verify-backend: report who holds the cache lock and keep build servers from inheriting it

ci/verify-backend.sh

@@ -33,7 +33,16 @@ chmod 700 "$NUGET_SCRATCH"
 export DOTNET_CLI_TELEMETRY_OPTOUT=1 DOTNET_NOLOGO=1 DOTNET_SKIP_FIRST_TIME_EXPERIENCE=1 DOTNET_GENERATE_ASPNET_CERTIFICATE=false
 
 exec 9> "$cache/.lock"
-flock -w 900 9 || { echo "could not get the backend test cache lock" >&2; exit 1; }
+# 拿不到锁时说一声是谁握着（不然步骤会静默等上 15 分钟）
+if ! flock -n 9; then
+  echo "waiting for the backend test cache lock ($cache/.lock); held by:" >&2
+  for fd in /proc/[0-9]*/fd/*; do
+    [[ "$(readlink "$fd" 2>/dev/null)" == "$cache/.lock" ]] || continue
+    pid="${fd#/proc/}"; pid="${pid%%/*}"
+    echo "  pid $pid: $(tr '\0' ' ' < "/proc/$pid/cmdline" 2>/dev/null | cut -c1-150)" >&2
+  done
+  flock -w 900 9 || { echo "could not get the backend test cache lock" >&2; exit 1; }
+fi
 
 sync_sources() {
   mkdir -p "$src"
@@ -61,7 +70,8 @@ sync_sources() {
   echo "synced sources: $changed changed, $stale removed"
 }
 
-build() { (cd "$src/backend" && dotnet build MoShiro.slnx -c Release --nologo -v:q); }
+# 9>&-：编译服务进程（VBCSCompiler、MSBuild 节点）会比步骤活得久，不能让它们继承锁的文件描述符，否则下一条流水线要一直等到它们自己退出
+build() { (cd "$src/backend" && dotnet build MoShiro.slnx -c Release --nologo -v:q 9>&-); }
 
 sync_sources
 if ! build; then
@@ -96,7 +106,7 @@ run_with_timeout() {
   # This avoids setsid(1) forking away from the PID tracked by the old shell
   # watchdog, which could leave testhosts alive (and holding the cache lock)
   # after Woodpecker canceled the step.
-  (cd "$src/backend" && python3 "$root/ci/run-with-timeout.py" "$seconds" "$@")
+  (cd "$src/backend" && python3 "$root/ci/run-with-timeout.py" "$seconds" "$@" 9>&-)
 }
 
 # Ignore exit 8 from other assemblies with no matching tests; both suites have a hard wall-clock cap.