• do we need to be more selective about which parts of WD entities we retrieve?

Absolutely! T382921
Even when large entities are successfully fetched, they often fail to be processed. Chances are high that a function will require only (one or a few) labels or (a selection of) statements, and hardly ever anything else.

We should also consider caching the call to the fetch functions directly, rather than relying on caching the outermost function calls. I think this could be helpful more generally within nested compositions, but the benefits seem most clearcut for the Wikidata fetch functions. As these are both builtin and terminal, the implementation is likely to be most straightforward here too.

Here are some findings:

So, the good-ish news is that we aren't massively inflating items with our ZObject normal form transformation (but we are indeed inflating a little bit--roughly, it seems that our format is BIGGER up to 500 KiB, after which point our format is more efficient than the Wikidata representation). The bad-ish news is that there is, therefore, no quick win here. These are genuinely large objects and we have to be smart about which bits of them to process.