I did the following experiment: using the SinkTestDocument that I attached at DOXIA-101 I generated two text documents, one by dumping the model directly into a text sink, the other by piping it through the current apt sink, parsing the result with the apt parser and dumping it into the same text sink as before. The results should be the same since the second chain corresponds to the 'identity transformation', ie piping a document through a parser and sink should give you the original document. I attach the two text files for comparison, here are the differences:

1. the parser swallows newlines between text elements
2. a paragraph within a list item is swallowed
3. verbatim text within a definition list item is not processed correctly
4. the closing of a definition list is not processed correctly
5. table header cells are not recognized and newlines within table cells are not processed correctly

Point 1 is not severe by itself because newlines are not significant in apt source documents, however, two newlines are, so I am not sure if it doesn't have consequences (eg within table cells).

Point 5 has partially been fixed by the patch Vincent attached at DOXIA-50.