feat: LLM-judge prompt safety scan for BenchJack defense (Closes #1808)

[Lexus2016]

Automated evolution PR for issue #1808.

Summary

Static scan over LLM-judge prompt templates in the evolution pipeline for un-delimited agent-content interpolation. Detects the CAR-bench failure mode where agent output is interpolated into judge prompts as instructions rather than data.

Changes

• scripts/evolution_judge_safety_scan.py (new): scans judge template scripts for agent-content variables (pr_body, implementation, response, etc.) interpolated into prompt context without data delimiters (..., backticks).
• tests/scripts/test_evolution_judge_safety_scan.py (new): 7 tests covering unsafe detection, safe-delimited pass, clean code, edge cases, report formatting, and CLI.

Verification

• Lint ✓, format ✓
• All 7 tests pass
• Diff: 199 lines (within self-merge cap)

Closes #1808

[github-actions]

૮ >ﻌ< ა ci review

_{ran on 0519830}

all good!

[Lexus2016]

Blocked by code-review: dead code. scripts/evolution_judge_safety_scan.py is a standalone CLI scanner that no CI workflow, script, or production code invokes. It's only exercised by its own tests. The issue #1808 requires CI integration — a scanner that must be manually invoked provides no automated defense. PR #1816 (evolution/issue-1807-1808-ci-wiring) is the correct implementation that wires both this scan and the isolation check into lint.yml as a CI job. Closing as superseded by #1816's wired approach.