[NSchatz]

What

Changes that make the sharded weekly mutation run safe for a library to adopt. Each change and the test that covers it:

1. The tool is checked before a sharded run. With more than one shard, each shard has a new step right after the install, "Check that cosyte-mutate can shard". It runs <mutate-cmd> --help with CI=false and fails within seconds unless every help it printed names COSYTE_MUTATE_SHARD. The error says sharding needs @cosyte/vitest-config 0.1.5 or later, or shard-count: 1. Before this, a tool that does not shard ran the whole suite on every shard and only then failed. Tests: test/testing-standard-jobs.test.mjs, "the probe asks the mutate command for its help...", "the probe passes the stand-in that shards and refuses the stand-in of 0.1.4..." (this one also covers several packages where one of them is old), and "with one shard the probe asks nothing".
2. One shard works with any tool. With shard-count: 1, a mutation.json written during the run (newer than a marker touched before the command) is read as the report of shard 1 of 1. A mutation.json left from before the run is never read. With more than one shard, mutation.json is still refused. Tests: "with one shard, the mutation.json a tool that does not shard writes is the report of shard 1 of 1", "with one shard, a mutation.json the run did not write is not read...", "with several shards, a mutation.json is never read as a shard report". On a real runner, self-check has a new weekly-mutation-one-shard job that runs the stand-in as 0.1.4, and quality-artifacts checks its merged report and note.
3. The cancel comment is corrected. The save step's comment (and the README) no longer say that a shard stopped at its time limit saves its results. GitHub cancels that job, and a step with !cancelled() does not run in a cancelled job. Test: "the save step says what it does not save...".
4. Restore across two runs. quality-artifacts has a new step, "The first weekly run started from an earlier run's results". For each shard of the first weekly call, it reads the run id in the restored cache key and requires that an earlier run saved it (an earlier attempt of the same run also counts). Test: "self-check proves each shard of its first weekly run restored an earlier run's results", which runs the step's script against made-up notes.
5. The README no longer names an internal path or id. The dash gate section's last paragraph is rewritten. A new test/public-surface.test.mjs reads every tracked file for the planning folder and its id shapes.
6. The gate test's commit email has no domain. test/gate-workflows.test.mjs now uses gate-tester. The shared PHI scan flags any address with a domain (example.com included; it refused that on the first push), and a domainless value passes. Proof: the pre-push scan passed on this branch.
7. The README's "What the real tool still needs" section is updated. It says sharding has landed in cosyte/config and ships in the next patch release of @cosyte/vitest-config, 0.1.5, which is not on npm yet. The sections "The sharded weekly run" and "The testing standard's jobs" describe the check, the one-shard path and the cross-run proof.

The stand-in cosyte-mutate in test/fixtures/quality-stubs/ now answers --help the way the real one does (stderr and exit 2 under CI). With COSYTE_MUTATE_STUB_AS=0.1.4 it plays the release that does not shard. Tests are in test/quality-stubs.test.mjs. The new self-check job is added to the two pinned context lists. In CHANGELOG, the lines for the six references published since the last heading are moved under their own headings, and a new Unreleased entry is added.

Why

Before any library adopts the sharded weekly run, a few problems needed fixing. A red weekly run makes the release mutation gate, which is on by default, refuse every release of that library. The old setup could go red without a useful message: an old tool ran for hours on every shard and then failed. The fix also gives a small library a way to adopt with today's released tool, and makes the comments and README match what the workflow actually does.

Proof

• harness gate (actionlint 1.7.12 with shellcheck, node --test "test/*.test.mjs", check-no-emdash, package-lists): passed.
• Red first: before the workflow and stub changes, the new tests in test/testing-standard-jobs.test.mjs and test/quality-stubs.test.mjs failed (13 failures). With the README from the base, test/public-surface.test.mjs failed on README.md:3303 (a planning path and a standards id). All of them pass now.
• The pre-push PHI scan refused gate@example.com and passed gate-tester.

Choices

• The probe uses the caller's mutate-cmd with --help added, and adds no new input. This works for pnpm exec cosyte-mutate full and pnpm -r exec cosyte-mutate full. With more than one shard, a command that cannot take --help at its end fails the check. To undo or widen this, add an optional probe-command input.
• The probe reads the help text, not a version number. That way it checks what the tool does, not its package version. To undo, compare the installed @cosyte/vitest-config version against 0.1.5 instead.
• The cancel comment is corrected, and saving on cancel is not added. A job cancelled at its limit runs no !cancelled() step, and we have not measured whether StrykerJS leaves a usable incremental file when it is stopped partway. To change this, bound the command inside the step with timeout so the job is not cancelled.
• The cross-run check is strict. If no earlier entry exists (a fresh fork, or entries GitHub evicted), quality-artifacts is red once, and the next run is green. quality-artifacts is not a required context. To undo, make that step warn instead of fail.

Kind: code