Feature: Add per-provider weight/priority configuration for routing decisions #114

Context

Currently, model="auto" selects models based on strategy (cheapest/balanced/fastest/quality) without any way for the operator to express provider preference. In my homelab setup, I have keys for OpenAI, Anthropic, and Groq, but I'd prefer to route to Groq first (cheapest + fastest for my workload) and only fall back to OpenAI/Anthropic when Groq is rate-limited or down.

Problem

The choose_auto_model function in app/auto_routing.py ranks models purely by the strategy's scoring logic. There's no way to say "prefer provider X" or "avoid provider Y unless necessary". The preferred_models list is model-level, not provider-level, and requires me to maintain a hardcoded list of model IDs that drifts as the catalog updates.

Proposed API

Add a provider_weights field to the routing config:

PUT /v1/routing
{
  "strategy": "balanced",
  "provider_weights": {
    "groq": 2.0,
    "openai": 1.0,
    "anthropic": 0.5
  }
}

The weight would multiply the normalized score (or add a bonus to the cost) so that a provider with weight 2.0 is preferred over one with weight 1.0, all else being equal. Weight 0 would exclude the provider entirely.

Implementation Sketch

In auto_routing.py, after computing the per-model score, apply the provider weight:

def _weighted_score(model_score: float, provider: str, weights: dict[str, float]) -> float:
    w = weights.get(provider, 1.0)
    if w <= 0:
        return float('-inf')  # exclude
    return model_score * w

This would slot into the existing sort key without major refactoring.

Why This Matters

Self-hosters often have cost/latency preferences that don't align with pure strategy optimization. Provider-level weighting gives fine-grained control without abandoning the model="auto" convenience.