute path here would nest the package under directories the
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 3434e374..7c3024dd 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -324,6 +324,10 @@ with v0.5.x or earlier peers.
 
 ### Changed
 
+- Transport address clones share immutable bytes instead of allocating a copy.
+  Socket addresses are formatted on the stack so construction still needs only
+  one allocation, including scoped IPv6 addresses.
+
 - `Degraded` is now a level rather than a latch. The supervisor's reason set
   was monotonic, which was correct while no child could recover; with recovery
   it would have meant "something broke at some point since boot" rather than
@@ -502,6 +506,19 @@ with v0.5.x or earlier peers.
   reaching it means a second build image and a second floor. `FIPS_BUILD_IMAGE`
   is still the oldest Debian-family distribution, which is no longer the oldest
   distribution outright.
+- A release-candidate `.deb` is versioned `X.Y.Z~rcN` instead of
+  `X.Y.Z-rcN`. dpkg read the old form as a revision of the release and sorted
+  it above `X.Y.Z`, so a host that installed a candidate was not upgraded by
+  the release. The tarball, artifact and `.deb` file names keep the tag's
+  `-rcN`.
+
+### Deprecated
+
+#### Sessions and rekey
+
+- Setting `node.rekey.enabled: false`. The disabled path is less tested and is
+  not supported, and the option is removed in v2. It still works as before;
+  the configuration reference and the mesh-layer design now say so.
 
 ### Removed
 
@@ -643,6 +660,24 @@ with v0.5.x or earlier peers.
   the Noise session, tree position and routes survive it. Linux and macOS (the
   platforms with the connected-socket fast path); elsewhere the heartbeat alone
   carries the new address.
+- The Ethernet transport on Linux no longer leaks a socket when the interface
+  is missing or its name is invalid at the moment the socket is opened, as when
+  the interface goes away between the presence check and the bind.
+- TCP connections try the remaining addresses for a hostname after a
+  connection fails, within the existing overall connection timeout.
+
+#### Gateway
+
+- The gateway retries failed firewall rebuilds every ten seconds without
+  waiting for another mapping change. Retries apply the latest desired
+  mappings and port forwards, preserving changes across transient failures.
+
+#### Packaging
+
+- The systemd tarball's `install.sh` restarts fips-dns and fips-gateway after
+  an upgrade when they were running. It stopped fips first, which stopped both
+  through `Requires=fips.service`, then restarted only fips, so `.fips`
+  resolution stayed down and the gateway stayed stopped until started by hand.
 
 - A node no longer relays away the answer to its own lookup. A request is
   flooded to every tree peer whose bloom filter claims the target, so a false
diff --git a/Cargo.toml b/Cargo.toml
index 70a5d296..8994c2de 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -144,3 +144,8 @@ path = "src/bin/fipstop/main.rs"
 name = "routing_next_hop"
 path = "benches/routing_next_hop.rs"
 harness = false
+
+[[bench]]
+name = "transport_addr_clone"
+path = "benches/transport_addr_clone.rs"
+harness = false
diff --git a/benches/transport_addr_clone.rs b/benches/transport_addr_clone.rs
new file mode 100644
index 00000000..603e0bcf
--- /dev/null
+++ b/benches/transport_addr_clone.rs
@@ -0,0 +1,107 @@
+//! Compare shared addresses with the former Vec backing, including the receive
+//! constructor and a producer cloning while a consumer drops on another thread.
+//! The cross-thread case includes bounded-channel overhead in both measurements;
+//! it is not an isolated measurement of the atomic reference count.
+
+use std::hint::black_box;
+use std::io::Write;
+use std::net::SocketAddr;
+use std::sync::{Barrier, mpsc};
+use std::time::{Duration, Instant};
+
+use criterion::{BenchmarkId, Criterion, criterion_group, criterion_main};
+use fips::TransportAddr;
+
+const ADDRESS_BYTES: [usize; 3] = [6, 18, 58];
+
+fn former_from_socket_addr(addr: SocketAddr) -> Vec<u8> {
+    let mut buf = Vec::with_capacity(56);
+    write!(&mut buf, "{addr}").expect("Vec<u8>::write_fmt is infallible");
+    buf
+}
+
+fn bench_clone(c: &mut Criterion) {
+    let mut group = c.benchmark_group("transport_addr_clone");
+    for len in ADDRESS_BYTES {
+        let former = vec![0x5a; len];
+        let shared = TransportAddr::from_bytes(&former);
+        group.bench_with_input(BenchmarkId::new("former_vec", len), &former, |b, addr| {
+            b.iter(|| black_box(addr.clone()))
+        });
+        group.bench_with_input(BenchmarkId::new("shared", len), &shared, |b, addr| {
+            b.iter(|| black_box(addr.clone()))
+        });
+    }
+    group.finish();
+}
+
+fn bench_socket_addr_and_clone(c: &mut Criterion) {
+    let mut group = c.benchmark_group("transport_addr_socket_and_clone");
+    for (name, text) in [
+        ("ipv4", "192.168.1.1:2121"),
+        ("ipv6", "[2001:db8::1]:2121"),
+        (
+            "scoped_ipv6",
+            "[ffff:ffff:ffff:ffff:ffff:ffff:ffff:ffff%4294967295]:65535",
+        ),
+    ] {
+        let socket: SocketAddr = text.parse().unwrap();
+        group.bench_with_input(BenchmarkId::new("former_vec", name), &socket, |b, addr| {
+            b.iter(|| {
+                let addr = former_from_socket_addr(black_box(*addr));
+                black_box((addr.clone(), addr))
+            })
+        });
+        group.bench_with_input(BenchmarkId::new("shared", name), &socket, |b, addr| {
+            b.iter(|| {
+                let addr = TransportAddr::from_socket_addr(black_box(*addr));
+                black_box((addr.clone(), addr))
+            })
+        });
+    }
+    group.finish();
+}
+
+fn cross_thread_clone_drop<T: Clone + Send + Sync>(addr: &T, iterations: u64) -> Duration {
+    let (tx, rx) = mpsc::sync_channel::<T>(256);
+    let ready = Barrier::new(2);
+    std::thread::scope(|scope| {
+        let consumer = scope.spawn(|| {
+            ready.wait();
+            for value in rx {
+                drop(black_box(value));
+            }
+        });
+        // Thread startup is outside the timed interval; completion is included.
+        ready.wait();
+        let start = Instant::now();
+        for _ in 0..iterations {
+            tx.send(black_box(addr.clone())).unwrap();
+        }
+        drop(tx);
+        consumer.join().unwrap();
+        start.elapsed()
+    })
+}
+
+fn bench_cross_thread(c: &mut Criterion) {
+    let mut group = c.benchmark_group("transport_addr_cross_thread");
+    for len in ADDRESS_BYTES {
+        let former = vec![0x5a; len];
+        let shared = TransportAddr::from_bytes(&former);
+        group.bench_with_input(BenchmarkId::new("former_vec", len), &former, |b, addr| {
+            b.iter_custom(|iterations| cross_thread_clone_drop(addr, iterations))
+        });
+        group.bench_with_input(BenchmarkId::new("shared", len), &shared, |b, addr| {
+            b.iter_custom(|iterations| cross_thread_clone_drop(addr, iterations))
+        });
+    }
+    group.finish();
+}
+
+criterion_group! {
+    name = benches;
+    config = Criterion::default().sample_size(50);
+    targets = bench_clone, bench_socket_addr_and_clone, bench_cross_thread
+}
+criterion_main!(benches);
diff --git a/docs/design/fips-gateway.md b/docs/design/fips-gateway.md
index d3530135..34d15f70 100644
--- a/docs/design/fips-gateway.md
+++ b/docs/design/fips-gateway.md
@@ -458,8 +458,10 @@ sequence is:
    batch, and a batch too large for any send buffer is refused
    before it is sent. A batch the kernel refuses leaves the previous
    table in place; the failure is logged, and the gateway keeps its
-   record of the change, so the next rebuild that succeeds applies
-   it.
+   record of the change. Pending changes are retried every ten seconds,
+   even when no new mapping event arrives. Each retry rebuilds the latest
+   desired state, so a newer change supersedes an earlier failed one.
+   Successful rebuilds clear the pending flag; clean tables are not retried.
 
 The rustables crate does not expose rule-handle tracking, so
 incremental update of individual rules is not available. Atomic
diff --git a/docs/design/fips-mesh-layer.md b/docs/design/fips-mesh-layer.md
index 0ac4bf16..9d016c77 100644
--- a/docs/design/fips-mesh-layer.md
+++ b/docs/design/fips-mesh-layer.md
@@ -384,8 +384,9 @@ A rekey is initiated when either threshold is reached on the link's current
 session: `node.rekey.after_secs` (default 120) elapsed since the link came
 up or last rekeyed, or `node.rekey.after_messages` (default 65536) frames
 sent. Either side can be the initiator independently. Rekey is on by
-default and can be disabled via `node.rekey.enabled: false` (the
-configuration tree is documented in
+default. `node.rekey.enabled: false` stops this node initiating rekey, but
+that configuration is unsupported: it is a less-tested path, and the option
+is removed in v2 (the configuration tree is documented in
 [../reference/configuration.md](../reference/configuration.md)).
 
 ### Mechanism
diff --git a/docs/reference/configuration.md b/docs/reference/configuration.md
index 5fd33b2d..b5b91fa9 100644
--- a/docs/reference/configuration.md
+++ b/docs/reference/configuration.md
@@ -498,7 +498,7 @@ cutover.
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `node.rekey.enabled` | bool | `true` | Initiate periodic Noise rekey on links and sessions. A peer-driven session rekey is still answered when this is off, so session keys can still rotate |
+| `node.rekey.enabled` | bool | `true` | Initiate periodic Noise rekey on links and sessions. A peer-driven session rekey is still answered when this is off, so session keys can still rotate. Disabling rekey is unsupported: it runs the node on a less-tested path, and the option is removed in v2. |
 | `node.rekey.after_secs` | u64 | `120` | Initiate rekey after this many seconds on a session |
 | `node.rekey.after_messages` | u64 | `65536` | Initiate rekey after this many messages sent on a session |
 
@@ -1366,6 +1366,7 @@ node:
     loss_threshold: 0.05             # MMP loss rate threshold for CE marking (5%)
     etx_threshold: 3.0               # MMP ETX threshold for CE marking
   rekey:
+    # enabled: false is unsupported, less tested, and removed in v2
     enabled: true                    # periodic Noise rekey for forward secrecy
     after_secs: 120                  # rekey interval (seconds)
     after_messages: 65536            # rekey after N messages sent
diff --git a/packaging/README.md b/packaging/README.md
index ae948cae..a989409c 100644
--- a/packaging/README.md
+++ b/packaging/README.md
@@ -393,9 +393,11 @@ pkg add ./fips-<version>-pfsense-ce2.8-amd64.pkg
 /usr/local/libexec/fips/fips-dns-setup   # edits config.xml; run deliberately
 ```
 
-Not a Netgate-supported package, and a pfSense firmware upgrade removes
-it. See [pfsense/README.md](pfsense/README.md) for the "Allow IPv6"
-prerequisite the mesh depends on, firewall-rule notes, and removal
+Not a Netgate-supported package. A pfSense firmware upgrade keeps it (it
+is a plain pkg, not a `pfSense-pkg-*`); after a major base change,
+reinstall the package built for the new base. See
+[pfsense/README.md](pfsense/README.md) for the "Allow IPv6" prerequisite
+the mesh depends on, firewall-rule notes, and upgrade and removal
 behaviour.
