This study tests whether defective Metal Active Gas (MAG) surfacing weld beads form distinct mechanism-specific clusters in multimodal sensor space, with implications for Machine Learning (ML) pipeline design. A synchronized acquisition platform integrating current and voltage signals at 10 kS/s, airborne acoustic sensing at 48-192 kHz, and high-resolution post-bead imaging was developed and validated, then used in a three-phase campaign yielding 105 weld beads on S355 steel. Current-based descriptors strongly separate acceptable and defective beads, with Cohen's |d| = 1.34 for the current coefficient of variation and 1.42 for the spike rate. The arc-stability response to voltage trim is strongly asymmetric, with current variability changing by a factor of 6.6 between trim = −3 V and +3 V. Defective beads form two distinct clusters in PCA and t-SNE projections, corresponding to electrical instability and insufficient material deposition. A baseline ML evaluation across five classifiers reached macro-F1 = 0.735 for the fused configuration. The findings provide quantitative pilot-scale evidence that defective weld beads do not form a homogeneous class, that the borderline class is similarly heterogeneous, and that supervised-learning pipelines for weld-bead quality assessment should be designed with mechanism-aware labels and sensor coverage rather than aggregated quality categories. The reported baseline macro-F1 = 0.735 should therefore be read as evidence that the extracted descriptors carry useful information for downstream supervised learning, not as a final model ranking.