I know this is not the answer you want but honestly I think your approach itself may need reconsidering, as to me it looks unreliable; you have to manually do a lot of steps very consistently, which for production work is usually a bad idea when the goal is error free repetition.
To your specific question: I would go by file names, and write a script to take the selected JPG names, and copy the corresponding named items from the card. That requires some programming, but not difficult programming.
But really I would start over and reconsider the core flow: Download time is not all that slow, at least in human time. Consider a product like Photo Mechanic (there are others but IMO that's the best; I have no relationship except as a customer). It will let you download both JPG and RAW, cull at your speed (I've never seen it be slow enough I had to wait while browsing through). You can then select and code metadata as you prefer. These (raw or jpg or both) can be handed off to lightroom or photoshop if you like, or you can directly process (e.g. upload) the JPG's.
It is not an editor, it is a front end to handle large volumes of shots quickly. I ingest, cull, crop, straighten (it does those) and then hand off only 5-10% of the shots to Lightroom for final post processing.
You can then set up whether JPG or RAW is used for specific functions, and both stay together. When done with a shoot just delete all the ones not used, raw and JPG both, in terms of space saving you mentioned.
I am not saying your workflow is invalid; it clearly works for you. But it is sideways to products out there used widely. Photo Mechanic is something of a de facto standard for photo journalists on deadline for fast, reliable work, but lots of people use other techniques and tools. Most of those involve some variation of ingest everything, sort things out on disk not cards, keep jpg and nef's associated don't separate them, and do not do things manually a computer can do more reliably.