 
 ### Windows (`.zip`)
diff --git a/packaging/pfsense/build-pkg.sh b/packaging/pfsense/build-pkg.sh
index 37efbbda..4840767e 100755
--- a/packaging/pfsense/build-pkg.sh
+++ b/packaging/pfsense/build-pkg.sh
@@ -24,8 +24,10 @@
 #     This package integrates through the DNS Resolver custom options.
 #   - The responder's bind address, for the reason recorded in
 #     fips.yaml.dns.
-#   - Lifetime. A pfSense firmware upgrade reinstalls the base image and
-#     takes third-party packages with it, so post-install says so.
+#   - Lifetime. A firmware upgrade keeps the package (pfSense-upgrade
+#     reinstalls only pfSense-pkg-* packages), but a major upgrade changes
+#     the FreeBSD base, so post-install says to reinstall the package
+#     built for the new base.
 #
 # Ships fips, fipsctl and fipstop. fips-gateway is excluded: its NAT
 # backend is nftables (Linux-only), and pfSense has pf for that anyway.
diff --git a/packaging/pfsense/fips-dns-setup b/packaging/pfsense/fips-dns-setup
index 41b33640..04ed4a42 100755
--- a/packaging/pfsense/fips-dns-setup
+++ b/packaging/pfsense/fips-dns-setup
@@ -13,10 +13,11 @@
 # options" box, which unbound.inc splices into the generated config
 # verbatim. It is stored base64-encoded in config.xml, which is the part
 # that makes it the right home: config.xml is what survives a reboot, a
-# firmware upgrade and a config restore, whereas everything this package
-# installs under /usr/local does not. So the .fips zone keeps resolving
-# across an upgrade that removes the daemon, which is a loud failure
-# (SERVFAIL on .fips) rather than a quiet one.
+# firmware upgrade and a config restore. What this package installs
+# under /usr/local survives a firmware upgrade too (pfSense-upgrade
+# reinstalls only pfSense-pkg-* packages), but not a removal of the
+# package. So the .fips zone can outlive the daemon, which is a loud
+# failure (SERVFAIL on .fips) rather than a quiet one.
 #
 # This edits the firewall's live configuration, so it is deliberately
 # NOT run from the package's post-install: installing a package should
diff --git a/packaging/systemd/install.sh b/packaging/systemd/install.sh
index f71513b7..89e78111 100755
--- a/packaging/systemd/install.sh
+++ b/packaging/systemd/install.sh
@@ -112,19 +112,20 @@ fi
 
 # --- Install systemd units ---
 
-was_active=false
-if systemctl is-active --quiet fips.service 2>/dev/null; then
-    was_active=true
-    echo "Stopping running fips service..."
-    systemctl stop fips.service
-fi
-
-dns_was_active=false
-if systemctl is-active --quiet fips-dns.service 2>/dev/null; then
-    dns_was_active=true
-    echo "Stopping running fips-dns service..."
-    systemctl stop fips-dns.service
-fi
+# fips-dns and fips-gateway carry Requires=fips.service, so stopping fips
+# stops them too. Record all three before stopping any of them.
+declare -A was_active=()
+for unit in fips fips-dns fips-gateway; do
+    if systemctl is-active --quiet "${unit}.service" 2>/dev/null; then
+        was_active[$unit]=true
+    fi
+done
+for unit in fips-gateway fips-dns fips; do
+    if [ "${was_active[$unit]:-}" = true ]; then
+        echo "Stopping running ${unit} service..."
+        systemctl stop "${unit}.service"
+    fi
+done
 
 install -m 0644 "${SCRIPT_DIR}/fips.service" "${SYSTEMD_DIR}/fips.service"
 install -m 0644 "${SCRIPT_DIR}/fips-dns.service" "${SYSTEMD_DIR}/fips-dns.service"
@@ -164,15 +165,13 @@ systemctl enable fips.service
 systemctl enable fips-dns.service
 echo "Services enabled (will start on boot)."
 
-# Restart if they were running before
-if $was_active; then
-    echo "Restarting fips service..."
-    systemctl start fips.service
-fi
-if $dns_was_active; then
-    echo "Restarting fips-dns service..."
-    systemctl start fips-dns.service
-fi
+# Restart each unit that was running before
+for unit in fips fips-dns fips-gateway; do
+    if [ "${was_active[$unit]:-}" = true ]; then
+        echo "Restarting ${unit} service..."
+        systemctl start "${unit}.service"
+    fi
+done
 
 echo ""
 echo "=== Installation complete ==="
diff --git a/src/bin/fips-gateway.rs b/src/bin/fips-gateway.rs
index 3d2b784d..4fa248c7 100644
--- a/src/bin/fips-gateway.rs
+++ b/src/bin/fips-gateway.rs
@@ -107,6 +107,30 @@ fn report_unreadable_conntrack(
     }
 }
 
+/// Log a pending NAT rebuild retry once per change of outcome.
+///
+/// A recovery is reported whether the retry applied the rules itself or a
+/// mapping change applied them in between and left nothing pending.
+#[cfg(target_os = "linux")]
+fn report_nat_retry(log: &mut nat::RetryLog, result: &Result<bool, nat::NatError>) {
+    match (log.observe(result.as_ref().err()), result) {
+        (nat::RetryReport::Failed, Err(e)) => {
+            warn!(error = %e, "Failed to retry pending NAT rules")
+        }
+        (nat::RetryReport::Repeated, Err(e)) => {
+            debug!(error = %e, "Pending NAT rules still failing to apply")
+        }
+        (nat::RetryReport::Recovered, Ok(true)) => {
+            info!("Applied pending NAT rules; retries recovered")
+        }
+        (nat::RetryReport::Recovered, _) => {
+            info!("Pending NAT rules were applied by a later update; retries recovered")
+        }
+        (nat::RetryReport::Clean, Ok(true)) => info!("Applied pending NAT rules"),
+        _ => {}
+    }
+}
+
 /// Check once at startup which conntrack source the tick will read, and say so.
 ///
 /// Without this, an operator on a kernel with no readable source learns that
@@ -545,8 +569,14 @@ async fn main() {
     info!("fips-gateway running");
 
     let mut exit_code = 0;
+    let mut nat_retry = tokio::time::interval(std::time::Duration::from_secs(10));
+    nat_retry.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
+    let mut nat_retry_log = nat::RetryLog::default();
     loop {
         tokio::select! {
+            _ = nat_retry.tick() => {
+                report_nat_retry(&mut nat_retry_log, &nat_mgr.retry_pending());
+            }
             Some(event) = event_rx.recv() => {
                 match event {
                     pool::PoolEvent::MappingCreated { virtual_ip, mesh_addr } => {
diff --git a/src/gateway/nat.rs b/src/gateway/nat.rs
index 7a0d2770..72de3a89 100644
--- a/src/gateway/nat.rs
+++ b/src/gateway/nat.rs
@@ -174,6 +174,8 @@ pub struct NatManager {
     mappings: HashMap<Ipv6Addr, NatMapping>,
     /// Inbound port-forward rules.
     port_forwards: Vec<PortForward>,
+    /// Desired state has not yet been acknowledged by the kernel.
+    rebuild_pending: bool,
 }
 
 impl NatManager {
@@ -199,6 +201,7 @@ impl NatManager {
             lan_interface,
             mappings: HashMap::new(),
             port_forwards: Vec::new(),
+            rebuild_pending: false,
         }
     }
 
@@ -222,7 +225,7 @@ impl NatManager {
     /// the nftables table atomically. Pass an empty slice to clear.
     pub fn set_port_forwards(&mut self, forwards: &[PortForward]) -> Result<(), NatError> {
         self.port_forwards = forwards.to_vec();
-        self.rebuild()?;
+        self.rebuild_desired()?;
         info!(
             count = self.port_forwards.len(),
             "Applied inbound port forwards"
@@ -307,6 +310,18 @@ impl NatManager {
         self.mappings.len()
     }
 
+    /// Retry a failed rebuild using the latest desired state.
+    ///
+    /// Returns whether a pending rebuild was applied. A clean manager does
+    /// not open a socket or rebuild the table.
+    pub fn retry_pending(&mut self) -> Result<bool, NatError> {
+        if !self.rebuild_pending {
+            return Ok(false);
+        }
+        self.rebuild_desired()?;
+        Ok(true)
+    }
+
     /// The objects a rebuild sends, grouped into the batches that carry them.
     ///
     /// One batch, always. The kernel applies a batch as a single transaction,
@@ -497,12 +512,85 @@ impl NatManager {
 
     /// Rebuild, returning the outcome with the time the rebuild took in
     /// microseconds, so a mapping change can log its cost on either path.
-    fn timed_rebuild(&self) -> (Result<(), NatError>, u64) {
+    fn timed_rebuild(&mut self) -> (Result<(), NatError>, u64) {
         let started = Instant::now();
-        let result = self.rebuild();
+        let result = self.rebuild_desired();
         let elapsed_us = u64::try_from(started.elapsed().as_micros()).unwrap_or(u64::MAX);
         (result, elapsed_us)
     }
+
+    /// Rebuild the desired state, leaving it marked pending unless the kernel
+    /// accepts it.
+    fn rebuild_desired(&mut self) -> Result<(), NatError> {
+        self.track_rebuild(Self::rebuild)
+    }
+
+    /// Run `apply` with the desired state marked pending, and clear the mark
+    /// only if it succeeds.
+    ///
+    /// Split from `rebuild_desired` so a test can drive both outcomes without
+    /// a netlink socket.
+    fn track_rebuild(
+        &mut self,
+        apply: impl FnOnce(&Self) -> Result<(), NatError>,
+    ) -> Result<(), NatError> {
+        self.rebuild_pending = true;
+        apply(self)?;
+        self.rebuild_pending = false;
+        Ok(())
+    }
+}
+
+/// How a pending-rebuild retry compares with the retry before it.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum RetryReport {
+    /// A failure that differs from the previous retry's outcome, or the first.
+    Failed,
+    /// The same failure as the previous retry.
+    Repeated,
+    /// The first success after a failed retry.
+    Recovered,
+    /// A success with no failed retry before it.
+    Clean,
+}
+
+/// Remembers the last failed retry of a pending rebuild.
+///
+/// A failure that does not clear on its own, for example the kernel refusing
+/// the batch while a module or capability is missing, fails every retry the
+/// same way, and a per-retry warning would repeat for the life of the
+/// process. Warning on a change of outcome, and reporting the recovery once,
+/// keeps the log readable. A retry that finds nothing pending after a failure
+/// counts as recovery, since a mapping change rebuilt the table in between.
