getRequestListener: uncaught `socket.destroySoon is not a function` in forceClose when the socket is mocked (e.g. fastify.inject via MCP SDK)

[MarketingMaatwerk]

What happened

getRequestListener's request-drain logic force-closes the socket after DRAIN_TIMEOUT_MS when the incoming body was not fully read. forceClose assumes the socket is a real net.Socket:

const forceClose = () => {
  cleanup();
  const socket = incoming.socket;
  if (socket && !socket.destroyed) {
    socket.destroySoon();
  }
};

When incoming is not a real http.IncomingMessage (for example a mocked request), socket.destroyed is undefined (so the guard passes) and socket.destroySoon does not exist, producing an uncaught exception on a timer that callers cannot catch:

TypeError: socket.destroySoon is not a function
 ❯ Timeout.forceClose node_modules/@hono/node-server/dist/index.mjs:390:14
 ❯ listOnTimeout node:internal/timers:585:17
 ❯ processTimers node:internal/timers:521:7

How we hit this

The MCP TypeScript SDK (@modelcontextprotocol/sdk, StreamableHTTPServerTransport) uses getRequestListener to bridge Node req/res to fetch Request/Response. We mount that transport inside a Fastify route and integration-test it with fastify.inject() (light-my-request). light-my-request's MockSocket is a plain EventEmitter with neither destroyed nor destroySoon, so every injected MCP request schedules a 500 ms timer that later crashes the (vitest) worker with the unhandled TypeError above. Because it fires on a timer, it is timing-dependent and shows up as flaky "Unhandled Errors" in otherwise green test runs.

Environment

• @hono/node-server 1.19.14 (via @modelcontextprotocol/sdk 1.29.0)
• Node v22.23.0
• light-my-request 6.6.0 (fastify inject)

Suggested fix

Make forceClose defensive about non-standard sockets:

const forceClose = () => {
  cleanup();
  const socket = incoming.socket;
  if (socket && !socket.destroyed) {
    if (typeof socket.destroySoon === 'function') {
      socket.destroySoon();
    } else {
      socket.destroy?.();
    }
  }
};

That keeps the drain behaviour for real sockets and turns the mocked-socket case into a no-op instead of an uncatchable crash. Happy to send a PR if that direction works for you.