diff --git a/.agents/notes/implemented/process/2026-07-26-ci-failover-runbook.i18n.yaml b/.agents/notes/implemented/process/2026-07-26-ci-failover-runbook.i18n.yaml
index 8615d248fc..ed819bc19e 100644
--- a/.agents/notes/implemented/process/2026-07-26-ci-failover-runbook.i18n.yaml
+++ b/.agents/notes/implemented/process/2026-07-26-ci-failover-runbook.i18n.yaml
@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/process/2026-07-26-ci-failover-runbook.md
-2026-07-26-ci-failover-runbook.md: 47901844f4ec581dff8500cc429c54a076b7642b
-2026-07-26-ci-failover-runbook.zh.md: 88a793e144a4d06718ffe323a94f8e81f8e62b82
+2026-07-26-ci-failover-runbook.md: 90ef2905bec86911697e552c0c1eac96ea0d3a18
+2026-07-26-ci-failover-runbook.zh.md: 5c8fd0e90ae31e3622ae6b49217591a54b84c351
diff --git a/.agents/notes/implemented/process/2026-07-26-ci-failover-runbook.md b/.agents/notes/implemented/process/2026-07-26-ci-failover-runbook.md
index 47901844f4..90ef2905be 100644
--- a/.agents/notes/implemented/process/2026-07-26-ci-failover-runbook.md
+++ b/.agents/notes/implemented/process/2026-07-26-ci-failover-runbook.md
@@ -12,6 +12,12 @@ The three required Linux worker jobs in [CI](../../../../.github/workflows/ci.ym
 
 Each of the three required Linux worker jobs, the independent native Windows job, and the `all checks passed` verdict job — which would otherwise stay queued on the failed pool even after every worker passed — resolves its runner pool through the `DSH_CI_FAILOVER` repository variable. Unset (normal), they run on the hosted enterprise pools. Set to `selfhosted` by any repository writer, all five retarget onto the in-house self-hosted pools: the Linux jobs and verdict onto the `vm-backup` pool, coverage and snapshot concurrency drop to shared-VM bounds, and the hosted-path pnpm cache restores are skipped; the native Windows job onto the `dsh-win-ci` pool. The switch is writer-manageable repository state, not a merge, so it works while every check is red. The in-house pools' readiness is continuously re-proven by the `serial / linux (self-hosted standby)` and `serial / windows (self-hosted standby)` lanes, which run the complete unsharded aggregates on every master push.
 
+`ci.yml` exempts exactly one event from `cancel-in-progress` (`${{ github.event_name != 'push' }}`), so one master push does not cancel the drill still running from the previous one. Each drill runs its complete unsharded aggregate with one gate worker, which takes longer than the interval between master merges; under unconditional cancellation a drill is superseded before reaching a verdict and the lane yields no readiness evidence for a responder to check.
+
+The exemption is narrower than "a drill always finishes", in two ways. GitHub keeps a single pending entry per group, so a newer pending run displaces an older one and intermediate push runs still end as `cancelled` during busy periods. And the expression is evaluated against the *newly triggered* run, so a run whose own event is not `push` — a benchmark dispatched on master, sharing the group `CI-<ref>` — evaluates to `true` and does cancel a drill that is mid-flight. That is a rare manual action and the next master push restores the evidence, so it does not warrant further mechanism. What the carve-out buys is that the lane periodically reaches a verdict at all, which is what makes it usable as evidence.
+
+The decision belongs at workflow level because cancellation applies to the whole superseded run: a job-level `concurrency` group does not exempt its job. The negated form is load-bearing rather than cosmetic: naming `pull_request` alone would also stop cancelling `workflow_dispatch`, and each runner benchmark fans out to twelve larger runners for up to fifteen minutes inside this same group on master, so a re-dispatch would queue ahead of a drill instead of replacing a stale measurement. What bounds the cost is that a master push carries only `wine-apt-cache` and these two drills; every other job is pull-request-gated, `workflow_dispatch`-gated, or `if: false`, and `scripts/ci-workflow.spec.ts` pins that set — classifying by exact condition, since a negated event test mentions the event it excludes — so a new push-reachable job cannot quietly start accumulating uncancelled runs.
+
 ### What the in-house pool is
 
 `vm-backup`: one 64-core VM, six always-on systemd-managed runner instances. Its image must preinstall Playwright Chromium's Linux system packages; CI downloads the lockfile-selected browser but never runs `apt` on this persistent shared host. Check the latest `serial / linux (self-hosted standby)` run before switching: its aggregate includes browser replay, so a green standby verifies both ordinary capacity and this browser prerequisite.
diff --git a/.agents/notes/implemented/process/2026-07-26-ci-failover-runbook.zh.md b/.agents/notes/implemented/process/2026-07-26-ci-failover-runbook.zh.md
index 88a793e144..5c8fd0e90a 100644
--- a/.agents/notes/implemented/process/2026-07-26-ci-failover-runbook.zh.md
+++ b/.agents/notes/implemented/process/2026-07-26-ci-failover-runbook.zh.md
@@ -12,6 +12,12 @@ Status: implemented
 
