perf(worker): 持久化准入快照提升队列吞吐

原因：线上 P24 验收显示 Worker 在每轮准入前逐条恢复媒体、重算候选和查询 attempts，跨地域查询导致 48 个执行槽长期只能使用约 8 至 16 个。\n\n影响：任务首次排队时保存鉴权后的 admission scope 快照，Worker 单次批量读取并只刷新动态总容量；显式重新选路保留完整慢路径。验收模拟器改到非数据库专用 Worker 节点，运行期 85% 作为立即中止硬门禁，80% 继续作为容量认证目标。\n\n验证：Go 全量测试、go vet、govulncheck、真实 PostgreSQL 迁移与跨 Store 集成测试、ShellCheck、发布及验收脚本、OpenAPI、前端 lint/test/build、pnpm audit high、Compose 与 Kubernetes 渲染均通过。
2026-08-01 21:25:54 +08:00
parent c89c56ca65
commit 56a5176c92
10 changed files with 274 additions and 100 deletions
@@ -2630,7 +2630,7 @@ pressure_row_live_gate_reason() {
    $4 >= 900 {print "oldest_wait_seconds=" $4 " limit=900"; exit}
    $6 <= 0 {print "postgres_max_connections_unavailable"; exit}
    $5 * 4 >= $6 * 3 {print "postgres_connections=" $5 " limit_75_percent_of=" $6; exit}
-    $7 >= 80 {print "node_memory_percent=" $7 " limit=80"; exit}
+    $7 >= 85 {print "node_memory_percent=" $7 " hard_limit=85"; exit}
    $8 >= 1536 {print "api_pod_memory_mib=" $8 " limit=1536"; exit}
    $14 > 0 {print "lease_renewal_failures=" $14; exit}
    $15 > 0 {print "lease_lost=" $15; exit}