diff --git a/src/lib/inference/vllm-serving-port.test.ts b/src/lib/inference/vllm-serving-port.test.ts
index bcb5d71463f..474cf867ad7 100644
--- a/src/lib/inference/vllm-serving-port.test.ts
+++ b/src/lib/inference/vllm-serving-port.test.ts
@@ -231,6 +231,7 @@ describe("managed vLLM serving-port guard (#8685)", () => {
     const profile = detectVllmProfile({ platform: "spark", type: "nvidia" })!;
     const managed = vllmContainerRow(profile.containerName);
     mockSuccessfulVllmInstall(mocks, profile.containerName, [() => managed, () => managed]);
+    mockDefaultDockerOwnership(managed);
     mocks.recoverHostLocalManagedVllmEndpoint.mockReturnValue({
       baseUrl: "http://127.0.0.1:8000",
       apiKey: "b".repeat(64),
@@ -286,6 +287,10 @@ describe("managed vLLM serving-port guard (#8685)", () => {
       () => vllmContainerRow(profile.containerName),
       () => vllmContainerRow(profile.containerName, { id: "c".repeat(64) }),
     ]);
+    mockDefaultDockerOwnership(
+      vllmContainerRow(profile.containerName),
+      vllmContainerRow(profile.containerName, { id: "c".repeat(64) }),
+    );
     mocks.recoverHostLocalManagedVllmEndpoint.mockReturnValue({
       baseUrl: "http://127.0.0.1:8000",
       apiKey: "b".repeat(64),
@@ -415,4 +420,149 @@ describe("managed vLLM serving-port guard (#8685)", () => {
     expect(mocks.dockerRunDetached).toHaveBeenCalled();
     expect(errSpy.mock.calls.flat().join("\n")).not.toContain("already in use");
   });
+  it("adopts its interrupted container on the default Docker context (#11426)", async () => {
+    process.env.DOCKER_HOST = "ssh://remote-builder.example.test";
+    process.env.DOCKER_CONTEXT = "remote-builder";
+    const profile = detectVllmProfile({ platform: "spark", type: "nvidia" })!;
+    const managed = vllmContainerRow(profile.containerName, { state: "running" });
+    // Three responses: the ownership check that classifies the port holder, then
+    // the ordinary replacement guard's own inspection.
+    mockSuccessfulVllmInstall(mocks, profile.containerName, [
+      () => managed,
+      () => managed,
+      () => managed,
+    ]);
+    mockDefaultDockerOwnership(managed);
+    // An interrupted install persists no runtime receipt, and a profile without
+    // managed bearer auth carries no auth label, so lifecycle recovery cannot
+    // admit the container this very install left behind.
+    mocks.recoverHostLocalManagedVllmEndpoint.mockReturnValue(null);
+    publishContainerBindings("0.0.0.0:8000", "[::]:8000");
+    const checkServingPort = vi.fn(async () => ({
+      ok: false,
+      reason: "port 8000 is held by docker-proxy (PID 4242)",
+    }));
+
+    const result = await installVllm(profile, {
+      hasImage: true,
+      nonInteractive: true,
+      promptFn: vi.fn<(q: string) => Promise<string>>(),
+      checkServingPort,
+    });
+
+    expect(result).toEqual({ ok: true });
+    expect(mocks.dockerForceRm).toHaveBeenCalledWith(
+      MANAGED_CONTAINER_ID,
+      expect.objectContaining({ ignoreError: true, suppressOutput: true }),
+    );
+    expect(mocks.dockerRunDetached).toHaveBeenCalled();
+    expect(errSpy.mock.calls.flat().join("\n")).not.toContain("another process");
+    const dockerOptions = [
+      ...mocks.dockerImageInspectFormat.mock.calls.map((call) => call[2]),
+      ...mocks.dockerPullWithProgressWatchdog.mock.calls.map((call) => call[1]),
+      ...mocks.dockerSpawn.mock.calls.map((call) => call[1]),
+      ...mocks.dockerForceRm.mock.calls.map((call) => call[1]),
+      ...mocks.dockerRunDetached.mock.calls.map((call) => call[1]),
+      ...mocks.dockerCapture.mock.calls.map((call) => call[1]),
+    ];
+    expect(dockerOptions.length).toBeGreaterThan(0);
+    expect(new Set(dockerOptions.map((options) => options?.env?.DOCKER_CONTEXT))).toEqual(
+      new Set(["default"]),
+    );
+    expect(new Set(dockerOptions.map((options) => options?.env?.DOCKER_HOST))).toEqual(
+      new Set([undefined]),
+    );
+  });
+
+  it("still refuses when an unlabeled container holds the serving port", async () => {
+    const profile = detectVllmProfile({ platform: "spark", type: "nvidia" })!;
+    const foreign = vllmContainerRow(profile.containerName, { label: "", state: "running" });
+    mockSuccessfulVllmInstall(mocks, profile.containerName, [() => foreign, () => foreign]);
+    mockDefaultDockerOwnership(foreign);
+    mocks.recoverHostLocalManagedVllmEndpoint.mockReturnValue(null);
+
+    const result = await installVllm(profile, {
+      hasImage: true,
+      nonInteractive: true,
