Close MCP transport gracefully on shutdown

[20syldev]

Problem

When the MCP host (e.g. Claude Code) sends SIGTERM to terminate the server on exit, users see:

1 MCP server failed

even though everything worked correctly. The exit code is 0, but the host interprets the shutdown as a failure.

Root cause

Two issues were combining to cause this:

1. MCP transport was never closed before process.exit()

index.ts's shutdown() called process.exit(0) immediately after spawning the telemetry flush process. The MCP Server object was never told to close, so the stdio transport received an abrupt EOF instead of a clean close notification. The host saw the broken pipe and reported a failure.

2. Duplicate signal handlers racing each other

browser-eval-manager.ts registers its own SIGINT/SIGTERM handlers that both stop the Playwright connection and call process.exit(0). These fired concurrently with the handlers in index.ts, causing process.exit(0) to be called before index.ts had a chance to flush telemetry or close the MCP transport.

Fix

src/index.ts — make shutdown() async and call await server.close() before process.exit():

- const shutdown = () => {
+ const shutdown = async () => {
    log('Server terminated')
+   try {
+     await server.close()
+   } catch {
+     // Ignore close errors during shutdown
+   }
    // ... telemetry flush ...
    process.exit(0)
  }

src/_internal/browser-eval-manager.ts — remove process.exit(0) from the signal handlers (owned exclusively by index.ts):

  process.on("SIGINT", async () => {
    await stopBrowserEvalMCP()
-   process.exit(0)
  })
  process.on("SIGTERM", async () => {
    await stopBrowserEvalMCP()
-   process.exit(0)
  })

Fixes #126, relates to #113