Thanks for the Internet Archive pointer. Hadn't thought of it (probably
because of a few past unsuccessful attempts to find archived pages.)
Tried BadgerFish (
http://libx.lib.vt.edu/services/code4lib/lccnrelay3/2004022563 which proxies
lccn.loc.gov's marcxml) and it meets the requirements of faithful
reproduction of the XML, albeit in a very verbose way that doesn't attempt
to do any minimization.
That leaves, indeed, two independent problems:
a) a free converter to GData's JSON format or another less redundant
convention than badgerfish. Looking at this for 1 second, I'm wondering if
this is even possible to implement without knowing the schema of the XML
document. It says, for instance, to use arrays  for elements that may
occur more than once.
b) something MARC-specific to express MARC records in JSON. I talked to
Nathan Trail from LOC at code4lib, and they're revamping their lccn server
this year to scale up and also serve more formats. Presumably, this effort
could lead to a de-facto standard of how to serve MARC in JSON.
Thinking out loud about this for a minute, I'm wondering if part a) is
really a worthwhile goal. Aside from the impromptu prototyping of an
XML-to-JSON gateway, I don't see any production use for a XML to JSON
converter that is agnostic to the schema; for performance reasons alone.