 三个必需的 Linux 工作作业、独立的原生 Windows 作业，以及 `all checks passed` 判定作业（若不随切换，即使全部工作作业通过，它仍会滞留在故障池的队列中）——各自通过仓库变量 `DSH_CI_FAILOVER` 解析运行器池。变量不存在（正常）时它们运行在托管企业池上；由任何具备写权限的协作者设为 `selfhosted` 时，五个作业全部切换到公司自有的自托管池：Linux 作业与判定作业切到 `vm-backup` 池，覆盖率与快照的并发降到共享虚拟机上限，并跳过托管路径的 pnpm 缓存恢复；原生 Windows 作业切到 `dsh-win-ci` 池。这个开关是写者可管理的仓库状态而非一次合并，因此在所有检查都是红色时仍然有效。自有池的就绪状态由 `serial / linux (self-hosted standby)` 与 `serial / windows (self-hosted standby)` 通道持续验证——每次 master 推送都在其上运行完整的未分片聚合流程。
 
+`ci.yml` 只豁免一个事件不做取消（`${{ github.event_name != 'push' }}`），因此一次 master 推送不会取消上一次推送留下的、仍在运行的演练。每次演练以单门禁工作进程执行完整的未分片聚合流程，耗时长于 master 合并的间隔；在无条件取消下，演练会在得出结论前被后续运行取代，该通道无法产出供响应者查看的就绪证据。
+
+这项豁免比「演练总能跑完」要窄，有两点限制。其一，GitHub 每个组只保留一个待运行条目，更新的待运行条目会顶掉更早的，繁忙时段中间的推送运行仍会以 `cancelled` 结束。其二，该表达式是针对**新触发的运行**求值的，因此自身事件不是 `push` 的运行——例如在 master 上派发的基准测试，与演练共用 `CI-<ref>` 组——求值为 `true`，会取消正在运行中的演练。这属于罕见的手动操作，且下一次 master 推送即可恢复证据，因此不值得为它再加机制。这项豁免换来的是该通道**周期性**地得出结论，而这正是它能作为证据的前提。
+
+这个决定必须放在工作流级：取消作用于被取代的整个运行，作业级 `concurrency` 组并不能豁免其所属作业。采用否定式写法而非仅指名 `pull_request`，是有实质作用的：后者会连 `workflow_dispatch` 一起停止取消，而每次运行器基准测试会在 master 上的同一并发组内同时占用 12 台大规格运行器、最长 15 分钟，届时重复派发会排在演练之前，而不是替换掉已过时的测量。成本之所以可控，是因为一次 master 推送只承载 `wine-apt-cache` 和这两条演练；其余作业都受拉取请求门控、`workflow_dispatch` 门控或 `if: false`，并且 `scripts/ci-workflow.spec.ts` 会锁定这个集合——按条件精确匹配，因为否定式事件判断会包含它所排除的事件名——使新的推送可达作业无法悄悄开始累积未取消的运行。
+
 ### 自有池是什么
 
 `vm-backup`：一台 64 核虚拟机，6 个常驻 systemd 管理的运行器实例。其镜像必须预装 Playwright Chromium 的 Linux 系统软件包；CI 会下载锁文件选定的浏览器，但绝不在这台持久化共享主机上运行 `apt`。切换前先看 `serial / linux (self-hosted standby)` 最近一次运行：其聚合流程包含浏览器回放，因此绿色热备同时验证常规容量和这项浏览器先决条件。
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 2095c139dd..08dbcc44f8 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -15,9 +15,21 @@ on:
           - larger-runner-benchmark
           - consolidated-runner-benchmark
 
