What each code system costs to build and serve, and how fast it answers

The harness (ferroterm-bench, in the repository under tools/) builds the release into an index with ferroterm-build, timing the build and reading its peak resident memory from the operating system; starts the server over the index and times the wait until /health answers; reads the server's resident memory after start and again after the warm requests; and, for each operation, sends one cold request followed by the warm requests, timing every HTTP round trip from the same machine. The latency columns are the cold request and the nearest-rank p50, p95, and p99 of the warm requests.

A request that answers anything other than a 2xx status fails the run for that system and no record is written, so a record never holds the latency of an error response. Every record states the machine it was taken on, and whether that machine was a container; the repository carries one (bench/compose.yaml), which builds the server, the build tool, and the harness on digest-pinned images and mounts the licensed releases read-only, but a record is only from it when the record says so. Resident memory is what starting the server loads: the hierarchy, the text index, and the other structures it reads whole, which is why the figure when the server reports ready and the figure after the warm requests are close. Units fit the value: seconds, milliseconds, or microseconds; GB, MB, or KB.

The repository ships no code system content. To take a record over the releases you are licensed for, follow the benchmarks chapter of the book; a record you take on your hardware is the number that matters for your deployment.