* Publish GitHub release from master branch
This ensures that changes made on `develop` since branching for the
release are not included. It also ensures that the final release
sourcemaps line-up correctly (they were always build on master)`.
* Consolidate publish jobs
The jobs `job-publish-release` and `create_github_release` both handle
different parts of publishing a release. They have been consolidated
into a single `job-publish-release` job.
* Update release script to expect a merge commit
The release script was originally written to be run on `develop`, so it
expected the current commit to be a result of `Squash & Merge`. Now
that it's run on `master`, it will generally be run against a merge
commit.
The version is now extracted from the commit message using a regular
expression that should work on all version of Bash v3+, and should be
tolerant of both merge commits and `Squash & Merge` commits.
* Target `master` with release PR
`master` is now targeted by the release PR instead of `develop`, as
the release has to be created from the master branch.
The update to `develop` is handled after the release by a PR from
`master` to `develop`, which is created automatically after the
release.
As a solution to the constant lockfile churn issues we've had with
`npm`, the project now uses `yarn` to manage dependencies.
The `package-lock.json` file has been replaced with `yarn.lock`, which
was created using `yarn import`. It should approximate the contents of
`package-lock.json` fairly well, though there may be some changes due to
deduplication. The codeowners file has been updated to reference this
new lockfile.
All documentation and npm scripts have been updated to reference `yarn`
rather than `npm`. Note that running scripts using `npm run` still works
fine, but it seemed better to switch those to `yarn` as well to avoid
confusion.
The `npm-audit` Bash script has been replaced with `yarn-audit`. The
output of `yarn audit` is a bit different than `npm audit` in that it
returns a bitmask to describe which severity issues were found. This
made it simpler to check the results directly from the Bash script, so
the associated `npm-audit-check.js` script was no longer required. The
output should be exactly the same, and the information is still sourced
from the same place (the npm registry).
The new `yarn-audit` script does have an external dependency: `jq`.
However, `jq` is already assumed to be present by another CI script, and
is present on all CI images we use. `jq` was not added to `package.json`
as a dependency because there is no official package on the npm
registry, just wrapper scripts. We don't need it anywhere exept on CI
anyway.
The section in `CONTRIBUTING` about how to develop inside the
`node_modules` folder was removed, as the advice was a bit dated, and
wasn't specific to this project anyway.
The npm audit script was auditing all dependencies, then filtering the
results to just the advisories concerning production dependencies. This
was done by checking the boolean `dev` and `optional` properties of each
`findings` entry in each advisory.
The `dev` and `optional` properties are now missing, which is resulting
in dev advisories being mistakenly identified as affecting production.
This check has been removed, and instead the `--production` flag is used
when calling `npm audit`. This accomplishes the same goal without
relying as much upon the audit output format.
The `--production` flag was added in `npm` `v6.10.0`, so `npm` has been
updated to the current latest stable (`v6.10.2`) for the `test-deps`
job. It was also updated on the `prep-deps-npm` job to ensure
consistency in behaviour. The other jobs only use `npm run` which hasn't
changed substantially in some time, so compatibility isn't really a
concern for those.
`audit.json` has also been added to `.gitignore`. It was accidentally
checked in once while working on this branch.
* ci: Rename full_test to test_and_release
* ci: Add scripts to automate GH releases
* Add .bak files to .gitignore
* ci: Add reviewer to the auto version PR
Two important notes:
1. The time it takes to download is negligble compared to e2e test runs
2. Since we cannot use environment variables in CircleCI cache keys we
can't cache the download correctly and have it update when we switch
firefox versions—this isn't the end of the world because of point 1
