fix(ci): watchdog polls + exits early instead of sleeping the full timeout

scripts/ci-gate-watchdog.sh

@@ -1,16 +1,22 @@
 #!/usr/bin/env bash
 # ci-gate-watchdog.sh — invoked by the `watchdog` job in _ci-gate.yml.
 #
-# After QUEUE_TIMEOUT_MINUTES, cancels every sibling job in the current
-# workflow run that is still in `queued` state, except for "Queue Watchdog"
-# (this script's own job) and "Merge Gate" (the gate job, which is still
-# pending here because `needs: watchdog`). Cancellation forces a terminal
-# state so `gate` can finally schedule and `re-actors/alls-green` can
-# evaluate, ensuring the required `Merge Gate` status always reports.
+# Poll the run's sibling jobs and EXIT AS SOON AS none is still `queued`
+# (everything got scheduled) — only cancel jobs that remain stuck in `queued`
+# after QUEUE_TIMEOUT_MINUTES. Cancelling forces a terminal state so `gate` can
+# finally schedule and `re-actors/alls-green` can evaluate, ensuring the required
+# `Merge Gate` status always reports.
+#
+# Why poll instead of `sleep $TIMEOUT`: a fixed sleep billed a full
+# QUEUE_TIMEOUT_MINUTES of runner time on EVERY run, even though jobs only get
+# stuck in `queued` on self-hosted runners that never pick them up. On GitHub-
+# hosted runners siblings schedule within seconds, so this now exits in seconds.
+# "Queue Watchdog" (this job) and "Merge Gate" (awaiting this job) are exempt —
+# their `queued`/pending state here is intentional.
 #
 # Required env:
 #   GH_TOKEN              — GitHub token with actions:write on the run
-#   QUEUE_TIMEOUT_MINUTES — minutes to wait before evaluating queued jobs
+#   QUEUE_TIMEOUT_MINUTES — minutes to wait before cancelling still-queued jobs
 #   REPO                  — owner/repo of the current workflow run
 #   RUN_ID                — workflow run id
 
@@ -21,19 +27,33 @@ set -euo pipefail
 : "${REPO:?required}"
 : "${RUN_ID:?required}"
 
-sleep "$(awk "BEGIN{printf \"%d\", $QUEUE_TIMEOUT_MINUTES * 60}")"
-
 # Names whose `status == "queued"` is intentional at this point in the run.
-# The watchdog itself is mid-execution; the gate is awaiting watchdog.
 exempt_filter='.name != "Queue Watchdog" and .name != "Merge Gate"'
 
-stuck=$(gh api --paginate "repos/${REPO}/actions/runs/${RUN_ID}/jobs?per_page=100" \
-  --jq ".jobs[] | select(.status == \"queued\" and ${exempt_filter}) | \"\(.id)\t\(.name)\"")
+queued_siblings() {
+  gh api --paginate "repos/${REPO}/actions/runs/${RUN_ID}/jobs?per_page=100" \
+    --jq ".jobs[] | select(.status == \"queued\" and ${exempt_filter}) | \"\(.id)\t\(.name)\""
+}
+
+# awk parses QUEUE_TIMEOUT_MINUTES so a float (e.g. 0.5) truncates to an integer
+# instead of crashing bash arithmetic, which only handles integers.
+limit_seconds=$(awk "BEGIN{printf \"%d\", $QUEUE_TIMEOUT_MINUTES * 60}")
+poll_interval=15
 
-if [ -z "$stuck" ]; then
-  echo "No stuck queued jobs found."
-  exit 0
-fi
+# $SECONDS is a bash builtin tracking elapsed script time — no subshell per poll.
+# Exit the instant nothing is queued; otherwise keep watching until the deadline.
+while :; do
+  stuck=$(queued_siblings)
+  if [ -z "$stuck" ]; then
+    echo "No queued sibling jobs — all scheduled. Nothing to cancel."
+    exit 0
+  fi
+  if [ "$SECONDS" -ge "$limit_seconds" ]; then
+    echo "Timeout (${QUEUE_TIMEOUT_MINUTES}m) reached; cancelling jobs still stuck in queued:"
+    break
+  fi
+  sleep "$poll_interval"
+done
 
 while IFS=$'\t' read -r job_id job_name; do
   echo "Cancelling stuck queued job: ${job_name} (id=${job_id})"