mohdel v0.104.0

Added

• Anthropic prompt caching. splitPrompt translates structured system
  messages with cache: '5m' | '1h' markers into Anthropic's typed
  cache_control: { type: 'ephemeral', ttl?: '1h' } blocks. The adapter
  now extracts cache_creation_input_tokens and cache_read_input_tokens
  from message_start.usage and surfaces them on AnswerResult and via
  costFor.
• OpenAI / chat-completions cache reporting. OpenAI Responses,
  cerebras, fireworks, and xAI adapters extract
  prompt_tokens_details.cached_tokens and convert the OpenAI "subset
  of prompt_tokens" convention into mohdel's additive
  cacheReadInputTokens field. Caller code sees one consistent shape
  regardless of provider semantics.
• AnswerResult.cacheWriteInputTokens / cacheReadInputTokens
  added to the type definition. Symmetric write/read pair matching
  catalog cacheWritePrice/cacheReadPrice; Anthropic's
  cache_creation_input_tokens is normalized into cacheWriteInputTokens
  at the adapter boundary.
• computeCost honours cacheWritePrice and cacheReadPrice in
  catalog specs, falling back to inputPrice when absent so non-caching
  providers degrade gracefully. Pricing is additive: i*ip + cw*cwp + cr*crp + o*op + t*tp.

Changed

• Deps: @google/genai ^2.0.0 → ^2.0.1.