feat: Two-stage evaluation pipeline for efficiency

[don-petry]

Problem

Every interaction is sent through full LLM evaluation, even non-rule interactions (casual questions, code generation, etc.). This wastes LLM calls and adds unnecessary latency/cost.

Proposed Solution

Implement a two-stage evaluation pipeline:

• Stage 1 (lightweight): Quick classification — "Does this interaction contain a persistent preference or rule?" — using a low-cost model or heuristic
• Stage 2 (full): Extract and professionalize the rule using the full evaluator

Expected 50%+ reduction in LLM calls since many interactions are routine code assistance with no rule content.