fix: [sc-106107] Make the wedged-stop assertion measure the stop it issues, and survive launchctl's exit code

The macOS job failed this step immediately after it had printed its own success
message. The pwsh shell ends by exiting with $LASTEXITCODE, and `launchctl stop`
returns non-zero in situations that are not failures here - the repo's own
stop-service action already guards it with `|| true`. Windows and Linux passed
only because sc.exe and systemctl happen to return zero. Report the stop
command's exit code and clear it: whether the stop was honored is decided by the
agent's log, not by the service manager's exit status. The step now exits
explicitly on both paths rather than inheriting whatever was left behind.

That failure also exposed a second, quieter bug: the assertion compared against
the scenario-wide "Service stopped" baseline, but the --update that points the
agent at the stub broker restarts the service and logs its own "Service
stopped". The delta was therefore already satisfied before the stop under test
was issued, which is why the step reported success after 0.2s - it was observing
the update's restart, not the wedged stop, on every platform. Count the lines
immediately before issuing the stop so the delta can only come from it.

Author: mlataza

.github/workflows/integration-test.yml

@@ -520,12 +520,10 @@ jobs:
           $subscribed = Get-PatternCount $content "Subscribed to messages"
           $timedOut   = Get-PatternCount $content "Failed to subscribe: timed out waiting for broker acknowledgement"
           $backoffs   = Get-PatternCount $content "Reconnecting in"
-          $stops      = Get-PatternCount $content "Service stopped"
           "subscribed=$subscribed" >> $env:GITHUB_OUTPUT
           "timed_out=$timedOut"    >> $env:GITHUB_OUTPUT
           "backoffs=$backoffs"     >> $env:GITHUB_OUTPUT
-          "stops=$stops"           >> $env:GITHUB_OUTPUT
-          Write-Output "Baseline -> subscribed=$subscribed timed_out=$timedOut backoffs=$backoffs stops=$stops"
+          Write-Output "Baseline -> subscribed=$subscribed timed_out=$timedOut backoffs=$backoffs"
 
       # Read off a logged path rather than re-deriving the platform path; see the
       # fuller note on the bounded-output scenario's copy of this step.
@@ -670,9 +668,21 @@ jobs:
         env:
           LOG_FILE: ${{ matrix.log_file }}
           SERVICE: ${{ matrix.service }}
-          BASE_STOPS: ${{ steps.suback_baseline.outputs.stops }}
         run: |
-          $baseStops = [int]$env:BASE_STOPS
+          function Get-StopCount {
+            $content = Get-Content $env:LOG_FILE -Raw -ErrorAction SilentlyContinue
+            if ($content) { ([regex]::Matches($content, [regex]::Escape("Service stopped"))).Count } else { 0 }
+          }
+
+          # Counted here rather than reusing the scenario-wide baseline: the
+          # --update that pointed the agent at the stub broker restarts the
+          # service, and that restart logs its own "Service stopped". Measured
+          # against the older baseline the delta was already satisfied before this
+          # stop was even issued, so the assertion "passed" in 0.2s without
+          # observing the stop it exists to test.
+          $before = Get-StopCount
+          Write-Output "Service stopped lines before issuing the stop: $before"
+
           $sw = [System.Diagnostics.Stopwatch]::StartNew()
 
           if ($IsWindows) {
@@ -683,14 +693,21 @@ jobs:
             sudo launchctl stop "system/$($env:SERVICE)"
           }
 
+          # launchctl exits non-zero in situations that are not failures here -
+          # the repo's own stop-service action already guards it with `|| true` -
+          # and the pwsh shell ends by exiting with $LASTEXITCODE, which failed
+          # this step on macOS even after the assertion below had passed. Report
+          # the code and clear it; whether the stop was honored is decided by the
+          # agent's log, not by the service manager's exit status.
+          Write-Output "Stop command exit code: $LASTEXITCODE"
+          $global:LASTEXITCODE = 0
+
           # A clean exit runs the agent's deferred teardown, whose last act is the
           # "Service stopped" line; a force-kill past the stop deadline never
           # produces one.
           $stopped = $false
           while ($sw.Elapsed.TotalSeconds -lt 30) {
-            $content = Get-Content $env:LOG_FILE -Raw -ErrorAction SilentlyContinue
-            $stops = if ($content) { ([regex]::Matches($content, [regex]::Escape("Service stopped"))).Count } else { 0 }
-            if ($stops -gt $baseStops) { $stopped = $true; break }
+            if ((Get-StopCount) -gt $before) { $stopped = $true; break }
             Start-Sleep -Milliseconds 500
           }
           $sw.Stop()
@@ -700,6 +717,7 @@ jobs:
             exit 1
           }
           Write-Output "Service stopped cleanly after $([math]::Round($sw.Elapsed.TotalSeconds, 1))s while wedged"
+          exit 0
 
       - name: Assert no orphaned scripts after the wedged stop
         uses: ./.github/actions/assert-no-orphaned-scripts