fix(acceptance): 联动压力采样与负载停止

零副本 Worker 站点没有 metrics，应从聚合采样中跳过；有副本的站点采样失败仍保持硬失败。补充缺失字段诊断，避免只输出无上下文的采样失败。\n\n容量轮次和混合长稳压测现在持续监视压力采样进程；采样器提前退出时立即终止当前负载并走失败收口，避免在资源监控失效后继续造任务。\n\n验证：bash -n、ShellCheck、零副本站点测试、采样失败联动停止测试、manual-release-test。

scripts/cluster/run-production-acceptance.sh

@@ -2074,12 +2074,12 @@ WHERE acceptance_run_id IS NULL
   sample_pressure "$pressure_report" "$pressure_stop" &
   pressure_pid=$!
   active_pressure_pid=$pressure_pid
-  run_mixed_load_profile \
+  run_with_pressure_monitor "$pressure_pid" run_mixed_load_profile \
     mixed-soak "$AI_GATEWAY_ACCEPTANCE_SOAK_DURATION" \
     "$soak_image_rate" "$soak_video_rate" \
     "$report_root/mixed-soak.json" 3h || status=$?
   if (( status == 0 )); then
-    run_mixed_load_profile \
+    run_with_pressure_monitor "$pressure_pid" run_mixed_load_profile \
       mixed-overload "$AI_GATEWAY_ACCEPTANCE_OVERLOAD_DURATION" \
       "$overload_image_rate" "$overload_video_rate" \
       "$report_root/mixed-overload.json" 30m || status=$?
@@ -2506,9 +2506,15 @@ verify_control_plane_and_recent_logs() {
 workload_metrics_for_site() {
   local workload=$1
   local site=$2
-  local host pod_ip
+  local host pod_ip replicas
   [[ $workload == api || $workload == worker ]]
   [[ $site == ningbo || $site == hongkong ]]
+  if [[ $workload == worker ]]; then
+    replicas=$(remote_kubectl get deployment "easyai-worker-$site" -n "$namespace" \
+      -o 'jsonpath={.spec.replicas}') || return 1
+    [[ $replicas =~ ^[0-9]+$ ]] || return 1
+    (( replicas > 0 )) || return 2
+  fi
   if [[ $site == ningbo ]]; then
     host=$CLUSTER_NINGBO_HOST
   else
@@ -2524,6 +2530,28 @@ workload_metrics_for_site() {
   cluster_ssh "$host" "curl -fsS --max-time 5 http://$pod_ip:8088/metrics"
 }
 
