diff --git a/scripts/run-tests.mjs b/scripts/run-tests.mjs
index bbd100b13..2b437111c 100644
--- a/scripts/run-tests.mjs
+++ b/scripts/run-tests.mjs
@@ -1281,6 +1281,7 @@ export const SUITES = {
     "src/lib/daemon-start.test.ts",
     "src/lib/daemon-socket-occupancy.test.ts",
     "src/lib/daemon-endpoint-faults.test.ts",
+    "src/lib/daemon-endpoint-churn.test.ts",
     "src/lib/daemon-startup-contract.test.ts",
     "src/lib/runtime-startup-throttle.test.ts",
     "src/lib/daemon-update-lifecycle.test.ts",
diff --git a/src/lib/daemon-endpoint-churn.test.ts b/src/lib/daemon-endpoint-churn.test.ts
new file mode 100644
index 000000000..330154c1e
--- /dev/null
+++ b/src/lib/daemon-endpoint-churn.test.ts
@@ -0,0 +1,188 @@
+import assert from "node:assert/strict";
+import { mkdtempSync, rmSync } from "node:fs";
+import { createServer, type Server } from "node:net";
+import { tmpdir } from "node:os";
+import path from "node:path";
+import test from "node:test";
+
+import { inspectDaemonAddress } from "./daemon-socket-occupancy.ts";
+
+/**
+ * Repeated startup/shutdown churn against a real endpoint (cave-58eoq.4).
+ *
+ * The bead asks for repeated lifecycle stress that "detects orphaned processes,
+ * stale endpoint state, unsafe retries, and false healthy/fixed diagnostics".
+ * Two of those are already covered elsewhere and deliberately not repeated here:
+ * the supervisor's sleep/wake, backoff ladder and abort semantics are pinned by
+ * daemon-connection-supervisor.test.ts, and single-transition endpoint faults by
+ * daemon-endpoint-faults.test.ts.
+ *
+ * What neither covers is the *repeated* case. A single bind/close cycle can pass
+ * while the tenth fails, and those are the interesting failures: an address that
+ * reports occupied after its owner left, a probe that caches its first verdict,
+ * a listener that cannot rebind because something held the path. Each of those
+ * shows up as a daemon that "won't start" or a duplicate daemon, and neither
+ * reproduces on the first try.
+ */
+
+const BOUND_MS = 1_000;
+const CYCLES = 12;
+
+const cleanups: Array<() => void> = [];
+test.after(() => {
+  for (const cleanup of cleanups.reverse()) {
+    try {
+      cleanup();
+    } catch {
+      // Teardown must never convert a passing assertion into a failing suite.
+    }
+  }
+});
+
+function tempDir(): string {
+  const dir = mkdtempSync(path.join(tmpdir(), "cave-endpoint-churn-"));
+  cleanups.push(() => rmSync(dir, { recursive: true, force: true }));
+  return dir;
+}
+
+/** Windows has no UNIX domain sockets; the daemon uses a named pipe there. */
+function addressFor(dir: string, name: string): string {
+  if (process.platform === "win32") {
+    return path.join("\\\\.\\pipe", `cave-endpoint-churn-${name}-${process.pid}`);
+  }
+  return path.join(dir, `${name}.sock`);
+}
+
+async function bind(socketPath: string): Promise<Server> {
+  const server = createServer(() => {});
+  // Registered before listen resolves, not after the caller unbinds it. These
+  // tests assert mid-cycle, so an assertion that throws would otherwise skip
+  // its unbind() and leave a listener holding the address — the suite then
+  // hangs on the failure path, which is the exact behaviour this whole fault
+  // harness exists to prevent. Closing twice is harmless; the cleanup runner
+  // swallows it.
