fix: 精确回收 NC01 内存并修复 zombie 泄露

config/hwlab-node-control-plane.yaml

@@ -553,10 +553,11 @@ targets:
             - COPY --from=golang-toolchain /usr/local/go /usr/local/go
             - COPY --from=docker-cli /usr/bin/kubectl /usr/local/bin/kubectl
             - ENV PATH=/usr/local/go/bin:$PATH
+            - RUN apt-get update && apt-get install -y --no-install-recommends tini && rm -rf /var/lib/apt/lists/*
             - RUN ln -sf /usr/local/bin/bun /usr/local/bin/bunx
             - ENV HWLAB_CI_NODE_DEPS=/opt/hwlab-ci-node-deps/node_modules
             - RUN set -eu; export HTTP_PROXY="${HTTP_PROXY:-${http_proxy:-}}"; export HTTPS_PROXY="${HTTPS_PROXY:-${https_proxy:-$HTTP_PROXY}}"; export ALL_PROXY="${ALL_PROXY:-${all_proxy:-}}"; export NO_PROXY="${NO_PROXY:-${no_proxy:-}}"; export http_proxy="$HTTP_PROXY"; export https_proxy="$HTTPS_PROXY"; export all_proxy="$ALL_PROXY"; export no_proxy="$NO_PROXY"; export npm_config_registry="https://registry.npmmirror.com/"; export BUN_CONFIG_REGISTRY="https://registry.npmmirror.com/"; export npm_config_noproxy="$NO_PROXY"; if [ -n "$HTTP_PROXY" ]; then export npm_config_proxy="$HTTP_PROXY"; fi; if [ -n "$HTTPS_PROXY" ]; then export npm_config_https_proxy="$HTTPS_PROXY"; fi; export npm_config_fetch_retries=2; export npm_config_fetch_retry_mintimeout=2000; export npm_config_fetch_retry_maxtimeout=16000; export npm_config_fetch_timeout=120000; proxy_label="${HTTP_PROXY:+HTTP_PROXY}"; proxy_label="${proxy_label:-none}"; mkdir -p /opt/hwlab-ci-node-deps; cd /opt/hwlab-ci-node-deps; printf '{"private":true,"dependencies":{}}\n' > package.json; ok=0; delay=2; for attempt in 1 2 3 4 5; do echo "{\"event\":\"tools-yaml-node-npm-install\",\"attempt\":\"$attempt/5\",\"registry\":\"$npm_config_registry\",\"proxy\":\"$proxy_label\"}" >&2; if timeout 180s npm install --package-lock=false --no-save --ignore-scripts --no-audit --no-fund --omit=dev yaml@2.8.3; then ok=1; break; fi; if [ "$attempt" = 5 ]; then break; fi; echo "{\"event\":\"tools-yaml-node-npm-install\",\"status\":\"retrying\",\"attempt\":\"$attempt/5\",\"sleepSeconds\":$delay}" >&2; sleep "$delay"; delay=$((delay * 2)); done; test "$ok" = 1; node --input-type=module -e 'import("/opt/hwlab-ci-node-deps/node_modules/yaml/browser/dist/index.js").then((yaml)=>console.log("yaml-ok", typeof yaml.parse))'
-            - RUN node --version && npm --version && bun --version && git --version && python3 --version && docker --version && kubectl version --client=true && ssh -V && go version
+            - RUN node --version && npm --version && bun --version && git --version && python3 --version && docker --version && kubectl version --client=true && ssh -V && go version && tini --version
         buildArgs: {}
         buildNetwork: host
         publicBaseImages:
@@ -771,7 +772,8 @@ targets:
             - RUN ln -sf /usr/local/bin/bun /usr/local/bin/bunx
             - ENV HWLAB_CI_NODE_DEPS=/opt/hwlab-ci-node-deps/node_modules
             - RUN set -eu; export HTTP_PROXY="${HTTP_PROXY:-${http_proxy:-}}"; export HTTPS_PROXY="${HTTPS_PROXY:-${https_proxy:-$HTTP_PROXY}}"; export ALL_PROXY="${ALL_PROXY:-${all_proxy:-}}"; export NO_PROXY="${NO_PROXY:-${no_proxy:-}}"; export http_proxy="$HTTP_PROXY"; export https_proxy="$HTTPS_PROXY"; export all_proxy="$ALL_PROXY"; export no_proxy="$NO_PROXY"; export npm_config_registry="https://registry.npmmirror.com/"; export BUN_CONFIG_REGISTRY="https://registry.npmmirror.com/"; export npm_config_noproxy="$NO_PROXY"; if [ -n "$HTTP_PROXY" ]; then export npm_config_proxy="$HTTP_PROXY"; fi; if [ -n "$HTTPS_PROXY" ]; then export npm_config_https_proxy="$HTTPS_PROXY"; fi; export npm_config_fetch_retries=2; export npm_config_fetch_retry_mintimeout=2000; export npm_config_fetch_retry_maxtimeout=16000; export npm_config_fetch_timeout=120000; proxy_label="${HTTP_PROXY:+HTTP_PROXY}"; proxy_label="${proxy_label:-none}"; mkdir -p /opt/hwlab-ci-node-deps; cd /opt/hwlab-ci-node-deps; printf '{"private":true,"dependencies":{}}\n' > package.json; ok=0; delay=2; for attempt in 1 2 3 4 5; do echo "{\"event\":\"tools-yaml-node-npm-install\",\"attempt\":\"$attempt/5\",\"registry\":\"$npm_config_registry\",\"proxy\":\"$proxy_label\"}" >&2; if timeout 180s npm install --package-lock=false --no-save --ignore-scripts --no-audit --no-fund --omit=dev yaml@2.8.3; then ok=1; break; fi; if [ "$attempt" = 5 ]; then break; fi; echo "{\"event\":\"tools-yaml-node-npm-install\",\"status\":\"retrying\",\"attempt\":\"$attempt/5\",\"sleepSeconds\":$delay}" >&2; sleep "$delay"; delay=$((delay * 2)); done; test "$ok" = 1; node --input-type=module -e 'import("/opt/hwlab-ci-node-deps/node_modules/yaml/browser/dist/index.js").then((yaml)=>console.log("yaml-ok", typeof yaml.parse))'
-            - RUN node --version && npm --version && bun --version && git --version && python3 --version && docker --version && kubectl version --client=true && ssh -V && go version
