LLMInference Should Delegate to Functional API

[eric-tramel]

Description

The LLMInference class (in module.py) should delegate to or be implemented in terms of the functional API (in functional.py) rather than having a separate, parallel implementation.

Currently there may be duplication between:

• LLMInference.forward() - the module-based approach
• Functions in functional.py - the stateless, Value-aware approach

Having two separate implementations risks:

1. Inconsistent behavior between module and functional styles
2. Duplicated bug fixes and improvements
3. Confusion about which approach is canonical
4. Divergent handling of Value/ValueRef semantics

The LLMInference class should be a thin wrapper that internally uses the functional API, ensuring a single source of truth for LLM call behavior.

Acceptance Criteria

• Audit LLMInference implementation in module.py
• Identify functional API functions that handle LLM calls (or create them if missing)
• Refactor LLMInference.forward() to delegate to functional API
• Ensure Value/ValueRef handling is consistent between both interfaces
• Maintain backward compatibility with existing LLMInference usage
• Update tests to verify delegation works correctly
• Update documentation to clarify relationship between module and functional styles

Notes

• The functional API in functional.py operates on Value objects and propagates errors as values
• LLMInference is defined in module.py as a subclass of InferenceModule
• This may require adding LLM-specific functions to functional.py if they don't exist