experimental.testProxy hangs raw TCP sockets (regression in 16.3.0, works in 16.2.12) #96768

Link to the code that reproduces this issue

https://github.com/antoinerousseau/next-testproxy-repro

To Reproduce

git clone <gist above into a folder> repro && cd repro
pnpm install
node echo-server.js &                              # plain TCP echo server on :3901

ENABLE_TEST_PROXY=true pnpm build
ENABLE_TEST_PROXY=true ECHO_PORT=3901 pnpm start    # next start -p 3900

curl http://localhost:3900/api/probe                # hangs forever

The route (app/api/probe/route.js) does nothing but open a raw
net.Socket (no fetch, no HTTP) to a plain TCP echo server, write a byte,
and await the echo:

import net from "node:net";

function probeRawSocket(port) {
  return new Promise((resolve, reject) => {
    const socket = net.createConnection({ host: "127.0.0.1", port }, () => {
      socket.write("ping");
    });
    socket.on("data", (data) => {
      socket.end();
      resolve(data.toString());
    });
    socket.on("error", reject);
  });
}

export async function GET() {
  const start = Date.now();
  const reply = await probeRawSocket(Number(process.env.ECHO_PORT));
  return Response.json({ reply, ms: Date.now() - start });
}

next.config.js only sets:

module.exports = {
  experimental: {
    testProxy: process.env.ENABLE_TEST_PROXY === "true",
  },
};

Current vs. Expected behavior

With experimental.testProxy: true on Next.js 16.3.0, the request to
/api/probe never resolves — no response, no error, no timeout. Confirmed
via lsof on the next-server process that the outbound TCP connection to
the echo server is never even opened at the socket level.

With the exact same code and env, only changing the version:

Expected: testProxy should only intercept fetch()/HTTP traffic (per its
documented purpose for next/experimental/testmode), and should not affect
unrelated raw net.Socket connections — matching the 16.2.12 behavior.

Real-world impact

We hit this via next/experimental/testmode/playwright/msw, which flips this
same flag on to let Playwright/MSW intercept fetch() calls in e2e tests.
One of our API routes also runs a normal pg (node-postgres) query in the
same request. After upgrading 16.2.12 → 16.3.0, every Playwright test whose
setup hits that route now hangs until Playwright's default 30s hook timeout,
and with retries + a single CI worker this silently burns ~90 minutes before
the job reports failure — with no actual error surfaced anywhere. We
confirmed with pg_stat_activity/lsof that Postgres never receives a
connection attempt while stuck.

This minimal repro shows it's not pg-specific — any raw socket traffic
appears to be swallowed by testProxy in 16.3.0 rather than passed through.

Provide environment information

Operating System:
  Platform: darwin
  Arch: arm64
  Version: Darwin Kernel Version 25.5.0
Node: v24.18.0
Next.js: 16.3.0 (regression; 16.2.12 unaffected)

Also reproduced in CI on ubuntu-24.04 (GitHub Actions), same result.

Which area(s) are affected?

Not sure / Other (experimental.testProxy / next/experimental/testmode)

Additional context

This was discovered while upgrading an app from 16.2.12 to 16.3.0. Reverting
only the next/@next/env version (keeping experimental.testProxy: true)
resolves it, confirming it's a version regression and not a config issue on
our side.

(Previously filed as #96766, auto-closed for linking a gist instead of a repo — refiling with a proper public repo link. Also likely related to #96521, which pins the root cause more precisely: the @mswjs/interceptors bump to socket-level interception in #96059.)