Changelog

[1.4.73] - 2026-06-22

Fixed

• The preferred Japanese site-language feature now waits for page locale metadata before making generic /en to /ja URL guesses, and skips the guess when a page declares supported alternates without Japanese. Japanese locale/cookie/browser hints and explicit hreflang="ja" or known-site redirects still apply, but multilingual sites without Japanese pages no longer get sent to broken guessed routes such as /en/ja.
• Selection popovers now close on outside click without clearing the native page text selection, and the same selected text stays dismissed instead of immediately reopening the modal.
• The selected-text lookup feature is now exposed clearly as Selection popups in Reader settings, so users can turn it off while keeping click, hover, annotations, and study tools enabled.

[1.4.71] - 2026-06-22

Changed

• Updated every Discord invite link to the new server invite.

[1.4.70] - 2026-06-22

Changed

• Default settings now open in English again, the homepage next-steps grid no longer includes the support card, and the docs header includes a Discord link beside GitHub.

[1.4.69] - 2026-06-22

Fixed

• Compact carousel and storefront media-card labels now suppress furigana when a narrow title link sits beside a sibling cover link, and direct text inside flex/grid controls is wrapped as one inline run. This keeps book shelves, icon menus, and product carousels from expanding, stacking, or overlapping after page annotation.

[1.4.68] - 2026-06-22

Fixed

• BookWalker canvas OCR now re-snapshots the current page after page turns when image OCR is set to tap/hover instead of automatic, so hovering or tapping the reused viewer canvas triggers OCR for the new page instead of reusing the previous snapshot.

[1.4.66] - 2026-06-22

Fixed

• The play/resume control on a paused video now appears immediately instead of waiting for OCR to finish, so you can always unpause right away. The captured frame snapshot and scanning indicator still stay hidden until OCR has text, keeping the player's own comment/like/scrubber controls reachable while it runs.

[1.4.65] - 2026-06-22

Fixed

• Compact image/navigation labels now keep their original inline layout and suppress furigana, preventing sites with icon menu links such as Travel Donkey from stacking labels or stretching header/sidebar rows.

[1.4.64] - 2026-06-22

Fixed

• Page-word highlights now refresh contrast before scan chunks yield and mirrored hosts hide before their overlay is attached, preventing dark first-paint highlights that only corrected after hover.

[1.4.63] - 2026-06-22

Fixed

• Mirrored input and textarea placeholders now hide the native placeholder while the annotated overlay is visible, preventing double-rendered text in compact controls such as Google/Gemini voice/search inputs. Placeholder mirrors suppress furigana in the 24px control line and clean up their temporary styles as soon as the field changes.
• YouTube title mirrors with furigana now use ruby-friendly line-height and the crop sweep detects absolute mirror overflow, so modern lockup/card titles reserve enough height instead of clipping or packing ruby into broken-looking line breaks.

[1.4.62] - 2026-06-22

Fixed

• Furigana-bearing reader words now reserve more vertical line-height on tight mobile Safari layouts such as the hosted よむ homepage, and page-word highlights stay bounded to the text glyph height instead of expanding into chunky blocks across the taller ruby line.
• Text mirrors now keep their host overflow visible while annotations are active, preventing mirrored ruby from being clipped by compact page chrome.
• YouTube feed continuations now load under YouTube's own scroll observer instead of being programmatically pulled into view by Yomu, fixing homepage jumps to the Shorts/continuation area on iPhone and avoiding stuck desktop/mobile skeleton rows while scrolling.
• The "Prefer Japanese site language and location" feature no longer wraps page fetch or XMLHttpRequest calls. Sites still receive Japanese URL/cookie/locale hints, but fragile app request pipelines such as YouTube continuations and Reddit Shreddit APIs are left untouched, fixing ReferenceError: isSameOriginRequestUrl is not defined, "Request failed" banners, and stuck infinite-scroll hydration.

[1.4.61] - 2026-06-22

Fixed

• Mobile YouTube reload loop ("A problem repeatedly occurred on https://m.youtube.com/?ra=m&hl=ja&gl=JP"). Root cause was the preferred Japanese site language feature (pre-existing, unrelated to the 1.4.58 work): (1) its Accept-Language fetch/XHR wrapper added the header to — and re-derived — cross-origin requests, breaking YouTube's gstatic icon/script loads in WebKit/Safari ("access control checks"); and (2) its alternate-URL redirect watcher re-fired on every m.youtube SPA navigation, full-reloading back to the Japanese URL forever. The header is now only added to same-origin requests, and the redirect runs at most once per host per tab session (the language cookie keeps the site Japanese afterward). Reproduced and verified in WebKit + a Japanese locale.
• Keyboard shortcuts no longer swallow normal typing (e.g. Shift+H) in inputs that live inside a shadow root, such as YouTube's search box — the keydown gate now checks the event's composed path, not just the retargeted shadow host.

Changed

• Re-introduced the 1.4.58 YouTube/mobile UX improvements (they were not the cause of the reload loop): portrait/Shorts subtitle sizing + position, frame-synced subtitle timing, paused-frame OCR overlay gating, reliable pause-on-lookup, snappier taps with fewer mispresses, iPad/Apple-Pencil touch targets, bounded comment pitch retry, and broader in-player YouTube text coverage.

[1.4.60] - 2026-06-22

Fixed

• Full rollback of the 1.4.58/1.4.59 YouTube-mobile-UX changes: they broke mobile YouTube page loading and the 1.4.59 partial revert did not fix it. This release restores the exact known-good 1.4.57 reader/page-scan/subtitle/OCR behavior. The reverted work (mobile subtitle sizing/sync, paused-frame OCR gating, pause-on-lookup, tap reliability, iPad touch targets, comment pitch, broader YouTube text coverage) will be re-introduced one change at a time only after on-device mobile validation, since the regression could not be reproduced in the headless test harness.

[1.4.59] - 2026-06-22

Fixed

• Hotfix: revert the 1.4.58 change that parsed Japanese inside YouTube's player/overlay wrappers (and added end-screen/pause-overlay scan roots). On real mobile YouTube (m.youtube.com) the broadened whole-page scan over the large, continuously-hydrating player subtree could freeze the page so it never finished loading. The page-scan/parsing behavior is restored to the known-good pre-1.4.58 state (the native-caption exclusion was always kept). Broader in-player text coverage will return only after on-device mobile validation.
• All other 1.4.58 fixes are unaffected and remain: mobile subtitle sizing/position, frame-synced subtitle timing, paused-frame OCR overlay gating, reliable pause-on-lookup, snappier taps with fewer mispresses, iPad/Apple-Pencil touch targets, and the bounded comment pitch-accent retry.

[1.4.58] - 2026-06-22

Fixed

• Mobile/Shorts subtitles are larger and sit higher by default so they clear the Shorts action rail and scrubber. Font scaling is now portrait-aware (it no longer collapses to ~17px on a phone), and the size/position settings stay authoritative — an explicit value always wins.
• Subtitle timing follows the video frame-by-frame (via requestVideoFrameCallback, with a requestAnimationFrame fallback) instead of a ~250 ms sampler, so cue changes and karaoke highlighting stay in sync, including at 1.5–2× speed. The per-frame sampler runs only while playing and is cancelled on pause, tab-hide, and teardown to protect battery.
• Paused-frame OCR no longer covers the player before it finishes: the captured frame, scan status, and resume button stay hidden until OCR produces text, so the native player and its comment/like/scrubber controls remain reachable while scanning. A dictionary lookup that pauses the video no longer spawns an OCR overlay over the player, and the overlay re-aligns on rotate/fullscreen/pinch-zoom.
• Opening the dictionary reliably pauses the exact video the subtitles track (instead of guessing the largest video on the page), keeps it paused while drilling into nested lookups, and resumes it on close.
• Tapping a word is snappier with fewer mispresses: a tap that lands just off a word is now resolved at finger-up instead of waiting for the delayed synthetic click, and word hit targets no longer overlap their neighbours.
• iPad with an Apple Pencil now gets the touch-sized tap targets, hit-slop, and 44 px controls (touch ergonomics key off any-pointer: coarse instead of the primary pointer, which a Pencil reports as fine).
• YouTube comments receive pitch-accent underlines more reliably on keyless setups: budget-denied words are retried through a bounded, paced public-pitch lane instead of being abandoned, without exceeding a per-page lookup cap.

Changed

• More YouTube text is parsed: Japanese inside the player/overlay wrappers, end-screen cards, and pause-overlay titles now receives furigana and pitch coloring. The native caption overlay stays excluded so it is never double-rendered, and the reader never re-ingests its own annotations.

[1.4.57] - 2026-06-22

Fixed

• YouTube and Google search Japanese text coverage is stricter: comments, metadata, action labels, AI/result cards, related chips, and compact video rows keep base text visible while receiving furigana, pitch coloring, and pitch underlines instead of being clipped or skipped.
• Touch lookup on mobile is more reliable: normal page/subtitle reader words now open on a tap-only pointer release, cancel cleanly during scroll drags, and have a slightly larger coarse-pointer hit target without moving page layout.
• Hover lookup audio now reuses a fresh tap/click-time media reservation for automatic playback, so iPad/Safari hover flows can play term audio after a recent reader gesture instead of being blocked as delayed non-gesture media.

[1.4.56] - 2026-06-22

Fixed

• YouTube watch titles and channel metadata no longer go missing, duplicate, or misalign after the page re-renders. The non-destructive furigana mirror hides the host text and overlays an annotated copy; a YouTube re-render that rewrote the host's style/class without changing its text used to strip the mirror's visibility:hidden/position:relative (so the native title reappeared beside the mirror, or the overlay drifted out of place). The mirror now watches the host's own attribute changes and re-pins those styles.

[1.4.55] - 2026-06-22

Changed

• Internal: extracted the new-tab immersion-example audio playback engine into a dedicated, unit-tested NewTabImmersionAudioPlayer module (request-id invalidation so a slow load never plays over a newer card), and converged a duplicated URL-dedupe helper onto the shared uniqueTrimmedStrings util. No change to behavior.

[1.4.54] - 2026-06-22

Fixed

• Parser-disabled/native site profiles now still run the final visible-Japanese sweep, so storefront chrome, buttons, nav, comments, and other visible Japanese text remain readable instead of silently opting out.
• YouTube guide and mini-guide labels are no longer excluded from the residual scan path; visible Japanese YouTube chrome keeps ruby, source coloring, and pitch underlines unless it is an actual native player control.
• YouTube desktop and mobile comments now stay in the inline parser path even after host re-renders, avoiding text mirrors, unwanted "詳細" expansion buttons, and missing furigana/pitch on comment text.
• The YouTube subtitle transcript resize interaction now clears cleanly on pointer cancellation, lost capture, and mouseup fallbacks, fixing stuck resize states, manual-scroll glitches, and highlight jumps during live playback.
• The YouTube mobile/fullscreen smoke harness now retries transient navigation races while installing userscript CSS, making the fullscreen sidebar regression check stable against YouTube page refreshes.

[1.4.53] - 2026-06-22

Changed

• Internal: extracted the new-tab offline grade write-behind queue (persist failed grades, dedupe, retry on reconnect) into a dedicated, unit-tested NewTabGradeQueue module, further shrinking the new-tab controller. No change to behavior.

[1.4.52] - 2026-06-22

Fixed

• The Japanese-site request now catches more generic multilingual URL patterns (en subdomains and /en or /en-US path segments) and the Yomu puck can toggle it on or off. Turning it off after Yomu redirects the page returns to the remembered original URL.

[1.4.51] - 2026-06-22

Changed

• Internal: extracted the new-tab kanji-detail multi-source fetch/cache subsystem into a dedicated, unit-tested KanjiDetailSource module (load/invalidate/clear) and lifted the shared promiseWithTimeout helper into core/async-utils, shrinking the new-tab controller. No change to behavior.
• Internal: moved pointer hit-testing geometry out of the new-tab controller into a focused dom/pointer-geometry module, converging duplicated rect math onto the shared dom/rect primitives. No change to behavior.

[1.4.50] - 2026-06-21

Fixed

• Hover lookup audio now attempts and preloads immediately even before the page has browser-level user activation, fixing hover cards that stayed in the audio loading state while clicked cards played.
• YouTube comments now render ruby/pitch inline without falling back to text mirrors that make every short comment look overflowed and show "詳細".
• YouTube chrome text, mobile controls, guide labels, and comment buttons stay in the ruby/pitch scan path while native caption windows remain reserved for YouTube itself.
• Generic page scanning now does a final visible-Japanese sweep across comments, nav, buttons, tabs, and other UI text on ordinary pages, while BookWalker storefront chrome and other parser-disabled/native profiles keep their DOM opt-outs.
• Subtitle primary text no longer flashes unparsed while parser enrichment is pending; lines render only after parsed furigana/pitch-ready HTML is available.

[1.4.49] - 2026-06-21

Changed

• Internal: extracted the live JPDB review bridge → card adapter and the live-card identity helper into a dedicated, unit-tested jpdb-live-card module, shrinking the new-tab controller and giving the live-review card model a single owner. No change to behavior.

[1.4.47] - 2026-06-21

Fixed

• ChatGPT, Claude, Gemini, and Grok-style markdown/message responses now scan as normal readable prose, so furigana, reader word color, and pitch underlines render on Japanese chatbot answers.
• Click lookup now resolves passive/message words from rendered geometry when the page reports a wrapper as the click target, opening the sticky dictionary while keeping native buttons and controls click-through.
• Hover lookup audio now attempts and preloads immediately even before the page has browser-level user activation, fixing hover cards that stayed in the audio loading state while clicked cards played.
• Reader ruby/furigana CSS now wins against more aggressive host markdown resets.
• YouTube Shorts subtitles now keep visible primary text while parser enrichment is pending, then rebake cached cue and transcript HTML with furigana/pitch when enrichment completes.
• Paused-video OCR snapshots now keep the native player visible until OCR text is ready, then reveal the still frame together with the parsed OCR overlay.
• YouTube mini-guide navigation stays native, while comments, buttons, mobile controls, and compact titles remain in the ruby/pitch annotation path.

[1.4.46] - 2026-06-21

Fixed

• The "Prefer Japanese site language and location" setting now actively redirects common multilingual sites to their Japanese URL variants instead of relying only on browser-language hints. YouTube, Google, Google News, MDN, GitHub Docs, Microsoft Learn, Apple Support, Wikipedia alternates, and other pages that expose Japanese alternate links now move to Japanese reliably when the setting is enabled.

[1.4.45] - 2026-06-21

Added

• The puck radial menu now has an OCR mode toggle that cycles Auto -> Tap/Hover -> Off. Tap/Hover keeps image OCR available on intentional pointer activation and keeps paused-video frame OCR available, since videos do not have a still image to tap.

Fixed

• Turning OCR fully off now releases any paused-video OCR frame overlay as well as image/canvas OCR overlays.

[1.4.44] - 2026-06-21

Changed

• Folded the remaining upstream release heads and current local topic branches into main as superseded integration merges, leaving no unmerged remote branches.

Fixed

• YouTube and Yomu-hosted video frames that request fullscreen on the bare video element now redirect fullscreen to the player container, so the subtitle overlay remains visible above the video in mobile/narrow fullscreen layouts.

[1.4.43] - 2026-06-20

Changed

• Removed Yomu's bundled public proxy fallback. Cross-origin requests now use direct/browser-readable paths, the userscript bridge, or a proxy URL the user explicitly configures.

Fixed

• JPDB vocabulary detail pages no longer re-render JPDB's native large headword, fixing malformed furigana/layout on entries such as 表示.
• Japanese Settings parsing now replaces overbroad fallback tokens with curated settings words, so labels such as ポップアップ表示 click and underline as 表示 instead of opening a broad synthetic selection.
• Settings lookup-link rows now keep the label, URL template, order controls, and remove controls aligned on wider layouts.
• Opening Settings no longer toggles inert on the host page, avoiding long modal-open pauses on large sites while preserving the dialog backdrop and focus trap.

[1.4.42] - 2026-06-20

Fixed

• Keyless public Jiten parsing now hydrates visible words with cached detail records so furigana and pitch accents appear without an API key while public requests stay capped and paced.
• Welcome splash reader words now open the dictionary popover on click or keyboard activation without stealing onboarding action button clicks.
• Japanese Settings no longer repeats every dropdown option below each select. Settings now keep only the parsed selected-value mirror where native controls need it, ignore stale option-list metadata, and apply furigana/pitch rendering from the current form controls in both hosted and userscript settings.

[1.4.41] - 2026-06-20

Fixed

• Not-in-deck words now use a quieter default slate-blue state and a softer highlight mix, so pages where most parsed words are outside the deck no longer turn into a bright wall of blue. Actual new and in-deck words keep the existing stronger blue.
• Headword furigana tests now compare the visible word surface instead of raw ruby text, matching the 1.4.40 popup headword rendering.

[1.4.40] - 2026-06-20

Fixed

• Popup headwords can now render furigana while keeping each kanji as a clickable drilldown button, so annotated spellings such as 大[たい]変[へん] show ruby instead of bare kanji-only buttons.

[1.4.39] - 2026-06-20

Fixed

• Apple Pencil handwriting on the Search tab no longer drops live strokes when Safari loses pointer capture mid-stroke. The shared kanji doodle canvas now keeps listening through document/window pointer events until the real pointerup/cancel, so Search handwriting, new-tab kanji practice, and popup kanji practice all use the same patched path.
• Search results now include the kanji reading/inline-word metadata and hidden canonical spellings that the current result cards expect, so ruby-rendered words still expose their plain surface text.

[1.4.38] - 2026-06-20

Added

• Added ASB-style subtitle timing controls under Tracks so loaded primary/native subtitle files can be nudged by 100 ms, aligned to the previous or next subtitle at the current playhead, and reset without reloading the track.

