Why Synthetic Defects Can Validate but Must Not Train

A junior teammate proposes: "We have almost no real defect images. Let's generate thousands of synthetic defective images with cut-paste and generative edits, label them 'defective,' and train a supervised binary classifier (normal vs. defective) directly on real normal images plus this synthetic defective set — it solves the data scarcity problem and lets us skip the one-class anomaly detection machinery entirely."

1. What goes wrong with this proposal, concretely — what will the resulting classifier actually learn to detect?
2. Contrast this with how this case actually uses synthetic defects, and explain why that use doesn't have the same failure mode.
3. This case's synthetic-defect vocabulary is cross-referenced from case-study-synthetic-data-generation-for-computer-vision. Which concept from that case applies almost directly here, and how?