Files
poli0981andClaude Opus 5.5 ffceb4bd54 fix: review findings — separators are not items, Replace counts what it deletes
An adversarial review of the Firefox port found:

- #edit treated a Firefox separator as an item in three places: a folder's
  count badge, whether a folder can be expanded, and whether it can be
  sorted. A folder of separators and one bookmark offered Sort A-Z.
- Replace's confirmation counted every writable root, so a signed-in
  Chrome's account roots, which Replace never empties, were counted among
  what it would delete. In every release since 1.1.0; it now counts the
  replaceable roots only.
- The store jobs checked their files against the SHA256SUMS.txt that came
  in the same artifact, with --ignore-missing. Each file is now checked by
  name against the published Release's checksums.
- Qodana's publish job ran only after a successful scan, so a scan failing
  on findings never published them.
- getRoots' position fallback had no test. It now has two, with unknown
  ids and a Bookmarks Menu listed first.

Docs follow: the two browser modules and the warning codes (02), where the
landed index is compared (05), the separator markup and folder count (06),
what the Firefox suites are (11), the release staging and environments
limited to v* tags (12, 00), and which loose entries reach the Bookmarks
Menu on Firefox (03, 04, README).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-18 22:10:35 +07:00

28 KiB
Raw Permalink Blame History

12 — CI/CD

Integrates with the centralized ops repo poli0981/.github, whose reusable workflows use the reusable-* naming family. There is no browser-extension-* family — the single Chrome-extension workflow is reusable-chrome-extension.yml.

1. Known constraints (from the ops-repo migration)

  • poli0981 is a user account → no org-level secrets; repo-level Actions secrets only.
  • Caller stubs MUST declare explicit permissions: blocks — reusable workflows do not inherit them (bug already burned once; never omit).
  • New repo joins in side-by-side mode is unnecessary (greenfield) — adopt the centralized workflows directly from day one.
  • reusable-chrome-extension.yml runs no quality gates. Its complete workflow_call surface is inputs package-manager (default 'none'), node-version ('24'), build-command ('npm run build'), source-dir ('.'), release (false) — and no secrets: block at all, so passing any named secret to it is a workflow-compile error. It installs, builds, asserts manifest_version == 3, zip -rqs source-dir and uploads an artifact. Lint, type-check, knip, tests, coverage and the grep gates are the caller's job (§2.1).
  • Sibling precedent for the same stack is poli0981/switch-every-tab-hotkey, which currently keeps its CI self-contained rather than calling the ops reusables. Do not assume its workflows are a drop-in template.

2. Workflows (caller stubs in this repo)

This chapter no longer copies the workflow YAML (decided 2026-10-28). It used to, and the copies drifted three times: a Dependabot bump moved actions/setup-node v6 → v7 in .github/workflows/ and never touched this doc, the check:manifest step added to the release job in 674ac34 was missing here, and a copy kept continue-on-error after the file had dropped it. Because docs win over code in this project, a stale copy is a standing invitation to "reconcile" a pinned action backwards or delete a security gate. The workflow files are the ground truth about themselves. What follows is what a YAML file cannot say: why each job exists, what it must keep doing, and the traps already stepped in.

Common to every job since v1.1.0:

  • actions/checkout with persist-credentials: false — no job pushes, so no job leaves a token in .git/config.
  • actions/setup-node with node-version-file: .nvmrc (Node 26, 01 §2). The ops reusable takes a string input and cannot read the file, so ci.yml passes it '26' — keep the two equal by hand.
  • npm ci --ignore-scripts, then npx --no-install wxt prepare. No dependency gets to run code at install time (SonarCloud S6505); our own postinstall is the one script we need, so it runs explicitly, from the local binary, never fetched. A local npm ci still runs postinstall normally. One exception we do not control: the build job delegates to the ops reusable, which runs a plain npm ci, so install scripts do run there. That job holds a read-only token and no secrets; an ops-repo input to skip install scripts is on the 15 backlog.
  • Reusable workflows get named secrets, never secrets: inherit: they are pinned to the ops repo's moving main, so each one receives only what it declares.

2.1 ci.yml — every push/PR to main

Three jobs. The two whose names branch protection requires (§5) are Quality gates and Build + package / build.

