La Peace

 changes since 1.41

FIXED

- groq models from other companies work now, like openai/gpt-oss and qwen. those are
  reasoning models: they think quietly before they answer, and the thinking uses the same
  token limit as the reply. at 90 tokens they spent the whole limit thinking and sent back
  nothing. reasoning models now get room to think, and gpt-oss is asked to keep it short
  so replies stay quick
- when a model ran out of room or sent back nothing, the error just said "no response".
  it now says what happened and what to change
- a model cut off in the middle of thinking could put its unfinished thoughts into chat as
  kai's reply. that no longer happens