Reliable corrosion monitoring becomes more difficult when a diagnostic model is expected to remain effective beyond the experiment used for its development. This study combines Lamb-wave measurements with machine learning to assess corrosion in 316L stainless steel, with particular emphasis on transfer between independent experiments. A public laboratory dataset containing two corrosion experiments was analyzed, comprising 429 corrosion states and 12 directed piezoelectric actuator-sensor paths per state. Damage indices, baseline-relative descriptors, and frequency-domain features were evaluated using conventional machine-learning models and a physics-guided path-attention multi-task learning framework (PG-PAMTL). Performance was assessed under within-experiment validation, strict zero-shot transfer, and unsupervised domain adaptation. In the strict cross-experiment setting, PG-PAMTL achieved a macro ROC-AUC of 0.830 and a PR-AUC of 0.871, indicating substantially stronger transferability than the conventional benchmarks. Domain adaptation increased the classification ROC-AUC only marginally to 0.832, but its effect was more evident for severity estimation, where the macro MAE decreased from 39.3 to 32.5 µm. Permutation analysis showed that predictive information was concentrated in a limited number of sensing paths and in both waveform-change and frequency-domain features. A separate sensitivity analysis also confirmed that a known bubble-related disturbance affected transfer performance. Experimental progression, however, was strongly associated with corrosion severity, limiting the interpretation of the regression results. The findings therefore support Lamb-wave machine learning for cross-experiment corrosion detection, while showing that reliable absolute severity estimation still requires validation under more independent experimental conditions.