Traditional greenhouse climate control requires actuators to maintain a suitable microclimate while limiting energy use and avoiding harmful command changes under uncertain weather and hardware constraints. Conventional controllers, such as PID and fuzzy logic, remain attractive because they are simple and computationally inexpensive, but they can respond poorly to stochastic disturbances and slow thermal inertia. Learning-based control can be more adaptive, but its use in agricultural cyber-physical systems is limited by training instability, weak safety guarantees, and mechanically aggressive actuation. This study presents GreenGuard, a neuro-symbolic control architecture that combines an edge-deployable language model, Qwen2.5-1.5B-Instruct, with deterministic runtime safety rules and actuator slew-rate limiting. In the proposed framework, the language model generates structured control proposals from filtered state summaries, while deterministic supervisory layers enforce operating bounds and mechanically feasible command transitions. We evaluate the method in a disturbance-driven thermodynamic greenhouse simulation using historical weather inputs from Dhaka, Bangladesh, with actuator saturation and sensor noise included to improve deployment realism. To make the evaluation more reliable and practically meaningful, we use 20 random seeds, include a MPC-style baseline, report paired Wilcoxon signed-rank tests with Holm correction, and quantify actuator movement using Total Variation. These additions help evaluate not only tracking accuracy and energy use, but also controller stability, repeatability, and potential actuator stress. Across 20 random seeds, GreenGuard achieved an RMSE of 4.35 ± 0.14 ∘C, compared with 12.42 ± 0.11 ∘C for PID, 8.80 ± 0.93 ∘C for PPO, and 5.26 ± 0.00 ∘C for the lightweight MPC-style baseline. The controller also reduced the energy-cost proxy by 45.4% relative to PPO, 58.7% relative to PID, and 16.0% relative to the MPC-style baseline while producing fewer temperature-bound violations. These findings suggest that tightly constrained LLM-based control may provide a practical middle ground between reactive heuristics and unconstrained learning-based control for slow-dynamics agricultural environments. The results should be interpreted within the limits of simulation-based evaluation.