fix(backend): inherit dry_run from execution_context in child graph validation

When AgentExecutorBlock spawns a child graph execution, it passes
execution_context (with dry_run=True) but the dry_run parameter
defaults to False. This caused validate_and_construct_node_execution_input
to reject missing credentials in the sub-agent, even though dry-run
should skip credential validation.

Fix: derive dry_run from execution_context.dry_run when an execution_context
is provided. Also propagate simulation_context to child graphs.

autogpt_platform/backend/backend/executor/utils.py

@@ -928,6 +928,14 @@ async def add_graph_execution(
             execution_context.parent_execution_id if execution_context else None
         )
 
+        # When execution_context is provided (e.g. from AgentExecutorBlock),
+        # inherit dry_run and simulation_context so child-graph validation
+        # skips credential checks and simulated blocks get the same hints.
+        if execution_context and execution_context.dry_run:
+            dry_run = True
+            if simulation_context is None and execution_context.simulation_context:
+                simulation_context = execution_context.simulation_context
+
         # Validate simulation_context *before* creating any DB records so we
         # don't leave orphaned INCOMPLETE graph_execution rows on failure.
         _SIMULATION_CONTEXT_MAX_BYTES = 16 * 1024