feat(epub): metadata front matter and figure caption dedupe

[mathurshubham]

What

Two EPUB-focused output improvements, targeting the completeness and
cleanliness axes where EPUB trailed the other formats in the benchmark
(README: EPUB 74, the only format anydoc did not lead).

A third change — stripping the whole-heading **bold** that pandoc's EPUB
CSS forced onto headings — was originally part of this PR but landed
upstream independently (2d3f923), so it has been dropped here. This PR now
covers metadata front matter and figure-caption dedupe only.

1. OPF metadata → YAML front matter (was: duplicate title, dropped metadata)

epub/mod.rs read only <title> and injected it as a body H1. Most books
already carry their own title page in the content, so the output was
double-titled; author, language, date, and publisher were dropped entirely.

Now Dublin Core metadata renders as a ----fenced YAML front matter block:

• New DocumentMeta on the model (title, authors, language, date,
  publisher, description), mirrored in the Node and Python bindings.
• Parses the OPF <metadata> descendants (so EPUB2's <dc-metadata>
  nesting is covered), scoped to <metadata> so a stray <title> elsewhere
  can't leak in.
• Main title chosen via the EPUB3 refinement
  <meta property="title-type" refines="#id">main</meta>, falling back to
  the first non-empty dc:title.
• Keeps every dc:creator; a single author is a scalar, multiple render as
  a YAML list. Repeated dc:date (EPUB2 opf:event) takes the first
  non-empty.
• All values are quoted and escaped, so colons/quotes/dashes stay valid YAML.

Only EPUB populates meta; every other format leaves it default and renders
no front matter.

2. Duplicated figure captions removed

pandoc wraps a lone image as
<figure><img alt="x"><figcaption>x</figcaption></figure>, which the
converter walked transparently, printing the text twice (tiny dot tiny dot).

<figure> now has its own handler: a <figcaption> matching a contained
image's alt (normalized: trimmed, whitespace-collapsed, case-folded) is
dropped; a distinct caption is kept as its own paragraph with its own CSS
styling, display: none respected, and its id kept as an anchor. Image
alts are collected from figure descendants, so wrapped <a><img></a>
counts.

Real-world before/after

Converted three Project Gutenberg EPUBs with main and with this branch;
every changed line is one of the intended improvements, with no other drift.
Conversion time is unchanged.

Alice's Adventures in Wonderland — author/date/language were dropped
entirely before:

-# Alice's Adventures in Wonderland
+---
+title: "Alice's Adventures in Wonderland"
+author: "Lewis Carroll"
+language: "en"
+date: "2008-06-27"
+---

Pride and Prejudice and Frankenstein gain their author, date, and
language the same way. (These books contain no <figcaption>, so the
caption dedupe is exercised by the fixtures and the pandoc-built book.epub
snapshot instead.)

Tests

• Rust: 202 tests pass (cargo test --locked); fmt + clippy -D warnings
  clean.
  • New renderer unit tests: front-matter shape / escaping / multi-author
    list / empty-meta.
  • New html.rs tests: caption dedupe, distinct caption kept, empty alt,
    wrapped <a><img>, hidden figcaption, figcaption's own CSS styling.
  • New integration test: epub_metadata_is_parsed_into_the_model.
• Node bindings: pass, index.d.ts regenerated with DocumentMeta.
• Python bindings: pass, incl. the stub-coverage guard (DocumentMeta
  registered in lib.rs, .pyi, and __init__.__all__).
• New handmade fixtures handmade-metadata.epub and handmade-figure.epub.

Snapshot changes

• EPUB snapshots gain front matter and lose the injected title H1; the
  book.epub image line loses its doubled caption.
• No other format's output moved — the new Document.meta field is inert
  everywhere except EPUB.

Notes

• Benchmark movement is a goal, not a gate: the EPUB judge/metrics ground
  truth is built from source XHTML and excludes OPF metadata, so front matter
  may not raise the score even though the output is more complete.
• book.epub's snapshot now includes pandoc's build timestamp
  (date: "2026-07-30T23:11:58Z", from the dc:date pandoc stamps at build
  time), so regenerating that fixture will drift this one snapshot line. That
  is inherent to reading dc:date, not a bug.
• The bench corpus isn't redistributable, so no before/after score is
  included; happy to iterate if a corpus run surfaces anything.

[mathurshubham]

@tomsideguide would appreciate a look at this one when you get a chance, I've kept it current with main and it's ready to review.
Since it was opened, your commit 2d3f923 landed the heading-emphasis cleanup that was originally this PR's second change, so I've dropped that part and rebased; the PR is now just the two EPUB pieces, both still unaddressed on main:

• OPF metadata → YAML front matter — EPUB is the one format in the README benchmark anydoc doesn't lead, and completeness is where it loses: author, language, and date are currently dropped entirely, and the title gets duplicated as a synthetic body H1. For the LLM-ingestion use case the README leads with, front matter is the difference between a consumer knowing what document it's reading and guessing.

• Figure-caption dedupe — pandoc-produced EPUBs (a large share of real ones) print every lone image's text twice (tiny dot tiny dot).

One design note worth your eyes specifically: the DocumentMeta model field is deliberately format-agnostic — docx (core.xml), ODF (meta.xml), and RTF (\info) all carry the same Dublin Core-ish metadata, so this is groundwork any of those formats can populate later without another model change. Node and Python bindings are mirrored so to_document stays consistent across surfaces.

Happy to split, trim, or rework any part of it if you'd rather take a different shape.