fix(channel): exit when parent dies instead of spinning on EPIPE

[JefBronze]

Problem

When the Claude Code session that spawned the MCP server exits, the server is orphaned with its stdio sockets closed. Every stdout/stderr write then throws EPIPE. The uncaughtException handler calls log(), which writes to the dead stderr, which throws EPIPE again, and the process spins in an infinite exception loop at 100% CPU.

Observed on macOS: three such orphans (one running 12 days, 88 CPU-hours) pinned three cores on a fanless MacBook Air, causing repeated Dark Wake Thermal Emergency sleeps and throttled charging. SIGTERM could not stop them either, because shutdown() also logs to the dead stderr.

A sample of a stuck process shows every frame in TriggerUncaughtException → handler → stderr.write → TriggerUncaughtException.

Fix

• log() becomes a no-op once stdio is known dead and swallows write errors.
• error handlers on stdin/stdout/stderr treat EPIPE, ECONNRESET, EIO, EBADF as parent death.
• stdin end/close triggers gracefulShutdown, so the offline heartbeat is still written.
• A process.ppid watchdog (same interval as heartbeat) catches the case where the parent is SIGKILLed and no EOF arrives promptly.
• uncaughtException handler exits on stdio errors instead of re-logging them.
• Hard process.exit(1) after 3 s if graceful shutdown hangs.

Testing

In an isolated CC2CC_BRIDGE_DIR:

1. Piped stdin closed after 2 s → server exited within the heartbeat interval.
2. Parent shell SIGKILLed → orphan exited within ~8 s via the ppid watchdog.

Both previously left the process alive at 100% CPU.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TxRbndHbg5wB5BQ591TZAP