fix(html): keep the whitespace of a whitespace-only inline element

[Lukas (L4XB)]

Closes #1539

The bug

markdownify.chomp() lifts the whitespace surrounding an inline element's text
out of that text, and the conversion functions then return '' once nothing is
left. A whitespace-only element therefore takes its whitespace with it and the
words on either side run together.

input	on main	with this branch
further<strong> </strong>reference	furtherreference	further reference
<b>First</b><b> </b><b>Last</b>	**First****Last**	**First** **Last**
Hello<code> </code>world	Helloworld	Hello world

This is #1539, which was closed pointing at an upstream fix. markdownify 1.2.3
is the current release and abstract_inline_conversion still reads:

prefix, suffix, text = chomp(text)
if not text:
    return ''

so the behaviour is unchanged on main today, with that version installed.

It is not only an HTML-input problem. A word processor keeps a differently
formatted space as a run of its own, so a .docx where the space between two
plain words is bold reaches mammoth as First<strong> </strong>Last and comes
out of MarkItDown as FirstLast. tests/test_docx_inline_whitespace.py builds
exactly that document; three of its four formattings fail on main.

The fix

#2477 solved this for <u> by overriding convert_u. There are twelve more
tags with the same conversion shape, so rather than twelve more overrides this
wraps the function markdownify resolves for a tag in get_conv_fn: a
whitespace-only element returns its text unchanged, everything else takes the
normal path.

b, code, del, em, i, kbd, s, samp, strike, strong, sub, sup, u

a is deliberately not in that set. #2396 is open for anchors and covers more
than the whitespace-only case, including <a> middle </a>, which this branch
does not touch.

get_conv_fn arrived in markdownify 1.0. On an older release the dispatcher
never calls it, so the override is simply never reached and nothing changes.

Tests

• tests/test_html_converter.py gains the whitespace-only case for all thirteen
  tags, the exact-character case (  stays  ), the empty-element
  case, the joined-emphasis case, and controls that an element with real content
  still converts to **word**, `word`, ~~word~~, <u>word</u>.
• tests/test_docx_inline_whitespace.py is new: it builds a minimal .docx
  whose middle run is a single space with bold, italic, underline or
  strikethrough, and asserts the result is First Last.

15 of the new cases fail on main. Seven mutants of the changed lines are
killed by them: the wrapper removed, the guard never firing, the guard inverted,
every element returning unconverted text, and b, i, s dropped from the tag
set one at a time.

packages/markitdown      919 passed, 14 skipped   (main: 892 passed, 14 skipped)
packages/markitdown-ocr   53 passed
packages/markitdown-mcp   21 passed
black 23.7.0 --check      97 files unchanged

Python 3.12, macOS, markdownify 1.2.3.

[Lukas (L4XB)]

@microsoft-github-policy-service agree