Delay-coupled systems often require low-latency decisions from sparse telemetry, where dense fixed-step neural inference is wasteful and can degrade near stability margins. We introduce Network-Optimised Spiking (NOS), a trainable two-state event-driven dynamical unit for delayed, graph-coupled streams, whose states map to a fast load variable and a slower recovery resource. NOS uses bounded excitability for finite buffers, explicit leak terms for service and damping, and graph-local coupling with per-link gates and communication delays, with differentiable resets compatible with surrogate-gradient training and neuromorphic execution. We prove existence and uniqueness of subthreshold equilibria, derive Jacobian-based stability conditions, and obtain a scalar network stability threshold that separates topology from node dynamics via a Perron-mode spectral condition. A stochastic arrival model aligned with telemetry smoothing explains increased variability as systems approach stability boundaries. On delayed graph forecasting and early-warning tasks from queue telemetry, NOS improves detection F1 and detection latency over MLP, RNN/GRU, and temporal GNN baselines under a common residual-based protocol, while providing calibration rules for resource-constrained deployments. Code and Demos: https://mbilal84.github.io/nos-snn-networking/

翻译：延迟耦合系统通常需要基于稀疏遥测数据进行低延迟决策，而密集的固定步长神经推断不仅会造成资源浪费，还可能降低系统在稳定性边界附近的鲁棒性。本文提出网络优化脉冲（NOS）单元，这是一种可训练的双状态事件驱动动力学单元，专为具有延迟的图耦合数据流设计，其状态映射为快速负载变量与慢速恢复资源。NOS采用有限缓冲区下的有界兴奋性机制，通过显式泄漏项实现服务与阻尼功能，并利用带逐链路门控与通信延迟的图局部耦合结构；其可微分重置机制兼容代理梯度训练与神经形态计算执行。我们证明了亚阈值平衡点的存在性与唯一性，推导了基于雅可比矩阵的稳定性条件，并通过佩龙模谱条件获得将拓扑结构与节点动力学解耦的标量化网络稳定性阈值。与遥测平滑处理对齐的随机到达模型解释了系统趋近稳定性边界时变异性增加的现象。在基于队列遥测的延迟图预测与预警任务中，采用统一残差协议时，NOS在检测F1分数与检测延迟方面均优于MLP、RNN/GRU及时序GNN基线模型，同时为资源受限部署场景提供了校准规则。代码与演示：https://mbilal84.github.io/nos-snn-networking/