diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index e1251fa..adaf5e6 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -92,16 +92,26 @@ jobs:
         run: pnpm typecheck
       - name: Lint
         run: pnpm lint
-      # `test:windows` caps the fork pool at 2 instead of `pnpm test`'s
-      # unbounded default. Five specs in this suite really start processes
-      # (tests/agent-pty, pty-deps, pty-helpers, install-powershell,
-      # smoke): node-pty's ConPTY helper and a cold `powershell.exe` are
-      # both console-bound, and this 2-core runner made parallel spawning
-      # the amplifier behind every red `Test` step in the 2026-09-09/10
-      # window — six real 5000 ms timeouts (agent-pty `:71`/`:106`,
-      # install-powershell `:105`, the latter measuring 12.1 s) and two
-      # silent vitest worker deaths. The test bodies now carry their own
-      # explicit timeouts as well; this just stops them starving each other.
+      # `test:windows` runs the whole suite in ONE fork (`--maxWorkers=1`),
+      # not `pnpm test`'s unbounded pool. Two distinct Windows-only failures
+      # motivated it, both recorded in docs/plans/2026-09-10-dsh-0.1.5-rc.2-adaptation.md:
+      #
+      #   1. Six red runs were real 5000 ms vitest timeouts in specs that
+      #      really start a process (tests/agent-pty:71/:106,
+      #      tests/install-powershell:105 — a cold powershell.exe start
+      #      measured 12.1 s, tests/git.spec.ts:85 at 9.6 s). The suite now
+      #      declares its own budgets (global testTimeout 15 s, 30 s in the
+      #      two heaviest describes), and serial execution stops the
+      #      process-spawning specs starving each other on a 2-core runner.
+      #   2. Four runs died at teardown with `[vitest-pool]: Worker forks
+      #      emitted error / Worker exited unexpectedly` while reporting
+      #      N-1 of N files passed and zero FAIL lines — a worker recycling
+      #      race, not a test failure. It survived --maxWorkers=2, so the
+      #      lane now recycles nothing at all.
+      #
+      # No retries: they would hide a real regression instead of removing the
+      # race. The cost is wall-clock (the Windows step runs several minutes
+      # against a 30-minute job budget), not signal.
       - name: Test
         run: pnpm test:windows
       - name: Build
diff --git a/docs/plans/2026-09-10-dsh-0.1.5-rc.2-adaptation.md b/docs/plans/2026-09-10-dsh-0.1.5-rc.2-adaptation.md
index d8c4195..c3e6a06 100644
--- a/docs/plans/2026-09-10-dsh-0.1.5-rc.2-adaptation.md
+++ b/docs/plans/2026-09-10-dsh-0.1.5-rc.2-adaptation.md
@@ -44,7 +44,8 @@ DSH `dsh-v0.1.5-rc.1..dsh-v0.1.5-rc.2`（上游 4 个 commit / 300 个变更文
 
 - 证据：run `34500290681`(295615abb)、`34448911025`(bbf953b8a) 报 `[vitest-pool]: Worker forks emitted error / Caused by: Error: Worker exited unexpectedly`（`Errors 1 error`），同一份日志里 `0 FAIL`、125 个文件全过；后者静默吞掉一个文件的 10 条结果。同一 commit `8ae527170` 的另一次 run 里 `ci-windows` 是**绿的** → 该步骤可证伪地非确定。
 - 归因（**推断，非确证**）：真起进程的 spec 有 5 个（`agent-pty` / `pty-deps` / `pty-helpers` / `install-powershell` / `smoke`），node-pty 的 ConPTY 辅助进程与冷启 `powershell.exe` 都绑定控制台，在默认**无界 fork 池**下并行抢占；本地 macOS 无法复现该模式，因此不做无证据的猜测性改动（不加 `retry`，不改默认 pool）。
-- 修法：`package.json` 新增 `test:windows`（`vitest run --maxWorkers=2`），`ci-windows` 的 `Test` 步骤改跑它——**只在 Windows CI 收敛并发，`pnpm test` 的本地语义不变**。若真实 Windows 结果仍复现，按"升级而非掩盖"的原则继续收敛（同 PR 内串行化这些 spec，仍不加 retry），并在 PR 中如实标注。
+- 修法（两次迭代，第一次被真实结果证否）：先上 `test:windows`（`vitest run --maxWorkers=2`）——**真实 Windows 上仍复现**（run `34574847774`：126/127 文件全过、`0 FAIL`、`Unhandled Error: Worker exited unexpectedly`，丢结果的那个 worker 正是跑 `tests/native-surface.spec.ts` 的）。该次日志里 `native-surface.spec.ts` **零结果**，与历史两次崩溃同形（`34500290681` 125/126、`34448911025` 123/124——都是「N-1 个文件有结果 + teardown 期 worker 静默退出」）。
+- 迭代后的修法：`--maxWorkers=1`（整个套件跑在一个 fork 里，**零 worker 回收**）。理由：跨 4 次崩溃的一致形态是「worker 回收/退出竞态」，而不是某个测试失败；单 worker 从根上取消这个竞态，代价只是墙钟时间（本地 14s → 41s，Windows 步骤仍在 30 分钟 job 预算内）。**仍不加 retry**：retry 会把真实回归一起吞掉。
 