+#[derive(Debug, Default)]
+pub struct RetryLog {
+    last_failure: Option<String>,
+}
+
+impl RetryLog {
+    /// Record a retry outcome and say how it compares with the last one.
+    ///
+    /// `None` is a success, whether or not anything was pending; `Some` is
+    /// the retry's error. Two failures are the same when their messages are.
+    pub fn observe(&mut self, failure: Option<&NatError>) -> RetryReport {
+        match (failure.map(ToString::to_string), self.last_failure.take()) {
+            (Some(now), Some(before)) => {
+                let report = if now == before {
+                    RetryReport::Repeated
+                } else {
+                    RetryReport::Failed
+                };
+                self.last_failure = Some(now);
+                report
+            }
+            (Some(now), None) => {
+                self.last_failure = Some(now);
+                RetryReport::Failed
+            }
+            (None, Some(_)) => RetryReport::Recovered,
+            (None, None) => RetryReport::Clean,
+        }
+    }
 }
 
 /// Largest batch, in bytes, that the kernel can admit in one send.
@@ -855,6 +943,13 @@ fn send_batch(bytes: &[u8]) -> Result<(), NatError> {
 // `SO_SNDBUF` fallback after `SO_SNDBUFFORCE` returns `EPERM` needs a process
 // without CAP_NET_ADMIN, and the gateway suite's container is privileged, so
 // nothing runs that either. The gateway suite covers only the success path.
+//
+// The pending-rebuild flag is covered here on both sides through
+// `track_rebuild`, with a stand-in for the rebuild: a failure leaves it set and
+// `retry_pending` still retries, and a success clears it so `retry_pending`
+// finds nothing to do. What stays only in the ignored kernel test, which no
+// suite runs, is `retry_pending` itself applying a pending rebuild and
+// returning `Ok(true)`, and a kernel rejection leaving the old table in place.
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -883,6 +978,185 @@ mod tests {
         mgr
     }
 
+    #[test]
+    fn failed_rebuild_retains_latest_desired_state_for_retry() {
+        let mut mgr = manager_with_mappings(2);
+        // Make encoding fail before opening a socket.
+        mgr.pre_chain = Chain::new(&mgr.table);
+        assert!(!mgr.retry_pending().unwrap(), "clean retry must not encode");
+
+        assert!(mgr.add_mapping(vip(3), mesh(3)).is_err());
+        assert!(mgr.rebuild_pending);
+        assert!(mgr.retry_pending().is_err());
+        assert!(mgr.rebuild_pending);
+        assert!(mgr.remove_mapping(vip(3)).is_err());
+        assert!(!mgr.mappings.contains_key(&vip(3)));
+        assert!(mgr.add_mapping(vip(2), mesh(99)).is_err());
+        assert_eq!(mgr.mappings[&vip(2)].mesh_addr, mesh(99));
+        assert!(mgr.remove_mapping(vip(4)).is_err());
+        assert!(
+            mgr.rebuild_pending,
+            "an absent mapping must not clear pending work"
+        );
+    }
+
+    #[test]
+    fn failed_tracked_rebuild_stays_pending_and_retry_still_rebuilds() {
+        let mut mgr = manager_with_mappings(1);
+        let refused = || NatError::Nftables("refused".into());
+
+        assert!(mgr.track_rebuild(|_| Err(refused())).is_err());
+        assert!(mgr.rebuild_pending, "a failed rebuild must stay pending");
+
+        // With encoding made to fail, a retry that is still pending attempts
+        // the rebuild, and fails, instead of reporting nothing to do.
+        mgr.pre_chain = Chain::new(&mgr.table);
+        assert!(mgr.retry_pending().is_err());
+        assert!(mgr.rebuild_pending);
+    }
+
+    #[test]
+    fn successful_tracked_rebuild_clears_pending_and_retry_does_nothing() {
+        let mut mgr = manager_with_mappings(1);
+        assert!(
+            mgr.track_rebuild(|_| Err(NatError::Nftables("refused".into())))
+                .is_err()
+        );
+        assert!(mgr.rebuild_pending);
+
+        let mut applied = 0;
+        mgr.track_rebuild(|m| {
+            assert!(m.rebuild_pending, "the rebuild runs with the mark set");
+            applied += 1;
+            Ok(())
+        })
+        .unwrap();
+        assert_eq!(applied, 1);
+        assert!(
+            !mgr.rebuild_pending,
+            "a successful rebuild must clear pending"
+        );
+
+        // Encoding would fail, so `Ok(false)` shows the retry did not rebuild.
+        mgr.pre_chain = Chain::new(&mgr.table);
+        assert!(!mgr.retry_pending().unwrap());
+    }
+
+    #[test]
+    fn retry_log_warns_on_a_new_failure_and_reports_recovery_once() {
+        let kernel = |errno| NatError::Kernel {
+            errno: Errno(errno),
+            seq: 3,
+        };
+        let mut log = RetryLog::default();
+
+        assert_eq!(log.observe(None), RetryReport::Clean);
+        assert_eq!(log.observe(Some(&kernel(22))), RetryReport::Failed);
+        assert_eq!(log.observe(Some(&kernel(22))), RetryReport::Repeated);
+        assert_eq!(log.observe(Some(&kernel(22))), RetryReport::Repeated);
+        assert_eq!(
+            log.observe(Some(&kernel(1))),
+            RetryReport::Failed,
+            "a different failure is a different outcome and is worth a line"
+        );
+        assert_eq!(log.observe(None), RetryReport::Recovered);
+        assert_eq!(log.observe(None), RetryReport::Clean);
+        assert_eq!(
+            log.observe(Some(&kernel(1))),
+            RetryReport::Failed,
+            "a failure after recovery warns again"
+        );
+    }
+
+    #[test]
+    fn failed_port_forward_rebuild_remains_pending() {
+        let mut mgr = manager_with_mappings(1);
+        mgr.pre_chain = Chain::new(&mgr.table);
+        let forward = PortForward {
+            proto: Proto::Tcp,
+            listen_port: 8080,
+            target: SocketAddrV6::new(Ipv6Addr::LOCALHOST, 80, 0, 0),
+        };
+        assert!(mgr.set_port_forwards(&[forward]).is_err());
+        assert_eq!(mgr.port_forwards.len(), 1);
+        assert!(mgr.rebuild_pending);
+        assert!(mgr.set_port_forwards(&[]).is_err());
+        assert!(mgr.port_forwards.is_empty());
+        assert!(mgr.rebuild_pending);
+    }
+
+    #[test]
+    #[ignore = "requires CAP_NET_ADMIN and nft in an isolated network namespace"]
+    fn kernel_rejection_retries_latest_state_without_another_mapping_event() {
+        let mut mgr = manager_with_mappings(2);
+        let table_name = format!("{TABLE_NAME}_retry_test_{}", std::process::id());
+        mgr.table = Table::new(ProtocolFamily::Inet).with_name(&table_name);
+        mgr.pre_chain = Chain::new(&mgr.table)
+            .with_name(PREROUTING_CHAIN)
+            .with_type(ChainType::Nat)
+            .with_hook(Hook::new(HookClass::PreRouting, DSTNAT_PRIORITY));
+        mgr.post_chain = Chain::new(&mgr.table)
+            .with_name(POSTROUTING_CHAIN)
+            .with_type(ChainType::Nat)
+            .with_hook(Hook::new(HookClass::PostRouting, SRCNAT_PRIORITY));
+        mgr.rebuild().unwrap();
+        let listing = || {
+            let output = std::process::Command::new("nft")
+                .args(["-j", "list", "table", "inet", &table_name])
+                .output()
+                .unwrap();
+            assert!(output.status.success());
+            output.stdout
+        };
+        let before = listing();
+        // Encoding succeeds, but the kernel rejects an overlong chain name.
+        let invalid = Chain::new(&mgr.table).with_name("x".repeat(300));
+        let original = std::mem::replace(&mut mgr.pre_chain, invalid);
+        assert!(matches!(
+            mgr.add_mapping(vip(3), mesh(3)),
+            Err(NatError::Kernel { .. })
+        ));
+        assert!(mgr.remove_mapping(vip(1)).is_err());
+        assert!(mgr.remove_mapping(vip(3)).is_err());
+        assert!(mgr.add_mapping(vip(2), mesh(99)).is_err());
+        assert!(mgr.retry_pending().is_err());
+        assert_eq!(
+            listing(),
+            before,
+            "rejection leaves the old kernel table intact"
+        );
+
+        mgr.pre_chain = original;
+        assert!(mgr.retry_pending().unwrap());
+        assert!(!mgr.retry_pending().unwrap());
+        assert_eq!(mgr.mapping_count(), 1);
+        assert_eq!(mgr.mappings[&vip(2)].mesh_addr, mesh(99));
+        let after = listing();
+        assert_ne!(after, before);
+        let table: serde_json::Value = serde_json::from_slice(&after).unwrap();
+        assert_eq!(
+            table["nftables"]
+                .as_array()
+                .unwrap()
+                .iter()
+                .filter(|entry| entry.get("rule").is_some())
+                .count(),
+            3
+        );
+        assert!(String::from_utf8(after).unwrap().contains("fd02::63"));
+
+        // A successful normal update also clears earlier pending work.
+        mgr.pre_chain = Chain::new(&mgr.table);
+        assert!(mgr.add_mapping(vip(4), mesh(4)).is_err());
+        mgr.pre_chain = Chain::new(&mgr.table)
+            .with_name(PREROUTING_CHAIN)
+            .with_type(ChainType::Nat)
+            .with_hook(Hook::new(HookClass::PreRouting, DSTNAT_PRIORITY));
+        mgr.remove_mapping(vip(4)).unwrap();
+        assert!(!mgr.retry_pending().unwrap());
+        mgr.cleanup().unwrap();
+    }
+
     #[test]
     fn rebuild_deletes_and_recreates_the_table_inside_one_batch() {
         let batches = manager_with_mappings(3).rebuild_batches();
diff --git a/src/testutil.rs b/src/testutil.rs
index 65874819..f55e42e8 100644
--- a/src/testutil.rs
+++ b/src/testutil.rs
@@ -19,6 +19,11 @@ pub(crate) fn make_node_addr(val: u8) -> NodeAddr {
 pub(crate) struct LogCapture(std::sync::Arc<std::sync::Mutex<Vec<String>>>);
 
 impl LogCapture {
+    /// Every captured line, each prefixed with its level.
+    pub(crate) fn lines(&self) -> Vec<String> {
+        self.0.lock().unwrap().clone()
+    }
+
     /// Only the captured lines emitted at WARN.
     pub(crate) fn warnings(&self) -> Vec<String> {
         self.0
diff --git a/src/transport/ethernet/io_linux.rs b/src/transport/ethernet/io_linux.rs
index cc0bae0c..5c26be69 100644
--- a/src/transport/ethernet/io_linux.rs
+++ b/src/transport/ethernet/io_linux.rs
@@ -1,7 +1,7 @@
 //! AF_PACKET socket creation, binding, and ioctl helpers (Linux).
