fix(libxce-failover): classify and propagate netio I/O errors on failover

Add io_classify_io_error and propagate endpoint/epoll errors through
endpoint_update_epoll and cbuf_get instead of swallowing them, so client
failover surfaces the real error rather than hanging. Adds netio arm-failure,
close-propagation, and request send-propagation tests.

Pre-existing uncommitted work, committed to establish a clean baseline before
ha-commit-durability implementation.

src/lib/testsuite/test_netio_arm_failure.c

@@ -0,0 +1,95 @@
+/*
+ * test_netio_arm_failure.c — libxce-failover: a poll-arm failure must fail LOUD.
+ *
+ * Root cause (bug-646): endpoint_update_epoll() swallowed iomgr_update_poll()
+ * (EPOLL_CTL_MOD) failures — it logged "rc1=-1 < 0" but returned 0 and left the
+ * poll interest unarmed.  On the synchronous client send path (endpoint_enq),
+ * a failed arm meant the request bytes were never flushed, no reply ever
+ * arrived, and the caller blocked forever in cbuf_get() while a perfectly
+ * healthy io_worker slept in epoll_wait().  (EPOLL_CTL_MOD fails with ENOENT
+ * when the fd is no longer registered — e.g. after a failed reconnect.)
+ *
+ * Contract under test: when the poll interest cannot be armed, endpoint_enq()
+ * must (1) return a negative, RETRIABLE error so the caller propagates it
+ * (request.c senders return it; the client failover loop rotates — bug-650),
+ * and (2) must NOT synchronously invoke the endpoint's err_handler.  endpoint_enq
+ * runs on the caller's thread, which may already hold a lock the err_handler
+ * needs (the controller's primary_heartbeat_fn holds ha->lock and
+ * vote_ep_err_handler re-acquires it → self-deadlock, bug-652).  The real
+ * teardown stays on the io_worker thread, off the next epoll error event.
+ */
+
+#define _GNU_SOURCE
+#include <errno.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/socket.h>
+#include <unistd.h>
+
+#include <unity/unity.h>
+#include "netio.h"
+
+/* ── err_handler invocation tracker ──────────────────────────────── */
+
+static int g_err_count = 0;
+
+static int err_handler_stub(void *ep)
+{
+    (void)ep;
+    g_err_count++;
+    return 0;
+}
+
+/*
+ * Arming EPOLL_CTL_MOD on an fd that was never EPOLL_CTL_ADD'd fails with
+ * ENOENT — the exact post-failed-reconnect condition.  endpoint_enq() must
+ * surface it (negative rc) WITHOUT calling the err_handler (no re-entrancy).
+ */
+static void test_enq_arm_failure_is_loud_no_errhandler(void)
+{
+    iomgr_t iomgr;
+    memset(&iomgr, 0, sizeof(iomgr));
+    TEST_ASSERT_EQUAL_INT(0, iomgr_init(&iomgr));
+
+    endpoint_t *ep = endpoint_get();
+    TEST_ASSERT_NOT_NULL(ep);
+
+    /* A real socket fd that is NOT registered in iomgr's epoll: EPOLL_CTL_MOD
+     * on it fails with ENOENT, mirroring a send after a failed reconnect. */
+    ep->fd = socket(AF_INET, SOCK_STREAM, 0);
+    TEST_ASSERT_TRUE(ep->fd >= 0);
+    ep->iomgr              = &iomgr;
+    ep->ssl_handshake_done = 1; /* past handshake: calc_flags arms EPOLLIN     */
+    ep->flags              = 0; /* force flags != calc_flags → iomgr_update_poll */
+    endpoint_set_err_handler(ep, (err_handler_t)err_handler_stub);
+
+    g_err_count = 0;
+
+    /* One packet on the outq so there is real outbound work to strand. The
+     * endpoint owns it after enq; endpoint_fini drains it via outq_item_killer. */
+    io_packet_t *pkt = (io_packet_t *)calloc(1, sizeof(*pkt));
+    TEST_ASSERT_NOT_NULL(pkt);
+
+    int rc = endpoint_enq(ep, &pkt->link);
+
+    /* Loud failure (negative, retriable) so the caller propagates and rotates —
+     * but the err_handler must NOT be invoked from this synchronous send path
+     * (re-entrancy would self-deadlock the controller, bug-652). */
+    TEST_ASSERT_TRUE(rc < 0);
+    TEST_ASSERT_EQUAL_INT(0, g_err_count);
+
+    endpoint_fini(ep);
+    free(ep);
+    close(iomgr.epoll_fd);
+    printf("test_enq_arm_failure_is_loud_no_errhandler: PASS\n");
+}
+
+/* ── suite entry point ───────────────────────────────────────────── */
+
+void test_netio_arm_failure(void)
+{
+    printf("\n>> netio poll-arm failure tests\n\n");
+
+    RUN_TEST(test_enq_arm_failure_is_loud_no_errhandler);
+}