+            - RUN apt-get update && apt-get install -y --no-install-recommends tini && rm -rf /var/lib/apt/lists/*
+            - RUN node --version && npm --version && bun --version && git --version && python3 --version && docker --version && kubectl version --client=true && ssh -V && go version && tini --version
         buildArgs: {}
         buildNetwork: host
         publicBaseImages:
@@ -991,7 +993,8 @@ targets:
             - RUN ln -sf /usr/local/bin/bun /usr/local/bin/bunx
             - ENV HWLAB_CI_NODE_DEPS=/opt/hwlab-ci-node-deps/node_modules
             - RUN set -eu; export HTTP_PROXY="${HTTP_PROXY:-${http_proxy:-}}"; export HTTPS_PROXY="${HTTPS_PROXY:-${https_proxy:-$HTTP_PROXY}}"; export ALL_PROXY="${ALL_PROXY:-${all_proxy:-}}"; export NO_PROXY="${NO_PROXY:-${no_proxy:-}}"; export http_proxy="$HTTP_PROXY"; export https_proxy="$HTTPS_PROXY"; export all_proxy="$ALL_PROXY"; export no_proxy="$NO_PROXY"; export npm_config_registry="https://registry.npmmirror.com/"; export BUN_CONFIG_REGISTRY="https://registry.npmmirror.com/"; export npm_config_noproxy="$NO_PROXY"; if [ -n "$HTTP_PROXY" ]; then export npm_config_proxy="$HTTP_PROXY"; fi; if [ -n "$HTTPS_PROXY" ]; then export npm_config_https_proxy="$HTTPS_PROXY"; fi; export npm_config_fetch_retries=2; export npm_config_fetch_retry_mintimeout=2000; export npm_config_fetch_retry_maxtimeout=16000; export npm_config_fetch_timeout=120000; proxy_label="${HTTP_PROXY:+HTTP_PROXY}"; proxy_label="${proxy_label:-none}"; mkdir -p /opt/hwlab-ci-node-deps; cd /opt/hwlab-ci-node-deps; printf '{"private":true,"dependencies":{}}\n' > package.json; ok=0; delay=2; for attempt in 1 2 3 4 5; do echo "{\"event\":\"tools-yaml-node-npm-install\",\"attempt\":\"$attempt/5\",\"registry\":\"$npm_config_registry\",\"proxy\":\"$proxy_label\"}" >&2; if timeout 180s npm install --package-lock=false --no-save --ignore-scripts --no-audit --no-fund --omit=dev yaml@2.8.3; then ok=1; break; fi; if [ "$attempt" = 5 ]; then break; fi; echo "{\"event\":\"tools-yaml-node-npm-install\",\"status\":\"retrying\",\"attempt\":\"$attempt/5\",\"sleepSeconds\":$delay}" >&2; sleep "$delay"; delay=$((delay * 2)); done; test "$ok" = 1; node --input-type=module -e 'import("/opt/hwlab-ci-node-deps/node_modules/yaml/browser/dist/index.js").then((yaml)=>console.log("yaml-ok", typeof yaml.parse))'
-            - RUN node --version && npm --version && bun --version && git --version && python3 --version && docker --version && kubectl version --client=true && ssh -V && go version
+            - RUN apt-get update && apt-get install -y --no-install-recommends tini && rm -rf /var/lib/apt/lists/*
+            - RUN node --version && npm --version && bun --version && git --version && python3 --version && docker --version && kubectl version --client=true && ssh -V && go version && tini --version
         buildArgs: {}
         buildNetwork: host
         publicBaseImages:
@@ -1252,7 +1255,8 @@ targets:
             - RUN ln -sf /usr/local/bin/bun /usr/local/bin/bunx
             - ENV HWLAB_CI_NODE_DEPS=/opt/hwlab-ci-node-deps/node_modules
             - RUN set -eu; unset HTTP_PROXY HTTPS_PROXY ALL_PROXY http_proxy https_proxy all_proxy; export npm_config_registry="https://registry.npmmirror.com/"; export BUN_CONFIG_REGISTRY="https://registry.npmmirror.com/"; export npm_config_fetch_retries=2; export npm_config_fetch_retry_mintimeout=2000; export npm_config_fetch_retry_maxtimeout=16000; export npm_config_fetch_timeout=120000; mkdir -p /opt/hwlab-ci-node-deps; cd /opt/hwlab-ci-node-deps; printf '{"private":true,"dependencies":{}}\n' > package.json; ok=0; delay=2; for attempt in 1 2 3 4 5; do echo "{\"event\":\"tools-yaml-node-npm-install\",\"attempt\":\"$attempt/5\",\"registry\":\"$npm_config_registry\",\"proxy\":\"none\"}" >&2; if timeout 180s npm install --package-lock=false --no-save --ignore-scripts --no-audit --no-fund --omit=dev yaml@2.8.3; then ok=1; break; fi; if [ "$attempt" = 5 ]; then break; fi; echo "{\"event\":\"tools-yaml-node-npm-install\",\"status\":\"retrying\",\"attempt\":\"$attempt/5\",\"sleepSeconds\":$delay}" >&2; sleep "$delay"; delay=$((delay * 2)); done; test "$ok" = 1; node --input-type=module -e 'import("/opt/hwlab-ci-node-deps/node_modules/yaml/browser/dist/index.js").then((yaml)=>console.log("yaml-ok", typeof yaml.parse))'
-            - RUN node --version && npm --version && bun --version && git --version && python3 --version && docker --version && kubectl version --client=true && ssh -V && go version
+            - RUN apt-get update && apt-get install -y --no-install-recommends tini && rm -rf /var/lib/apt/lists/*
+            - RUN node --version && npm --version && bun --version && git --version && python3 --version && docker --version && kubectl version --client=true && ssh -V && go version && tini --version
         buildArgs: {}
         buildNetwork: host
         publicBaseImages:

