Senior-level debugging specifically for scientific code failure modes. Covers CUDA out-of-memory errors, numerical instability (NaN/Inf propagation), reproducibility failures, multiprocessing/GIL issues, segmentation faults in C extensions (pydicom, SimpleITK, ITK), race conditions in dataloaders, silent correctness bugs, HPC/SLURM job failures, and the methodology of debugging "wrong but plausible" results. Use when a pipeline fails, gives unexpected results, or differs from a previous run.
