diff --git a/apps/web/design.md b/apps/web/design.md
index 1fb6db3..dff5998 100644
--- a/apps/web/design.md
+++ b/apps/web/design.md
@@ -12,3 +12,5 @@
 - 320px과 모바일 프로젝트에서도 문서 가로 넘침이 없어야 하며, 오류·재시도 상태는 기존 관리자/리뷰 알림 패턴을 따른다.
 
 국소 구현 소유자는 `src/pages/Admin.tsx`, `src/pages/admin/admin-console.css`, `src/pages/SessionReview.tsx`, `src/pages/session-review/`다. 공통 토큰과 앱 셸은 각각 `src/styles/tokens.css`, `src/components/shell/`의 소유권을 유지한다.
+
+- 2026-08-28 · `/learn` 성장 지표의 라포 헤더에서 중첩 inset surface를 제거해 같은 패널의 일반 헤더 위계와 정렬했다. 토큰 변경 없음.
diff --git a/apps/web/e2e/uc-learner-home-dashboard.spec.ts b/apps/web/e2e/uc-learner-home-dashboard.spec.ts
index b5e5599..4776599 100644
--- a/apps/web/e2e/uc-learner-home-dashboard.spec.ts
+++ b/apps/web/e2e/uc-learner-home-dashboard.spec.ts
@@ -700,6 +700,39 @@ test.describe("uc learner home dashboard", () => {
     await expect(page).toHaveURL(/\/learn\/practice$/);
   });
 
+  test("성장 지표 라포 헤더는 중첩 inset 표면을 사용하지 않는다", async ({
+    page,
+  }) => {
+    await installFixtures(page);
+    await page.goto("/learn");
+
+    await page.getByRole("tab", { name: "성장 지표" }).click();
+    const rapportHeading = page
+      .locator("#lh-panel-growth")
+      .getByRole("heading", { name: "회기를 건너 누적됩니다." });
+    const headingContent = rapportHeading.locator("..");
+    await expect(rapportHeading).toBeVisible();
+    await expect(headingContent).not.toHaveClass(/vg-surface/);
+    await expect
+      .poll(() =>
+        headingContent.evaluate((element) => {
+          const style = getComputedStyle(element);
+          return {
+            backgroundColor: style.backgroundColor,
+            borderTopWidth: style.borderTopWidth,
+            borderRadius: style.borderRadius,
+            padding: style.padding,
+          };
+        }),
+      )
+      .toEqual({
+        backgroundColor: "rgba(0, 0, 0, 0)",
+        borderTopWidth: "0px",
+        borderRadius: "0px",
+        padding: "0px",
+      });
+  });
+
   // usecase: 첫 방문 학습자가 기록·성장 탭에서 빈 상태 안내 문구를 확인한다
   test("신규 학습자의 기록·성장 탭이 빈 상태 안내 문구를 보여준다", async ({
     page,
diff --git a/apps/web/src/pages/LearnerHome.tsx b/apps/web/src/pages/LearnerHome.tsx
index f1d1785..40512d7 100644
--- a/apps/web/src/pages/LearnerHome.tsx
+++ b/apps/web/src/pages/LearnerHome.tsx
@@ -1554,12 +1554,7 @@ export default function LearnerHome({ view = "dashboard" }: LearnerHomeProps) {
                   className={surfaceClassName("lh-panel lh-panel--growth")}
                 >
                   <div className="lh-panel__head">
-                    <div
-                      className={surfaceClassName(undefined, {
-                        variant: "inset",
-                        flat: true,
-                      })}
-                    >
+                    <div>
                       {/* D1: 섹션 라벨 — dot 제거 */}
                       <Kicker dot={false}>내담자와 쌓은 라포</Kicker>
                       <h2>회기를 건너 누적됩니다.</h2>
diff --git a/scripts/boot-public-runtime.ps1 b/scripts/boot-public-runtime.ps1
index 44c3ed0..c5776d2 100644
--- a/scripts/boot-public-runtime.ps1
+++ b/scripts/boot-public-runtime.ps1
@@ -327,10 +327,15 @@ if ((Test-ApiControlPlaneHealthy) -and (Test-EngineHealthy) -and (Test-VoiceApiH
     Write-BootLog "web preview down on 127.0.0.1:$WebPort; including web in restart"
   }
   Write-BootLog ("running start-public-runtime.ps1 " + ($startArgs -join " "))
-  $out = & powershell.exe -NoProfile -ExecutionPolicy Bypass -File $startScript @startArgs 2>&1
-  $out | ForEach-Object { Write-BootLog ("  pub> " + $_) }
-  if ($LASTEXITCODE -ne 0) {
-    Write-BootLog "ERROR: start-public-runtime.ps1 exit $LASTEXITCODE"
+  $startFailed = $false
+  try {
+    & $startScript @startArgs |
+      ForEach-Object { Write-BootLog ("  pub> " + $_) }
+  } catch {
+    $startFailed = $true
+    Write-BootLog "ERROR: start-public-runtime.ps1 failed: $($_.Exception.Message)"
+  }
+  if ($startFailed) {
     exit 1
   }
 }
diff --git a/scripts/install-public-runtime-task.ps1 b/scripts/install-public-runtime-task.ps1
index f3a3082..5dea5a6 100644
--- a/scripts/install-public-runtime-task.ps1
+++ b/scripts/install-public-runtime-task.ps1
@@ -127,10 +127,8 @@ $repeatTrigger = New-ScheduledTaskTrigger `
 $settings = New-ScheduledTaskSettingsSet `
   -AllowStartIfOnBatteries `
   -DontStopIfGoingOnBatteries `
-  -ExecutionTimeLimit (New-TimeSpan -Minutes 10) `
+  -ExecutionTimeLimit (New-TimeSpan -Minutes 60) `
   -MultipleInstances IgnoreNew `
-  -RestartCount 3 `
-  -RestartInterval (New-TimeSpan -Minutes 1) `
   -StartWhenAvailable `
   -WakeToRun
 
