We present AURA-ST, a three-stage modular pipeline for low-resource speech-to-text translation submitted to the IWSLT 2026 African-Celtic Track 1. The architecture bypasses traditional cross-attention between audio and text modalities by treating projected acoustic representations as a native token prefix to a frozen large language model. A dual-stream encoder captures linguistic and paralinguistic features via a jointly trained semantic and a paralinguistic encoder. A convolutional subsampler then bridges the modality gap through a 4x temporal compression and a linear projection into the LLM embedding space. Finally, a MLP-targeted Low-Rank Adaptation adapter fine-tunes the frozen Gemma-4-E2B backbone for translation without catastrophic forgetting of base language model knowledge. We further identify and resolve the incompatibility between standard PEFT attention-level adapter injection and the Gemma-4 Per-Layer Embedding architecture that tends to cause gradient isolation. Trained on the IWSLT 2026 Track 1 data covering Hausa, Igbo, and Yoruba, the final system achieves a best proxy teacher-forced SacreBLEU of 91.29 on the validation set at Phase 3, with Phase 1 speech encoder validation loss converging to 0.651.