mohdel v0.105.0

Added

• classifyProviderError now splits 429 responses into
  RATE_LIMIT_TIER (caller quota exhausted — retrying inside the
  rate-limit window cannot succeed) and RATE_LIMIT_LOAD (provider
  shedding load for reasons unrelated to the caller's quota — the
  next attempt may succeed immediately). RATE_LIMIT (no suffix)
  remains as a fallback when the signal is ambiguous.
• New opts.provider parameter on classifyProviderError lets
  adapters opt into provider-specific 429 disambiguation. Generic
  x-ratelimit-remaining-* header detection covers OpenAI-compatible
  providers; per-provider overrides cover anthropic
  (overloaded_error vs rate_limit_error), gemini
  (RESOURCE_EXHAUSTED), and cerebras (generic "high traffic" body
  → LOAD).
• Adapter wiring done for openai, anthropic, gemini, and every
  provider going through _chat_completions.js (cerebras, xai,
  deepseek, mistral, fireworks, groq, novita, openrouter).