I'm not convinced your L/S dichotomy applies. The concern there is that the natural world (or some objective target domain) has natural joints, and the job of the scientist (, philosopher, et al.) is to uncover those joints. You want to keep 'hair splitting' until the finest bones of reality are clear, then grouping hairs up into lumps, so their joints and connections are clear. The debate is whether the present categorisation objectively under/over-generates , and whether there is a factor of the matter. If it over-includes, then real structure is missing.

In the case of embeddings vs. vectors, classical vs., baroque, transpiler vs., compiler -- i think the apparent 'lumper' is just a person ignorant of classification scheme offered, or at least, ignorant of what property it purports to capture.

In each case there is a real objective distinction beneath the broader category that one offers in reply, and that settles the matter. There is no debate: a transpiler is a specific kind of compiler; an embedding vector is a specific kinds of vector; and so on.

There is nothing at stake here as far as whether the categorisation is tracking objective structure. There is only ignorance on the part of the lumper: the ignorant will, of course, always adopt more general categories ("thing" in the most zero-knowledge case).

A real splitter/lumper debate would be something like: how do we classify all possible programs which have programs as their input and output? Then a brainstorm which does not include present joint-carving terms, eg., transformers = whole class, transformer-sourcers = whole class on source code, ...