 
 use crate::transport::TransportError;
-use std::os::unix::io::{AsRawFd, RawFd};
+use std::os::unix::io::{AsRawFd, FromRawFd, IntoRawFd, OwnedFd, RawFd};
 
 /// Wrapper around an AF_PACKET SOCK_DGRAM file descriptor.
 ///
@@ -39,6 +39,9 @@ impl PacketSocket {
                 err
             )));
         }
+        // SAFETY: socket returned a new descriptor, owned only by this guard
+        // until construction succeeds. Every setup error closes it on drop.
+        let owned_fd = unsafe { OwnedFd::from_raw_fd(fd) };
 
         // Look up interface index
         let if_index = get_if_index(fd, interface)?;
@@ -58,7 +61,6 @@ impl PacketSocket {
         };
         if ret < 0 {
             let err = std::io::Error::last_os_error();
-            unsafe { libc::close(fd) };
             // The interface can disappear between the index lookup and the
             // bind. That is absence arriving a few microseconds late, not a
             // configuration fault, so it reports as absence.
@@ -77,7 +79,6 @@ impl PacketSocket {
         let flags = unsafe { libc::fcntl(fd, libc::F_GETFL) };
         if flags < 0 {
             let err = std::io::Error::last_os_error();
-            unsafe { libc::close(fd) };
             return Err(TransportError::StartFailed(format!(
                 "fcntl(F_GETFL) failed: {}",
                 err
@@ -86,7 +87,6 @@ impl PacketSocket {
         let ret = unsafe { libc::fcntl(fd, libc::F_SETFL, flags | libc::O_NONBLOCK) };
         if ret < 0 {
             let err = std::io::Error::last_os_error();
-            unsafe { libc::close(fd) };
             return Err(TransportError::StartFailed(format!(
                 "fcntl(F_SETFL, O_NONBLOCK) failed: {}",
                 err
@@ -94,7 +94,7 @@ impl PacketSocket {
         }
 
         Ok(Self {
-            fd,
+            fd: owned_fd.into_raw_fd(),
             if_index,
             ethertype,
         })
@@ -354,3 +354,69 @@ fn get_if_mtu(fd: RawFd, if_index: i32) -> Result<u16, TransportError> {
     let mtu = unsafe { ifr.ifr_ifru.ifru_mtu } as u16;
     Ok(mtu)
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn invalid_interface_does_not_leak_socket() {
+        const CHILD: &str = "FIPS_TEST_PACKET_SOCKET_CHILD";
+        if std::env::var_os(CHILD).is_none() {
+            if PacketSocket::open("lo", 0x88b5).is_err() {
+                // Unprivileged: every open fails at socket() before the paths
+                // under test. Loud on a runner that claims privilege, so this
+                // cannot go quiet.
+                assert!(
+                    std::env::var_os("FIPS_TEST_PRIVILEGED").is_none(),
+                    "this runner declares FIPS_TEST_PRIVILEGED but cannot open a \
+                     raw socket on lo: the descriptor leak check went unexercised"
+                );
+                return;
+            }
+            // Count descriptors in a separate process so concurrently running
+            // tests cannot affect the result.
+            let output = std::process::Command::new(std::env::current_exe().unwrap())
+                .args([
+                    "--exact",
+                    "transport::ethernet::io::platform::tests::invalid_interface_does_not_leak_socket",
+                    "--include-ignored",
+                    "--nocapture",
+                ])
+                .env(CHILD, "1")
+                .output()
+                .unwrap();
+            let stdout = String::from_utf8_lossy(&output.stdout);
+            let stderr = String::from_utf8_lossy(&output.stderr);
+            assert!(output.status.success(), "{stdout}\n{stderr}");
+            // libtest exits 0 when the filter selects nothing, so require
+            // that the child actually ran this test.
+            assert!(
+                stdout.contains("1 passed"),
+                "child did not run the test:\n{stdout}\n{stderr}"
+            );
+            return;
+        }
+
+        let open_fds = || std::fs::read_dir("/proc/self/fd").unwrap().count();
+        let before = open_fds();
+        for interface in ["fips-missing-interface", "invalid\0interface"] {
+            for _ in 0..8 {
+                let result = PacketSocket::open(interface, 0x88b5);
+                let expected = match &result {
+                    Err(TransportError::InterfaceUnavailable { .. }) => true,
+                    Err(TransportError::StartFailed(message)) => {
+                        message.starts_with("invalid interface name:")
+                    }
+                    _ => false,
+                };
+                assert!(
+                    expected,
+                    "unexpected result for {interface:?}: {:?}",
+                    result.as_ref().map(|_| "PacketSocket")
+                );
+            }
+        }
+        assert_eq!(open_fds(), before, "failed setup leaked a socket");
+    }
+}
diff --git a/src/transport/mod.rs b/src/transport/mod.rs
index 81a25fee..585508ac 100644
--- a/src/transport/mod.rs
+++ b/src/transport/mod.rs
@@ -483,9 +483,12 @@ impl TransportAddr {
     /// 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::new(buf)
+        // The longest form is 58 bytes: [39-byte IPv6%10-digit scope]:65535.
+        let mut buf = [0u8; 64];
+        let mut cursor = &mut buf[..];
+        write!(&mut cursor, "{addr}").expect("64 bytes is enough for any SocketAddr");
+        let len = 64 - cursor.len();
+        Self::from_bytes(&buf[..len])
     }
 }
 
@@ -1269,28 +1272,23 @@ impl TransportHandle {
 pub(crate) async fn resolve_socket_addr(
     addr: &TransportAddr,
 ) -> Result<SocketAddr, TransportError> {
+    resolve_socket_addrs(addr).await?.next().ok_or_else(|| {
+        TransportError::InvalidAddress(format!("DNS resolution returned no addresses for {}", addr))
+    })
+}
+
+/// Resolve every socket address in resolver order, bypassing DNS for numeric IPs.
+pub(crate) async fn resolve_socket_addrs(
+    addr: &TransportAddr,
+) -> Result<impl Iterator<Item = SocketAddr>, TransportError> {
     let s = addr
         .as_str()
         .ok_or_else(|| TransportError::InvalidAddress("not valid UTF-8".into()))?;
 
-    // Fast path: numeric IP address — no DNS lookup
-    if let Ok(sock_addr) = s.parse::<SocketAddr>() {
-        return Ok(sock_addr);
-    }
-
-    // Slow path: DNS resolution
-    tokio::net::lookup_host(s)
-        .await
-        .map_err(|e| {
-            TransportError::InvalidAddress(format!("DNS resolution failed for {}: {}", s, e))
-        })?
-        .next()
-        .ok_or_else(|| {
-            TransportError::InvalidAddress(format!(
-                "DNS resolution returned no addresses for {}",
-                s
-            ))
-        })
+    // lookup_host handles numeric addresses without allocating or querying DNS.
+    tokio::net::lookup_host(s).await.map_err(|e| {
+        TransportError::InvalidAddress(format!("DNS resolution failed for {}: {}", s, e))
+    })
 }
 
 // ============================================================================
@@ -1400,11 +1398,28 @@ mod tests {
 
     #[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"));
+        for text in [
+            "0.0.0.0:0",
+            "127.0.0.1:2121",
+            "255.255.255.255:65535",
+            "[::1]:2121",
+            "[2001:db8::1]:2121",
+            "[::ffff:255.255.255.255]:65535",
+            "[fe80::1%4294967295]:65535",
+            "[ffff:ffff:ffff:ffff:ffff:ffff:ffff:ffff%4294967295]:65535",
+        ] {
+            let socket: std::net::SocketAddr = text.parse().unwrap();
+            let addr = TransportAddr::from_socket_addr(socket);
+            assert_eq!(addr.as_str(), Some(text));
+            assert_eq!(addr.as_bytes(), socket.to_string().as_bytes());
+            assert_eq!(
+                addr.as_str()
+                    .unwrap()
+                    .parse::<std::net::SocketAddr>()
+                    .unwrap(),
+                socket
+            );
+        }
     }
 
     #[test]
diff --git a/src/transport/tcp/mod.rs b/src/transport/tcp/mod.rs
index 4b6c67bd..82450626 100644
--- a/src/transport/tcp/mod.rs
+++ b/src/transport/tcp/mod.rs
@@ -25,7 +25,7 @@
 mod pool;
 pub mod stats;
 
-use super::resolve_socket_addr;
+use super::resolve_socket_addrs;
 use super::{
     ConnectionState, DiscoveredPeer, PacketTx, ReceivedPacket, Transport, TransportAddr,
     TransportError, TransportId, TransportState, TransportType,
@@ -388,25 +388,22 @@ impl TcpTransport {
     /// Configures socket options, reads TCP_MAXSEG for MTU, splits the
     /// stream, spawns a receive task, and stores in the pool.
     async fn connect(&self, addr: &TransportAddr) -> Result<mpsc::Sender<Vec<u8>>, TransportError> {
-        let socket_addr = resolve_socket_addr(addr).await?;
+        let socket_addrs: Vec<_> = resolve_socket_addrs(addr).await?.collect();
         let timeout_ms = self.config.connect_timeout_ms();
 
-        // Connect with timeout
-        let stream = match tokio::time::timeout(
-            Duration::from_millis(timeout_ms),
-            TcpStream::connect(socket_addr),
-        )
-        .await
-        {
-            Ok(Ok(stream)) => stream,
-            Ok(Err(_)) => {
+        let connected =
+            connect_to_any_addr(self.transport_id, addr, &socket_addrs, timeout_ms).await;
+        let stream = match connected {
+            Ok(stream) => stream,
+            Err(error @ TransportError::ConnectionRefused) => {
                 self.stats.record_connect_refused();
-                return Err(TransportError::ConnectionRefused);
+                return Err(error);
             }
-            Err(_) => {
+            Err(error @ TransportError::Timeout) => {
                 self.stats.record_connect_timeout();
-                return Err(TransportError::Timeout);
+                return Err(error);
             }
+            Err(error) => return Err(error),
         };
 
         // Configure socket options via socket2
@@ -557,10 +554,8 @@ impl TcpTransport {
         }
 
         // Validate address is UTF-8 before spawning (fail fast on bad input)
-        let addr_string = addr
-            .as_str()
-            .ok_or_else(|| TransportError::InvalidAddress("not valid UTF-8".into()))?
-            .to_string();
+        addr.as_str()
+            .ok_or_else(|| TransportError::InvalidAddress("not valid UTF-8".into()))?;
         let timeout_ms = self.config.connect_timeout_ms();
         let config = self.config.clone();
         let transport_id = self.transport_id;
@@ -575,51 +570,22 @@ impl TcpTransport {
 
         let task = tokio::spawn(async move {
             // Resolve address (may involve DNS for hostnames)
-            let socket_addr: SocketAddr = if let Ok(sa) = addr_string.parse() {
-                sa
-            } else {
-                tokio::net::lookup_host(&addr_string)
-                    .await
-                    .map_err(|e| {
-                        TransportError::InvalidAddress(format!(
-                            "DNS resolution failed for {}: {}",
-                            addr_string, e
-                        ))
-                    })?
