fix(acceptance): 按自适应容量执行单节点压测

修正单节点验收将 hard limit 等同空闲分配容量的问题，从冷启动容量 4 开始逐级加压，并要求实际分配容量和运行任务峰值均达到目标后才计入稳定并发。\n\n异常退出时同步清理模拟器和回调收集器资源。已执行 bash -n、ShellCheck 和 production acceptance 脚本测试。

scripts/cluster/run-production-acceptance.sh

@@ -132,10 +132,10 @@ fi
 : "${AI_GATEWAY_ACCEPTANCE_OVERLOAD_DURATION:=10m}"
 : "${AI_GATEWAY_ACCEPTANCE_GATEWAY_TLS_SERVER_NAME:=}"
 : "${AI_GATEWAY_ACCEPTANCE_GATEWAYS:=}"
-: "${AI_GATEWAY_ACCEPTANCE_SINGLE_NODE_BASELINE_SLOTS:=8}"
-: "${AI_GATEWAY_ACCEPTANCE_SINGLE_NODE_GEMINI_SLOTS:=8}"
+: "${AI_GATEWAY_ACCEPTANCE_SINGLE_NODE_BASELINE_SLOTS:=4}"
+: "${AI_GATEWAY_ACCEPTANCE_SINGLE_NODE_GEMINI_SLOTS:=4}"
 : "${AI_GATEWAY_ACCEPTANCE_SINGLE_NODE_GEMINI_REQUESTS:=96}"
-: "${AI_GATEWAY_ACCEPTANCE_SINGLE_NODE_VIDEO_SLOTS:=8 12 16 24 32 40 48}"
+: "${AI_GATEWAY_ACCEPTANCE_SINGLE_NODE_VIDEO_SLOTS:=4 8 12 16 24 32 40 48}"
 : "${AI_GATEWAY_ACCEPTANCE_REAL_IMAGE_URLS:=}"
 : "${AI_GATEWAY_ACCEPTANCE_SNAPSHOT_DATABASE_URL:=}"
 : "${AI_GATEWAY_ONLINE_ACCOUNT:=}"
@@ -383,6 +383,9 @@ cleanup() {
   if [[ $single_node_emulator_installed == true ]]; then
     cleanup_single_node_emulator >/dev/null 2>&1 || true
   fi
+  if [[ $single_node_acceptance == true ]]; then
+    cleanup_single_node_kubernetes_resources >/dev/null 2>&1 || true
+  fi
   if (( status != 0 )) && [[ -n $run_id && $failure_recorded != true ]]; then
     set +e
     if [[ $single_node_acceptance == true ]]; then
@@ -3500,10 +3503,14 @@ apply_single_node_capacity() {
   local deadline=$((SECONDS + 90)) allocation
   while (( SECONDS < deadline )); do
     allocation=$(database_query "
-SELECT count(*)||':'||COALESCE(max(allocated_capacity),0)
+SELECT count(*)||':'||
+  COALESCE(max(hard_capacity_limit),0)||':'||
+  COALESCE(max(allocated_capacity),0)||':'||
+  COALESCE(max(safe_capacity),0)||':'||
+  COALESCE(max(pressure_state),'unknown')
 FROM gateway_worker_instances
 WHERE status='active' AND heartbeat_at > now()-interval '30 seconds';")
-    if [[ $allocation == "1:$slots" ]]; then
+    if single_node_capacity_ready "$allocation" "$slots"; then
       local metrics_ready=false
       for _ in {1..30}; do
         if remote_kubectl top node easyai-ningbo --no-headers >/dev/null 2>&1 &&
@@ -3520,15 +3527,30 @@ WHERE status='active' AND heartbeat_at > now()-interval '30 seconds';")
         return 1
       }
       active_profile=S$slots
-      echo "single_node_capacity=PASS slots=$slots worker_pool=$worker_pool active_instances=1"
+      local active_instances hard_limit allocated safe_capacity pressure_state
+      IFS=':' read -r active_instances hard_limit allocated safe_capacity pressure_state <<<"$allocation"
+      echo "single_node_capacity=PASS hard_limit=$slots allocated=$allocated safe_capacity=$safe_capacity pressure_state=$pressure_state worker_pool=$worker_pool active_instances=1"
       return 0
     fi
     sleep 2
   done
-  echo "single-node Worker allocation did not converge: expected=1:$slots actual=${allocation:-missing}" >&2
+  echo "single-node Worker capacity did not converge: expected=active:1 hard_limit:$slots allocated_and_safe:1..$slots pressure:not-critical actual=${allocation:-missing}" >&2
   return 1
 }
 
+single_node_capacity_ready() {
+  local state=$1
+  local expected_hard_limit=$2
+  local active_instances hard_limit allocated safe_capacity pressure_state extra
+  IFS=':' read -r active_instances hard_limit allocated safe_capacity pressure_state extra <<<"$state"
+  [[ -z $extra &&
+     $active_instances =~ ^[0-9]+$ && $active_instances -eq 1 &&
+     $hard_limit =~ ^[0-9]+$ && $hard_limit -eq $expected_hard_limit &&
+     $allocated =~ ^[0-9]+$ && $allocated -ge 1 && $allocated -le $expected_hard_limit &&
+     $safe_capacity =~ ^[0-9]+$ && $safe_capacity -ge 1 && $safe_capacity -le $expected_hard_limit &&
+     $pressure_state != critical ]]
+}
+
 install_single_node_load_driver() {
   [[ $remote_load_drivers_installed == false ]] || return 0
   local binary_path env_file remote_env local_digest remote_digest
@@ -3676,6 +3698,17 @@ record_single_node_result() {
       load:$load[0]
     }' >"$summary"
   chmod 0600 "$summary"
+  if (( peak_allocated < slots || peak_running < slots )); then
+    echo "single-node target concurrency was not observed: configured=$slots peak_allocated=$peak_allocated peak_running=$peak_running" >&2
+    return 1
+  fi
+}
+
+cleanup_single_node_kubernetes_resources() {
+  remote_kubectl delete deployment easyai-acceptance-emulator easyai-acceptance-callback-collector \
+    -n "$namespace" --ignore-not-found >/dev/null || true
+  remote_kubectl delete service easyai-acceptance-emulator easyai-acceptance-callback-collector \
+    -n "$namespace" --ignore-not-found >/dev/null || true
 }
 
 run_single_node_profile() {
@@ -3768,10 +3801,7 @@ FROM gateway_tasks WHERE acceptance_run_id='$run_id'::uuid;")
   admin_request POST "/api/admin/system/acceptance/runs/$run_id/finish" "$finish_body" "$finish_response"
   failure_recorded=true
   cleanup_single_node_emulator
-  remote_kubectl delete deployment easyai-acceptance-emulator easyai-acceptance-callback-collector \
-    -n "$namespace" --ignore-not-found >/dev/null || true
-  remote_kubectl delete service easyai-acceptance-emulator easyai-acceptance-callback-collector \
-    -n "$namespace" --ignore-not-found >/dev/null || true
+  cleanup_single_node_kubernetes_resources
 }
 
 run_single_node_acceptance() {