`get_model_name_from_config` raises `KeyError` instead of returning empty string

backend/utils/config_utils.py:41-49:

def get_model_name_from_config(model_config: Dict[str, Any]) -> str:
    """Get model name from model id"""
    if model_config is None:
        return ""
    model_repo = model_config["model_repo"]
    model_name = model_config["model_name"]
    if not model_repo:
        return model_name
    return f"{model_repo}/{model_name}"

The function defends against None but not against partial dicts. Two callers — backend/utils/llm_utils.py:77 (OpenAIModel(model_id=get_model_name_from_config(llm_model_config) if llm_model_config else "")) and backend/utils/memory_utils.py:47, 55 — pass dicts that come straight from tenant_config_manager.get_model_config(). That manager returns default={} whenever the model id can't be parsed (config_utils.py:101), which means a sneaky path exists where you get an {} dict rather than None, and accessing model_config["model_repo"] then raises KeyError.

Concrete repro: configure a tenant with a non-numeric LLM_ID. get_model_config returns default which the caller may have passed as {} or as a partial dict like {"model_name": "x"}. The next call into get_model_name_from_config blows up.

Suggested fix

def get_model_name_from_config(model_config: Dict[str, Any]) -> str:
    if not model_config:
        return ""
    model_repo = model_config.get("model_repo") or ""
    model_name = model_config.get("model_name") or ""
    if not model_repo:
        return model_name
    return f"{model_repo}/{model_name}"

Severity: Medium. The crash propagates up and the user sees a 500.

Start in backend/utils/config_utils.py:41-49 and trace the callers in backend/utils/llm_utils.py:77 and backend/utils/memory_utils.py:47,55. Reproduce the failure with a non-numeric LLM_ID, then verify that empty or partial model configuration dictionaries return an empty or usable model name without raising KeyError.