here is cleaned/finished previous implementation of RAW_TEXT/RAW_BINARY formats for COPY statements.
The RAW with text formats means unescaped data, but with correct encoding - input/output is realised with input/output function. RAW binary means content produced/received by sending/received functions.
As much as I know you and some other people would like it to be otherwise, this patch clearly does not have a sufficient degree of consensus to justify committing it to PostgreSQL 9.6. I'm marking it Returned with Feedback.
I should add that I've been thinking about this some more, and that I now agree that something should be done to support this at the SQL level, mainly so that clients can manage very large pieces of data in a stream-oriented fashion rather than having to marshall the data in memory to load/unload via INSERT/SELECT. Anything that is client-side only is likely to have this memory issue.
At the same time I'm still not entirely convinced that COPY is a good vehicle for this. It's designed for bulk records, and already quite complex. Maybe we need something new that uses the COPY protocol but is more specifically tailored for loading or sending large singleton pieces of data.
Now it is little bit more time to think more about. But It is hard to design some more simpler than is COPY syntax. What will support both directions.
My implementation has same limit like COPY BINARY - it isn't worse. It should be good enough for VARLENA types that should not be higher than 1GB. It is not designed for LOB replacement.