ci: proposal — notify #team-alpha on merge-train/spartan MQ failures

.github-new/workflows/ci3-spartan-mq-notify.md

@@ -0,0 +1,56 @@
+# Proposal: Notify #team-alpha on merge-train/spartan merge-queue failures
+
+This change could not be applied directly because the ClaudeBox session was not
+started with `ci-allow`, so `.github/` edits are blocked. Apply the snippet
+below manually, or restart the request prefixed with `ci-allow`.
+
+## Why
+
+The only existing failure notification for `merge-train/spartan` in the merge
+queue is `merge-queue-dequeue-notify.yml`, which fires on the
+`pull_request: dequeued` event. If GitHub's merge queue retries an attempt and
+the retry succeeds, the dequeue event never fires and the team gets no signal
+about the underlying flake or failure. This adds a per-attempt notification so
+every MQ failure for `merge-train/spartan` is surfaced in #team-alpha
+(`C0AU8BULZHC`).
+
+## Patch
+
+Apply to `.github/workflows/ci3.yml` — insert the new step immediately after
+the existing `Notify Slack on backport CI failure` step (before the
+`Upload benchmarks` step) in the `ci` job:
+
+```yaml
+      # Fires on every merge-queue CI failure for merge-train/spartan, not just on dequeue
+      # (the queue may retry, so merge-queue-dequeue-notify.yml alone misses individual failures).
+      # CI_MODE=merge-queue-heavy is only set for merge-train/spartan in the merge queue
+      # (see .github/ci3_labels_to_env.sh).
+      - name: Notify Slack on merge-train/spartan merge-queue failure
+        if: failure() && env.CI_MODE == 'merge-queue-heavy'
+        env:
+          SLACK_BOT_TOKEN: ${{ secrets.SLACK_BOT_TOKEN }}
+          RUN_URL: https://github.com/${{ github.repository }}/actions/runs/${{ github.run_id }}
+        run: |
+          if [ -n "${SLACK_BOT_TOKEN}" ]; then
+            TEXT="CI3 failed in merge queue for *merge-train/spartan*: <${RUN_URL}|View Run>"
+            curl -X POST https://slack.com/api/chat.postMessage \
+              -H "Authorization: Bearer $SLACK_BOT_TOKEN" \
+              -H "Content-type: application/json" \
+              --data "$(jq -n --arg c "C0AU8BULZHC" --arg t "$TEXT" '{channel:$c, text:$t}')"
+          fi
+```
+
+## Detection logic
+
+`env.CI_MODE == 'merge-queue-heavy'` is the cleanest signal for "this CI run
+is for `merge-train/spartan` inside the merge queue". `.github/ci3_labels_to_env.sh`
+only sets `CI_MODE=merge-queue-heavy` when the merge-group event corresponds to
+a PR whose head branch is `merge-train/spartan`; standard merge-queue runs for
+other PRs use `CI_MODE=merge-queue`.
+
+## Test plan
+
+- After merge, wait for a `merge-train/spartan` PR to fail in the merge queue
+  and confirm a notification appears in #team-alpha.
+- Verify non-spartan merge-queue runs do **not** post to this channel (their
+  `CI_MODE` is `merge-queue`, not `merge-queue-heavy`).