The Room class moves to src/room/php_room.c

[EdmondDantes]

http_server_class.c carried the whole PHP object layer over rooms: the Room
class with its object handlers and transfer_obj, the delivery/knob helpers, and
the six HttpServer methods that address a room by name — 850 lines of a
subsystem the server owns only a hub of, in the largest file in the tree. They
move to src/room/php_room.c, beside the core they drive.

The move is verbatim. Four substitutions were applied to the moved text, because
http_server_object's layout is private to http_server_class.c:

Z_HTTP_SERVER_P(ZEND_THIS)  ->  http_server_object_from_zend(Z_OBJ_P(ZEND_THIS))
server->running             ->  http_server_is_running(server)
server->topic_hub           ->  http_server_get_topic_hub(server)
server->topic_hub = room_hub_create(…)  ->  http_server_topic_hub_ensure(server)

The last two are new pinpoint accessors, declared in php_http_server.h — the
pattern src/http3/ already uses against the same opaque struct.
http_server_topic_hub_ensure() is now the only place a hub is created, and
the worker-clone rule that used to be spelled at one of the three creation sites
and omitted at another is stated once, inside it. A clone inherits its parent's
hub through the transfer shell; one minted on a clone would be attached by
nobody, and every publish through it would report served=0 with nothing to show
for it.

What review caught, and it was not visible from any green build

php_room.c first went into config.m4 inside the PHP_WEBSOCKET=yes block,
where its neighbours live. But the generated HttpServer.php_arginfo.h registers
all six room methods unconditionally, so a --disable-websocket build linked
cleanly — a shared object may carry undefined symbols — and then failed at
dlopen:

undefined symbol: zim_TrueAsync_HttpServer_enableRooms

The file is in the unconditional source list now. Measured after the fix, in a
separate build tree: --disable-websocket compiles with no warnings, nm -u
shows no undefined symbol of ours, the extension loads, and
class_exists('TrueAsync\Room') is correctly false while
TrueAsync\HttpServer is true.

No CI job builds --disable-websocket. That is S6.4's business, and this is the
second finding in two steps that only a build of that configuration could reach.

Also taken from the review: the ensure() comment claimed an exclusivity it did
not have; php_room_minit()'s stated MINIT dependency was on a class it does not
touch, while the real one — room_delivery_exception_ce, registered earlier by
ws_php_classes_register() — was unwritten; and three pieces of dead scaffolding
went (an empty #ifdef/#endif pair, a NULL check the accessor already makes, an
unused include).

Left deliberately

RoomDeliveryException's stub declares it extends WebSocketException, so
php_room.c still includes websocket/php_websocket.h for its class entry, and
every send()/trySend() failure path goes through it. Giving that exception a
base of its own is a PHP-visible contract change, not a file move; the file header
names it as the blocker where the next step will hit it.

Measured

• phpize && ./configure && make -j12: no warnings of ours; nm -u clean.
• tests/phpt: 381 passed, 0 failed, 1 warned (an XFAIL section on a passing
  test, pre-existing).
• tests/valgrind-rooms.sh: 9 tests, 0 leaked.
• Regenerating the stubs produces no content change (only the gen_stub.php
  version hash differs locally, so nothing is committed there).

[github-actions]

Coverage

Total lines: 82.32% → 82.19% (-0.12 pp)

File	Baseline	Current	Δ	Touched
src/core/http_connection.c	75.74%	75.55%	-0.19 pp
src/http2/http2_session.c	88.68%	88.55%	-0.13 pp
src/http3/http3_listener.c	77.33%	76.88%	-0.45 pp
src/http3/http3_packet.c	90.43%	85.22%	-5.22 pp
src/http_request.c	83.94%	80.83%	-3.11 pp
src/http_server_class.c	75.87%	75.03%	-0.84 pp	●
src/http_server_config.c	92.41%	92.47%	+0.06 pp
src/room/php_room.c	0.00%	82.37%	+82.37 pp	●
src/websocket/ws_session.c	90.59%	90.04%	-0.55 pp