Fixed

• YouTube Shorts subtitle rows now wrap long parsed/ruby cues instead of clipping after the first visible segment, and native YouTube translation overlays recover when the translated timedtext response is empty by translating from the source captions.

[1.4.37] - 2026-06-20

Fixed

• Installed frequency dictionaries, including the Jiten frequency dictionary, now live under Settings -> Sources -> Lookup pills and render their popup frequency badges from the same enabled/order controls instead of a separate frequency-dictionary subsection.
• Hover term audio now treats a fresh hover as a new auto-play attempt for the same word, so re-hovering or reopening a hover popup is no longer skipped by duplicate-render guards while repeated renders from one hover still stay quiet.
• Fullscreen subtitle fallback now ignores ordinary page metadata such as title links, category chips, navigation labels, and centered text just below the player. This prevents fake one-line subtitles like 生成 フルボイス from appearing after fullscreen on sites without real subtitle tracks, while preserving real custom caption overlays.
• Clicking subtitle words in ASB and other subtitle-style overlays now pauses the playing video again, then resumes it when the lookup closes.
• Tapping parsed words inside popup dictionary prose now opens a nested lookup instead of leaving those words render-only; dictionary links and source summaries still keep their native click behavior.

[1.4.35] - 2026-06-20

Added

• Added a Reader setting to turn off Yomu's own lookup popup when using jpdb reader, Jiten Reader, or Yomitan for popups, while keeping Yomu annotations, media tools, mining, and study features available.

[1.4.34] - 2026-06-20

Changed

• Imported dictionary sources now render directly in the definition stack instead of sitting behind a combined "Dictionaries" wrapper panel.

Fixed

• Settings now keep furigana parsing and pitch underline styling in the API, furigana, and underline controls after tab changes and late status refreshes, including hosted settings text that only has local fallback readings.

[1.4.33] - 2026-06-20

Fixed

• Recommended dictionary buttons in Settings now say "Update" immediately when the matching dictionary is already installed, including saved/imported title variants such as Jitendex.org and JPDB frequency dictionaries.
• Historical lookup-link settings now migrate cleanly after retiring the goo lookup link, preserving the intended Jiten, JPDB, then Yomu order.
• The Help tab Donate button now keeps its accent fill, border, and readable text instead of inheriting the dark base button background.

[1.4.32] - 2026-06-20

Fixed

• Pitch accent now loads for every word on a page at once instead of trickling in a few words at a time. On non-YouTube pages, local-dictionary pitch lookups (which use no network) were capped at 12 words per scan and the rest were silently dropped, so dense pages only ever showed pitch on a handful of words until you scrolled or clicked each one. The whole visible batch is now enriched together — in idle-paced chunks that keep the page responsive — while public/online lookups stay throttled exactly as before.

[1.4.31] - 2026-06-20

Fixed

• The Media → Audio sources list is now authoritative again. Playback follows the order you configure (the first enabled source that has audio wins) instead of picking a source at random or racing to whichever responds fastest. Custom sources — including a local audio server such as http://localhost:9090/?term={term}&reading={reading} — now play from the position you place them in the list. "Shuffle audio" continues to vary the individual clips a single source offers, but no longer reshuffles the source priority list itself.
• Hover audio no longer lags behind clicking. Any word that auto-plays on hover now prefetches its audio while the popup loads — matching the always-warm click/modal path — so playback starts promptly. This especially fixes common single-character words (火, 水, 人 …), whose audio was previously fetched cold on hover.

[1.4.30] - 2026-06-20

Fixed

• Discord and other page-word surfaces now keep highlights, underlines, and text colors on particle and short-token spans by preserving the shared state and pitch classes instead of leaving those tokens visually inert.
• Strict or stale-CSS pages now get the same critical fallback for status, JPDB/Jiten, Anki, and pitch word-color channels while the full yomu.css resource is unavailable.

[1.4.29] - 2026-06-20

Changed

• Internal: separate the embedded grammar dataset (GRAMMAR_PATTERN_DATA, GRAMMAR_RULE_EXAMPLES) from the grammar-detection logic in study/tools.ts, moving ~2.85k lines of data into study/grammar-data.ts. The logic file drops from 3197 to 736 lines. No behavior change; the 635-case grammar suite stays green.

[1.4.28] - 2026-06-20

Fixed

• Rolled the direct recommended dictionary installs onto the latest release line so WTY JA-JA, Pixiv Light, and JPDB Kanji remain available from direct ZIP install buttons alongside the Anki module refactor.

[1.4.27] - 2026-06-20

Changed

• Internal: split the oversized Anki integration module (anki/client.ts, ~2.9k lines) into focused modules — mobile handoff, card templates, media files, model/field mapping, field rendering — and removed duplicated cache and array helpers (one shared core/public-cache and core/array-utils). No user-facing behavior change; the userscript bundle is ~1.8 KB smaller and gains regression tests for the deep-link, cache, and array utilities. Also realigned settings/JPDB test copy that had drifted from the shipped i18n.

[1.4.26] - 2026-06-20

Fixed

• Recommended local dictionaries now publish with direct Yomitan ZIP install links, including WTY JA-JA, Pixiv Light, and JPDB Kanji, while dropping the homepage-only MarvNC collection card.
• Mobile YouTube comment text keeps its reader parsing and scan coverage while preserving the settings copy required by the generated form checks.

[1.4.18] - 2026-06-20

Fixed

• Discord and other strict or stale-CSS pages now keep the default reader word highlight, pitch underline, and pitch text-color fallback when the full yomu.css resource is unavailable; the inline critical fallback now carries those shared word visual channels instead of only base pitch underline styling.

[1.4.17] - 2026-06-20

Fixed

• Recommended local dictionaries now install directly from usable Yomitan ZIP links, with WTY JA-JA fixed, generic homepage-only MarvNC removed, and Pixiv Light plus JPDB Kanji added from the local export.

[1.4.16] - 2026-06-19

Fixed

• Recommended local dictionaries no longer show homepage-only cards for monolingual entries: WTY JA-JA now installs from its documented Yomitan zip, the generic MarvNC collection card was replaced with direct Pixiv Light, and JPDB Kanji was added from the local dictionary export because it has a usable direct zip.

[1.4.15] - 2026-06-19

Fixed

• The JPDB/Jiten grading-provider switcher now sits in the review-target gutter, stays available for users with both API keys, and uses the provider label itself as part of the tap target.
• Single-provider setups no longer show redundant "JPDB" or "Jiten" review-target text next to the grade buttons.

[1.4.14] - 2026-06-19

Fixed

• On YouTube, the Japanese subtitle line no longer goes missing while the English (native) line shows alone. Auto-generated captions and their auto-translated Japanese track are segmented independently, so the Japanese cue can start a beat after — or fall into a gap relative to — the English line that's already on screen; the pair now appears together by surfacing the Japanese cue aligned to the active English one. This also steadies the transcript side panel, which flickered when the active line momentarily had no Japanese cue.

[1.4.13] - 2026-06-19

Fixed

• Jiten speaker buttons embedded in JPDB page add-ons now play audio again, matching the same Jiten audio behavior already available inside lookup popovers.

[1.4.11] - 2026-06-19

Fixed

• Opening Settings on YouTube no longer makes the immersion filter rescan the page just because よむ mounted its own dialog, so the panel appears much faster on busy YouTube pages.

[1.4.10] - 2026-06-19

Fixed

• BookWalker (and any cross-origin "tainted" image) OCR now works again on iPad/Safari. The clean source image fetched for the canvas mirror was wrapped in a Blob with no MIME type, and WebKit refuses to decode an <img> from such a blob URL — so on iPad the rebuild produced nothing and no OCR overlay, highlights, or spinner ever appeared (Chrome content-sniffs and was unaffected). The fetched image now carries an image MIME type, detected from its magic bytes and falling back to the URL extension.

[1.4.9] - 2026-06-19

Fixed

• Clicking a hovered page word now pins the lookup popover immediately, so it stays open while the clicked lookup renders instead of being removed by the hover close timer.

[1.4.8] - 2026-06-19

Fixed

• Generic chat and comment prose now keeps furigana even when nearby avatars or thumbnails make the row look like compact media metadata, fixing missing ruby on Discord without adding a Discord-specific parser.
• Reader word underlines are thinner and sit closer to the text across normal page text, OCR, and hosted reader surfaces.

[1.4.7] - 2026-06-19

Fixed

• YouTube channel suggestions no longer show subscribe rows from preview metadata that lacks a current subscription state, preventing already-subscribed channels from reappearing as "Subscribe" recommendations.
• The channel suggestion shelf now has a single opt-out action, "Hide"; the route-only "Dismiss" button was removed.

[1.4.6] - 2026-06-18

Changed

• Jiten now comes before JPDB in fresh installs, migrated default installs, definition-source rows, Study combined source labels, settings copy, docs copy, and dictionary/source ordering surfaces. Custom source priorities are preserved.
• The homepage now focuses on よむ itself: lookup, local dictionaries, study cards, audio, OCR, PDFs, subtitles, and video, without promoting JPDB in the landing-page copy.
• The hosted Settings menu warms the よむ runtime on hover, focus, touch, and pointer down, then retries immediately when the runtime script finishes loading so opening settings feels snappier on mobile.
• The hosted video player now includes a localized Jimaku subtitle-search link next to the Subtitles button.

Fixed

• The subtitle overlay no longer latches onto page chrome posted next to a video — e.g. a Discord message author's handle that contains Japanese (such as "Canna波蘭") no longer gets stuck as the subtitle while scrolling past the clip. The generic page-caption scan now only accepts text that is centered on the player and overlaid within the video frame, instead of any nearby Japanese text.
• OCR no longer resurrects overlays for images that are removed before idle scanning runs, preventing stale page-turn scans from painting old or missing content.
• Hosted docs Japanese mode now localizes the Tools/Guides nav, theme-toggle labels, and home-page Next Steps cards, including the PDF reader card.
• Hosted light/dark switching now treats the stored よむ theme as authoritative and cleans up Greasemonkey storage listeners, avoiding flashes when settings or multiple tabs change theme.
• Keyless public JPDB search, vocabulary, and pitch lookups now persist successful results in a local browser cache, reducing repeat public/proxy requests for hosted visitors.

[1.4.5] - 2026-06-18

Changed

• PDF reader is now more native to よむ and far smoother to read:
  • No more competing OCR overlay. PDF.js exposes an accurate selectable text layer, so the runtime reads it natively (popups, furigana, mining) and no longer also runs image-OCR over the same page — the chaotic double-painted OCR boxes are gone. Manual OCR still works for genuinely scanned pages with no text layer.
  • Native page scrolling with a sticky toolbar. The document scrolls with the page (mouse wheel / trackpad / keyboard just work) instead of an inner scroll box, and the brand bar + page/zoom controls stay pinned at the top.
  • Sensible default zoom. Fit-width is capped to a readable column so a small page no longer balloons to 300%+ on wide monitors.
  • No black flash when zooming. Pages are re-rendered into an off-screen canvas and swapped in only when ready, so the visible page never blanks to black mid-zoom.

[1.4.4] - 2026-06-18

Added

• When both a JPDB and a Jiten API key are set, the lookup popover shows a small ⇄ toggle next to the Jiten/JPDB status. It switches which SRS the deck and grade buttons act on (persisted), so a word present in both services can be graded into either. The settings connection check now reports "Connected to Jiten and JPDB." when both keys are configured.
• The kanji card now shows Jiten and JPDB kanji facts side by side (when both are available) instead of only the active provider.

Changed

• Kanji-fact sources are labelled simply "JPDB" and "Jiten" instead of "Readings and components" / "Jiten kanji facts".
• Jiten words now use the same Add to deck / Never forget / Blacklist actions as JPDB; the separate Mining / Suspended / Forget row was removed.

Fixed

• Grading a word in the toggled-to service no longer reads the other service's deck state: the chosen provider's card state is refreshed on switch, and a JPDB word that also exists in Jiten is only treated as gradable in Jiten on an exact spelling+reading match.

[1.4.3] - 2026-06-18

Fixed

• PDF reader: large and scanned PDFs (e.g. 90–130 MB image-heavy workbooks) could crash the tab and render upside-down with the page stuck in a black corner. The renderer now bakes device-pixel scaling into the page viewport instead of a canvas transform (fixes the orientation/black-corner artifact), caps each canvas to a safe pixel budget and maximum dimension so it never exceeds the browser's canvas limit, cancels in-flight renders when you zoom/fit/resize (fixes the "Cannot use the same canvas during multiple render() operations" crash), evicts off-screen pages and limits render concurrency so memory stays bounded on 200+ page books, and renders the first page immediately instead of a batch. Opening, zooming, and scrolling large PDFs are now fast and stable.

[1.4.2] - 2026-06-18

Fixed

• The Jiten and JPDB definition panels are no longer shown on their own sites — on jiten.moe the redundant Jiten panel (definitions + "used in vocabulary") is suppressed, and on jpdb.io the JPDB panel is suppressed, since the native page already shows that source. Other added sources (your dictionaries, Immersion Kit, kanji practice, study tools) still appear, and both panels remain available everywhere else. The native-page addon no longer leaves an empty "No definitions" box when the self-site source was its only content.

[1.4.1] - 2026-06-18

Added

• Traffic-funnel content for SEO: a new /guides/ hub plus four how-to guides — read manga in Japanese (free OCR setup), mine sentences from anime & YouTube to Anki, comprehensible-input YouTube (with a levelled N5–N1 channel list drawn from the in-app starter guide), and Yomitan vs JPDB vs Anki — and a /compare/migaku-alternative page. Each targets a distinct search intent, ships matching FAQPage structured data, and links down to the relevant tool page and install. Added a Guides nav entry and sidebar group, plus "Related guide" cross-links from the OCR, subtitle, YouTube, and JPDB tool pages and a guides card on the home page.

[1.4.0] - 2026-06-18

Added

• Hosted PDF reader (/pdf-reader/): open or drop any PDF and read it with よむ — no userscript install required. Pages render fully client-side with PDF.js (Apache-2.0, vendored), drawing each page to a canvas for full fidelity (images, figures, multi-column layouts, CJK fonts via cMaps, and scanned-image codecs JBIG2/JPEG2000) with a selectable text layer over it that the よむ runtime scans for popup lookup, mining, and furigana. Image-only/scanned pages are flagged and fall through to よむ's OCR. Files are opened locally in the browser tab and never uploaded. Includes page navigation, zoom/fit-width, continuous lazy rendering for large documents, per-document reading-position memory, and theme/accent/interface-language synced with your よむ settings.
• New yomu-pdf-reader site-parser and /pdf-reader/ route so the runtime recognises and enhances the hosted reader's text layer, plus a PDF Reader link in the docs navbar, the in-app overflow menus, the Settings help links, and the docs home/support pages.

[1.3.30] - 2026-06-18

Changed

• SEO overhaul for the docs site and hosted apps. Every page now emits a correct per-page canonical URL and og:url (previously every page declared itself the home page, a duplicate-content signal and wrong social card), plus per-page titles, meta descriptions, and Open Graph / Twitter tags. Added an auto-generated sitemap.xml, a robots.txt, and JSON-LD structured data (SoftwareApplication + WebSite on the home page, FAQPage + BreadcrumbList on the tool pages). Internal planning/research docs (ADRs, backlogs) are now noindex and excluded from the sitemap.
• New /tools/ hub plus six keyword-targeted free-tool landing pages (Japanese OCR & manga reader, furigana reader, kanji stroke order, subtitle miner & video reader, JPDB study, YouTube immersion filter), each with its own FAQ schema and cross-links, so individual tools can rank on their own.
• The hosted new-tab study page and video player now carry descriptive titles, meta descriptions, canonical URLs, and Open Graph tags so those standalone tool URLs are indexable.

[1.3.29] - 2026-06-18

Fixed

• The YouTube subtitle transcript panel no longer overlaps the video when docked to the left on the single-column watch layout (narrow windows / iPad). YouTube hoists the player out of #primary into an absolutely-positioned full-bleed container in that layout, so shifting only the metadata column left the player covering the panel; the full-bleed player container now gets the same side inset and slides clear of the panel.

[1.3.28] - 2026-06-17

Fixed

• Furigana for inflected words now stays aligned to the kanji when the reader falls back to a dictionary-form card reading. For example, 読んで now shows よ over 読 instead of centering よむ over the full surface.

[1.3.27] - 2026-06-17

Fixed

• Selection token-choice popovers now show a compact translation by default and no longer show the redundant parsed-source line. Added an Appearance setting to turn selection-popover translations off.
• Mokuro catalog card titles are scan targets again, keeping line-clamped titles lookupable with ruby annotations.
• The OCR concurrency regression test now matches the current behavior restored in 1.3.21: iPad follows the configured OCR concurrency instead of forcing serial scanning.
• The visible-page scanner batching regression test now waits deterministically for continuation scans under slower CI runners.

[1.3.26] - 2026-06-17

Fixed

• Selection token-choice popovers now show a compact translation by default and no longer show the redundant parsed-source line. Added an Appearance setting to turn selection-popover translations off.
• Mokuro catalog card titles are scan targets again, keeping line-clamped titles lookupable with ruby annotations.
• The OCR concurrency regression test now matches the current behavior restored in 1.3.21: iPad follows the configured OCR concurrency instead of forcing serial scanning.

[1.3.25] - 2026-06-17

Fixed

• Selection token-choice popovers now show a compact translation by default and no longer show the redundant parsed-source line. Added an Appearance setting to turn selection-popover translations off.
• The OCR concurrency regression test now matches the current behavior restored in 1.3.21: iPad follows the configured OCR concurrency instead of forcing serial scanning.

[1.3.24] - 2026-06-17

Fixed

• Support-page cards now use relative links for the video player, study page, and docs home instead of hard-coded /yomu-reader/ production paths. This keeps the generated video-player link working across deploy contexts.

