Qwen3.6-27B falters near 90K context

This looks less like a surprise bug than a familiar long-context cliff: the window exists, but agentic reliability still drops as prompts get crowded.

• –The failure mode matters: tool calling breaks before plain code generation does, which suggests structured output and planning are the first things to degrade.
• –Q4_K_XL plus q8_0 KV cache on AMD/llama.cpp is a real-world stress test, so the hardware/runtime stack may be amplifying the context cliff.
• –Qwen's own docs recommend tool-use parsers and newer serving stacks like vLLM or SGLang for this model, which hints the benchmark-friendly path is not always the best path for local agent loops.
• –For devops-style work at 90K+, retrieval, shorter working sets, or a stronger long-context serving setup will likely beat brute-forcing a larger nominal window.