feat: generate convo summary with llm

[Dakuan]

Summary by cubic

Generate conversation titles with an LLM based on recent user messages, with safe thought-tag stripping and truncation. Titles update on the 3rd user message and fall back to the first user message for early chats.

• New Features

  • Regenerate title on every 3rd user message using the last 3 user messages and the previous title.
  • Sanitize model output by stripping tags and truncate to 120 chars.
  • Fallback to the first user message (truncated) when fewer than 3 user messages or when generation returns empty.

• Refactors

  • Added helpers: getUserMessageTexts, shouldRegenerateTitle, stripThoughtTags.
  • Reused a wrapped model with reasoning extraction for both streaming and title generation.
  • Added unit tests for new helpers and title logic.

^{Written for commit 532f3b9. Summary will update on new commits.}