Keep llama-server alive across turns #16

Problem. llama-server is started and killed per process, and AgentMux spawns one process per turn, so the model reloads every turn.

Do. Run llama-server as a reusable daemon keyed by model + settings (lockfile + port under the config root), shared across Mux Code processes, with an idle timeout and a health check. Clean shutdown command.

Spec §5.2 #1.

Spec: SPEC_MUXCODE_HARNESS_AND_FEATURE_GAPS_2026_09_26.md