ci: scope the refresh lane's concurrency group by ref

One global group would have two PRs touching this lane cancel each other's dry
run and report the cancellation as though the lane were busy. Two SCHEDULED
refreshes still cannot overlap, which is the race the group exists for.

Also corrects a comment that described a `continue-on-error` the step does not
use: the AFTER leg's exit is captured with `set +e` and read before any pipe.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019RfFHiRCSs3JXLK4cwcfox

Author: claude

.github/workflows/shard-timings-refresh.yml

@@ -104,7 +104,12 @@ permissions:
 # the same bot branch; the newer inputs are the better ones, so the in-flight run
 # yields.
 concurrency:
-  group: shard-timings-refresh
+  # Scoped BY REF, not one global group. Two scheduled refreshes cannot overlap
+  # (they would race on the same bot branch, and the newer inputs are the better
+  # ones, so the in-flight run yields) — but two PRs touching this lane are
+  # unrelated runs, and a single global group would have each cancel the other's
+  # dry run and report a cancellation as though the lane were busy.
+  group: shard-timings-refresh-${{ github.ref }}
   cancel-in-progress: true
 
 jobs:
@@ -310,11 +315,13 @@ jobs:
       # verdict line rather than recomputed here — a second implementation of the
       # binning would be a second answer to grade against.
       #
-      # ⛔ `continue-on-error` on the AFTER leg only, and it is not leniency: a
-      # red here is the partitioner refusing an honest measurement, which is
-      # information the PR must carry, not a reason to withhold the PR. The
-      # verdict text goes into the body either way and the PR's own lint job
-      # grades it again.
+      # ⛔ The AFTER leg's non-zero exit is CAPTURED, not propagated, and that is
+      # not leniency: a red there is the partitioner refusing an honest
+      # measurement, which is information the PR must carry rather than a reason
+      # to withhold the PR. `set +e` around that one command, with `$?` read
+      # immediately and before any pipe, is what keeps the verdict readable
+      # without letting the step's own status swallow it; the text goes into the
+      # body either way and the PR's own lint job grades it again.
       - name: Predicted bins, before and after
         id: bins
         if: steps.compare.outputs.changed == 'true'