fix: tighten convergence threshold to prevent false-positive early stop

Changed threshold from -500.0 to -500.5. The fitness returns
-500.0 - (100/nll) on full match, so real matches always score < -500.0.
But partial matches with max keyword+substring bonuses (500 total) could
theoretically reach -500.0 with very low NLL, causing premature stopping.

attacks/method2_cma_es.py

@@ -150,7 +150,7 @@ class CMAESTokenOptimizer:
             es.tell(solutions, fitnesses)
             print(f"[Gen {gen}] fitness: {best_loss:.4f} | cache_hits: {cache_hits} | trigger: {repr(best_trigger_text)}")
 
-            if best_loss <= -500.0:
+            if best_loss <= -500.5:
                 print("[!] Attack converged - full target string found in output!")
                 break