v0.2.0

Everything below is about what happens around the captions: keeping them, reading them,
looking words up in them, and understanding who said what.

Added

• The Library window (menu bar → Library…, ⌘L). Every session is kept with its transcript
  and can be browsed, searched across, renamed, filed into folders, exported and annotated.
  Three columns: a folder tree of any depth, the sessions inside it, and the transcript.
• Persistent history in ~/Library/Application Support/Live Subtitles/history.sqlite —
  SQLite with FTS5, so search runs over everything ever recorded.
• ⌘K, one field over sessions, lines, words, notes, folders and commands. The same field
  finds a line and runs “Translate This Session”.
• Word lookup. Double-click any word for the system dictionary's entry — pronunciation,
  part of speech, senses — its meaning in Chinese, every other line it appears in, and a
  second opinion from Wiktionary. All of it inside the card; nothing opens a browser.
• Paragraph translation into Simplified Chinese, on demand and on-device. The first run
  asks macOS for the language pairs and the toolbar says so.
• Notes in Markdown, with the source line quoted in automatically and that line's words
  one tap away.
• Audio retention. Each session keeps the sound it was transcribed from — 16 kHz mono AAC,
  about 12 MB for a 45-minute episode, against the 6–8 % of a core the recogniser already uses.
• Speaker separation, run after each session and shown as a coloured, numbered chip on
  every line, with a switch between the enhanced reading and the raw transcript.
• LiveSubtitlesKit, a local package holding the storage layer, deliberately free of
  AppKit so it can be tested without launching the app. 12 tests, run in CI.

Changed

• The whole interface follows the macOS 26/27 design language: Liquid Glass where the system
  provides it and materials below that, a sectioned status menu, a Library menu with keys,
  and settings rebuilt as glass cards.
• Caption typography: SF Pro semibold, tracking scaled to the size, leading at 1.18×, and a
  wide soft shadow in place of the hard one, which read as an outline rather than as type.
• Study icons redrawn as one shape each at one weight, so the labels start on one line.
• The sidebar is a real folder tree: any depth, the disclosure triangle on the trailing edge
  as the Finder has it, and a New Folder row that is always there. Folders and sessions drag.

Fixed

• The library did not follow a recording. Cues were always written as they were finished, but
  the window read the archive once, when it opened, and never again.
• The sidebar width was never applied at all: a toolbar modifier sitting between the list and
  its width setting stopped it, so the sidebar stayed at its 192 pt default however large the
  window was — growing the window only widened the reading column.
• Double-clicking a word re-translated the whole episode. The scope and the request id were
  separate values, and a word lookup could leave the scope saying “paragraphs”.

Planned

• Batch re-recognition: re-transcribe a finished session from its retained audio, which
  brings punctuation, sentence casing and better accuracy. The evidence, the API to use and
  the traps already paid for are written up in
  docs/design/analysis-pipeline.md.
• Offline transcription of local files, pre-transcribed with a batch model and exported as a
  perfectly timed .srt with zero live latency.

---

Full Changelog: v0.1.6...v0.2.0