fix: a captioned quote is a figure in both directions again - #69

The spec withdrew the quote attribution model in markup-carve/carve#1213, so a
captioned quote is a `figure` whose target is a `block_quote` and whose caption
is the `^ ` line. This bridge still carried the withdrawn model, and that was
not merely stale bookkeeping - it was two live defects against the pinned
engine, both measured rather than argued.

The exchange AST did not validate. `pandocToCarveAst` put an `attribution`
field on a `block_quote`, and `resources/ast-schema.json` refuses it: the node
declares `additionalProperties: false`, so reversing a `Figure[BlockQuote]`
produced a document the published schema rejects with
`{"keyword":"additionalProperties","params":{"additionalProperty":"attribution"}}`.
The one thing an exchange AST has to be is the exchange AST.

Caption numbering disagreed with the engine. For a captioned quote followed by
a captioned image, both spelling `Figure #`, the engine renders `Figure 1` then
`Figure 2`. The bridge kept the quote's `#` literal and started the sequence at
the image, so pandoc's HTML writer printed `Figure #: Hamlet` and
`Figure 1: Second` - a placeholder that never resolves and an off-by-one on
every later figure and cross-reference.

A quote figure now lowers through the ordinary figure path, and a
`Figure[BlockQuote]` reverses to `figure{target: block_quote}`. The runtime
probe of the engine's serializer and the lowering that rewrote the shape on the
way out are both gone with the model they served.

The attribution lowering existed for a real reason and the reason still holds:
pandoc's plain and rst writers drop a `Figure`'s caption. Measured on pandoc
3.5, that is not quote-specific. The plain writer drops the caption of EVERY
figure, image figures included; the rst writer keeps one only through the
`.. figure::` directive, which is the image case. A figure wrapping a code block
loses it identically. So the loss belongs to pandoc's writers for non-image
figures, and routing around it by moving the caption inside the quote puts the
attribution in the one place the HTML Standard forbids - which is what the spec
withdrew. latex, html, markdown, docbook and org all keep it, and the roundtrip
test now asserts both halves of that table so the day pandoc changes, the test
says so.

Nothing had to be kept working: `block_quote.attribution` never appeared in any
published engine (the carve npm package is at 0.1.3, which has no such field),
and this bridge has no tag and no npm release, so no document produced by the
withdrawn model exists outside a working tree. Reading the
`[X]{.attribution}` span back in was therefore dropped rather than kept: nothing
emits it anymore, and re-importing it would move content out of a blockquote on
the strength of a class name this bridge itself invented.

Measuring the new path surfaced a separate crash that predates it. Pandoc's own
HTML reader emits a `Figure` with an EMPTY caption for
`<figure><img src="a.png"></figure>`, and both figure branches built a `figure`
node with no `caption` field from it - refused by the schema, which requires the
field, and fatal in the writer with
`Cannot read properties of undefined (reading 'forEach')`. Emitting `caption: []`
instead is not the fix: `renderCarve` writes a lone `^` line for it, and that
line re-parses as a lazy continuation, so `> q` comes back as the two-line
paragraph `q\n^` inside the quote. An uncaptioned figure is emitted as its host
instead - a shape Carve source can spell - and the dropped wrapper is reported.