New buffer-based transcription → mind map core

Live transcription is no longer turned into branches phrase-by-phrase.
Now it first goes into a buffer, and when a new topic starts, the previous one is cleanly converted into a full mind map section.
This gives structure, clarity, and far more accurate mind maps with less manual cleanup.

Details

1. New core workflow

• Transcription is collected in a buffer
• When topic changes, the previous topic becomes a section in the mind map
• Sections are finalized automatically at the end of recording
• Mind map is now built from meaningful chunks, not random fragments

2. Cleaner and more focused mind maps

• Irrelevant pieces of speech are skipped, reducing noise
  Screenshot
• Empty final sections are removed (no fake "end" branches)
  Example
• Secondary branches are auto-collapsed to keep focus and reduce CPU
  Video
• Mind map is now the main view — transcript panel no longer pops up
  Before: https://youtu.be/bTwstnrgZUo
  After: https://youtu.be/ANSqZYrNQgc

3. Better reading & comprehension

• Paragraph size can be adjusted for fast or deep reading
• Still split by meaning, so the flow stays natural
  Screenshot

4. Smarter automatic section splitting

• Speech pauses help detect natural topic changes
• Embedding-based scoring improves accuracy
• ONNX and OpenAI models deliver cleaner, more human-like sections
• Result: fewer wrong cuts, less manual cleanup, more reliable structure

---

Fixes

• Fixed zoom when scrolling transcription in a branch
• Fixed speech finalization on empty paragraphs
• Fixed autoscroll when user scrolls manually
• Restored paragraph highlighting in transcription view
• Fixed audio playback in duplicated mind maps
• Fixed selection sync between branch and paragraph
• Removed raw HTML in notes, improved scrolling behavior

And other minor fixes