v2 service crash-loops on EPIPE when a local MCP command dies instantly

Description

If a local MCP server in config points at a command that exits immediately (in my case a bun run --cwd <dir that no longer exists> start entry), the v2 background service crashes on startup with an uncaught EPIPE: broken pipe, send. The client keeps respawning it, so the TUI is stuck showing "restarting service" forever and v2 is unusable until the MCP entry is removed by hand.

The EPIPE comes from flushing the MCP initialize frame into the dead child's stdin. It surfaces asynchronously through the stream's destroy path after the stdin sink has already completed, so nothing is listening for the error event on proc.stdin and it takes down the whole process:

EPIPE: broken pipe, send
 syscall: "send",
   errno: -32,
    code: "EPIPE"

      at end (unknown:1:1)
      at internal:fs/streams:402:23
      at _destroy (internal:streams/destroy:63:18)
      ...

Reproduces on the Bun 1.4 runtime that the v2 binary ships with; Bun 1.3.14 doesn't surface the late EPIPE.

Plugins

No response

OpenCode version

v0.0.0-beta-17595

Steps to reproduce

1. Add a local MCP server whose command dies instantly, e.g. "command": ["bun", "run", "--cwd", "/nonexistent/dir", "start"]
2. Launch the v2 TUI
3. Service exits 1 with the EPIPE above; TUI loops on "restarting service"

Screenshot and/or share link

No response

Operating System

macOS 26

Terminal

No response