v0.1.6

Added

• Text-encoded tool calls are recovered — models that emit tool calls as plain text instead of
  the structured tool_calls field (Gemma, Qwen/Hermes-style, and OpenAI-compatible endpoints that
  render calls via their chat template) now work. Recognized formats: <tool_call>{json}</tool_call>
  (and the <|tool_call> … <tool_call|> and [TOOL_CALL] variants), the call:NAME{key:<|">v<|">}
  DSL, and ```tool_code Python-style calls. Only runs when the provider returned no
  structured call, so native function-calling is untouched.