ADR-009: Tests are sorted by what they need, and none of them skips

A unit test needs nothing beyond the machine it runs on. task test:unit runs every package under cmd/, internal/ and tools/; the pre-commit hook runs that task, and CI runs it on Linux, Windows and macOS, and under the race detector on Linux. A test that needs more carries a build tag and has a task of its own: live for a real Mattermost, in tests/live (task test:live), and, when the MCP Apps views arrive, views for a browser (ADR-022). Without a tag, go test ./... needs no Mattermost and no browser. task test:live:compile builds the live suite without running it, so a change that breaks it fails in seconds rather than in the live job.

Tests run on every operating system. A test decides by the capability it needs and asserts on every system rather than skipping one by name. A test that cannot reach its dependency fails and says what is missing: the live suite exits before running anything when no Mattermost answers, naming the address it tried and the task that starts one.

Put a test that needs an outside dependency behind a tag and a task, so the fast suite stays fast and needs no setup. Do not call t.Skip on a missing dependency, and do not write a test that runs on one operating system only unless the code it tests is built for that system alone.

A skipped test reads as a passing one: a run that skipped everything reports success having proven nothing, and a skip on one system is coverage that never ran there. mm-mcp ships for all three systems.

Not chosen

• Skip when a dependency is missing: The run stays green on a machine that cannot run the test, so nobody learns that it did not.
• Live and unit tests in one suite: Every commit would wait for a Mattermost, and a contributor without Docker could not run the tests at all.