Knowledge-aware recommendation (KR) systems can enhance recommendation performance by utilizing factual knowledge in knowledge graphs (KGs). Existing methods typically adopt graph connectivity as the only propagation principle, spreading all factual knowledge into the recommendation process. However, this injects non-preferred facts into recommendation signals, driving the results away from user preferences and severely limiting knowledge-aware recommendation performance. To address this, we propose a knowledge-aware recommendation framework called PreFact, which introduces fact-preference relevance as an additional propagation principle to regulate knowledge propagation toward preferred facts. Specifically, to evaluate the fact-preference relevance, we design a relevance evaluation mixture-of-experts that simultaneously specializes in fact semantic encoding and user preference mining, which establishes cross-space associations in a heterogeneity-coordinated manner. To propagate knowledge around preferred facts, we design a self-adaptive switch network that continuously maintains a relevance-proportionate knowledge propagation flow and delicately cuts off the spread of non-preferred facts, which integrates factual knowledge into recommendation in a preferred-fact-centric manner. Extensive experiments on three public datasets demonstrate that PreFact consistently surpasses state-of-the-art models. The implementation of PreFact is available at https://github.com/fonzoin/PreFact .