-                    .next()
-                    .ok_or_else(|| {
-                        TransportError::InvalidAddress(format!(
-                            "DNS resolution returned no addresses for {}",
-                            addr_string
-                        ))
-                    })?
-            };
+            let socket_addrs: Vec<_> = resolve_socket_addrs(&remote_addr).await?.collect();
 
-            // Connect with timeout
-            let stream = match tokio::time::timeout(
-                Duration::from_millis(timeout_ms),
-                TcpStream::connect(socket_addr),
-            )
-            .await
-            {
-                Ok(Ok(stream)) => stream,
-                Ok(Err(e)) => {
-                    debug!(
-                        transport_id = %transport_id,
-                        remote_addr = %remote_addr,
-                        error = %e,
-                        "Background TCP connect refused"
-                    );
-                    return Err(TransportError::ConnectionRefused);
-                }
-                Err(_) => {
+            // A refusal is logged with its OS error by connect_to_any_addr.
+            let connected =
+                connect_to_any_addr(transport_id, &remote_addr, &socket_addrs, timeout_ms).await;
+            let stream = match connected {
+                Ok(stream) => stream,
+                Err(error @ TransportError::Timeout) => {
                     debug!(
                         transport_id = %transport_id,
                         remote_addr = %remote_addr,
                         "Background TCP connect timed out"
                     );
-                    return Err(TransportError::Timeout);
+                    return Err(error);
                 }
+                Err(error) => return Err(error),
             };
 
             // Configure socket options via socket2
@@ -1280,6 +1246,43 @@ async fn tcp_receive_loop(
 // Socket Configuration Helpers
 // ============================================================================
 
+/// Connect to the first reachable of a peer's resolved addresses within one timeout.
+async fn connect_to_any_addr(
+    transport_id: TransportId,
+    remote_addr: &TransportAddr,
+    socket_addrs: &[SocketAddr],
+    timeout_ms: u64,
+) -> Result<TcpStream, TransportError> {
+    if socket_addrs.is_empty() {
+        return Err(TransportError::InvalidAddress(format!(
+            "DNS resolution returned no addresses for {}",
+            remote_addr
+        )));
+    }
+
+    // Try candidates in resolver order within one overall connection timeout.
+    match tokio::time::timeout(
+        Duration::from_millis(timeout_ms),
+        TcpStream::connect(socket_addrs),
+    )
+    .await
+    {
+        Ok(Ok(stream)) => Ok(stream),
+        Ok(Err(error)) => {
+            // Tokio returns the last candidate's error; keep it in the log,
+            // since the returned variant does not carry it.
+            debug!(
+                transport_id = %transport_id,
+                remote_addr = %remote_addr,
+                error = %error,
+                "TCP connect failed"
+            );
+            Err(TransportError::ConnectionRefused)
+        }
+        Err(_) => Err(TransportError::Timeout),
+    }
+}
+
 /// Configure a TCP socket with the transport's settings.
 fn configure_socket(
     stream: &std::net::TcpStream,
@@ -1432,6 +1435,130 @@ mod tests {
         }
     }
 
+    #[tokio::test]
+    async fn test_connect_tries_later_candidates() {
+        let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+        let good_addr = listener.local_addr().unwrap();
+
+        // No listener can bind port zero: binding it allocates an ephemeral port.
+        let bad_addr = "127.0.0.1:0".parse().unwrap();
+
+        let remote = TransportAddr::from_string(&format!("localhost:{}", good_addr.port()));
+        let stream =
+            connect_to_any_addr(TransportId::new(1), &remote, &[bad_addr, good_addr], 1_000)
+                .await
+                .expect("second TCP candidate should connect");
+        let (accepted, _) = timeout(Duration::from_secs(1), listener.accept())
+            .await
+            .unwrap()
+            .unwrap();
+        assert_eq!(stream.peer_addr().unwrap(), good_addr);
+        assert_eq!(accepted.peer_addr().unwrap(), stream.local_addr().unwrap());
+    }
+
+    #[tokio::test]
+    async fn test_connect_candidates_fail() {
+        let id = TransportId::new(1);
+        let remote = TransportAddr::from_string("localhost:0");
+        assert!(matches!(
+            connect_to_any_addr(id, &remote, &[], 1_000).await,
+            Err(TransportError::InvalidAddress(ref message)) if message.ends_with("localhost:0")
+        ));
+        assert!(matches!(
+            connect_to_any_addr(id, &remote, &["127.0.0.1:0".parse().unwrap()], 1_000).await,
+            Err(TransportError::ConnectionRefused)
+        ));
+    }
+
+    #[tokio::test]
+    async fn test_connect_failure_logs_os_error_with_peer() {
+        let remote = TransportAddr::from_string("localhost:0");
+        let refused = ["127.0.0.1:0".parse().unwrap()];
+        // Register the log callsite before installing the capture. Other
+        // tests reach it in parallel, and one that registers it first while
+        // this capture is the only subscriber caches its own thread's "never"
+        // interest; installing the capture rebuilds already-registered
+        // callsites, so registering first makes the capture see the event.
+        let _ = connect_to_any_addr(TransportId::new(7), &remote, &refused, 1_000).await;
+        let (logs, guard) = crate::testutil::capture_logs_scoped();
+        let result = connect_to_any_addr(TransportId::new(7), &remote, &refused, 1_000).await;
+        drop(guard);
+
+        assert!(matches!(result, Err(TransportError::ConnectionRefused)));
+        // The returned variant always reads "connection refused"; the log
+        // line must carry the OS error and the peer instead.
+        let lines = logs.lines();
+        assert!(
+            lines.iter().any(|line| line.starts_with("DEBUG")
+                && line.contains("TCP connect failed")
+                && line.contains("transport_id=transport:7")
+                && line.contains("remote_addr=localhost:0")
+                && line.contains("os error")),
+            "expected a debug line with the OS error and peer, got {lines:?}",
+        );
+    }
+
+    async fn send_to_hostname(background: bool) {
+        // Prefer the last localhost address so dual-stack hosts also
+        // exercise fallback through the full transport connection path.
+        let mut addresses: Vec<_> = tokio::net::lookup_host("localhost:0")
+            .await
+            .unwrap()
+            .collect();
+        addresses.reverse();
+        let listener = TcpListener::bind(addresses.as_slice()).await.unwrap();
+        let (tx, _rx) = packet_channel(10);
+        let config = make_outbound_config();
+        // A refused localhost connection can take over two seconds on Windows.
+        // Allow the configured connection timeout, plus DNS/scheduling margin.
+        let connect_deadline =
+            Duration::from_millis(config.connect_timeout_ms()) + Duration::from_secs(2);
+        let mut sender = TcpTransport::new(TransportId::new(1), None, config, tx);
+        sender.start_async().await.unwrap();
+        let remote = TransportAddr::from_string(&format!(
+            "localhost:{}",
+            listener.local_addr().unwrap().port()
+        ));
+
+        if background {
+            sender.connect_async(&remote).await.unwrap();
+            assert!(
+                wait_until(
+                    || sender.connection_state_sync(&remote) == ConnectionState::Connected,
+                    connect_deadline,
+                )
+                .await
+            );
+        }
+        timeout(
+            connect_deadline,
+            sender.send_async(&remote, &build_msg1_frame()),
+        )
+        .await
+        .expect("hostname send exceeded connection deadline")
+        .unwrap();
+        let (mut stream, _) = timeout(Duration::from_secs(2), listener.accept())
+            .await
+            .unwrap()
+            .unwrap();
+        let packet = timeout(Duration::from_secs(2), read_fmp_packet(&mut stream, 1400))
+            .await
+            .unwrap()
+            .unwrap();
+        assert_eq!(packet, build_msg1_frame());
+        sender.stop_async().await.unwrap();
+    }
+
+    #[tokio::test]
+    async fn test_connect_on_send_hostname() {
+        send_to_hostname(false).await;
+    }
+
+    #[tokio::test]
+    async fn test_connect_async_hostname() {
+        send_to_hostname(true).await;
+    }
+
     fn make_outbound_config() -> TcpConfig {
         TcpConfig {
             bind_addr: None,
diff --git a/src/transport/types.rs b/src/transport/types.rs
index 4e95f4ce..7557279c 100644
--- a/src/transport/types.rs
+++ b/src/transport/types.rs
@@ -8,6 +8,7 @@
 //! via `pub use types::*`, so existing `crate::transport::{LinkId, ...}`
 //! imports are unaffected.
 
+use alloc::{string::String, sync::Arc, vec::Vec};
 use core::fmt;
 use core::time::Duration;
 
@@ -91,23 +92,28 @@ impl fmt::Display for LinkDirection {
 /// Each transport type interprets this differently:
 /// - UDP/TCP: "host:port" (IP address or DNS hostname)
 /// - Ethernet: MAC address (6 bytes)
+///
+/// The immutable bytes are shared across clones.
 #[derive(Clone, PartialEq, Eq, Hash)]
-pub struct TransportAddr(Vec<u8>);
+pub struct TransportAddr(Arc<[u8]>);
 
 impl TransportAddr {
     /// Create a transport address from raw bytes.
+    ///
+    /// Copies the bytes into shared storage; the vector's allocation is not
+    /// reused. Prefer [`Self::from_bytes`] when a byte slice is already available.
     pub fn new(bytes: Vec<u8>) -> Self {
-        Self(bytes)
+        Self(bytes.into())
     }
 
     /// Create a transport address from a byte slice.
     pub fn from_bytes(bytes: &[u8]) -> Self {
-        Self(bytes.to_vec())
+        Self(Arc::from(bytes))
     }
 
     /// Create a transport address from a string.
     pub fn from_string(s: &str) -> Self {
-        Self(s.as_bytes().to_vec())
+        Self::from_bytes(s.as_bytes())
     }
 
     /// Get the raw bytes.
@@ -158,7 +164,7 @@ impl fmt::Display for TransportAddr {
                 Ok(())
             }
             None => {
-                for byte in &self.0 {
+                for byte in self.0.iter() {
                     write!(f, "{:02x}", byte)?;
                 }
                 Ok(())
@@ -175,7 +181,36 @@ impl From<&str> for TransportAddr {
 
 impl From<String> for TransportAddr {
     fn from(s: String) -> Self {
-        Self(s.into_bytes())
+        Self::from_string(&s)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::TransportAddr;
+
+    #[test]
+    fn transport_addr_clone_shares_immutable_bytes() {
+        let addr = TransportAddr::from_string("192.168.1.1:2121");
+        let cloned = addr.clone();
+
+        assert_eq!(addr, cloned);
+        assert_eq!(addr.as_bytes().as_ptr(), cloned.as_bytes().as_ptr());
+    }
+
+    #[test]
+    fn transport_addr_equality_and_hash_use_byte_values() {
+        // Only tests use std; the transport primitives remain alloc-only.
