I think my original thought was that if we pick a variant and it's just a little too short (<= short fall), then it's not worth picking a way larger one, but if we create a construction, we might as well create the exact desired size.

The interpretation here instead is that it's generally good to make the stretched thing a little less wide/large than the base/content. That's more consistent I guess. Do we happen to know what LuaLaTeX does? Probably the same as this here?

I think my original thought was that if we pick a variant and it's just a little too short (<= short fall), then it's not worth picking a way larger one, but if we create a construction, we might as well create the exact desired size.

I think I gathered this as well. But I didn't like the idea of either having the short fall hard coded (and so appears as a bit of magic to the end user), or adding another parameter along with the size...

The interpretation here instead is that it's generally good to make the stretched thing a little less wide/large than the base/content. That's more consistent I guess. Do we happen to know what LuaLaTeX does? Probably the same as this here?

Yeah my reasoning is really just the above. I believe *TeX uses the following formula calc.max(x - 5pt, x * 0.901), and then just picks among the variants or assembly to get the smallest one at least the calculated size. So basically the same as this PR then.

I could see it being nicer to only apply the short fall for prebuilt variants, but exposing this will just be confusing to end users I think. Also, the prebuilt variants and their sizes are entirely font dependent anyways...

Alright, I think I'm good with this change. The updated reference images also actually look a bit nicer imo.

I'll sort out the other PR now.