Provision Copilot CLI for code-review judge in Claude workflow

[gggdttt]

Problem

Running the code-review category through the Claude evaluation workflow fails in the scoring step with:

LLMJudgeError: Copilot CLI not found; cannot run the semantic judge

The code-review semantic judge (judge_verdicts in src/bcbench/evaluate/codereview_judge.py) always runs on Copilot CLI as a fixed, agent-independent judge. But claude-evaluation.yml only installs Claude Code — it never provisions Copilot CLI, the copilot-requests permission, or COPILOT_GITHUB_TOKEN — so judging dies for any Claude code-review run.

This is a pre-existing gap in the Claude workflow, unrelated to the BCQuality work in #696.

Fix

In .github/workflows/claude-evaluation.yml:

• Add copilot-requests: write to the evaluation job permissions.
• Install @github/copilot@1.0.57, gated on category == 'code-review' (skipped for bug-fix / test-generation to save time).
• Pass COPILOT_GITHUB_TOKEN: ${{ github.token }} to the run step and mask it.

This mirrors the judge environment already present in copilot-evaluation.yml.

Testing

Workflow-only change. To validate: run the Claude evaluation workflow with category: code-review (test-run) and confirm the judge step no longer errors.