Splitting mixed content nodes on particular regex match with xslt 3

My simplified input looks like this:

<stuff>
    <p>CAPITALWORD is part of <i>mixed</i> content.</p>
    <p>ANOTHER is <i>here</i> but it's not the only one. SOMEWORDS are <i>mixted up</i> in the same
        paragraph. SOMETIMES even <i>multiple times.</i></p>
</stuff>

Now, my goal is to split paragraphs on each full-caps word. I thought I would go for grouping text starting with at least two capital letters like this:

<xsl:output method="xml" indent="true"></xsl:output>
<xsl:mode on-no-match="shallow-copy"/>
    
<xsl:template match="p">
  <xsl:for-each-group select="node()" group-starting-with="text()[matches(., '[A-Z]{2,}')]">
    <xsl:element name="p" >
      <xsl:apply-templates select="current-group()"/>
    </xsl:element>  
  </xsl:for-each-group>
</xsl:template>

but this won't work because I'm dealing with mixed content rather than strings only. So I get this:

<stuff>
   <p>CAPITALWORD is part of <i>mixed</i> content.</p>
   <p>ANOTHER is <i>here</i>
   </p>
   <p> but it's not the only one. SOMEWORDS are <i>mixed up</i> in the <i>same</i>
   </p>
   <p>
        paragraph. SOMETIMES even <i>multiple times.</i>
   </p>
</stuff>

instead of the desired output:

<stuff>
    <p>CAPITALWORD is part of <i>mixed</i> content. </p>
    <p>ANOTHER is <i>here</i> but it's not the only one. </p>
    <p>SOMEWORDS are <i>mixed up</i> in the <i>same</i> paragraph. </p>
    <p>SOMETIMES even <i>multiple times.</i></p>
</stuff>

I will be most grateful for tips on how to achieve the desired output.

One approach is a two step transformation, the first step uses analyze-string on text nodes to wrap your capitalized word into an element, the second step then can easily use group-starting-with on those wrapper elements:

<xsl:stylesheet xmlns:xsl="http://www.w3.org/1999/XSL/Transform" version="3.0"
  xmlns:xs="http://www.w3.org/2001/XMLSchema"
  xmlns:fn="http://www.w3.org/2005/xpath-functions"
  exclude-result-prefixes="#all"
  expand-text="yes">

  <xsl:mode on-no-match="shallow-copy"/>
  
  <xsl:template match="p">
    <xsl:variable name="capitalized-marked-up" as="node()*">
      <xsl:apply-templates mode="markup-capitalized"/>
    </xsl:variable>
    <xsl:for-each-group select="$capitalized-marked-up" group-starting-with="capitalized-word">
      <p>
        <xsl:apply-templates select="current-group()"/>
      </p>
    </xsl:for-each-group>
  </xsl:template>
  
  <xsl:template match="capitalized-word">
    <xsl:apply-templates/>
  </xsl:template>
  
  <xsl:mode name="markup-capitalized" on-no-match="shallow-copy"/>
  
  <xsl:template mode="markup-capitalized" match="text()">
    <xsl:apply-templates select="analyze-string(., '\p{Lu}{2,}')" mode="wrap"/>
  </xsl:template>
  
  <xsl:template mode="wrap" match="fn:match">
    <capitalized-word>{.}</capitalized-word>
  </xsl:template>

  <xsl:output indent="yes"/>

</xsl:stylesheet>

There are basically two approaches to this. One is to convert all the information to a node structure by adding markup to the text, and then process it as a tree of nodes using grouping mechanisms and the like. That's what @MartinHonnen has done. The other is to convert all the information to text, for example by converting <i>italic</i> to {italic}, and then process it using regular expressions (typically xsl:analyze-string), finally converting {italic} back to<i>italic</i> as a post-processing step.

I would usually use the first technique, but if the only markup that occurs within the mixed content is a single element type (i) with no attributes, then you could consider the second.

An xsl:for-each-group is never going to split text up into fragments, which is what you seem to be attempting to do.