Love this idea! Although unfortunately it ended up tripling the CI run time. It's also currently only checking the circuits used in integration tests, so we're not seeing gate counts for the full set, which would take even longer to process.

One thing that could help: caching gate counts to disk using a filename based on a hash of the flattened Noir code, along with the nargo and bb versions. That way, if nothing has changed we can skip recalculating the gate count and hopefully speed up CI quite a bit.