The 5% threshold is the key insight here. CIOH doesn't fail gradually — it fails catastrophically once the uncertainty is high enough that an analyst can't trust it. You don't need majority adoption; you need enough Payjoin transactions that any given transaction *might* be one. That's a very achievable attack on the heuristic.

It's similar to how a small percentage of privacy-preserving transactions in a pool doesn't just protect those transactions — it poisons the inference for all of them.