fix(runtime): settle Pi RPC shutdown on close when the spawn itself failed

[itswl]

What

PiRpcClient.beginShutdown() waits on the child's exit event. Node emits error + close and never exit when the spawn itself fails, so a failed pi spawn made close() take ~4 seconds (two shutdownGraceMs timers, since kill() on a failed spawn returns false) — and with nothing else keeping the event loop alive, the promise never settled at all. this.exited was also only set in the exit handler, so the existing fast-path (if (this.exited) return Promise.resolve()) could not trigger.

Change

• src/runtime/pi-rpc-client.ts: treat close as a terminal event alongside exit — both in the constructor's termination handler (this.exited + fail-closed) and in beginShutdown() (finish on either event). Normal exit → close sequences and the SIGTERM→SIGKILL escalation for stubborn children are unchanged.
• src/runtime/runtime-adapters.ts: the local ProcessLike child-process contract gains the matching close overload.
• test/unit/runtime/pi-rpc-client.test.mjs: FailedSpawnProcess double (emits error + close, no exit, kill() returns false) asserting shutdown settles in under a second instead of waiting out the grace timers.

Validation

• Repro before the change (real pi binary, nonexistent cwd, live event loop): t=3ms request rejected: Pi RPC process failed: spawn <path> ENOENT → t=9ms child close code=-2 → t=4003ms close resolved. After the change, the catalog call fails in 50ms with the same error.
• bun run build ✅
• bun run typecheck ✅
• bun run licenses:check ✅ · bun run publication:check:tree ✅
• bun test test/unit/runtime/pi-rpc-client.test.mjs test/unit/runtime/pi-model-catalog.test.mjs → 21 pass / 0 fail
• bun test test/unit/runtime/ → 232 pass / 0 fail
• bun run test:unit → 896 pass / 1 skip / 1 fail. The remaining failure, three-Agent live acceptance is opt-in, hermetic by default, and fixture-verifiable, reproduces on unmodified main in this environment and is unrelated to this change; a one-off real dashboard … SIGINT timeout in an earlier run did not recur and passes standalone in ~320 ms.

Fixes #218

🤖 Generated with Claude Code

[itswl]

补充一条真实场景的耗时证据：在 0.5.7 的真实 setup 里，pi 探测因 spawn 失败（cwd 缺失）而失败时，调用方等了 4020ms 才拿到错误——正是 beginShutdown 等不到 exit、由两个 shutdownGraceMs 定时器兜底结算所致；进程内还有其它句柄（setup 交互），所以这次不是永不 settle，而是固定 4 秒延迟。

[eddiearc]

MERGE-WITH-NITS（Fable 5.1 审合）

Node 失败 spawn 只有 error+close、没有 exit；shutdown 只等 exit 会拖满 ~4s grace。把 close 也当终态是对症修复，FailedSpawnProcess 用例到位。SIGTERM→SIGKILL 路径未动。

请先于 #224 合（同改 pi-rpc-client + 测末尾）。缺 package.json bump，跟整批 release 策略一起定即可。