Nice work. A lightweight confusion matrix tool is surprisingly hard to find.

Technical question: did you consider support for class imbalance scenarios (e.g. thresholding or cost-weighted confusion matrices)? In many real datasets, F1/accuracy hide useful signal unless you vary the decision rule.

Clean implementation, and good call on keeping it simple.