ÞÝÐING Revisited: Empirical Limits of Translation-Based Prompt Injection Defense and the Deobfuscation Effect | Synapse