feat(listener): bind a TCP listener on port 0 and report what it got

[EdmondDantes]

A caller that names a free port before binding it leaves a window in which the port belongs to nobody. The phpt suite picks ports that way, and under -j4 it loses the race: core/015 starts a second server after the first has run its whole lifecycle, and start() answers Failed to acquire TCP listener for 127.0.0.1:58864 (bind). Over thirty tests take ports from that helper, so narrowing the window only makes the failure rarer.

Port 0 removes it. addListener(), addHttp1Listener() and addHttp2Listener() accept 0, the kernel assigns at bind time, and HttpServer::getBoundListeners() reports what the server holds — one entry per configured listener, in configuration order, empty while the server is not running. A TCP entry carries host, port and tls; a UNIX entry carries path and no port key at all, so an absent answer cannot be mistaken for a number.

The answer lives on the server rather than in the config: the point of port 0 is that one config can be handed to a second server, and a config that recorded the first server's port would send the second one to bind it — the very bug this fixes. A restarted server rebinds and reports the new port for the same reason.

Across threads. Such a listener is bound once into the shared set and every thread adopts a duplicate, SO_REUSEPORT or not — a per-thread bind would give each thread a different port. The set keys its entries on the listener's position in the configuration instead of on host and port, since two listeners asking for 0 are indistinguishable until the kernel answers, and it records the bound port beside the requested one: a pool parent builds no listen event of its own, so the set holds its only answer. Two listeners naming the same host:port still share one socket, as before.

HTTP/3 keeps requiring an explicit port. It binds through ZEND_ASYNC_UDP_BIND, which reports no local address, so 0 is refused there rather than answered with the number that was asked for.

An old expectation changed. core/032-config-validation asserted that port 0 throws on the three TCP adders — that is the behaviour this changes. It now expects acceptance for TCP and a throw for H3, and a -1 case keeps the lower bound covered.

Evidence: the extension builds clean on Windows (MSVC, /W4, no new warnings); core/078-listener-port-zero covers a single server and a two-worker pool, each serving a request on the assigned port, and passes 5 of 5 runs; the Windows suite reads 380 tests, 169 passed, 0 warned, 0 failed, 211 skipped.

Open, and named rather than hidden: the arginfo header was hand-edited because gen_stub.php needs the tokenizer extension, which this build lacks — a reviewer regenerating it should get a byte-identical file.

[github-actions]

Coverage

Total lines: 82.98% → 83.13% (+0.15 pp)

File	Baseline	Current	Δ	Touched
src/core/http_connection_tls.c	73.83%	73.55%	-0.28 pp
src/http3/http3_callbacks.c	83.44%	84.18%	+0.74 pp
src/http3/http3_listener.c	77.36%	76.76%	-0.60 pp
src/http3/http3_packet.c	90.43%	85.22%	-5.22 pp
src/http_server_class.c	74.09%	75.97%	+1.88 pp	●
src/http_server_config.c	92.48%	92.58%	+0.10 pp	●
src/websocket/ws_session.c	90.41%	89.67%	-0.74 pp