graphify/extractors/markdown.py (extract_markdown, on v8 at fe66389 and in the 0.9.58 wheel)
documents this edge:

heading --references--> other node (when backtick `Name` matches a known pattern)

Nothing implements it. The extractor emits page and heading nodes, contains edges, and
references edges only for inline, reference-style and wiki links to other markdown files. A code
span in a heading or body line never produces an edge, so an AST build has no document-to-code edge
at all.

Reproduction (0.9.58):

d=$(mktemp -d)
printf 'class Widget:\n    pass\n' > $d/widget.py
printf '# Doc\n\n## `Widget`\n\nThe `Widget` class.\n' > $d/doc.md
python -c "from pathlib import Path; from graphify.extract import extract; \
r = extract([Path('$d/widget.py'), Path('$d/doc.md')]); \
print(sum(1 for e in r['edges'] if e.get('relation') == 'references'))"

prints 0.

Proposed fix, implemented and tested on a fork branch (PR to follow):

1. extract_markdown collects inline code spans outside fences and frontmatter as raw_calls
   tagged language: "markdown", owned by the enclosing heading (the page node before the first
   heading). Forms: bare Name, Name(), dotted pkg.Name, and path.py::Name(::Name)*.
2. A markdown_mentions language resolver (registered through register_language_resolver, so it
   runs after the id-remap passes) resolves them. The path::Symbol form matches the cited file
   doc-relative, then exact, then as a unique path suffix, and walks contains/method edges to the
   symbol: EXTRACTED 1.0. A bare name resolves only when exactly one callable code node carries the
   label: INFERRED 0.95. Ambiguous names, builtins, missing files and non-callable matches emit
   nothing.
3. The shared cross-file call pass skips markdown raw_calls, so prose never emits calls edges and
   document nodes stay out of call resolution.
4. The docstring is corrected.

Results on the branch: the six new tests pass; the existing suite is unchanged (5,569 passed, 14
skipped; the eight failures, four test_backup_* in tests/test_export.py and four
test_pathless_postgres_extract_initializes_empty_detection cases in tests/test_extract_cli.py,
fail identically on clean v8). A dry run over a 1,581-document, 11,297-symbol Python repo produced 3,532 document-to-code
references edges (3,409 INFERRED, 123 EXTRACTED) from 10,809 mentions; 985 ambiguous names were
skipped. The only recurring false positive found was a prose word (source) matching a same-named
test fixture function.