Default freebuff to DeepSeek V4 Pro, switch lite to Kimi

- Freebuff picker now offers DeepSeek (default, smartest, with "Collects data
  for training" warning), Kimi K2.6, and MiniMax. Both Kimi and DeepSeek run
  24/7 with 5/18h rate limits and 1000-slot instant-admit capacity.
- Codebuff Lite (paid) defaults to Kimi instead of DeepSeek to avoid silently
  routing user prompts through a model whose provider trains on them.
- Generalised the gemini-thinker session bypass from Kimi-only to any smart
  parent model, with `canFreebuffModelSpawnGeminiThinker` as the helper.
- Editor variant table replaces the chained ternary; only Opus retains
  <think>-tag scaffolding.
- Disabled INCLUDE_REASONING_IN_MESSAGE_HISTORY pending broader testing; the
  new stream-parser-reasoning tests skip while it's off.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>