feat/Add OpenReward hosted env execution path-- OpenRewards PR 3/3

[AmyTao]

Summary

This draft PR brings the OpenReward hosted-environment work into the v0.5-integration line.

It was prepared from the local port-pr-684-to-v05 branch, using v0.5-integration as the integration base and cherry-picking/adapting the relevant OpenReward hosted-env changes from release/v0.6.0.

What changed

• Added OpenReward as a first-class hosted environment execution path.
• Added OpenReward metadata/model-policy wiring for OpenAI Responses tool calls.
• Added OpenReward rollout recording so runs appear in the OpenReward runs UI.
• Added local rollout artifact metadata at artifacts/openreward_rollout.json.
• Added hosted-env reward event helpers and ORS/OpenReward adapter plumbing.
• Added trajectory redaction coverage and hosted-env artifact contract tests.
• Added the openreward dependency and updated the lockfile.

Validation

• uv run pytest tests/test_openreward_env.py -q
  • 16 passed
• uv run pytest tests/test_openreward_env.py tests/test_hosted_env.py tests/test_hosted_env_rollout_contract.py tests/test_adapters.py tests/test_trajectory_redaction.py -q
  • 74 passed
• uv run ruff check src/benchflow/openreward_env.py tests/test_openreward_env.py
  • All checks passed!
• End-to-end OpenReward smoke test:

uv run bench eval create \
  --source-env openreward:GeneralReasoning/CTF \
  --source-env-arg split=train \
  --source-env-arg index=0 \
  --model gpt-5.4-mini \
  --jobs-dir /private/tmp/benchflow-openreward-pr3

Result: