fix: improve AI corrective prompt clarity and add global next_node_id validation

- Rewrote CORRECTIVE_PROMPT_TEMPLATE to clearly distinguish option→child
  vs action→sibling next_node_id semantics with concrete examples
- Added global check in ai_tree_validator that action next_node_ids
  actually reference existing nodes in the tree (was silently unchecked)
- Added max_tokens truncation warning to branch_detail logger
- Added test for action next_node_id referencing nonexistent node

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

backend/tests/test_ai_tree_validator.py

@@ -124,6 +124,13 @@ class TestReferenceIntegrity:
         errors = validate_generated_tree(tree)
         assert any("non-existent child" in e for e in errors)
 
+    def test_action_next_node_id_references_nonexistent_node(self):
+        """Action next_node_id pointing to a node that doesn't exist anywhere in the tree."""
+        tree = _make_valid_tree()
+        tree["children"][1]["next_node_id"] = "ghost-node"
+        errors = validate_generated_tree(tree)
+        assert any("ghost-node" in e for e in errors)
+
     def test_duplicate_option_ids(self):
         tree = _make_valid_tree()
         tree["options"][0]["id"] = "same"