Map figure_group to pandoc's subfigure model, both directions

[dereuromark]

Closes #62.

Carve PART 9 section 4c gives a bare ::: figure container the figure_group node: one figure of ordered panels, whose ^ line after the closing fence captions the whole group. Pandoc has that model natively - a Figure whose blocks are themselves Figures is its subfigure shape - so this maps structurally in both directions rather than degrading.

Carve to pandoc

The group becomes the outer Figure, its caption the outer caption, its children the blocks. The panels - the direct figure and table children - lower to nested Figures and Tables, and everything else stays in place between them.

Input:

{#fig-x .columns-2}
::: figure
{#fig-x-a}
![one](a.png)
^ (a) One

{#fig-x-b}
![two](b.png)
^ (b) Two
:::
^ Figure #: Group caption

Through pandoc -t latex:

\begin{figure}
\centering
\begin{subfigure}[t]{0.45\linewidth}
\centering
\pandocbounded{\includegraphics[keepaspectratio]{a.png}}
\caption{(a) One}\label{fig-x-a}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.45\linewidth}
\centering
\pandocbounded{\includegraphics[keepaspectratio]{b.png}}
\caption{(b) Two}\label{fig-x-b}
\end{subfigure}
\caption{Figure 1: Group caption}\label{fig-x}
\end{figure}

This replaces the Div ["admonition","figure"] a bare ::: figure used to cross as. A filter keyed on that Div has to key on the Figure nesting instead; the changelog says so.

Pandoc to Carve

A Figure holding nested Figure/Table blocks used to hit the "general figure content unwrapped" path, which dropped the grouping and left the caption as a trailing paragraph. It imports as a figure_group now. Single-block Figures keep their existing mapping on purpose: a lone captioned image or table is one figure, not a group of one, and nothing in the document distinguishes the two if it were.

Numbering follows section 4c, not source order

• The group draws ONE number at its OPENING fence, which is why its caption converts before its children. Corpus 318-composite-figures-11 numbers the group "Figure 1" and a captioned figure nested in its stray content "Figure 2", even though the group's ^ line is the construct's last.
• A panel draws nothing, and neither does anything a panel CONTAINS. Its # prints as written - the visible failure the language prefers to a silent one. Suppressing the draw matters as much as suppressing the digit: a panel that consumed a number would shift every later caption in the document by one.
• A panel id resolves </#id> as the group's number plus a bijective base-26 letter by panel order - "Figure 2a", corpus 318-composite-figures-2. A group with no numbered caption lends its panels nothing.

One asymmetry is deliberate and section 4c states it: inside a group a captioned quote is a PANEL, not a section 4a attribution, because "the quote is not a special host inside the group either" (corpus 318-composite-figures-10). The 4a reroute still governs everywhere else, including deeper inside the group's own stray content.

Engine pin

No published @markup-carve/carve carries the node - 0.1.3 parses ::: figure as a generic container, so no arm for the composite node compiles against it. The dependency is a git pin at carve-js 3f5dd8c (markup-carve/carve-js#1073) until an engine release carries it.

The lockfile's resolved url is written in the git+https:// form by hand. Hosts with a git insteadOf rule rewriting HTTPS to SSH make npm install record git+ssh://git@github.com/..., and npm ci follows the LOCKED url - which fails in CI with a publickey error on a public repository. Verified by deleting node_modules and running npm ci.

Also fixed

A pandoc Figure wrapping a single Table was dropping the wrapper's attributes when the two collapse into one Carve table. Pandoc's readers put the label on the Figure rather than on the Table it wraps, so the id a cross-reference resolves against went with the wrapper. They merge now - outer id wins, classes union, key/values merge with the outer taking precedence - which is the rule the 4a quote branch already applied to the same kind of collapse. It bites hardest on a section 4c table panel, whose id is what resolves with a letter.

Note on the ticket's numbering line

The ticket asked the round trip to pin "a panel caption's # placeholder survives as the typed node without n". It does not, and the test records what actually happens instead: a caption's # is a request and the rendered number is the answer, this bridge resolves on the way out for every figure, and the reverse importer never rebuilds a caption_number from printed text for anything. So the group's # comes back as "1" and a panel's comes back as the escaped character it printed - visually identical, still drawing nothing. Reconstructing the placeholder would mean guessing it back out of a rendered # and misreading an author's genuine one.

Every case has a CONTROL beside it, because the two spellings are one character apart and an engine without section 4c renders them identically - a group test can pass against an implementation that never discriminated at all. Both controls were mutated to feed them the group spelling and both then failed, so they discriminate rather than pass on anything.