config/unidesk-cli.yaml

@@ -223,7 +223,7 @@ gc:
               minAgeHours: 336
               maxDeletePerRun: 5
         memoryPressure:
-          planPreviewLimit: 6
+          planPreviewLimit: 4
           processPatterns:
             - k3s-server
             - containerd
@@ -291,12 +291,86 @@ gc:
             targetMemAvailableBytes: 5368709120
             minimumSwapFreeBytesAfterReclaim: 34359738368
             targets:
-              - id: kubepods-inactive-anon
+              - id: kafka-broker-inactive-anon
-                path: /sys/fs/cgroup/kubepods.slice
+                workloadSelector:
-                reclaimBytes: 3221225472
+                  namespace: platform-infra
+                  matchLabels:
+                    strimzi.io/name: platform-infra-kafka-kafka
+                  maxMatchedPods: 1
+                reclaimBytes: 268435456
                 swappiness: 200
-                minimumInactiveAnonBytes: 3221225472
+                minimumInactiveAnonBytes: 268435456
                 priority: 10
+              - id: kafka-entity-operator-inactive-anon
+                workloadSelector:
+                  namespace: platform-infra
+                  matchLabels:
+                    strimzi.io/name: platform-infra-kafka-entity-operator
+                  maxMatchedPods: 1
+                reclaimBytes: 134217728
+                swappiness: 200
+                minimumInactiveAnonBytes: 134217728
+                priority: 20
+              - id: argocd-controller-inactive-anon
+                workloadSelector:
+                  namespace: argocd
+                  matchLabels:
+                    app.kubernetes.io/name: argocd-application-controller
+                  maxMatchedPods: 1
+                reclaimBytes: 134217728
+                swappiness: 200
+                minimumInactiveAnonBytes: 134217728
+                priority: 30
+              - id: gitea-inactive-anon
+                workloadSelector:
+                  namespace: devops-infra
+                  matchLabels:
+                    app.kubernetes.io/name: gitea
+                  maxMatchedPods: 1
+                reclaimBytes: 134217728
+                swappiness: 200
+                minimumInactiveAnonBytes: 134217728
+                priority: 40
+              - id: hwlab-v03-hwpod-worker-inactive-anon
+                workloadSelector:
+                  namespace: hwlab-v03
+                  matchLabels:
+                    app.kubernetes.io/name: hwlab-hwpod-worker
+                  maxMatchedPods: 2
+                reclaimBytes: 134217728
+                swappiness: 200
+                minimumInactiveAnonBytes: 134217728
+                priority: 50
+              - id: sub2rank-development-worker-inactive-anon
+                workloadSelector:
+                  namespace: platform-infra-development
+                  matchLabels:
+                    app.kubernetes.io/name: sub2rank-worker
+                  maxMatchedPods: 1
+                reclaimBytes: 134217728
+                swappiness: 200
+                minimumInactiveAnonBytes: 134217728
+                priority: 60
+              - id: hwlab-v03-harnessrl-worker-inactive-anon
+                workloadSelector:
+                  namespace: hwlab-v03
+                  matchLabels:
+                    app.kubernetes.io/name: hwlab-harnessrl-worker
+                  maxMatchedPods: 2
+                reclaimBytes: 134217728
+                swappiness: 200
+                minimumInactiveAnonBytes: 134217728
+                priority: 70
+              - id: hwlab-v03-workbench-worker-inactive-anon
+                workloadSelector:
+                  namespace: hwlab-v03
+                  matchLabels:
+                    app.kubernetes.io/name: hwlab-workbench-worker
+                  maxMatchedPods: 2
+                reclaimBytes: 134217728
+                swappiness: 200
+                minimumInactiveAnonBytes: 134217728
+                priority: 80
         kubernetesObjectRetention:
           enabled: true
           aggregatePreviewLimit: 1

project-management/PJ2026-01/specs/PJ2026-01060107-nc01-ci-resource-governance.md

@@ -96,10 +96,17 @@ CoreDNS 和 k3s 控制面使用 Kubernetes 系统关键优先级。Sub2API 和
 
 中心 PaC owning YAML 必须声明 `deliveryTrigger.mode=manual-plan-confirm` 和 `deliveryTrigger.mechanism=pac-webhook`。PaC webhook 接收端继续作为唯一 PipelineRun 创建入口，但 Gitea push hook、PR merge callback、branch follower、poller 和其他自动发送方不得调用它；PR merge 后只允许 source mirror 或 authority 更新。受控 apply 必须删除或禁用 Gitea 自动 hook，并在 status 中把自动发送方启用视为配置漂移。
 
+### 5.7 CI-RESOURCE-REQ-007 子进程回收
+
+长期运行且会启动 Git、SSH、编译器或其他子进程的容器必须具有明确的子进程回收者。业务进程不得在没有 init/reaper 的情况下直接作为容器 PID 1；已经正确等待直接子进程的实现仍须处理其退出后被 PID 1 接管的孤儿后代。init/reaper 必须只负责信号转发和子进程回收，不得成为第二业务生命周期 authority。
+
+运行面诊断必须按父 PID 聚合 zombie，并映射到 namespace、Pod、稳定 workload owner 和修复类别。修复通过正常自动交付滚动 owner；直接终止 zombie、批量杀进程或只重启 Pod 不能作为根因修复。
+
 ## 6. 验收
 
