The breadcrumb approach right now is simple invisible markers, not paraphrase-resistant watermarking. You're right that semantic watermarking that survives LLM rephrasing is the harder and more interesting problem. It's on the radar but not in scope for v1.