diff --git a/scripts/register-boot-task.ps1 b/scripts/register-boot-task.ps1
index c13e584..ed9168f 100644
--- a/scripts/register-boot-task.ps1
+++ b/scripts/register-boot-task.ps1
@@ -113,8 +113,9 @@ $principal = New-ScheduledTaskPrincipal `
 $settings = New-ScheduledTaskSettingsSet `
   -AllowStartIfOnBatteries `
   -DontStopIfGoingOnBatteries `
+  -MultipleInstances IgnoreNew `
   -StartWhenAvailable `
-  -ExecutionTimeLimit (New-TimeSpan -Minutes 15)
+  -ExecutionTimeLimit (New-TimeSpan -Minutes 60)
 
 Register-ScheduledTask `
   -TaskName $TaskName `
diff --git a/scripts/start-public-runtime.ps1 b/scripts/start-public-runtime.ps1
index c3ec44f..eb029af 100644
--- a/scripts/start-public-runtime.ps1
+++ b/scripts/start-public-runtime.ps1
@@ -6,6 +6,15 @@
   [int]$WhisperPort = 9882,
   [int]$MeloTtsPort = 9883,
   [int]$VoiceSidecarReadySeconds = 300,
+  [ValidateRange(30, 600)]
+  [int]$ApiReadySeconds = 180,
+  [ValidateRange(30, 300)]
+  [int]$VoiceApiReadySeconds = 90,
+  [ValidateRange(60, 1800)]
+  [int]$WebBuildTimeoutSeconds = 600,
+  [string]$RecoveryLockPath = "$env:LOCALAPPDATA\Vignette\public-runtime-start.lock",
+  [ValidateRange(0, 300)]
+  [int]$RecoveryLockWaitSeconds = 0,
   [string]$Python = "$env:LOCALAPPDATA\Programs\Python\Python311\python.exe",
   [string]$Cloudflared = "$env:LOCALAPPDATA\Microsoft\WinGet\Links\cloudflared.exe",
   [string]$CloudflaredConfig = "$env:USERPROFILE\.cloudflared\vignette-config.yml",
@@ -29,6 +38,46 @@
 
 $ErrorActionPreference = "Stop"
 
