feat: top-level routing_preferences with selection_policy and metrics fetch (v0.4.0)

[adilhafeez]

Adds a new routing configuration format gated to v0.4.0+.

What's changing

New top-level routing_preferences — group multiple candidate models under a named route with a cost/latency-based selection policy:

version: v0.4.0

model_providers:
  - model: openai/gpt-4o
    access_key: $OPENAI_API_KEY
  - model: openai/gpt-4o-mini
    access_key: $OPENAI_API_KEY
    default: true

routing_preferences:
  - name: code understanding
    description: understand and explain existing code snippets
    models:
      - openai/gpt-4o
      - openai/gpt-4o-mini
    selection_policy:
      prefer: cheapest   # cheapest | fastest | random

model_metrics_sources:
  cost:
    url: https://example.com/model-costs   # { "openai/gpt-4o": 0.005, ... }
    refresh_interval: 300
  latency:
    url: https://example.com/model-latency
    refresh_interval: 60

Inline override — request body can include routing_preferences to override config for a single request.

Rules

• v0.4.0 forbids routing_preferences inside model_providers (startup error if present)
• Models referenced in routing_preferences must be declared in model_providers
• model_metrics_sources provides live cost/latency data via HTTP GET, refreshed in the background