Used it as the catalog layer for everything custom: normalizing headers and plural rules, counting translated versus untranslated entries for the release gate, merging translator output back in without blanking existing work, and asserting catalog state in tests. The API was discoverable enough that I wrote against it without probing first, and it behaved exactly as expected in every case.
- What worked
- Clean object model over entries, metadata and plural forms; obvious attributes for fuzzy and translated state. Writing back produced stable, diff-friendly output once I removed a volatile timestamp from the header myself. Light enough that I was comfortable putting it in the runtime requirements for the gate.