+function Enter-RecoveryLock {
+  param(
+    [string]$LockPath,
+    [int]$WaitSeconds
+  )
+
+  if ([string]::IsNullOrWhiteSpace($LockPath)) {
+    throw "Public runtime recovery lock path is empty"
+  }
+  $resolvedLockPath = [System.IO.Path]::GetFullPath($LockPath)
+  $lockDirectory = [System.IO.Path]::GetDirectoryName($resolvedLockPath)
+  if ([string]::IsNullOrWhiteSpace($lockDirectory)) {
+    throw "Public runtime recovery lock path has no parent directory"
+  }
+  [System.IO.Directory]::CreateDirectory($lockDirectory) | Out-Null
+  $deadline = (Get-Date).AddSeconds($WaitSeconds)
+  do {
+    try {
+      return [System.IO.File]::Open(
+        $resolvedLockPath,
+        [System.IO.FileMode]::OpenOrCreate,
+        [System.IO.FileAccess]::ReadWrite,
+        [System.IO.FileShare]::None
+      )
+    } catch [System.IO.IOException] {
+      if ((Get-Date) -ge $deadline) {
+        throw "Another public runtime recovery is already in progress"
+      }
+      Start-Sleep -Seconds 1
+    }
+  } while ($true)
+}
+
+# boot task, watchdog, 수동 승격이 같은 포트와 프로세스를 동시에 교체하지 못하게 한다.
+# lock 파일은 stable Git root 밖에 두고 FileShare.None 핸들 수명으로만 소유권을 가진다.
+$recoveryLock = Enter-RecoveryLock `
+  -LockPath $RecoveryLockPath `
+  -WaitSeconds $RecoveryLockWaitSeconds
+
+try {
 # 엔진 readiness 캐시 TTL. 기본 30초는 워치독 주기(5분)보다 짧아 매 헬스체크마다
 # 실제 claude -p 생성을 새로 돌리게 만든다(재시작 폭풍의 근본 원인). 크게 늘려
 # /ready 가 거의 항상 캐시를 반환하게 한다 → 헬스체크가 LLM 호출에 묶이지 않는다.
@@ -344,6 +393,37 @@ function Stop-ProcessesBounded {
   throw "Timed out stopping $Role process IDs: $($remaining -join ',')"
 }
 