  • quality — "Quality gates". Local, because the ops reusable runs no quality gates at all (§1). In order: lint, check (svelte-check on the TS 6 API — failing on warnings too since v1.2.1, which is what makes Svelte's a11y_* checks a gate, 10 §3 — then the TS 7 tsc), knip, coverage (11 §1 thresholds), guard (08 §3), check:licenses (every licence in the lockfile, 09 §3.2), npm audit --omit=dev --audit-level=low (the zero-runtime-dependencies claim, 09 T5), build, check:manifest, and since v1.3.0 build:firefox + check:manifest -- --browser firefox — the second package, held to its own key set (08 §1, 08 §9). The build/check pairs exist because the unit test over wxt.config.ts proves what we asked for and only the built manifest proves what WXT emitted — a dev build of the same config carries tabs/scripting/host_permissions.
  • windows — "Windows smoke (guard + unit tests)". Not required. The owner develops on Windows, and the failures that only happen there — drive letters and percent-encoding in script paths, CRLF checkouts — had no CI coverage.
  • build — "Build + package". Delegates packaging to reusable-chrome-extension.yml with package-manager: 'npm', node-version: '26', build-command: 'npm run build', source-dir: '.output/chrome-mv3', release: false. All four are required in practice: leaving package-manager at its 'none' default skips install and build, then fails on test -f ./manifest.json, a file WXT never writes to the repo root.

⚠️ contents: write on the build job is not a mistake. The reusable declares two mutually exclusive jobs — build (contents: read, if: !inputs.release) and release (contents: write, if: inputs.release). GitHub validates the caller's grant against every job the reusable declares, before evaluating any if:, so release: false does not exempt the caller from the release job's requirement. Granting contents: read makes the entire run fail with startup_failure: no jobs, no log, and only the generic "this run likely failed because of a workflow file issue" in the UI. Verified 2026-07-26 on the first real run of this workflow.

npm run guard runs the same script locally and in CI, so the gate cannot drift between the two (10 §3).

2.2 codeql.yml

Calls the current reusable, reusable-codeql.yml (CodeQL Action v4). Its input is languages — a JSON array encoded as a string ('["javascript-typescript"]'), not the singular language of the older codeql.yml. Weekly schedule plus push/PR to main; the permission block (security-events: write, actions: read, contents: read, packages: read) is repeated on the job because reusable workflows do not inherit it.

⚠️ Prerequisite: repo Settings → Advanced Security → CodeQL "Default setup" must be DISABLED, or this advanced-setup workflow is rejected at SARIF upload time.

2.3 release.yml — on tag v*

Standalone, not the ops reusable: the reusable's release job only does gh release create --generate-notes. It produces no SHA256SUMS.txt (required by 09 §5), performs no store submission, and exposes no outputs:, so a caller cannot reference the zip it built.

Jobs, least privilege first (split 2026-10-28; the store job became three in v1.3.0):

Job Token Does
build "Gate + package" contents: read install, every CI gate again (a tag skips the branch triggers, and must never ship what a PR would have blocked), zip and (since v1.3.0) zip:firefox — the Chrome/Edge package, the Firefox package and its sources archive — check:manifest on both packages, the tag ↔ manifest version check on both, SHA256SUMS.txt over all three files, each staged by its exact name so a missing one fails the run (a glob would pass with one of three); copies the three files and SHA256SUMS.txt into a plain dist/ and uploads that as the artifact release-<tag> — upload-artifact skips hidden paths such as .output/ by default, which failed the first v1.1.0 tag run before any Release existed
publish "GitHub Release" contents: write downloads the artifact, sha256sum -c it, creates the Release with --verify-tag (or re-uploads to an existing one). Installs nothing and runs no repository code — the only job that ever holds a write token
store-chrome, store-firefox, store-edge contents: read since v1.3.0, one job per store, each only when its own repo variable is 'true' (CWS_AUTOPUBLISH, AMO_AUTOPUBLISH, EDGE_AUTOPUBLISH) and each in its own GitHub Environment holding only that store's secrets (§3). They install tools/publish/ — publish-browser-extension and its three dependencies, from their own lockfile, scripts off — instead of the whole dev toolchain, check the downloaded files by exact name against the SHA256SUMS.txt of the published Release (not the artifact's own copy), and run its publish-extension binary. Before v1.3.0 a single store job ran wxt submit after a full npm ci: four Chrome secrets shared a runner with ~320 packages
announce contents: read Discord, only when DISCORD_ANNOUNCE is 'true'

Before the split, one job ran npm ci — every dependency's install script plus our own tests — while holding contents: write, with the token also left in .git/config by checkout.

