Multimodal recommendation leverages auxiliary signals such as images and text to enhance personalization. However, two key challenges remain: semantic noise from auxiliary modalities, and the difficulty of modeling both long sequences efficiently and short or sparse behaviors robustly. We present CAMeRA, a structure-aware framework that repositions ID embeddings as semantic anchors to guide multimodal fusion. It introduces an ID-Guided Dual Modulation mechanism that aligns visual and textual features with user-specific behaviors while suppressing irrelevant noise. To preserve structural integrity, CAMeRA adopts a dual-phase optimization strategy that decouples ID-based sequential modeling from multimodal enhancement. Furthermore, we design a Context-Aware Adaptive Mamba (CA-Mamba) encoder, which extends Mamba with bidirectional context modeling and a GRU-based temporal compensation unit to address sparsity and instability in long-tail scenarios. Experiments on four Amazon benchmarks demonstrate that CAMeRA consistently outperforms strong baselines in both accuracy and robustness across diverse user behaviors.