fix(telemetry): restore worker agent delivery and truthful system status

deploy/telemetry-plane/telegraf/Get-NdcMissionCorePipelineTelemetry.ps1

@@ -1,7 +1,8 @@
 [CmdletBinding()]
 param(
     [string]$ContainerName = "ndc-mission-core-perception-worker",
-    [string]$HealthUrl = "http://127.0.0.1:18020/health"
+    [string]$HealthUrl = "http://127.0.0.1:18020/health",
+    [string]$RuntimeSnapshot = "C:\ProgramData\NDC\MissionCore\telemetry-agent\perception\current.json"
 )
 
 $ErrorActionPreference = "Stop"
@@ -18,6 +19,46 @@ $stageIds = @(
 )
 $health = $null
 $collectorState = "unavailable"
+# A fresh full-profile export supersedes the historical service. Presence of an
+# expired/broken export is NOT permission to relabel a legacy model as current.
+if (Test-Path -LiteralPath $RuntimeSnapshot -PathType Leaf) {
+    $runtime = $null
+    try {
+        $item = Get-Item -LiteralPath $RuntimeSnapshot
+        $age = ([DateTime]::UtcNow - $item.LastWriteTimeUtc).TotalSeconds
+        if ($item.Length -le 8192 -and $age -ge -1 -and $age -le 5) {
+            $stream = [IO.File]::Open($RuntimeSnapshot, 'Open', 'Read', 'ReadWrite')
+            try {
+                $bytes = New-Object byte[] 8193
+                $count = $stream.Read($bytes, 0, $bytes.Length)
+                if ($count -le 8192) {
+                    $runtime = [Text.Encoding]::UTF8.GetString($bytes, 0, $count) | ConvertFrom-Json
+                }
+            } finally { $stream.Dispose() }
+            if ($runtime.schema_version -ne 'missioncore.perception-runtime-observation/v1') {
+                $runtime = $null
+            }
+        }
+    } catch { $runtime = $null }
+    $samples = foreach ($stageId in $stageIds) {
+        $sample = [ordered]@{
+            stage_id = $stageId
+            stage_state = 'unavailable'
+            service_state = 'unavailable'
+            collector_state = 'unavailable'
+        }
+        if ($runtime) {
+            $sample.collector_state = 'live'
+            foreach ($name in @('service_state', 'profile_name', 'active_request_id',
+                'input_state', 'wait_reason', 'live_children', 'input_pauses', 'pending_bundles', 'buffer_bytes')) {
+                $sample[$name] = $runtime.$name
+            }
+        }
+        [pscustomobject]$sample
+    }
+    @($samples) | ConvertTo-Json -Compress -Depth 4
+    exit 0
+}
 try {
     $healthJson = docker exec $ContainerName python3 -c `
         "import urllib.request;print(urllib.request.urlopen('$HealthUrl',timeout=2).read().decode())" `
@@ -68,13 +109,13 @@ $completedRuns = if ($health -and $null -ne $health.completed_runs) {
     [int64]$health.completed_runs
 }
 else {
-    [int64]0
+    [int64]-1
 }
 $failedRuns = if ($health -and $null -ne $health.failed_runs) {
     [int64]$health.failed_runs
 }
 else {
-    [int64]0
+    [int64]-1
 }
 $modelLoadSeconds = if ($health -and $null -ne $health.model_load_seconds) {
     [double]$health.model_load_seconds

deploy/telemetry-plane/telegraf/mission-core-linux.conf.tmpl

@@ -1,5 +1,8 @@
 [agent]
   interval = "${MISSIONCORE_TELEMETRY_INTERVAL}"
+  flush_interval = "${MISSIONCORE_TELEMETRY_INTERVAL}"
+  metric_batch_size = 200
+  metric_buffer_limit = 2000
   round_interval = true
   omit_hostname = false
 
@@ -26,6 +29,7 @@
   total = true
 
 [[outputs.mqtt]]
+  startup_error_behavior = "retry"
   servers = ["tcp://${MISSIONCORE_MQTT_HOST}:${MISSIONCORE_MQTT_PORT}"]
   topic = "mission-core/v1/contours/${MISSIONCORE_CONTOUR_ID}/agents/${MISSIONCORE_AGENT_ID}/host"
   username = "${MISSIONCORE_MQTT_USERNAME}"

deploy/telemetry-plane/telegraf/mission-core-windows.conf.tmpl

@@ -1,5 +1,8 @@
 [agent]
   interval = "${MISSIONCORE_TELEMETRY_INTERVAL}"
+  flush_interval = "${MISSIONCORE_TELEMETRY_INTERVAL}"
+  metric_batch_size = 200
+  metric_buffer_limit = 2000
   round_interval = true
   omit_hostname = false
 
@@ -30,6 +33,7 @@
 [[inputs.docker]]
   endpoint = "npipe:////./pipe/docker_engine"
   container_name_include = []
+  container_state_include = ["running", "exited", "restarting", "paused"]
 
 [[inputs.http_response]]
   urls = ["http://127.0.0.1:8000/v2/health/ready"]
@@ -59,6 +63,9 @@
     "current_stage",
     "active_request_id",
     "collector_state",
+    "profile_name",
+    "input_state",
+    "wait_reason",
   ]
 
 [[inputs.tail]]
@@ -73,6 +80,7 @@
   data_type = "string"
 
 [[outputs.mqtt]]
+  startup_error_behavior = "retry"
   servers = ["tcp://${MISSIONCORE_MQTT_HOST}:${MISSIONCORE_MQTT_PORT}"]
   topic = "mission-core/v1/contours/${MISSIONCORE_CONTOUR_ID}/agents/${MISSIONCORE_AGENT_ID}/host"
   username = "${MISSIONCORE_MQTT_USERNAME}"
@@ -83,6 +91,7 @@
   namedrop = ["missioncore_pipeline", "missioncore_pipeline_event"]
 
 [[outputs.mqtt]]
+  startup_error_behavior = "retry"
   servers = ["tcp://${MISSIONCORE_MQTT_HOST}:${MISSIONCORE_MQTT_PORT}"]
   topic = "mission-core/v1/contours/${MISSIONCORE_CONTOUR_ID}/agents/${MISSIONCORE_AGENT_ID}/pipeline"
   username = "${MISSIONCORE_MQTT_USERNAME}"