This reminds me of a precompressed template system I built for an old Python Tornado webserver. The system precompressed the static parts of the template, with

This reminds me of a precompressed template system I built for an old Python Tornado webserver. The system precompressed the static parts of the template, with matches only self-referencing inside each span of static content. The dynamic parts of the template were prefixed with 'literal' tags. This meant that serving a gzipped stream was as simple as concatenation, just as it was for normal uncompressed templates. It could have been a lot cleverer - it could have huffman'ed the dynamic parts, for example - but it was an easy win. Later, our Tornado servers were migrated behind a caching proxy and the whole purpose of serving up (partially) compressed pages vanished.