Blockchain technology has transformed financial systems, but it also facilitates new types of fraud. Ponzi schemes constitute a persistent category of financial fraud that has manifested across multiple blockchain architectures, including both UTXO-based (e.g., Bitcoin) and account-based (e.g., Ethereum) systems. Accurately identifying such schemes remains a formidable challenge, particularly on Non-Turing-complete platforms where the lack of programmable expressiveness precludes in-depth smart contract analysis. Additionally, the inherent anonymity of participants, coupled with frequent address re-generation, conceals transactional linkages and significantly complicates pattern-oriented detection approaches. We present Ponzitracker, a unified framework for detecting Ponzi schemes via transaction graph analysis. Ponzitracker leverages the Dynamic Transaction Graph (DTG), a unified representation of account interactions that abstracts both UTXO-based and account-based blockchain paradigms, facilitating consistent fraud detection across heterogeneous blockchain environments. The DTG captures pyramid-shaped transaction structures observed in real-world Ponzi operations. Based on DTG, we present Hunter, a unified model that integrates structural and temporal features to effectively identify fraudulent entities. Evaluation on multiple real-world blockchain datasets shows that Ponzitracker achieves superior detection accuracy and efficiency compared to state-of-the-art baselines, demonstrating its effectiveness for large-scale, cross-platform Ponzi scheme analysis.