Active Memory fallback model now configurable

Active Memory had a problem: when no model was explicitly configured, it silently fell back to a hardcoded GPT-5.4-mini reference, routing requests through OpenAI regardless of what provider a user had configured. For teams using bring-your-own-key setups or cheaper alternatives, this meant unexpected expenses and traffic flowing to an unintended provider.

The fix removes the hardcoded default entirely. Instead of letting the extension quietly substitute an unconfigured model, the system now stops model resolution at the agent primary model and skips memory recall if nothing is found. Administrators who want a fallback can now configure one explicitly using config.modelFallback, specifying any provider and model combination they prefer.

The old modelFallbackPolicy setting is deprecated. It remains in config files without breaking anything, but the runtime no longer reads it—users will see a warning logged if they have it set. This gives teams time to migrate to the explicit configuration without requiring an immediate config change.

In the active-memory extension, the getModelRef function was simplified. Rather than returning a source indicator alongside the model reference, it now returns either a valid model object or undefined, making the calling code cleaner and the behavior more predictable.