v1.16.1

Fixed

• Unified MITM-aware retry policy across all provider families: get_chat_response on anthropic, gemini and mistral now retries connection failures with exponential backoff (wait_exponential, 4 attempts, ~30/60/120s waits), matching the openai family; the openai family's fixed 10/25/45s waits were replaced with the same exponential schedule. Long single-shot requests killed by TLS-inspecting corporate proxies (WinError 64 / connection reset) are now retried automatically instead of failing on the first attempt.
• Rate limits (429) and transient server errors (5xx) retried on all provider families: anthropic retries RateLimitError/InternalServerError, gemini retries ServerError (5xx) via the decorator — its internal 429 loop with Retry-Info delays is unchanged, and the 5xx re-raise is scoped to the chat path so undecorated methods keep their user-facing diagnostics — mistral retries SDKError with status 429/>=500. After retries are exhausted the user gets a clear "connection interrupted" message instead of a generic error.
• 18 new tests covering retry-then-success per family, retry exhaustion, policy configuration, no-double-retry for gemini 429, and mistral predicate behavior; mutation-verified non-tautological.
• README troubleshooting section for corporate TLS-inspecting proxies.