HTML is a tree of nodes, before all. It's not just a text.
However, most templating engines handle their input and output as it was just a text; they don't care what happens around their tags, their {$foo}'s and <% bar() %>'s; also they don't care about what are they outputting. Sometimes they happen to produce a correct html, but that's just a coincidence; they didn't aim for that, all they wanted is to replace some funny marks in the text stream with their evaluation.
There are a few templating engines which do treat their output as a set of nodes; XSLT and Haml come to mind. For some tasks, this has advantages: for example, you can automatically reformat (like, delete all empty text nodes; auto-indent; word-wrap). The result is guaranteed to be a correct xml/sgml unless you use a strict subset of operations that can break that. Also, such templating engine would automatically quote strings, differently in text nodes and in attributes, because it strictly knows whether you're writing an attribute or a text node. Moreover, it can conditionally remove a node from output because it knows where it does begin and end, which is useful, and do other non-trivial node operations.
You might not like XSLT for its verbosiness or functionalness, but it's damn helps that your template is xmllint-able XML, and your output is a good sgml/xml.
So the question is: Which template engines do you know that treat their output as a set of correct nodes, not just an unstructured text? I know XSLT, Haml and some obscure python-based one. Moar!