Everything hot in wr is measured locally and deterministically: the network
is a loopback httptest server, the fixtures are synthetic articles and
documents, animations are off, no wall-clock or external dependency. Run it
with the Go tooling you already have:
go test -run '^$' -bench . -benchmem # the whole suite
go test -run '^$' -bench Render -benchmem # one area
sh scripts/bench.sh # everything + profiles into bench-out/
The profiles answer “where does it get stuck”:
go tool pprof -top -nodecount=20 bench-out/wr.test bench-out/cpu.prof
go tool pprof -top -sample_index=alloc_space bench-out/wr.test bench-out/mem.prof
The numbers below are the current baseline: scripts/bench.sh baseline
(count=3, medianas) on a Ryzen 5 3600 (6c/12t), Go 1.27.2, Linux, ext4. They
are a comparison line, not a promise: rerun on your machine before and after
any change, and read “The workflow” at the end about what counts as a
regression on a shared machine.