The Release step is idempotent deliberately. gh release create fails outright when a Release for the tag already exists, so any failure after it — realistically the store job, once CWS_AUTOPUBLISH is on — stranded the tag: a re-run could not get past that line, and recovery meant deleting the Release and the tag by hand. docs/13 §7 documents the manual escape; this removes the need for it in the common case.

The announcement is gated because the reusable exits 1 on "No webhooks configured". On the real v1.0.0 run the Release, the zip and SHA256SUMS.txt all published correctly and the run still showed red. A red run on a good release trains you to ignore the colour, so the notification must not be able to fail the release.

⚠️ It is gated with if: on a repo VARIABLE, not with continue-on-error. GitHub rejects continue-on-error on a job that calls a reusable workflow (Property continue-on-error is not allowed), and a rejected file does not fail one job — the whole run is a startup_failure with zero jobs and no log. That sat undetected from 5f4c6e7 until 2026-08-03 because the commit landed after the v1.0.0 tag and no tag was pushed in between. A variable and not a secret: secrets cannot be referenced in a job-level if:.

⚠️ The webhook secrets are mapped by name. This repo stores RELEASES_WEBHOOK and REPO_WEBHOOK; the reusable reads DISCORD_RELEASES_WEBHOOK and DISCORD_REPO_WEBHOOK. Under the old secrets: inherit the reusable received neither, so the day DISCORD_ANNOUNCE was set the job would have failed on "No webhooks configured". It now passes each one explicitly under the name the reusable reads — found 2026-10-28 by comparing the repo's secret list with the reusable's workflow_call block.

tag_override is required: the reusable resolves TAG="${INPUT_TAG:-$EVENT_TAG}" with EVENT_TAG = github.event.release.tag_name, which is empty under on: push: tags, so without it the job exits 1 on "No release tag available". The announcement is a job inside this workflow, not a separate on: release: [published] caller, because a Release created by gh release create under the default GITHUB_TOKEN emits no event that starts new workflow runs.

2.4 notify-ci-failure.yml

Runs on workflow_run of CI, gated on conclusion == 'failure' && vars.DISCORD_CI_NOTIFY == 'true', and passes the one secret the reusable declares (DISCORD_CI_WEBHOOK) by name. The reusable declares that secret required: true and GitHub validates it before any of the reusable's own logic, so without the gate this caller failed on every CI completion, success included. DISCORD_CI_WEBHOOK exists (set 2026-07-26); the variable is still unset, so notifications stay off until the owner opts in. A notification must never be able to change the colour of the thing it reports on.

2.5 sonarcloud.yml — SonarCloud CI analysis

Added 2026-10-28, replacing SonarCloud's Automatic Analysis so the project gets coverage and a scan of exactly what CI builds. Project poli0981_bookmarkmagic, organisation poli0981; settings live in sonar-project.properties.

  • A separate workflow, not a job in ci.yml: a Sonar outage or a quality-gate opinion must never turn CI red. The verdict reaches PRs as the SonarCloud GitHub App's own check. It is not a required check.
  • Two jobs, so the token never shares a runner with dependency code. A job's secrets are loaded into its runner for the whole job, and any step can read them from the runner process whatever its own env: says (the shape of the March 2025 tj-actions compromise). coverage does the common install and npm run coverage with no secret, and uploads coverage/lcov.info and the generated .wxt/ (which tsconfig.json extends, and which defines the @/ alias) as a one-day artifact; scan installs nothing — full-history checkout (fetch-depth: 0, for new-code detection), the artifact, then the scan with -Dsonar.projectVersion=<package.json version>, so "new code" means "since the last release".
  • The price of that isolation: the scan job has no node_modules, so Sonar's type-aware rules resolve our own modules (through .wxt/) but not the dependencies' types.
  • The scan action is pinned to a commit (v8.4.1): it is the step that holds SONAR_TOKEN. Dependabot and fork PRs get no repository secrets, so the scan is skipped with a notice instead of failing.
  • Coverage: vitest.config.ts measures every src/**/*.ts module and reports text + lcov; the CI gate is still the core-only glob threshold of 11 §1. sonar.coverage.exclusions drops components, entrypoints and scripts — the same split as 11 §1, where those are covered by the manual pass and the script tests. The three locale files are excluded from duplication detection: they are parallel by design.
  • No sonar.qualitygate.wait: an unreviewed Security Hotspot can only be cleared by a person in the Sonar UI, and a CI step that waits on a human is a step that is red for a reason nobody can fix from the repo.

