html userdiff is not showing all my changes

> I am attempting to do a word diff of an html source file. Part of the removed 
> html is disappearing from the diff when I enable the fancy html word diff.
> 
> Here's the output from basic `git diff`:
> diff --git a/adv_layout_source.html b/adv_layout_source.html
> index 18a81dd..c4ed609 100644
> --- a/adv_layout_source.html
> +++ b/adv_layout_source.html
> @@ -42,8 +42,8 @@
>        <ul>
>          <li class="ydn-patterns"><em></em><a href="#">ydn-patterns</a></li>
>          <li class="ydn-mail"><em></em><a href="#">ydn-mail</a></li>
> -        <li class="yws-maps"><em></em><a href="#">yws-maps</a></li>
> -        <li class="ydn-delicious"><em></em><a href="#">ydn-delicious</a></li>
> +        <li><em></em><a href="#">yws-maps</a></li>
> +        <li><em></em><a href="#">ydn-delicious</a></li>
>          <li class="yws-flickr"><em></em><a href="#">yws-flickr</a></li>
>          <li class="yws-events"><em></em><a href="#">yws-events</a></li>
>        </ul>
> 
> 
> Here's the default `git diff --word-diff`:
> diff --git a/adv_layout_source.html b/adv_layout_source.html
> index 18a81dd..c4ed609 100644
> --- a/adv_layout_source.html
> +++ b/adv_layout_source.html
> @@ -42,8 +42,8 @@
>       <ul>
>         <li class="ydn-patterns"><em></em><a href="#">ydn-patterns</a></li>
>         <li class="ydn-mail"><em></em><a href="#">ydn-mail</a></li>
>         [-<li class="yws-maps"><em></em><a-]{+<li><em></em><a+} 
> href="#">yws-maps</a></li>
>         [-<li class="ydn-delicious"><em></em><a-]{+<li><em></em><a+} 
> href="#">ydn-delicious</a></li>
>         <li class="yws-flickr"><em></em><a href="#">yws-flickr</a></li>
>         <li class="yws-events"><em></em><a href="#">yws-events</a></li>
>       </ul>
> 
> Which is correct, but less than ideal because it highlights much more than the 
> actual changes.
> 
> So I create a .gitattributes file with one line:
> *.html diff=html
> 
> And rerun `git diff --word-diff`:
> diff --git a/adv_layout_source.html b/adv_layout_source.html
> index 18a81dd..c4ed609 100644
> --- a/adv_layout_source.html
> +++ b/adv_layout_source.html
> @@ -42,8 +42,8 @@
>       <ul>
>         <li class="ydn-patterns"><em></em><a href="#">ydn-patterns</a></li>
>         <li class="ydn-mail"><em></em><a href="#">ydn-mail</a></li>
>         <li[-class="yws-maps"-]><em></em><a href="#">yws-maps</a></li>
>         <li><em></em><a href="#">ydn-delicious</a></li>
>         <li class="yws-flickr"><em></em><a href="#">yws-flickr</a></li>
>         <li class="yws-events"><em></em><a href="#">yws-events</a></li>
>       </ul>
> 
> Yikes! What happened to the second line of changes? The removed code is not 
> displayed at all.
> 
> This is running git 1.7.3.3.
> 
> I suspect the problem is in the html patterns in userdiff.c, but I don't 
> understand the word-diff-regex well enough to fix it.

> Hi Michael,
> 
>> If a wordRegex can make parts of diff disappear than there is problem
>> deeper in the diff machinery.
> It can do exactly that. The word regex determines what is a word, but
> everything else is counted as "whitespace". The word diff view shows
> only differences in words, not in whitespace (which is intentional,
> since whitespace changes in things like LaTeX or HTML are not
> interesting). Note that it doesn't show whitespace _differences_, but it
> does show the whitespace itself (taken from the "new" version of the
> file).

> So, if the word regex somehow doesn't match the second line at all (or
> at least not the differen part), the differences could get ignored.
> 
>> Can you trim this down to a minimal example?
> That would be useful in any case.