Hybrid model routing: local + cloud task routing for cost optimization

[kovtcharov]

Problem

GAIA supports multiple LLM providers (Lemonade local, Claude, OpenAI) but has no intelligent routing between them. Users must manually choose a backend. There is no way to route simple tasks to local/cheap models and complex reasoning to cloud frontier models.

Strategic Context

This is Track C's #1 deliverable and AMD's strongest cost differentiation story. From the strategy doc:

• A single heartbeat generates 2-3M tokens/day — costing $1-3/day on Sonnet, $30-45/day on Opus
• Routing to a local model saves 50-90%
• "This is the most concrete 'AMD saves you money' proof point"

Proposed Architecture

• Task classifier (small local model or heuristic) determines task complexity
• Simple tasks (status checks, classification, formatting) → local Lemonade model
• Complex tasks (multi-step reasoning, code generation, analysis) → cloud model
• User-configurable routing rules and cost budgets
• Telemetry: track tokens saved, cost avoided

Dependencies

• Multi-provider LLM support (already exists in `src/gaia/llm/`)
• Task complexity classification (new)

Acceptance Criteria

• Automatic routing based on task complexity
• User can configure routing preferences
• Cost savings telemetry visible to user
• Fallback to cloud when local model quality is insufficient