Llama.cpp samplers fail on Gemma 4 architecture

The "no variance" bug in Gemma 4 isn't just a settings issue; it's an architectural mismatch between legacy samplers and new logit dynamics.

• –Gemma 4's logit soft-capping requires specialized handling that was only recently stabilized in PR #21390.
• –The migration of sampling logic directly into the CUDA computation graph has introduced state synchronization errors for floating-point parameters.
• –Users are seeing coherent output at temperature 1000 because the sampler is defaulting to greedy decoding when it encounters invalid logit states.
• –Re-downloading GGUFs is mandatory for many as early April 2026 quants lacked the necessary `add_bos` and metadata for Gemma 4's specific reasoning budget.
• –The community is pivoting toward "Min P" and the new "Reasoning Budget Sampler" as the only reliable way to maintain quality on these newer architectures.