diff --git a/crates/tracedecay/tests/common/mod.rs b/crates/tracedecay/tests/common/mod.rs
index 28bbc596c4..2f0774b287 100644
--- a/crates/tracedecay/tests/common/mod.rs
+++ b/crates/tracedecay/tests/common/mod.rs
@@ -657,6 +657,9 @@ pub fn http_agent_with_timeout(timeout: Duration) -> ureq::Agent {
 /// panic while the child is still running, `Drop` force-stops and reaps it.
 pub struct TestChildProcess {
     child: Child,
+    /// Unix socket this child bound. Retired once the child is reaped so a
+    /// descriptor a subprocess inherited cannot keep the path connectable.
+    owned_unix_endpoint: Option<PathBuf>,
 }
 
 /// Daemon-specific name retained for test fixtures that keep a daemon alive.
@@ -664,7 +667,39 @@ pub type DaemonProcess = TestChildProcess;
 
 impl TestChildProcess {
     pub fn new(child: Child) -> Self {
-        Self { child }
+        Self {
+            child,
+            owned_unix_endpoint: None,
+        }
+    }
+
+    /// Record the Unix socket this child bound.
+    ///
+    /// `SIGKILL` skips the daemon's own endpoint cleanup (`drop(listener)` then
+    /// unlink). A subprocess that inherited the listen descriptor across `fork`
+    /// still accepts on that path after the owner is reaped, which is a leaked
+    /// descriptor, not a live daemon. Retiring the path with the owner is what
+    /// the graceful shutdown already does, and what a forced stop must do too.
+    #[cfg(unix)]
+    pub fn own_unix_endpoint(&mut self, path: PathBuf) {
+        self.owned_unix_endpoint = Some(path);
+    }
+
+    /// Unlink the recorded endpoint, but only after a stop this call forced.
+    ///
+    /// A child that exited on its own already ran its own endpoint cleanup, so
+    /// the path is either gone or has since been rebound by a replacement. The
+    /// restart journeys reassign the guard (`daemon = spawn_...`), which drops
+    /// the predecessor *after* the successor has bound the same path; unlinking
+    /// there would retire a live daemon's socket.
+    fn retire_owned_unix_endpoint(&mut self, forced: bool) {
+        let path = self.owned_unix_endpoint.take();
+        if !forced {
+            return;
+        }
+        if let Some(path) = path {
+            let _ = std::fs::remove_file(path);
+        }
     }
 
     pub fn id(&self) -> u32 {
@@ -752,9 +787,13 @@ impl TestChildProcess {
     /// Force-stops the daemon and reaps its process before returning.
     ///
     /// `Child::kill` maps to `SIGKILL` on Unix and the platform termination
-    /// primitive elsewhere, keeping fault-injection tests portable.
+    /// primitive elsewhere, keeping fault-injection tests portable. The owned
+    /// Unix endpoint, when recorded, is unlinked only after that reap.
     pub fn kill_and_wait(&mut self) -> std::io::Result<ExitStatus> {
-        terminate_and_reap(&mut self.child)
+        let forced = self.is_running();
+        let status = terminate_and_reap(&mut self.child)?;
+        self.retire_owned_unix_endpoint(forced);
+        Ok(status)
     }
 
     fn drain_stderr(&mut self) {
@@ -778,15 +817,30 @@ impl TestChildProcess {
 
 impl Drop for TestChildProcess {
     fn drop(&mut self) {
-        let _ = terminate_and_reap(&mut self.child);
+        let forced = self.is_running();
+        if terminate_and_reap(&mut self.child).is_ok() {
+            self.retire_owned_unix_endpoint(forced);
+        }
     }
 }
 
