Fix error on emphasis around whitespace character reference (#492)

[apoorvdarshan]

Fixes #492

Root cause

For emphasis whose content is a whitespace character reference, e.g.

* *

markdown-it decodes the reference to literal whitespace in the text token's
content (" "), keeping the original source only in the token's markup
field. The em/strong renderers emitted that literal whitespace directly
between the emphasis markers, producing * *. On re-parse, a
Unicode-whitespace flanking character can no longer open or close emphasis
(CommonMark emphasis rules), so the formatted Markdown renders to different
HTML than the input and the AST safety check aborts with an error.

Reproduction

On the pristine master clone (mdformat 1.0.0):

$ printf -- '* *' | mdformat -
Error: Could not format "-".

Formatted Markdown renders to different HTML than input Markdown. This is a bug
in mdformat or one of its installed plugins. ...

Internally (validation disabled) mdformat produced \* *, whereas:

• input * * renders to <p><em> </em></p>
• output \* * / * * renders to <p>* *</p> (no emphasis) — hence the mismatch.

With the fix:

$ printf -- '* *' | mdformat -
* *
$ printf -- '* *' | mdformat -
* *

Both round-trip to the same HTML as the input.

Fix

In em() and strong() (src/mdformat/renderer/_context.py), re-decimalify
leading/trailing Unicode whitespace of the rendered inner text using the
existing decimalify_leading / decimalify_trailing helpers. This mirrors
exactly what paragraph() already does for whitespace at paragraph
boundaries, so boundary whitespace is kept in character-reference form and the
emphasis markers stay functional. The change is localized to the two emphasis
renderers; non-whitespace content is untouched.

Tests

Added two round-trip cases to tests/data/default_style.md (driven by
tests/test_style.py, which runs the real CLI including the AST check):

• * * → * *
• ** ** → ** **

Confirmed these fail on unmodified code (the CLI exits non-zero via the AST
check) and pass with the fix. The full test suite passes locally
(4286 passed, 1 skipped), including all CommonMark spec fixtures, and
black / isort / flake8 report no issues on the changed file.

Disclosure: prepared with AI assistance; reviewed and verified locally.