[mafredri]

make static builds the GitHub Pages files (coverage report and badge) with the new ci/static.sh, and static.yml runs it instead of inline commands, so CI and a local run build the same output. The inline step ran under bash -e without pipefail, so a failing go tool cover would publish a badge without a number; ci/static.sh fails instead.

ci/test.sh collects coverage with -test.gocoverdir, which go tool covdata can merge across separate test runs, and converts it to ci/out/profile.txt.

The arm64 steps in ci/test.sh and ci/bench.sh install qemu-user instead of qemu-user-static. GitHub's ubuntu-latest runners are now Ubuntu 26.04, where qemu-user-static has no installation candidate, so the test and bench jobs failed at that step. qemu-user installs /usr/bin/qemu-aarch64 on Ubuntu 24.04 and 26.04 (checked in containers, and in this PR's CI on 26.04). A separate commit formats ci/bench.sh with shfmt (whitespace only).

AUTOBAHN=1 make test on master and on this branch ran and passed the same 853 tests and subtests, including the 299 Autobahn cases and TestWasm. Coverage was 92.3% on master and 92.5% here, over the same 1284 statements; identical runs differ by more than that (see the decision log). make static writes ci/out/static/coverage.html and coverage.svg. static.yml runs only on pushes to master and on manual dispatch, so the checks on this PR do not run it.

Refs #539

Decision log

• Excluded coverage matches master: stringer.go is still deleted from the profile, internal/test/... is left out of -coverpkg, and master's examples filter matched nothing because internal/examples is a separate module.
• ci/static.sh uses curl -fsSL instead of wget, from the review on ci: remove dev branch jobs from daily workflow #549. ci/test.sh uses pipefail too, because go list feeds -coverpkg through a pipe.
• qemu-user, not qemu-user-binfmt (which apt lists as providing qemu-user-static on 26.04): the scripts run qemu-aarch64 directly, so they do not need binfmt registration.
• Coverage difference: the 3 statements only this branch covered are conn.go:176-179 (a write ctx cancelled during the write) and frame.go:68-70 (a read error after one header byte). Without Autobahn, 8 runs of each command covered 1129 to 1149 statements, and conn.go:176-179 was covered in 2 of 8 runs of each. Most of the spread is in compression code; the tests pick compression modes at random. frame.go:68-70 was covered only in this branch's Autobahn run.

🤖 This PR was created with the help of Coder Agents.

[code-asher]

Awesome looks good to me!