Register the connection listener explicitly, so re-emitted sockets are handled

[gkurt]

Server currently passes its connection handler to net.Server's constructor:

super((socket) => this.connectionListener(socket));

In Node this is exactly equivalent to registering a 'connection' listener — the
constructor argument is added with this.on('connection', …) — so re-emitting a
socket onto the server dispatches to it. That behaviour is load-bearing for
httpolyglot's main consumer: mockttp re-injects the tunnelled socket after CONNECT
with server.emit('connection', socket), so that a tunnelled connection gets sniffed
and routed exactly like a direct one.

The equivalence doesn't hold on every runtime. Bun handles the constructor callback
internally rather than registering it as a listener: listeners('connection') is empty,
and emit('connection', socket) never reaches connectionListener. Genuine inbound
connections still work, so a polyglot server looks healthy right up until something
re-injects a socket — at which point the socket is silently never read and the peer
hangs with no error on either side. (Filed upstream at oven-sh/bun#41060.)

This switches to the explicit form:

super();
this.on('connection', (socket) => this.connectionListener(socket));

which is a no-op on Node — same registration, same ordering, nothing can attach a
listener between super() and the .on() call — and makes the re-injection contract
hold on runtimes that treat the constructor argument specially.

Tests

Adds test/reinjected-connection.spec.ts, which stands up a plain net.Server in the
role of a proxy's CONNECT handler, hands the accepted socket back to the polyglot
server via emit('connection', …), and asserts the request is routed to the HTTP
listener.

	before	after
full suite, Node 24.20.0	19 passing	20 passing
new test, Node 24.20.0	passes	passes
new test, Bun 1.4.0	fails (2s timeout)	passes

The new test passes on Node either way — it can't fail there, since the two forms are
equivalent. Its value is that it pins a contract that mockttp depends on and that is
currently only implicit.

One note if you reproduce the Bun column: bun x mocha runs mocha under Node (the
binary's shebang wins), so the test appears to pass unpatched. bun --bun x mocha is
what actually runs it under Bun.

Notes

I'm not asking you to support Bun — the ALPN gap below means httpolyglot's TLS path
can't fully work there yet regardless. This is narrower than that: the constructor
form's equivalence to .on('connection') is a Node implementation detail, and the
explicit form costs nothing to prefer.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[gkurt]

@pimterry sorry for the PR. I intended to review this before opening a PR, but Claude opened it anyway. This indeed needs to be fixed in Bun, and seems like it's already fixed on main. I am so sorry for this. Closing.