Cross-lingual Speech Emotion Recognition (SER) presents a formidable challenge in affective computing, largely due to the complex entanglement of linguistic content with paralinguistic emotional cues. Traditional domain adaptation techniques frequently fail to differentiate language-specific phonetic variations from distinct prosodic patterns. This issue is particularly acute in transfer tasks between non-tonal languages like English and tonal languages like Mandarin, where lexical tones often confound emotional pitch. To bridge this gap, we introduce the Contrastive Prosody-Phoneme Disentanglement Network (CPPD-Net), a novel framework that systematically decouples content from style. CPPD-Net utilizes a Hierarchical Dual-Stream architecture, integrating an Adversarial Style Erasure (ASE) mechanism to eliminate domain-specific language attributes from the content representation. Additionally, we employ a contrastive InfoNCE loss to maximize the mutual information shared between emotion-dependent views. Validated on the large-scale MSP-Podcast and BIIC-Podcast corpora, CPPD-Net achieves an Unweighted Average Recall (UAR) of 65.18% on the English-to-Mandarin transfer task, significantly surpassing the state-of-the-art SAPA method (59.25%). This work establishes a new benchmark for robust, language-invariant emotion recognition.