fix(ci): preserve manual label retry through gh pr view failure + correct stale-event comment

Two follow-up issues from round-3 Copilot review:

1. **`gh pr view` failure path repeats round-2's manual-label silent
   no-op bug**: round-3 split the stale-SHA branch by event type, but
   the `gh pr view --json headRefOid` failure path right above kept a
   single `exit 0` that applies to all event types. On the manual
   label path the earlier step has already removed the
   `copilot-review` label, so an exit here is a silent no-op (no
   `@claude` trigger, no feedback). Fix: split the failure path the
   same way as the stale-SHA branch. Manual label path falls back to
   the payload SHA and proceeds (operator gets a CI run for what we
   can verify, the marker dedup downstream still skips cleanly if the
   payload SHA was stale, and the operator at least sees a real run).
   Auto event paths keep the fail-closed `exit 0` (Copilot's next
   review or the cron fallback drives the PR forward).

2. **Stale-event comment overstates push semantics**: round-3's
   "stale auto event" comment said "a newer push triggers a fresh
   review/comment event that will reach this workflow". A bare push
   does NOT emit `pull_request_review` / `pull_request_review_comment`
   events; only Copilot's review/comment does. Reworded to "a fresh
   Copilot review/comment for the newer HEAD will fire its own
   event" and added the explicit cron-fallback escape hatch
   ("otherwise cron fallback handles within ~30–60 min") so the
   reader understands the actual recovery path, not an aspirational
   one.

Author: kotakanbe

.github/workflows/copilot-review-fix.yml

@@ -134,9 +134,30 @@ jobs:
           # error: skipping is safer than posting a possibly-stale
           # trigger.
           payload_sha="${HEAD_SHA:-}"
+          current_sha=""
           if ! current_sha=$(gh pr view "$PR_NUMBER" --repo "$REPO" --json headRefOid --jq '.headRefOid'); then
-            echo "::warning::PR #$PR_NUMBER: gh pr view --json headRefOid failed — skipping @claude trigger (will be re-driven by cron / next event)"
-            exit 0
+            current_sha=""
+            if [ "${GITHUB_EVENT_NAME:-}" = "pull_request" ]; then
+              # Manual label path: the earlier step has already
+              # removed `copilot-review`, so an exit here would be
+              # a silent no-op (no `@claude` trigger, no PR
+              # feedback). Fall back to the payload SHA and
+              # proceed; if the payload SHA was stale, the marker
+              # dedup further down will skip cleanly. The operator
+              # gets a CI run for what we can verify (payload
+              # HEAD), which is better than nothing.
+              echo "::warning::PR #$PR_NUMBER: gh pr view --json headRefOid failed on manual label path — falling back to payload HEAD ($payload_sha) and proceeding (label was already removed; silent exit would break the manual retry path)"
+              current_sha="$payload_sha"
+            else
+              # Auto event paths: skipping is safe — the same
+              # Copilot review/comment that fired this run will
+              # re-fire on the next cron tick / re-request, so
+              # cron / a fresh Copilot review will eventually
+              # drive this PR forward without us posting a
+              # possibly-stale trigger now.
+              echo "::warning::PR #$PR_NUMBER: gh pr view --json headRefOid failed on auto event path — skipping @claude trigger (will be re-driven by cron / next Copilot review)"
+              exit 0
+            fi
           fi
           if [ -z "$current_sha" ] || [ "${#current_sha}" -lt 40 ]; then
             echo "::warning::missing/short current PR HEAD '$current_sha' — skip @claude trigger"
@@ -156,12 +177,14 @@ jobs:
               echo "stale label event: payload HEAD=$payload_sha, current HEAD=$current_sha — proceeding with current HEAD (manual label path: no newer event will reach this workflow)"
             else
               # Auto event paths (`pull_request_review` /
-              # `pull_request_review_comment`): a newer push triggers
-              # a fresh review/comment event that will reach this
-              # workflow with the up-to-date HEAD, so it's safe (and
-              # correct) to skip this stale-SHA run rather than post
-              # a duplicate trigger.
-              echo "stale auto event: payload HEAD=$payload_sha, current HEAD=$current_sha — skip @claude trigger (newer event will handle current HEAD)"
+              # `pull_request_review_comment`): a fresh Copilot
+              # review/comment for the newer HEAD will fire its
+              # own event reaching this workflow with the up-to-
+              # date HEAD; if Copilot does NOT re-review, the
+              # cron fallback in `copilot-clean-label.yml` covers
+              # this PR within ~30–60 min. Either way, skipping
+              # this stale-SHA run is correct.
+              echo "stale auto event: payload HEAD=$payload_sha, current HEAD=$current_sha — skip @claude trigger (newer Copilot review/comment for current HEAD will fire its own event, otherwise cron fallback handles within ~30–60 min)"
               exit 0
             fi
           fi