Using LibreOffice I whipped up the following minimal example, edited down to the minimum needed:

<?xml version="1.0"?>
<office:document-content office:version="1.2"
    xmlns:office="urn:oasis:names:tc:opendocument:xmlns:office:1.0"
    xmlns:text="urn:oasis:names:tc:opendocument:xmlns:text:1.0">
<office:body><office:text>
<text:p text:style-name="P1">
    Fo<text:span text:style-name="T1">o<text:line-break/>
    B</text:span><text:span text:style-name="T3">a</text:span>
    <text:span text:style-name="T5">r<text:line-break/></text:span>
</text:p>
</office:text></office:body>
</office:document-content>
[download]

The transform needed of the XML document is something like this: <r><p>a<x>b<s/>c</x>d</p></r>

<r>
   `-- <p>
          `+- a
           +- <x>
           |     `+- b
           |      +- <s>
           |      `- c
           `- d
[download]

To this: <r><p>a<x>b</x></p><p><x>c</x>d</p></r>

<r>
   `+- <p>
    |     `+- a
    |      `- <x>
    |            `-- b
    `- <p>
          `+- <x>
           |     `-- c
            `- d
[download]

Where I think the elements surrounding <s/>, represented here as <x>, could even be nested more than one level, i.e. maybe <r><p>a<x>b<y>c<x>d<s/>e</x>f</y>g</x>h</p></r>

A very interesting problem, unfortunately I don't have enough time to invest at the moment. You definitely shouldn't try to do this with regexes. If the files aren't too big, I'd probably approach this with XML::LibXML...

Update 2: On second thought, a stream-based parser might be easier in this case... hmmm...

use warnings;
use strict;
use XML::LibXML;

my $dom = XML::LibXML->load_xml(string => <<'END_XML');
<?xml version="1.0"?>
<office:document-content office:version="1.2"
    xmlns:office="urn:oasis:names:tc:opendocument:xmlns:office:1.0"
    xmlns:text="urn:oasis:names:tc:opendocument:xmlns:text:1.0">
<office:body><office:text>
<text:p text:style-name="P1">
    Fo<text:span text:style-name="T1">o<text:line-break/>
    B</text:span><text:span text:style-name="T3">a</text:span>
    <text:span text:style-name="T5">r<text:line-break/></text:span>
</text:p>
</office:text></office:body>
</office:document-content>
END_XML

my $xpc = XML::LibXML::XPathContext->new($dom);
$xpc->registerNs('office',
    'urn:oasis:names:tc:opendocument:xmlns:office:1.0');
$xpc->registerNs('text',
    'urn:oasis:names:tc:opendocument:xmlns:text:1.0');

while (1) {
    my ($lb) = $xpc->findnodes('//text:p//text:line-break') or last;
    die "can't handle <text:line-break> with children: $lb"
        if $lb->hasChildNodes;
    my ($p) = $xpc->findnodes('ancestor::text:p[1]',$lb)
        or die "failed to find <text:p> ancestor of $lb";
    print "$p\n"; #DB
    # ... do something useful here ...
    $lb->unbindNode; # so this loop terminates
}

print "#####\n";
print $dom;
[download]