OpenCode agent temps bypass llama.cpp defaults

The likely bug here is config precedence, not broken sampling. OpenCode can specify `temperature` per agent, but llama.cpp also has its own defaults, so the server log alone is not proof that the agent setting was ignored.

• –OpenCode’s docs explicitly support agent-scoped `temperature`, and say omitted values fall back to model-specific defaults.
• –llama.cpp’s server has its own baseline sampler defaults, so startup/verbose output can be misleading if you’re expecting it to mirror per-request values.
• –The practical fix is to set a sensible default in llama.cpp and then tune agent/subagent temperatures in OpenCode.
• –For local models, the model card’s recommended sampler settings still matter more than generic “best” temperatures.
• –This is a classic client-vs-server config mismatch: if you want to verify behavior, inspect the actual request payload, not just the server’s default log line.