Rather than duplicating FTL code in fixtures_reference and fixtures_structure and testing both separately, we could generate the structure JSONs from the reference fixtures.

fixtures_reference
    *.ftl // copied from projectfluent/fluent
    *.json // copied from projectfluent/fluent
fixtures_structure
    *.json // generated from fixtures_reference/*.ftl

We'd still run two sets of tests:

• Reference tests compare the output of the reference parser (stored in fixtures_reference/*.json) with the current output of the tooling parser, ignoring Spans and Annotations.
• Structure tests compare the expected output of the tooling parser (stored in fixtures_structure/*.ftl) with its current output.