I don't like those kind of benchmarks too: you tune your product to maximum and don't touch your competitor at all.
That's why the mentioned benchmarks imply NO TUNING or very light one like enabling mlock in Elasicsearch / enabling secondary indexes in Manticore. Otherwise it works as it's shipped. If you read the articles here https://db-benchmarks.com/posts/ you'll understand that the authors really wanted to do fair benchmarks and to use as little tuning as possible.
> ES has a proven history of horizontal scaling (throw more servers to scale). And I didn't see any architectured benchmark.
This is true. ES is better in terms of this. Manticore has replication and distribute indexes, but doesn't have automatic sharding and shards orchestration. This is a work in progress.