 - L0：配置解析、资源 quantity、requests/limits 关系、matrix 收敛、有界 worker 队列和依赖保持通过轻量验证。
 - L1：受控 `plan/status/diagnosis` 显示 YAML 来源、并发预算、批次、资源和保护状态，不依赖裸 Kubernetes 命令。
+- L1：所有会启动子进程的常驻 Pod 均显示 init/reaper 为 PID 1；在真实 Git/SSH 操作前后，owner 的 zombie 数不增长。
 - L2：正常 HWLAB source PR merge 后不产生 PipelineRun。受控 `plan` 显示 env reuse、镜像构建数量、精确构建与 rollout 范围且没有非预期扩大；随后使用手动 `trigger` 向 PaC 发送一次 webhook，并由 PaC 创建一次 PipelineRun。观察到同一时刻构建 TaskRun 不超过 YAML 预算，相同发布意图不产生第二次构建，所有构建 Pod 带资源合同，并且 NC01 Node、CoreDNS、Sub2API、Public Edge 和公开 API 在构建期间保持健康。
 
 验收失败时保留计划、手动触发链和运行证据，并修复 owning YAML、planner 或 renderer。重启 k3s、删除 Pod/PipelineRun、手工 Argo sync、运行时 patch、mirror flush 或裸补跑均不能作为最终通过证据。

project-management/PJ2026-01/specs/PJ2026-010603-yaml-first-ops.md

@@ -71,6 +71,8 @@ YAML运维负责 HWLAB/UniDesk 自有平台配置的真相源、解析、渲染
 | 有界默认输出 | CLI 在未显式请求机器输出或完整输出时返回的短摘要、表格和 drill-down 命令集合。 |
 | host proxy client | 新节点 k3s 安装前由 `trans` 分发的 0 依赖出网代理客户端，用于给 Docker、k3s、containerd、包管理器、Git 和 git-mirror 提供初始出网。 |
 | 配置组合 | 通过 baseline、configRef、继承、覆盖和变量渲染生成有效计划的配置方式；组合结果是执行视图，不是第二份可编辑真相源。 |
+| 稳定工作负载选择器 | 由 namespace 和 Kubernetes labels 组成、跨 Pod 重建保持不变的工作负载身份；Pod UID 只能是运行面解析结果。 |
+| 精确 cgroup 回收 | 受控 CLI 将稳定工作负载选择器解析为当前 Running Pod cgroup，并在执行前重新解析身份后请求有界 `memory.reclaim`。 |
 
 ## 4. 系统边界和接口
 
@@ -103,6 +105,7 @@ YAML运维负责 HWLAB/UniDesk 自有平台配置的真相源、解析、渲染
 | PJ2026-01060312 | 配置组合 | 本规格 6.10 | baseline、configRef、继承、覆盖、变量渲染和重复块治理 | 配置真相、目标解析、公共原语 | 多节点 YAML 复用、长期维护 |
 | PJ2026-01060313 | K3s本地Registry | 本规格 6.11 | node-local registry 的 k3s workload/PVC、零种子迁移、镜像预加载和 Docker 依赖分类 | Host引导代理、目标解析、发布流水 | 纯净 k3s bring-up、HWLAB/AgentRun 构建 |
 | PJ2026-01060314 | Hyper-V Ubuntu虚拟机 | [PJ2026-01060314 Hyper-V Ubuntu虚拟机](PJ2026-01060314-hyperv-ubuntu-vm.md) | Windows Hyper-V 宿主上的 Ubuntu VM、网络、初始化、常驻和 Provider 接入 | 配置真相、目标解析、Host引导代理 | Docker、k3s、provider-gateway、trans |
+| PJ2026-01060315 | 精确工作负载回收 | 本规格 6.12 | 稳定工作负载选择器、当前 Pod cgroup 解析和有界内存回收 | 配置真相、目标解析、公共原语 | 平台 GC、核心运行面保护 |
 
 ## 6. 原子需求
 
@@ -239,3 +242,15 @@ YAML运维应允许新节点在没有旧 registry 数据和没有长期 host Doc
 当迁移目标是证明从零部署能力时，受控 CLI 应先停止旧 host Docker registry，并在状态中明确 `seededFromOldRegistry=false`、`zeroSeeded=true`、PVC bound、workload ready 和 endpoint ready。迁移后所需的 BuildKit、runtime/base、tools image 和 service build artifact 必须通过 YAML-first `runtime-image preload`、tools image build/status 或等价 k3s 构建入口补齐，不能通过旧 registry 拷贝伪装成从零成功。
 
 YAML运维状态输出应区分三类依赖：必须保留的 host bootstrap 面（provider-gateway/trans、0 依赖 host proxy、k3s installer/systemd）、待迁移的 host Docker/one-shot build 面，以及已经进入 k3s 的长期服务。后续减少 Docker 依赖时，应优先把长期服务、registry、git-mirror cache、web-probe/哨兵和构建执行迁入 k3s；host proxy client 不属于迁移目标。
+
+### 6.12 OPS-YAML-REQ-012 精确工作负载内存回收
+
+| 编号 | 短名 | 主责模块 | 关联模块 |
+| --- | --- | --- | --- |
+| OPS-YAML-REQ-012 | 精确工作负载回收 | PJ2026-01060315 精确工作负载回收 | [平台运维](PJ2026-0106-platform-ops.md)、[NC01 CI资源治理](PJ2026-01060107-nc01-ci-resource-governance.md) |
+
+平台 GC 应允许 owning YAML 使用稳定工作负载选择器声明精确 cgroup 回收目标。选择器至少包含 namespace、非空 labels 和最大匹配 Pod 数；目标可用内存、单 Pod 回收上限、swappiness、inactive anonymous memory 下限和 swap 保留线继续由 YAML 声明。YAML 不得保存 Pod UID、容器 ID、当前 cgroup 路径或从运行面反解出的可调数值。
+
+受控 plan 只选择 Running Pod，并披露 selector、解析出的 namespace、Pod、UID、cgroup、当前内存、inactive anonymous memory 和保护原因。零匹配、超过最大匹配数、API 不可读、cgroup 缺失、symlink、内存下限不足或 swap 保留线不足时必须保护，禁止回退到父 cgroup、按名称猜测或扩大 selector。
+
+执行前必须使用同一 YAML selector 重新解析当前 Running Pod，并逐项核对 Pod UID、cgroup 路径和全部回收参数。身份变化、目标已满足或候选不再合格时跳过且结构化披露；只有身份闭合的当前 cgroup 才能写入有界 `memory.reclaim`。执行不得停止、重启、驱逐或迁移 workload，也不得把内核请求上限宣称为实际收益；终态继续分别披露 cgroup 下降、宿主 `MemAvailable` 变化、目标缺口和 swap 排除状态。

