🧪 test(bench): decide the measurement contracts in peryx-bench-core

[gaborbernat]

.cargo/mutants.toml excludes crates/*/src/bench/** on the grounds that a benchmark workload only builds load, so mutating one moves what CodSpeed measures rather than whether anything is correct. peryx-bench-core sits outside that exclusion and earns it. The crate reduces rounds to a median and a coefficient of variation, ranks parties on a logarithmic colour ladder, converts CPU percentages into seconds, and answers the boolean that gates an A/B comparison. The nightly found 75 mutants alive across its six modules; a run against current main put it at 50 alive, 8 of them cfg twins this host cannot build, plus 4 that finished with no verdict.

None of the six modules only shapes load. stats.rs, report.rs and compare.rs reduce, format and classify, with no I/O. usage.rs reports resident memory and CPU seconds for a process tree, and servers.rs decides when a spawned server is ready. machine.rs is the closest call, since four of its functions exist to build load, but each returns a rate that lands in the published machine.toml, so the arithmetic behind that rate carries a contract like any other. This PR adds no exclusion and proposes none.

Four survivors resist any test as written, so the code changes instead. bytes / workers and each * workers appeared twice each in machine.rs and moved how much load a measurement built while leaving the rate it reported alone; folding them into shares gives the round-down a value a test can name. write_one drove its loop off a running total, so a mutated increment ran forever and no timeout could book it; walking spans off the offsets bounds each write by construction. tree_of walked up from every process and needed a self-parent comparison to terminate, which no test can construct; walking down from the root and taking each parent's children once terminates by construction and visits each process once instead of once per descendant. wait_ready bounded its own poll loop with an Instant comparison that tokio::time::timeout says with no arithmetic to get wrong.

Three more needed the code to hand back what it had already worked out. drain reports the bytes it read, so the loopback rate comes from what arrived. The CPU-percent conversion moved into cpu_millis and the printed change column into describe, since a value that reaches stdout inside a format string and nowhere else is a value nothing can read back. kill_process_group returns the io::Result it was discarding, and Drop goes on ignoring it.

Two more belonged to the harness. A server that never became ready left an accept pending in the fixture thread, and the guard joined that thread forever, so three mutants that should have failed a test hung it instead. The guard now releases the accepts it did not use.

That leaves eight rows in a local missed.txt, all of them the not(target_os = "macos") halves of model, model_at, cores and describe_cores. Rewriting those gates to cfg(all()) and building the host that way settles all eight: the existing suite catches model_at and describe_cores, and model and cores are unviable, because replacing either body orphans its only caller and dead_code fails the build. Nothing else in the crate survives on hardware that can compile it.

Refs #1893

[codspeed-hq]

Merging this PR will not alter performance

✅ 29 untouched benchmarks
⏩ 133 skipped benchmarks¹

---

_{Comparing test/bench-core-mutants-1893 (1967611) with main (34eca99)}

Footnotes

1. 133 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