Detect agent tool abuse from call sequence, permission jump, output intent, retry pattern, and data movement