Antes de 30I-C, el runtime respondía correctamente pero de forma ciega a telemetría fina. Usaba principalmente:

• inference_nodes
• inventory de modelos
• topología básica

Sin exponer con consistencia:

• temperatura
• potencia
• carga GPU
• VRAM usada/libre
• fan RPM
• source_of_truth
• freshness
• confidence

build_gpu_operational_summary() y luego gpu_operational_summaries.

flowchart LR
    GX[GPU exporter] --> P[Prometheus]
    P --> SF[sensor_fusion.py]
    SF --> SNAP[sensor_snapshot]
    SNAP --> GPS[gpu_operational_summaries]
    GPS --> ANSWER[short GPU answer]

• inventory_state = known
• observed_state = online
• operational_state = active
• métricas vivas disponibles

• inventory_state = known
• observed_state = expected_offline
• operational_state = inactive
• sin métricas vivas inventadas

• no activa Multi-GPU
• no implica scheduler
• no convierte inventory en runtime activo