Treat invalid SWE-Lancer pytest runs as rollout errors

[sylvesterkaczmarek]

Summary

• require at least one real pytest pass/fail result before scoring an IC SWE-Lancer task
• treat a retry set containing only the -1.0 missing-status sentinel as a grading/infrastructure failure
• preserve ordinary failing test runs (0.0) as valid model failures

Previously, repeated missing pytest_exit_code values left test_results non-empty, so the evaluator returned score=0.0 and reported that tests ran smoothly. The validity check now accepts only run sets containing at least one real pass/fail status.

The code change is intentionally limited to the existing post-retry guard.

Fixes #151.