fix: unify child-process stop semantics with await-exit and SIGKILL escalation

[LeXwDeX]

Why

深度审计确认（孤儿进程向量，无 zombie——Bun/Node 自动收割，但孤儿成立）：

1. packages/opencode/src/util/process.ts:149-155：Process.stop POSIX 路径 proc.kill()（SIGTERM）后立即返回——不 await proc.exited、无 SIGKILL 升级。调用点 lsp/client.ts:640、lsp/lsp.ts:198-246 的实例 finalizer 因此可能「正常完成但进程还活着」；忽略 SIGTERM 的 language server 永久孤儿（Windows 路径有 taskkill /T /F 对照，POSIX 缺失等价语义）
2. packages/opencode/src/mcp/index.ts:533-558：MCP state finalizer 对 descendants(pid)（:416-438）只发一轮 SIGTERM，无升级、不等待；closeClient/storeClient（:564-589）与 create 失败回滚（:400-403）只调 client.close()，完全跳过孙进程清理；pgrep -P 快照遍历期间新派生的后代逃逸；Windows 上 descendants 恒为 []
3. 放大器：packages/opencode/src/index.ts:141 CLI finally 无条件 process.exit() 放弃异步 finalizer——在注释声明的取舍（docker MCP 不响应 SIGTERM）之上，前置的异步 kill+await 若未完成，子进程直接孤儿化

Scope

• packages/opencode/src/util/process.ts（注意文件头注释：packages/sdk/js/src/process.ts 有同步副本，必须一起改）
• packages/opencode/src/mcp/index.ts
• packages/opencode/src/index.ts 退出路径（给异步清理一个有界窗口，而非无限等待）

Approach

• Process.stop：SIGTERM → 有界等待 exit（如 3-5s）→ SIGKILL 升级 → await exit（与 Windows taskkill /T /F 语义对齐）；SDK 副本同步
• MCP：descendants 清理提取为可复用例程，kill 走升级语义并等待；closeClient/storeClient/create 回滚路径同样覆盖孙进程
• index.ts：finally 中对实例清理给有界窗口（超时后仍 process.exit()，保持既有取舍但缩小竞态面）

Acceptance

• 单测（注入假 ChildProcess / 测试专用 server）：SIGTERM-忽略的子进程在 Process.stop 升级窗口内被 SIGKILL 终止且 exit 被 await
• MCP 动态 disconnect/reconnect 路径孙进程被清理的回归测试
• 既有 process/mcp/lsp 测试全绿；SDK 副本 diff 一致

[LeXwDeX]

SpecGit delivery branch: fix/500-hook-command-grandchildren (draft pull request #505).

[LeXwDeX]

Implementation record for the index.ts clause (Approach point 3):

A bounded drain window was implemented (f40bf17: unref'd 5s force-exit timer replacing the unconditional process.exit()) and reverted (479469a) on CI evidence: three CLI subprocess lifecycle tests failed on linux — acp lifecycle > stdin EOF exits cleanly (7.1s), run > --format json emits a pure error record (8.0s), run > attach mode (180s timeout). The ~5s signature shows real session paths keep ref'd handles alive (server/sqlite/watcher), so the event loop never drains and every CLI exit degrades to the 5s fallback; the previous unconditional exit was masking that residue.

Conclusion: the bounded window's precondition (a drainable loop) does not hold today. Making it safe requires eliminating the residual handles first — that is leak-audit scope beyond this issue, proposed as a separate work item. The unconditional process.exit() stays as the hanging-subprocess backstop (its original documented purpose). All other #503 clauses (Process.stop SIGKILL escalation, MCP shutdownClient, SDK copy) shipped as specified in b78cdb0.