⚠️ Prerequisite: Automatic Analysis must be off in the SonarCloud project (Administration → Analysis Method). With it on, the scanner refuses to run.

Snyk is connected only as a GitHub App (the security/snyk (poli0981) PR check); there is no SNYK_TOKEN and no Snyk step in any workflow.

2.6 triage.yml — maintenance-mode triage

Added 2026-11-18 with /MAINTENANCE.md. On issues: [opened, labeled] and pull_request_target: [opened] it runs scripts/triage.mjs, which reads the runner's event file and prints a decision, then carries it out with the gh CLI the runner ships — no third-party action ever holds its write token (issues: write, pull-requests: write, contents: read).

  • The decision is code with tests, not shell: decide() is a pure function (tests/unit/scripts/triage.test.ts) — a feature-form issue answering "Something new" is labelled maintenance-mode, commented and closed not planned; the two in-policy answers get needs-triage; an enhancement label the owner adds by hand closes the issue the same way; an outside pull request titled feat… is closed. The owner's own issues, members, collaborators, bots and anything labelled roadmap are exempt. The test also pins the dropdown's wording against feature_request.yml, which is the only thing the decision parses.
  • ⚠️ pull_request_target holds a write token for fork pull requests. The checkout is the base branch — the event's default — and nothing from the pull request runs. Titles and bodies reach the script as the event file and are only compared with constants; the shell reads back an action from a fixed set, an integer, a label from a fixed list and a comment the script wrote, and validates the first two before using them.
  • Idempotent. The feature form fires opened and labeled for one new issue. concurrency runs them one at a time per number and the second re-reads the live labels and state before acting.
  • The labels must exist (§5); gh … --add-label fails on a missing one.
  • Order: comment and close first, label last. The label is what a re-run reads as "done", so it must not exist before the work it stands for.
  • A hand label closes only when the owner adds it. An enhancement issue with no answer to the form's question is the owner's one-touch close when the owner labelled it; from anyone else, or on opened, it goes to needs-triage instead of being closed on a guess.
  • actions/checkout and actions/setup-node are pinned to commits here, unlike the other workflows, because this one runs them with a write token on events anyone can trigger; and setup-node has no cache:.

2.7 qodana.yml — Qodana for JS

Added 2026-11-18. qodana-js (JetBrains' JavaScript/TypeScript linter, which needs the Ultimate licence QODANA_TOKEN carries) on push to main, pull requests, a weekly schedule and workflow_dispatch. Configuration is qodana.yaml: native mode (withinDocker: false), profile qodana.recommended, fixtures and generated output excluded.

  • Its own workflow, never required, for the reason Sonar has one (§2.5).
  • Two jobs. scan has a read-only GITHUB_TOKEN, installs the dev dependencies (install scripts off) because qodana-js resolves types through node_modules, runs the action with PR comments and annotations off, and keeps the SARIF report as the qodana-report artifact. publish installs and runs nothing: it uploads that SARIF to code scanning (category: qodana), so the findings sit beside CodeQL's under Security and quality and on pull requests.
  • ⚠️ The documented exception to §2.5. Sonar's token never shares a runner with dependency code; Qodana's cannot avoid it, since the analysis needs the dependencies. The token only uploads reports to one Qodana Cloud project. If a dependency is ever found compromised, rotate it in qodana.cloud.
  • The action is pinned to the commit of v2026.2.2, not the vendor's moving v2026.2 branch: it holds a token, and the newest tag was younger than the 7-day cooldown when it was added. Dependabot proposes updates.
  • pr-mode: false: the whole project every time. It is small, and a pull request that touches only docs would otherwise scan nothing.
  • Dependabot and fork pull requests have no secrets; the scan is skipped with a notice and publish does not run.
  • The report artifact is a zip inside the artifact, with absolute entry names; publish extracts the one entry /qodana.sarif.json by name. A glob for the file found nothing on the first run.
  • First run (2026-11-18): 27 findings, none a defect. Three regular expressions that could be simpler — fixed — and 24 that contradict a rule of this repository, switched off in qodana.yaml with the reason beside each: JSDoc completeness (doc comments here are prose, 10 §6), bitwise operators in the PNG CRC-32 of gen-icons.mjs, and the deliberately doubled dark palette in tokens.css. A finding is fixed when it is a defect and switched off, with its reason, when it is a convention; it is never left standing to be ignored.

