Pluggable LLM inference observability with layer timelines, heatmaps, and KV-cache metrics for vLLM, Ollama, and Hugging Face.