ci(shard-timings): zero candidate runs is NOT MEASURED, not a finding

The refresh lane read every non-zero from the run selector as a refusal, so
an EMPTY candidate list printed "NO ELIGIBLE RUN among the 0 ... every one
was censored, failed, or has lost its artifacts" and failed the job. Nothing
had been examined, and none of those three causes had occurred.

Split into two legs with two exits and two messages: an empty candidate list
exits EXIT_PREREQUISITE_NOT_MET (3, the repo-wide code), candidates that
existed and were all rejected keep exit 1 and keep the sentence naming the
causes that now apply. The workflow reads 3 as NOT MEASURED, leaves the
dataset alone, stays green, and says loudly in its annotation and its step
summary that a persistent NOT MEASURED is a defect rather than a steady
state. The step's own run block is lifted out of the YAML and driven under
bash against a stub node on all three exit codes.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012GKcPZbMoGq7WPzKLfRBTU

Author: claude

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

@@ -268,8 +268,74 @@ jobs:
           # is coverage left on the table; the cost is two API reads per run
           # examined and the loop below still stops at the first accumulation
           # that covers the workspace.
+          #
+          # ⭐ THREE EXITS, THREE READINGS (#16467). The selector answers 0, 1
+          # or 3, and this step must not flatten them:
+          #
+          #   0  eligible runs on stdout — carry on.
+          #   1  candidates existed and every one was censored, failed or lost
+          #      its artifacts. A FINDING. The job goes red, as it always did.
+          #   3  EXIT_PREREQUISITE_NOT_MET — the API held NO completed
+          #      `schedule` run of ci.yml at all. NOTHING was measured. The
+          #      dataset is untouched, which is the correct outcome, so the job
+          #      must not go red for it.
+          #
+          # Measured, which is why this branch exists: on the pull request that
+          # landed the hourly run, this lane ran with `event=schedule` before
+          # that trigger existed on main, printed "NO ELIGIBLE RUN among the 0"
+          # and named three causes none of which had occurred. It also does not
+          # clear at merge — there is a bootstrap window of at least an hour
+          # before the first hourly run finishes.
+          #
+          # `|| SELECT_EXIT=$?` and not a bare call: this step runs under
+          # `bash -e`, where the non-zero would kill it before the code could
+          # be read at all.
+          #
+          # stderr goes to a FILE and is echoed back immediately: the log keeps
+          # every line it had, and the refusal text becomes quotable into the
+          # step summary below, where somebody reading a green job will see it.
+          SELECT_EXIT=0
           node scripts/ci/select-shard-timings-run.mjs --candidates --limit 24 \
-            > "$RUNNER_TEMP/candidates.json"
+            > "$RUNNER_TEMP/candidates.json" 2> "$RUNNER_TEMP/select-stderr.txt" || SELECT_EXIT=$?
+          cat "$RUNNER_TEMP/select-stderr.txt"
+          echo "select_exit=$SELECT_EXIT" >> "$GITHUB_OUTPUT"
+          if [ "$SELECT_EXIT" -eq 3 ]; then
+            # ⛔ NEVER QUIET. A 3 that scrolls past in a green job is how this
+            # lane ends up passing because it never looked — the shape that has
+            # already cost this repo two cards. It is an annotation, a step
+            # summary section and a job-level notice, and every one of them says
+            # that a PERSISTENT 3 is a defect rather than a steady state.
+            echo "not_measured=true" >> "$GITHUB_OUTPUT"
+            echo "::warning::Shard timings NOT MEASURED: no completed \`schedule\` run of ci.yml exists on main yet. The dataset was left untouched. If this repeats once the hourly full run has been live for a few hours, the trigger is gone or every hourly run is being cancelled — file it."
+            {
+              echo "### Shard timings: NOT MEASURED (exit 3) — nothing was regenerated, and nothing failed"
+              echo
+              echo "\`select-shard-timings-run --candidates\` found **no completed \`schedule\` run of"
+              echo "\`ci.yml\` on \`main\` at all**. That is not \"every candidate was rejected\": there were"
+              echo "no candidates, so no run was censored, none failed and none lost its artifacts."
+              echo "\`scripts/test-shard-timings.json\` is untouched, which is the correct outcome for this"
+              echo "reading, and this job is green because nothing went wrong — not because anything passed."
+              echo
+              echo "Expected exactly once: while the hourly full run bootstraps. The trigger has to be on"
+              echo "\`main\` and one run has to finish, and run-summary artifacts live 1 day."
+              echo
+              echo "⛔ **A PERSISTENT NOT MEASURED IS A DEFECT, NOT A STEADY STATE.** If this section is"
+              echo "still here after the hourly run has been live for a few hours, the \`schedule\` trigger"
+              echo "has been removed from \`ci.yml\` or every hourly run is being cancelled — and the"
+              echo "balancing dataset is quietly ageing out while this job reports green. File a card."
+              echo
+              echo "The selector's own refusal text, verbatim:"
+              echo
+              echo '```'
+              cat "$RUNNER_TEMP/select-stderr.txt" 2>/dev/null || echo '(the refusal text is in this job log)'
+              echo '```'
+            } >> "$GITHUB_STEP_SUMMARY"
+            exit 0
+          fi
+          if [ "$SELECT_EXIT" -ne 0 ]; then
+            echo "::error::select-shard-timings-run refused with exit $SELECT_EXIT — candidates existed and none was eligible. Nothing was regenerated."
+            exit "$SELECT_EXIT"
+          fi
           echo "Eligible runs, newest first:"
           node -e '
             const runs = JSON.parse(require("fs").readFileSync(process.env.RUNNER_TEMP + "/candidates.json", "utf8"));
@@ -304,6 +370,11 @@ jobs:
       # quiet week costs one run's download and a busy one costs a few.
       - name: Regenerate the dataset, accumulating runs until the workspace is covered
         id: generate
+        # Skipped on the NOT MEASURED reading: there is nothing to download.
+        # Every step after `compare` is already gated on
+        # `steps.compare.outputs.changed`, which is the empty string when
+        # `compare` itself is skipped — so guarding these two guards the tail.
+        if: steps.select.outputs.not_measured != 'true'
         env:
           GITHUB_TOKEN: ${{ github.token }}
         run: |
@@ -411,6 +482,7 @@ jobs:
       # file is what has to differ for a PR to be worth anyone's attention.
       - name: Compare against the committed dataset
         id: compare
+        if: steps.select.outputs.not_measured != 'true'
         run: |
           if cmp -s scripts/test-shard-timings.json "$RUNNER_TEMP/refresh/refreshed.json"; then
             echo "changed=false" >> "$GITHUB_OUTPUT"