+function Stop-ProcessTreeBounded {
+  param(
+    [int]$RootProcessId,
+    [int]$TimeoutSec,
+    [string]$Role
+  )
+
+  $taskkill = Join-Path $env:SystemRoot "System32\taskkill.exe"
+  $previousErrorActionPreference = $ErrorActionPreference
+  try {
+    $ErrorActionPreference = "Continue"
+    & $taskkill @("/PID", "$RootProcessId", "/T", "/F") | Out-Null
+    $taskkillExit = $LASTEXITCODE
+  } finally {
+    $ErrorActionPreference = $previousErrorActionPreference
+  }
+  if ($taskkillExit -ne 0) {
+    throw "Failed to stop $Role process tree rooted at PID $RootProcessId (taskkill exit=$taskkillExit)"
+  }
+
+  $deadline = (Get-Date).AddSeconds($TimeoutSec)
+  do {
+    if ($null -eq (Get-Process -Id $RootProcessId -ErrorAction SilentlyContinue)) {
+      return @($RootProcessId)
+    }
+    Start-Sleep -Milliseconds 200
+  } while ((Get-Date) -lt $deadline)
+
+  throw "Timed out stopping $Role process tree rooted at PID $RootProcessId"
+}
+
 function Get-UvicornProcessesByPort {
   param(
     [string]$AppImport,
@@ -1470,11 +1550,11 @@ if ($apiControlPlaneReady -and -not $ForceApiRestart) {
       $health.db -eq $true -and
       $health.engine -eq $true
     } `
-    -TimeoutSec 30
+    -TimeoutSec $ApiReadySeconds
   $voiceHealth = Wait-JsonHealth `
     -Uri "http://127.0.0.1:$ApiPort/voice/health" `
     -IsHealthy { param($health) Test-VoiceApiReady -Health $health } `
-    -TimeoutSec 30
+    -TimeoutSec $VoiceApiReadySeconds
 }
 if ($health.environment -ne "prod" -or -not $health.db -or -not $health.engine) {
   throw "Admin/auth control plane is not production-safe: $($health | ConvertTo-Json -Compress)"
@@ -1493,8 +1573,17 @@ if (!$SkipWebRestart) {
     -ArgumentList @("/c", "npm run build") `
     -WorkingDirectory $WebDir `
     -NoNewWindow `
-    -Wait `
     -PassThru
+  if (-not $build.WaitForExit($WebBuildTimeoutSeconds * 1000)) {
+    $null = @(
+      Stop-ProcessTreeBounded `
+        -RootProcessId $build.Id `
+        -TimeoutSec $ProcessStopTimeoutSeconds `
+        -Role "web build"
+    )
+    throw "Web build timed out after $WebBuildTimeoutSeconds seconds"
+  }
+  $build.Refresh()
   if ($build.ExitCode -ne 0) {
     throw "Web build failed with exit code $($build.ExitCode)"
   }
@@ -1734,3 +1823,12 @@ if (!$SkipWebRestart) {
 }
 Write-Output "Health: $($health | ConvertTo-Json -Compress)"
 Write-Output "Voice health: $($voiceHealth | ConvertTo-Json -Compress)"
+} finally {
+  if ($null -ne $recoveryLock) {
+    try {
+      $recoveryLock.Dispose()
+    } catch {
+      Write-Warning "Public runtime recovery lock release failed: $($_.Exception.Message)"
+    }
+  }
+}
diff --git a/scripts/test_public_runtime_watchdog_provenance.py b/scripts/test_public_runtime_watchdog_provenance.py
index bffe07a..db8b877 100644
--- a/scripts/test_public_runtime_watchdog_provenance.py
+++ b/scripts/test_public_runtime_watchdog_provenance.py
@@ -98,7 +98,7 @@ class PublicRuntimeWatchdogProvenanceTest(unittest.TestCase):
         )
         gate = BOOT_SOURCE.index("Assert-StableSourceProvenance", gate_comment)
         docker_mutation = BOOT_SOURCE.index("docker.exe update")
-        runtime_recovery = BOOT_SOURCE.index("-File $startScript")
+        runtime_recovery = BOOT_SOURCE.index("& $startScript @startArgs |")
         self.assertLess(gate, docker_mutation)
         self.assertLess(gate, runtime_recovery)
 
@@ -120,6 +120,115 @@ class PublicRuntimeWatchdogProvenanceTest(unittest.TestCase):
         task_registration = BOOT_REGISTER_SOURCE.index("Register-ScheduledTask")
         self.assertLess(dirty_gate, task_registration)
 
