HTML Table in PDF - `\multirow` is used but somehow, `\usepackage{multirow}` is missing when table is crossreferenced

[cderv]

So the issue is that \multirow is used, but somehow, \usepackage{multirow} is missing when

::: {#tbl-test}

```{=html}
<table>
  <thead>
    <tr>
      <th>species_spanner</th>
      <th rowspan="2">island</th>
    </tr>
    <tr>
      <th rowspan="1">species</th>
    </tr>
  </thead>
</table>
```

:::

but not if the table is not crossreferenced

```{=html}
<table>
  <thead>
    <tr>
      <th>species_spanner</th>
      <th rowspan="2">island</th>
    </tr>
    <tr>
      <th rowspan="1">species</th>
    </tr>
  </thead>
</table>
```

So probably related to multirow variable not set in one of the scenario

quarto-cli/src/resources/formats/pdf/pandoc/tables.tex

Lines 6 to 8 in 2f24776

	$if(multirow)$
	\usepackage{multirow}
	$endif$

Originally posted by @cderv in #13084 (reply in thread)

Full reprex

---
format: pdf
keep-tex: true
keep-md: true
---

::: {#tbl-test}

```{=html}
<table>
  <thead>
    <tr>
      <th>species_spanner</th>
      <th rowspan="2">island</th>
    </tr>
    <tr>
      <th rowspan="1">species</th>
    </tr>
  </thead>
</table>
```

:::

Git diff between two tex files

diff --git "a/.\\index-ori.tex" "b/.\\index.tex"
index d3add64..b1718c8 100644
--- "a/.\\index-ori.tex"
+++ "b/.\\index.tex"
@@ -67,7 +67,6 @@
 
 
 \usepackage{longtable,booktabs,array}
-\usepackage{multirow}
 \usepackage{calc} % for calculating minipage widths
 % Correct order of tables after \paragraph or \subparagraph
 \usepackage{etoolbox}
@@ -167,6 +166,11 @@
 
 
 \begin{longtable}[]{@{}ll@{}}
+
+\caption{\label{tbl-test}}
+
+\tabularnewline
+
 \toprule\noalign{}
 species\_spanner & \multirow{2}{=}{island} \\
 species \\
@@ -174,6 +178,7 @@ species \\
 \endhead
 \bottomrule\noalign{}
 \endlastfoot
+
 \end{longtable}

[cscheid]

What's the show-ast-trace output of these two?

[cderv]

I am onto it. I haven't yet done the investigation.

[cderv]

I believe this happens because when this is a floatref, Quarto will write the table to raw LaTeX inside the Lua processing using pandoc.write(), and then insert raw LaTeX output.

quarto-cli/src/resources/filters/customnodes/floatreftarget.lua

Line 415 in 2f24776

local raw_output = pandoc.RawBlock("latex", pandoc.write(pandoc.Pandoc({tbl}), "latex"))

The problem with that is all that the Pandoc writer is doing for the writer context is lost. Especially metadata that could be set based on written content.

For multirow, this is the Pandoc LaTeX writer that detects it, and then set to true the metadata
https://github.com/jgm/pandoc/blob/826f487818e6b36ec1e6feeafaa2e9449ce47695/src/Text/Pandoc/Writers/LaTeX/Table.hs#L354C9-L355

  when (rowspan /= RowSpan 1) $
    modify (\st -> st{ stMultiRow = True })

And this value is used to defined the template variable
https://github.com/jgm/pandoc/blob/826f487818e6b36ec1e6feeafaa2e9449ce47695/src/Text/Pandoc/Writers/LaTeX.hs#L214

                  defField "multirow" (stMultiRow st) $

So basically, we are not retrieving all the metadata that would be used by the writer to pass to the template.

Hopefully we have a way to do that. I'll look into this, but @tarleb I could get your insight

[tarleb]

I think it's save to simply always set the multirow variable to true: LaTeX compilation will be a tiny bit slower on average, because the respective LaTeX package will be loaded even when it's not needed. But I'd be surprised if it was more than a few milliseconds.

[cderv]

What bothers me is that

• multirow may not be the only thing that a Table would activate for pandoc's template.
• Adding multirow in all templates even if not used is a bit too heavy to me - we would force LaTeX users to have a package they would not use, as long as there is a table in the document. This could cause trouble for Journal's article or other LaTeX publication where they have rules for dependencies.

I am thinking I could try a workaround by using a custom template for pandoc.write so that output can be parsedto get the table on one side, and all values for variables on the other...
I would need to extract all variables in advance, but this could work as a way to get access to the values and then reset them for later processing.

[cscheid]

Quarto has text-based latex postprocessors. Because it's brittle and hard to reason about, it's not a tool we should use lightly. But we could add a text-based postprocessor that attempts to detect the combination of

• multirow feature usage
• no package inclusion

When that combination is detected, we could inject the appropriate \usepackage call.