+# Cancel a superseded run on every event EXCEPT push. A push run carries the two
+# self-hosted standby drills, which take longer than the interval between master
+# merges, so cancelling supersedes a drill before it reaches a verdict and the
+# lane yields no readiness evidence. Must be decided here: cancellation applies
+# to the whole superseded run, so a job-level group cannot exempt its job.
+# Negated rather than `== 'pull_request'` so workflow_dispatch keeps cancelling:
+# a re-dispatched runner benchmark holds up to 12 larger runners for 15 minutes
+# and shares this group with the drills on master, so queueing it would delay
+# them. The guarantee is narrow — evaluated against the newly triggered run, so a
+# dispatch on master still cancels a mid-flight drill, and a newer pending push
+# displaces an older one. Bounds and rationale:
+# .agents/notes/implemented/process/2026-07-26-ci-failover-runbook.md
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
+  cancel-in-progress: ${{ github.event_name != 'push' }}
 
 permissions:
   contents: read
diff --git a/scripts/ci-workflow.spec.ts b/scripts/ci-workflow.spec.ts
index fe0c7b87e5..1073b35018 100644
--- a/scripts/ci-workflow.spec.ts
+++ b/scripts/ci-workflow.spec.ts
@@ -84,6 +84,73 @@ describe('CI workflow', () => {
     expect(aggregate.needs).not.toContain('serial-windows')
   })
 
+  it('exempts push from cancellation, so one master merge does not cancel the running drill', () => {
+    const workflow = loadWorkflow('.github/workflows/ci.yml')
+    if (!isRecord(workflow.jobs) || !isRecord(workflow.concurrency)) {
+      throw new TypeError('CI workflow must define jobs and a workflow-level concurrency block')
+    }
+
+    // Cancellation applies to the whole superseded RUN, so this has to be
+    // decided at workflow level and gated on the event: a job-level group
+    // cannot exempt its job from its run being cancelled. Only push is exempt —
+    // a drill takes longer than the interval between master merges. The negated
+    // form is load-bearing: `== 'pull_request'` would also stop cancelling
+    // workflow_dispatch, and a re-dispatched runner benchmark holds up to 12
+    // larger runners for 15 minutes in this same group on master. The
+    // expression is evaluated against the NEWLY TRIGGERED run, so a dispatch on
+    // master still cancels a mid-flight drill; the runbook records that bound.
+    expect(workflow.concurrency['cancel-in-progress']).toBe("${{ github.event_name != 'push' }}")
+
+    // Neither drill may carry a job-level group: it would not exempt the job
+    // from run-scoped cancellation.
+    for (const name of ['serial-linux-selfhosted', 'serial-windows']) {
+      const job = workflow.jobs[name]
+      if (!isRecord(job)) throw new TypeError(`${name} must be defined`)
+      expect(job.concurrency).toBeUndefined()
+      // Both stay master-push-only; that is what makes the push carve-out safe.
+      expect(job.if).toBe("github.event_name == 'push' && github.ref == 'refs/heads/master'")
+    }
+
+    // What bounds the cost of exempting push: a master push may only carry the
+    // cache seeder and the two drills. Any job reachable on push would start
+    // accumulating uncancelled runs, so the set is pinned here.
+    //
+    // Classification is an exact allowlist of the conditions in use, not a
+    // substring match: `github.event_name != 'pull_request'` mentions
+    // `pull_request` yet IS push-reachable, so matching on the event name alone
+    // would silently misclassify it as gated.
+    const NOT_PUSH_REACHABLE = new Set([
+      "github.event_name == 'pull_request'",
+      "always() && github.event_name == 'pull_request'",
+      "github.event_name == 'workflow_dispatch' && inputs.suite == 'larger-runner-benchmark'",
+      "github.event_name == 'workflow_dispatch' && inputs.suite == 'consolidated-runner-benchmark'",
+    ])
+    const pushReachable = Object.entries(workflow.jobs)
+      .filter(([, job]) => {
+        if (!isRecord(job)) return false
+        if (job.if === undefined) return true // unconditional: runs on every event
+        if (job.if === false) return false // `if: false` parses as a boolean
+        if (typeof job.if !== 'string') return true // unrecognized shape: surface it
+        return !NOT_PUSH_REACHABLE.has(job.if.trim())
+      })
+      .map(([name]) => name)
+      .sort()
+    expect(pushReachable).toEqual(['serial-linux-selfhosted', 'serial-windows', 'wine-apt-cache'])
+
+    // Why workflow_dispatch must keep cancelling: each benchmark fans out to a
+    // dozen larger runners at once, in this same group on master. If it stopped
+    // cancelling, a re-dispatch would queue ahead of a drill instead of
+    // replacing the stale measurement.
+    for (const name of ['larger-runner-benchmark', 'consolidated-runner-benchmark']) {
+      const job = workflow.jobs[name]
+      if (!isRecord(job) || !isRecord(job.strategy)) {
+        throw new TypeError(`${name} must define a matrix strategy`)
+      }
+      expect(job.strategy['max-parallel']).toBe(12)
+      expect(job['timeout-minutes']).toBe(15)
+    }
+  })
+
   it('keeps supported LSP source under native Windows coverage', () => {
     const config = readFileSync(resolve(root, 'vitest.config.ts'), 'utf8')