It seems like a stretch to argue that the model isn't "a work based on" GPL code when that GPL code is an input to a deterministic algorithm from which the model is produced. So, my bet is on point #1.

The only ambiguity as far as I can tell is GPL covers "source code", "machine-readable Corresponding Source", and "object code form", and it's not explicit whether vector-fields count as any of those things. I doubt anyone would seriously argue that zipping and then un-zipping some GPL source code means you don't need to respect the original license. LLMs are different in that they're lossy compared to the zip format - does the nature of this lossiness invalidate the intent of the GPL's original language? I doubt it.