[1.3.23] - 2026-06-17

Fixed

• Selection token-choice popovers now show a compact translation by default and no longer show the redundant parsed-source line. Added an Appearance setting to turn selection-popover translations off.
• Closing a selection popover continues to keep the highlighted text without immediately reopening the same selection lookup.

[1.3.22] - 2026-06-17

Fixed

• BookWalker/iPad bottom-sheet lookups now keep their own vertical scroll gestures instead of letting the host viewer's page-level touch handlers swallow them. This restores scrolling inside long dictionary sheets while keeping the sticky sheet modeless.

[1.3.21] - 2026-06-17

Fixed

• Mokuro continuous/autoscroll readers no longer keep OCR frames and overlays for old offscreen CSS-background pages. Yomu now keeps only the active viewport-near raster surfaces, so long manga strips do not accumulate hundreds of hidden frame images while scrolling.
• Mokuro native .textBox scanning is now viewport-near and bounded, with visible boxes prioritized over old offscreen pages. This keeps Yomu responsive even when image OCR is disabled and mokuro has mounted hundreds of text boxes.
• Reverted the previous iPad-only OCR budget throttle: OCR concurrency, image-reader prefetch, and dark-panel second-pass behavior again follow the user's settings.

[1.3.20] - 2026-06-17

Fixed

• Greasy Fork companion @require URLs now include the Yomu package version as a query parameter before the SRI hash. This forces userscript managers to refetch the matching companion bundle when the main script updates, avoiding stale-cache SRI failures such as Firefox loading an older yomu-video.user.js against a newer yomu.user.js.

[1.3.19] - 2026-06-17

Fixed

• BookWalker Firefox/iPad mirror replay now fetches clean source images through the same GM_xmlhttpRequest arraybuffer path proven by the mirror probe, then wraps the bytes in a Blob locally. This avoids userscript-manager differences around responseType: "blob" that could leave the tainted canvas path with no rebuilt OCR frame and no spinner.

[1.3.18] - 2026-06-17

Fixed

• BookWalker Firefox/iPad clean-source mirror replay is now used when clicking or tapping a tainted viewer canvas: the manual pointer path waits for the async mirror snapshot before enqueuing OCR, so the hidden OCR frame can be created and the loading spinner can start instead of the click returning before any frame exists. Added focused coverage for clean-source mirror replay and manual tainted-canvas activation.

[1.3.17] - 2026-06-17

Fixed

• iPad OCR now uses a lighter automatic budget: serial OCR, no extra manga-page lookahead beyond the normal margin, and no automatic dark-panel second pass. This keeps mokuro/BookWalker OCR from monopolizing Safari's main thread and makes tap-to-open dictionary popovers responsive again.
• The puck radial menu spaces actions farther apart and places the audio toggle immediately after the power toggle, keeping the green controls together.
• Closing a text-selection lookup now keeps the page selection without immediately reopening the same selection popover.

[1.3.16] - 2026-06-17

Fixed

• Video, OCR, and other Greasy Fork companion libraries now publish their registry on both the userscript sandbox global and window when those objects differ. This keeps the main reader from falling back to “Video companion is missing; related features are disabled.” in userscript managers that split @require companion code and the main script across globals.

[1.3.15] - 2026-06-17

Fixed

• BookWalker OCR never started on Firefox (no spinner): the recorder installer called state() before injecting the page-world recorder, which created a sandbox-compartment state object on the page window; the injected page recorder then reused that object, so its recorded ops lived in the sandbox compartment and the main-world reader read an empty map. The Firefox/different-realm path no longer touches state() before injection — the injected page script creates the page-compartment state itself, so the reader can read the recorded descramble ops. iPad/Chrome (same realm) are unchanged.

[1.3.14] - 2026-06-17

Added

• The puck radial menu gains a mute toggle for auto-play term audio. The node swaps between a speaker and a muted-speaker icon (accent when on, grey when muted) and pauses/restores Auto-play term audio without opening settings; unmuting from a fully-off auto-play mode restores playback.

Fixed

• Hosted docs runtime now appends the settings/video companion scripts before its early-return, so the companions stay available when the core reader runtime already exists (e.g. an installed userscript). A refactor had moved the append after the early-return, dropping companions in that case.
• The new-tab study "No cards." empty state no longer splits its wordmark and message into opposite corners on coarse-pointer landscape tablets.

Verified

• Full typecheck/test/build suite (the hosted-companion regression test now passes), plus a live browser check of the puck audio toggle on the docs runtime.

[1.3.13] - 2026-06-17

Fixed

• BookWalker OCR still didn't fire on Firefox (it works on iPad): the canvas-mirror recorder ran in the Tampermonkey content sandbox while the OCR reader runs in the page main world, and Firefox won't let a sandbox-created state object be defined on / read from the page window (Xray compartments). The recorder is now injected into the page main world when the realms differ (via a Trusted-Types-safe <script>), so its state and recorded ops are page-compartment objects the reader can read directly; the same-realm direct patch is kept for iPad/Chrome. Fixed an arity bug in the injected recorder (it reads arguments, which includes the source image, so the 9/5/3-arg drawImage forms map coordinates from index 1).

[1.3.12] - 2026-06-17

Fixed

• Hosted docs cold starts no longer wait several seconds on optional public Jiten enrichment before pitch/furigana classes can land. Profiling the live homepage showed the Try Me sample hitting public Jiten vocabulary/parse through the proxy and receiving upstream 5xx responses; that route now uses a short 1.5s background timeout, shared parse backoff, and abort-aware transient detection so fallback rendering and JPDB pitch enrichment can continue promptly.
• The hosted docs loader now preloads the core reader script and lets normal docs load it before optional settings/video companions, reducing contention before the first visible annotations.
• The hosted Try Me samples now opt into visible furigana for demo text, so the landing page shows the reading aid clearly even though the default reader setting still hides ruby on easier kanji elsewhere.

[1.3.11] - 2026-06-17

Fixed

• BookWalker two-page spreads only OCR'd the right page on Firefox/iPad. A spread is composited from one reused buffer (render right → composite right → render left → composite left), so the rebuild visits that buffer twice at different sequence bounds. The source-image URL collector used a shared "seen" set and skipped the second visit, so the left page's image was never fetched and rendered blank. The collector now uses per-path "seen" copies (matching the rebuild), so both pages' source images are fetched and both halves OCR.

[1.3.10] - 2026-06-17

Fixed

• BookWalker OCR still didn't fire on Firefox/iPad after 1.3.8: the canvas-mirror recorder is installed from the Tampermonkey content sandbox while the OCR reader runs in the page main world — different JS realms with different globalThis, so the "shared" recorder state wasn't actually shared (recorded: false). The recorder state is now anchored on the page window (unsafeWindow/window, the same object from both realms) and records are keyed by a DOM-attribute id (the DOM is shared across realms) instead of per-realm object references. captureCanvasMirror also lazily installs the recorder as a safety net. A diagnostic probe confirmed the sandbox/main-world hooks all capture the viewer's descramble ops and the rebuild is readable on Firefox.

[1.3.9] - 2026-06-17

Added

• The settings puck now opens a radial menu of context actions instead of jumping straight to settings. It blooms out of the puck with the kanji-map visual language — accent-haloed circular nodes, soft elevation, a springy staggered reveal, and a faint page scrim — and always fans into the open screen quadrant so a corner puck never throws nodes off-screen. Actions: a power toggle that pauses/resumes all annotations (the puck colour goes grey when paused and the page reads natively until resumed), open settings, scan the page, open the study page, and — only on YouTube — toggle the immersion filter. Each node shows its label as a pill above it on hover/focus.
• A new "Manual scan only" option (Settings → Reader, and the welcome page) stops automatic scanning so a page is read only when you tap the puck's Scan action or press the scan shortcut.

Verified

• Full typecheck/test/build suite, plus a live browser check on the docs runtime: open, pause (annotations stripped, puck greyed, scan disabled), resume (page re-scanned), and the tooltip-above-node layout.

[1.3.8] - 2026-06-17

Fixed

• BookWalker OCR on Firefox/iPad (1.3.7's canvas mirror) did nothing because the document-start recorder ships in the main userscript bundle while the capture path ships in the OCR companion bundle — separate module instances, so the reader queried an empty op map. The recorder state now lives on a shared global, so the companion sees the recorded descramble ops. Added an opt-in diagnostic (localStorage['yomu.canvasMirrorDebug'] = '1') that logs [Yomu][canvas-mirror] capture metrics.
• Favicons (apple-touch-icon, favicon-32x32, favicon-16x16) are now rasterized from the canonical yomu-icon.svg via a new scripts/generate-favicons.mjs, replacing stale art that no longer matched the logo.

Verified

• canvas-mirror unit tests plus the full typecheck/test/build/verify suite.

[1.3.7] - 2026-06-17

Fixed

• BookWalker OCR now works in userscript managers on Firefox and iPad/Safari, where the DRM page canvas is cross-origin "tainted" and cannot be read. A document-start recorder captures the viewer's own descramble drawImage tile-copies, then rebuilds the current page onto a fresh, untainted canvas by replaying those copies from an origin-clean image fetched via GM_xmlhttpRequest. No descramble keys or crypto are reimplemented — it mirrors whatever the engine actually draws, so it survives viewer updates and reconstructs two-page spreads correctly. The path is gated to BookWalker hosts and only runs when the canvas is tainted, so Chrome and the Yomu extension are unaffected.

Verified

• New canvas-mirror unit tests (op selection, recursive buffer reconstruction, recorder hook) plus the full typecheck/test/build/verify suite.

[1.3.6] - 2026-06-17

Fixed

• Auto-translated subtitle tracks now update primary and secondary cue state in the same tick, and the primary line stays visible while its aligned translation is still active.
• The generated extension build keeps the BookWalker visible-tab screenshot bridge available while avoiding an unused exported helper in the userscript bundle.

Verified

• Focused subtitle-controller, OCR raster, extension hardening, and typecheck gates.

[1.3.5] - 2026-06-17

Fixed

• BookWalker Firefox/iPad extension OCR now captures protected rendered canvases through a visible-tab screenshot bridge instead of falling back to fetched NFBR source images, because those source assets can be scrambled before the viewer composites the real page. Readable rendered canvases still wait for stable page pixels before OCR, avoiding stale transition captures.
• Compact media tiles outside YouTube now suppress furigana only through generic layout/media heuristics, not custom site parsers or URL/class-name allowlists, so clipped grids keep pitch colouring and dictionary lookup without ruby breaking the card height.
• Yomu floating controls, settings tabs, and settings buttons now keep their own colours and layout against aggressive host-page CSS.

Verified

• Focused canvas-reader, OCR raster, extension hardening, settings CSS, and compact media/YouTube target tests.

[1.3.4] - 2026-06-17

Fixed

• BookWalker/iPad OCR: tainted canvas readers now fall back to the GM-fetchable source page image when Firefox/WebKit cannot read canvas pixels, page turns release stale OCR frames, and storefront pages opt out of generic DOM scanning so Yomu no longer breaks the carousel/navigation layout.
• OCR interactions: tapping an OCR word on iPad opens the lookup immediately, OCR hover popovers no longer flicker when the pointer crosses non-Japanese text inside the same OCR line, and fallback parsing keeps single-kanji godan-s verbs such as 騙した in dictionary form with pitch/furigana.
• YouTube stability: non-destructive mirrors wait until replacement text is ready before hiding native text, reuse stable attributed-string hosts, restore hidden titles if YouTube rerenders the host, and reduce no-key parsing contention on live/comment-heavy pages.
• Settings dialog Japanese: passive parsed settings copy can open dictionary popovers without stealing or clearing the user's selection, command/tab labels still activate the settings UI, and the settings-dialog tests now mock the slow form/deck seams correctly.
• Site scanning: Bloomee landing-page headings such as 季節のお花を、かんたんに飾れる are now eligible for furigana/pitch annotation, while image-only copy still requires OCR.

Added

• Release probes/tests for BookWalker live canvas diagnostics, source-image fallback OCR, BookWalker storefront scan suppression, repeated mirror rescans, YouTube no-key/keyed parsing behavior, Bloomee styled headings, and settings passive lookup/selection preservation.

[1.3.3] - 2026-06-17

Added

• Welcome/onboarding panel: its Japanese (eyebrow, intro copy, the five feature descriptions, option labels, and action buttons) is now annotated with furigana and pitch-accent colouring through the same nested-parse path used for popovers and the settings dialog. The panel is a reader root (excluded from page scanning), so it now opts in explicitly via jpdb-reader-parseable and re-annotates after a settings-language switch. Only runs when the interface language resolves to Japanese; English copy is left untouched.

[1.3.2] - 2026-06-17

Fixed

• BookWalker (NetFront/NFBR viewer): the on-screen page-buffer marker (.currentScreen) is now matched on the page’s own #viewport container rather than any ancestor, so a .currentScreen that ever lands on a shared ancestor such as #renderer can no longer select both double-buffered pages at once (a duplicate, shared-quota Google Lens call plus a stale overlay for the off-screen page). Composes with the rendered-content preference and still falls back to all page canvases before any buffer is marked current (e.g. the cover).
• YouTube kana-only titles and metadata now keep contiguous hiragana words together during fallback parsing, avoiding over-segmentation such as にほんご becoming separate mini-tokens while preserving real particle and inflection boundaries.
• Hover popovers now re-anchor to the equivalent refreshed word node when YouTube replaces a mirrored title/metadata node under a stationary cursor, preventing reactive page updates from auto-dismissing the popup.

Verified

• Live injected-userscript probe on the real viewer.bookwalker.jp viewer (opened via the free read flow) and the trial viewer, single page and landscape double-page spread: one canvas captured per visible screen, OCR text + clickable words render, the dictionary popover opens on click, and the per-page OCR loading spinner shows during the scan.

[1.3.1] - 2026-06-17

Fixed

• Hosted-page image OCR cache hits now render immediately after refresh and pending OCR cache writes flush on pagehide/visibility changes, so stable docs screenshots do not wait for another recognizer pass before their overlays come back.
• OCR cache persistence now ignores volatile data: and blob: frame keys, keeping refresh-ready cached OCR focused on stable page images.
• BookWalker canvas OCR now falls back from a blank .currentScreen buffer to a painted sibling page buffer, recovering OCR on NFBR viewer pages where the visible buffer marker changes before the page is drawn.
• OCR furigana now anchors to each normalized base span instead of the wider word/line wrapper, so readings such as いばしょ stay over the specific kanji/base text they annotate.
• YouTube home no longer shows the curated channel subscription shelf, and failed/unknown channel preview probes no longer count as unsubscribed channels that can flash the shelf back into view.
• YouTube channel shelf refreshes are now structurally stable: preview hydration updates a row in place without rebuilding the action bar or replacing the full list.
• Compact YouTube feed/search/suggested titles keep lookup/color styling without furigana, preventing title clipping and the add/remove furigana churn caused by YouTube's frequent title rerenders.

[1.3.0] - 2026-06-17

Fixed

• YouTube attributed-string UI text no longer collapses mirrored furigana/underline rendering into narrow vertical stacks after YouTube re-renders feed titles, watch metadata, buttons, or comments. Inline hosts are promoted to a stable inline-block mirror container, then restored when the mirror is removed.
• Mokuro native OCR toggle changes are now handled at runtime: when Mokuro OCR/text boxes turn on, Yomu clears auto-painted image/canvas/background OCR overlays and suppresses cached or in-flight auto OCR results from repainting stale layers; when Mokuro OCR turns off, Yomu auto OCR resumes.
• Mokuro canvas/background reader frames now preserve the difference between automatic scans and user-requested scans, so automatic raster OCR is cleared when the native text layer takes over while manual scans remain visible.
• OCR async cleanup now quietly ignores stale scan state after toggle/destroy timing gaps instead of surfacing stale-state errors during cleanup.
• Image OCR highlight panels are a little more transparent by default, letting the page artwork show through more while keeping the active line readable.
• Jiten kanji facts now load in kanji dictionary popovers even when a JPDB key is configured alongside the Jiten key, so the Jiten frequency/readings/vocabulary panel appears on kanji drilldowns.
• Release/CI stability: the generated probe scripts are registered as fallow entry points, and the previously failing OCR stale-state test shard is covered by focused regressions.

Verified

• Re-checked the bundled release against live signed-in YouTube home/search/watch/comment surfaces, ComicWalker canvas OCR, blank-canvas controls, status-dot/subtitle behavior, vertical OCR/pitch fixtures, and mobile docs hero breakpoints.

[1.2.3] - 2026-06-17

Fixed

• OCR interaction unit tests: mocked getBoundingClientRect on word spans in hover-lookup.test.ts to cover the hover/click coordinates (40, 24) and satisfy the new strict hit-testing.
• Test state leakage: added localStorage.clear() to the global beforeEach in jpdb.test.ts to prevent cached OCR data from bleeding between tests.

[1.2.2] - 2026-06-17

Fixed

• OCR hover interaction: restricted OCR line hover/click popovers to strictly trigger when pointer is over the actual text bounding box, rather than anywhere inside the line frame wrapper.
• Dark mode background parity: adjusted the light-mode OCR active background to be semi-transparent, aligning with the dark-mode layout.
• Removed unwanted text-shadow/glow styling from furigana elements inside OCR layers.

[1.2.1] - 2026-06-17

Fixed

• Landing page hero: shortened the Japanese action-button labels (よむをインストール → インストール, 学習アプリを開く → 学習) so the pills no longer wrap onto extra rows on mobile, and added breathing room between the よむ wordmark and the headline so the wordmark's underline (and its tap/hover target) no longer crowds the first line of the headline.
• Reverted the forced line break in the Japanese hero headline. Inserting a hard newline switched the headline to white-space: pre-wrap, which broke the in-place furigana rendering (readings went missing / misaligned until the line was re-selected). The headline now wraps with the reader's default behaviour so furigana renders consistently with the rest of the page.

