I think external referencing is the best approach for using binary data with XML. The CDATA tag doesn't live up to expectations and base 64 encoding just makes things 33% bigger.
Just as with texture image data a URL to the file is preferred over embedding in COLLADA documents. I envision that large chunks of binary data can be in well known formats just as image data is stored in .png, .tif, .jpg, .bmp and so forth. None of those binary formats are within the scope of the COLLADA specification. I think everyone understands that that is not necessary.
In COLLADA, a URL from an <accessor> element expects to parse an <array> element. The <array> element has the meta data to describe the block of data. If that URL refers to an external, binary file we can expect the file extension to indicate the format of the data, just as with imagery. If we want to read .png data then everyone needs a library of code for that. It's the same thing with external data... if the data is in .xls format, for example, then we need those routines etc.
We can use the <array> schema definition to create a binary form of that storage. In this case then I agree that COLLADA should define this schema. I think that is what some of you have been saying while ignoring the more general well known binary format use-case that is prevalent with image data.