+run_with_pressure_monitor() {
+  local pressure_pid=$1
+  shift
+  local load_pid load_status=0
+  "$@" &
+  load_pid=$!
+  active_load_pid=$load_pid
+  while kill -0 "$load_pid" >/dev/null 2>&1; do
+    if ! kill -0 "$pressure_pid" >/dev/null 2>&1; then
+      kill "$load_pid" >/dev/null 2>&1 || true
+      wait "$load_pid" >/dev/null 2>&1 || true
+      active_load_pid=
+      echo 'pressure sampler exited before the active load completed' >&2
+      return 1
+    fi
+    sleep "${AI_GATEWAY_ACCEPTANCE_PROCESS_WATCH_SECONDS:-2}"
+  done
+  wait "$load_pid" || load_status=$?
+  active_load_pid=
+  return "$load_status"
+}
+
 current_max_pod_memory_mib() {
   remote_kubectl top pods -n "$namespace" \
     -l 'app.kubernetes.io/part-of=easyai-ai-gateway' --no-headers |
@@ -2801,7 +2829,7 @@ sample_pressure() {
   local output=$1
   local stop_file=$2
   local database_state node_memory_percent pod_memory_mib worker_resources pool_state metrics
-  local consecutive_failures=0
+  local consecutive_failures=0 missing_fields metrics_status
   echo 'timestamp,queued,running,oldest_wait_seconds,db_connections,db_max_connections,node_memory_percent,pod_memory_mib,pool_acquired_max,pool_max,pool_idle_min,empty_acquire_total,canceled_acquire_total,lease_failure_total,lease_lost_total,active_instances_min,allocated_capacity_max,critical_pool_acquired_max,critical_pool_max,critical_pool_idle_min,critical_empty_acquire_total,critical_canceled_acquire_total,river_pool_acquired_max,river_pool_max,river_pool_idle_min,river_empty_acquire_total,river_canceled_acquire_total,worker_memory_mib,worker_cpu_millicores' >"$output"
   while [[ ! -f $stop_file ]]; do
     if ! database_state=$(database_query "
@@ -2828,7 +2856,12 @@ WHERE acceptance_run_id='$run_id'::uuid;"); then
     if ! pool_state=$(
       for workload in api worker; do
         for site in ningbo hongkong; do
-          metrics=$(workload_metrics_for_site "$workload" "$site") || exit
+          metrics_status=0
+          metrics=$(workload_metrics_for_site "$workload" "$site") || metrics_status=$?
+          if (( metrics_status == 2 )); then
+            continue
+          fi
+          (( metrics_status == 0 )) || exit 1
           awk -v workload="$workload" '
             $1=="easyai_gateway_postgres_pool_acquired_connections" {acquired=$2}
             $1=="easyai_gateway_postgres_pool_max_connections" {pool_max=$2}
@@ -2900,9 +2933,16 @@ WHERE acceptance_run_id='$run_id'::uuid;"); then
     ) 2>/dev/null; then
       pool_state=
     fi
+    missing_fields=
+    [[ -n $database_state ]] || missing_fields+=database_state,
+    [[ -n $node_memory_percent ]] || missing_fields+=node_memory_percent,
+    [[ -n $pod_memory_mib ]] || missing_fields+=pod_memory_mib,
+    [[ -n $worker_resources ]] || missing_fields+=worker_resources,
+    [[ -n $pool_state ]] || missing_fields+=pool_state,
     if [[ -z $database_state || -z $node_memory_percent ||
           -z $pod_memory_mib || -z $worker_resources || -z $pool_state ]]; then
       ((consecutive_failures += 1))
+      echo "pressure sampling unavailable: attempt=$consecutive_failures fields=${missing_fields%,}" >&2
       if (( consecutive_failures >= 3 )); then
         echo 'pressure sampling failed three consecutive times' >&2
         return 1
@@ -3053,11 +3093,24 @@ run_capacity_round() {
   sample_pressure "$pressure_report" "$pressure_stop" &
   pressure_pid=$!
   active_pressure_pid=$pressure_pid
-  run_load_profile gemini-baseline "$prefix-gemini-baseline.json" || status=$?
-  if (( status == 0 )); then run_load_profile gemini-large "$prefix-gemini-large.json" || status=$?; fi
-  if (( status == 0 )); then run_load_profile gemini-peak "$prefix-gemini-peak.json" || status=$?; fi
-  if (( status == 0 )); then run_load_profile video-throughput "$prefix-video-throughput.json" || status=$?; fi
-  if (( status == 0 )); then run_recovery_profile "$prefix-video-recovery.json" || status=$?; fi
+  run_with_pressure_monitor "$pressure_pid" \
+    run_load_profile gemini-baseline "$prefix-gemini-baseline.json" || status=$?
+  if (( status == 0 )); then
+    run_with_pressure_monitor "$pressure_pid" \
+      run_load_profile gemini-large "$prefix-gemini-large.json" || status=$?
+  fi
+  if (( status == 0 )); then
+    run_with_pressure_monitor "$pressure_pid" \
+      run_load_profile gemini-peak "$prefix-gemini-peak.json" || status=$?
+  fi
+  if (( status == 0 )); then
+    run_with_pressure_monitor "$pressure_pid" \
+      run_load_profile video-throughput "$prefix-video-throughput.json" || status=$?
+  fi
+  if (( status == 0 )); then
+    run_with_pressure_monitor "$pressure_pid" \
+      run_recovery_profile "$prefix-video-recovery.json" || status=$?
+  fi
   touch "$pressure_stop"
   wait "$pressure_pid" || pressure_status=$?
   active_pressure_pid=

tests/release/production-acceptance-script-test.sh

@@ -1,4 +1,5 @@
 #!/usr/bin/env bash
+# shellcheck disable=SC2034 # Variables are consumed by dynamically sourced functions.
 set -euo pipefail
 
 root=$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)
@@ -15,6 +16,15 @@ awk '
 # shellcheck source=/dev/null
 source "$tmp/is-release-ancestor.sh"
 
+awk '
+  /^workload_metrics_for_site\(\)/ { capture = 1 }
+  capture && /^current_max_pod_memory_mib\(\)/ { exit }
+  capture { print }
+' "$script" >"$tmp/pressure-monitor.sh"
+
+# shellcheck source=/dev/null
+source "$tmp/pressure-monitor.sh"
+
 cluster_root=$tmp/repository
 git init -q "$cluster_root"
 git -C "$cluster_root" config user.name 'Acceptance Script Test'
@@ -35,4 +45,39 @@ if is_release_ancestor invalid "$third"; then
   exit 1
 fi
 
+remote_kubectl() {
+  printf 0
+}
+
+cluster_ssh() {
+  echo 'zero-replica Worker attempted to collect metrics' >&2
+  return 1
+}
+
+namespace=easyai
+if workload_metrics_for_site worker ningbo; then
+  echo 'zero-replica Worker site returned metrics success' >&2
+  exit 1
+else
+  metrics_status=$?
+fi
+[[ $metrics_status == 2 ]]
+
+long_running_load() {
+  while :; do
+    sleep 1
+  done
+}
+
+(sleep 1; exit 1) &
+pressure_pid=$!
+AI_GATEWAY_ACCEPTANCE_PROCESS_WATCH_SECONDS=1
+active_load_pid=
+if run_with_pressure_monitor "$pressure_pid" long_running_load; then
+  echo 'load continued after the pressure sampler failed' >&2
+  exit 1
+fi
+wait "$pressure_pid" >/dev/null 2>&1 || true
+[[ -z $active_load_pid ]]
+
 echo 'production_acceptance_script_tests=PASS'