[1.2.0] - 2026-06-17

Added

• Light-on-dark manga text is now read on mixed pages. A page often mixes black-on-white speech bubbles with white-on-black caption boxes/panels; inverting the whole page would just swap which half fails. Instead, when a page has a dark region that the normal recognition pass left unread, a second inverted pass runs and only its lines that fall over genuinely dark areas are merged in. So a dark caption box (e.g. 「嘘だろ……！！」) is recognized alongside the normal bubbles, with no extra request or latency on ordinary pages or dark panels the recognizer already read, and no spurious lines on light pages. Toggle: "Read light text on dark panels".
• Persistent OCR cache: recognized text for stable-src reader/article images is mirrored to localStorage (bounded by entry count and total size) so a page refresh re-renders recognized text instantly instead of re-running every recognizer request. Volatile data: frames (paused-video / canvas snapshots) are excluded.
• Image-based manga readers (Mokuro, MangaDex, etc.) now prefetch a sliding window of upcoming pages and raise the per-page image budget on reader-like pages — the same look-ahead canvas readers already get — so the next pages are recognized in the background before you reach them. Ordinary pages, where auto-OCR stays near the viewport, are unaffected.

Changed

• Google Lens recognition falls through to the cookie-authenticated lens.google.com upload endpoint (per-user quota, sent with the Lens web client's Origin/Referer) when the shared keyless endpoint is throttled or returns nothing — reducing rate-limiting and recovering more text. The extra request only fires when the primary endpoint comes back empty, so request volume on pages that already read is unchanged.
• Furigana now uses a single tight outline instead of the word contrast glow, which was bleeding into the thin reading strokes and hurting legibility on busy manga backgrounds (tunable via --jpdb-reader-furigana-shadow).

Fixed

• The YouTube channel shelf now clears its pending refresh when every channel is confirmed subscribed, so it doesn't keep re-testing/re-rendering after the shelf is removed.

[1.1.2] - 2026-06-17

Fixed

• Landing page hero now reads correctly on phones and tablets. The よむ wordmark is left-aligned (and sized as a label rather than the oversized display text it used to inherit on every breakpoint); the app-icon logo stays centered on phones; the headline and the action buttons hug the left edge instead of being centered/stacked as full-width blocks. On tablet widths the layout switches to the desktop-style two-column form (copy on the left, logo on the right) instead of stacking the logo on top with empty space beside it.
• The Japanese hero headline now breaks at the clause boundary — 好きなものを読んで / 日本語を学ぶ — instead of splitting a word (e.g. 読んで) across the line break. Keeping the word whole also restores its pitch-accent underline, which a mid-word wrap had been hiding.

[1.1.1] - 2026-06-17

Fixed

• Keyless parser no longer glues numeric episode/volume counters onto the following title words: a run like 1〜5話おまとめ版 now keeps 話 as its own counter token instead of merging into 話おまとめ / 話おまとめ版. The split is driven by the number-then-counter boundary, not a per-title list.
• Names such as 紫音 are no longer split into kanji component readings (紫=むらさき + 音=おと) by a hand-coded reading table. Readings now come only from a dictionary the user has loaded: when a name dictionary (e.g. JMnedict) supplies 紫音, the whole compound is kept with the dictionary's verified reading; when no dictionary knows the name, the parser faithfully reflects the single-kanji lookups instead of inventing a reading. (A single baked-in reading is wrong as often as right — 紫音 alone reads しおん / しいん / しのん / むらさき depending on the person.) Regression tests pin both behaviours.

[1.1.0] - 2026-06-17

Added

• Image and manga OCR now runs in parallel. A small concurrency pool (new "Parallel OCR requests" setting, default 3) replaces the old strictly one-at-a-time queue, so a reader page full of page images or canvases is read far faster. Work is deduplicated by image content, so a re-snapshotted canvas frame (canvas readers poll for page turns) never fires a redundant OCR call — the cache fills the duplicate in once the first scan resolves.
• Canvas and CSS-background manga readers (BookWalker, comic-walker / カドコミ, Mokuro, and other viewers) prefetch a sliding window of upcoming pages. A new "Canvas prefetch pages" setting (default 2) snapshots and OCRs the next few spreads in the background — extending the look-ahead to ocrPrefetchPages viewport-heights — so a page's text is ready by the time you scroll to it. Combined with parallel OCR, several spreads are read at once.
• New "Read light text on dark panels" setting (default on): dark panels (white text on a black background) are inverted once before OCR so polarity-sensitive recognizers (a local MangaOCR/PaddleOCR/Apple Vision server, etc.) can read them. The check is conservative — only dark-dominant panels with a minority of bright pixels are inverted, so normal pages are never touched and OCR is never run twice. It is a no-op for Google Lens, which already reads light-on-dark text.

Verified

• Live on comic-walker.com (カドコミ): canvas-reader detection snapshots only the rendered page canvases (skipping the viewer's not-yet-painted blank placeholders), the sliding-window prefetch captures upcoming spreads while scrolling, and the parallel pool reads them concurrently — 70+ text lines recognized across several pages of a chapter.

Fixed

• YouTube channel recommendations now cancel queued shelf refreshes when the shelf is removed and pause background preview probes during subscribe-all writes, preventing all-subscribed or dismissed shelves from briefly reappearing while YouTube subscription checks are still settling.
• YouTube channel shelf tests now wait for stable row/subscription states, keeping the release gate deterministic under GitHub's Node 24 runners.

[1.0.5] - 2026-06-17

Fixed

• JPDB review and search page enhancements now keep Immersion Kit examples stable, position example subtitles over the image correctly, show kanji doodle practice on unrevealed kanji fronts with the trace hidden by default, remember the native examples toggle state, and preserve native JPDB audio links.
• JPDB passive link-style words now open Yomu hover dictionaries reliably, while Jiten and JPDB study prompts suppress Yomu-added furigana for the word being quizzed.
• JPDB word-card backs now render the full local/Jiten/JPDB definition information, matching kanji cards and Jiten word pages.

[1.0.4] - 2026-06-17

Fixed

• YouTube channel recommendations now hide cleanly when the signed-in account is already subscribed to every curated channel, and the shelf no longer shows stale Subscribe rows or the "Previews load from YouTube on this page." footer copy.
• YouTube text mirrors no longer let ruby-room sizing write height onto YouTube description expanders, metadata rows, or action buttons, preventing watch descriptions from growing, metadata from stacking vertically, and action labels/channel metadata from flickering or clipping.

[1.0.3] - 2026-06-17

Fixed

• Jiten definition headwords no longer render the bracketed furigana annotation (e.g. 以[い]前[ぜん]) as base text under the reading. The annotated mainReading is now stripped to clean base text and distributed as per-kanji ruby, matching the related-words rendering.

[1.0.2] - 2026-06-17

Fixed

• Image OCR text now reveals its pitch-accent underline and furigana on hover, not only when focused or pinned — matching how normal reader words behave on hover.
• Tapping an OCR text line on touch/pointer devices reliably activates its markup (pitch and furigana) before the dictionary popup's lookup handlers run, and a short post-tap guard keeps the follow-up click from immediately toggling the line back off. Previously the first tap could open the popup without the line ever showing its prepared pitch/furigana.

[1.0.1] - 2026-06-17

Fixed

• Audio-source rows in settings now keep the source dropdown and voice/URL box top-aligned, including Japanese option helper text.

[1.0.0] - 2026-06-16

Added

• Image OCR now treats manga reader raster pages as first-class scan surfaces: BookWalker/ComicWalker-style <canvas> pages and Mokuro CSS background-image pages are converted into invisible OCR anchors and aligned with the existing overlay pipeline.

Fixed

• Mokuro pages now trigger Yomu OCR when Mokuro's own OCR layer is turned off, including installed-library reader pages that render each page as a CSS background image instead of an <img> or <canvas>.
• BookWalker and ComicWalker OCR now wakes up reliably when canvases mount after startup, waits until a canvas has real page pixels before scanning, and captures every visible spread canvas instead of only the first readable side.
• Hosted Yomu docs no longer annotate the oversized hero brand text, avoiding the huge/shifted ruby and pitch overlay on the main logo.

[0.7.90] - 2026-06-16

Changed

• Image OCR now uses the compact spinner indicator only: no expanded scanning card and no dismiss control. The same compact indicator covers normal images and paused-video frames.
• OCR image/video loading copy is consistently "Scanning..." in accessible labels.

Fixed

• OCR text keeps pitch and furigana prepared as soon as it renders, so tapping an image text segment reveals the furigana and pitch underline instantly; clicking away hides both again without rebuilding the line.
• Brand/logo/icon images, including the hosted Yomu logo, are skipped by image OCR, and images that return no OCR text are remembered so repeat hovers do not rescan them.

[0.7.89] - 2026-06-16

Added

• Settings now sync between the hosted reader (hrussellzfac023.github.io/yomu-reader/newtab/) and the userscript. The hosted page runs in the page's main world, which has no GM_* APIs — those live only in the userscript's content world — so settings saved on the hosted page previously stayed in that origin's localStorage and never reached the shared GM storage the userscript reads on jpdb.io and everywhere else. A new GM storage event-bridge (mirroring the existing HTTP bridge) lets the hosted page route gmStorageGet/set/delete/listValues to the userscript's GM storage, scoped to Yomu-owned keys. Edits on the hosted page now reach the userscript, and if the userscript is installed after settings were already saved on the hosted page, the existing localStorage→GM migration seeds them into GM on the next hosted load.

Fixed

• Audio sources (Jisho, JapanesePod101, jiten, etc.) now play on the hosted reader with the userscript installed. The reader was routing cross-origin audio fetches through the userscript HTTP bridge even on the hosted page, but that bridge serialises responses as JSON and cannot carry binary audio Blobs across the content/page world boundary — the Blob arrived empty, playback failed, and the error was swallowed (No playable audio found with an empty errors list). On the hosted (newtab) runtime the reader now prefers a direct fetch through the public worker proxy, which serves that same media with CORS headers and returns a real Blob; the bridge remains a fallback. Off the hosted runtime (a real userscript on jpdb.io etc.) the bridge is still preferred, since its GM_xmlhttpRequest is exempt from the page's CSP.
• The audio settings preview no longer shows a "Playing よむ…" toast when nothing audible actually plays. The toast was shown optimistically before playback resolved, and play()'s boolean result was ignored — so a silent miss (no source resolved and the chime fallback disabled) still claimed playback. The toast now appears only when playback succeeds; otherwise an "Audio preview failed" toast is shown.
• No playable audio found now reports the underlying error(s) instead of an empty list. Candidate-preparation failures (CORS, non-audio responses, blob fetch failures) were being swallowed in two code paths, so genuine audio failures looked like "nothing found" with no diagnosis. Those errors are now collected and logged.

[0.7.88] - 2026-06-16

Fixed

• YouTube page furigana no longer duplicates text or thrashes (a word repeatedly "flashing" while it re-rendered), which had also degraded watch-page performance. The non-destructive furigana text-mirror was being re-ingested by the passive-interaction scan path (used for every site-profile root, including YouTube): a rescan of a mirror host re-collected the mirror's bare gap text nodes — punctuation/ASCII such as （Googleによる翻訳） — alongside the still-present (hidden) original host text, doubling the collected text (e.g. 原文を見る（Googleによる翻訳） rendered twice) and self-perpetuating into a rebuild loop on YouTube's constantly-mutating DOM (caption/translation strip, chrome). PASSIVE_AWARE_FRAGMENT_SKIP_SELECTOR now skips .jpdb-reader-text-mirror (matching the base skip list), so the reader's own mirror is never re-scanned — the chrome is still annotated, just not doubled. The mirror host's visibility:hidden/position:relative are also re-asserted when YouTube re-renders and strips our inline styles, preventing the original text from reappearing beside the mirror (duplication) or the mirror anchoring to the wrong ancestor (misalignment).

[0.7.87] - 2026-06-16

Changed

• On mokuro readers, mokuro's own "OCR enabled" overlay now defaults off (once, before mokuro loads its settings) so the reader's sharper OCR runs on the page by default; turning mokuro's toggle back on is respected and the reader defers to mokuro's text boxes again. The "OCR enabled" toggle label is annotated to make clear it now switches between Yomu OCR (off) and mokuro OCR (on).

[0.7.86] - 2026-06-16

Changed

• Canvas-based manga readers are now detected generically, not just on a host allowlist. Any reader that paints page images onto a <canvas> qualifies when the canvas is large and page-shaped, prominent in the viewport, and carries a rendered raster image (the decoded page) — verified live on Shonen Jump+ (shonenjumpplus.com) in addition to BookWalker and ComicWalker. The rendered-image test keeps the path off WebGL games, charts, vector/UI canvases and blank buffers, and skips cross-origin-tainted canvases that can't be read; known reader hosts and reader page-counters keep a lenient size/shape-only fast-path.

[0.7.85] - 2026-06-16

Changed

• On mokuro readers the image OCR now follows mokuro's own "OCR enabled" (displayOCR) setting: when mokuro OCR is off, the reader runs its own sharper, more touch-friendly image OCR on the page instead of deferring to mokuro's built-in engine; when it is on, the reader keeps using mokuro's native text boxes.
• The OCR loading/ready status card (with the dismiss-to-compact spinner) now appears on every OCR'd image, not just paused video frames, so slower image OCR shows progress.

Fixed

• The "title flips between plain and annotated" loop on reconciling single-page apps (e.g. the mokuro.moe catalog): a host whose annotation is repeatedly reverted now switches to the non-destructive text mirror, which overlays the annotation without mutating the app's own node, ending the loop for any such site.

[0.7.84] - 2026-06-16

Fixed

• The word-card headword now shows the pitch-accent underline (matching words on the page) when a pitch pattern is known. The headword was rendered as .jpdb-reader-spelling without a pitch class, so it only showed the pitch graph and never the colored underline. It now carries its jpdb-pitch-* class and reuses the same per-pattern pitch colors (e.g. odaka teal, heiban blue); words without a known pitch are left undecorated.

[0.7.83] - 2026-06-16

Fixed

• Jiten related-words / used-in lists now distribute furigana per kanji using the annotated reading (e.g. 読み取る renders 読(よ)み取(と)る) instead of placing the whole reading (よみとる) over the whole word. Okurigana stays as plain base text and is no longer shown as if it were a kanji reading.

[0.7.82] - 2026-06-16

Added

• YouTube channel shelf now records when every curated channel is subscribed (or can no longer be resolved) and stops re-testing subscription status on each render. The flag is keyed by a signature of the channel list, so editing the curated list automatically re-tests against the new set. Channels that were deleted, moved, or renamed are treated as unresolvable and never block the all-subscribed state or raise errors.

[0.7.81] - 2026-06-16

Fixed

• A Jiten API key (starts with ak_) pasted into the JPDB API key box now moves to the Jiten API key box once the field is committed, instead of staying displayed in the JPDB box. The key was already routed to the Jiten credential internally; now the visible inputs match where each key belongs (a genuine JPDB key stays put).

[0.7.80] - 2026-06-16

Fixed

• Jisho.org audio now works on the hosted reader. Without a userscript, jisho.org cannot be fetched directly (CORS) and the public proxy fails its TLS handshake with jisho (525), so the fallback scraped the text proxy's markdown rendering — which drops the audio <source>, leaving jisho silent. The reader now requests the raw jisho HTML from the text proxy and parses the <audio id="audio_{word}:{reading}"><source> element with the same logic as the userscript path and yomitan, so the pronunciation resolves and plays.
• Guarded toast cleanup against a torn-down document so a toast's exit timer can no longer throw an unhandled "document is not defined".

[0.7.79] - 2026-06-16

Changed

• Moved the text-selection popover's action pills (Jiten/JPDB/Jisho/Copy/Yomu) to the top, directly under the title, so they are reachable without scrolling past a long parsed token list.
• Compacted the Japanese docs hero: the in-place parsed tagline now uses a smaller display size with balanced wrapping, so furigana no longer pushes it into a tall, sparse column (less wasted space).

Fixed

• Selection popover word buttons no longer fall through to the page word underneath the popover. Because the popover overlaps page text, clicking a parsed token could resolve the underlying word by point geometry and open its dictionary at the wrong location (shifting the popover); document click handling now ignores clicks on the popover's own controls.
• Text selection now opens the lookup popover consistently once the selection settles, including iPad selection-handle and loupe adjustments that do not emit a fresh mouseup/touchend — a debounced selectionchange trigger covers them.

[0.7.78] - 2026-06-16

Fixed

• Showed the subtitle control rail on tall portrait video players (e.g. iPad reels/shorts-style pages) that fill most of the viewport height. The player-frame resolver previously treated any viewport-sized wrapper as a page container, so on portrait players the frame collapsed to the bare <video>, player affordances were not detected, and the rail was hidden (display:none); landscape players were unaffected. Tight wrappers that hug the video now resolve as the player frame, while oversized page containers that leave room for other content are still ignored.
• Isolated keyless YouTube subtitle pitch enrichment from the shared visible-page lookup budget, bounding public fallback pitch lookups through Jiten's batched lookupMany so long transcripts do not starve visible page text.
• Rebuilt visible keyless YouTube transcript rows after cheap provisional parsing, and kept provisional subtitle cache entries out of session storage until they have been enriched with furigana/pitch.
• Tightened Jiten public batch parsing so ambiguous short terms are separated during unauthenticated lookup, preventing missing characters in Jiten-derived words.
• Split leading particles from Segmenter particle+noun compounds such as 日本語の森, preserving the expected word boundaries for hover and pitch/furigana enrichment.

