Fix cost attribution edge cases

[orban]

Summary

• preserve skill-generation cost attribution on repo-task timeout and empty-run returns
• recompute resumed multi-run cost totals from serialized instead of median
• add regression coverage for both failure-path propagation and resume summary aggregation

Testing

• pytest eval-harness/tests/test_reporter.py eval-harness/tests/test_task_runner.py eval-harness/tests/test_resume.py eval-harness/tests/test_agentbench.py -q

[orban]

Superseded by #31, which includes the edge-case fixes plus the resumed multi-run cost median fix. Closing to consolidate review.