3. Store publishing automation (Chrome Web Store, Firefox Add-ons, Edge Add-ons)

Since v1.3.0 the release workflow can submit to all three stores, each step off until its owner switches it on with a repo variable. Uploaded builds still go through each store's review; automation removes the clicking, not the review.

Job Gate (repo variable) Environment Secrets in that environment Uploads
store-chrome CWS_AUTOPUBLISH=true chrome-web-store CHROME_PUBLISHER_ID, CHROME_EXTENSION_ID (= the item id, 00 §9), CHROME_SERVICE_ACCOUNT_CLIENT_EMAIL, CHROME_SERVICE_ACCOUNT_PRIVATE_KEY the Chrome package
store-firefox AMO_AUTOPUBLISH=true firefox-add-ons FIREFOX_EXTENSION_ID (= the gecko id), FIREFOX_JWT_ISSUER, FIREFOX_JWT_SECRET the Firefox package and its sources, listed, Firefox desktop only
store-edge EDGE_AUTOPUBLISH=true edge-add-ons EDGE_PRODUCT_ID, EDGE_CLIENT_ID, EDGE_API_KEY the Chrome package
  • The tool is publish-browser-extension (6.2.0 — what wxt submit wraps), installed from tools/publish/'s own lockfile: four packages, not the dev toolchain. A job's secrets are readable by everything on its runner for the whole job (§2.5), and these can push an update to every user.
  • One job per store, so a store's credentials never share a runner with another's, and one store refusing an upload does not stop the others.
  • Environments carry the secrets, not the repository, so nothing else in any workflow can read them. Give each one a required reviewer (the owner) to approve every upload, and restrict its deployment branches and tags to the tag pattern v*: an environment GitHub creates on first use has no rule at all, so any branch with a workflow that names it — a pushed branch, not only a release — could reach its secrets. The reviewer is the second lock, not the only one.
  • Chrome Web Store API v2. v1.1 — what the old store job's client-id/secret/refresh-token secrets used — was shut down on 2026-10-15; that job would have failed the day CWS_AUTOPUBLISH was set. v2 takes a publisher id and a Google Cloud service account (CHROME_API_VERSION=v2 is set by the job, never left to the tool's default).
  • AMO: the first version must be created by hand. The API (and the tool) can add a version to an add-on that exists, not create one; 13 §8. AMO JWT credentials come from addons.mozilla.org/developers/addon/api/key/.
  • Edge: the first submission is made in Partner Center, by hand (13 §9). API keys (v1.1, with a Client ID) expire every 72 days — a scheduled chore, not an incident; until rotated, store-edge fails and the others do not.
  • No re-run of a store job that succeeded: every store refuses a version number it has seen. Re-run only the failed job (13 §7).
Version Method
v1.0.x–v1.2.x Chrome Web Store by hand (13 §1b) — the CHROME_* secrets never existed.
v1.3.0 The first AMO and Edge versions must be uploaded by hand (13 §8, 13 §9); the Chrome one by hand until the owner sets up §3's environment.
later Each store automated once its environment has its secrets and its variable is true; switch a variable off to stop submitting there instantly (a review dispute, a rollout to watch).

Consider a staged (percentage) rollout on the Chrome Web Store for anything touching the write paths — CHROME_DEPLOY_PERCENTAGE for the automated job, the dashboard otherwise.

4. Branch/version flow

feat branch → PR (CI green + review) → squash-merge to main
release: bump package.json (semver) → CHANGELOG.md (Keep a Changelog) →
         git tag -s vX.Y.Z → push tag → release.yml
  • Patch = fixes / translation fixes / accessibility fixes — including a keyboard or assistive-technology route to something the product could already do (v1.2.1's Alt+↑/↓ beside drag-to-reorder): it closes a barrier, it does not add a capability · Minor = features, a new UI language (it changes the listing and what the switcher offers — 07 §6) / any permission or privacy-tab change (per 09 §4) · Major = breaking format/schema changes (BM JSON version bump).

5. Repo settings checklist (one-time ⚠️)

Only the repo-visible items can be verified from a checkout; the rest live in GitHub settings and stay ⚠️ until the owner confirms them. An unticked box here means "unknown", not "not done".

