What every repository is held to

The same conformance standard applies across every organisation in the map. That is what makes them one ecosystem rather than a pile of unrelated modules: a caller can assume the same things about any of them without reading its CI configuration.

CGO_ENABLED=0

No cgo, and no shelling out to a command-line tool in place of a library. A dependency is a Go module or it is written.

The rule is stricter than it first sounds, because exec.Command to a CLI is the usual way a cgo dependency comes back in disguise — the binary is still C, it is just C the build no longer sees. A library that runs mkfs.ext4 has not removed the C dependency; it has removed the compiler’s knowledge of it.

What the rule buys, concretely: cross-compilation to six architectures from one machine, a js/wasm build that runs the same code in a browser tab, and a bare-metal target with no libc underneath.

Six 64-bit targets

Build and test on amd64, arm64, riscv64, loong64, ppc64le and s390x — native where runners exist, under qemu-user (Debian trixie, per-architecture QEMU_CPU) otherwise.

s390x earns its place by being big-endian. Any on-disk or on-wire encoding that was quietly written little-endian-first fails there and nowhere else, which is exactly the bug that is invisible on a developer’s machine and fatal in a file format.

100% statement coverage

A CI gate, error branches included.

Two clarifications that matter more than the number:

• A skipped test is not a passing one. Jobs fail when a toolchain or a device is missing, rather than turning green with the interesting case unexecuted. A green run whose -run filter matched nothing is worse than a red one, because it is believed.
• Coverage is measured the way CI measures it — with -coverpkg over the packages under test, not per-package defaults that can report 100% while leaving a package unexercised.

Measured, not asserted

A performance claim comes with a benchmark against the reference implementation — MRI for the Ruby stack, the C library for a codec, numpy or scipy for a numeric kernel. A claim that an implementation is faithful to a format comes with a byte comparison against that reference, not a round-trip through itself.

Decoders are fuzzed. On the filesystem drivers, 70 million executions found two denial-of-service panics that no hand-written test had reached.

Where a claim is about something that runs — a window opening, a widget drawing, a page rendering — it is verified by capture, not by reasoning about the code.

Reference libraries first

Where a good pure-Go library already exists, it is used rather than rewritten. The ecosystem is large because a great many capabilities had no pure-Go implementation, not because everything was rebuilt on principle — and one organisation in it has already been retired for duplicating a reference library that was itself CGO-free.

One landing, one docs site

Each organisation publishes a Hugo landing page at <org>.github.io and versioned MkDocs Material documentation at <org>.github.io/docs/, with the same brand, an 88-pixel hero logo, a favicon, and a light/dark/system theme toggle that defaults to system.

BSD-3-Clause

Every repository, every organisation. Copyright is held by the authors of the organisation the code lives in.

Go 1.26.4 floor

New repositories set the go directive to 1.26.4 — the full patch version, not a bare 1.26. A low go directive is never deliberate; where one is found it is a drift to fix, not a compatibility decision to respect.