McpServer.layerStdio drops in-flight responses when stdin closes #8710

RpcServer.makeProtocolStdio interrupts the server fiber as soon as stdin ends (Effect.ensuring(... Fiber.interrupt(fiber)), packages/effect/src/rpc/RpcServer.ts ~L1395 on main). Responses to requests that were already read but have not finished are never written. A client that writes its requests and then closes stdin, such as a one-shot shell pipeline or a test harness, gets partial output or none. The process also exits with 130 (interrupted) on an ordinary EOF.

Reproduction (effect 4.0.0, @effect/platform-node 4.0.0, Node 24.16 and Bun 1.4.2):

// server.ts
import { NodeRuntime, NodeServices } from "@effect/platform-node"
import { Effect, Layer, Schema } from "effect"
import { McpProtocol, McpServer, Tool, Toolkit } from "effect/ai"

const Lookup = Tool.make("lookup", {
  parameters: Schema.Struct({ id: Schema.String }),
  success: Schema.Struct({ name: Schema.String })
})
const Tools = Toolkit.make(Lookup)

McpServer.toolkit(Tools).pipe(
  Layer.provide(Tools.toLayer({
    lookup: () => Effect.sleep("100 millis").pipe(Effect.as({ name: "first" }))
  })),
  Layer.provide(McpServer.layerStdio({
    name: "repro", version: "0.0.0", protocols: [McpProtocol.v2025_06_18]
  })),
  Layer.provide(NodeServices.layer),
  Layer.launch,
  NodeRuntime.runMain
)

// frames.jsonl
{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"repro","version":"0"}}}
{"jsonrpc":"2.0","method":"notifications/initialized"}
{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"lookup","arguments":{"id":"1"}}}

node server.ts < frames.jsonl                 # only the initialize response, exit 130
(cat frames.jsonl; sleep 1) | node server.ts  # initialize and tools/call responses, exit 130

On Bun the first command prints nothing, not even the initialize response.

Expected: after EOF the server finishes the requests it has already read and writes their responses, then exits 0. A bounded grace period would be fine.

Workaround: wrap Stdio so stdin stays open after EOF until each request id read from it has a response on stdout or a notifications/cancelled, up to a timeout.