Gemma 4 context shift, quantized KV crashes

This is the kind of edge-case failure that tends to show up first when a model adds longer context and inference optimizations at the same time. If you're planning to run Gemma 4 locally, assume the plain path is safer than KV-cache quantization until upstream runtimes harden the code.

• –Google’s launch materials say Gemma 4 supports llama.cpp and quantized deployments, so the issue looks like a backend regression or unsupported interaction, not an officially intended behavior.
• –The crash only showing up with `--quantkv 1` points at KV-cache remapping or shifting logic, which is exactly where long-context inference gets fragile.
• –For local-agent workloads, this is a warning to benchmark the specific runtime/flag combination you actually plan to ship, not just the base model.
• –Even a strong open model loses practical value if the “efficient” inference path is unstable; reliability matters as much as raw benchmark wins.