v0.9.3 — Fix: AI term correction model (Groq removed qwen3-32b)

Fixed

• AI term correction failed with HTTP 404 because Groq removed the hardcoded
  qwen/qwen3-32b model. Switched to the production model llama-3.3-70b-versatile
  and dropped the qwen-specific reasoning_effort parameter.
• The model-availability check now reports a 404 clearly (“model unavailable — Groq may
  have renamed or removed it”) instead of a bare HTTP 404.