I may need some help debugging why the tests are failing. For TexLive 2021, I'm not sure why class="ltx_figure_panel" is getting added. For TexLive 2022, siV3 having 10.0 instead of 10 indicates that the binding is operating in $VERSION_TWO mode, which means that \fmtversion reported a year prior to 2021.
@dginev , do you have any insight?

@teepeemm first thought before looking at any of your changes -- try syncing with the upstream master branch before continuing:

$ cd your-fork-directory
$ git remote add upstream https://github.com/brucemiller/LaTeXML
$ git fetch --all
$ git checkout siunitx-tpmm
$ git rebase upstream/master
$ git push --force

We have CI passing now, so getting to the latest state will make this easier to review by everyone.

Catching up to upstream/master did occur to me, and I think I did that (somewhere around the 6afbc9f commit). git diff --name-status upstream/master reports these 16 files, and nothing else.

It seems like I was failing because of the class ltx_figure_panel, which comes from a width calculation in a table or longtable. I'm not sure why that wasn't happening on my end, but I've taken out the tables (but left in the tabulars).

I've completely disabled the siV3 test, so we'll need to look into why that wasn't working. I'm not seeing \fmtversion in any of the test files. Is it possible it expands to something different in GitHub actions? (I don't suppose they share the make formats?)

I'm not sure why 2021 is still failing. I'm not seeing what would cause sectional-block instead of logical-block, and why 2021 would be special in that regard.

Congrats on passing CI! I will take some time to review next week, apologies for being a little short on time.

I figured out why siV3 was having trouble, and left a comment in t/80_complex.t: lib/LaTeXML/Engine/latex_base.pool.ltxml sets \fmtversion to 2018/12/01. If make formats is called, blib/lib/LaTeXML/Engine/latex_dump.pool.ltxml overrides \fmtversion to the actual value. The CI calls make formats only for 2023 onward, so I restricted siV3 to just that. This also means make test CI=true will fail without make formats (unless we have expl3 always make formats if it's not already present?).

I'm still not sure why t/complex/si was failing 2021, and that last commit just removed that test. So passing CI is a bit of a cheat. I'll try to look into that.

I'm unable to get the sectional-block tag that the failing test (t/complex/si with TexLive2021) is getting. I've installed TL2021, run make clean, make, and used LATEXML_KPSEWHICH=/usr/local/texlive/2021/bin/universal-darwin/kpsewhich, but I can only get output identical to LT2022--25 with logical-block tags. (I've also tried TL2020.)

@dginev Were you able to get around to reviewing this, or were you waiting for me to figure out the failing test? I'm not quite sure how to debug why a GH action is differing than my local result. Is there an easier way than editing the action to focus on that test?

@teepeemm I have disappeared in a maze of corners to examine since March 2026, so no, no updates from me on your PR.

I will get back to the open PRs from a very intriguing perspective come August (post-release) that may be a bit controversial. Basically, the Rust rewrite of LaTeXML is completed (core-complete since April), and is likely to be released in summer (and made public). At which point I plan to transpile all PR suggestions over as Rust suggestions.

The controversial bit is that this effort succeeded in very large part due to AI assistance, in particular the Claude Opus series of models from Anthropic. And I plan on continuing to use them until we map out all of the required coverage for arXiv. So, with all of those caveats, I will be happy to look at your PR in a couple of months!

In the meantime Bruce may beat me to it, which would also be great for the perl side.
I am giving you the long story since you tagged me. More "soon"!