+      promptFn: vi.fn<(q: string) => Promise<string>>(),
+      checkServingPort: async () => ({ ok: false, reason: "port 8000 is held" }),
+    });
+
+    expect(result).toEqual({ ok: false });
+    expect(mocks.dockerForceRm).not.toHaveBeenCalled();
+    expect(mocks.dockerRunDetached).not.toHaveBeenCalled();
+    expect(errSpy.mock.calls.flat().join("\n")).toContain("already in use");
+  });
+  /** Return each ownership row from the physical host's default Docker context. */
+  function mockDefaultDockerOwnership(...rows: string[]): void {
+    const base = mocks.dockerCapture.getMockImplementation();
+    let rowIndex = 0;
+    mocks.dockerCapture.mockImplementation(
+      (args: readonly string[], options?: { env?: NodeJS.ProcessEnv }) =>
+        args[0] === "container" && options?.env?.DOCKER_CONTEXT === "default"
+          ? (rows[Math.min(rowIndex++, rows.length - 1)] ?? "")
+          : (base?.(args, options) ?? ""),
+    );
+  }
+
+  /** Report the host bindings the managed container publishes for container 8000. */
+  function publishContainerBindings(...bindings: string[]): void {
+    const base = mocks.dockerCapture.getMockImplementation();
+    mocks.dockerCapture.mockImplementation(
+      (args: readonly string[], options?: { env?: NodeJS.ProcessEnv }) =>
+        args[0] === "port" ? `${bindings.join("\n")}\n` : (base?.(args, options) ?? ""),
+    );
+  }
+
+  it("refuses a managed container published on a different host port", async () => {
+    const profile = detectVllmProfile({ platform: "spark", type: "nvidia" })!;
+    const managed = vllmContainerRow(profile.containerName, { state: "running" });
+    mockSuccessfulVllmInstall(mocks, profile.containerName, [
+      () => managed,
+      () => managed,
+      () => managed,
+    ]);
+    mockDefaultDockerOwnership(managed);
+    mocks.recoverHostLocalManagedVllmEndpoint.mockReturnValue(null);
+    // Our managed container serves another port, so an unrelated process owns
+    // the port that failed. Removing this container would free nothing.
+    publishContainerBindings("0.0.0.0:19000");
+
+    const result = await installVllm(profile, {
+      hasImage: true,
+      nonInteractive: true,
+      promptFn: vi.fn<(q: string) => Promise<string>>(),
+      checkServingPort: async () => ({ ok: false, reason: "port 8000 is held" }),
+    });
+
+    expect(result).toEqual({ ok: false });
+    expect(mocks.dockerForceRm).not.toHaveBeenCalled();
+    expect(mocks.dockerRunDetached).not.toHaveBeenCalled();
+    expect(errSpy.mock.calls.flat().join("\n")).toContain("already in use");
+  });
+
+  it("refuses a managed container bound away from the probed loopback address", async () => {
+    const profile = detectVllmProfile({ platform: "spark", type: "nvidia" })!;
+    const managed = vllmContainerRow(profile.containerName, { state: "running" });
+    mockSuccessfulVllmInstall(mocks, profile.containerName, [
+      () => managed,
+      () => managed,
+      () => managed,
+    ]);
+    mockDefaultDockerOwnership(managed);
+    mocks.recoverHostLocalManagedVllmEndpoint.mockReturnValue(null);
+    publishContainerBindings("127.0.0.2:8000", "192.168.1.10:8000");
+
+    const result = await installVllm(profile, {
+      hasImage: true,
+      nonInteractive: true,
+      promptFn: vi.fn<(q: string) => Promise<string>>(),
+      checkServingPort: async () => ({ ok: false, reason: "port 8000 is held" }),
+    });
+
+    expect(result).toEqual({ ok: false });
+    expect(mocks.dockerForceRm).not.toHaveBeenCalled();
+    expect(mocks.dockerRunDetached).not.toHaveBeenCalled();
+    expect(errSpy.mock.calls.flat().join("\n")).toContain("already in use");
+  });
 });
diff --git a/src/lib/inference/vllm.ts b/src/lib/inference/vllm.ts
index a455c88fc93..f92c101e60d 100644
--- a/src/lib/inference/vllm.ts
+++ b/src/lib/inference/vllm.ts
@@ -1242,7 +1242,7 @@ function startContainer(
   // avoids deleting an unrelated same-name container if the name changes hands.
   const replacement = vllmContainerReplacementTarget(
     profile.containerName,
-    model.managedBearerAuth ? dockerEnv : undefined,
+    model.managedBearerAuth || expectedReplacementContainerId !== undefined ? dockerEnv : undefined,
     expectedReplacementContainerId,
   );
   if (!replacement.ok) return replacement;
@@ -1949,6 +1949,46 @@ function applyRequestedVllmGpuDevice(
   return requestedGpuDevice ? selectVllmGpuDevice(profile, requestedGpuDevice) : profile;
 }
 
+/**
+ * Container id of this install's own managed container when that container is
+ * what holds the serving port.
