fix: wrap toxigen samples_start past dataset end instead of crashing

The auto-offset progressed past the 50-question dataset, causing
'no items in range' → RuntimeError. Now wraps modulo dataset size
so the farm loops continuously through the question set.

Co-Authored-By: Virgil <virgil@lethean.io>

eval.py

@@ -300,10 +300,15 @@ def _run_generative_rounds(model_name, task, n_questions, rounds, samples_start=
     system_prompt = GENERATIVE_SYSTEM_PROMPTS[task]
     items = _load_bench_items(task)
 
-    window = items[samples_start:samples_start + n_questions]
-    if not window:
-        print(f"  WARNING: no items in range [{samples_start}, {samples_start + n_questions})")
+    total_items = len(items)
+    if total_items == 0:
+        print(f"  WARNING: benchmark has no items")
         return []
+    wrapped_start = samples_start % total_items
+    window = items[wrapped_start:wrapped_start + n_questions]
+    if not window:
+        window = items[0:n_questions]
+    print(f"  samples [{wrapped_start}, {wrapped_start + len(window)}) of {total_items} (canon offset {samples_start})")
 
     all_rounds = []
     for r in range(1, rounds + 1):