fix alignment-token early return inside nested boxing groups

[dginev]

Summary (generated)

digestNextBody() was prematurely returning on \cr/\lx@hidden@cr tokens even when deep inside nested hbox/vbox groups within a tabular cell. This caused cascading errors when \adjustbox{valign=t}{\begin{subfigure}...} was used inside \tabular, as the \\ inside subfigure leaked out as \lx@hidden@cr and triggered the alignment early-return check.

Fix: guard the alignment-token check with a boxing-depth comparison so it only fires at the original alignment nesting level, not inside deeper boxing groups.

Also includes minor whitespace reformatting by latexmllint.

Minimal test

\documentclass{article}
\usepackage{subcaption}
\usepackage{adjustbox}
\begin{document}
\begin{figure}
\begin{tabular}{c}
  \adjustbox{valign=t}{\begin{subfigure}{2cm}B\\C\end{subfigure}}
\end{tabular}
\end{figure}
\end{document}

Easy diff

https://github.com/brucemiller/LaTeXML/pull/2775/changes?w=1

Disclosure

Diagnostic and patch via Claude Opus 4.6 "heavy effort".

I have manually verified this patch recovers arXiv:2401.06604v3 out of regression status.

[dginev]

Note that if this PR looks good, there is a second fix using the same method. I have branched it off from here to ensure there isn't cross-talk between the bugs (there wasn't). If it looks promising I can point the other PR over to the master branch, when this one is merged:

dginev#12

[brucemiller]

Another curious one. I still had my tentative patch to #2776 (which essentially is fixing \hbox) in place when I ran the above test and it gave me no errors, but it's missing the B in the output. Hmm. I got exactly the same result with your patch (also missing the B). That suggests to me that my proposed patch is a more fundamental one; although, possibly the same sort of test should mask the alignment tokens in digestNextBody? At any rate, there's still something off.

[dginev]

Interesting. Still quite slippery.

To salvage the B we likely need to avoid running into the alignment behavior of \\ within the subfigure? Maybe a Let('\\\\', '\lx@newline'); added to the start of sub beforeFloat, so that we override any outer rebindings of par ?

(Tried the suggestion, B is recovered. Though naturally the subfigure contents get yanked out of the ltx:tabular since subfigure tries to float up.)

[brucemiller]

I was pondering along the same lines (rebinding \\): unclear why the "wrong" \\ should disappear the B, though. At any rate, the whole thing's got schema issues. Would seem to want a smarter Alignment strategy (can't put this stuff in a table, so use css/flex/whatever)

[dginev]

Would seem to want a smarter Alignment strategy (can't put this stuff in a table, so use css/flex/whatever)

Just to be clear -- that sounds completely out of scope for my "arXiv regression patching" sprint I am trying to complete for v0.9. If we recover content but "get bruised" on markup fidelity, that is ok for now.

So I'm happy with this kind of PR, if you are :-)

[brucemiller]

Looks good! Thanks!!