Crate metrale_bench 

2026-09-26: Metrale Engine plugins, and the benchmark suite the met serve TUI and met benchmark drive.

Plugin is the general abstraction: Plugin::load receives a PluginHandle (status, log, progress and glow events, the artifact store, the target endpoint and a cancellation flag). Benchmark specialises it into a drivable state machine: the implementor does one step per Benchmark::next, and Benchmark::run streams the BenchmarkResult frames.

impl Stream and async fn in a trait are not dyn-compatible, so DynBenchmark carries the same contract with boxed futures and is blanket-implemented for every Benchmark; executor::drive consumes it.

Owner: bench. Invariants: none beyond the types.

Re-exports

pub use artifacts::ArtifactStore;

pub use benchmark::Benchmark;

pub use benchmark::BenchmarkDescriptor;

pub use coherence::CoherencePolicy;

pub use dynamic::DynBenchmark;

pub use executor::BenchmarkExecutor;

pub use executor::ExecutorMessage;

pub use executor::RunHandle;

pub use hardware::Hardware;

pub use hardware::HardwareState;

pub use hardware::HardwareStateReport;

pub use hardware::Sensitivity;

pub use history::RunRecord;

pub use history::RunSource;

pub use metadata::PluginMetadata;

pub use params::ParamKind;

pub use params::ParamSpec;

pub use params::ParamValue;

pub use params::ParamValues;

pub use plugin::Plugin;

pub use plugin::PluginEvent;

pub use plugin::PluginHandle;

pub use plugin::TargetEndpoint;

pub use result::Align;

pub use result::BenchmarkResult;

pub use result::Cell;

pub use result::CellStyle;

pub use result::Column;

pub use result::LogLevel;

pub use result::LogLine;

pub use result::ResultTable;

pub use result::RunStatus;

pub use result::Stat;

pub use result::Verdict;

pub use result::VerdictKind;

Modules

artifacts

2026-09-26: The Metrale Engine home (ArtifactStore): the per-plugin artifact and per-benchmark run directories, plus write-if-changed assets and provisioning stamps.

benchmark

2026-09-26: The Benchmark trait, a Plugin driven one step at a time, and BenchmarkDescriptor, its static identity.

benchmarks

2026-09-26: The benchmark drivers, and the text helpers they share.

coherence

2026-09-26: A known-answer probe of the endpoint, run by the executor before a benchmark loads.

dynamic

2026-09-26: DynBenchmark, the object-safe form of Benchmark.

executor

2026-09-26: drive, the loop that turns a benchmark into a stream of frames, and BenchmarkExecutor, which runs a benchmark’s lifecycle as a task on a given tokio runtime.

gate

2026-09-26: The pull-request gate: benchmark records committed beside the code they measured.

hardware

2026-09-26: The hardware fingerprint: which box a benchmark number was measured on, and the box-class key gate baselines are indexed by.

headless

2026-09-26: Driving a benchmark without a terminal: start it on the executor, drain its messages until it finishes, build the run record and save it with crate::history::save, as the dashboard also does.

history

2026-09-26: Run history: RunRecords written by save and read by load, load_all and find, under the artifact store’s runs/<benchmark> directory. A record built by RunRecord::new carries every parameter, the target, the source and the engine version beside the terminal frame. The headless driver and the TUI both save through save.

http

2026-09-26: A minimal OpenAI-compatible client over raw HTTP/1.1 on tokio::net::TcpStream, with no TLS: one streaming chat call, measured (chat_stream), plus the blocking, GET, gap and failure helpers in the submodules. Chunked transfer-encoding is decoded by the private reader module before SSE lines are split, so a chunk boundary inside a data: line does not lose a token.

metadata

2026-09-26: Plugin provenance: who wrote a plugin, where it came from, where to report it. Returned by crate::Plugin::metadata.

param_text

2026-09-26: Text ↔ ParamValues: the conversions a run record (history.rs) and a --param flag need.

params

2026-09-26: Parameter schema: what a benchmark exposes for editing before it runs.

plugin

2026-09-26: Plugin, the abstraction a crate::Benchmark specialises, with its PluginHandle and the TargetEndpoint it drives.

python

2026-09-26: Subprocess runner, Python preflight and venv provisioning. BFCL provisions its venv with these during load(); run is also used to call other tools.

registry

2026-09-26: The benchmark suite, in the order the Benchmarks pane lists it, and lookup by id.

result

2026-09-26: What a benchmark emits on every next(): the frame the TUI renders.

serve_env

2026-09-26: The METRALE_* environment a gate’s server is measured under.

serve_identity

2026-09-26: What makes one met serve “the same server” as another: the bytes of the binary, the arguments it was started with, and its METRALE_* serve levers.