@dchud I had the same idea and published rmarc (claude wrote it, main difference was I decided to be more aggressively backwards compatible, i.e. only performance, no particular improvements).
When I was deep in using marc, my pain points
-speed
-needing chardet to run on each field because the declared encoding was a lie
-fields shifts, where a indicator or subfield was omitted and the first letter of the value shifted into the subfield.
-just a complete lack of a schema for the binary format