scripts/src/agentrun-managed-repository-reconciler.test.ts

@@ -59,6 +59,7 @@ describe("AgentRun YAML-owned managed repository reconciler", () => {
     expect(documents[2].spec.template.spec.automountServiceAccountToken).toBe(false);
     expect(documents[2].spec.template.spec.hostNetwork).toBe(spec.gitMirror.repositoryReconciler.hostNetwork);
     expect(documents[2].spec.template.spec.dnsPolicy).toBe(spec.gitMirror.repositoryReconciler.dnsPolicy);
+    expect(documents[2].spec.template.spec.containers[0].command.slice(0, 2)).toEqual(["tini", "--"]);
     expect(documents[2].spec.template.spec.containers[0].readinessProbe.exec.command).toContain("--health");
     const authVolume = documents[2].spec.template.spec.volumes.find((item: any) => item.name === "git-auth");
     expect(authVolume.projected.sources).toHaveLength(2);

scripts/src/agentrun-managed-repository-reconciler.ts

@@ -235,6 +235,8 @@ function managedRepositoryObjects(spec: AgentRunLaneSpec): Record<string, unknow
               image: spec.gitMirror.toolsImage,
               imagePullPolicy: reconciler.imagePullPolicy,
               command: [
+                "tini",
+                "--",
                 "node",
                 `${controllerMountPath}/controller.mjs`,
                 "--config",

scripts/src/gc-remote-memory-pressure.test.ts

@@ -123,6 +123,13 @@ test("memory-pressure async runner uses a parent-child barrier and locked worker
   assert.match(runnerSource, /worker-identity-unavailable|worker-exited-without-terminal-state/u);
 });
 
+test("workload-selected cgroup reclaim re-resolves the Pod UID and path before execution", () => {
+  assert.match(webObserveSource, /_CGROUP_RECLAIM_RUNTIME_INVENTORY = None/u);
+  assert.match(webObserveSource, /cgroup reclaim candidate identity changed before execution/u);
+  assert.match(webObserveSource, /item\.get\("path"\) == path/u);
+  assert.match(webObserveSource, /item\.get\("pod"\).*get\("uid"\).*resolved_uid/u);
+});
+
 test("memory-pressure plan includes only YAML allowlisted cgroup reclaim stages with swap reserve", () => {
   const result = runPythonFixture(String.raw`
 MEMORY_CONFIG = {
@@ -239,7 +246,7 @@ test("default memory-pressure plan projects an actionable YAML-bounded preview b
   assert.equal(projected.runEligibility.allowed, true);
   assert.equal(projected.runEligibility.allCandidatesEvaluated, true);
   assert.equal(projected.candidatePreview.limit, remoteTarget.memoryPressure.planPreviewLimit);
-  assert.equal(projected.candidatePreview.candidates.length, 6);
+  assert.equal(projected.candidatePreview.candidates.length, Math.min(candidates.length, remoteTarget.memoryPressure.planPreviewLimit));
   assert.equal(projected.candidatePreview.candidates[2].identityClosure.identitiesComplete, true);
   assert.equal(projected.candidatePreview.candidates[2].processIdentities, undefined);
   assert.equal(projected.policy.runCommand, "bun scripts/cli.ts gc remote NC01 run --confirm --memory-pressure-only");

scripts/src/gc-remote-web-observe.py

@@ -359,15 +359,40 @@ def configured_cgroup_reclaim_targets():
             return []
         target_id = item.get("id")
         path = item.get("path")
+        workload_selector = item.get("workloadSelector")
         reclaim_bytes = item.get("reclaimBytes")
         swappiness = item.get("swappiness")
         minimum_inactive_anon = item.get("minimumInactiveAnonBytes", 0)
         if not isinstance(target_id, str) or not re.match(r"^[A-Za-z0-9._-]{1,80}$", target_id) or target_id in seen_ids:
             _MEMORY_RECLAIM_CONFIG_ERROR = "yaml-cgroup-reclaim-target-id-invalid:%s" % index
             return []
+        if (path is None) == (workload_selector is None):
+            _MEMORY_RECLAIM_CONFIG_ERROR = "yaml-cgroup-reclaim-target-identity-invalid:%s" % target_id
+            return []
+        if path is not None:
             if not isinstance(path, str) or (path != "/sys/fs/cgroup" and not path.startswith("/sys/fs/cgroup/")) or os.path.normpath(path) != path:
                 _MEMORY_RECLAIM_CONFIG_ERROR = "yaml-cgroup-reclaim-target-path-invalid:%s" % target_id
                 return []
+        if workload_selector is not None:
+            if not isinstance(workload_selector, dict):
+                _MEMORY_RECLAIM_CONFIG_ERROR = "yaml-cgroup-reclaim-workload-selector-invalid:%s" % target_id
+                return []
+            namespace = workload_selector.get("namespace")
+            match_labels = workload_selector.get("matchLabels")
+            max_matched_pods = workload_selector.get("maxMatchedPods")
+            if not isinstance(namespace, str) or not re.match(r"^[a-z0-9]([-a-z0-9]*[a-z0-9])?$", namespace):
+                _MEMORY_RECLAIM_CONFIG_ERROR = "yaml-cgroup-reclaim-workload-namespace-invalid:%s" % target_id
+                return []
+            if not isinstance(match_labels, dict) or not match_labels or any(
+                not isinstance(key, str) or not re.match(r"^[A-Za-z0-9./_-]{1,253}$", key)
+                or not isinstance(value, str) or not re.match(r"^[A-Za-z0-9._-]{1,63}$", value)
+                for key, value in match_labels.items()
+            ):
+                _MEMORY_RECLAIM_CONFIG_ERROR = "yaml-cgroup-reclaim-workload-labels-invalid:%s" % target_id
+                return []
+            if isinstance(max_matched_pods, bool) or not isinstance(max_matched_pods, int) or max_matched_pods < 1 or max_matched_pods > 20:
+                _MEMORY_RECLAIM_CONFIG_ERROR = "yaml-cgroup-reclaim-workload-max-pods-invalid:%s" % target_id
+                return []
         if isinstance(reclaim_bytes, bool) or not isinstance(reclaim_bytes, int) or reclaim_bytes <= 0 or reclaim_bytes > 8 * 1024 * 1024 * 1024:
             _MEMORY_RECLAIM_CONFIG_ERROR = "yaml-cgroup-reclaim-bytes-invalid:%s" % target_id
             return []
@@ -388,6 +413,7 @@ def configured_cgroup_reclaim_targets():
         validated.append({
             "id": target_id,
             "path": path,
+            "workloadSelector": workload_selector,
             "reclaimBytes": reclaim_bytes,
             "swappiness": swappiness,
             "minimumInactiveAnonBytes": minimum_inactive_anon,
@@ -397,6 +423,92 @@ def configured_cgroup_reclaim_targets():
         })
     return sorted(validated, key=lambda item: item.get("priority"))
 
