Make TimerCallbackTest match libcurl 8.5 timer updates

.okf/api/multi.md

@@ -5,7 +5,7 @@ description: Static 1:1 bindings for libcurl multi calls ext-curl leaves out
 resource: /pcurl/multi.zep
 tags: [pcurl, api, libcurl]
 status: draft
-generated: { by: claude-code/claude-opus-5.5, at: "2026-10-11T12:00:00Z" }
+generated: { by: cursor-agent/grok-4.6, at: "2026-10-14T13:30:00Z" }
 sources:
   - id: zep
     resource: /pcurl/multi.zep
@@ -41,7 +41,7 @@ Every method is named after the C call it binds and takes ext-curl's `CurlMultiH
 
 * Long options (`< CURLOPTTYPE_OBJECTPOINT`) go straight to `curl_multi_setopt`.
 * `CURLMOPT_SOCKETFUNCTION` (20001): `callable(?CurlHandle $easy, int $s, int $what, mixed $clientp, mixed $socketp): int`. `$easy` is the ext-curl object found in the multi's handle list, or `null`.
-* `CURLMOPT_TIMERFUNCTION` (20004): `callable(CurlMultiHandle $multi, int $timeoutMs, mixed $clientp): int`.
+* `CURLMOPT_TIMERFUNCTION` (20004): `callable(CurlMultiHandle $multi, int $timeoutMs, mixed $clientp): int`. libcurl invokes it when the timeout it wants changes, not on every `curl_multi_socket_action(CURL_SOCKET_TIMEOUT)`. On 8.5 that TIMEOUT action clears `timer_lastcall` then skips the callback if the new timeout is `-1`; 8.11+ keeps `last_timeout_ms` and still reports `-1`. The suite's timer tests set `CURLOPT_TIMEOUT_MS` so both trees have an expire to report.
 * `CURLMOPT_SOCKETDATA` (10002) / `CURLMOPT_TIMERDATA` (10005): any value, passed back as `$clientp`. libcurl's own data pointer is always pcurl's state, never the PHP value.
 * `null` uninstalls a function; a non-callable returns `CURLM_BAD_FUNCTION_ARGUMENT` (10) and leaves the old one in place.
 * Any other object/function option → `CURLM_UNKNOWN_OPTION` (6). `CURLMOPT_PUSHFUNCTION` stays with ext-curl's `curl_multi_setopt`.

.okf/log.md

@@ -1,5 +1,9 @@
 # Directory Update Log
 
+## 2026-10-14
+
+* **TimerCallbackTest vs libcurl 8.5**: Ubuntu 24.04 CI (libcurl 8.5.0) failed 4 tests that assumed `curlMultiSocketAction(CURL_SOCKET_TIMEOUT)` always invokes `CURLMOPT_TIMERFUNCTION`. It doesn't: 8.5's `multi_socket` memsets `timer_lastcall` to `{0,0}` on `CURL_SOCKET_TIMEOUT`, then `Curl_update_timer` skips the callback when the new timeout is `-1` (already "none"). A localhost transfer with no `CURLOPT_TIMEOUT` looks exactly like that after `EXPIRE_RUN_NOW` is consumed — add_handle still fires `0`, `curl_multi_exec` still fires, completion never sends `-1`. 8.11 (`a58b50f`) split that bookkeeping into `last_expire_ts` (cleared) and `last_timeout_ms` (kept), so the same TIMEOUT call reports `-1`; Homebrew 8.22 and the Pi's 8.14.1 follow that path and were green. The tests now put `CURLOPT_TIMEOUT_MS` on the easy so 8.5 has an expire left to report. C unchanged: the trampoline was firing whenever libcurl asked.
+
 ## 2026-10-12
 
 * **Version**: 0.10.5 in `config.json`, `composer.json`, `PHP_PCURL_VERSION`. No API change; `ide/0.10.5/` generated with `zephir stubs`, identical to `ide/0.10.1/`. Steps past the 0.10.3 and 0.10.4 numbers to line up with the sibling extensions.

tests/TimerCallbackTest.php

@@ -27,7 +27,7 @@
 
 it('is asked to cancel with -1 once the transfer completes', function (): void {
     $timeouts = [];
-    $easy = easyFor('/cancel');
+    $easy = easyWithTimeout('/cancel');
     drive($this->multi, [$easy], null, function (CurlMultiHandle $m, int $timeoutMs) use (&$timeouts): void {
         $timeouts[] = $timeoutMs;
     });
@@ -38,14 +38,30 @@
 });
 
 /*
- * libcurl calls the timer callback only when the timeout it wants changes. Each test below installs a
- * no-op callback before adding the handle, so libcurl has a recorded timeout (0) that the next call
- * moves, then swaps in the callback under test.
+ * libcurl calls TIMERFUNCTION only when the timeout it wants changes.
+ *
+ * curl_multi_socket_action(CURL_SOCKET_TIMEOUT) on libcurl 8.5 (Ubuntu 24.04) memsets
+ * timer_lastcall to {0,0} first, then skips the callback if the new timeout is -1 — which
+ * is what a localhost transfer looks like once RUN_NOW has been consumed and nothing else
+ * is in the expire tree. 8.11+ keeps last_timeout_ms across that memset, so the same call
+ * reports -1 even without an easy timeout.
+ *
+ * CURLOPT_TIMEOUT_MS leaves an expire in the tree: TIMEOUT then reports the remaining ms,
+ * and completion reports -1. Each test below installs a no-op callback before adding that
+ * easy, so libcurl records 0, then swaps in the callback under test for the next change.
  */
+function easyWithTimeout(string $path): CurlHandle
+{
+    $easy = easyFor($path);
+    curl_setopt($easy, CURLOPT_TIMEOUT_MS, 5_000);
+
+    return $easy;
+}
+
 function multiWithPendingTransfer(CurlMultiHandle $multi, string $path): void
 {
     Multi::curlMultiSetopt($multi, CURLMOPT_TIMERFUNCTION, fn (): int => 0);
-    curl_multi_add_handle($multi, easyFor($path));
+    curl_multi_add_handle($multi, easyWithTimeout($path));
 }
 
 it('treats a null return as 0, so libcurl carries on', function (): void {