[0.7.77] - 2026-06-16

Added

• Automatic OCR for canvas-based manga readers (notably the BookWalker browser viewer at viewer.bookwalker.jp / viewer-trial.bookwalker.jp): each page <canvas> is snapshotted to a pointer-transparent, invisible overlay and read by the existing OCR pipeline, re-snapshotting automatically on page turns. The snapshot lets the host's own page-turn taps/swipes pass straight through.

Fixed

• Remove duplicate clearTranscriptVirtualRender method implementation to resolve typescript compilation errors.
• Adapted Playwright QA integration test assertions and locators to support ruby/furigana text nodes in the DOM.
• Prevented userscript network globals leakage by resetting GM_xmlhttpRequest and GM in Vitest beforeEach setup.
• Mokuro readers (reader.mokuro.app, opened from mokuro.moe) now read mokuro's accurate native text boxes instead of also re-running image OCR (Google Lens) on the same artwork. The redundant pass dropped characters the native layer already had (e.g. 事) and painted a competing overlay over the page. Manual scanning is still available for panels mokuro itself missed.

Changed

• Clearer OCR provider settings so setup is obvious: "Google Lens — free, no setup (recommended)", "Google Cloud Vision — needs API key", and "Local OCR server — advanced", each with provider-specific help; the local engine list labels MangaOCR (best for manga) and Apple Vision (macOS).
• Larger tap targets for manga words on touch devices, making lookups easier to hit without a stylus.

[0.7.76] - 2026-06-16

Fixed

• Fixed YouTube DOM instability on real watch pages where decorated title, channel, viewer-count, and live-chat text could disappear, reappear, or duplicate after YouTube/Polymer rerendered managed text nodes.
• YouTube watch metadata now uses DOM-safe mirrors for ruby and pitch, avoids hidden rolling-number text in watch-info rows, prevents nested enhancement wrappers, and covers current live-chat teaser and /live_chat iframe surfaces.
• Added real-YouTube smoke coverage for the target watch page, a normal watch page, the home feed, and a live-chat watch page, including sustained rerender/scroll/hover checks, duplicate-wrapper checks, missing-parsing checks, and idle reparse sanity.

[0.7.75] - 2026-06-16

Fixed

• Kept YouTube transcript resize/orientation responsive by aborting transcript row hydration and background parse warmup during drag, then resuming it after the panel settles; this removes the parse/DOM work that could fight the resize loop on iPad.
• Added a Playwright resize profiler for iPad landscape/portrait, iPhone, and desktop that measures open, drag, page scroll while the panel is open, and orientation changes for right/left/bottom transcript placements.
• Revalidated iPad YouTube left/right/bottom placement with screenshot-backed DOM evidence: left and right shift the video plus title/actions/description beside the panel, while bottom overlays without resizing the player.

[0.7.73] - 2026-06-16

Fixed

• Initial YouTube DOM-safe text decoration release for watch/home surfaces; superseded by 0.7.76 after real-site verification found YouTube watch-info rolling-number and live-chat iframe coverage needed the additional hardening above.

[0.7.72] - 2026-06-16

Fixed

• YouTube bottom transcript mode no longer mutates YouTube player dimensions or dispatches resize nudges while opening or resizing the drawer, removing the expensive reflow path that made panel open/resize feel slow and keeping the video frame stable as the drawer moves.
• iPad/tablet YouTube left/right transcript placements now stay side-docked when explicitly selected, including the m.youtube.com tablet DOM, shifting the player and watch metadata beside the panel instead of falling back to bottom or covering the video.
• YouTube bottom mode now caps its default drawer height to the space below the video when possible, stays flush with the visual viewport, and keeps title/metadata/action/description widths normal; the Playwright matrix now proves iPad/mobile/desktop bottom resizing does not call YouTube setSize or touch player sizing.

[0.7.71] - 2026-06-16

Fixed

• Repaired the 0.7.70 release lane CI failures by aligning YouTube safe-DOM expectations with the scanner behavior, rerunning the previously failing shards, and restoring the missing smoke script entries used by CI.
• Cleared the remaining dead-code finding in the kanji study companion registry without suppressions.
• Tightened Jiten public vocabulary cache invalidation when runtime stores are reset or rescanned, and restored OCR word hover lookup coverage for active image text.

[0.7.70] - 2026-06-16

Fixed

• Integrated the remaining P0 worker fixes into one release lane with proof-backed coverage for YouTube text enhancement, fullscreen subtitle rails, iPad sidebar layout, settings self-enhancement/alignment, Jiten/JPDB dictionary coexistence, keyless Jiten definitions, popover action pills, study cards, and batched fallback enrichment.
• YouTube text enhancement now skips only internal ripple/icon chrome instead of whole renderer subtrees, restoring ruby and pitch on titles, chips, guide labels, comments, transcripts, watch chrome, and video metadata while keeping framework internals untouched.
• Added and strengthened Playwright smoke coverage for hosted settings without userscript injection, desktop/tablet/mobile settings layout, Jiten keyless definitions, Jiten+JPDB source matrices, popover pill navigation, YouTube fullscreen/subtitle rails, YouTube sidebar layout, YouTube ruby coverage, popover/new-tab audio, and concurrent enrichment.

[0.7.69] - 2026-06-16

Fixed

• Rebased and republished the iPad/tablet YouTube transcript sidebar performance and layout fixes on top of the latest dictionary-source release, keeping the preview-first panel open path covered by CI so full transcript hydration stays deferred instead of blocking panel open.
• Current Immersion Kit example-target clicks now use explicit active-popover and ancestor checks, avoiding redundant popup navigation when the active example word is clicked inside dictionary details.

[0.7.68] - 2026-06-16

Fixed

• JPDB public dictionary details now render alongside Jiten details even when the clicked/search result card came from Jiten parsing, so keyless, Jiten-key-only, JPDB-key-only, and dual-key lookups can all show the real Jiten and JPDB source panels together.
• Definition source toggles now persist through settings normalization, so turning off Jiten or JPDB removes that source from both popover dictionary cards and the /search detail page.

[0.7.67] - 2026-06-15

Fixed

• YouTube transcript sidebars now open with a lightweight preview first and defer the full transcript render/hydration, removing the long synchronous panel-open pause on iPad/tablet watch pages while keeping Auto pause-panel reopening immediate.
• Left-side YouTube transcript placement now anchors YouTube's flex watch columns to the viewport start and shifts only the primary watch column, so the video, title, and actions sit against the panel instead of sliding too far left or underneath it. The Playwright iPad/mobile/desktop layout matrix now asserts this flex anchoring, preview render path, resize behavior, Auto behavior, and bottom safe-area alignment.

[0.7.66] - 2026-06-15

Fixed

• Corrected keyless Jiten dictionary rendering so the Jiten source uses the same real vocabulary-detail renderer as API-key mode, including readings, meanings, example sentences, composed-of / used-in words, and speech-bubble audio buttons. The keyless path no longer renders the local Jitendex inner substitute card or the Jitenで開く / Open in Jiten button inside the Jiten source.
• New-tab search detail fallback loading now fetches Jiten vocabulary details when the full card-render data loader is unavailable, so keyless dictionary searches still hydrate the real Jiten panel.

[0.7.65] - 2026-06-15

Fixed

• Tapping the video transcript side/bottom panel resize handle now closes the panel instead of resizing it, similar to the popover dictionary's sheet handle behavior.

[0.7.64] - 2026-06-15

Fixed

• Enabled Jiten dictionary definitions to work fully without a configured Jiten API key, including meanings, readings, example sentences, related words (composite / used in), and speech bubble TTS buttons. The card spelling is resolved using a public search fallback when a key is absent, and public Jiten endpoints are successfully queried without the Authorization header.
• Coexisted the external open button as a pill link rather than replacement content when keyless definitions are rendered.
• Fixed a potential JSDOM TypeError where location.pathname is missing or undefined under test environments.

[0.7.63] - 2026-06-15

Fixed

• Dictionary, kanji, study, popup, JPDB-page, new-tab, and docs Immersion Kit players now share the same example toolbar/action renderer, so audio, next, and previous controls update the active example in place without reparsing or reloading the page.
• Immersion Kit controls now wrap within their toolbar in light and dark layouts, preventing the action backdrop from overflowing narrow dictionary/study cards.

[0.7.62] - 2026-06-15

Fixed

• Tapping the video transcript side/bottom panel resize handle now closes the panel instead of resizing it, similar to the popover dictionary's sheet handle behavior.

[0.7.61] - 2026-06-15

Fixed

• YouTube watch-page transcript sidebars now keep iPad/tablet left placement from covering the video, title, or action area by shifting the watch content column alongside the player inset. Right placement keeps the existing side-dock behavior, while bottom placement is flush with the visible viewport and only resizes YouTube player nodes so metadata, descriptions, actions, chips, and recommendations do not stretch or clip.
• YouTube transcript layout changes now debounce the private player resize nudge during drag/keyboard resize interactions, avoiding repeated synchronous resize work while still settling the player dimensions after the panel changes. Added a Playwright iPad/mobile/desktop sidebar-layout matrix with screenshots and timing/DOM evidence.

[0.7.60] - 2026-06-15

Fixed

• YouTube Shorts now skips non-Japanese reels on the desktop player too (iPad's "Request Desktop Website" serves ytd-shorts, not m.youtube.com). The active short is classified by its URL video id + original (oEmbed/tab) title instead of the per-reel title, which on a non-English UI locale is auto-translated into Japanese and made every English short look Japanese. A dropped first nav-click no longer parks the player on a short forever.
• YouTube Shorts subtitles now appear on autoplay without needing a manual pause/resume. Swiping between reels reuses one <video> element and emits no navigation event, so the overlay used to latch "out of view"; the player now re-detects the active reel and self-heals the overlay when the video becomes visible again.
• The reader now starts with a disabled Anki client when the Anki companion library is missing, so dictionary popovers and action pills still initialize instead of failing the whole page with "Yomu Anki companion is unavailable."
• Removed an accidentally tracked node_modules symlink from the repository so fresh worktrees and worker setup are not blocked by a self-referential dependency path.

[0.7.59] - 2026-06-15

Fixed

• OCR image text now keeps furigana and pitch markup inactive until the user activates an OCR line or word, and clears that rich markup when selection moves or the lookup closes so docs/image OCR surfaces do not leave stale ruby or underline overlays.
• Expanded search word details now omit fetched-empty Jiten panels while preserving the generic keyless Jiten fallback link elsewhere.

[0.7.58] - 2026-06-15

Fixed

• No-key Jiten definition panels now render real imported Jitendex/Jiten term entries, including readings and structured example text, instead of replacing the section with only an external Jiten button. The same Jiten source renderer is used for API-backed details and local no-key entries, while the external Jiten link remains an extra action when real content is available.

[0.7.57] - 2026-06-15

Fixed

• Restored green CI by reconciling two contradictory JPDB definition-source tests. Example sentences render the headword occurrence as a rich passive JPDB word (ruby/pitch + identity for accurate lookup), and renderedJpdbRelatedWords scopes example headwords out so only used-in vocabulary is enriched/counted as related words.

[0.7.56] - 2026-06-15

Fixed

• Keyless Jiten dictionary sources stay visible in new-tab/search definition stacks when Jiten is enabled, rendering the passive headword and an external Jiten lookup link instead of disappearing while still avoiding duplicated JPDB meanings.

[0.7.55] - 2026-06-15

Fixed

• JPDB definition example sentences no longer double-count the headword as a "used-in" related word. The example sentence is parsed in place (nested, clickable) and highlights the headword inline, instead of also emitting a passive related-word token that inflated the related-word list.
• Integrated the in-progress P0 end-to-end fixes (action-pill navigation handlers, Jiten/JPDB definition source rendering, scanner coverage, and nested-parse refinements) onto the latest upstream main.

[0.7.54] - 2026-06-15

Fixed

• Supersedes 0.7.53 with the deferred YouTube public-vocabulary hydration release validation fixed, so live YouTube titles and chrome regain ruby/pitch without blocking first paint.

[0.7.53] - 2026-06-15

Fixed

• YouTube visible-page enhancement now hydrates fallback text with deferred public vocabulary lookups instead of disabling public lookup entirely, restoring ruby/pitch on live YouTube titles and chrome without blocking first paint.
• Visible-page scans now await async pre-render enrichment for fallback tokens that need readings, so resolved ruby is present before the DOM is wrapped instead of racing behind the initial render.

[0.7.52] - 2026-06-15

Fixed

• Japanese settings helper text inside audio source rows is now parsed for furigana and pitch while leaving the adjacent preview audio button untouched.

[0.7.51] - 2026-06-15

Fixed

• Hosted new-tab Settings now rebinds parsed Japanese labels to the current settings DOM before rendering, so relocalized labels such as 設定 and 設定の表示言語 keep furigana and pitch underlines instead of only enhancing preview/sample words.

[0.7.50] - 2026-06-15

Fixed

• Hosted new-tab settings now self-enhance Japanese settings chrome and appearance controls with furigana and pitch underlines, including stable labels such as 設定, 外観, 表示言語, 自動, and 日本語.
• Public JPDB fallback enrichment now re-applies furigana when a hosted new-tab fallback word resolves to a public vocabulary card, so no-key/settings UI parses gain readings instead of only updating hidden card metadata.
• Settings parsing now prioritizes visible settings controls before inactive panels while still covering hidden panels up front, so broad self-enhancement does not starve the current view.

[0.7.49] - 2026-06-15

Fixed

• Settings self-enhancement now skips transient status lines while preserving settings panel text, keeping hosted/new-tab parsing aligned with the concurrent enrichment release gates.

[0.7.48] - 2026-06-15

Fixed

• Japanese settings in ruby mode now use compact, grouped media controls so audio, Immersion Kit, OCR, video, and YouTube settings align cleanly without the large empty gaps and scattered checkbox/select rows seen in the reopened layout report.
• Added a responsive browser smoke for Japanese ruby settings at desktop, tablet, and mobile widths to catch control overlap, overflow, and excessive row gaps.

[0.7.47] - 2026-06-15

Fixed

• Compact stats chart gap taps now route through the resolved nearest day element, restoring coarse-pointer day selection and keeping the generated release shard aligned with the shipped interaction.

[0.7.46] - 2026-06-15

Fixed

• Hosted new-tab Settings now finishes self-enhancing Japanese settings panels without the userscript bridge by parsing hidden panels up front, using the runtime JPDB parser with timeout fallback, and avoiding redundant tab-switch reparses while keeping status notices and native controls untouched.

[0.7.45] - 2026-06-15

Fixed

• Hosted settings chrome enhancement now uses the local/segmented parser fallback instead of hitting JPDB, keeping parsed settings labels responsive while preserving the concurrent visible-page enrichment release.

[0.7.44] - 2026-06-15

Fixed

• Hosted new-tab searches such as 読み取る and よむ load keyless Word and Kanji study cards when switching from Search into Study, instead of getting stuck on "Looking for more words...".
• Empty study sources show actionable Starter words, Settings, and Search controls instead of inactive review navigation buttons.

[0.7.43] - 2026-06-15

Fixed

• Hosted new-tab searches such as 読み取る and よむ now load keyless Word and Kanji study cards when switching from Search into Study, instead of getting stuck on "Looking for more words...".
• Empty study sources now show actionable Starter words, Settings, and Search controls instead of inactive review navigation buttons.

[0.7.42] - 2026-06-15

Fixed

• Large visible furigana/pitch parse batches now fetch with bounded concurrency instead of waiting strictly one JPDB batch at a time, so big pages such as YouTube feeds, hosted settings/search chrome, and Immersion Kit reparses can progressively enhance without serial enrichment stalls.

[0.7.41] - 2026-06-15

Fixed

• Hosted new-tab Settings now enhances its own Japanese title, tabs, search label/results, panel labels/helper text, and Cancel/Save footer controls with reader parsing when the userscript is not active on the hosted page. Settings tabs, inputs, selects, checkboxes, links, and footer buttons remain native while passive reader words can still be hovered where appropriate.

[0.7.40] - 2026-06-15

Fixed

• Japanese new-tab chrome keeps native actions after reader ruby/pitch parsing. Clicking the parsed 設定 overflow item now opens settings instead of searching for nested parsed text, while stats chart gap taps still select the nearest day.

[0.7.39] - 2026-06-15

Fixed

• Dictionary popover source headers now keep enhanced Japanese labels such as イマージョンキット compact after ruby and pitch parsing, instead of stretching each parsed token across the accordion row.

[0.7.38] - 2026-06-15

Fixed

• YouTube feed and watch page scanning now covers signed-in feed channel names, metadata, watch titles/descriptions, sidebar recommendations, and transcript rows with reader color, pitch underline, and ruby where lookup data is available. Native caption overlays remain untouched to avoid disrupting YouTube playback timing and caption styling, while Yomu transcript rows continue to be enhanced.

[0.7.37] - 2026-06-15

Fixed

• Parsed Japanese new-tab chrome now preserves the original button action before nested word lookup. The Settings menu item can render reader ruby/pitch markup without turning clicks on 設定 into a search for parsed inner text such as 統計.

[0.7.36] - 2026-06-15

Fixed

• Jiten now appears as its own source panel in popover dictionaries and expanded new-tab search results whenever the Jiten definition source is enabled. If no Jiten detail result is available, the panel shows a clear empty state instead of disappearing behind JPDB, Anki, or local-dictionary availability.

[0.7.35] - 2026-06-15

Fixed

• New-tab Immersion Kit example navigation now swaps the visible sentence, count, translation, media URL, and controls before sentence parsing finishes, so next/previous clicks no longer feel frozen while ruby/pitch work catches up.
• New-tab Immersion Kit audio now tries the streamable media URL before waiting on slower blob/proxy hydration, so pressing play gives immediate playback feedback on hosted search pages while retaining the fallback path.

