LLM feature runners

LLM runners share the setup/cli/llm/<feature>.sh namespace. Each feature owns its own facts and managed artifacts while consuming earlier contracts.

• Host readiness — CPU, NUMA, memory, reserve, and GPU-producer prerequisites.
• llama.cpp — exact-source CPU/CUDA build and opt-in local-GGUF smoke.
• KTransformers — exact-source KT-Kernel and SGLang-KT model-free toolchain.

The accepted order is host readiness, standalone llama.cpp build/smoke, KTransformers toolchain, model catalog/acquisition, then the bounded 80 GB MoE smoke. llama.cpp and KTransformers accept exact reviewed repository URLs and full commits so future builds can be added through compatibility-matrix updates without following floating upstream branches.