<p>Rumor propagation exhibits significant variations in patterns and user behaviors across different topics, which limits the generalizability of traditional detection methods and has motivated research on cross-topic rumor detection. Existing approaches primarily employ transfer learning or continual learning to facilitate knowledge transfer across topics, mitigating domain shifts and improving detection performance. However, these methods still face two main limitations: first, current cross-topic datasets are often constructed merely by aggregating data from multiple platforms, neglecting differences in data scale, platform sources, and topic-specific characteristics, which undermines the reliability of evaluation. Second, most studies lack a systematic analysis of feature importance across topics, resulting in limited model interpretability. To this end, we leverage a novel dataset from the Twitter platform, containing news cascades spanning seven topics. By explicitly analyzing various cascade-aware features and complementing them with implicit comparisons in the feature embedding space, we find substantial differences across topics in terms of root node influence and retweet time distributions. Based on these insights, we propose a two-stage cross-topic rumor detection model, TFTA. In the first stage, we design an unsupervised feature representation learning framework that combines contrastive learning with sparsification strategies to identify and highlight key features. The resulting cascade sense features are then projected into an embedding space to form an interpretable matrix, enhancing model interpretability. In the second stage, we quantify the importance of topic-specific and topic-agnostic features using the interpretable matrix and develop a topic-adaptive module to enable node-level cross-topic feature tuning. Furthermore, we introduce a time-aware graph attention mechanism with adversarial training to further optimize feature fusion. Extensive experiments demonstrate that TFTA significantly outperforms state-of-the-art methods in both cross-topic rumor detection and early detection tasks.</p>