* refactor: optimize LLM fallback strategy using a unified attempt queue (Strategy B)
* refactor: apply exponential backoff, fix format state bleeding, and update typing to Any
* fix: avoid false breaker failures in LLM fallback
* style: format llm_utils.py with ruff
* fix: resolve possibly unbound variable warning in llm_utils.py
* refactor: replace loose Any type annotations with precise Context, ConfigManager, JSONValue, and LLMResponse
* docs: update call_provider_with_retry docstring and timeout comments
---------
Co-authored-by: Rat0323 <261020116+Rat0323@users.noreply.github.com>
Co-authored-by: SXP-Simon <sxp20061207@163.com>