In the 32000 PDF spec, in "14.3.2 Metadata Streams" it is almost suggested not to compress metadata, because
PDF documents are often made available on the Web or in other environments, where many tools routinely examine, catalogue, and classify documents. These tools should be able to understand the self-contained description of the document even if they do not understand PDF.
The implementation of PDMetadata produces an uncompressed metadata, but PDFMergerUtility saves compressed metadata. This breaks a required PDF/A-1b property. (happened at work today)