Expand deterministic danger coverage

[LE0-Lin]

Summary

• detect four narrowly defined high-risk command patterns: PowerShell download-and-execute, forced Git directory cleanup, aggressive Docker system prune, and Kubernetes namespace deletion
• add nearby safe forms and explicit prohibitions as negative controls
• expand Adversarial Benchmark v1 from 74 to 122 deterministic contracts across six instruction formats
• keep the four unresolved semantic challenges visible

Verification

• PYTHONPATH=src python -m unittest discover -s tests -v (126 passed)
• PYTHONPATH=src python -m agent_config_score.cli examples/demo`n- PYTHONPATH=src python -m agent_config_score.cli . (A 100)
• PYTHONPATH=src python scripts/run_adversarial_benchmark.py (122/122 contracts, 50/50 clean controls, 4/8 challenges detected)
• PYTHONPATH=src python scripts/run_real_world_benchmark.py (all three pinned repositories matched reviewed expectations)