0.8.0

Changed

• Clip boundaries are chosen on sentences. The model now reads the
  transcript as one sentence per line (derived from word punctuation, with
  unpunctuated rambles split at their longest pauses) instead of Whisper's
  segments, which break mid-sentence, and is told to start and end every clip
  on a line. The ending and opening reviews work on the same sentences. When
  a start still lands inside a sentence, the clip opens that sentence rather
  than skipping to the next; an end inside a sentence completes it.
• Tightened clips keep their tails. Removing pauses used to trim the room
  after the last word down to 0.3 s, so the export's 0.4 s fade ducked the
  final syllable. Tightening now keeps 0.7 s after the last word and 0.3 s
  before the first.
• Long videos come back in minutes, not an hour. Speaker-framing analysis
  (face tracking plus active speaker detection at 25 fps) is the slowest
  per-clip stage by a wide margin. The pipeline now runs it only for the top
  tier of clips (the eight highest-scoring, extended to twelve for scores of
  80 or more) and leaves the rest pending. A pending clip is analysed the
  moment it is opened in the editor, or before it is exported, and the
  editor says so while it waits. The analysis uses the clip's current trim,
  so a clip extended before opening is tracked end to end.
• Captions lay out by width, not word count. Groups are broken into lines
  from a per-style character budget derived from the font's measured glyph
  widths and the output aspect ratio, capped at two lines. The same layout
  feeds the preview and the ASS export (which now carries explicit line
  breaks with libass wrapping off), so a caption can no longer wrap
  differently in the file than it did in the editor.
• Captions hold briefly after a sentence. A finished group stays up for
  up to 1.5 s, cut short by the next group, instead of leaving an empty frame
  on every pause.
• Loudness normalisation is two-pass. Exports measure the clip first and
  apply one linear gain instead of the single-pass gain rider, which pumped
  audibly on speech: through loudnorm's linear mode when the peaks allow it
  (with the range target raised to the measured range, which that mode
  requires), otherwise a plain gain into a true-peak limiter. Falls back to
  single-pass only when the source cannot be measured.

Fixed

• The "fit under N MB" field could not be cleared or retyped. It was bound
  straight to the saved setting, so an empty field snapped back and every
  keystroke rewrote the settings file. It now edits a draft and commits on
  blur or Enter.
• Exported captions sat higher than the preview. The ASS style was
  bottom-aligned on the anchor line while the preview centred its block
  there, so two-line captions rendered about half a block too high. Events
  are now positioned middle-centre on the anchor.
• Whisper hallucinations reached captions and the clip picker. Segments
  Whisper itself flags as silence (high no-speech probability with
  low-confidence text) or as looped output (high compression ratio) are now
  dropped at stitch time, along with their words, and never primed into the
  next chunk.
• Words with zero or negative duration never lit up in karaoke captions.
  Word timings are now made monotonic and given a minimum on-screen duration
  when the transcript is stitched.

Added

• scripts/eval-clips.ts. Measures clip boundaries on saved projects
  (mid-sentence opens and closes, clipped or dead-air tails, length spread),
  and with --rerun compares them against fresh detection on the same
  transcript, so prompt changes can be checked instead of guessed.

• "Fit under N MB" export. A size cap in Settings and the editor export
  panel, so clips land under Discord, email or WhatsApp limits. The encoder
  already knew how; this is the missing control, plus the achieved file size
  (and a note when the planner had to downscale) after export.

• OpenAI-compatible API endpoints. Settings now takes a chat base URL
  (Azure, OpenRouter, Groq, LM Studio, Ollama) and an optional separate
  transcription URL, so a local Whisper server can sit next to a hosted LLM.
  Structured-output calls fall back from json_schema to json_object (and
  then a plain JSON completion) when the provider does not support OpenAI's
  strict schema mode. OPENAI_BASE_URL still overrides Settings when set.

• winget publishing path. A manifest generator (scripts/print-winget-manifest.mjs)
  and a Release workflow job that updates
  microsoft/winget-pkgs once the
  first listing exists. See docs/winget.md.

Removed

• The WorkVivo posting integration. It was specific to one organisation's
  internal comms platform and depended on an undocumented endpoint set, which
  made it an odd fit for a general-purpose tool. Everything it did that was not
  WorkVivo-specific stayed: the size-targeted two-pass encode
  (src/shared/uploadBudget.ts) is still here and still tested, and brand voice
  now steers the TikTok/Reels/Shorts post captions rather than only the internal
  ones.

Changed

• Brand voice settings (name, tone, style, things to avoid) now feed the AI post
  caption writer. Previously they only affected the internal posting captions,
  so the setting appeared to do nothing for most users.