0.7.3

[0.7.1] - 2026-07-28

Fixed

• Glossary terms now reliably reach the model. The glossary is sent inside
  the user-turn task statement as well as the system prompt (translation-tuned
  models weigh the former), reworded per task (translation vs. summary), and
  emitted as one terse line instead of an instruction paragraph. The term list
  is capped relative to the text length so a large glossary can no longer
  drown a short paragraph and derail the translation.
• Glossary matching misses fewer occurrences. Terms whose edges are CJK
  characters match inside unspaced Japanese/Chinese prose (the word-boundary
  guard no longer applies there), and English entries also match simple
  plural/possessive forms ("Caesars", "Caesar's").
• Blocks that are exactly one glossary term skip the model. Headings such
  as a bare person's name are rendered directly with the glossary target — in
  page translation, selection translation, and EPUB export — instead of
  risking a model round-trip mangling the mandated form.

0.7.0

[0.7.0] - 2026-07-19

Added

• Bilingual UI (Japanese / English). All menu, toolbar, dialog, and status
  strings are now translatable. The UI language follows the system locale
  (Japanese system → Japanese UI, anything else → English) and can be forced
  via the new 表示 → 言語 / Language menu (applied after restart). Source
  strings remain Japanese; the English catalog lives in i18n/spindle_en.ts
  (refresh with the update_translations CMake target, compiled into the
  binary by lrelease at build time).

0.6.6

[0.6.6] - 2026-07-19

Fixed

• Facing pages meet cleanly at the spine. Each page of a spread now
  edge-aligns toward the center instead of centering in its own half, the 2px
  seam between the two views is gone, and the page-fit measurement no longer
  loses the width of the pre-fit scrollbar — which had left a theme-colored
  strip at the spine.
• Page turns can no longer leave pages unscaled. Navigating quickly could
  occasionally skip the reader-script injection, showing the raw page at
  natural size with scrollbars. The injected-script collections are now only
  rewritten when their content actually changes, and every finished load
  verifies the fixed-layout fit, re-injecting the script if it never ran.
• Backward navigation keeps the established spread pairing. Spread pairs
  are anchored at the front of the book, so stepping back from a lone final
  page rejoins the canonical pairs instead of shifting every spread by one.
• Arrow keys follow the reading direction on fixed-layout pages. In
  right-bound books ← advances and → goes back, matching the edge-click
  mapping.

0.6.5

[0.6.5] - 2026-07-18

Changed

• The selection-translation popup closes on any click outside it. Clicks
  landing on the reading view (a native child window invisible to the popup's
  mouse grab) previously left the popup open; only clicking outside the app
  window or pressing Escape dismissed it.

Fixed

• Selection translation works in fixed-layout books. The facing-page
  spread view was created after the web channel was wired up, so selections on
  the companion page could never reach the app — no menu appeared. The
  companion page now connects to the bridge; selecting text there opens the
  text-action menu (翻訳 / 要約 / コピー / Web で検索). Highlight colors and
  notes are offered only on the primary page, where the selection can be
  anchored to the current chapter.

0.6.3

[0.6.3] - 2026-07-18

Added

• Fixed-layout EPUBs can be read as facing-page spreads. The new
  表示 → 固定レイアウトを見開き表示 option is enabled by default and honors
  the book's left/right page-progression direction plus page-spread-*
  placement metadata. Clicking the outer 15% of either page also turns the
  spread in the corresponding physical direction. The first spine page is
  always treated as a cover and shown alone. Binding direction can be set to
  automatic, right-bound, or left-bound; automatic uses right binding for
  vertical writing and otherwise follows the EPUB.

Fixed

• Fixed-layout arrow navigation turns the page on the first key press.
  Left/Right are intercepted before Chromium can use the first press to
  horizontally scroll an oversized page. The complete fixed page is fitted to
  the reading pane while retaining its authored aspect ratio. Books that omit
  rendition:layout are also recognized when every spine document is exactly
  one image page with no surrounding text.
• Single-image chapters now always open fitted to the reading pane. The
  source image's aspect ratio determines whether its width or height meets the
  window edge, independently of the retained text zoom. Image zoom and mouse
  drag panning now use the image's actual bounds, avoiding scrollable empty
  space on the non-overflowing axis.
• Book-search results now jump to the selected occurrence. Repeated matches
  in the same chapter use their stored text offsets instead of always opening
  Chromium's first match. Pressing Enter in the search field opens the first
  result immediately.

0.6.2

[0.6.2] - 2026-07-15

Added

• Configurable line wrapping. 表示 > 折り返し設定… can keep text fitted to
  the window or limit each line to a chosen 10–120 character measure (initially
  40). The setting is saved across launches and also works with vertical writing.

0.6.1

[0.6.1] - 2026-07-14

Fixed

• Glossary extraction no longer times out when a model keeps generating
  malformed JSON. Responses now use a strict schema and a bounded output
  length, so a runaway model cannot occupy Ollama until the five-minute
  request timeout.

