tools: add a test wrapper that replays crashing tests under gdb

meson test --wrapper hook to print a gdb backtrace inline in the test
log when a test exits with an actual crash signal (SIGSEGV, SIGABRT,
SIGBUS, SIGFPE, SIGILL). Wired into the default add_test_setup() so it
runs automatically on every `meson test`.

Environmental terminations (SIGTERM/SIGKILL/SIGPIPE/SIGALRM) are passed
through without replay, and the original signal is re-raised so the
parent's wait() observes WIFSIGNALED rather than a plain exit code.

tools/test-crash-trace.sh

@@ -0,0 +1,56 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: LGPL-2.1-or-later
+#
+# meson test wrapper that replays crashing tests under gdb and prints a
+# backtrace, so SIGSEGV/SIGABRT/etc. failures in CI show a stack trace inline
+# in the test log instead of just an exit code.
+#
+# Usage: meson test --wrapper=$PWD/tools/test-crash-trace.sh
+
+set -euo pipefail
+
+if [[ $# -eq 0 ]]; then
+    echo "usage: $0 <command> [args...]" >&2
+    exit 2
+fi
+
+rc=0
+"$@" || rc=$?
+
+# Replay only on actual crash signals (128 + signal): SIGILL=132, SIGABRT=134,
+# SIGBUS=135, SIGFPE=136, SIGSEGV=139. SIGTERM/SIGKILL/SIGPIPE/SIGALRM mean the
+# test was killed by the environment (timeout, etc.), not that it crashed —
+# replaying those just makes gdb hit the same kill.
+case "$rc" in
+    132|134|135|136|139)
+        if command -v gdb >/dev/null 2>&1; then
+            echo "===== exit $rc — replaying under gdb =====" >&2
+            style_args=()
+            if [[ -t 2 ]]; then
+                style_args=(--ex 'set style enabled on')
+            fi
+            gdb --batch \
+                --ex 'set pagination off' \
+                ${style_args[@]+"${style_args[@]}"} \
+                --ex 'run' \
+                --ex 'thread apply all bt full' \
+                --args "$@" >&2 || true
+        else
+            echo "===== exit $rc — install gdb for a backtrace =====" >&2
+        fi
+        ;;
+esac
+
+# If the child died by signal, re-raise it so the parent's wait() observes
+# WIFSIGNALED instead of a plain exit code. Best-effort: SIGKILL can't be
+# delivered to ourselves and would have killed us already.
+if [[ $rc -gt 128 ]]; then
+    sig=$((rc - 128))
+    trap - "$sig" 2>/dev/null || true
+    # Suppress the wrapper bash's own core dump from the re-raised crash signal;
+    # the original test binary already had its chance to dump core on line 18.
+    ulimit -c 0
+    kill -s "$sig" $$ 2>/dev/null || true
+fi
+
+exit "$rc"