feat: a composite figure is an editable figure in the editor

[dereuromark]

Closes #15.

markup-carve/carve-grammars#225 landed both halves this was waiting on: an
engine pin that parses a bare ::: figure into a figure_group, and a
carveFigureGroup entry in the CarveKit schema. This bumps the pin onto it and
wires the editor up.

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

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

loads as one editable carveFigureGroup: its direct carveFigure and table
children are the panels in source order, the ^ line below the closing fence
is the group's own caption, and anything else in the body stays where it was
written. Before this the whole document became one opaque carveUnsupported
atom - lossless to load and save, and gone on the first edit anywhere.

tests/composite-figure.test.ts said, in as many words, that it would go red
the day carve-grammars gave the group a schema entry and that this was the
signal to model it here. It went red. It now asserts the rich shape, the round
trip through the app's own import and serialize path, and an edit, which is
the case the opaque atom could never survive.

The titled and labelled openers are a different production and keep their old
reading. That control used to prove nothing - the engine could not tell the two
spellings apart, so both parsed to the same generic container and the pair
agreed with itself. The test now asserts the two READINGS DIFFER, so it fails
rather than passes if the pin ever moves back.

The bump surfaced a silent regression, and it is not about figures

pruneDefaults in src/editor.ts stripped attributes the editor materialized
from schema defaults, so an unedited document could be recognized and its source
envelope honored. It stripped only the null ones.

A default does not have to be null. markup-carve/carve-grammars#221 gave
carveCaption a short attribute defaulting to false, so from that pin
onward every document holding a caption came back carrying {"short":false},
never compared equal to the document it was loaded from, lost its envelope, and
was written back from the lossy rich model. The visible loss was {#fig-x}
above a titled ::: figure disappearing on save. Nothing failed - the round
trip simply became lossy, which is the failure the envelope exists to prevent.

The check now compares the mounted document with the one that was loaded, both
reduced to what the author wrote, and hands the serializer the original
document when they match. Two things about that:

• Both sides are reduced. The bridge sets some attributes to a value that is
  also the schema default - carveComment writes block: false for every %%
  line - so reducing only the mounted side breaks exactly the documents the
  reduction exists to keep.
• The original document is handed over, not a reconstruction of it. The
  envelope's fingerprint was taken over that exact JSON, and the envelope is
  only honored while the fingerprint matches, so anything less is a guess at
  what the fingerprint will accept.

Also here

• Preview and editor styles for the group. The preview pane holds the engine's
  HTML (figure.carve-figure-group wrapping div.carve-figure-panels) and the
  editor holds CarveKit's node, which has no panels row; both are styled so the
  group reads as one figure on either side of the split.
• tests/blockquote-attribution.test.ts is now blockquote-caption.test.ts.
  Its assertions were deliberately shape-independent and passed unchanged, but
  its docblock claimed the engine carries a quote's caption as an attribution
  field on block_quote. PART 9 §4a said that; §4b withdrew it
  (spec: withdraw the quote attribution, a captioned quote is a figure again carve#1213), and a captioned quote is a figure whose target is
  the quote again - the loader builds the carveFigure/carveCaption pair.
  Verified against the spec's current main, not from the ticket text.
• Two README "lossy" limitations retired. The composite-figure one is this
  change. The {:TAG}-typed-in-the-source-pane one was already false before it:
  fix: bump carve-grammars onto the build that parses {:TAG} and keeps a quote's attribution #14 moved the pin onto a build that parses it and
  tests/language-attribute.test.ts has asserted it ever since, so the README
  was contradicting a passing test.

Proofs

Every claim above was checked in both directions - mutate, watch it go red,
restore from HEAD, watch it go green:

• convertFigureGroup reduced to the generic container (an engine without §4c)
  • red, on the "the two differ" control
• the null-only pruning restored - red, on the titled round trip
• the envelope reduction applied to the mounted side only - green at first,
  which is why a test was added
• the envelope trusted unconditionally, never falling through after an edit -
  red

The one-sided reduction is worth naming. {#fig-x} above a ::: note
round-trips either way; the same document with a %% line inside it does not,
because that is where an attribute set to its own default appears. The pair in
tests/roundtrip.test.ts is what pins the symmetry, and a third case edits the
document and asserts the output changed, so a serializer that just returned the
loaded source could not pass the first two.

The spec submodule pin

npm run check:pins warns that carve-grammars' build was written against a spec
53 commits behind the engine's. That gap is inside carve-grammars and no pin
here can move it; bumping its spec/ submodule is 58 commits and 316 corpus
files, which is its own change. The warning is the watchdog naming the
repository that can close it, which is what it is for.