fix: clean OpenClaw template files from judge workspace to prevent grading failures

[JackyCSer]

Problem

When openclaw agents add creates the judge agent, it scaffolds AGENTS.md, SOUL.md, BOOTSTRAP.md and other template files into the workspace. These files instruct the agent to perform a bootstrap/personality flow (read SOUL.md, introduce itself, etc.) before processing any prompt.

This causes the judge agent to ignore the grading prompt entirely and output bootstrap text like I'm Claw 🦾, bootstrap complete... instead of the expected JSON scores. The JSON parser then fails with Failed to parse judge JSON response, giving 0 scores to all llm_judge and hybrid tasks — regardless of how well the tested model actually performed.

Impact

In our testing with doubao-seed-2.0-pro:

• 13 out of 23 tasks were affected (7 pure llm_judge + 6 hybrid)
• Reported score: 40% (9.19/23)
• Estimated actual score: 70%+ (after excluding judge failures)
• All affected tasks had breakdown: {} and notes: "" — the judge never evaluated them

Root Cause

openclaw agents add judge-agent
→ scaffolds AGENTS.md, SOUL.md, BOOTSTRAP.md into workspace
→ AGENTS.md says: "Before doing anything, read SOUL.md, USER.md..."
→ Judge receives grading prompt but executes bootstrap flow instead
→ Outputs personality text, not JSON
→ _parse_judge_response finds no JSON → returns {}
→ score = 0.0 for all judge-graded tasks

Note: prepare_task_workspace (for the tested agent) does shutil.rmtree before each task, so the tested agent is unaffected. Only the judge workspace persists these template files.

Fix

Add _clean_judge_workspace() that removes OpenClaw-scaffolded template files (AGENTS.md, SOUL.md, BOOTSTRAP.md, etc.) after agent creation, ensuring the judge operates as a pure grading function.

The cleanup runs on every _ensure_judge_agent call (idempotent), so it also handles cases where the agent already exists but the workspace was previously polluted.

Testing

• Verified that after cleaning template files mid-run, subsequent llm_judge tasks received proper JSON scores
• The fix is backward-compatible — if the workspace is already clean, it's a no-op