[examples] fix: use task-level swe-bench reward

[zackcxb]

What does this PR do?

Fix the Claude Code blackbox recipe after the runtime/task refactor in #83.
The recipe still loaded the removed uni_agent.reward registry, so in-sandbox
SWE-bench evaluation failed on current main. This PR migrates the recipe to
the task-level uni_agent.tasks.swe_bench.reward.compute_reward API and adds
the exec_shell adapter required by that evaluator.

Checklist Before Starting

• Searched for related changes: [BREAKING][core] refactor: unify agent runtime and training interfaces #83 introduced the task-level reward API and removed the legacy registry.
• Formatted the PR title as [examples] fix: migrate Claude Code reward to task API.

Test

• Focused development regression probe: reproduced the missing uni_agent.reward import and missing exec_shell, then passed all 4 reward-adapter cases after the fix. The probe is not committed because this repository treats executable recipes as integration coverage and does not run recipe pytest suites in CI.
• python -m pytest tests/uni_agent/agents/test_claude_code_agent.py -q — 8 passed.
• pre-commit run --all-files --show-diff-on-failure --color=always — all hooks passed.
• Integration evidence from the earlier rollback validation rollout, with this reward compatibility patch applied: 32 sessions produced 50 trajectories; all 50 reward records reported eval_completed=True, with no reward-path crash.

API and Usage Example

No new user-facing configuration is required. The existing Claude Code
run_infer.sh and training recipe now evaluate SWE-bench through the post-#83
task reward API.

result = await compute_reward(metadata, sandbox, eval_timeout=600)
score = 1.0 if result.get("resolved", False) else 0.0

Design & Code Changes

• Remove the obsolete recipe-local reward registry lookup.
• Lazily call the canonical task-level SWE-bench reward function.
• Reject unsupported reward data sources explicitly; this recipe is SWE-bench-specific.
• Add SandboxEnvForReward.exec_shell, preserving quoted working directories and timeouts.

Checklist Before Submitting

• Read the Contribute Guide and PR template requirements.
• Ran full-repository pre-commit checks.
• Explained why recipe integration evidence is used instead of adding a new pytest file.
• Confirmed the PR title matches the required format.
• Replaced all PR template placeholder text.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.