server: --batch-size 0 aborts in output_reserve #37

Upstream: ggml-org#28525

llama-server accepts --batch-size 0 and then aborts during context initialization instead of rejecting the invalid value cleanly.

Environment: current main 67672dc5b, Linux aarch64, NVIDIA GB10, CUDA 13.0, Gemma 4 12B Q4_K_M.

llama-server -m model.gguf -ngl 999 -c 512 --batch-size 0

Observed: exit 134 with GGML_ASSERT(n_outputs >= 1) from llama_context::output_reserve().

The CLI setter assigns zero without validation. llama_init_from_model() only rejects the case where both n_batch and n_ubatch are zero, so the default nonzero n_ubatch lets n_batch=0 reach context allocation.