Gearbox vibration responses vary with speed and load, posing a challenge to reliable cross-condition fault diagnosis. This paper proposes Balanced Fusion, a framework that integrates per-sensor vibration features with pairwise sensor coherence under an unseen-target cross-condition protocol. In the Node branch, a weight-shared temporal convolutional encoder extracts temporal features from each sensor, while a compact spectral representation combines six global spectral descriptors with four normalized band-energy ratios. The Edge branch represents inter-sensor relationships using four-band magnitude-squared coherence (MSC) features computed for every physically indexed sensor pair. By normalizing the squared magnitude of the cross-spectrum by the product of the corresponding autospectra, MSC emphasizes frequency-specific inter-sensor dependence rather than absolute vibration amplitudes. At the decision level, a two-stage fusion scheme using equal weights fixed in advance first averages the temporal and spectral class-probability vectors to obtain the Node probabilities and then averages the Node and Edge probabilities to produce the final prediction. The target condition is reserved for final testing; model training and checkpoint selection use only source-domain data. Averaged over the two transfer directions, Balanced Fusion achieves Macro-F1 scores of 99.86% and 78.37% on XJTU_Spurgear and SEU Gearbox, respectively. On the more challenging SEU Gearbox dataset, the final fusion increases Macro-F1 from 69.08% for the Edge branch to 78.37%, a gain of 9.29 percentage points. Sensor-pair identity ablations show that fusion performance depends on preserving pair identity and sample-specific coherence rather than on feature dimensionality alone, while prediction-level analysis reveals complementary errors between the Node and Edge branches.