+        use std::collections::HashSet;
+
+        let original = TransportAddr::from_string("192.168.1.1:2121");
+        let same_value = TransportAddr::from_bytes(b"192.168.1.1:2121");
+        let mut addrs = HashSet::new();
+
+        assert_ne!(original.as_bytes().as_ptr(), same_value.as_bytes().as_ptr());
+        assert_eq!(original, same_value);
+        addrs.insert(original);
+        assert!(addrs.contains(&same_value));
     }
 }
 
diff --git a/testing/README.md b/testing/README.md
index 779d0d67..f1d96569 100644
--- a/testing/README.md
+++ b/testing/README.md
@@ -110,6 +110,14 @@ what the real `build-apk.sh` and `build-ipk.sh` package. No router,
 opkg or apk-tools is involved. Part of both CI runners as
 `openwrt-scripts`.
 
+### [tarball-install/](tarball-install/) -- systemd Tarball Upgrade
+
+Runs the systemd tarball's `install.sh` twice in a Debian 12 systemd
+container, the second time as an upgrade, with stub binaries. Checks
+that the units running before the upgrade (fips, fips-dns and
+fips-gateway, in three combinations) are the ones running after it.
+Part of both CI runners as `tarball-install`.
+
 ### [native-api/](native-api/) -- Native Datagram API
 
 Checks the experimental native datagram API: a client process opens a
diff --git a/testing/check-deb-version.sh b/testing/check-deb-version.sh
new file mode 100755
index 00000000..a18b747b
--- /dev/null
+++ b/testing/check-deb-version.sh
@@ -0,0 +1,175 @@
+#!/bin/bash
+# ── Debian package version derivation check ─────────────────────────────────
+# A release candidate is tagged vX.Y.Z-rcN, because git refuses '~' in a ref
+# name. dpkg reads X.Y.Z-rcN as revision rcN of X.Y.Z and sorts it ABOVE the
+# release, so a host that installed the candidate is never upgraded by the
+# release. package-linux.yml's "Derive Linux package version" step therefore
+# maps the tag's pre-release suffix to '~' for the .deb, which dpkg sorts below.
+#
+# This runs that step's own text, taken from the workflow, rather than a copy,
+# so the check and the workflow cannot drift apart. Cases:
+#   refs/tags/v0.5.3      both versions 0.5.3
+#   refs/tags/v0.5.3-rc1  deb 0.5.3~rc1, tarball and artifact 0.5.3-rc1, and
+#                         dpkg orders 0.5.2 < 0.5.3~rc1 < 0.5.3
+#   refs/heads/maint      deb equals the tarball version, which is
+#                         <Cargo version>+maint.<height>.<hash>
+#
+# It also checks the wiring the derivation needs to have any effect: the job
+# declares the deb_package_version output, the .deb build passes it as
+# --version, and that step renames a '~' in the package file name to '-' (a
+# GitHub release renames an asset with special characters in its name, which
+# would leave checksums-linux.txt naming a file the release does not have).
+# Those three are read from the text, not executed.
+#
+# Exit 0 = clean. Exit 1 = a case or a wiring check failed. Exit 2 = the check
+# could not run (no dpkg, no PyYAML, the step not found, or an output missing);
+# never treated as a pass.
+# ─────────────────────────────────────────────────────────────────────────────
+set -uo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+PROJECT_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
+WORKFLOW="$PROJECT_ROOT/.github/workflows/package-linux.yml"
+
+cant() {
+    echo "check-deb-version: $*; cannot verify the Debian version derivation" >&2
+    exit 2
+}
+
+[ -f "$WORKFLOW" ] || cant "missing $WORKFLOW"
+command -v dpkg >/dev/null 2>&1 || cant "dpkg not found"
+command -v python3 >/dev/null 2>&1 || cant "python3 not found"
+python3 -c "import yaml" >/dev/null 2>&1 || cant "python3 module 'yaml' not found"
+
+WORK=$(mktemp -d) || cant "mktemp failed"
+trap 'rm -rf "$WORK"' EXIT
+
+# Pull the derivation step's run text, the job's declared outputs and the
+# .deb build step's run text out of the workflow.
+if ! python3 - "$WORKFLOW" "$WORK" <<'PY'
+import sys
+from pathlib import Path
+
+import yaml
+
+workflow, work = sys.argv[1], Path(sys.argv[2])
+doc = yaml.safe_load(Path(workflow).read_text(encoding="utf-8"))
+jobs = doc.get("jobs") or {}
+
+
+def step_run(job, name):
+    for step in (jobs.get(job) or {}).get("steps") or []:
+        if step.get("name") == name:
+            return step.get("run")
+    return None
+
+
+derive = step_run("determine-versioning", "Derive Linux package version")
+build = step_run("build", "Build Debian package in the pinned container")
+if not derive or not build:
+    missing = "derivation" if not derive else "Debian build"
+    print(f"check-deb-version: the {missing} step was not found in {workflow}",
+          file=sys.stderr)
+    sys.exit(2)
+(work / "derive.sh").write_text(derive, encoding="utf-8")
+(work / "build.sh").write_text(build, encoding="utf-8")
+outputs = (jobs.get("determine-versioning") or {}).get("outputs") or {}
+(work / "outputs").write_text(
+    "".join(f"{k}={v}\n" for k, v in outputs.items()), encoding="utf-8")
+PY
+then
+    exit 2
+fi
+
+FAILED=0
+ok()  { echo "  PASS  $*"; }
+bad() { echo "  FAIL  $*"; FAILED=$((FAILED + 1)); }
+
+# Run the derivation for one ref and load its outputs into LINUX and DEB.
+# Exits 2 when the step fails or does not write both outputs: no version was
+# established, which is not a failed case.
+derive() {
+    local ref="$1" out="$WORK/github_output"
+    : > "$out"
+    if ! (cd "$PROJECT_ROOT" && GITHUB_OUTPUT="$out" GITHUB_REF="$ref" \
+            GITHUB_REF_NAME="${ref#refs/*/}" bash -eo pipefail "$WORK/derive.sh") >"$WORK/derive.log" 2>&1; then
+        echo "check-deb-version: the derivation failed for $ref:" >&2
+        cat "$WORK/derive.log" >&2
+        exit 2
+    fi
+    LINUX=$(sed -n 's/^linux_package_version=//p' "$out")
+    DEB=$(sed -n 's/^deb_package_version=//p' "$out")
+    if [ -z "$LINUX" ] || [ -z "$DEB" ]; then
+        cant "the derivation for $ref did not write both outputs (linux '$LINUX', deb '$DEB')"
+    fi
+    return 0
+}
+
+# Record whether dpkg orders $1 $2 $3 (lt, gt, ...).
+expect_order() {
+    if dpkg --compare-versions "$1" "$2" "$3"; then
+        ok "dpkg: $1 $2 $3"
+    else
+        bad "dpkg: $1 $2 $3 does not hold"
+    fi
+    return 0
+}
+
+# Record whether a derived version equals the expected one.
+expect_eq() {
+    local label="$1" got="$2" want="$3"
+    if [ "$got" = "$want" ]; then
+        ok "$label: $got"
+    else
+        bad "$label: $got (want $want)"
+    fi
+    return 0
+}
+
+echo "Release tag refs/tags/v0.5.3"
+derive refs/tags/v0.5.3
+expect_eq "linux_package_version" "$LINUX" "0.5.3"
+expect_eq "deb_package_version" "$DEB" "0.5.3"
+
+echo "Candidate tag refs/tags/v0.5.3-rc1"
+derive refs/tags/v0.5.3-rc1
+expect_eq "linux_package_version" "$LINUX" "0.5.3-rc1"
+expect_eq "deb_package_version" "$DEB" "0.5.3~rc1"
+expect_order "$DEB" lt 0.5.3
+expect_order "$DEB" gt 0.5.2
+
+echo "Branch refs/heads/maint"
+derive refs/heads/maint
+CRATE=$(awk -F'"' '/^version = /{print $2; exit}' "$PROJECT_ROOT/Cargo.toml")
+HEIGHT=$(git -C "$PROJECT_ROOT" rev-list --count HEAD) || cant "git rev-list failed"
+HASH=$(git -C "$PROJECT_ROOT" rev-parse --short HEAD) || cant "git rev-parse failed"
+[ -n "$CRATE" ] || cant "no version in Cargo.toml"
+expect_eq "linux_package_version" "$LINUX" "${CRATE}+maint.${HEIGHT}.${HASH}"
+expect_eq "deb_package_version" "$DEB" "$LINUX"
+
+echo "Wiring"
+# shellcheck disable=SC2016  # the ${{ }} expressions are workflow text, not shell
+if grep -qxF 'deb_package_version=${{ steps.linux_version.outputs.deb_package_version }}' "$WORK/outputs"; then
+    ok "determine-versioning declares the deb_package_version output"
+else
+    bad "determine-versioning does not declare deb_package_version from the derivation step"
+fi
+# shellcheck disable=SC2016
+if grep -qF -- '--version "${{ needs.determine-versioning.outputs.deb_package_version }}"' "$WORK/build.sh"; then
+    ok "the .deb build passes deb_package_version as --version"
+else
+    bad "the .deb build does not pass deb_package_version as --version"
+fi
+if grep -qF "tr '~' '-'" "$WORK/build.sh"; then
+    ok "the .deb build renames a '~' in the package file name"
+else
+    bad "the .deb build does not rename a '~' in the package file name"
+fi
+
+echo
+if [ "$FAILED" -eq 0 ]; then
+    echo "check-deb-version: all checks passed"
+    exit 0
+fi
+echo "check-deb-version: $FAILED check(s) failed"
+exit 1
diff --git a/testing/ci-local.sh b/testing/ci-local.sh
index d985b144..07b573ae 100755
--- a/testing/ci-local.sh
+++ b/testing/ci-local.sh
@@ -224,6 +224,7 @@ CHAOS_SUITES=(
 #     testing/chaos/scripts/chaos.sh bloom-storm.
 GATEWAY_SUITES=(gateway)
 OPENWRT_SUITES=(openwrt-scripts)
+TARBALL_INSTALL_SUITES=(tarball-install)
 SIDECAR_SUITES=(sidecar)
 FIREWALL_SUITES=(firewall)
 IFACE_BINDING_SUITES=(iface-binding)
@@ -266,6 +267,9 @@ list_suites() {
     echo "  OpenWrt packaging:"
     for s in "${OPENWRT_SUITES[@]}"; do echo "    $s"; done
     echo ""
+    echo "  systemd tarball:"
+    for s in "${TARBALL_INSTALL_SUITES[@]}"; do echo "    $s"; done
+    echo ""
     echo "  Firewall baseline:"
     for s in "${FIREWALL_SUITES[@]}"; do echo "    $s"; done
     echo ""
@@ -1402,6 +1406,15 @@ run_integration() {
         return
     fi
 
+    # Also ahead of the build context, for the same reason: the tarball
+    # install scenarios use stub binaries and their own systemd image.