[0.7.34] - 2026-06-15

Fixed

• Hosted new-tab search now applies reader parsing and pitch enrichment to Japanese search chrome, autocomplete suggestions, kanji result cards, word result cards, and the Japanese settings dialog opened from search. Searches such as 学習能力 now hydrate from unknown pitch to concrete pitch underline/color classes when local or public pitch data is available, and settings chrome such as よむ 設定 can render furigana/pitch via the same bounded parser.

[0.7.33] - 2026-06-15

Fixed

• Dictionary popover action pills remain live in hover popovers, so Jiten, JPDB, Jisho, Yomu, and Copy actions open or provide feedback instead of rendering as disabled pill-shaped controls.
• YouTube dictionary popovers now have smoke coverage that clicks configured action pills and verifies the userscript tab-opening path without leaking the click to the page.

[0.7.32] - 2026-06-15

Fixed

• YouTube filter chips, mini-guide labels, and topbar controls such as 作成 now use the passive reader scan path, so stable Japanese control text can show ruby, pitch underlines, and source colors while native YouTube clicks still pass through.

[0.7.31] - 2026-06-15

Fixed

• Lookup popovers and expanded search detail headers now use the same duplicate-reading suppression as new-tab search results, so a kanji headword with visible furigana no longer repeats the identical kana reading beside the title while alternate readings remain visible.

[0.7.30] - 2026-06-15

Fixed

• New-tab search results no longer repeat the same kana reading in the metadata when that reading is already visible as furigana on the result term, while still showing the plain reading when furigana is disabled or suppressed by settings.

[0.7.29] - 2026-06-15

Fixed

• Twitter/X no longer shows a paused-frame OCR card on videos. Twitter plays every clip inline in the timeline and keeps the same markup on the tweet detail page, so there is no separate "watch" player to distinguish from a feed preview; all Twitter/X videos now opt out of paused-frame OCR. Posted photos still flow through the normal image OCR path.

[0.7.28] - 2026-06-15

Fixed

• YouTube no longer runs OCR on feed thumbnails. The inline hover preview (ytd-video-preview) reuses the real player markup, so it was being treated as the main player and snapshotted — leaving a "No text found" card pinned over a thumbnail; it is now classified as a thumbnail and skipped. Static feed/Shorts thumbnail images are likewise excluded from image OCR, so the extension no longer auto-sends YouTube thumbnails to the OCR provider. Paused-frame OCR on the actual watch-page player is unaffected.

[0.7.27] - 2026-06-15

Fixed

• Immersion Kit next/previous controls on the new tab now update the sentence, count, translation, parsed text, image, and audio source immediately while media finishes hydrating in the background.
• Immersion Kit cards now use a compact subtitle overlay across popovers, hosted search/new-tab views, and JPDB page addons instead of stretching a large floating caption box over screenshots.
• Mixed-script Immersion Kit fallback searches now keep useful compound verb fragments such as 読み取 and 取る, improving examples for terms like 読み取る.

[0.7.26] - 2026-06-15

Changed

• Recommended local dictionary downloads now include Japanese monolingual homepage entries for WTY JA-JA and the MarvNC monolingual collection.
• Local kanji-only dictionaries now stay scoped to kanji drilldowns and kanji-oriented study/detail cards instead of appearing on ordinary word definition cards.

Fixed

• Yomitan structured-content dictionary images are imported as inline image data and rendered with alt/fallback text when available.
• Monolingual structured Japanese glossary content remains parseable by the reader scanner without turning dictionary image fallback labels into nested lookup targets.

[0.7.25] - 2026-06-15

Fixed

• Japanese parsing now repairs incomplete kana spans inside continuous text, preventing dangling stems like やや さし and keeping context words such as 読んで and 読み取る coherent across Jiten, JPDB, local dictionary, and fallback parses.
• Selected sentence lookup now reparses long or fragmented rendered selections with sentence context, so large selections still open a popup and sentences like 好きなものを読んで日本語を学ぶ keep coherent token choices.

[0.7.24] - 2026-06-14

Fixed

• Page-reader scanning now uses one active/passive interaction model across supported sites, so links, tabs, summaries, buttons, and compact controls can render color, pitch underline, and ruby while native clicks still pass through.
• JPDB search links, dictionary-site controls, settings/new-tab controls, and hosted-docs chrome now stay passive for click/tap while hover lookup remains available outside Yomu's own dictionary popover.
• Dictionary popover summaries such as translation headings now get visual parsing but remain render-only, preserving details toggles and preventing nested popover lookup loops.

[0.7.23] - 2026-06-14

Fixed

• Settings changes now propagate across userscript, hosted docs, and new-tab contexts so open reader pages stay in sync after a save.
• Removed obsolete dictionary source toggles and result-limit controls from Settings; stale saved values no longer disable JPDB or local dictionary results.
• Furigana settings now show concrete selected modes instead of the vague automatic option, and legacy automatic values migrate to the effective behavior.
• Japanese settings labels, including parsed ruby labels, align cleanly with their controls, and the Appearance preview is centered.
• Empty local dictionary status no longer warns about Safari evicting data from the GM-backed store.

[0.7.22] - 2026-06-14

Fixed

• Study mode now falls back to usable word and kanji practice material based on the active tab, preventing keyless or degraded queues from staying stuck while looking for more cards.
• The Anki Study deck selector's "All vocabulary" option now loads the whole enabled collection instead of querying a literal deck named all.
• Connected Anki cards now render through the split Anki companion during the release smoke, covering existing-card popovers and new-tab review actions.
• The in-page Study deck selector now has Jiten deck scoping coverage alongside JPDB and Anki so the source selector behavior stays consistent.

[0.7.21] - 2026-06-14

Fixed

• YouTube channel recommendations now disappear entirely once every curated channel is already subscribed, instead of showing an empty/all-subscribed shelf.

[0.7.20] - 2026-06-14

Fixed

• Hosted new-tab searches now keep the q query parameter in sync for typed searches, nested modal lookups, handwriting candidates, and cleared searches, so browser back/forward restores the searched term and rendered results.
• New-tab search results no longer render the redundant global external-search link row when each expanded dictionary entry already has its own lookup pills.
• Search-result entry pills now dispatch Copy, Yomu, Jiten/JPDB, and Anki actions through the shared card-action path; the Anki pill now gives the same success/error feedback in new-tab search details as it does in lookup popovers.

[0.7.19] - 2026-06-14

Fixed

• YouTube feed, Shorts gallery, watch title, sidebar, transcript, caption, comment, and live-chat text now keep Yomu ruby, pitch underlines, and source colors on signed-in-style surfaces while preserving YouTube title/ruby layout protections.
• YouTube thumbnail images are excluded from automatic and hover OCR scheduling, preventing thumbnail OCR status overlays and reducing feed-scroll work.
• YouTube virtualized feed filtering no longer blanks visible pending cards or collapses filtered cards immediately, reducing black rows and scroll content-shift while still hiding offscreen pending items.
• YouTube comment and live-chat controls stay unwrapped while surrounding Japanese text remains parseable.

[0.7.18] - 2026-06-14

Fixed

• Large Japanese text selections now keep the selection popover open, remain sentence-aware, and show the configured Yomu/Jiten/JPDB/copy action pills on desktop and mobile.

[0.7.17] - 2026-06-14

Fixed

• YouTube channel recommendations now continue checking subscribed state through the expanded 100-channel shelf in small batches, so already-subscribed channels past the first visible preview rows are removed instead of being recommended again.

[0.7.16] - 2026-06-14

Fixed

• New-tab review now keeps Jiten available anywhere it is enabled, including auto-source review queues and Jiten-only source labels without requiring an unrelated JPDB API key.
• Hosted docs/newtab settings now mirror managed userscript settings back to same-origin app storage, keeping source settings and local dictionary preferences coherent where the hosted store can be shared.
• Hosted search now checks the actual local dictionary store before searching it, so imported Yomitan dictionaries appear when present and empty installs show the add-dictionary state instead of stale local results.

[0.7.15] - 2026-06-14

Fixed

• Hosted docs and new-tab Japanese UI text now receive ruby, pitch colorization, and pitch underlines across hero text, cards, install links, buttons, headings, and settings controls.
• Pitch accent underlines now use a consistent baseline across inline text, cards, buttons, headings, and settings surfaces, including words that render without furigana.
• Furigana settings are now respected exactly in hosted new-tab search terms, and default/difficult-kanji modes no longer inflate line spacing when ruby is suppressed.
• Added smoke coverage for enhancement coverage and underline baselines on hosted docs and new-tab fixtures.

[0.7.14] - 2026-06-14

Fixed

• YouTube popover action pills now own their clicks and open Jiten, JPDB, Yomu, and custom lookup links through userscript-safe tab APIs without leaking the click to YouTube's native navigation handlers.
• Firefox/Tampermonkey no longer needs to define the companion registry on YouTube page globals, avoiding XrayWrapper console errors from Yomu's own startup path.
• YouTube filter channel-subscription and mobile Shorts tests now clean up failed retries and wait for deterministic action completion, keeping release CI from inheriting async work across retries.

[0.7.13] - 2026-06-14

Fixed

• New-tab search kanji cards are now compact while staying touch-friendly, so kanji results no longer dominate the result grid.
• Expanded new-tab kanji details now use clearer section headers, compact component buttons, and per-kanji grouping for multi-kanji words so readings/components, RTK, component graph, Uchisen, and Immersion Kit sections remain scannable.

[0.7.12] - 2026-06-14

Fixed

• YouTube popover action pills now own their clicks and open Jiten, JPDB, Yomu, and custom lookup links through userscript-safe tab APIs without leaking the click to YouTube's native navigation handlers.
• Firefox/Tampermonkey no longer needs to define the companion registry on YouTube page globals, avoiding XrayWrapper console errors from Yomu's own startup path.

[0.7.11] - 2026-06-14

Fixed

• Hosted docs links now stay under the /yomu-reader/ GitHub Pages base, including the Getting Started and study-app paths that were escaping to root-level 404s.
• The docs landing hero has stronger responsive spacing on desktop, tablet, and mobile so the icon, copy, actions, and following install panel remain readable.
• The hosted new-tab fallback shell now opens in Japanese by default, localizes its document title and visible chrome, and keeps fallback card meanings localized before the full app hydrates.

Changed

• New-tab Japanese localization now has a parity test so future visible study-page copy cannot silently fall back to English.

[0.7.10] - 2026-06-14

Changed

• Page-reader kanji drilldowns now use the Kanji/Study companion for the heavier JPDB, RTK, stroke, and origin renderers, trimming the core userscript bundle while keeping local dictionary kanji details available.
• Reader page parsing now treats visible interactive labels across supported sites as passive targets instead of excluding them, so hover lookup, ruby, pitch underlines, and color classes can render while native clicks still pass through.
• Hosted docs hero and card links keep the same active/passive scan rendering path as other sites, including ruby and pitch underline coverage for stable headings and passive links.
• Project package metadata now uses the MIT license.

Fixed

• Audio fallback mode now keeps playable recorded/custom audio ahead of browser/API text-to-speech, including random replay, duplicate resolved media URLs, and nested custom JSON source pools.
• Hosted Yomu new-tab audio works without a userscript HTTP bridge, fetching custom JSON/audio through page fetch, playing via blob URLs, and keeping fallback text-to-speech out while recorded clips are playable.
• iPad blob playback, Wikipedia hover/tap replay, and YouTube hover playback now have source-backed smoke coverage for the randomized no-immediate-repeat pool.
• Passive words inside native buttons no longer start press lookup, keeping button taps and clicks native while hover lookup remains available.
• Page-reader kanji drilldowns no longer crash when the Kanji/Study companion is absent; they now show a dictionary-only fallback instead of an install notice.
• Hosted Yomu docs settings surfaces now follow the visible docs light/dark theme, preventing a dark settings panel over a light docs page.

[0.7.9] - 2026-06-14

Changed

• The default lookup pills now open Yomu search instead of Jisho, while Jisho remains available as an optional disabled pill.
• The hosted search page reads q, query, and search URL parameters and renders search links from the user's enabled lookup pills, excluding the redundant Yomu self-search link.

Fixed

• Hosted search results load again from query-param searches, including public JPDB fallback results when no local dictionary result matches.

[0.7.8] - 2026-06-14

Fixed

• Subtitle overlay text now uses the vivid status/source color on dark video surfaces instead of the UI-readable tint, keeping colored subtitles legible while preserving transcript panel readability.

[0.7.7] - 2026-06-14

Fixed

• YouTube watch titles, descriptions, comments, OCR text, and subtitles keep ruby/pitch rendering while virtualized homepage, sidebar, and Shorts gallery title chrome stays native to prevent the "only furigana remains" DOM corruption.
• YouTube comment "more" buttons, live-chat reply buttons, and recommendation chrome are no longer parsed as hover dictionary targets.
• Paused-frame OCR overlays, status cards, and rail resume buttons are torn down on YouTube SPA navigation so homepage preview frames cannot stick over the next watch page or add duplicate play controls.
• Subtitle detection is retried immediately on YouTube video load/metadata changes and the first cue renders during the initial lead-in, reducing blank overlay/sidebar time on short-form clips.
• YouTube Shorts filtering now recognizes reel overlay/card video ids more reliably and keeps non-Japanese Shorts out of desktop and mobile galleries.
• The YouTube smoke suite now covers homepage-to-watch OCR cleanup, side-panel rail/toggle visibility, player resizing, mobile feed filtering, Shorts filtering, and iPad-sized touch playback layouts.

Changed

• Native subtitle blur now uses transparent text fill plus text-shadow instead of a CSS blur filter, keeping hover reveal responsive on mobile/tablet GPUs.
• Reader underline alignment now uses a consistent line-height for non-passive reading surfaces so words with and without furigana share the same underline baseline.

[0.7.6] - 2026-06-14

Fixed

• The settings dialog on the documentation page now respects the pink accent color setting on startup and during subsequent mode transitions instead of overriding it to the default green.
• JPDB pointer candidate generation now resolves deinflected dictionary terms correctly when performing public dictionary lookups.

[0.7.5] - 2026-06-14

Fixed

• Word highlights and underlines now leave a visible break between adjacent same-status words across page text, ruby-enhanced text, subtitle overlays/transcripts, settings previews, and OCR active lines.

[0.7.4] - 2026-06-14

Changed

• Merged the verified 0.7.3 onboarding, host-style isolation, parser coverage, accent picker, passive contrast, and audio-source fixes back onto main.
• Anki mining is opt-in on first run while existing migrated Anki setups stay deliberate and preserved.

Fixed

• Audio replay selection now skips duplicated resolved media URLs and repeated browser text-to-speech voices before falling back to another available source.
• Hosted docs hero text keeps Yomu pitch underlines aligned when words render without furigana.
• Page-reader kanji drilldown popovers now hydrate enabled Uchisen mnemonic cards instead of leaving the source mount empty.

[0.7.3] - 2026-06-14

Changed

• The first-run welcome screen copy is shorter: image OCR now says to tap any image, and the built-in study page description is a single concise sentence.

Fixed

• The welcome screen no longer shows an internal scrollbar on tall desktop viewports just because of the old 720px height cap.
• Host-page heading styles, including Wikipedia-style borders, no longer add a horizontal rule under the よむ title on the welcome screen.
• The welcome screen now leaves a clearer gap between the immersion default checkboxes and the action divider.
• The custom accent color picker now previews rapid drags once per animation frame and only publishes the selected color when the picker commits, preventing preset/custom color churn and sluggishness while trying many colors quickly.
• Host-page styles can no longer force Yomu settings inputs, number fields, selects, textareas, checkboxes, or onboarding checkboxes into site colors such as Discourse's dark controls or blue checkmarks.
• Reader-owned settings chrome, including the theme label and settings tabs, now receives the same nested ruby and colorized word rendering as settings labels and help text.
• Passive site UI chrome coverage now includes role tabs, checkbox/radio/switch controls, titled controls, and Wikipedia Vector pinnable buttons so visible Japanese labels keep ruby, underlines, highlights, and text-color classes without stealing native clicks.

[0.7.2] - 2026-06-14

Fixed

• Hover dictionary autoplay now ignores stale hover lookups before they start playback or mark the popover as loading. Moving quickly between words no longer leaves old audio work competing with the current card.
• Immediate autoplay de-duplication now uses the full card identity instead of only the visible spelling, so distinct same-spelling lookup cards can still play their own hover audio.
• The performance profiler now includes word-to-word hover timing while the current popover's inner sections are expanded, covering the laggy dictionary case that does not show up when sections stay collapsed.

[0.7.1] - 2026-06-14

Fixed

• Re-published the Discord strict-content-security-policy pitch accent stylesheet fallback on the 0.7.x release line. The userscript version now reflects the current release train while preserving the fix that keeps pitch accent colorization and pitch underlines available when full reader CSS has to load through the fallback path.

[0.6.201] - 2026-06-14

Fixed

• Discord and other strict-content-security-policy pages now keep pitch accent colorization and pitch underlines when the full reader stylesheet has to be loaded through the fallback path. Yomu now asks the userscript HTTP bridge for the hosted stylesheet before trying page fetch, and the critical inline stylesheet includes the minimal pitch word paint rules needed while the full CSS is unavailable.

[0.6.200] - 2026-06-14

Fixed

• Hosted documentation pages now honor the configured accent color for dark-mode brand buttons. Setting the accent to pink no longer leaves the "Install よむ" button on the old green hover/active CSS path.

[0.6.199] - 2026-06-14

Fixed

• AnkiConnect is no longer contacted by a direct cross-origin fetch when the userscript/extension request bridge is absent. Opening the hosted yomu site (or any content page) without the userscript installed used to fire a doomed request to the default http://127.0.0.1:8765, which the browser blocks and logs as "Cross-Origin Request Blocked" on every attempt. Yomu now skips that request and reports Anki as needing the bridge instead. Same-origin endpoints and bridge-backed requests (Tampermonkey/Violentmonkey and the browser extension) are unchanged, so configured Anki setups keep working on hosted pages.

