🟠 LOW SEV · Inference Consume Latency p99 High 🔥 device-events consume latency p99 > 3s <|📈 graph> · <|📊 service dashboard> · :rotating_light: alarm definition · :no_bell: silence _Use consume_latency p99 as the real lag signal - the devices_behind / max_device_lag / inference_consumer_live gauges are miscalibrated by the bursty iOS report cadence and will lie. Prod baseline is 0.06-0.46s with a single 2.05s spike in 7d. NoData is intentionally OK here: histogram_quantile returns NaN when the consumer is gone, and that case is already covered by the Down / Coverage Lost criticals._ Grafana vmain-8eb09b4