It also reminded me of a vague inverse of this hack. In old versions of Qemu (possibly it is still implemented, but I have vague memories it got ripped out), you could point Qemu at a directory on disk and it'd produce an emulated floopy disk drive with a virtual FAT12 image containing the directory contents. AFAIK it didn't keep the actual data in memory, I guess all it needed was file sizes to know how to build a virtual memory mapping that contained the filesystem metadata + proxied reads from the underlying files for data sectors. I look forward to seeing your implementation of this concept in a virtualized SQLite file <-> GraphQL proxy ;)
edit: insane, it still exists and apparently supports write mode?! https://en.wikibooks.org/wiki/QEMU/Devices/Storage#Virtual_F...
If you want it to fetch less data for querying a subset of columns, you could create create an index on those columns - then SQLite will do an COVERING INDEX scan and thus read only the necessary data (with the B-Tree itself and the start / end page alignment being the only overhead).