+_CGROUP_RECLAIM_RUNTIME_INVENTORY = None
+
+def cgroup_reclaim_runtime_inventory():
+    global _CGROUP_RECLAIM_RUNTIME_INVENTORY
+    if _CGROUP_RECLAIM_RUNTIME_INVENTORY is not None:
+        return _CGROUP_RECLAIM_RUNTIME_INVENTORY
+    attribution = MEMORY_CONFIG.get("attribution") or {}
+    timeout = safe_int(attribution.get("commandTimeoutSeconds"))
+    runtime = command(["k3s", "crictl", "pods", "-o", "json"], timeout)
+    if runtime.get("exitCode") != 0:
+        _CGROUP_RECLAIM_RUNTIME_INVENTORY = {
+            "ok": False,
+            "reason": "workload-runtime-inventory-failed",
+            "command": bounded(runtime),
+            "pods": [],
+        }
+        return _CGROUP_RECLAIM_RUNTIME_INVENTORY
+    try:
+        payload = json.loads(runtime.get("stdout") or "{}")
+    except (TypeError, ValueError, json.JSONDecodeError):
+        _CGROUP_RECLAIM_RUNTIME_INVENTORY = {"ok": False, "reason": "workload-runtime-inventory-invalid", "pods": []}
+        return _CGROUP_RECLAIM_RUNTIME_INVENTORY
+    cgroup_root = attribution.get("cgroupRoot")
+    pod_paths = {}
+    scan_limit = safe_int(attribution.get("cgroupScanLimit"))
+    scanned = 0
+    for current, directories, files in os.walk(os.path.join(cgroup_root, "kubepods.slice"), topdown=True, followlinks=False):
+        directories[:] = sorted(name for name in directories if not os.path.islink(os.path.join(current, name)))
+        scanned += 1
+        if scanned > scan_limit:
+            _CGROUP_RECLAIM_RUNTIME_INVENTORY = {"ok": False, "reason": "workload-cgroup-scan-truncated", "pods": []}
+            return _CGROUP_RECLAIM_RUNTIME_INVENTORY
+        match = re.search(r"pod([0-9a-f_]{36})\.slice$", os.path.basename(current))
+        if match is not None and "memory.current" in files:
+            pod_paths[match.group(1).replace("_", "-")] = current
+            directories[:] = []
+    pods = []
+    for pod in payload.get("items") or []:
+        metadata = pod.get("metadata") if isinstance(pod, dict) else None
+        uid = metadata.get("uid") if isinstance(metadata, dict) else None
+        name = metadata.get("name") if isinstance(metadata, dict) else None
+        namespace = metadata.get("namespace") if isinstance(metadata, dict) else None
+        labels = pod.get("labels") if isinstance(pod, dict) else None
+        if pod.get("state") != "SANDBOX_READY" or not all(isinstance(value, str) and value for value in [uid, name, namespace]):
+            continue
+        pods.append({
+            "namespace": namespace,
+            "name": name,
+            "uid": uid,
+            "labels": labels if isinstance(labels, dict) else {},
+            "path": pod_paths.get(uid),
+        })
+    _CGROUP_RECLAIM_RUNTIME_INVENTORY = {"ok": True, "pods": pods}
+    return _CGROUP_RECLAIM_RUNTIME_INVENTORY
+
+def resolve_cgroup_reclaim_target(item):
+    path = item.get("path")
+    if isinstance(path, str):
+        return {"ok": True, "matches": [{"path": path, "pod": None}]}
+    selector = item.get("workloadSelector") or {}
+    labels = selector.get("matchLabels") or {}
+    label_selector = ",".join("%s=%s" % pair for pair in sorted(labels.items()))
+    inventory = cgroup_reclaim_runtime_inventory()
+    if inventory.get("ok") is not True:
+        return {"ok": False, "reason": inventory.get("reason"), "command": inventory.get("command"), "matches": []}
+    pods = [pod for pod in inventory.get("pods") or [] if
+        pod.get("namespace") == selector.get("namespace")
+        and all((pod.get("labels") or {}).get(key) == value for key, value in labels.items())
+    ]
+    if len(pods) == 0:
+        return {"ok": False, "reason": "workload-selector-no-running-pods", "matches": []}
+    if len(pods) > safe_int(selector.get("maxMatchedPods")):
+        return {"ok": False, "reason": "workload-selector-max-pods-exceeded", "matchedPodCount": len(pods), "matches": []}
+    matches = []
+    for pod in pods:
+        uid = pod.get("uid")
+        name = pod.get("name")
+        resolved_path = pod.get("path")
+        if not isinstance(uid, str) or not isinstance(name, str) or not isinstance(resolved_path, str):
+            return {"ok": False, "reason": "workload-pod-cgroup-unavailable", "matches": []}
+        matches.append({
+            "path": resolved_path,
+            "pod": {"namespace": selector.get("namespace"), "name": name, "uid": uid, "labelSelector": label_selector},
+        })
+    return {"ok": True, "matches": matches}
+
 def collect_cgroup_reclaim_candidates():
     global _MEMORY_RECLAIM_CONFIG_ERROR, _CGROUP_RECLAIM_DIAGNOSTICS
     targets = configured_cgroup_reclaim_targets()
@@ -412,7 +524,18 @@ def collect_cgroup_reclaim_candidates():
     swap_free = safe_int((host.get("swap") or {}).get("freeBytes"))
     candidates = []
     for item in targets:
-        path = item.get("path")
+        resolution = resolve_cgroup_reclaim_target(item)
+        if resolution.get("ok") is not True:
+            diagnostics["protected"].append({
+                "id": item.get("id"),
+                "workloadSelector": item.get("workloadSelector"),
+                "reason": resolution.get("reason"),
+                "matchedPodCount": resolution.get("matchedPodCount"),
+            })
+            continue
+        for resolved in resolution.get("matches") or []:
+            path = resolved.get("path")
+            pod = resolved.get("pod")
             reclaim_path = os.path.join(path, "memory.reclaim")
             protected_reason = None
             if os.path.realpath(path) != path or not os.path.isdir(path) or os.path.islink(path):
@@ -432,20 +555,28 @@ def collect_cgroup_reclaim_candidates():
                 diagnostics["protected"].append({
                     "id": item.get("id"),
                     "path": path,
+                    "workloadSelector": item.get("workloadSelector"),
+                    "resolvedPod": pod,
                     "reason": protected_reason,
                     "memoryCurrentBytes": current,
                     "inactiveAnonBytes": inactive_anon,
                 })
                 continue
             reclaim_bytes = safe_int(item.get("reclaimBytes"))
+            identity_suffix = (pod or {}).get("uid")
+            candidate_id = "cgroup-memory-reclaim:%s" % item.get("id")
+            if identity_suffix:
+                candidate_id += ":%s" % identity_suffix
             candidates.append({
-            "id": "cgroup-memory-reclaim:%s" % item.get("id"),
+                "id": candidate_id,
                 "configId": item.get("id"),
                 "kind": "cgroup-memory-reclaim",
                 "risk": "medium",
-            "description": "Request bounded cgroup v2 memory reclaim from an exact YAML allowlist path",
+                "description": "Request bounded cgroup v2 memory reclaim from a YAML-resolved workload",
                 "reason": "host-memavailable-below-yaml-target",
                 "path": path,
+                "workloadSelector": item.get("workloadSelector"),
+                "resolvedPod": pod,
                 "priority": safe_int(item.get("priority")),
                 "reclaimBytes": reclaim_bytes,
                 "swappiness": safe_int(item.get("swappiness")),
@@ -462,7 +593,7 @@ def collect_cgroup_reclaim_candidates():
                 "expectedMemoryRss": fmt_bytes(reclaim_bytes),
                 "estimatedDiskBytes": 0,
                 "estimatedReclaimBytes": 0,
-            "action": {"op": "cgroup-v2-memory-reclaim", "allowlist": "yaml-exact-cgroup-path"},
+                "action": {"op": "cgroup-v2-memory-reclaim", "allowlist": "yaml-workload-selector-or-exact-path"},
             })
     diagnostics["eligibleCount"] = len(candidates)
     return candidates
@@ -679,14 +810,23 @@ def extend_with_allowlisted_session_members(identities, processes, sid, browser_
     return added
 