+  cleanups.push(() => server.close());
+  await new Promise<void>((resolve, reject) => {
+    server.once("error", reject);
+    server.listen(socketPath, () => resolve());
+  });
+  return server;
+}
+
+async function unbind(server: Server): Promise<void> {
+  await new Promise<void>((resolve) => server.close(() => resolve()));
+}
+
+test("an endpoint survives repeated bind/release cycles without drifting", async () => {
+  const dir = tempDir();
+  const address = addressFor(dir, "churn");
+
+  for (let cycle = 1; cycle <= CYCLES; cycle += 1) {
+    const server = await bind(address);
+    assert.equal(
+      await inspectDaemonAddress({ socketPath: address, timeoutMs: BOUND_MS }),
+      "occupied",
+      `cycle ${cycle}: a live listener must read occupied`,
+    );
+
+    await unbind(server);
+    assert.equal(
+      await inspectDaemonAddress({ socketPath: address, timeoutMs: BOUND_MS }),
+      "free",
+      `cycle ${cycle}: a released address must read free, not stale-occupied`,
+    );
+  }
+});
+
+test("a released address never blocks the next daemon from binding it", async () => {
+  // The stale-endpoint failure a user actually hits: the daemon exits, and the
+  // next launch cannot bind because the path was left behind. If any cycle here
+  // throws EADDRINUSE, that is the bug, and it is invisible in a single cycle.
+  const dir = tempDir();
+  const address = addressFor(dir, "rebind");
+
+  for (let cycle = 1; cycle <= CYCLES; cycle += 1) {
+    const server = await bind(address);
+    await unbind(server);
+  }
+
+  const finalServer = await bind(address);
+  assert.equal(
+    await inspectDaemonAddress({ socketPath: address, timeoutMs: BOUND_MS }),
+    "occupied",
+    "the address is still usable after repeated churn",
+  );
+  await unbind(finalServer);
+});
+
+test("concurrent probes during churn never disagree with the endpoint's real state", async () => {
+  // Several surfaces can probe at once — a launcher, a status poll, a recovery
+  // banner. Under churn they must all see the same truth, because one probe
+  // reporting occupied after shutdown is how a launch gets wrongly refused.
+  const dir = tempDir();
+  const address = addressFor(dir, "concurrent");
+
+  for (let cycle = 1; cycle <= 6; cycle += 1) {
+    const server = await bind(address);
+    const whileUp = await Promise.all(
+      Array.from({ length: 4 }, () =>
+        inspectDaemonAddress({ socketPath: address, timeoutMs: BOUND_MS })),
+    );
+    assert.deepEqual(
+      new Set(whileUp),
+      new Set(["occupied"]),
+      `cycle ${cycle}: concurrent probes disagreed while the listener was up`,
+    );
+
+    await unbind(server);
+    const whileDown = await Promise.all(
+      Array.from({ length: 4 }, () =>
+        inspectDaemonAddress({ socketPath: address, timeoutMs: BOUND_MS })),
+    );
+    assert.ok(
+      !whileDown.includes("occupied"),
+      `cycle ${cycle}: a probe claimed occupied after the listener was gone`,
+    );
+  }
+});
+
+test("churn leaves no connection behind on any listener it probed", async () => {
+  // The orphan check. A probe that holds its connection open would accumulate
+  // one per cycle, and the daemon would slowly run out of accept slots — a
+  // failure that only appears after long uptime, never in a single-shot test.
+  const dir = tempDir();
+  const address = addressFor(dir, "orphans");
+  let everOpened = 0;
+  let stillOpen = 0;
+
+  for (let cycle = 1; cycle <= CYCLES; cycle += 1) {
+    const server = createServer(() => {});
+    // Same registration discipline as bind() above: this test counts
+    // connections, so it cannot use the helper, but an assertion throwing
+    // mid-cycle must still not strand a listener.
+    cleanups.push(() => server.close());
+    server.on("connection", (socket) => {
+      everOpened += 1;
+      stillOpen += 1;
+      socket.on("close", () => {
+        stillOpen -= 1;
+      });
+    });
+    await new Promise<void>((resolve, reject) => {
+      server.once("error", reject);
+      server.listen(address, () => resolve());
+    });
+
+    await inspectDaemonAddress({ socketPath: address, timeoutMs: BOUND_MS });
+    await unbind(server);
+  }
+
+  // Bounded drain: close events are asynchronous, but they are not slow.
+  const deadline = Date.now() + 500;
+  while (stillOpen > 0 && Date.now() < deadline) {
+    await new Promise((resolve) => setTimeout(resolve, 5));
+  }
+
+  assert.equal(everOpened, CYCLES, "every cycle must have been probed for real");
+  assert.equal(stillOpen, 0, "a probe left a connection open on a departed daemon");
+});