Do not duplicate a non-ASCII space when wrapping

[dylanpulver]

convert_p's wrap branch captures trailing whitespace with a bare rstrip() and re-appends it after fill():

line_no_trailing = line.rstrip()
trailing = line[len(line_no_trailing):]
line = fill(line, ...)
new_lines.append(line + trailing)

Bare rstrip() removes everything str.isspace() accepts, but fill() only strips the ASCII set — so a trailing U+00A0 is left in place by fill() and appended again. One   in, two out.

<p>a&nbsp;<br>c</p>   wrap=False -> 'a\xa0  \nc'
                      wrap=True  -> 'a\xa0\xa0  \nc'

What I ran. With a wrap_width too wide to wrap anything, wrap=True should equal wrap=False. 96 documents — 8 trailing characters × 3 bodies × 2 newline_style × with/without <br> — compared between the two settings, same command on develop and on the patch:

	mismatches
develop (bc98c9e)	18/96
patched	0/96

All 18 are duplications, in six characters: U+00A0, U+202F, U+2007, U+2003, U+2009, U+3000. ASCII space and the no-trailing-character case are the controls and match in both trees. pytest 83 passed before and after; the added assertion fails on unpatched source. flake8 --ignore=E501,W503 clean.

The alternative I rejected. Wrapping line_no_trailing instead of line also gives 0/96, and it's arguably the more obvious reading. It differs once the line actually wraps: at wrap_width=8, <p>aaa bbbb&nbsp;<br>c</p> gives

• this patch: aaa\nbbbb\xa0 — longest content line 5
• that one: aaa bbbb\xa0 — longest content line 9, over the requested width

because it hands fill() a string with the U+00A0 already removed, so the width calculation doesn't see it. A no-break space occupies a column like any other, so I kept it inside the fill() input. Happy to switch if you'd rather.

Deliberately not touched. The rstrip() in process_text is the other half of the pair #188 converted, and #188 said explicitly that the trailing side is an open question — "or, conversely, where trailing such spaces could safely be stripped even if they no longer are after this change" — so that one is yours to decide, and this change doesn't depend on it. This also does not fix #175: that report goes through chomp(), which is untouched here (<div>this is a <i>test&nbsp;</i>with whitespaces</div> is unchanged by this patch).

AI assistance: found and patched with Claude Code (Claude Opus 5, claude-opus-5); I reviewed and ran everything above myself.