Diagnosing Sudden Latency Spikes in `llama.cpp` Inference Pods: Distinguishing Application Delays from Resource Contention