+    def test_task_limits_cover_a_full_cold_start_without_overlap(self) -> None:
+        self.assertIn("-ExecutionTimeLimit (New-TimeSpan -Minutes 60)", BOOT_REGISTER_SOURCE)
+        self.assertIn("-MultipleInstances IgnoreNew", BOOT_REGISTER_SOURCE)
+        self.assertIn("-ExecutionTimeLimit (New-TimeSpan -Minutes 60)", INSTALLER_SOURCE)
+        self.assertIn("-MultipleInstances IgnoreNew", INSTALLER_SOURCE)
+        self.assertNotIn("-RestartCount", INSTALLER_SOURCE)
+        self.assertNotIn("-RestartInterval", INSTALLER_SOURCE)
+
+        # 최초 health 진단 + DB 복구 + engine/voice/API/web의 각 bounded wait와
+        # process-stop 여유를 모두 합쳐도 작업 스케줄러 60분보다 5분 이상 짧다.
+        initial_health_seconds = 210
+        recovery_seconds = 60 + 90 + (300 * 2) + 180 + 90 + 600 + 30 + (15 * 4)
+        safety_margin_seconds = 300
+        task_limit_seconds = 60 * 60
+        self.assertLess(
+            initial_health_seconds + recovery_seconds + safety_margin_seconds,
+            task_limit_seconds,
+        )
+        boot_prerequisite_seconds = 360 + 90
+        self.assertLess(
+            boot_prerequisite_seconds + recovery_seconds + safety_margin_seconds,
+            task_limit_seconds,
+        )
+
+    def test_hard_down_bypasses_transient_failure_debounce(self) -> None:
+        hard_down = WATCHDOG_SOURCE.index("$hardDown =")
+        debounce = WATCHDOG_SOURCE.index(
+            "if ($failCount -lt $FailuresBeforeRestart -and -not $hardDown)"
+        )
+        restart = WATCHDOG_SOURCE.index("& $startScript @startArgs")
+        self.assertLess(hard_down, debounce)
+        self.assertLess(debounce, restart)
+        self.assertIn('($FailedNames -contains "cloudflared")', WATCHDOG_SOURCE)
+        self.assertIn('($FailedNames -contains "api")', WATCHDOG_SOURCE)
+        self.assertIn('($FailedNames -contains "web-preview")', WATCHDOG_SOURCE)
+        self.assertIn('($FailedNames -contains "voice-sidecars")', WATCHDOG_SOURCE)
+        self.assertIn(
+            "immediate restart: public runtime hard-down detected",
+            WATCHDOG_SOURCE,
+        )
+
+    def test_hard_down_classifier_distinguishes_total_and_transient_failures(self) -> None:
+        powershell = shutil.which("powershell.exe")
+        if powershell is None:
+            self.skipTest("Windows PowerShell 5.1 is not available")
+
+        classifier = WATCHDOG_SOURCE[
+            WATCHDOG_SOURCE.index("function Test-PublicRuntimeHardDown") :
+            WATCHDOG_SOURCE.index("# engine 판정은", WATCHDOG_SOURCE.index("function Test-PublicRuntimeHardDown"))
+        ].strip()
+        with tempfile.TemporaryDirectory() as temporary_directory:
+            harness = Path(temporary_directory) / "hard-down.ps1"
+            harness.write_text(
+                classifier
+                + r'''
+if (-not (Test-PublicRuntimeHardDown -FailedNames @('cloudflared'))) { exit 2 }
+if (-not (Test-PublicRuntimeHardDown -FailedNames @('api','web-preview','voice-sidecars'))) { exit 3 }
+if (Test-PublicRuntimeHardDown -FailedNames @('engine')) { exit 4 }
+exit 0
+''',
+                encoding="utf-8-sig",
+            )
+            completed = subprocess.run(
+                [
+                    powershell,
+                    "-NoLogo",
+                    "-NoProfile",
+                    "-NonInteractive",
+                    "-ExecutionPolicy",
+                    "Bypass",
+                    "-File",
+                    str(harness),
+                ],
+                capture_output=True,
+                text=True,
+                encoding="utf-8",
+                errors="replace",
+                timeout=30,
+                check=False,
+            )
+            self.assertEqual(
+                completed.returncode,
+                0,
+                msg=f"stdout={completed.stdout}\nstderr={completed.stderr}",
+            )
+
+    def test_check_only_exits_before_failcount_log_or_recovery_mutation(self) -> None:
+        failed = WATCHDOG_SOURCE.index("$failed = @(")
+        check_only = WATCHDOG_SOURCE.index("if ($CheckOnly)", failed)
+        inner_health = WATCHDOG_SOURCE.index("if ($failed.Count -eq 0)", check_only)
+        normal_health = WATCHDOG_SOURCE.index(
+            "if ($failed.Count -eq 0)", inner_health + 1
+        )
+        fail_count = WATCHDOG_SOURCE.index("$failCount =", check_only)
+        restart = WATCHDOG_SOURCE.index("& $startScript @startArgs", fail_count)
+        self.assertLess(check_only, fail_count)
+        self.assertLess(check_only, restart)
+        check_only_contract = WATCHDOG_SOURCE[check_only:normal_health]
+        self.assertIn("Write-Output", check_only_contract)
+        self.assertNotIn("Set-FailCount", check_only_contract)
+        self.assertNotIn("Write-WatchdogLog", check_only_contract)
+        self.assertNotIn("Restore-DatabaseContainer", check_only_contract)
+
+    def test_boot_streams_start_script_progress_without_nested_powershell_buffering(self) -> None:
+        self.assertIn("& $startScript @startArgs |", BOOT_SOURCE)
+        self.assertIn('Write-BootLog ("  pub> " + $_)', BOOT_SOURCE)
+        self.assertIn("start-public-runtime.ps1 failed", BOOT_SOURCE)
+        self.assertNotIn("$out = & powershell.exe", BOOT_SOURCE)
+
     def test_watchdog_and_boot_probe_exact_local_voice_stack(self) -> None:
         for source in (WATCHDOG_SOURCE, BOOT_SOURCE):
             for expected in (
diff --git a/scripts/test_start_public_runtime_contract.py b/scripts/test_start_public_runtime_contract.py
index b55442e..81fef78 100644
--- a/scripts/test_start_public_runtime_contract.py
+++ b/scripts/test_start_public_runtime_contract.py
@@ -16,6 +16,160 @@ WHISPER_START = (SCRIPTS / "start-local-whisper-stt.ps1").read_text(
 