  • Branch protection on main — configured 2026-08-03. Two corrections to what this line used to say, both learned by applying it: - Required checks are Quality gates and Build + package / build. The old text said Build + package, the job name in ci.yml. That is not what gets reported: a job that calls a reusable workflow publishes its check as <caller job> / <reusable job>. Requiring the caller's name alone matches nothing, so every PR waits forever on a check that can never arrive — main becomes permanently unmergeable. - required_approving_review_count: 0, not 1. "Require review" is right for a team and a lockout for one maintainer, who cannot approve their own pull request. Zero still forces every change through a PR and through the checks above; it only drops the second pair of eyes that does not exist. Raise it the day a second maintainer does.

    Also set: linear history (matches the rebase-merge practice), no force
    pushes, no branch deletion, conversation resolution required. **Admins are
    deliberately NOT enforced** — with a single maintainer, an escape hatch
    from your own protection rules is the difference between a safety net and
    a locked door.
    
  • CodeQL Default setup DISABLED (Settings → Advanced Security) — confirmed 2026-07-26; codeql.yml uploads SARIF cleanly, which is only possible with it off (§2.2).

  • ⚠️ Store environments (§3), each with the owner as required reviewer, deployments limited to tags matching v*, and only its own secrets: chrome-web-store (the four API v2 values; CHROME_EXTENSION_ID is the item id in 00 §9), firefox-add-ons (FIREFOX_EXTENSION_ID = bookmarkmagic@poli0981.dev and the AMO JWT pair), edge-add-ons (product id, Client ID, API key — rotate every 72 days). GitHub creates an environment the first time a job names it, with no protection; add the reviewer, the v* tag rule and the secrets in Settings → Environments.

  • Actions secrets for Discord notifications — present as of 2026-10-28: DISCORD_CI_WEBHOOK, RELEASES_WEBHOOK, REPO_WEBHOOK (the last two under names the reusable does not read; release.yml maps them, §2.3). DISCORD_PING_ROLE_ID is optional and unset.

  • SONAR_TOKEN — set 2026-10-28, used only by sonarcloud.yml (§2.5).

  • QODANA_TOKEN — set by the owner 2026-11-18 (the Qodana Cloud project token, Ultimate), used only by qodana.yml's scan job (§2.7).

  • ⚠️ SonarCloud Automatic Analysis OFF (SonarCloud → bookmarkmagic → Administration → Analysis Method). Required for §2.5; with it on, the CI scan errors out. The owner agreed to switch it off on 2026-10-28.

  • Private vulnerability reporting enabled (verified 2026-10-28 via the API) — SECURITY.md and the issue chooser send reporters there first.

  • ⚠️ Repo variables CWS_AUTOPUBLISH, AMO_AUTOPUBLISH, EDGE_AUTOPUBLISH — unset, so no store job runs. Note that when a variable does not exist the expression is simply false, so "not configured" and "configured off" are indistinguishable from a run's output.

  • ⚠️ Repo variables DISCORD_ANNOUNCE and DISCORD_CI_NOTIFY — both still unset (2026-10-28), so no notification fires. The webhook secrets now exist, so either can be flipped to true whenever the owner wants the posts. Unset means skipped; a reusable-workflow caller cannot use continue-on-error to absorb a failure (§2.3).

  • Dependabot: npm weekly, grouped dev minor+patch, github-actions too — .github/dependabot.yml — and since v1.3.0 tools/publish/ (the store-upload tool's own lockfile, §3) as a third ecosystem entry. Both ecosystems have a 7-day cooldown (matching .npmrc min-release-age=7, 09 §3.1); WXT is excluded from the group so every WXT bump arrives alone and gets its built-manifest diff; majors of the TypeScript pair and @types/node are ignored.

  • Labels the forms and triage.yml apply — maintenance-mode, needs-triage, roadmap (new with MAINTENANCE.md) and parser, i18n (named by the parser and translation forms since 2026-08-03, and missing until 2026-11-18: a form silently drops a label that does not exist).

  • Issue templates: .github/ISSUE_TEMPLATE/ — bug, parser report (requires a sanitized sample), translation, feature, plus a config.yml chooser. Landed 2026-08-03; 13 §6 had asserted they existed since before launch.

  • ⚠️ GPG signing required for tags (matches tracker-repo practice). The v1.0.0 tag object is in fact PGP-signed, so the practice held once.