fix(api/task): guard empty results list in process_results

[Luodian]

Summary

For generate_until / generate_visual_cot tasks, when a sample's generation failed upstream (model raised, retries exhausted, returned []), the results list handed to ConfigurableTask.process_results can be empty. The existing isinstance check then falls through to results[0] on the list-of-list branch and raises IndexError — which aborts the whole postprocess loop for that task, not just the missing sample.

Change

Add a leading results and so an empty list routes to the else branch (which does [res.strip() for res in results] and produces []). Downstream task-level process_results then receives an empty list and can decide how to score the missing sample without taking down the whole task.

if self.OUTPUT_TYPE in ("generate_until", "generate_visual_cot"):
-   if isinstance(results, list) and isinstance(results[0], list):
+   # Guard empty results so results[0] below does not IndexError for
+   # samples whose generation failed. Downstream process_results then
+   # receives an empty list and can decide how to score the miss.
+   if results and isinstance(results, list) and isinstance(results[0], list):
        results = [res.strip() for res in results[0]]
    else:
        results = [res.strip() for res in results]

Test plan

• Any task run where at least one generation produces [] — verify the task finishes and the missing sample is scored by downstream process_results.
• Happy-path run — verify no change in scores.