Just thinking out loud and assuming you only have this extension and want to do a quick poc: you could try to delegate the preliminary line parsing to awk (which is very powerful and works in a streaming fashion).

CNAB file -> awk -> select from line_read().

Awk could, for example, transform the fixed-width file in a delimited one (I believe this extension does not strictly require json-formatted inputs). You vould also perform pre-filtering of the input data on awk side (see the "/ /" syntax).

If this extension supports reading from stdin, you could build a traditional unix command pipe, otherwise you would have to glue together the two commands with a named pipe.

Depending on your sensibility, this could feel ugly or very elegant. I am confident it would be very fast, not memory bound, and can be made strict and correct.