+    if [[ -z "$ONLY_SUITE" ]]; then
+        run_tarball_install
+    elif [[ "$ONLY_SUITE" == "tarball-install" ]]; then
+        run_tarball_install
+        return
+    fi
+
     # Populate THIS run's build context, then install the binaries into it.
     # Everything but the binaries is copied from the tracked context directory;
     # the binaries are installed fresh, and a previous run's are deliberately
@@ -1576,6 +1589,8 @@ run_suite() {
             run_gateway ;;
         openwrt-scripts)
             run_openwrt_scripts ;;
+        tarball-install)
+            run_tarball_install ;;
         firewall)
             run_firewall ;;
         iface-binding)
@@ -1676,6 +1691,18 @@ run_openwrt_scripts() {
     return $rc
 }
 
+# The systemd tarball's install.sh, run as an upgrade under real systemd with
+# stub binaries: the units that were running before the upgrade must be the
+# ones running after it, including fips-dns and fips-gateway, which stop with
+# fips through Requires=.
+run_tarball_install() {
+    local rc=0
+    info "[tarball-install] Running the systemd tarball upgrade scenarios"
+    bash "$SCRIPT_DIR/tarball-install/test.sh" || rc=$?
+    record "tarball-install" $rc
+    return $rc
+}
+
 run_ci_parity() {
     local rc=0
     info "[ci-parity] Comparing the local suite set against the GitHub matrix"
@@ -1755,6 +1782,17 @@ run_wait_converge() {
     record "wait-converge" $rc
 }
 
+# The .deb version package-linux.yml derives for a release tag, a candidate
+# tag and a branch. A candidate must sort below its release under dpkg, which
+# the tag's -rcN does not, so the workflow maps it to ~rcN; this runs the
+# workflow's own step text. Static, about a second, and needs nothing built.
+run_deb_version() {
+    local rc=0
+    info "[deb-version] Checking the Debian version derived for tags and branches"
+    bash "$SCRIPT_DIR/check-deb-version.sh" || rc=$?
+    record "deb-version" $rc
+}
+
 # ── Main ───────────────────────────────────────────────────────────────────
 
 main() {
@@ -1776,6 +1814,7 @@ main() {
     run_action_pins
     run_comment_refs
     run_wait_converge
+    run_deb_version
 
     if [[ "$TEST_ONLY" == true ]]; then
         run_tests
diff --git a/testing/interop/interop-test.sh b/testing/interop/interop-test.sh
index 913518fe..857e4b59 100755
--- a/testing/interop/interop-test.sh
+++ b/testing/interop/interop-test.sh
@@ -574,14 +574,19 @@ count_node_pattern() {
     docker logs "${CONTAINER[$node]}" 2>&1 | grep -cE "$pattern" || true
 }
 
+# Wait until the count of a pattern across all node logs reaches
+# min_count. Returns 0 when it does, 1 on timeout, and 2 at once when a
+# node's logs cannot be read, since no count is established then; the
+# caller reports that when it takes its own count.
 wait_for_log_pattern_count() {
     local pattern="$1" min_count="$2" timeout="$3"
-    local start=$SECONDS
-    while (( SECONDS - start < timeout )); do
-        [ "$(count_log_pattern "$pattern")" -ge "$min_count" ] && return 0
+    local start=$SECONDS count
+    while :; do
+        count="$(count_log_pattern "$pattern")" || return 2
+        [ "$count" -ge "$min_count" ] && return 0
+        (( SECONDS - start < timeout )) || return 1
         sleep "$LOG_POLL_INTERVAL"
     done
-    [ "$(count_log_pattern "$pattern")" -ge "$min_count" ]
 }
 
 # One node's bloom-union mesh-size estimate, or "null" when the daemon
@@ -866,8 +871,11 @@ PASSED=0; FAILED=0
 wait_for_log_pattern_count \
     "Rekey cutover complete \(initiator\), K-bit flipped" 1 \
     "$FIRST_REKEY_TIMEOUT" || true
-fmp_cutovers="$(count_log_pattern 'Rekey cutover complete \(initiator\), K-bit flipped')"
-if [ "$fmp_cutovers" -ge 1 ]; then
+if ! fmp_cutovers="$(count_log_pattern 'Rekey cutover complete \(initiator\), K-bit flipped')"; then
+    echo "  FAIL  FMP rekey cutovers: node logs unreadable ($fmp_cutovers), not established"
+    FAILED=$((FAILED + 1))
+    INTEROP_FAILURES+=("[log] FMP rekey cutovers: node logs unreadable ($fmp_cutovers)")
+elif [ "$fmp_cutovers" -ge 1 ]; then
     echo "  PASS  FMP rekey initiator cutovers: $fmp_cutovers"
     PASSED=$((PASSED + 1))
 else
@@ -897,10 +905,17 @@ echo "Phase 4: Second rekey cycle (waiting up to ${SECOND_REKEY_WAIT}s for the n
 # the same pre/post cutover-count delta convention as the control window
 # (Phase 1b), instead of a blind sleep. Bounded by SECOND_REKEY_WAIT so a
 # stalled rekey falls through to the strict Phase 5/6 assertions.
-fmp_cutovers_before="$(count_log_pattern 'Rekey cutover complete \(initiator\), K-bit flipped')"
-wait_for_log_pattern_count \
-    "Rekey cutover complete \(initiator\), K-bit flipped" \
-    "$((fmp_cutovers_before + 1))" "$SECOND_REKEY_WAIT" || true
+# An unreadable baseline gives nothing to wait beyond, so do not wait.
+# Phase 4 keeps no PASSED/FAILED of its own; the INTEROP_FAILURES entry
+# is what turns the run red.
+if fmp_cutovers_before="$(count_log_pattern 'Rekey cutover complete \(initiator\), K-bit flipped')"; then
+    wait_for_log_pattern_count \
+        "Rekey cutover complete \(initiator\), K-bit flipped" \
+        "$((fmp_cutovers_before + 1))" "$SECOND_REKEY_WAIT" || true
+else
+    echo "  FAIL  second rekey cycle: node logs unreadable ($fmp_cutovers_before), baseline not established"
+    INTEROP_FAILURES+=("[log] second rekey cycle: node logs unreadable ($fmp_cutovers_before)")
+fi
 echo ""
 
 echo "Phase 5: Post-second-rekey connectivity (reconverge within ${POST_REKEY_TIMEOUT}s)"
@@ -1015,16 +1030,22 @@ done
 # Positive checks — the rekey machinery actually exercised both layers.
 echo ""
 echo "  -- Rekey machinery exercised --"
-fmp_total="$(count_log_pattern 'Rekey cutover complete \(initiator\), K-bit flipped')"
-fsp_total="$(count_log_pattern 'FSP rekey cutover complete')"
-if [ "$fmp_total" -ge 1 ]; then
+# An unreadable log means the harness observed nothing, so it is a FAIL
+# on either layer; the FSP WARN below is for a readable zero only.
+if ! fmp_total="$(count_log_pattern 'Rekey cutover complete \(initiator\), K-bit flipped')"; then
+    echo "    FAIL  FMP rekey cutovers: node logs unreadable ($fmp_total), not established"
+    FAILED=$((FAILED + 1))
+elif [ "$fmp_total" -ge 1 ]; then
     echo "    PASS  FMP rekey cutovers across mesh: $fmp_total"
     PASSED=$((PASSED + 1))
 else
     echo "    FAIL  FMP rekey cutovers: $fmp_total (expected >= 1)"
     FAILED=$((FAILED + 1))
 fi
-if [ "$fsp_total" -ge 1 ]; then
+if ! fsp_total="$(count_log_pattern 'FSP rekey cutover complete')"; then
+    echo "    FAIL  FSP rekey cutovers: node logs unreadable ($fsp_total), not established"
+    FAILED=$((FAILED + 1))
+elif [ "$fsp_total" -ge 1 ]; then
     echo "    PASS  FSP rekey cutovers across mesh: $fsp_total"
     PASSED=$((PASSED + 1))
 else
diff --git a/testing/tarball-install/test.sh b/testing/tarball-install/test.sh
new file mode 100755
index 00000000..e3474095
--- /dev/null
+++ b/testing/tarball-install/test.sh
@@ -0,0 +1,367 @@
+#!/bin/bash
+# Test the systemd tarball's install.sh as an upgrade, under real systemd.
+#
+# fips-dns.service and fips-gateway.service carry Requires=fips.service, so
+# stopping fips stops both. An upgrade that checks which units are running
+# only after stopping fips sees the other two as already stopped, and starts
+# fips alone again. This suite installs the tarball once, puts the units in a
+# known state, runs install.sh a second time (the upgrade), and checks that
+# exactly the units that were running before are running after.
+#
+# The tarball is staged from the real packaging/systemd/install.sh and unit
+# files, and the real packaging/common configuration, the same file set
+# build-tarball.sh ships. The binaries and DNS helpers are stubs: the fips and
+# fips-gateway stubs sleep, so their units stay active, and the DNS helpers
+# exit 0. What is under test is install.sh against real units and a real
+# systemd, which carries the Requires= stop propagation; nothing here depends
+# on what the binaries do.
+#
+# Scenarios, each on a fresh container:
+#   all-active  fips, fips-dns and fips-gateway running before the upgrade.
+#               All three must be running after it, and fips must have been
+#               restarted (its InvocationID changes).
+#   fips-only   only fips running. fips must be running after, and the other
+#               two must still be stopped.
+#   none-active nothing running. Nothing may be running after.
+#
+# Usage: ./test.sh [--install-sh PATH] [scenario ...]
+#   No scenarios = run all of them.
+#   --install-sh PATH runs another install.sh in place of the tree's, so the
+#   suite can be pointed at an older script without editing the tree.
+#
+# Requirements: Docker able to grant SYS_ADMIN and NET_ADMIN and an unconfined
+# AppArmor profile (the container is not privileged; see
+# testing/lib/systemd-container.sh), and /dev/net/tun on the host.
+#
+# Exit 0 only when every check passed. A container that cannot boot or be
+# inspected is a FAIL, never a skip.
+
+set -uo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+# shellcheck source=SCRIPTDIR/../lib/systemd-container.sh
+source "$SCRIPT_DIR/../lib/systemd-container.sh"
+# shellcheck source=SCRIPTDIR/../lib/image-build.sh
+source "$SCRIPT_DIR/../lib/image-build.sh"
+REPO_ROOT="$(cd "$SCRIPT_DIR/../.." && pwd)"
+
+IMAGE="fips-tarball-test:debian12"
+BOOT_TIMEOUT=60
+# Bounds one install.sh run. It starts at most three units, and the gateway's
+# ExecStartPre returns at once because fips0 exists before the first start.
+INSTALL_TIMEOUT=120
+UNIT_START_TIMEOUT=60
+TARBALL_DIR=/opt/fips-tarball
+
+PASS=0
+FAIL=0
+
+INSTALL_SH="$REPO_ROOT/packaging/systemd/install.sh"
+STAGE=""
+
+log()  { echo "=== $*"; }
+pass() { echo "  PASS: $*"; PASS=$((PASS + 1)); }
+fail() { echo "  FAIL: $*"; FAIL=$((FAIL + 1)); }
+
+# Remove a container, ignoring one that does not exist.
+cleanup_container() {
+    local name="$1"
+    docker rm -f "$name" >/dev/null 2>&1 || true
+}
+
+# Remove every scenario's container and the staged tarball.
