feat(scheduler): Cost-Aware & Latency-Aware Dynamic Multi-Provider Routing

[sdageltc]

Summary & Motivation

Different coding tasks require different levels of model capability. Running high-end frontier models (Claude 3.5 Sonnet / GPT-4o) on simple boilerplate tasks wastes budget, while running lightweight models on complex architectural tasks leads to repeated failed attempts.

We want to expand orchestrator/provider_scheduler.py into a Cost- & Risk-Tiered Dynamic Provider Router.

---

Proposed Routing Logic

1. Risk-Tiered Model Selection:
   • Tier 1 (Trivial / Boilerplate / Format): Gemini 2.5 Flash Lite / Qwen 2.5 Coder (Fast, sub-cent cost).
   • Tier 2 (Standard Task Implementation / Unit Tests): Gemini 2.5 Flash / Claude 3.5 Haiku / GPT-4o-mini.
   • Tier 3 (Crucible / Architecture / High-Risk Refactor / 3-Strike Repair): Claude 3.5 Sonnet / GPT-4o / DeepSeek R1.
2. Dynamic Fallback Escalation:
   • If a contract fails Attempt 1 with a Tier 2 model, Attempt 2 automatically escalates to a Tier 3 frontier model with the failed attempt evidence.
3. Cost Ceiling Enforcement:
   • Respects max cost limits configured in BudgetGuard.

---

Acceptance Criteria

• Routes contract prompts based on declared risk_tier or failure history.
• Seamless fallback on provider rate limits (HTTP 429) or timeouts.
• Full simulation unit tests in tests/test_provider_scheduler.py.