 
 class PublicRuntimeVoiceContractTest(unittest.TestCase):
+    def test_recovery_is_serialized_and_lock_is_always_released(self) -> None:
+        lock = PUBLIC_RUNTIME.index("$recoveryLock = Enter-RecoveryLock")
+        main_try = PUBLIC_RUNTIME.index("try {", lock)
+        first_runtime_mutation = PUBLIC_RUNTIME.index("Stop-UvicornByPort", main_try)
+        finalizer = PUBLIC_RUNTIME.rindex("} finally {")
+        dispose = PUBLIC_RUNTIME.index("$recoveryLock.Dispose()", finalizer)
+        self.assertLess(lock, main_try)
+        self.assertLess(main_try, first_runtime_mutation)
+        self.assertLess(finalizer, dispose)
+        self.assertIn(
+            '"$env:LOCALAPPDATA\\Vignette\\public-runtime-start.lock"',
+            PUBLIC_RUNTIME,
+        )
+
+    def test_recovery_lock_is_exclusive_and_reusable_in_windows_powershell(self) -> None:
+        powershell = shutil.which("powershell.exe")
+        if powershell is None:
+            self.skipTest("Windows PowerShell 5.1 is not available")
+
+        lock_function = PUBLIC_RUNTIME[
+            PUBLIC_RUNTIME.index("function Enter-RecoveryLock") :
+            PUBLIC_RUNTIME.index(
+                "# boot task, watchdog, 수동 승격이 같은 포트와 프로세스를 동시에 교체하지 못하게 한다."
+            )
+        ].strip()
+        with tempfile.TemporaryDirectory() as temporary_directory:
+            lock_path = str(Path(temporary_directory) / "runtime.lock").replace("'", "''")
+            harness = Path(temporary_directory) / "recovery-lock.ps1"
+            harness.write_text(
+                lock_function
+                + f"""
+$first = Enter-RecoveryLock -LockPath '{lock_path}' -WaitSeconds 0
+try {{
+  try {{
+    $second = Enter-RecoveryLock -LockPath '{lock_path}' -WaitSeconds 0
+    exit 2
+  }} catch {{
+    if ($_.Exception.Message -notlike '*already in progress*') {{ exit 3 }}
+  }}
+}} finally {{
+  $first.Dispose()
+}}
+$third = Enter-RecoveryLock -LockPath '{lock_path}' -WaitSeconds 0
+$third.Dispose()
+exit 0
+""",
+                encoding="utf-8-sig",
+            )
+            completed = subprocess.run(
+                [
+                    powershell,
+                    "-NoLogo",
+                    "-NoProfile",
+                    "-NonInteractive",
+                    "-ExecutionPolicy",
+                    "Bypass",
+                    "-File",
+                    str(harness),
+                ],
+                capture_output=True,
+                text=True,
+                encoding="utf-8",
+                errors="replace",
+                timeout=30,
+                check=False,
+            )
+            self.assertEqual(
+                completed.returncode,
+                0,
+                msg=f"stdout={completed.stdout}\nstderr={completed.stderr}",
+            )
+
+    def test_cold_start_waits_are_bounded_and_configurable(self) -> None:
+        for expected in (
+            "[int]$ApiReadySeconds = 180",
+            "[int]$VoiceApiReadySeconds = 90",
+            "[int]$WebBuildTimeoutSeconds = 600",
+            "-TimeoutSec $ApiReadySeconds",
+            "-TimeoutSec $VoiceApiReadySeconds",
+            "$build.WaitForExit($WebBuildTimeoutSeconds * 1000)",
+            "Web build timed out after $WebBuildTimeoutSeconds seconds",
+        ):
+            with self.subTest(expected=expected):
+                self.assertIn(expected, PUBLIC_RUNTIME)
+
+    def test_web_build_timeout_stops_the_owned_process_tree(self) -> None:
+        powershell = shutil.which("powershell.exe")
+        if powershell is None:
+            self.skipTest("Windows PowerShell 5.1 is not available")
+
+        tree_functions = PUBLIC_RUNTIME[
+            PUBLIC_RUNTIME.index("function Stop-ProcessTreeBounded") :
+            PUBLIC_RUNTIME.index("function Get-UvicornProcessesByPort")
+        ].strip()
+        with tempfile.TemporaryDirectory() as temporary_directory:
+            harness = Path(temporary_directory) / "process-tree.ps1"
+            child = Path(temporary_directory) / "child-sleeper.ps1"
+            child_pid = Path(temporary_directory) / "child.pid"
+            quoted_child_pid = str(child_pid).replace("'", "''")
+            child.write_text(
+                f"$PID | Set-Content -LiteralPath '{quoted_child_pid}' -Encoding ascii\n"
+                "Start-Sleep -Seconds 300\n",
+                encoding="utf-8-sig",
+            )
+            quoted_child = str(child).replace("'", "''")
+            quoted_child_pid_for_harness = str(child_pid).replace("'", "''")
+            harness.write_text(
+                tree_functions
+                + f'''
+$root = Start-Process -FilePath 'cmd.exe' `
+  -ArgumentList @('/c', 'powershell.exe -NoProfile -ExecutionPolicy Bypass -File "{quoted_child}"') `
+  -WindowStyle Hidden `
+  -PassThru
+try {{
+  $deadline = (Get-Date).AddSeconds(10)
+  while (-not (Test-Path -LiteralPath '{quoted_child_pid_for_harness}') -and (Get-Date) -lt $deadline) {{
+    Start-Sleep -Milliseconds 100
+  }}
+  if (-not (Test-Path -LiteralPath '{quoted_child_pid_for_harness}')) {{ exit 2 }}
+  $childProcessId = [int](Get-Content -LiteralPath '{quoted_child_pid_for_harness}' -Raw)
+  $stopped = @(Stop-ProcessTreeBounded -RootProcessId $root.Id -TimeoutSec 10 -Role 'test tree')
+  if ($null -ne (Get-Process -Id $root.Id -ErrorAction SilentlyContinue)) {{ exit 3 }}
+  if ($null -ne (Get-Process -Id $childProcessId -ErrorAction SilentlyContinue)) {{ exit 4 }}
+  exit 0
+}} finally {{
+  Stop-Process -Id $root.Id -Force -ErrorAction SilentlyContinue
+}}
+''',
+                encoding="utf-8-sig",
+            )
+            completed = subprocess.run(
+                [
+                    powershell,
+                    "-NoLogo",
+                    "-NoProfile",
+                    "-NonInteractive",
+                    "-ExecutionPolicy",
+                    "Bypass",
+                    "-File",
+                    str(harness),
+                ],
+                capture_output=True,
+                text=True,
+                encoding="utf-8",
+                errors="replace",
+                timeout=30,
+                check=False,
+            )
+            self.assertEqual(
+                completed.returncode,
+                0,
+                msg=f"stdout={completed.stdout}\nstderr={completed.stderr}",
+            )
+
     def test_public_provider_model_and_loopback_environment_are_explicit(self) -> None:
         for expected in (
             '$WhisperModel = "small"',
diff --git a/scripts/watch-public-runtime.ps1 b/scripts/watch-public-runtime.ps1
index 5ee9cdb..6563cc7 100644
--- a/scripts/watch-public-runtime.ps1
+++ b/scripts/watch-public-runtime.ps1
@@ -311,7 +311,7 @@ function Test-DatabaseListening {
 # 재시작 전에 컨테이너를 직접 되살린다.
 function Restore-DatabaseContainer {
   try {
-    $output = & docker.exe start $DbContainer 2>&1
+    $output = & docker.exe start $DbContainer 2>$null
     $output | ForEach-Object { Write-WatchdogLog "  docker> $_" }
     if ($LASTEXITCODE -ne 0) {
       Write-WatchdogLog "db restore: docker start exit=$LASTEXITCODE"
@@ -329,6 +329,17 @@ function Restore-DatabaseContainer {
   return $false
 }
 
