## Problem

The router's cooldown logic applies a single global `allowed_fails` threshold across all deployments. Deployments with different reliability characteristics (e.g., a slow experimental model vs. a stable production model) cannot have independent failure tolerance policies. Additionally, a TTL mismatch between the in-memory cache and Redis cache for cooldown state causes deployments to recover from cooldown inconsistently across proxy instances.

## Proposed fix

Add an `allowed_fails_policy` field to per-deployment config that overrides the global threshold for that deployment. Fix the DualCache TTL so in-memory and Redis cooldown entries expire at the same time, ensuring consistent recovery behavior across proxy replicas.

## Files

- `litellm/router.py` (or router cooldown logic)