arle-upstream-runtime-scanlisted

# arle-upstream-runtime-scan This is an ARLE source-survey skill, not a replacement for local measurement. Use it to decide what upstream evidence to read before an ARLE runtime change, then license or kill the local decision with ARLE logs, benches, traces, and tests. ## Scope Use for: - Comparing ARLE against SGLang, vLLM, TensorRT-LLM, or MLX-style serving. - Choosing a Qwen3.5, Qwen3-Next, DeepSeek V3/V4, MoE, attention, cache, sampler, loader, or quantization precedent before local implementation. - Explaining capacity, KV-cache budget, memory fraction, max tokens, continuous batching, prefix-cache, or queue-growth behavior. - Reading external profiler traces where prefill/decode stage separation matters. - Designing OPD/eval serving baselines that need a fair external reference. Do not use for: - Local kernel retuning. Use `kernel-optimization`; BBuf KernelWiki and KernelPilot are merged there. - Pure SGLang PR review, production incident response, or architecture-diagram lookup. Those BBuf skills are not ARLE-specific enough. - Any claim that can be answered by local ARLE code, docs, or measurement without upstream context. ## Non-Negotiables - Start from ARLE truth surfaces: `docs/index.md`, relevant `AGENTS.md`, current code, and recent `docs/experience/{wins,errors}`. - Treat upstream source survey as hypothesis-grade. It may suggest a design or risk; it does not prove ARLE behavior. - Fetch external repos into `/tmp` for inspection when need