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

src/k1link/web/compute_contour_network.py

@@ -486,7 +486,7 @@ def _run_compose_broker(telemetry_plane_root: Path) -> None:
     ]
     for suffix, timeout in (
         (["config", "--quiet"], 20),
-        (["up", "-d", "--no-build", "broker"], 45),
+        (["up", "-d", "--no-deps", "--no-build", "--force-recreate", "broker"], 45),
     ):
         try:
             completed = subprocess.run(
@@ -525,11 +525,17 @@ def apply_broker_network(
         target.mqtt_bind_address,
     )
     changed = after != before
+    repaired = False
     mode = environment_path.stat().st_mode
     try:
         if changed:
             _atomic_private_write(environment_path, after, mode=mode)
             _run_compose_broker(telemetry_plane_root)
+        elif not _tcp_reachable(target.mqtt_bind_address, target.mqtt_port, timeout=3):
+            # An unchanged file does not prove Docker still publishes the port
+            # after a host/network restart. This is an explicit Apply, not a GET.
+            _run_compose_broker(telemetry_plane_root)
+            repaired = True
         if not _tcp_reachable(
             target.mqtt_bind_address,
             target.mqtt_port,
@@ -553,6 +559,7 @@ def apply_broker_network(
         "contour_id": target.contour_id,
         "target": "broker",
         "changed": changed,
+        "repaired": repaired,
         "applied_at_utc": _utc_now(),
         "ready": True,
     }

src/k1link/web/system_telemetry_api.py

@@ -478,7 +478,7 @@ def _agent_raw_document(document: dict[str, Any]) -> dict[str, Any]:
     perception: dict[str, Any] = raw["perception"]
     pipeline_stage_metrics: dict[str, dict[str, Any]] = {}
     pipeline_active_stages: set[str] = set()
-    for sample in samples:
+    for sample in sorted(samples, key=lambda item: str(item.get("observed_at_utc", ""))):
         measurement = sample.get("measurement")
         payload, fields, tags = _metric_payload(sample)
         if sample.get("kind") == "pipeline":
@@ -497,6 +497,9 @@ def _agent_raw_document(document: dict[str, Any]) -> dict[str, Any]:
                     "service_state",
                     "current_stage",
                     "active_request_id",
+                    "profile_name",
+                    "input_state",
+                    "wait_reason",
                 ):
                     value = fields.get(name)
                     if isinstance(value, str):
@@ -511,6 +514,8 @@ def _agent_raw_document(document: dict[str, Any]) -> dict[str, Any]:
                 if model_load_seconds is not None:
                     perception["model_load_seconds"] = model_load_seconds
                 perception["collector_state"] = fields.get("collector_state")
+                for name in ("live_children", "input_pauses", "pending_bundles", "buffer_bytes"):
+                    perception[name] = _number(fields.get(name))
             elif sample.get("source_schema") == PIPELINE_TELEMETRY_SCHEMA:
                 # Native lifecycle rows are immutable run evidence. The current
                 # service snapshot remains owned by the periodic health sample,
@@ -593,6 +598,9 @@ def _agent_raw_document(document: dict[str, Any]) -> dict[str, Any]:
                 "memory_total_mib": memory_total,
                 "power_watts": _number(fields.get("power_draw")),
                 "temperature_celsius": _number(fields.get("temperature_gpu")),
+                "sm_clock_mhz": _number(fields.get("clocks_current_sm")),
+                "memory_clock_mhz": _number(fields.get("clocks_current_memory")),
+                "driver_version": fields.get("driver_version"),
             }
         elif measurement == "missioncore_triton_health":
             raw["triton"] = {
@@ -637,7 +645,15 @@ def _agent_raw_document(document: dict[str, Any]) -> dict[str, Any]:
                     "image": tags.get("container_image"),
                 },
             )
-            if measurement == "docker_container_cpu":
+            if measurement == "docker_container_status":
+                container_states[container_name]["state"]["Status"] = (
+                    tags.get("container_status") or "unavailable"
+                )
+            elif measurement == "docker_container_health":
+                container_states[container_name]["state"]["Health"] = {
+                    "Status": fields.get("health_status") or "unknown",
+                }
+            elif measurement == "docker_container_cpu":
                 usage = _number(fields.get("usage_percent"))
                 if usage is not None:
                     stats["CPUPerc"] = f"{usage:.3f}%"
@@ -906,7 +922,12 @@ def _pipeline_document(raw: dict[str, Any]) -> dict[str, Any]:
     )
 
     def stage_state(stage_id: str) -> str:
-        if not perception:
+        if not perception or perception.get("state") == "unavailable":
+            return "unavailable"
+        if perception.get("state") in ("waiting", "synchronizing", "starting"):
+            return "waiting"
+        # A full-profile heartbeat is not a measurement of each individual stage.
+        if perception.get("profile_name"):
             return "unavailable"
         if not busy:
             return "ready"
@@ -970,6 +991,13 @@ def _pipeline_document(raw: dict[str, Any]) -> dict[str, Any]:
             else None
         ),
         "model_load_seconds": _number(perception.get("model_load_seconds")),
+        "profile_name": perception.get("profile_name"),
+        "input_state": perception.get("input_state"),
+        "wait_reason": perception.get("wait_reason"),
+        "live_children": _number(perception.get("live_children")),
+        "input_pauses": _number(perception.get("input_pauses")),
+        "pending_bundles": _number(perception.get("pending_bundles")),
+        "buffer_bytes": _number(perception.get("buffer_bytes")),
         "stages": [stage_document(stage_id, label) for stage_id, label in stages],
     }