fix: skip SA_NODEFER when CHAIN_AT_START is active

SA_NODEFER (added in #1446) is incompatible with the CHAIN_AT_START
signal handler strategy. When chaining to the runtime's signal handler
(e.g. Mono), the runtime may reset the signal to SIG_DFL and re-raise.
With SA_NODEFER the re-raised signal is delivered immediately, killing
the process before our handler can regain control.

Without SA_NODEFER, the re-raised signal is blocked during handler
execution, allowing the runtime handler to return and sentry-native
to proceed with crash capture.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: jpnurmi

src/backends/sentry_backend_inproc.c

@@ -507,7 +507,16 @@ startup_inproc_backend(
     // running. This is needed for recursive crash detection to work -
     // without it, a crash during crash handling would block the signal
     // and leave the process in an undefined state.
-    g_sigaction.sa_flags = SA_SIGINFO | SA_ONSTACK | SA_NODEFER;
+    // However, SA_NODEFER is incompatible with CHAIN_AT_START: when we
+    // chain to the runtime's signal handler (e.g. Mono), it may reset
+    // the signal to SIG_DFL and re-raise. With SA_NODEFER the re-raised
+    // signal is delivered immediately (killing the process) before our
+    // handler can regain control.
+    g_sigaction.sa_flags = SA_SIGINFO | SA_ONSTACK;
+    if (g_backend_config.handler_strategy
+        != SENTRY_HANDLER_STRATEGY_CHAIN_AT_START) {
+        g_sigaction.sa_flags |= SA_NODEFER;
+    }
     for (size_t i = 0; i < SIGNAL_COUNT; ++i) {
         sigaction(SIGNAL_DEFINITIONS[i].signum, &g_sigaction, NULL);
     }