Gemma 4 31B loops in llama.cpp

Gemma 4's specialized "thinking" mode is exposing deep-seated tokenizer assumptions in local inference engines.

• –A critical bug in llama.cpp was splitting double newlines into separate tokens, causing the model to lose coherence in long-form reasoning sessions.
• –Specialized tool-call tokens were incorrectly classified as user-defined instead of control tokens, preventing the parser from identifying reasoning boundaries.
• –Users on build b8693 should verify their GGUF files; those exported before April 4, 2026, lack essential tokenizer fixes regardless of the runtime version.
• –Stability requires the `--jinja` flag to correctly process the model's internal templates and setting `--min-p 0.0` to avoid interference with Gemma 4's sampling logic.
• –A new "Unified KV Cache" update has significantly improved VRAM efficiency, but users must update to the latest master branch to benefit from the reduced memory footprint.