-/// PID-directed stop: survives `process_group(0)` / `setsid` detachment.
+/// Stop the child and the process group it leads.
+///
+/// Test daemons are started with `process_group(0)`, so the spawned pid is the
+/// group leader and is outside the test's group. `Child::kill` signals only
+/// that pid. A child still between `fork` and `exec` inherits the listen
+/// descriptor and stays in the leader's group, so the path keeps accepting
+/// after the leader is reaped. Signal that group first, while the pid is still
+/// the live leader, then reap the leader. Do not signal after `try_wait` has
+/// reaped: the pid can be recycled. A child that left the group (its own
+/// `process_group(0)`) is not this signal's target; unlinking the owned socket
+/// retires the path those descriptors were bound to.
 fn terminate_and_reap(child: &mut Child) -> std::io::Result<ExitStatus> {
+    let pid = child.id();
     if let Ok(Some(status)) = child.try_wait() {
         return Ok(status);
     }
+    signal_owned_process_group(pid);
 
     if let Err(kill_err) = child.kill() {
         if let Some(status) = child.try_wait()? {
@@ -798,6 +852,23 @@ fn terminate_and_reap(child: &mut Child) -> std::io::Result<ExitStatus> {
     child.wait()
 }
 
+#[cfg(unix)]
+fn signal_owned_process_group(pid: u32) {
+    let Ok(pid) = i32::try_from(pid) else {
+        return;
+    };
+    if pid <= 0 {
+        return;
+    }
+    // Safety: `pid` is the still-live child returned by `Child::id` before
+    // `try_wait` reaps it, and `process_group(0)` made that pid the group id.
+    // A non-positive argument would signal this process's own group.
+    let _ = unsafe { libc::kill(-pid, libc::SIGKILL) };
+}
+
+#[cfg(not(unix))]
+fn signal_owned_process_group(_pid: u32) {}
+
 /// Detach a test child from the test process group.
 ///
 /// Nextest (and other harness timeouts) signal the test's process group.
@@ -1089,13 +1160,6 @@ pub fn spawn_tracedecay_daemon_with(
     spawn_tracedecay_daemon_process(&home, &binary, configure)
 }
 
-/// How long a replacement daemon waits for a stopped predecessor's endpoint to
-/// stop accepting before reporting it as still live.
-///
-/// Generous on purpose: the wait only costs time when a predecessor is
-/// genuinely still reachable, and a real leak still fails rather than hangs.
-const PREDECESSOR_DAEMON_VACATE_TIMEOUT: Duration = Duration::from_secs(10);
-
 fn spawn_tracedecay_daemon_process(
     home: &Path,
     binary: &Path,
@@ -1118,42 +1182,20 @@ fn spawn_tracedecay_daemon_process(
             })
             .is_some_and(|address| TcpStream::connect(address).is_ok())
     };
-    // Stopping a predecessor daemon is asynchronous with respect to its
-    // endpoint: `kill` plus `wait` reaps the PID the harness spawned, but the
-    // kernel keeps the listening socket alive while *any* duplicate of that
-    // descriptor survives, including one a subprocess inherited across `fork`
-    // and still holds because it has not reached its own `exec` yet. Asserting
-    // instantaneously therefore reports an ordinary teardown tail as a live
-    // daemon, which is what `init_project_fixture` journeys (spawn, init, drop,
-    // spawn again) hit on a loaded runner. Wait a bounded time for the endpoint
-    // to stop accepting; a daemon that keeps accepting still fails with the
-    // same refusal.
-    poll_until(
-        Instant::now() + PREDECESSOR_DAEMON_VACATE_TIMEOUT,
-        Duration::from_millis(25),
-        || {
-            #[cfg(unix)]
-            let live = std::os::unix::net::UnixStream::connect(&socket_path).is_ok();
-            #[cfg(not(unix))]
-            let live = portable_daemon_connectable();
-            (!live).then_some(())
-        },
-        || {
-            #[cfg(unix)]
-            {
-                format!(
-                    "refusing to replace a live test daemon at {}",
-                    socket_path.display()
-                )
-            }
-            #[cfg(not(unix))]
-            {
-                format!(
-                    "refusing to replace a live test daemon recorded at {}",
-                    authority_path.display()
-                )
-            }
-        },
+    // A reaped predecessor has already unlinked its socket. A path that still
+    // accepts is a daemon this call did not stop, and replacing it would bind
+    // over a live owner.
+    #[cfg(unix)]
+    assert!(
+        std::os::unix::net::UnixStream::connect(&socket_path).is_err(),
+        "refusing to replace a live test daemon at {}",
+        socket_path.display()
+    );
+    #[cfg(not(unix))]
+    assert!(
+        !portable_daemon_connectable(),
+        "refusing to replace a live test daemon recorded at {}",
+        authority_path.display()
     );
 
