Graphs are pervasive across diverse domains, requiring a Graph Foundation Model (GFM) pre-trained to capture generalizable knowledge across heterogeneous domains. However, multi-domain pre-training poses two challenges: (i) domain discrepancy—as graphs from different domains possess distinct feature dimensions and semantics—and (ii) adaptation difficulty to unseen target domains with limited supervision. Many prior studies relied on PCA-based preprocessing to unify multi-domain feature dimensions. However, such static feature alignment risks the loss of information essential for the model optimization and the acquisition of generalizable knowledge. Moreover, since it does not exploit structural information and inter-domain relationships, it often yields suboptimal alignment. We propose Dynamic Domain-aware Feature Alignment (DDFA), a novel framework that jointly considers intra-domain dimension alignment and inter-domain semantic alignment. We also introduce WB regularization to reduce domain discrepancy efficiently, thus guaranteeing upper-bound of generalization error. In the adaptation stage, we design a target-specific dual-space dimension encoder with a novel node sampling strategy and a prompting method to fully exploit limited target supervision and pre-trained knowledge. Extensive experiments on seven datasets demonstrate the superior performance on most datasets under few-shot node classification scenarios.