GPU free-tier arbitrage routing policy with observable fallback

[Spherrrical]

Plano already sits in the routing layer but it doesn't take advantage of that position to automatically reduce cost for developers. We should build a first-class GPU free-tier arbitrage policy: routing low-stakes or bursty agent traffic to free/low-cost providers when available, with deterministic fallback to the primary when they're unavailable or overloaded, and full trace visibility into every routing decision.

Requirements

• Configurable arbitrage policy at the model_providers level: specify a ranked list of free/low-cost providers with fallback ordering
• Deterministic fallback: when a free-tier provider is unavailable, rate-limited, or errors, Plano falls back to the primary predictably
• All routing decisions surfaced in traces: which provider was selected, why, when it fell back, and what the next selection was
• Reliability guardrails: free-tier providers should not silently degrade: failures must be explicit and logged

What "done" looks like

A developer can add a minimal config block to enable arbitrage, run a request, and see in the trace: provider selected, reason (free-tier available), fallback chain if applicable.

[arlyon]

linking to #697 as well since this is essentially just fallbacks with ranking + perhaps a pre-made config on top