docs: document the retry-until shim input

CLAUDE.md

@@ -285,6 +285,15 @@ jobs:
     permissions: { pages: write, id-token: write, contents: read, actions: write, statuses: write }
 ```
 
+Your `publish-dispatch.yml` wrapper's `workflow_dispatch` inputs must include `pr`
+**and `retry-until`**, forwarding both into `publish.yml`'s `with:` — copy the shim
+in [`docs/reference/workflows.md`](docs/reference/workflows.md) or the
+[tutorial](docs/tutorials/adding-to-a-fresh-repo.md) verbatim rather than
+hand-rolling it. `retry-until` is never set by a human; `publish.yml`'s
+`re-dispatch` job always passes it when re-firing the shim after a tag release, so
+a shim missing the input rejects that dispatch (breaking every tagged release's
+re-publish, not just the wedged-origin retry it exists for).
+
 Add a `release` job that `uses:` this repo's `release.yml@<tag>` (with `permissions:
 contents: write`) to attach each tag's built `docs.zip` (bare `html/` root) as a
 Release asset so `assemble` can reconstruct released versions, and the

docs/reference/workflows.md

@@ -90,6 +90,7 @@ re-dispatch, a fork-PR preview, a manual re-deploy). A reusable workflow can't b
 | `version-name` | no | `""` | Version name of **this run's** `docs` artifact to stage directly, instead of gathering it from durable sources. Set by `ci.yml`'s inline publish (the run isn't a completed success yet, so the gather can't discover it — or would find a stale previous build). Empty → pure durable gather (the dispatch paths). |
 | `pr` | no | `""` | Fork PR number to approve (pins its head SHA as `preview-approved`) and preview. Set on the shim's `workflow_dispatch` path. |
 | `dispatch-workflow` | no | `publish-dispatch.yml` | Filename of the shim in the consumer's repo, re-fired by the `re-dispatch` job. Override only if you renamed the shim. |
+| `retry-until` | no | `""` | Internal — epoch-seconds deadline for auto-retrying a wedged Pages origin. Set only by the `re-dispatch` job (from the release's publish time) when it re-fires the shim; carried unchanged through any retries. **The shim must declare + forward this input** (`workflow_dispatch` + the `with:` block below) even though consumers never set it themselves — `re-dispatch` always passes it, and a shim missing the input rejects the dispatch. Don't set manually. |
 
 ## What `publish.yml` gathers
 

docs/tutorials/adding-to-a-fresh-repo.md

@@ -132,6 +132,10 @@ on:
         description: "Fork PR to approve + preview (empty = re-deploy)"
         required: false
         default: ""
+      retry-until:
+        description: "Internal — epoch-seconds retry deadline. Don't set manually."
+        required: false
+        default: ""
 jobs:
   publish:
     uses: DiamondLightSource/myst-version-switcher-plugin/.github/workflows/publish.yml@__LATEST_TAG__
@@ -142,6 +146,8 @@ jobs:
       pr: ${{ inputs.pr }}
       # the file the tag re-dispatch re-fires
       dispatch-workflow: publish-dispatch.yml
+      # forwarded so the re-dispatch job's auto-retry can re-fire this shim
+      retry-until: ${{ inputs.retry-until }}
     permissions:
       contents: read
       actions: write