0.6.0

[0.6.0] - 2026-07-14

Added

• Glossary generation (用語集を生成). 翻訳 > 用語集を生成… extracts proper
  nouns and recurring terms from the current chapter or the whole book with
  Ollama (JSON-forced output, summary model) and writes their target-language
  wordings into <book>.glossary.json, merging with any existing entries
  (existing ones win). Extracted terms are kept only if they actually occur in
  the scanned text; progress dialog with cancel (cancel keeps the terms found
  so far), and the updated glossary applies to translations and summaries
  immediately.

0.5.2

[0.5.2] - 2026-07-12

Added

• Audio-file export (章を音声ファイルへ書き出し). The 読み上げ menu can
  render the current chapter to a single audio file using the same voices and
  原文/訳文 rule as playback, with a progress dialog and cancel. Blocks
  synthesized by different engines/sample rates (VOICEVOX 24 kHz, Piper
  22.05 kHz, OS voices) are resampled into one stream with a 300 ms pause
  between paragraphs. OS voices are used via QTextToSpeech's synthesize
  capability where the engine supports it; VOICEVOX and Piper always support
  export. WAV (16-bit mono) is written directly; MP3 / M4A are encoded
  through ffmpeg (auto-detected on PATH, or set 読み上げ > 音声設定 >
  ffmpeg) — MP3 at ~130 kbps VBR (-q:a 4), M4A as 96 kbps AAC.

[0.5.1] - 2026-07-12

Added

• Local AI voices for read-aloud: VOICEVOX and Piper. Besides the OS
  voices, the voice pickers list the speakers of a running
  VOICEVOX server (VOICEVOX: 四国めたん （ノーマル）, Japanese) and the .onnx voice models of a local
  Piper install (Piper: en_US-lessac-medium, multi-language). Picking a prefixed voice routes that
  language to that engine — so e.g. Japanese can be spoken by VOICEVOX while
  English in the same book falls back to an OS voice. The VOICEVOX endpoint
  (default http://localhost:50021) and the Piper executable / voices folder
  are set in 読み上げ > 音声設定. Speech rate maps to VOICEVOX speedScale /
  Piper --length_scale; synthesized audio plays through Qt Multimedia (the
  engines are built only when that module is present).

[0.5.0] - 2026-07-12

Added

• Read-aloud (読み上げ). The current chapter can be spoken block by block
  with the OS speech voices (Qt TextToSpeech; WinRT/SAPI on Windows,
  AVSpeechSynthesizer on macOS, speech-dispatcher on Linux). Reading follows
  the translation display mode — 原文 and 対訳 read the source text, 訳文
  reads the translation (falling back to the source where none is cached
  yet) — with the voice matched to the spoken language. The spoken paragraph
  is tinted and auto-scrolled into view,
  <ruby> text is spoken as its <rt> reading instead of base + furigana
  concatenated, reading starts from the paragraph currently on screen, and the
  chapter end can auto-advance into the next chapter (読み上げ menu toggle).
  Speech rate and a per-language voice are configurable in 読み上げ > 音声設定;
  play/pause/stop live on the toolbar and in the new 読み上げ menu
  (Ctrl+Shift+S toggles). Builds without Qt TextToSpeech still compile, with
  the actions reporting that no speech engine is available.

0.4.5

[0.4.5] - 2026-07-12

Added

• Bare-text chapters are now translatable. Some EPUBs (often converted
  ones, e.g. Aozora/Gutenberg conversions) carry chapter prose as bare text
  directly inside <div>/<body>, separated by <br/><br/> instead of
  <p> elements. Such text had no block structure, so translation,
  highlighting, search jumps, and Kindle-note matching only saw the headings.
  Chapters are now normalized on load: each run of bare inline text becomes a
  synthetic paragraph (split at 2+ consecutive <br>, spacing preserved),
  consistently across the reading view, sidecar caches, and translated-EPUB
  export.
• Wrong-language translations are caught. When the model answers in a
  language whose script clearly doesn't match the target (e.g. Chinese
  returned for a Japanese target), the request is retried once with a
  stronger instruction; if it still doesn't comply the paragraph shows an
  error instead of caching the wrong text.

Fixed

• Chinese ⇄ Japanese translation reliability. The translation prompt
  told the model to return text "already in the target language" unchanged,
  which made kanji-heavy Japanese pass through untranslated with a Chinese
  target (and vice versa) — and the untranslated result was then cached.
  The prompt now states the task in the user turn in a form that
  translation-tuned models (e.g. TranslateGemma) also follow, and no longer
  offers the pass-through escape hatch.
• Reasoning models' chain of thought leaked into translations. Content
  ending with a </think> marker (e.g. Qwen3 with thinking nominally
  disabled) now has everything before the marker stripped.