hermes - 💡(How to fix) Fix Feature Request: Configurable model provider retry interval / backoff

Feature Description

Add a configurable retry/backoff interval setting for model provider API calls. Currently when a provider returns a transient error or times out, the retry behavior has no user-configurable delay between attempts.

Motivation

Some providers (especially smaller or high-traffic inference endpoints) can be flaky with intermittent timeouts. Having a configurable delay between retries would help in two ways:

1. Reduce provider rate-limit hammering — when the endpoint is under load, immediate retries make things worse. A 5-10s backoff gives it breathing room.
2. Better UX on unstable connections — instead of rapid-fail -> fallback chain, the user could tune a small delay that lets the primary provider recover naturally before falling over to an auxiliary model.

Proposed Solution

A new config key under model section, something like:

model:
  retry_interval: 3     # seconds between retry attempts (default: 0)
  max_retries: 3         # max retries before fallback kicks in
  backoff_multiplier: 2  # exponential: 3s -> 6s -> 12s (default: 1, no backoff)

This would slot in naturally with the existing fallback_providers / fallback_models system — exhaust retries first, then fall through to the auxiliary models.

Alternatives Considered

• Keep status quo — rapid retries then immediate fallback. Works but can be wasteful when the primary provider would recover with a small pause.
• Provider-side fix — not every inference endpoint offers configurable retry behavior at the API level.

Happy to help test or provide more detail if this direction makes sense!