[0.6.198] - 2026-06-14

Changed

• The YouTube channel recommendation shelf is now hidden entirely when every compact-view channel is already subscribed, instead of showing an empty list with a "browse all channels for more" message.

Performance

• YouTube feed/search/description scans no longer fan out background public JPDB pitch lookups while the page is mutating. Ruby/text rendering stays first, hover lookups remain urgent, and desktop/mobile profiles now keep public pitch requests to the hovered words instead of hundreds of background scans.
• The YouTube homepage performance profiler now records thumbnail readiness, scroll stress, OCR frame counts, and word-to-word hover popover latency with a 250 ms near-instant SLA across JPDB, Anki, Jiten, combined, and no-key settings.

Fixed

• YouTube hover-preview videos in feed cards are no longer treated like the main paused player for OCR, preventing preview loops and useless OCR work on homepage thumbnails.
• Paused-frame OCR overlays now have regression coverage for rendered ruby and pitch/color classes, matching the parsed subtitle/text rendering path.

[0.6.197] - 2026-06-14

Performance

• YouTube homepage, feed, watch-page, description, and comment rescans now use the narrow YouTube site parser with a shorter mutation/initial delay while the broad generic fallback stays disabled for YouTube chrome. This restores fast ruby coverage for newly loaded video cards without re-attaching to player controls.
• Hover lookups now interrupt in-flight visible-page scans before doing dictionary work, and stale scans stop before applying DOM tokens. This keeps word-to-word popovers responsive while long descriptions/comments continue annotating in the background.
• Hover popover shells yield one animation frame before heavier dictionary hydration, so the card can appear promptly under scan stress instead of waiting behind rendering work.
• Term-audio hover playback now coalesces duplicate in-flight requests and suppresses only successful immediate duplicate autoplay attempts. Manual replay still works, and random audio selection keeps rotating through shuffled candidates instead of always using the first returned clip.

Fixed

• YouTube OCR/status text and subtitles remain eligible for ruby/pitch colorization while avoiding repeated raw-text removal/addition loops from stale parser output.

[0.6.196] - 2026-06-13

Performance

• YouTube visible-page scanning now splits long text targets safely and applies chunks from the end of each text node back toward the start, so expanded descriptions and comment churn no longer corrupt text offsets or force giant apply turns.
• Mobile YouTube scanning uses smaller cooperative batches when running without a JPDB API key, while API-backed parsing keeps larger mobile batches to avoid excessive continuation work. The stress harness now covers expanded descriptions, comment scrolling, play/pause, side-panel state, desktop hover, and a throttled mobile viewport.
• YouTube page parsing now avoids player controls, chip tabs, and other button chrome while still scanning titles, descriptions, comments, recommendations, chat, subtitles, and OCR text. This keeps native YouTube controls responsive and prevents lookup/ruby work from attaching to "Auto" and similar UI labels.
• Hover popovers defer study/Immersion loader setup until after the initial shell paints, and hover-card hydration repositions are coalesced into animation frames to reduce first-hover and word-to-word latency.
• Hover/audio warmup no longer prepares the same lookup audio repeatedly for a card; candidate-only warmups can still upgrade to playable audio once.
• Repeated YouTube render-rejection rescans are debounced, reducing remove/add loops when the host rehydrates scanned description/comment text.

Fixed

• JPDB API reads now retry transient connection resets and back off briefly when jpdb.io keeps closing connections, while the all-decks scan spaces per-deck requests instead of firing a burst that can trip Firefox PR_END_OF_FILE_ERROR failures on the same network.
• Partial-fragment rendering preserves the untouched prefix/suffix text around long scanned chunks, preventing the "raw text, ruby, then only ruby remains" failure mode in YouTube panels.

[0.6.195] - 2026-06-13

Added

• The first-run welcome splash now includes the same dark/light theme toggle used elsewhere, so new users can switch appearance before finishing setup.

Changed

• Settings now opens with Appearance first, followed by API, Sources, Media, Mining, Study, Shortcuts, and Help.
• The welcome splash layout now matches the rest of Yomu's UI more closely: setup controls are grouped as form rows, the decorative gradient is gone, and the feature cards have been replaced with a compact feature list.

[0.6.194] - 2026-06-13

Added

• OCR on paused video frames is now a user-facing toggle ("Read paused video frames") in the OCR settings panel. Previously it was always enabled with no way to turn it off without disabling OCR entirely.

Fixed

• YouTube Shorts and other portrait YouTube videos on desktop keep their native player size when the side transcript panel is open, instead of being stretched/cropped by the transcript layout resize.
• Fixed CI: the pitch-underline smoke script was not wired into package.json, causing the fallow:dead-code check to fail.

[0.6.193] - 2026-06-13

Performance

• Hovering words is snappier (live-profiled on a signed-in YouTube watch page). The dictionary popover no longer re-scans the whole popover for native title tooltips on every reposition (it's set up once at mount and kept current by its observer) — repositioning got dramatically cheaper. The page-word contrast pass memoizes the page-background computation per parent instead of walking each word's ancestors with getComputedStyle, so hovering words in dense text no longer recomputes the same background per sibling.
• The dictionary popover fades in with a subtle ~110ms opacity animation (opacity only — no layout shift; respects prefers-reduced-motion).

Fixed

• ImmersionKit examples now load in the popover even when the immersion section sits below the visible popover fold. Loading was gated on the section scrolling into view, so on tall cards the examples often never appeared until you scrolled; the open section for the looked-up word now loads on a short debounce regardless of fold position (in-view sections load faster, and abort-on-close still prevents redundant requests while moving between words).

[0.6.192] - 2026-06-13

Fixed

• Mobile YouTube Shorts filtering keeps advancing through adjacent non-Japanese Shorts until it reaches a Japanese-looking Short again, instead of stopping after the first auto-advance throttle window.

[0.6.191] - 2026-06-13

Fixed

• YouTube watch titles, mobile watch titles, home/search/sidebar/Shorts titles, metadata, and visible button labels are parsed for JPDB ruby, color, and pitch again instead of being excluded or stripped by the YouTube filter. Decorative hidden touch-feedback chrome stays ignored generically through aria-hidden.
• Long YouTube titles and other cropped text rows now keep furigana visible by growing every clipped ancestor that needs ruby room, including stale or nested clipped rows.

[0.6.190] - 2026-06-13

Performance

• Keyless (no API key) subtitle parsing no longer floods IndexedDB. Each cue's pitch + per-kanji reading lookups now run through a shared concurrency gate, so warming several cues at once can't fire thousands of simultaneous IndexedDB requests that starved the main thread and left lines half-annotated as they played. Live-profiled on a signed-in YouTube watch page: dense 12-word ASR cues now get full ruby/pitch in time.
• Hovering to look up a word is skipped while a mouse button is held (i.e. while dragging). Live profiling showed the hover probe (elementFromPoint + a querySelectorAll/getClientRects sweep over every transcript word) was a dominant cost of subtitle-sidebar resize lag; it no longer runs during a drag.
• The subtitle sidebar resize drag reuses the latched player rect instead of re-measuring it every frame (which toggled inset styles and forced two synchronous layouts). Measured getBoundingClientRect cost during a resize drag dropped ~60%.
• The page text-scan target dedup is no longer O(n²): outermost-match filtering uses an ancestor-membership walk instead of a nested native contains() scan, which dominated the auto-scan on YouTube's churning DOM during playback.
• Pointer-move hover probing is coalesced to once per animation frame instead of running on every raw pointer event.

[0.6.189] - 2026-06-13

Fixed

• Opening the popup dictionary now temporarily suppresses native browser title tooltips on the active lookup word's page ancestor path, so host-page title tooltips no longer cover the dictionary. Unrelated page titles are left alone and every suppressed title is restored when the popup closes.
• Keyless YouTube subtitle warmup now enriches future cached lines with public/local vocabulary and pitch before they can be reused during playback, so background transcript parsing no longer locks in plain half-ready subtitle HTML ahead of the active cue.

[0.6.188] - 2026-06-13

Added

• The transcript auto-scroll resume window is now configurable in Settings → Media ("Resume transcript auto-scroll after manual scroll (s)", default 4s). After you manually scroll the Lines panel, auto-follow stays paused for this many seconds before snapping back to the active line.

Fixed

• Jiten and JPDB now stay as separate dictionary definition sources in settings and popup ordering. Both source rows are enabled by default even before API keys are added, and adding a Jiten key no longer relabels or replaces the JPDB source.

[0.6.187] - 2026-06-13

Fixed

• Apple Pencil / pen hover once again opens hover lookups on iPad when hover lookup is enabled. Pen contact still suppresses immediate hover re-open briefly, so a real tap or touch-down does not produce the old rapid open/close behavior.

[0.6.186] - 2026-06-13

Fixed

• Manually scrolling the subtitle Lines panel no longer gets yanked back to the active line when the video advances to the next cue. Auto-follow pauses for a few seconds after a manual scroll and resumes on its own (or immediately when you click a line or use Previous/Next). The list's own snap-to-active scroll is no longer mistaken for a manual scroll.

[0.6.185] - 2026-06-13

Fixed

• Pausing a YouTube video on mobile now captures the OCR snapshot again. The main player on m.youtube.com is wrapped in a generic /watch link, which a v0.6.182 thumbnail guard mistook for a hover-preview tile and skipped the snapshot ("the auto doesn't work on pause"). Real feed/preview tiles are still ignored; the guard now only treats link-wrapped videos as thumbnails when they are not player-sized.
• Dragging the subtitle sidebar resize handle is smooth again: the layout-heavy panel reposition now runs at most once per animation frame instead of on every raw pointer-move event.
• The per-tick asbplayer move-handle sync (added in v0.6.176, running every ~250ms on every video on every site) now takes a fast early-out when no asbplayer subtitle overlay is present, instead of doing a document-wide drag-handle scan each tick.
• Resolving the generic (non-YouTube) video layout target during resizes no longer re-scans each candidate frame's whole subtree for player controls on every pointer move — the result is memoized per element with a short TTL.

Notes

• Exact-cue subtitle caching: the cache keying and identical-render guard were verified intact since 0.6.175; the "caching looks broken" reports trace to the render/layout jank fixed above, not to re-parsing or cache-key churn.

[0.6.184] - 2026-06-13

Fixed

• Hardened the settings dialog import-queue regression test so the release workflow can publish the verified YouTube, ASB-player, JPDB, and settings fixes without CI timing out under load.

[0.6.183] - 2026-06-13

Fixed

• YouTube watch controls, player settings submenus, suggested videos, and view/count metadata are parsed for ruby, color, and pitch styling again, while native clicks stay intact and hover popover lookup CTAs remain visible but inert.
• ASB-player style subtitles are enriched with ruby and deck/pitch color before cue HTML is displayed, so pre-rendered cues keep their styling when moved onscreen instead of visibly loading in afterward.
• Video subtitle/control rails now render only inside an actual visible video frame and stay hidden on no-video pages.
• JPDB native ruby headwords such as 発行 stay as one lookup word, kanji-used links receive furigana, and Yomu-injected JPDB panels fit the host column width.
• Appearance settings are simplified: the confusing default preset is gone, the old phone/tablet copy is removed, and the larger colorized preview now lives in the Appearance panel.

[0.6.182] - 2026-06-13

Added

• Playwright smoke coverage for YouTube auto-translated caption tracks whose translated timedtext endpoint returns no rows.

Fixed

• YouTube auto-translated subtitle tracks no longer leave the Lines panel stuck at "0 lines" / "Loading subtitle lines" when the translated timedtext response is empty: Yomu now translates usable source captions itself, and preserves YouTube's native translationLanguage when falling back to DOM captions.
• Jiten/example sentences with a highlighted word split across inline elements now keep the ruby/furigana render instead of dropping into plain text.
• Paused-video OCR controls now live in the subtitle rail when available, use a compact play icon fallback elsewhere, and ignore YouTube hover-preview thumbnail videos.

[0.6.181] - 2026-06-13

Changed

• Furigana in cramped rows is never stripped anymore: when ruby makes a clamped or fixed-height row overflow, the row gets room instead (line-clamp boxes keep their line count and lose only the plain-text height cap; other clipped boxes grow to their content height). Verified end-to-end — furigana stays and the full base text shows.
• The jpdb/jiten page add-on (Immersion Kit, dictionary entries) uses the width the page actually has instead of inheriting a narrow host column, and example media can render up to ~480px tall.

Fixed

• Words Jiten doesn't track were silently treated as "mature" — coloring them as known and hiding their furigana on jiten.moe pages. They are now neutral (not-in-deck), so untracked words keep honest colors and visible furigana.

[0.6.180] - 2026-06-13

Changed

• The clamp sweep is now evidence-based instead of list-based: it strips furigana only from rows whose container measurably overflows (any site, no per-site whitelists), so ruby that fits stays — including inside expandable descriptions and comments — and ruby that actually crops is removed wherever it happens. Verified across YouTube watch, jpdb.io, and NHK News Easy with zero measurably-cropped ruby rows.

[0.6.179] - 2026-06-13

Fixed

• Paused-video OCR usability: a floating "Resume video" pill now sits on the snapshot (recognized text swallows clicks for lookups, so text-dense frames made the player hard to unpause — the pill always works and also clears the overlay when playback is blocked), and stepping next/previous subtitle line while paused re-captures the snapshot instead of showing the stale frame.
• Restored pieces dropped in a release rebase: all-variant pitch patterns from imported dictionaries, the persistent-storage request + eviction notice for imported dictionaries, and the aligned grammar-row label column.

[0.6.178] - 2026-06-12

Added

• Subtitle player Playwright coverage now includes BBC-style article video, Video.js, JW Player, Plyr, Vimeo-style, Wistia-style, Mux/Kaltura-style, CIJ, mobile YouTube, desktop YouTube placements, and a live-site player discovery smoke.

Fixed

• Subtitle overlays and transcript drawers now anchor to the player frame/chrome instead of the centered raw video or surrounding article body on generic video sites.
• Generic player frames now resize with side/bottom transcript drawers on all non-YouTube sites, not only CIJ, and dispatch resize events so embedded players can refit themselves.
• Subtitle move handles are actually hit-testable and draggable above the subtitle text, with a mouse fallback for environments that do not deliver pointer events.
• ASBPlayer subtitle dragging now composes with the player’s existing transform instead of relying on the individual translate property, and the default move handle no longer paints a shadow bar.

[0.6.177] - 2026-06-12

Fixed

• Mobile YouTube watch metadata, action chips, descriptions, and transcript controls now keep furigana after late hydration instead of briefly showing ruby and then losing it to the compact-row clamp cleanup.

[0.6.176] - 2026-06-12

Added

• Lookup pitch graphs now show every accepted accent variant from imported pitch dictionaries, and append imported dictionary variants to the one jpdb supplies.

Fixed

• YouTube channel suggestions now keep Yomu's short level/topic summaries after YouTube preview hydration, instead of replacing them with long channel bios in compact suggestion rows.
• YouTube filtering now continues loading mobile homepage cards when the visible continuation item is reached, avoids annotating recommendation-grid titles as page text, and keeps the current Shorts watch item visible while hiding non-Japanese neighboring Shorts.
• Keyless subtitle playback now treats provisional parses as warm for lookahead and transcript warmups, shares in-flight parses across tiers, caches empty parses in the retry TTL, and re-anchors warmup after long seeks into cue gaps.
• YouTube native-caption fallback now warms the normalized text that actually renders, renews captions the page keeps showing, and can re-apply identical captions after a seek.
• Late pitch/vocabulary enrichment now re-bakes cached subtitle cue HTML and hydrated transcript rows, so stepping back keeps pitch colors.
• Public jpdb pitch lookups now run four-wide, so keyless pitch underlines fill in faster during subtitle bursts.
• ASBPlayer subtitle overlays no longer inherit Yomu subtitle underline color channels, avoiding dark shadow underlines when API-backed status/pitch styling is enabled.
• ASBPlayer subtitle overlays now get the same deliberate temporary move handle as Yomu subtitles, so their position can be nudged per video without dragging ordinary subtitle text.
• Deck membership styling now applies across Jiten, JPDB, and Anki rendered words, with provider-specific deck classes and merged Anki metadata preserved.
• Word lookups now resolve the real pointer target through shadow-DOM event retargeting, so framework re-renders that replace annotated nodes are re-annotated by the next scan.
• Imported dictionaries on iPad now request persistent storage on import, and the Dictionaries panel explains when Safari has cleared remembered IndexedDB data.
• Grammar rows in the lookup sheet now indent consistently across alternate forms.

[0.6.175] - 2026-06-12

Fixed

• Yomu-owned roots now inherit Yomu text color through their descendants as part of the host-page reset, so broad site CSS cannot tint settings sections while the floating puck and panel surfaces stay correctly sized and themed on Reddit, mobile Reddit, and other aggressive CSS hosts.

[0.6.174] - 2026-06-12

Added

• Subtitle overlays can now be temporarily repositioned per video with a dedicated move handle, with keyboard nudging from the focused handle for deliberate adjustments.
• Added a configurable subtitle-overlay shortcut, defaulting to Shift+H, to toggle Yomu/ASB-style subtitle lines without opening settings.

Fixed

• Unsetting primary/secondary subtitle tracks now clears the current rendered Yomu/ASB-style subtitle line instead of leaving the previous line in the player.
• Apple Pencil hover no longer rapidly opens and closes dictionary popovers over OCR/image-reader words; pen hover can scan, while lookup popovers require mouse hover or an explicit tap/click.
• Native page ruby is preserved when a scanned word fully covers the ruby base text, avoiding split or duplicated furigana in annotated pages.
• Kanji/Study companion features now degrade to no-op clients with a visible notice when the companion bundle is missing, instead of importing companion-only modules into the base runtime.