     let mut command = Command::new(binary);
@@ -1169,6 +1211,8 @@ fn spawn_tracedecay_daemon_process(
     detach_from_test_process_group(&mut command);
     let child = command.spawn().expect("tracedecay daemon should start");
     let mut daemon = DaemonProcess::new(child);
+    #[cfg(unix)]
+    daemon.own_unix_endpoint(socket_path.clone());
 
     let deadline = Instant::now() + Duration::from_secs(10);
     poll_until(
diff --git a/crates/tracedecay/tests/transport_acceptance_suite/daemon_fault_harness_test.rs b/crates/tracedecay/tests/transport_acceptance_suite/daemon_fault_harness_test.rs
index dcf97a9b1d..447b47d06f 100644
--- a/crates/tracedecay/tests/transport_acceptance_suite/daemon_fault_harness_test.rs
+++ b/crates/tracedecay/tests/transport_acceptance_suite/daemon_fault_harness_test.rs
@@ -321,6 +321,75 @@ fn configured_daemon_can_be_killed_and_reaped() {
     }
 }
 
+/// A listen descriptor inherited across `fork` must not outlive its owner.
+///
+/// `branch_search_serves_a_committed_generation_behind_dirty_worktree_state`
+/// drops the init daemon and immediately spawns the next one. Killing only the
+/// leader leaves the inherited listener accepting on the same path, so the
+/// next spawn reports a live daemon. The owner is the group leader; retiring
+/// its socket with it makes that path refuse the moment the owner is reaped.
+#[cfg(unix)]
+#[test]
+fn reaped_owner_releases_an_inherited_listen_socket() {
+    use std::os::unix::process::CommandExt;
+    use std::process::Stdio;
+    use std::time::{Duration, Instant};
+
+    let home = tempdir_or_panic();
+    let socket_path = home.path().join("inherited-listen.sock");
+    let script = r#"
+import os, socket, sys, time
+sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+sock.bind(sys.argv[1])
+sock.listen(1)
+if os.fork() == 0:
+    time.sleep(60)
+else:
+    time.sleep(60)
+"#;
+    let mut command = std::process::Command::new("python3");
+    command
+        .arg("-c")
+        .arg(script)
+        .arg(&socket_path)
+        .stdin(Stdio::null())
+        .stdout(Stdio::null())
+        .stderr(Stdio::null())
+        .process_group(0);
+    let child = command.spawn().expect("python listener should start");
+    let mut owner = common::TestChildProcess::new(child);
+    owner.own_unix_endpoint(socket_path.clone());
+
+    let deadline = Instant::now() + Duration::from_secs(5);
+    loop {
+        if std::os::unix::net::UnixStream::connect(&socket_path).is_ok() {
+            break;
+        }
+        if let Some(status) = owner
+            .try_wait()
+            .expect("listener status should be readable")
+        {
+            panic!("listener exited before accepting: {status}");
+        }
+        assert!(
+            Instant::now() < deadline,
+            "inherited listener never accepted on {}",
+            socket_path.display()
+        );
+        std::thread::sleep(Duration::from_millis(20));
+    }
+
+    drop(owner);
+    assert!(
+        std::os::unix::net::UnixStream::connect(&socket_path).is_err(),
+        "an inherited listen descriptor must not keep the owner's path accepting"
+    );
+    assert!(
+        !socket_path.exists(),
+        "reaping the owner must unlink the socket it bound"
+    );
+}
+
 #[cfg(all(unix, tracedecay_observation_fault_harness, feature = "test-transport"))]
 async fn assert_daemon_crash_stage(
     barrier_stage: &str,