Add plain text and transcript import with auto segmentation

[laubonghaudoi]

Summary

Add a browser-only import flow for pasted plain text or .txt transcript files, then split the text into draft subtitle cues. This is a high-value way to support caption creation without adding ASR or server-side processing.

Why this is in scope

• Runs entirely in the browser.
• Complements the existing "start from scratch" flow.
• Helps users who already have a transcript from another tool.
• Keeps AI transcription as a non-goal while making generated transcripts easier to edit.

Implementation plan

1. Add segmentation utilities.

   • Create lib/transcript-segmentation.ts.
   • Support these modes first:
     • Split by non-empty lines.
     • Split by sentence punctuation.
     • Split by maximum character count.
   • Output Subtitle[] with generated IDs, UUIDs, draft start/end times, and text.

2. Add timing options.

   • Default cue duration: 3 seconds.
   • Default gap: 0.2 seconds.
   • Optional max characters per cue.
   • Optional start time offset.
   • Keep the first version simple; alignment happens later on the waveform.

3. Add import UI.

   • Add an Import transcript option in the load subtitles / manage tracks dialog.
   • Allow paste into a textarea and loading a .txt file.
   • Show a preview table with cue number, start, end, and text.
   • Let the user create a new track or replace an empty active track.

4. Integrate with existing track actions.

   • Use loadSubtitlesIntoTrack or setInitialSubtitles depending on current track state.
   • Name generated tracks from the file name or Transcript.
   • Preserve undo behavior consistently with other imports.

5. Tests.

   • Add tests/transcript-segmentation.test.ts.
   • Cover line mode, sentence mode, max-character mode, empty input, and timing generation.

Acceptance criteria

• Users can paste text or load a .txt file and create draft subtitles locally.
• Segmentation is previewed before import.
• Imported cues can be edited, retimed, and exported like normal SRT/VTT cues.
• No ASR, cloud call, account, or backend is introduced.