Qwen3.6-35B-A3B turns gibberish on RAM spill

This looks more like a llama.cpp memory-management edge case than a prompt or sampling problem. The combination of Unified Memory, aggressive fit settings, and long-context offload is exactly where backend bugs tend to surface.

• –Upstream llama.cpp has known reports that CUDA Unified Memory is flaky for models that spill beyond VRAM, especially under heavy decode/load pressure.
• –The posted config is very aggressive: `--fit`, `--fit-target 256`, `--fit-ctx 204800`, `--kv-offload`, and large batch settings all increase the chance of allocator or paging issues.
• –Qwen3.6 is a sparse MoE model, so total parameter count is not the whole story, but long-context KV cache pressure can still push a run into unstable memory behavior.
• –Recent llama.cpp chatter also suggests some IQ3_S/CUDA combinations have regressions, so changing quantization or CUDA/toolchain version is a plausible next test.
• –For developers serving large local models, this is a reminder that “fits in RAM” is not the same as “runs correctly once pages start migrating.”