+ *
+ * Lifecycle recovery admits only a completed authenticated install: it needs
+ * the runtime receipt written after startup and the auth label that a profile
+ * adds only for managed bearer auth. An install interrupted before either
+ * exists leaves a running managed container that recovery can never claim, so
+ * ownership labels decide instead.
+ *
+ * Ownership alone is not enough. A managed container published on a different
+ * host port is not the process holding this port, and removing it would free
+ * nothing while destroying an unrelated runtime, so one published binding must
+ * cover the probed loopback address and match the port that failed. Every other
+ * state — a foreign or unlabeled holder, an ambiguous inspection, a distributed
+ * head or worker, a container that is not running, and an unreadable or
+ * address-mismatched binding — remains a conflict.
+ */
+function adoptableServingPortHolder(
+  containerName: string,
+  servingPort: number,
+  dockerEnv: Record<string, string>,
+): string | undefined {
+  const ownership = inspectVllmContainerOwnershipInDockerEnv(containerName, dockerEnv);
+  if (ownership.kind !== "managed" || !ownership.running) return undefined;
+  // The managed container always publishes the fixed container port 8000.
+  const published = dockerCapture(["port", containerName, "8000"], {
+    env: dockerEnv,
+    ignoreError: true,
+    timeout: 10_000,
+  })
+    ?.split(/\r?\n/u)
+    .some((binding) => {
+      const endpoint = binding.trim().match(/^(127[.]0[.]0[.]1|0[.]0[.]0[.]0):(\d+)$/u);
+      return endpoint !== null && Number(endpoint[2]) === servingPort;
+    });
+  if (!published) return undefined;
+  return ownership.containerId;
+}
+
 /**
  * Name the process holding the serving port so the operator can act, matching
  * how the Ollama auth proxy reports its own port conflict.
@@ -2315,6 +2355,7 @@ async function runVllmInstall(
   // Port 25000 is not checked here: it belongs to the managed-cluster
   // rendezvous contract and this single-node path never binds it.
   let recoveredHostLocalContainerId: string | undefined;
+  let recoveredHostLocalDockerEnv: Record<string, string> | undefined;
   const servingPort = await opts.checkServingPort?.(VLLM_PORT);
   if (servingPort && !servingPort.ok) {
     // An interrupted host-local install can leave its authenticated managed
@@ -2323,14 +2364,24 @@ async function runVllmInstall(
     // credential fingerprint. The replacement guard below then removes the
     // inspected container ID immediately before the new launch.
     try {
+      const hostLocalDockerEnv = buildLocalManagedVllmDockerEnv();
       const recovered = recoverHostLocalManagedVllmEndpoint();
       if (recovered?.baseUrl === `http://127.0.0.1:${String(VLLM_PORT)}`) {
         recoveredHostLocalContainerId = recovered.containerId;
         // Continue through the ordinary managed-container replacement path.
       } else {
-        printServingPortConflict(servingPort);
-        return { ok: false };
+        const adopted = adoptableServingPortHolder(
+          runtimeProfile.containerName,
+          VLLM_PORT,
+          hostLocalDockerEnv,
+        );
+        if (adopted === undefined) {
+          printServingPortConflict(servingPort);
+          return { ok: false };
+        }
+        recoveredHostLocalContainerId = adopted;
       }
+      recoveredHostLocalDockerEnv = hostLocalDockerEnv;
     } catch (error) {
       console.error(
         `  vLLM install failed: managed host-local vLLM recovery could not verify the container: ${(error as Error).message}`,
@@ -2373,9 +2424,8 @@ async function runVllmInstall(
   let hostLocalApiKey: string | null = null;
   let localDockerEnv = dualStationPlan
     ? buildLocalDualStationDockerEnv()
-    : model.managedBearerAuth
-      ? buildLocalManagedVllmDockerEnv()
-      : buildVllmDockerEnv();
+    : (recoveredHostLocalDockerEnv ??
+      (model.managedBearerAuth ? buildLocalManagedVllmDockerEnv() : buildVllmDockerEnv()));
   let gpuMemoryWarningShown = false;
   const reportGpuMemoryWarning = (result: GpuMemoryPreflightResult): void => {
     if (!result.ok || !result.warning || gpuMemoryWarningShown) return;
@@ -2426,7 +2476,9 @@ async function runVllmInstall(
   } else {
     const replacement = vllmContainerReplacementTarget(
       runtimeProfile.containerName,
-      model.managedBearerAuth ? localDockerEnv : undefined,
+      model.managedBearerAuth || recoveredHostLocalContainerId !== undefined
+        ? localDockerEnv
+        : undefined,
       recoveredHostLocalContainerId,
     );
     if (!replacement.ok) {