 ### C. `plugin-mount` 的 npm 安装失败（4 次：2× ETARGET + 2× OOM）
 
@@ -84,7 +85,7 @@ DSH `dsh-v0.1.5-rc.1..dsh-v0.1.5-rc.2`（上游 4 个 commit / 300 个变更文
 - **真机挂载冒烟（rc.2）**：`DSH_CMD` 指向 `npx -y --package @deepseek-ai/dsh@0.1.5-rc.2 dsh`；见"落地"节的实测计数。
 - **真实 Windows 收敛**：`ci-windows` **连续两次全绿**（同一 commit `5451f5c`）——`pull_request` 自动触发的 run [34561189320](https://github.com/omdsh-dev/DSH-better-sidebar/actions/runs/34561189320) 与随后 `workflow_dispatch` 的 run [34561457067](https://github.com/omdsh-dev/DSH-better-sidebar/actions/runs/34561457067)，两次都是 **126 files / 1318 passed / 3 skipped / 0 errors**（`Unhandled Errors` 与 `Worker exited unexpectedly` 出现次数均为 **0**；同环境的 2026-09-09/10 窗口里这两行出现过 2 次）。同一分支的 `plugin-mount` 两次都绿：`added 521 packages in 24s`（**0 条 ERESOLVE**，与历史 OOM run 的 3062 条对比）、`test:mount` **7 passed**、聚合双挂载通过。
 - **首次真实结果的证伪（已修正）**：本 PR 的第一版修复在同一 commit 上仍红两处，`.github/workflows/ci.yml` 的改动因此被真实证据改过一轮——见 C 节的「一次被真实结果否掉的中间方案」与 A 节第三个文件（`tests/git.spec.ts:85`，9607ms）。
-- **B 类的实测结论**：连续两次 Windows 全绿、0 次 worker 死亡，但这**不等于证明了 B 类的根因**（只有 2 次历史样本、无可比对照）；记为「并发收敛 + 超时预算到位后未再复现」，若将来复现按 §B 的升级路径继续收敛。
+- **B 类的实测结论（已按"升级而非掩盖"走完两轮）**：`--maxWorkers=2` 下仍崩溃（run `34574847774`），故升级为 `--maxWorkers=1`（单 fork、零回收）。这**不等于证明了根因**（4 次崩溃的一致形态是「N-1 文件有结果 + teardown 期静默退出」，只能说明是 worker 回收竞态而非测试失败）；收敛判据是单 worker 下连续两次 `ci-windows` 全绿。
 - **本地 3080 真机**（DSH 0.1.5-rc.2 + 插件 0.19.1）：部署与路由探针已完成（见"落地"节），GUI 人工验证由用户执行。
 
 ## 不做
diff --git a/package.json b/package.json
index 8d3c7d5..a2c4de2 100644
--- a/package.json
+++ b/package.json
@@ -80,7 +80,7 @@
     "watch": "tsdown --watch",
     "prepare": "tsdown",
     "test": "vitest run",
-    "test:windows": "vitest run --maxWorkers=2",
+    "test:windows": "vitest run --maxWorkers=1",
     "test:mount": "bash scripts/e2e-mount.sh",
     "test:mount:aggregate": "bash scripts/e2e-aggregate-mount.sh",
     "clean": "node -e \"require('node:fs').rmSync('lib',{recursive:true,force:true})\"",