And I don't think the performance issue exists.
Computers are fast nowadays. Parsing recfiles is straightforward.
Also you could easily archive historic/old/probably irrelevant records.
>>jdemle+(OP)
This is why I was very careful to say "well-defined subset". I wrote a full CSV library[1], and so I'm well aware of how deceptively difficult CSV is to deal with. However with a well-defined subset (and perhaps not using "," as a separator as well) it should be editable for at least simple changes.