feat: check reference-style links

[silversurfer562]

Closes the known gap left open at 0.3.0: only inline [text](target) links were resolved, so every reference form was a silent pass — a dead target behind a label went unreported.

What now works

All three CommonMark forms resolve against the document's [label]: target definitions:

Form	Example
full	[the guide][guide]
collapsed	[guide][]
shortcut	[guide]

Labels match case-insensitively with whitespace collapsed. Definitions may carry a title or <angle-bracket> target, may be indented, and the first definition of a repeated label wins.

An explicit reference with no definition is an error. [text][label] without a matching definition renders as literal text — the link does not exist, which is a refuted claim rather than an unverifiable one. Findings quote the reference form ([text][label]) rather than inline syntax the reader would not find in their document.

Three false-positive classes closed while building it

Each found by probing real markdown shapes, not by the corpus:

• Undefined shortcut references are ordinary prose. [3] and [TODO] only become links when a definition exists. Flagging them would fire on any bracketed text — including every [Unreleased] heading in a Keep a Changelog file. This is the whole reason shortcut support is safe to have.
• GFM footnotes share reference syntax exactly. [^1] against [^1]: Sourced parsed as a link whose target was the footnote body, flagging Sourced as a missing file. The ^ namespace is now excluded.
• Definition indentation is accepted beyond CommonMark's three-space cap. A definition nested under a list item is real and common, and missing one turns its reference into an error-severity false positive — the costlier direction. Fenced blocks are masked before definitions are read; indented code blocks aren't modelled anywhere in the extractor, so this stays consistent with the rest of it.

Definitions inside code fences correctly do not define — a renderer doesn't read them out of a code block, so a reference depending on one is properly reported as undefined.

Verification

Gate	Result
Tests	153 passed (was 130)
Coverage	96%, _extract.py 100%
Corpus precision/recall	passing, 6 new regression cases
Mutation score	81.3% against the 75% gate (890 mutants, up from 786)
Lint	ruff + black clean
Dogfood	verify over its own README/CHANGELOG — no new findings

Version left at 0.3.0 with the entry under [Unreleased]; this is additive, so the next release is a 0.4.0 minor.

🤖 Generated with Claude Code