what's needed is a application barrier() call

But now you're synchronizing all writes by all processes everywhere on the disk, not just the two blocks that actually got written to in this one file somewhere. If apps used that sort of barrier() too often then I think you'd end up with disk throughput that looks like you mounted with -o sync -- very little reordering would ever be allowed.

I guess this is complicated by the question of when dirty blocks get flushed, and in how large batches; maybe it's solvable. But I don't think memory is the main concern, at least.