LiteLLM proxy integration via pre-call hook

[rlippmann]

Goal

Integrate Context Compiler into LiteLLM proxy using async_pre_call_hook to enforce deterministic conversational state before model execution.

Approach

• Intercept /v1/chat/completions requests
• Extract latest user message
• Run compiler decision

Behavior:

• clarify → return string (short-circuit, no model call)
• update/state → inject compiled state as system message
• passthrough → forward request unchanged

Scope

• Minimal reference implementation
• No changes to core engine
• No grammar changes
• Optional use of cache for per-conversation state

Outcome

Demonstrate that Context Compiler can act as a drop-in reliability layer for LLM proxy infrastructure.

[rlippmann]

Closed via PR #51