That's a good point, but the basic idea is to send the local query almost-as-is to the remote server if possible. For example, if the local query is "INSERT INTO foreign_table(a,b,c) VALUES (1, 2, 3), (4, 5, 6)", send the remote query "INSERT INTO remote_table(a,b,c) VALUES (1, 2, 3), (4, 5, 6)" to the remote server where remote_table is the table name for the foreign table on the remote server. So, wouldn't the query string length be a problem in many cases? Maybe I'm missing something, though.
FDWs don't operate at that level. They don't see the original query string. They're plan nodes that operate with a row-by-row push/pull model. The foreign table node in question has no idea you're doing a multivalued insert and doesn't care if it's INSERT INTO ... SELECT, INSERT INTO ... VALUES, or COPY.
That's why I think using batching is the way to go here. Each operation remains isolated, but you don't force a round trip for each one, you just queue them up on the wire and you flush only at end-of-statement. A failure will cause the statement to ERROR and abort the tx, so the effect is the same, though the failure might be a bit later than if you forced a flush each time.