Generator training mode: train pool, eval overlap check, category mix, error-targeted tasks

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

harness/generator.py

@@ -393,7 +393,8 @@ def failure_summary(rep):
 
 
 def generate(task_id, pool, object_type, category, difficulty, model, base_url, run_base, topic=None,
-             max_repairs=3):
+             max_repairs=3, extra_check=None):
+    """extra_check(bundle) -> list of problems; it runs with the static checks (training mode: overlap)."""
     check_budget()
     pool_root = os.path.join(ROOT, "tasks_gen", pool)
     task_dir = os.path.join(pool_root, task_id)
@@ -411,6 +412,8 @@ def generate(task_id, pool, object_type, category, difficulty, model, base_url,
             b = parse_bundle(text)
             b.get("task", {}).setdefault("category", category)
             found = {"structure": check_bundle(b), "static": lint_files(b), "abaplint": abaplint_files(b)}
+            if extra_check and not found["structure"]:
+                found["extra"] = extra_check(b)
         except Exception as e:  # noqa: BLE001
             b, found = None, {"json": [f"invalid JSON: {e}"]}
         errs = [e for v in found.values() for e in v]