University counseling centers increasingly face high demand, limited service hours, lengthy waitlists, and constrained staffing capacity. This study presents a safety-aware psychoeducation chatbot designed to complement, rather than replace, professional counseling services by providing evidence-based mental health information and guided self-reflection outside regular service hours. The deployed system runs the base Qwen3-8B model with retrieval-augmented generation over 17 curated public-health documents from sources including NIMH and WHO, and a four-tier safety-routing layer for general questions, sensitive topics, out-of-scope requests, and crisis disclosures. High-risk messages are routed to UAE-specific emergency and crisis resources before language-model generation is invoked. QLoRA fine-tuning on 750 filtered CounselChat examples was evaluated and found to degrade response quality relative to the prompted base model, an ablation finding that informed the deployed configuration. Safety routing is evaluated on a 181-case adversarial benchmark with per-tier metrics, developed with independent labeling by university counseling staff (inter-rater agreement κ = 0.80). The results quantify the failure modes of lexical crisis detection, particularly for indirect ideation, and highlight the importance of deterministic safety controls in mental health chatbot design.