> Based on the change of the Subject: line, you want to do it through a
> subprocess connector?  Sounds good.

Yeah, I automatically assumed that was meant. Is there another approach
that makes sense?

For JVM Schemes like Kawa, it would make more sense to do a Scheme
wrapper that runs in the same JVM, but to use JDBC from other Schemes a
subprocess written in Java and using only standard JVM features is lightest.

> Can the binary format be implemented by generic R6RS/R7RS-small
> implementations?

Yes, it is already implemented in portable R6RS, R7RS, and Common Lisp
and Python :) Source files binary.sls, binary.scm, binary.lisp,
binary.py in <https://github.com/lispunion/database-subprocess>.

> Are the hazards of text so great they outweigh the
> presumed advantages for debugging, i.e. using a serialization library
> which allows transparent use of either?

Yes. Binary is a lot simpler and faster when shipping blobs and
weirdly-encoded text strings around (both of which we'll be sure to get
from queries). And simpler/faster/less error-prone in practically all
other ways too.

The debuggability point you make is valid. That's why we're planning
with John to make dual text and binary formats that can encode the same
set of data. Then you can read the binary data from the subprocess into
Scheme and convert it into text at will to inspect and edit it. And you
can easily turn S-expressions to binary data in Scheme code.

I've already done all of that in Scheme and Common Lisp and it's the
best of both worlds.
<https://github.com/lispunion/database-subprocess/blob/master/test-sqlite.sps>