This study evaluates a compact dual-branch network for multi-sensor machinery fault diagnosis under additive noise. Three waveform descriptors produce sample-specific channel weights; robust time-frequency and finite structured state-space encoders are then combined by a descriptor-conditioned gate. Under the original within-distribution protocol, mean noisy accuracy is 93.58±1.40% on HIT and 86.99±2.14% on WT, below the strongest baseline on both datasets. New WT ablation, matched-parameter Conv1D controls, representation analysis, and unseen-SNR tests show that adaptive channel weighting is beneficial, whereas the state-space branch and learned gate do not yield stable cross-dataset gains. Anti-aliased HIT preprocessing raises full-model accuracy to 97.31±0.99%, while record-disjoint WT evaluation reduces it to 55.79±2.48%. The results identify the supported components and delimit the conditions under which the architecture remains reliable.