I have seen your kind of setup before. The reasons may be different than your particular scenario, but nonetheless it made sense. One thing to consider has to do with the skills and tools people have available and so whatever makes the job done quick and well, goes.
You seem to describe an offset based, "flat" data structure. In my case, people used COBOL copybooks which are very good at describing this. IBM Rational Developer had a built-in wizard which allows the creation of Java Data Bindings from a COBOL copybook. This is to say that within a minute one gets a Java class which can create a record for your flat file in no time (it comes with all the logic required to do the padding, etc.)
To get the data generated, there are tools capable of generating XML files which cover all constraints defined by an XSD (e.g. alternate content i.e. xsd:choice, enumerated values, etc.) Now, assuming you have a proper XSD describing your logical model of your flat file, one can get 10s, 100s, even 100K XML generated from an XSD spec. It takes a click, plus the time spent by the tool to create those files.
Next, to get the XML files in your generated Java class, and so avoid going through XSLT or whatever (many shops don't have the skills) it may be as simple as writing Java mapping code between a JAXB generated class and the one created above, or if matching is possible, simply annotate the generated class to support JAXB unmarshalling. This last step may take longer to code, but it would be trivial code any Java developer would know how to do it.
This could possibly give you a view into why someone may have recommended Java and XSD for this task. XSD is a modeling language with built in support for constraints, which may prove helpful in generating test data through combinatorial techniques.