feat: a bare figure fence highlights as a composite figure

[dereuromark]

Carries the composite-figures vocabulary (PART 9 §4c, landed as markup-carve/carve#1215, design markup-carve/carve#1122) into the queries this repo ships.

The grammar needed nothing, and that is the finding

Measured before writing any rule. The parser already tells the two openers apart by which FIELDS each carries:

Input:

::: figure
![one](a.png)
^ (a) One
:::
^ Figure #: Group caption

Tree:

(document
  (div
    (div_marker_begin)
    class: (class_name)
    content: (content
      (paragraph (inline_image ...))
      (caption (caption_marker) content: (caption_content)))
    (div_marker_end))
  (caption (caption_marker) content: (caption_content)))

Input:

::: figure "A titled figure div"
x
:::

Tree:

(document
  (div
    (div_marker_begin)
    class: (class_name)
    title: (div_title)
    content: (content (paragraph))
    (div_marker_end)))

A bare opener carries neither title: nor label:; a titled one does. And the ^ line after the closing fence already parses as a SIBLING of the container, not inside it, which is exactly where §4c puts the group caption - §4's sixth host. So a reserved kind word wants a reserved CAPTURE, not a new node. Three parse shapes are pinned in test/corpus/carve.txt so a later grammar change cannot quietly take the distinction away.

What changed

File	Change
queries/highlights.scm	a bare ::: figure opener's class_name captures @type.builtin; four patterns in total, one for the group and three restoring @type inside it
test/corpus/carve.txt	three parse tests: the bare group with its caption after the closer, the titled and labelled openers, the nested pair
scripts/highlight-captures.mjs	new, resolves what an editor would actually paint
.github/workflows/ci.yml, package.json	run it

!title !label is the whole distinction and is why this is expressible as a query at all. A titled or labelled opener matches nothing new and keeps @type.

GROUPS DO NOT NEST is three more patterns at a higher priority, restoring @type on a bare opener inside a group. A tree-sitter query has no transitive closure, so "any descendant" cannot be written; what can is a wildcard chain rooted at the group, one per intervening level. Three levels covers every shape the language produces - the direct child, div/block_quote > content > div, and list > list_item > list_item_content > div. The residual, a bare opener more than three levels down, is stated at the rule.

A check this repo did not have

The queries are the other half of what this package ships - vim-carve, helix-carve and zed-carve pin this repo and load these files - and nothing here read them. A capture naming a node that no longer exists, or two patterns claiming one node with nobody deciding which wins, left every check green.

scripts/highlight-captures.mjs resolves what a consumer resolves: highest (#set! priority N) wins, later patterns break a tie, 100 by default. Reporting every match instead - which is what tree-sitter query prints - would call the nested cases green while the editor painted the generic colour over them.

Two things it has to get right, both found by watching it fail first: spell, nospell, conceal and none are not colours and land on the same nodes the colour patterns do; and #offset! is a directive the node binding refuses to build a query with, so it is stripped rather than checking a hand-copied subset of the file, which would have passed while the real file was broken.

Downstream

vim-carve and helix-carve carry their own copies of these queries and pin this repo, so both need a pin bump plus the same two-file query change once this lands. zed-carve likewise. Those are separate PRs.