 def execute_cgroup_reclaim_candidate(candidate):
+    global _CGROUP_RECLAIM_RUNTIME_INVENTORY
     config_id = candidate.get("configId")
     configured = next((item for item in configured_cgroup_reclaim_targets() if item.get("id") == config_id), None)
     if configured is None:
         raise RuntimeError("cgroup reclaim candidate is no longer present in YAML allowlist")
-    for key in ["path", "reclaimBytes", "swappiness", "minimumInactiveAnonBytes", "targetMemAvailableBytes", "minimumSwapFreeBytesAfterReclaim"]:
+    for key in ["workloadSelector", "reclaimBytes", "swappiness", "minimumInactiveAnonBytes", "targetMemAvailableBytes", "minimumSwapFreeBytesAfterReclaim"]:
         if configured.get(key) != candidate.get(key):
             raise RuntimeError("cgroup reclaim candidate no longer matches YAML: %s" % key)
-    path = configured.get("path")
+    _CGROUP_RECLAIM_RUNTIME_INVENTORY = None
+    resolution = resolve_cgroup_reclaim_target(configured)
+    if resolution.get("ok") is not True:
+        raise RuntimeError("cgroup reclaim workload selector no longer resolves: %s" % resolution.get("reason"))
+    path = candidate.get("path")
+    resolved_uid = ((candidate.get("resolvedPod") or {}).get("uid"))
+    current_match = next((item for item in resolution.get("matches") or [] if item.get("path") == path and ((item.get("pod") or {}).get("uid")) == resolved_uid), None)
+    if current_match is None:
+        raise RuntimeError("cgroup reclaim candidate identity changed before execution")
     reclaim_path = os.path.join(path, "memory.reclaim")
     if os.path.realpath(path) != path or not os.path.isdir(path) or os.path.islink(path):
         raise RuntimeError("cgroup reclaim path is unavailable or redirected")

scripts/src/gc-remote.ts

@@ -411,6 +411,23 @@ function positiveInteger(value: unknown): number | null {
   return typeof value === "number" && Number.isInteger(value) && value > 0 ? value : null;
 }
 
