Bottom routing

We are not a “best model” picker. We are a cheapest viable model picker.

1. If you set a concrete model id → pin (no fishing).
2. Else start in Tier‑0 (DeepSeek / Flash‑Lite / Haiku-class).
3. Escalate to Tier‑1 only on failure or a hard hint (huge prompt, effort=high).
4. Mid / Frontier only after lower tiers fail — or you pin them.
5. Prefer retry-next-cheapest over jumping to Opus.

Target: ≥85% of unpinned traffic stays on Tier‑0 / Tier‑1 (bottom_stay_rate).

Agents use the same policy per step so loops do not silently burn frontier tokens.