feat(preview): expire preview workspaces and prune expired ones on cleanup

[toiroakr]

Summary

A preview workspace whose close-time cleanup never ran (the job failed or was disabled, or the PR stayed open) was never deleted. This lets preview workspaces expire and lets every PR close sweep the expired ones.

• preview-deploy takes ttl (e.g. 7d). The workspace records an expiry when created, and every later push restarts it, so it counts from the last push.
• preview-cleanup takes prune-expired, organization-id and folder-id. After deleting the per-PR workspace it runs tailor workspace prune --expired.
• A PR that sat idle past ttl has its workspace swept while it is still open. The next push now creates a fresh workspace under the same name instead of deploying to the deleted one, and closing such a PR no longer fails.
• Separately, the workspace ID lookup in both actions used (?P<id>…) in a jq capture, which the jq 1.7.1 and 1.8.2 release binaries reject. try hid the error, so an existing workspace went unrecognized. It is now (?<id>…).

Both new inputs are off by default, so existing workflows behave as before. Related to issue #1940 in platform-planning.

Usage

- uses: tailor-platform/actions/preview-deploy@v2
  with:
    workspace-name-prefix: my-app
    region: us-west
    folder-id: ${{ vars.TAILOR_PLATFORM_FOLDER_ID }}
    ttl: 7d
    # ...

- uses: tailor-platform/actions/preview-cleanup@v2
  with:
    workspace-name-prefix: my-app
    prune-expired: "true"
    folder-id: ${{ vars.TAILOR_PLATFORM_FOLDER_ID }}
    # ...

Behavior worth reviewing

• Scope of the sweep. Only workspaces in the given folder (or, without folder-id, directly under the organization) whose whole name is {prefix}-pr-{number}, with the prefix escaped. Other apps sharing the location are untouched. The folder wins over the organization, as in preview-deploy. A workspace without a recorded expiry is never deleted.
• --limit 0. workspace prune aborts without deleting anything when more than its default 20 workspaces match. With that, a backlog of 21 would fail every later PR close. The name and location filters keep the sweep narrow.
• No location known. With none of folder-id, organization-id, TAILOR_PLATFORM_FOLDER_ID or TAILOR_PLATFORM_ORGANIZATION_ID (the two variables workspace create itself reads), the sweep warns and is skipped instead of failing, because prune --expired rejects an empty location and generated workflows pass an empty string for an unset variable. This is a provisional rule.
• Order and failure. The sweep is the last step and runs under !cancelled(), so a failed deletion is still swept and a failed sweep cannot undo the deletion or the comment update.
• Recreating a pruned workspace. preview-deploy runs workspace get on the recorded ID. Not found means pruned, so it creates a new workspace; the PR comment is updated with the new ID. Any other failure stops the step so a transient error cannot create a duplicate. If workspace create --ttl prints the new workspace and then exits nonzero because it could not confirm the expiry, the ID is still taken from its output and the expiry is set again, so the workspace is not left untracked. Workspace names are not unique on the platform, so reusing the name is fine.
• Not found is matched on CLI output (not_found or workspace … not found), because workspace get surfaces the raw Connect error without a stable code. A permission loss also reports not found.

Verification

• pnpm test:preview runs the real step scripts from the action files with a mocked gh and tailor, and is added to CI. Existing test:cli-bin-contract and test:deploy-outputs still pass; zizmor and ghalint are clean.
• The jq fix was checked against the jq 1.7.1 and 1.8.2 release binaries; distribution builds were not checked.
• The sed that escapes the prefix was only run with BSD sed locally. The CI run on ubuntu covers GNU sed.
• Pruning against a real workspace was not run.