+function compactWorkloadSelector(value: unknown): string | undefined {
+  const selector = recordOrEmpty(value);
+  const namespace = typeof selector.namespace === "string" ? selector.namespace : null;
+  const labels = recordOrEmpty(selector.matchLabels);
+  const pairs = Object.entries(labels)
+    .filter((entry): entry is [string, string] => typeof entry[1] === "string")
+    .sort(([left], [right]) => left.localeCompare(right))
+    .map(([key, label]) => `${key}=${label}`);
+  return namespace !== null && pairs.length > 0 ? `${namespace}:${pairs.join(",")}` : undefined;
+}
+
+function compactResolvedPod(value: unknown): Record<string, unknown> | undefined {
+  const pod = recordOrEmpty(value);
+  if (typeof pod.namespace !== "string" || typeof pod.name !== "string" || typeof pod.uid !== "string") return undefined;
+  return { name: `${pod.namespace}/${pod.name}`, uid: pod.uid };
+}
+
 function compactMemoryPressureCandidate(value: unknown): Record<string, unknown> {
   const candidate = recordOrEmpty(value);
   const action = recordOrEmpty(candidate.action);
@@ -426,14 +443,14 @@ function compactMemoryPressureCandidate(value: unknown): Record<string, unknown>
       ...base,
       configId: candidate.configId,
       path: candidate.path,
+      workloadSelector: compactWorkloadSelector(candidate.workloadSelector),
+      resolvedPod: compactResolvedPod(candidate.resolvedPod),
       reclaimBytes: candidate.reclaimBytes,
       requestedMemoryReclaimBytes: candidate.requestedMemoryReclaimBytes,
       requestedMemoryReclaim: candidate.requestedMemoryReclaim,
       estimateKind: candidate.estimateKind,
       swappiness: candidate.swappiness,
       minimumInactiveAnonBytes: candidate.minimumInactiveAnonBytes,
-      targetMemAvailableBytes: candidate.targetMemAvailableBytes,
-      minimumSwapFreeBytesAfterReclaim: candidate.minimumSwapFreeBytesAfterReclaim,
       memoryCurrentBytes: candidate.memoryCurrentBytes,
       memorySwapCurrentBytes: candidate.memorySwapCurrentBytes,
     };
@@ -478,7 +495,10 @@ function memoryPressureCandidateAllowed(value: unknown): boolean {
       && swappiness !== null && swappiness >= 0 && swappiness <= 200
       && (swappiness === 0 || positiveInteger(candidate.minimumInactiveAnonBytes) !== null)
       && action.op === "cgroup-v2-memory-reclaim"
-      && action.allowlist === "yaml-exact-cgroup-path";
+      && (action.allowlist === "yaml-exact-cgroup-path"
+        || (action.allowlist === "yaml-workload-selector-or-exact-path"
+          && typeof recordOrEmpty(candidate.workloadSelector).namespace === "string"
+          && typeof recordOrEmpty(candidate.resolvedPod).uid === "string"));
   }
   if (candidate.kind !== "browser-orphan-process-tree" && candidate.kind !== "stale-web-observer-process-tree") return false;
   const compact = compactMemoryPressureCandidate(candidate);
@@ -526,7 +546,15 @@ function compactMemoryPressureCandidateScan(value: unknown): Record<string, unkn
       protectedCount: cgroupProtected.length,
       protected: cgroupProtected.map((item) => {
         const protectedItem = recordOrEmpty(item);
-        return { id: protectedItem.id, reason: protectedItem.reason, memoryCurrentBytes: protectedItem.memoryCurrentBytes };
+        return {
+          id: protectedItem.id,
+          reason: protectedItem.reason,
+          matchedPodCount: protectedItem.matchedPodCount,
+          memoryCurrentBytes: protectedItem.memoryCurrentBytes,
+          inactiveAnonBytes: protectedItem.inactiveAnonBytes,
+          workloadSelector: compactWorkloadSelector(protectedItem.workloadSelector),
+          resolvedPod: compactResolvedPod(protectedItem.resolvedPod),
+        };
       }),
     },
     policySource: scan.policySource,

scripts/src/hwlab-node-control-plane.ts

@@ -2423,7 +2423,7 @@ function gitMirrorDeployment(name: string, namespace: string, labels: Record<str
     containers: [{
       name: "git-mirror",
       image: target.tekton.toolsImage.output,
-      command: ["node", "/etc/git-mirror/server.js"],
+      command: ["tini", "--", "node", "/etc/git-mirror/server.js"],
       env: [
         { name: "PORT", value: String(containerPort) },
         { name: "GIT_PROJECT_ROOT", value: "/cache" },

src/components/microservices/todo-note/Dockerfile

@@ -3,7 +3,7 @@ FROM ${TODO_NOTE_BASE_IMAGE}
 
 WORKDIR /app/src/components/microservices/todo-note
 RUN apt-get update \
-  && apt-get install -y --no-install-recommends git openssh-client ca-certificates \
+  && apt-get install -y --no-install-recommends git openssh-client ca-certificates tini \
   && apt-get clean \
   && rm -rf /var/lib/apt/lists/*
 COPY src/components/microservices/todo-note/package.json src/components/microservices/todo-note/bun.lock ./
@@ -13,4 +13,5 @@ COPY src/components/shared /app/src/components/shared
 COPY src/components/microservices/todo-note/src ./src
 
 EXPOSE 4288
+ENTRYPOINT ["tini", "--"]
 CMD ["bun", "run", "src/index.ts"]