Link reference definitions are read as paragraphs #92

Summary

#87 fixed the hard-break half of #25. The other half, which #25 also mentions, is still open: a CommonMark link reference definition ([label]: url) is read as an ordinary paragraph. So a [text][label] link that uses it does not resolve.

Reproduction (0.3.0)

from legaldown import parse_document

src = "---\ntitle: T\n---\n\n# A\n\nSee [the site][x].\n\n[x]: https://example.com\n"
print([(b.kind, b.text) for b in parse_document(src).sections[0].blocks])
# [('paragraph', 'See [the site][x].'), ('paragraph', '[x]: https://example.com')]

In CommonMark, the second block is a definition. It renders nothing, and it makes [the site][x] a link to https://example.com.

Impact

legaldown-render builds only from this model, so it renders the definition as visible text and the link as literal brackets. A renderer could parse definitions out of paragraph text itself, but that re-derives a CommonMark block rule the parser owns.

Proposal

The parser should recognise link reference definitions, and the model should expose them. For example:

• a link_definition block kind with label, target and title, or
• a document-level link_definitions: dict[str, (url, title)], keyed by the normalised label.

A definition can also start a paragraph or follow another one, and it can appear in list items and quotes. Serialising the model must keep it.

Found while adopting 0.3.0 in legaldown-render.