0.7.0 was the last version on this listing; 0.8.0 to 0.14.0 came out on GitHub only (https://github.com/Multysquid/shisu-ko/releases has every release's full notes). What changed since 0.7.0:

• 0.14.1: the first version on this listing since 0.7.0, with everything below.
• 0.14.0: Word colours: particles count as known and show green (a popup switch, on by default); names and Latin text are blue; your own list of known words, and Alt+Shift+K puts the word under the pointer on it; katakana words can count as known; a card for a noun also colours its する forms; the pitch overbar reads Yomitan's pitch graph.
• 0.13.0: a word mined in kanji is coloured where the subtitle writes it in kana (更に in さらに); a long line is never cut inside a word.
• 0.12.0: lines that are sentences, with their punctuation, never two sentences on one row; real lines are no longer thrown away as noise.
• 0.11: sung lyrics get subtitles; "No speech found in this video" when nothing was heard; a mined card gets the line that was on screen; lines merge across Whisper's segments and break between words; Left steps back a whole line.
• 0.11.0: word colours by the state of each word's Anki card (green, yellow, orange, red) and a pitch accent overbar, both off by default; setup asks which Whisper model to download.
• 0.10: only the tab you are watching is transcribed; subtitles pause when the speech is not Japanese.
• 0.9.0: the popup offers to update the server (the new notifications permission is for that); 0.8.0: a Start server button (optional nativeMessaging permission).
• Also: live streams; a master switch (Alt+Shift+S); the Whisper model picked in the popup; any installed font for the subtitles; a new card matched to its subtitle by sentence and word, with the screenshot and audio prepared while the line plays; Chrome support.

Model picker: the popup's "Transcription model" field takes any faster-whisper size or Hugging Face CTranslate2 repo id; sent with every /sync; the server validates the name (never a path), canonicalises aliases, downloads on a side thread while the old model keeps subtitling, swaps between windows, restarts sessions, keeps the other model's cue cache as <id>.<model>.cues.json, reports model/model_loading/model_error (+ /health lists downloaded models). Overlay shows "Loading model X…" / the error.
Font family: any installed font in front of the preset stack, with a live preview and an installed/not-found probe in the popup; sanitised before CSS.
Launcher update step (server/update.py): run.cmd/run.sh fast-forward a git checkout or fetch the newest release zip first; --no-update skips it.
Stale .part downloads that YouTube answers with 416 are discarded and downloaded afresh.
AMO listing package under docs/amo/ + publish-addon.cmd; homepage_url in the manifest; .amo-upload-uuid untracked.
Docs: README/AGENTS.md sections for all of the above, live streams marked Firefox-only, README/reviewer notes/privacy policy aligned with what 0.6.0 actually sends.
Tests: 290 server (from 190), 136 addon (from 73).