Publishes to npm from a version tag push

[johnnyt]

What

The npm publish moves from a person's hand to a GitHub Actions workflow.
.github/workflows/release.yml publishes @riddler/predicator when a
v*.*.* tag is pushed, and only when every check before the publish passes
at the tagged commit:

• the tagged commit is an ancestor of the default branch's head, the branch
  named by the push event (github.event.repository.default_branch), never a
  literal name;
• the tag without its v equals .version in package.json at that commit;
• npm does not already show that version (a version already published is
  reported, never published again; this also answers a re-run of a run that
  did publish);
• the full quality gate is green: the toolchain, cache, dependency and gate
  steps are copied byte for byte from ci.yml, and the gate command is read
  from gate.full in .claude/wurk.json as CI reads it.

The publish runs under npm trusted publishing: the job's permissions are
contents: read and id-token: write, no token exists in the file, the
repository's secrets or anywhere else, and npm adds provenance itself. Before
the publish, node and npm are checked against trusted publishing's floor and
npm is upgraded only when it is below it. npm publish runs under
mise exec --, so the existing prepack publish guard runs unchanged. The
run ends by printing the published version's address. Nothing else triggers
the workflow: no branch push, no pull request, no manual dispatch.

CLAUDE.md's bold npm publish row, two clauses of its relay paragraph
and the publish sentence of its Release preps paragraph, and the publish
sentences of .claude/wurk/release.md, now say the same thing: an agent or a
session never runs npm publish; the release workflow publishes on the tag
push the release-prep row already allows; a failed workflow is re-run from its
Actions page, never worked round by a local publish. The rewording is the
maintainer's own words, ruled by the operator, 2026-10-04.

ADR-0006 records the decision at proposed: the three conditions, the
registry, the trust model by name, what a failed publish does, the
documentation decision and that a published version stands. It also says a
re-run of a failed publish repeats the checks and the gate, and that a
prerelease tag is not a case it decides. It stays proposed
until a version has been published through the workflow.

No file under src/ changes, and no version or tag moves.

Checks

• Full gate (mise exec -- pnpm run gate) green locally on this tree.
• actionlint was not run: it is not installed on the machine this was
  written on. Instead the file was parsed as YAML, every run: script passed
  bash -n, and the four check steps were exercised under bash in a scratch
  clone at the v0.6.0 tag with the step's environment set by hand: the
  ancestry and version checks passed, the registry check stopped with
  "already on npm", a mismatched tag name stopped the version check, an extra
  commit off the default branch stopped the ancestry check, and an empty
  default-branch name stopped the fetch. No step from the toolchain on, and
  no publish, has run.
• The copied block was compared with ci.yml and is identical.

Direction check of the record

Every claim in ADR-0006 was checked against main and this branch:
package.json's repository.url, files, publishConfig and prepack
script; scripts/publish-guard.mjs as the guard prepack runs;
.github/workflows/ci.yml's toolchain-to-gate steps; gate.full in
.claude/wurk.json; and each workflow step the record names, cited by its
step name because the workflow exists only on this branch. The record cites no
commit. git diff origin/main -- docs/adr/ removes no line: the record is a
new file and the index gains one row.

Provenance

• Engineering choices (the workflow runs the gate itself, the default branch
  read from the event, the version read from package.json at the tag, the
  toolchain copied from ci.yml, npm trusted publishing, one record, no manual
  dispatch): decided by the conductor under a standing consent, 2026-10-03.
• The failed-publish handling and the documentation decision: decided by the
  conductor under a standing consent, 2026-10-04.
• The npm upgrade is guarded rather than unconditional: the node mise.toml
  pins already bundles an npm above the floor, so an unconditional upgrade
  would only add an unpinned major to every release; the floor is still
  enforced. Decided by the conductor under a standing consent, 2026-10-04.
• The registry check before the toolchain is the implementation of "a version
  the registry already shows is reported, not re-published".
• concurrency is set at the workflow's top level over its single job, keyed
  by the tag ref, with cancel-in-progress: false.
• Beyond the named row and sentences, more text moved because the change made
  it false or ambiguous: in CLAUDE.md's relay paragraph, "publishing" in
  the list of what stays forbidden now reads "running npm publish", and its
  last clause names the workflow as what publishes; and the closing publish
  sentence of .claude/wurk/release.md. The ADR
  index in docs/adr/README.md gains the new record's row, as every new
  record here has.
• No changelog fragment: changelog.d/README.md excludes documentation, ADRs
  and CI changes.