This paper addresses the HEXACO personality trait regression task in Asynchronous Video Interviews (AVI) for Track 1 of the ACM Multimedia AVI Grand Challenge 2025, and proposes a lightweight yet robust scattering-augmented multimodal prediction framework termed SA-RADE. We first employ SigLIP2, emotion2vec+, and SFR-Embedding-Mistral to extract visual semantic, affective acoustic, and textual semantic representations, respectively, followed by question-level temporal pooling. We then introduce the Wavelet Scattering Transform (WST) as a complementary visual stream, computing high-frequency facial texture statistics–mean, standard deviation, and salient energy–over five semantic facial regions to capture micro-expressions and subtle deformations that are informative for personality inference. To effectively fuse heterogeneous cues, we design a Residual-Adapter Deep Ensemble (RADE) regressor, which improves training stability and generalization by averaging multiple MLP sub-models and learning complementary cross-modal residuals via a gated adapter branch. Experiments on the official AVI-2025 dataset show that SA-RADE delivers stable performance across four HEXACO traits, achieving a markedly lower average MSE than baseline methods and yielding strong overall results, thereby demonstrating the effectiveness of scattering-based texture cues and lightweight ensemble learning for AVI personality computing.