Guide the agent through planning the integration of a custom CUDA or Triton kernel into the vLLM inference engine before any integration code is written — covering where the op plugs into the engine, paged KV cache and continuous batching compatibility, CUDA graph capture constraints, tensor parallelism implications, and the testing and benchmarking strategy. This skill produces an integration plan, not a kernel implementation.
