Summary

Fixes an issue where configuring a custom SQLAlchemy registry stored config_table into model_config["registry"] instead of config_registry, and ensures that concrete table models subclassing a custom base model are properly mapped.

Problem

1. In SQLModelMetaclass.__new__:

   config_registry = get_config("registry")
   if config_registry is not Undefined:
       config_registry = cast(registry, config_registry)
       # If it was passed by kwargs, ensure it's also set in config
       new_cls.model_config["registry"] = config_table  # BUG: should be config_registry
       setattr(new_cls, "_sa_registry", config_registry)
       setattr(new_cls, "metadata", config_registry.metadata)
       setattr(new_cls, "__abstract__", True)

   new_cls.model_config["registry"] was assigned config_table (which is Undefined or a bool), rather than the actual config_registry object.

2. Furthermore, because model_config is inherited by subclasses in Pydantic v2, reading get_config("registry") from inherited model_config would trigger the registry block on every subclass, unconditionally setting __abstract__ = True on concrete table classes inheriting from the custom base.

Solution

1. Gate the custom registry definition to classes that explicitly declare registry via class kwargs (kwargs.get("registry")) or explicit class_dict["model_config"].
2. Correctly assign new_cls.model_config["registry"] = config_registry.
3. Add an end-to-end regression test in tests/test_main.py testing custom base registry definition, table subclassing, mapper initialization, and DB session operations.