[framework, docs] feat: forward TaskResult.extra_info metadata

[qiangyupei]

Summary

Forward compact, use-case-specific TaskResult.extra_info fields through the generic Task Runner into a Gateway session's reward_info. This lets verifier metrics reach finalized trajectory metadata without adding task knowledge to
Gateway or the Agent Framework.

TaskResult.extra_info already exists. This PR only defines how the generic Runner forwards it when the existing report_reward=true option is enabled.

Changes

• Merge JSON-serializable extra_info into session reward_info.
• Keep reward, acc, and finished reserved for canonical TaskResult fields; ignore colliding metadata keys.
• Omit all non-reserved metadata, with a warning, when it is not JSON serializable or exceeds 64 KiB after UTF-8 serialization. Canonical fields remain in the reward payload.
• Document the contract in Task and Reward, with a short link from the Gateway documentation.
• Add CPU tests for forwarding, reserved keys, invalid JSON, recursion, and the serialized-size boundary.

API

return TaskResult(
    reward=score,
    accuracy=pass_rate,
    finished=agent_result.finished,
    extra_info={
        "compile_ok": compile_ok,
        "passed_cases": passed_cases,
        "latency_ms": latency_ms,
    },
)

The 64 KiB limit is a Task Runner policy, not a Gateway, HTTP, Ray, or trajectory schema requirement. Session metadata is copied into finalized trajectories and training records, so large logs, source code, and evaluator artifacts should use artifact storage instead.

Compatibility

• report_reward=false preserves the original Runner behavior and does not forward extra_info.
• With report_reward=true and no extra_info, the canonical payload is unchanged.
• Invalid, colliding, or oversized metadata cannot replace or remove canonical reward fields.

Validation

Validated on Ubuntu with Python 3.11 against 28174fdab3787d307ae3a96d32d3737b600575a0.

python -m pytest -q \
  tests/uni_agent/framework/test_task_runner.py \
  tests/uni_agent/tasks/test_inference_task_routing.py

# 21 passed, 3 warnings in 9.89s

The focused tests cover successful metadata forwarding, reserved-key protection, invalid JSON values, serialization failures, and the 64 KiB serialized-size boundary. Existing Task Runner and Task routing behavior is also covered by the same test run.

Focused lint and format validation:

python -m ruff check \
  uni_agent/framework/task_runner.py \
  tests/uni_agent/framework/test_task_runner.py

python -m ruff format --check \
  uni_agent/framework/task_runner.py \
  tests/uni_agent/framework/test_task_runner.py

The focused CPU tests, Ruff check, and format check passed. The three warnings come from third-party OpenTelemetry, SciPy/NumPy, and Ray dependencies and do not affect the test result.

Checklist

• The PR is focused and explains why no issue is needed.
• The title follows the repository format and names the owning layers.
• Tests cover metadata forwarding, reserved-key protection, invalid values,
  serialization failures, and the serialized-size boundary.
• Existing behavior without TaskResult.extra_info remains covered.
• The metadata contract, compatibility behavior, and size-limit rationale
  are documented.
• Logs, fixtures, and examples contain no credentials or private data.
• The focused CPU tests pass (21 passed).
• The focused Ruff check and format check pass.