🪲 BUG-#18: Fix permanent provider fallback across run() calls

[FernandoCelmer]

Description

Fixes permanent provider fallback across run() calls (issue #18). _provider_index and self.provider are reset to the primary at the start of every run() call, so each new user message retries the primary provider first. Fallback stickiness is preserved within a single run() to avoid adding retry delays to every internal tool-call turn while the primary is down.

Motivation and Context

Closes #18. Once the agent fell back to a secondary provider, it stayed on it for the entire process lifetime — even after the primary recovered. This caused unintended permanent use of fallback providers.

Types of changes

• Bug fix
• New feature
• Documentation

Checklist

• Self-review done
• Tests added
• CHANGELOG updated
• Docs updated

[FernandoCelmer]

The fix is correct and minimal. Resetting _provider_index and self.provider at the top of run() ensures each new user message gives the primary provider another chance, while fallback stickiness within a single run() (across _complete() turns) is preserved intentionally. The test covers the key invariants — primary is retried on a new run(), provider pointer is correct after recovery, and call counts are exact. No issues found.