pre-compute entry visibility once via count-all-refs()

[joelsa]

I've had some problems with non-convergent layouts. I've found in many of these cases the problem is an oscillation due to conditional content emission.

From my point of view, the ref counts can be pre-computed once. Is there a reason why this was not done before?

This PR does just that, pre-computing the reference counts. I think this will make the layout more stable and should have no negative consequences.

[joelsa]

I am unsure whether using count-all-refs() like this is the right choice, as it recomputes __glossary_final_counts() N times (once per entry) when a single call would be enough. Using __glossary_final_counts() directly would be cleaner, right?

[quachpas]

Hi @joelsa, thanks for the patch!
Can you provide a test/showcase as well? I'd like to be able to confirm that your divergence test case is indeed fixed by this patch. I can confirm on my side that that tests pass locally. I'll run the workflow once you've got a test written up.

Using __glossary_final_counts() directly would be cleaner, right?

Yes, use internal functions as you like as long as the result is clear and concise. count-refs/count-all-refs are just more practical. In practice, I don't think the optimization is even visible because the Typst compiler does its job pretty well.
count-refs is called from print-gloss which means it's already called for each entry².

Let's not modify show-all to mean something else other than the current argument. Instead, you should add a new argument to user-print-reference and propagate it. I'll help with double checking that it's done correctly.

Then we can write in default-print-gloss

-  if show-all == true or count-refs(entry.at(KEY)) >= minimum-refs {
+ if show-all == true or entry-visible == true {

This way, things are indeed pre-computed.