<p>The proliferation of rumor online has heightened the focus on rumor detection in recent times. Advances in automatic rumor detection have been substantial, yet the task of how to pinpoint rumors in low-resource cross-lingual/domains settings remains formidable. Researchers introduce zero-shot learning based on fine-tuning or prompt such as RPL, as potential remedies to these issues, but they rely on the discrete verbalizer and the virtual response to map encoder outputs to class labels. Such label-word choices are often ambiguous and language-dependent, leading to fragile decision boundaries and poor calibration under cross-lingual/domain shifts. To address these limitations, we propose PARD, a few-shot prototype augmentation framework for low-resource rumor detection. We reuse hierarchical prompt encoding to obtain language/domain-agnostic thread representations, but instead of zero-shot verbalization we adopt an N-way K-shot metric learning protocol where the few labeled target supports directly form class prototypes. Moreover, a prototype augmentation module with query-conditioned sample-level attention and feature-level gating further refines these prototypes and the distance metric to cope with noisy supports and sparse features. Extensive experiments conducted on four groups of cross-language/domain rumor datasets demonstrate that our proposed model achieves state-of-the-art performance and showcases advanced capabilities for early detection. In addition, real-world cross-domain rumor detection systems must handle high-concurrency streams of multilingual posts under strict latency constraints. Such platforms are typically deployed on high-performance computing infrastructures with parallel processing. Our few-shot learning paradigm and prototype augmentation operate through fully vectorized operations over a small set of prototypes, making the model amenable to efficient parallelization in these high-performance environments.</p>