fix: temp dir race condition in MCP Dockerfile swap

Background cleanup watcher used wait() on PIDs from the parent shell,
but wait() only works for child processes of the calling subshell. The
cleanup subshell's wait() returned immediately, deleting the temp dir
before Harbor could read it. Every MCP launch hit FileNotFoundError.

Fix: collect temp dirs in _MCP_TEMP_DIRS array, clean them up in
_drain_pool() which runs in the parent shell and can properly wait
for all harbor PIDs.

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

Author: sjarmak

configs/run_selected_tasks.sh

@@ -337,13 +337,21 @@ _wait_for_slot() {
     done
 }
 
-# Wait for all pending jobs to complete
+# Temp dirs created for MCP Dockerfile swap — cleaned up in _drain_pool
+_MCP_TEMP_DIRS=()
+
+# Wait for all pending jobs to complete, then clean up temp dirs
 _drain_pool() {
     local p
     for p in "${_PIDS[@]}"; do
         wait "$p" 2>/dev/null || true
     done
     _PIDS=()
+    # Clean up MCP temp dirs now that all harbor processes are done
+    for d in "${_MCP_TEMP_DIRS[@]}"; do
+        rm -rf "$d" 2>/dev/null || true
+    done
+    _MCP_TEMP_DIRS=()
 }
 
 # Launch a task PAIR: baseline + full simultaneously for the same task.
@@ -431,15 +439,10 @@ _launch_task_pair() {
         sleep 2
     fi
 
-    # Background watcher: clean up temp dir after MCP config completes.
-    # Not counted in _PIDS (it's cleanup, not a work slot).
-    if [ -n "$_mcp_temp_dir" ] && [ "${#pair_pids[@]}" -gt 0 ]; then
-        (
-            for p in "${pair_pids[@]}"; do
-                wait "$p" 2>/dev/null || true
-            done
-            rm -rf "$_mcp_temp_dir" 2>/dev/null || true
-        ) &
+    # Track temp dir for cleanup in _drain_pool (after all harbor processes finish).
+    # Cannot use background watcher because wait() only works for child processes.
+    if [ -n "$_mcp_temp_dir" ]; then
+        _MCP_TEMP_DIRS+=("$_mcp_temp_dir")
     fi
 }