+function Test-PublicRuntimeHardDown {
+  param([string[]]$FailedNames)
+
+  $allLocalSurfacesDown = (
+    ($FailedNames -contains "api") -and
+    ($FailedNames -contains "web-preview") -and
+    ($FailedNames -contains "voice-sidecars")
+  )
+  return ($FailedNames -contains "cloudflared") -or $allLocalSurfacesDown
+}
+
 # engine 판정은 /health(프로세스 liveness)가 아니라 /ready(실제 claude -p 생성)로 한다.
 # /health는 ok:true만 보므로 "프로세스는 살아 있고 그 프로세스의 claude 세션만 죽은"
 # 상태를 통과시킨다(2026-08-07 공개 런타임: engine=false인데 워치독 lastResult=0).
@@ -376,6 +387,14 @@ if (!$SkipPublicHealth) {
 }
 
 $failed = @($checks | Where-Object { -not $_.Ok })
+if ($CheckOnly) {
+  if ($failed.Count -eq 0) {
+    Write-Output "healthy: $($checks.Name -join ', ')"
+    exit 0
+  }
+  Write-Output "unhealthy: $((($failed | ForEach-Object { "$($_.Name)=$($_.Detail)" }) -join '; '))"
+  exit 1
+}
 if ($failed.Count -eq 0) {
   Set-FailCount 0
   Write-WatchdogLog "healthy: $($checks.Name -join ', ')"
@@ -385,17 +404,20 @@ if ($failed.Count -eq 0) {
 $failCount = (Get-FailCount) + 1
 Set-FailCount $failCount
 Write-WatchdogLog "unhealthy ($failCount/$FailuresBeforeRestart): $((($failed | ForEach-Object { "$($_.Name)=$($_.Detail)" }) -join '; '))"
-if ($CheckOnly) {
-  exit 1
-}
 
 # 연속 실패 debounce: claude -p readiness probe 는 콜드 스폰 시 10~20초가 정상이라
 # 단발 timeout 을 장애로 오판해 전체 재시작하던 것이 재시작 폭풍의 원인이었다.
-# 연속 $FailuresBeforeRestart 회 실패해야 실제 재시작한다.
-if ($failCount -lt $FailuresBeforeRestart) {
+# 다만 cloudflared가 사라졌거나 API·web·voice sidecar가 함께 죽은 경우는 transient
+# readiness가 아니라 공개 서비스 전체 소실이므로 첫 감지에서 즉시 복구한다.
+$failedNames = @($failed | ForEach-Object { $_.Name })
+$hardDown = Test-PublicRuntimeHardDown -FailedNames $failedNames
+if ($failCount -lt $FailuresBeforeRestart -and -not $hardDown) {
   Write-WatchdogLog "defer restart: $FailuresBeforeRestart 연속 실패 전까지 대기 (현재 $failCount)"
   exit 1
 }
+if ($hardDown -and $failCount -lt $FailuresBeforeRestart) {
+  Write-WatchdogLog "immediate restart: public runtime hard-down detected"
+}
 
 # DB가 죽어 있으면 start-public-runtime.ps1으로는 절대 복구되지 않는다(DB 기동은 boot 담당).
 # 먼저 되살리고, 그래도 안 되면 재시작을 아예 시도하지 않는다 — 고칠 수 없는 대상에
@@ -429,7 +451,7 @@ if ($SkipCloudflaredRestart) {
 }
 
 try {
-  & $startScript @startArgs 2>&1 | ForEach-Object {
+  & $startScript @startArgs | ForEach-Object {
     Write-WatchdogLog "$_"
   }
 } catch {