benchmark

README.md holds the results and how to run each benchmark; the benchmark skill refreshes it.

Two measures live here, and they answer different questions:

• Wall clock, through core:benchmark: Wado against other implementations. Only a quiet machine can time it, so it runs locally and fills README.md.
• Fuel, through core:fuel: whether a change made Wado's own output do more work. Each benchmark's test "fuel" block measures its phases once. Only -O2 is a performance claim: fuel does not depend on the host, so CI holds every pull request to the -O2 series in benchmark/fuel.json exactly (mise run test-fuel). gh-pages records -O1, -O2 and -O3 per commit, the other two for reference. Fuel leaves out GC and other host work by design: the wall clock is what shows them.

Report a change between two measurements with an improvement as +% and a regression as -%, whichever way the raw number moved.

Each benchmark's #[workload] test trains the profile in its pgo/. wado test --pgo benchmark package-gale-highlight-html records them again.

Read the wado-performance skill before measuring a change, and its section "A/B-ing a stdlib change" before comparing two versions of wado-compiler/lib. Without WADO_BIN, wado.sh runs cargo run --release, and a release build embeds the stdlib. So an A/B that swaps stdlib files into the tree rebuilds the compiler for every run.

Dead ends — a change the benchmark ignored, whose WIR diff shows no fewer instructions, and which does not even come out smaller — live in the wado-performance skill's dead-ends.md. A change that is kept belongs in its commit, not there.