+cleanup_all() {
+    local s
+    for s in $ALL_SCENARIOS; do
+        cleanup_container "$(container_name "$s")"
+    done
+    [ -n "$STAGE" ] && rm -rf "$STAGE"
+}
+
+# The container name for a scenario, scoped to this CI run.
+container_name() {
+    echo "fips-tarball-test-$1${FIPS_CI_NAME_SUFFIX:-}"
+    return 0
+}
+
+# Build the systemd image. The build context is an empty directory: the
+# tarball is copied into the container after it starts, not baked in.
+build_image() {
+    local ctx rc=0
+    ctx=$(mktemp -d)
+    retry_build "docker build -t $IMAGE" build_inline "$IMAGE" "$(cat <<'DOCKERFILE'
+FROM debian:12
+ENV DEBIAN_FRONTEND=noninteractive
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    systemd iproute2 procps && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+CMD ["/lib/systemd/systemd"]
+DOCKERFILE
+    )" "$ctx" || rc=$?
+    rm -rf "$ctx"
+    return "$rc"
+}
+
+# Stage a tarball directory: the real install script, units and configuration,
+# with stub binaries and DNS helpers.
+stage_tarball() {
+    local f
+    STAGE=$(mktemp -d)
+    cp "$INSTALL_SH" "$STAGE/install.sh" || return 1
+    for f in fips.service fips-dns.service fips-gateway.service fips-firewall.service; do
+        cp "$REPO_ROOT/packaging/systemd/$f" "$STAGE/" || return 1
+    done
+    for f in fips.yaml hosts fips.nft; do
+        cp "$REPO_ROOT/packaging/common/$f" "$STAGE/" || return 1
+    done
+    printf '#!/bin/sh\nexec sleep infinity\n' > "$STAGE/fips"
+    printf '#!/bin/sh\nexec sleep infinity\n' > "$STAGE/fips-gateway"
+    printf '#!/bin/sh\nexit 0\n' > "$STAGE/fipsctl"
+    printf '#!/bin/sh\nexit 0\n' > "$STAGE/fips-dns-setup"
+    printf '#!/bin/sh\nexit 0\n' > "$STAGE/fips-dns-teardown"
+    chmod 0755 "$STAGE"/install.sh "$STAGE"/fips "$STAGE"/fips-gateway \
+        "$STAGE"/fipsctl "$STAGE"/fips-dns-setup "$STAGE"/fips-dns-teardown
+    return 0
+}
+
+# Start the scenario's systemd container. Not privileged: see
+# testing/lib/systemd-container.sh for the flags and why. The TUN device is
+# passed so the scenario can create the fips0 that fips-gateway.service's
+# ExecStartPre waits for.
+start_container() {
+    local name="$1"
+    cleanup_container "$name"
+    run_quiet "docker run $name" \
+        docker run -d --name "$name" \
+        --label com.corganlabs.fips-ci=1 \
+        "${SYSTEMD_CAPS[@]}" \
+        --cgroupns=host \
+        --device /dev/net/tun \
+        -v /sys/fs/cgroup:/sys/fs/cgroup:rw \
+        --tmpfs /run --tmpfs /run/lock \
+        "$IMAGE" || return
+    check_isolation "$name"
+    return
+}
+
+# Wait for systemd to finish booting. `degraded` counts as booted: units such
+# as systemd-modules-load cannot succeed in a container, and the system still
+# finished starting. The state string is tested directly rather than piped, so
+# pipefail cannot turn a degraded boot into a failure.
+wait_for_systemd() {
+    local name="$1" state
+    for _i in $(seq 1 "$BOOT_TIMEOUT"); do
+        state=$(docker exec "$name" systemctl is-system-running --wait 2>/dev/null || true)
+        case "$state" in
+            running | degraded) return 0 ;;
+        esac
+        sleep 1
+    done
+    echo "  ERROR: systemd did not reach running state in ${BOOT_TIMEOUT}s" >&2
+    return 1
+}
+
+# Run the staged install.sh in the container, under a bound.
+run_install() {
+    local name="$1" label="$2" out rc=0
+    out=$(timeout "$INSTALL_TIMEOUT" docker exec "$name" "$TARBALL_DIR/install.sh" 2>&1) || rc=$?
+    if [ "$rc" -ne 0 ]; then
+        fail "$label: install.sh exited $rc"
+        echo "$out" | sed 's/^/    /'
+        return 1
+    fi
+    pass "$label: install.sh exits 0"
+    return 0
+}
+
+# Print a unit's active state ("unknown" when the container cannot be asked).
+unit_state() {
+    local name="$1" unit="$2" state
+    state=$(docker exec "$name" systemctl is-active "$unit" 2>/dev/null)
+    echo "${state:-unknown}"
+    return 0
+}
+
+# Print a unit's InvocationID, which changes on every start.
+invocation_id() {
+    docker exec "$1" systemctl show -p InvocationID --value "$2" 2>/dev/null
+    return
+}
+
+# Record a PASS when a unit is in the wanted state, a FAIL otherwise.
+expect_state() {
+    local name="$1" label="$2" unit="$3" want="$4" got
+    got=$(unit_state "$name" "$unit")
+    if [ "$got" = "$want" ]; then
+        pass "$label: $unit is $got"
+    else
+        fail "$label: $unit is $got (want $want)"
+    fi
+    return 0
+}
+
+# Start a unit under a bound, recording a FAIL when it does not start.
+start_unit() {
+    local name="$1" unit="$2" out
+    if ! out=$(timeout "$UNIT_START_TIMEOUT" docker exec "$name" systemctl start "$unit" 2>&1); then
+        fail "setup: could not start $unit: ${out:-no output}"
+        return 1
+    fi
+    return 0
+}
+
+# Boot a fresh container, copy the tarball in and run the first install.
+# Returns 1, having recorded a FAIL, when any of that does not happen.
+prepare() {
+    local name="$1" label="$2"
+    start_container "$name" || { fail "$label: container did not start"; return 1; }
+    wait_for_systemd "$name" || {
+        fail "$label: systemd did not boot; the checks would read an unstarted system"
+        return 1
+    }
+    if ! docker cp "$STAGE/." "$name:$TARBALL_DIR" >/dev/null; then
+        fail "$label: could not copy the tarball into $name"
+        return 1
+    fi
+    if ! docker exec "$name" ip tuntap add fips0 mode tun >/dev/null 2>&1; then
+        fail "$label: could not create fips0 in $name"
+        return 1
+    fi
+    run_install "$name" "$label: first install" || return 1
+    return 0
+}
+
+scenario_all_active() {
+    local label="all-active" name before after
+    name=$(container_name "$label")
+    log "$label: fips, fips-dns and fips-gateway running before the upgrade"
+    if prepare "$name" "$label" \
+        && start_unit "$name" fips.service \
+        && start_unit "$name" fips-dns.service \
+        && start_unit "$name" fips-gateway.service; then
+        expect_state "$name" "$label: before" fips.service active
+        expect_state "$name" "$label: before" fips-dns.service active
+        expect_state "$name" "$label: before" fips-gateway.service active
+        before=$(invocation_id "$name" fips.service)
+        if run_install "$name" "$label: upgrade"; then
+            expect_state "$name" "$label: after" fips.service active
+            expect_state "$name" "$label: after" fips-dns.service active
+            expect_state "$name" "$label: after" fips-gateway.service active
+            after=$(invocation_id "$name" fips.service)
+            if [ -z "$before" ] || [ -z "$after" ]; then
+                fail "$label: could not read fips.service's InvocationID (before '${before}', after '${after}')"
+            elif [ "$before" != "$after" ]; then
+                pass "$label: the upgrade restarted fips.service"
+            else
+                fail "$label: fips.service kept InvocationID $before, so the upgrade did not stop and restart it"
+            fi
+        fi
+    fi
+    cleanup_container "$name"
+    return 0
+}
+
+scenario_fips_only() {
+    local label="fips-only" name
+    name=$(container_name "$label")
+    log "$label: only fips running before the upgrade"
+    if prepare "$name" "$label" && start_unit "$name" fips.service; then
+        expect_state "$name" "$label: before" fips.service active
+        expect_state "$name" "$label: before" fips-dns.service inactive
+        expect_state "$name" "$label: before" fips-gateway.service inactive
+        if run_install "$name" "$label: upgrade"; then
+            expect_state "$name" "$label: after" fips.service active
+            expect_state "$name" "$label: after" fips-dns.service inactive
+            expect_state "$name" "$label: after" fips-gateway.service inactive
+        fi
+    fi
+    cleanup_container "$name"
+    return 0
+}
+
+scenario_none_active() {
+    local label="none-active" name
+    name=$(container_name "$label")
+    log "$label: nothing running before the upgrade"
+    if prepare "$name" "$label"; then
+        expect_state "$name" "$label: before" fips.service inactive
+        if run_install "$name" "$label: upgrade"; then
+            expect_state "$name" "$label: after" fips.service inactive
+            expect_state "$name" "$label: after" fips-dns.service inactive
+            expect_state "$name" "$label: after" fips-gateway.service inactive
+        fi
+    fi
+    cleanup_container "$name"
+    return 0
+}
+
+ALL_SCENARIOS="all-active fips-only none-active"
+
+_args=()
+while [ $# -gt 0 ]; do
+    case "$1" in
+        --install-sh)
+            INSTALL_SH="${2:?--install-sh requires a path}"
+            shift 2
+            ;;
+        -h|--help)
+            echo "usage: test.sh [--install-sh PATH] [scenario ...]"
+            echo "scenarios: $ALL_SCENARIOS"
+            exit 0
+            ;;
+        -*)
+            echo "Unknown option: $1" >&2
+            exit 1
+            ;;
+        *)
+            _args+=("$1")
+            shift
+            ;;
+    esac
+done
+set -- ${_args[@]+"${_args[@]}"}
+
+if [ $# -eq 0 ]; then
+    scenarios="$ALL_SCENARIOS"
+else
+    scenarios="$*"
+fi
+
+for scenario in $scenarios; do
+    case " $ALL_SCENARIOS " in
+        *" $scenario "*) ;;
+        *)
+            echo "Unknown scenario: $scenario (available: $ALL_SCENARIOS)" >&2
+            exit 1
+            ;;
+    esac
+done
+
+if [ ! -f "$INSTALL_SH" ]; then
+    echo "install script not found: $INSTALL_SH" >&2
+    exit 1
+fi
+
+trap cleanup_all EXIT
+
+log "Building $IMAGE"
+if ! build_image; then
+    fail "image build failed"
+elif ! stage_tarball; then
+    fail "could not stage the tarball"
+else
+    for scenario in $scenarios; do
+        case "$scenario" in
+            all-active) scenario_all_active ;;
+            fips-only) scenario_fips_only ;;
+            none-active) scenario_none_active ;;
+        esac
+        echo
+    done
+fi
+
+echo "═══════════════════════════════════════"
+echo "Results: $PASS passed, $FAIL failed"
+echo "═══════════════════════════════════════"
+
+[ "$FAIL" -eq 0 ] && [ "$PASS" -gt 0 ]