[framework, tasks] feat: retain runner scoring diagnostics in trajectories

[tongyx361]

Summary

Preserve runner scoring evidence in trajectory output so task failures can be diagnosed independently of the final training reward.

Changes

• Framework writes the runner reward, accuracy, and task context to extra_fields.runner_reward_info.
• SWE-bench retains evaluator exit codes, per-test results, and explicit agent errors.
• Document the added fields and cover TransferQueue output, training masks, and SWE resolution behavior.
• Attach the test log-capture handler to Uni-Agent's non-propagating namespace so existing logging assertions capture the intended records.

Validation

• PYTHONPATH=.:verl uv run --no-project --python /usr/bin/python python -m pytest -q tests/uni_agent/tasks/test_swe_agent_error_on_cpu.py tests/uni_agent/tasks/test_swe_reward_on_cpu.py tests/uni_agent/framework/test_generate_sequences_on_cpu.py: 75 passed. Tests used the runtime interpreter with Uni-Agent and the pinned public verl gitlink explicitly selected.
• pre-commit run --all-files --show-diff-on-failure: passed.

Internal validation and observed benefit

A historical internal GPU smoke evaluated four tasks across four model-version evaluations, producing 16 completed sessions. Persisted results contained runner reward, evaluator exit code and parsed test-status evidence for 16/16 sessions, and runner rewards matched the corresponding evaluation results in 16/16 cases. This made the task-side scoring evidence inspectable alongside the final result.

This is diagnostic coverage from an integrated implementation of the feature, not a before/after reward-quality comparison or an end-to-end validation of this split PR's exact head. No accuracy or throughput gain is claimed. The isolated PR regression tests are listed above.

Compatibility

The diagnostic fields are additive. Reward-worker scores, training masks, and SWE-bench resolution criteria retain their existing behavior. No migration is needed. This is a standalone observability improvement without a linked issue.

Checklist

• The PR is focused and explains why no issue is needed.
• The title follows the owning-layer format.
• Behavior is covered by tests; skipped validation is stated above.
• User-facing configuration and behavior changes are documented where applicable.
• Compatibility and migration requirements are documented.
• Logs, fixtures, and examples contain no credentials or private data.
• pre-commit run --all-files --show-diff-on-failure passes.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!