See this transcript:

$ cat /tmp/wt
{{1x<invalid>
|{{2x<invalid>y}}|
}}

{{1x<invalid>
|{{2x<invalid>
y}}|
}}

{{1x<invalid>
|{{2x<invalid>
y}}
}}
$ php bin/parse.php < /tmp/tpiwt 
<p data-parsoid='{"dsr":[18,53,0,0]}'>{{1x&lt;invalid>
|{{2x&lt;invalid>y}}|
}}</p>

<p data-parsoid='{"dsr":[55,73,0,0]}'>{{1x&lt;invalid>
|
}}</p>

<p data-parsoid='{"dsr":[75,110,0,0]}'>{{1x&lt;invalid>
|{{2x&lt;invalid>
y}}
}}</p>

So the middle wikitext snippet cause the output to be broken. This is because of the 'splitNlTokens' and associated code in AttributeExpander. The first two snippets lead to the tokenizer creating a <td> with complex attributes ("{{2x<invalid>..."). The last one is a non-attribute td and so doesn't break.

Of the first two snippets, the second snippet expands to a set of tokens that has a newline which triggers the splitTokens code in AttributeExpander which just messes up the <td> token, and when the TSP kicks in later, it drops a bunch of content based on the output of the AttributeExpander.

This effectively is the reason for Parsoid's broken output on https://tpi.wikipedia.org/wiki/Templet:Convert/LoffAonSoff?useparsoid=1