test(websocket): skip the room retry tests on the shared listen fd (#329)

dev/PLAN.md

@@ -36,6 +36,7 @@ an item marked "reproduce" gets its failing run before any code.
 9. `test_static_decoders` back in the unit suite (issue to open).
 10. #314 — no action until it recurs.
 11. HTTP/3 `chunk_queue` primer: with the next change to `h3_stream_append_chunk`.
+12. `http_server_pause_listeners` and the accept it may not stop: reproduce.
 
 - [ ] **Drop the streaming exemption in laravel-spawn.** `TrueAsyncServer::streamContent`
   calls `setNoCompression()` on every `StreamedResponse` as the workaround for #170
@@ -1673,6 +1674,36 @@ a release yet except the first two, which are fixed.
   still has none, so its summation is proved only on a run where the six bad
   requests did land on more than one worker. The file's header comment says so.
 
+## The room retry tests on the shared listen fd (#329)
+
+- [x] **`064`–`066` skip on the shared listen fd.** They need a subscriber on each
+  of two workers, which the server does not promise (#240), and on the shared fd
+  it measurably does not happen: `065`/`066` failed there three times in two runs
+  of PR #328, and locally `065` read `no subscriber on a remote worker` on 3 of
+  10, 5 of 20 and 2 of 20 runs against 0 of 16 on SO_REUSEPORT. The instrumented
+  helper showed the failing runs at `[0,4]` after round 0 and `[0,32]` after
+  round 7; in passing ones the first worker held 1 to 3. The retry queue they
+  test (`room_hub_post_locked`) reads nothing about the listener and stays
+  covered by the first CI pass. `tests/phpt/websocket/_ws_spread_skipif.inc`
+  mirrors `core/076`'s condition. Evidence: with
+  `TRUE_ASYNC_SERVER_SHARED_LISTEN_FD=1` the three tests skip with the reason;
+  without it each passes 10 of 10. Decided with Edmond on 2026-09-29.
+
+  Left open: the cause of the skew is not traced — every shared-fd run gave the
+  majority to the second worker, which the idea of a busy worker winning does
+  not explain; the measurements and the tracing step are in #329, together with
+  the question whether #240's answer still holds. `telemetry/012` rests on the
+  same premise in the second pass and has not failed yet; the same predicate
+  applies if it does. Opening connections in a burst, so both workers drain one
+  backlog, is an untried alternative to skipping.
+
+- [ ] **`http_server_pause_listeners` may not stop accepting on the socket.** It
+  calls the listen event's `stop`, and `libuv_listen_stop` in `ext/async`'s
+  reactor only drops a reference count ("uv_listen doesn't have a stop
+  function"); the pause tests check counters, not the fd (`h1/011`: "Don't
+  check listeners_paused"). Found while reading for #329, not run: the first step
+  is a test that stays connected-and-accepted while the listeners are paused.
+
 ## A WebSocket close is lost when the peer is still sending (#303 fallout)
 
 - [x] **The 1013 close never reaches a client that overflowed the inbound

tests/phpt/websocket/064-room-retry-delivered.phpt

@@ -5,6 +5,8 @@ Rooms reliable-send: a full target parks then a retry lands it (retry_queued ->
 if (!function_exists('TrueAsync\\__test_force_topic_post_full')) {
     echo "skip requires --enable-tas-test-hooks (fault-injection hook absent)";
 }
+require __DIR__ . '/_ws_spread_skipif.inc';
+ws_spread_skipif();
 ?>
 --EXTENSIONS--
 true_async_server

tests/phpt/websocket/065-room-retry-expired.phpt

@@ -5,6 +5,8 @@ Rooms reliable-send: a target that stays full past the deadline expires and send
 if (!function_exists('TrueAsync\\__test_force_topic_post_full')) {
     echo "skip requires --enable-tas-test-hooks (fault-injection hook absent)";
 }
+require __DIR__ . '/_ws_spread_skipif.inc';
+ws_spread_skipif();
 ?>
 --EXTENSIONS--
 true_async_server

tests/phpt/websocket/066-room-retry-rejected.phpt

@@ -5,6 +5,8 @@ Rooms reliable-send: enqueue past the outbound-queue cap is refused (retry_rejec
 if (!function_exists('TrueAsync\\__test_force_topic_post_full')) {
     echo "skip requires --enable-tas-test-hooks (fault-injection hook absent)";
 }
+require __DIR__ . '/_ws_spread_skipif.inc';
+ws_spread_skipif();
 ?>
 --EXTENSIONS--
 true_async_server

tests/phpt/websocket/_ws_spread_skipif.inc

@@ -0,0 +1,26 @@
+<?php
+/* SKIPIF for a test that needs a subscriber on every worker.
+ *
+ * Which worker accepts a connection is not something the server promises
+ * (true-async/server#240). Where the kernel load-balances SO_REUSEPORT, the
+ * rounds of ws_open_spread_subscribers reach every worker in practice. On the
+ * shared listen fd — the default outside Linux and FreeBSD, and the second
+ * Linux CI run through the env — they measurably do not: on 2026-09-29, 32
+ * WebSocket connections to 2 workers split 2:30 to 11:21 there against 10:22
+ * to 22:10 on SO_REUSEPORT, and 065 ended at 0:32 on 2 to 5 runs of 20. The
+ * condition mirrors core/076's, which selects the same camp.
+ *
+ * Usage:
+ *     <?php
+ *     require __DIR__ . '/_ws_spread_skipif.inc';
+ *     ws_spread_skipif();
+ *     ?>
+ */
+
+function ws_spread_skipif(): void
+{
+    if (!((PHP_OS_FAMILY === 'Linux' || PHP_OS_FAMILY === 'BSD')
+          && getenv('TRUE_ASYNC_SERVER_SHARED_LISTEN_FD') !== '1')) {
+        die('skip shared listen fd: no subscriber on every worker is promised (#240)');
+    }
+}