[0.6.174] - 2026-06-12

Added

• Copying a subtitle line now flashes a small check on the copy button, "Include the translation when copying a line" is a Media setting (default on, matching the old behavior), and rows with an aligned translation get an eye toggle to peek it while the list stays Japanese-only.

Fixed

• Blank bands in the subtitle Lines panel: auto-translated tracks ship literal HTML entities, and a cue containing only   rendered as an empty full-height row. Entities are now decoded, so empty cues are dropped.
• The hosted video-player page called an undefined hook on every language/theme application; the error aborted page setup. The subtitle player smoke also injects the companion scripts now (the player moved into the video companion long ago, so the smoke could never mount it).

[0.6.173] - 2026-06-12

Fixed

• The extension-only "Set Study as the new tab" toggle no longer drags the boot module into the settings form (an import cycle that broke runtime detection in 0.6.172); the runtime probe lives in its own dependency-free module and the toggle labels are localized (en/ja).

[0.6.172] - 2026-06-12

Fixed

• Image OCR now starts on large image-reader surfaces even when the page has no Japanese DOM text, covering image feeds and manga/gallery viewers rather than only standalone image documents.
• OCR overlays now refresh when an inner scrolling container moves, so image-feed pages that keep window.scrollY fixed still keep text regions aligned.

[0.6.172] - 2026-06-12

Fixed

• Disappearing/clipped text on YouTube (feed titles ending in bare "…", channel rows clipping mid-glyph, comment bodies vanishing, end-screen tiles): hosts that hydrate progressively apply their line-clamp/ellipsis styles after Yomu annotated the text, and rescans never revisit already-annotated words — so the grown furigana line stayed and the crop swallowed the base text. A clamp sweep now runs after every scan (and again once hydration settles) and strips ruby from rows that became layout-sensitive, keeping the colors and lookups. Verified end-to-end with a Playwright reproduction of the late-clamp race.

Changed

• "Set Study as the new tab" is back as a real toggle in extension builds (extensions can override the new tab); userscript installs keep the guidance text instead of a checkbox that can't work.

[0.6.171] - 2026-06-12

Fixed

• jpdb.io "Kanji used" glyphs and the pitch-accent diagram are no longer word-annotated — the kanji link was matching rare alt-form words (穏 rendered a しずか reading under the glyph).
• The paused-video OCR snapshot now pins to the video's real content box instead of stretching across the letterbox bars.
• Clicking OCR'd text on a video thumbnail no longer navigates to the video — the click stays with the lookup.
• The study-address help ("set as your browser's start/new-tab page…") is no longer overwritten by the offline-cache help; both texts are now accurate and live in their own spots.

Changed

• Removed the "Enable Yomu study page" checkbox — it had no effect (a userscript cannot override the browser's new tab); the address guidance now explains the real options.
• The Shortcuts panel now lists the fixed Study-page keys (Space/Enter reveal, 1–9 grades, U undo, ←/P previous-undo, →/N next) so every shortcut is documented in one place.

[0.6.170] - 2026-06-12

Added

• Words that live in both the Jiten and JPDB review queues now keep both identities after merging, so the study grade-target selector offers Both / Jiten / JPDB (and Anki) per card — grading just one provider works exactly like the jpdb+anki pairing always has.

Changed

• Triaged the 15-screenshot iPad feedback wave into the backlog (UT-64…UT-78): jpdb kanji-used furigana position, lookup pitch gaps, keyless subtitle completeness, lines-panel gap, copy-subtitle UX, end-screen clipping, disappearing feed/comment text, shortcuts-settings completeness, jitendex regression, study-settings copy, sidebar-left player void, unannotated chrome, paused-OCR stretch, reddit styling.

[0.6.169] - 2026-06-12

Fixed

• Removed the remaining "one credential or the other" assumptions: with both Jiten and JPDB keys, Jiten features no longer pretend JPDB is absent, settings labels read "Jiten + JPDB" where both apply, and an open jpdb.io review tab no longer hijacks the merged study queue in Auto mode — the explicit "Live JPDB review session" mode is the only thing that preempts, exactly as chosen.

[0.6.168] - 2026-06-12

Added

• With both Jiten and JPDB connected, the study page's deck picker now has one-tap "JPDB" and "Jiten" entries to study a single provider's queue — "All vocabulary" keeps mixing both, and individual decks keep scoping further.

[0.6.167] - 2026-06-12

Added

• Undo now works after every grade, on every provider: Jiten reviews reverse server-side as before, while JPDB and Anki grades — which cannot be reversed upstream — return the card to the front of the queue with an honest note that the recorded review still counts. The U key and the undo button appear for all of them.
• Pressing Previous (or the browser Back button) right after grading performs that undo — stepping back across a grade boundary is the natural undo gesture.
• Every study card has a stable URL (#card=…): advancing pushes browser history, so back/forward walk the session, a reload lands on the same card, and links to a specific card can be shared or bookmarked.

[0.6.166] - 2026-06-12

Changed

• Rendered Anki card headings no longer show raw database card ids: a multi-card note reads "Deck · Template" (or "Deck · Card 2"), while grade-target labels keep the #id where it genuinely disambiguates duplicate cards.

[0.6.165] - 2026-06-12

Added

• The study page's due summary now includes kanji reviews waiting on jpdb.io ("+N kanji on jpdb.io") whenever a jpdb.io learn or review tab is open: kanji reviews only exist on jpdb.io itself (the API has no access to them), so the page previously under-reported dues compared to jpdb Learn without saying why. The bridge that powers live reviews now also reads the Learn page's due composition.

[0.6.164] - 2026-06-12

Added

• Jiten study decks appear in the study page's deck picker (labelled "Jiten · deck name") when a Jiten key is connected. Picking one scopes the queue to that deck's words — Jiten's API has no per-deck review call, so Yomu intersects the global batch with the deck's word list. Picking a JPDB deck now scopes to JPDB only, and "All vocabulary" merges both providers as before.

[0.6.163] - 2026-06-12

Fixed

• Filtering no longer shifts the feed under your finger on mobile: the scroll compensation that keeps the viewport steady when cards hide now finds the actual scrolling container (m.youtube.com scrolls inside one, not the window), so it engages on phone and tablet layouts where it previously never fired.

[0.6.162] - 2026-06-12

Changed

• Grading with the buttons now advances with the same brief card-enter motion as a committed swipe, so the two grading paths feel identical (disabled under reduced-motion preferences). The failed-card requeue uses it too.
• The appearance preset and "Color words" options are now localized in Japanese.

[0.6.161] - 2026-06-12

Added

• First-run guidance on the study page: when no SRS source is connected, the practice-words view now shows a "Connect Jiten / JPDB / Anki" button that opens settings directly — instead of silently showing random practice words.

Verified

• Live queue-mirroring check against Jiten with a seeded account: the study page's card order matched the Jiten API's own study batch exactly (due cards first, then the server's new-card order), and the count line matched the API's due summary.

[0.6.160] - 2026-06-12

Added

• Press U on the study page to undo the last review where an undo exists (Jiten), matching the on-screen undo button — keyboard-only reviewing no longer needs the mouse for corrections.
• The Anki deck picker on the study page now shows each deck's waiting count (reviews + learning) straight from Anki's scheduler, refreshed at most once a minute.

[0.6.159] - 2026-06-12

Changed

• Loading Anki reviews on the study page is substantially faster: candidates from disabled decks are dropped with one cheap call before any card rendering, and card details now stream in small batches that stop as soon as the queue is full — instead of rendering an entire overfetched window up front (AnkiConnect renders every card's templates at ~110ms per card, so this cuts seconds to tens of seconds off study startup on large collections).

[0.6.158] - 2026-06-12

Fixed

• Anki kanji cards (RTK-style decks) no longer show the same kanji repeated several times: those templates render the glyph in multiple decorative fonts that cannot load outside Anki, so the repeats collapsed into identical copies. Yomu now keeps one glyph per card side and drops the duplicates (verified against a real RRTK deck), along with dead text-to-speech placeholders.

Changed

• The Anki status index rebuild is dramatically faster on large collections: it now derives note, deck and state information from three fast bulk queries instead of asking Anki to render every card's templates (~110ms per card — about 20 minutes for a 12k-card collection, now seconds). Classification verified identical against a live collection; the old path remains as a fallback for older AnkiConnect versions.

[0.6.157] - 2026-06-12

Fixed

• Compact UI text is no longer excluded from annotation: tight rows, small headings and media-bearing text links (like YouTube channel names next to avatars) now get colour/state annotation without furigana, instead of being skipped entirely. Ruby still stays off those rows so it can never crop or reflow them; true icon buttons and UI chrome remain untouched.

[0.6.156] - 2026-06-12

Fixed

• Jiten and JPDB API keys now coexist: settings has separate fields for each, and the study queue mixes both providers' reviews in one session. Previously saving a Jiten key silently erased the JPDB key, which made the study page mirror only Jiten while jpdb.io Learn kept counting dues — the main source of "my queue doesn't match jpdb" reports. A Jiten-format key pasted into the JPDB field still routes to the right place.

[0.6.155] - 2026-06-12

Fixed

• The Shorts shelf now fills its visible row properly after filtering: Yomu pages the shelf until it has as many Japanese items as the row advertises (or the paging cap is reached), and a shelf whose hydrated items are all non-Japanese collapses entirely instead of stranding one or two cards. Fully-filtered shelves are paged too — previously they froze empty because hiding happened before backfilling.
• Keyboard hint pills are no longer rendered at all on touch-only devices (previously they were only hidden with CSS).

[0.6.154] - 2026-06-12

Fixed

• Refreshing a video page no longer loses subtitle ruby and colouring: parsed cue annotations now persist for the browser session (six-hour cap) and restore instantly on reload instead of waiting for a fresh parse. This also covers keyless use, where the local parse result is the final one — it is persisted and reused the same way.

[0.6.153] - 2026-06-12

Added

• Appearance presets with a live preview (user request): one-click configurations covering the popular jpdb CSS recipes without writing CSS — don't color words, only color new words, underline new words instead of coloring, show all furigana, hide furigana for chosen card states, furigana on hover only, or no furigana. A sample annotated sentence in Settings → Reading restyles instantly as you change options.
• Furigana hiding is now per state group: choose any combination of Known, Due, Failed, Learning and New under "Hide for chosen states". A new "Show on hover only" furigana mode renders ruby invisibly until the word is hovered.
• New "Color words" option: keep Yomu's coloring on every card state, or only on new/not-in-deck words (everything you're already studying inherits the page's own text style).

Changed

• The default furigana mode is now Automatic: with a Jiten/JPDB key or Anki connected it hides furigana for words you know (known, due and failed states by default — your request), and falls back to difficult-kanji-only furigana otherwise. An explicitly chosen mode is never changed.

[0.6.152] - 2026-06-12

Fixed

• Mix and playlist stacks no longer clutter the filtered YouTube feed: the modern lockup-style cards (stacked collection thumbnails linking to watch?v=…&list=RD…) evaded the playlist detection because their links look like ordinary videos and their Japanese titles (ミックスリスト …) passed the language filter. They are now always treated as non-video items and hidden while filtering. (Live-verified on a signed-in feed: visible mix stacks went from several to zero.)

[0.6.151] - 2026-06-12

Added

• Swipe-to-grade on the study page can now be turned off: Settings → Study → "Swipe cards to grade (left = fail, right = pass)".

Fixed

• Study-page keyboard shortcuts (Space to reveal, 1-9 to grade) now work from a fresh page load and after every button press. They previously only worked while focus happened to sit inside the page content, which broke after each re-render — most visits had effectively dead shortcuts.
• The inline keyboard hint pills (like "Space" under the Reveal button) disappear permanently after the first time a shortcut is used — shortcuts stay discoverable in settings. Touch-only devices never see them (already the case).

[0.6.150] - 2026-06-12

Fixed

• Sparse YouTube shelves after filtering: when the Japanese-content filter leaves a shelf (e.g. the Shorts row) with fewer than three visible items, Yomu now pages the shelf forward (its own "show more" / next control) so YouTube hydrates more items for the filter to keep — instead of the shelf shrinking to one or two cards. Capped at four pages per shelf and throttled so a genuinely non-Japanese shelf is not paged forever. (The shelf "show more" hydration was verified live on a signed-in feed: 5 → 9 rendered items per click.)

[0.6.149] - 2026-06-12

Fixed

• YouTube feed gaps: shelf carousels (Shorts, news and similar rows) render their items lazily, so when the Japanese-content filter collapsed the rendered neighbours, YouTube's still-unrendered slots slid into view as blank full-height boxes. Those unhydrated slots now stay out of the flow until YouTube fills them, removing the empty holes in filtered feeds. (Verified on a signed-in feed: blank shelf boxes collapsed, grid reflows cleanly.)

[0.6.148] - 2026-06-12

Added

• Pausing a video now OCRs the paused frame (user request): the frame is snapshotted over the player and read with the same OCR pipeline as images, so on-screen Japanese (captions burned into the video, slides, signs) becomes readable words while paused. The overlay disappears the moment playback resumes or the video changes. Respects the OCR provider/enabled settings, skips small players, and can be turned off with the new ocrVideoPauseFrames setting. DRM-protected frames that cannot be captured are silently skipped.

[0.6.147] - 2026-06-12

Fixed

• Study-page sentences no longer show furigana over words you already know (known, mature, never-forget and similar states), even when the global furigana mode is "all" — the study page is a review surface, so readings only appear for words still being learned. Explicit stricter modes (off, difficult kanji, known status) behave exactly as before.

[0.6.146] - 2026-06-12

Fixed

• Study page swipe-to-grade works on iOS Safari: the card now declares touch-action: pan-y, so Safari no longer claims horizontal pans (which cancelled the gesture before the grade threshold), and the active pointer is captured once a drag starts so fast flicks cannot drop their release.
• The swipe edge glows are now real grading affordances: red on the left (fail) and green on the right (pass), visible only while dragging and scaling with drag distance. The old always-on green glows on both edges of review mode and the static green strip at the top of the page (which read as a stuck swipe indicator) are gone, including the first-paint copy in the page template.
• The study session clock stops itself if its environment disappears mid-tick instead of throwing from the timer.

[0.6.145] - 2026-06-12

Changed

• Study page internals: the search detail-expansion renderers (header, definitions, loading state, fallback definition sources, inline kanji section shell) moved out of the controller hotspot into the search view module as pure helpers — no behavior change.

Fixed

• The study-page Playwright smoke harness works again (it had silently drifted after the session-progress, grade-target and stop-at-batch reworks): it now also pins the failed-card requeue loop, the undo-review button after a Jiten grade, the JPDB all-decks union with due-time lookups, and the "Batch complete" breather, browser-level. Its API mocks are stateful (a reviewed card leaves the due pool) and the impossible combined JPDB+Jiten credential scenario was replaced with a JPDB-only one (by design, a Jiten key takes over the single API credential slot).

[0.6.144] - 2026-06-12

Changed

• Page scanning: a new scan now cancels the previous one between batches (fast scrolling or navigating no longer wastes time parsing regions that already left the screen), removed or changed text is skipped before parsing, and very long paragraphs no longer stall a scan turn.

[0.6.143] - 2026-06-12

Changed

• Anki setup: the status line now shows an explicit state badge (Off / Probing / Unreachable / Connected / Scanning / Mapped / Ready) and, after a library scan, lists each suggested field mapping with how confident the match is (high match / fuzzy match / unmapped) — no more guessing what the automatic setup concluded.

[0.6.142] - 2026-06-12

Added

• Failed cards now loop within your study session, jpdb-style: grading Nothing/Something (or Fail in two-button mode) puts the card at the back of the current queue so it comes around again until you pass it. Hard still advances the card.

[0.6.141] - 2026-06-12

Added

• Stop at the end of each batch (optional, off by default): when your loaded review batch runs out, the study page now shows a 'Batch complete' summary with your review count and session time, and waits for you to press Continue instead of silently fetching more cards.

[0.6.140] - 2026-06-12

Added

• Undo review (Jiten): after grading a Jiten card on the study page, an Undo button appears for a few minutes — it reverses the review on Jiten and puts the word back in front of you, unrevealed. (JPDB's API and AnkiConnect expose no review-undo, so this is Jiten-only for now.)

[0.6.139] - 2026-06-12

Changed

• Internal hygiene: dead-code analysis is back to zero findings (post-refactor internals privatized, companion-seam types documented as intentional).

[0.6.138] - 2026-06-12

Changed

• Anki card audio: media files are now cached after the first play, so replaying a card's audio (or revisiting the same note) no longer re-downloads it from AnkiConnect each time.

[0.6.137] - 2026-06-12

Fixed

• Study page: the session clock now stops when you leave the Word tab, and daily-goal time only counts seconds actually spent studying words (it previously kept ticking on the Stats/Search tabs and while idling).

[0.6.136] - 2026-06-12

Changed

• Internal: study-page search result rendering moved into its own module (smaller controller, no behavior change). Jisho audio source-selection verified to already match Yomitan's exactly (live-checked against jisho.org); Jiten deck-based word styling parked pending observable study-deck data.

[0.6.135] - 2026-06-12

Changed

• Toasts redesigned: multiple notifications now stack neatly instead of overlapping, fade in and out (animations respect reduced-motion preferences), and repeated identical messages extend the visible toast rather than piling up duplicates.

[0.6.134] - 2026-06-12

Added

• Mass review (Jiten): press Alt+M (configurable under Settings → Shortcuts) to review every visible due/learning Jiten word on screen as Good in a single batch — the same 'review everything on screen' flow JitenReader users know. Word colors refresh right after.

[0.6.133] - 2026-06-12

Changed

• Internal cleanup: learner-glossary summarization is now a single shared helper used by both the dictionary grouping and study-page meaning cleanup; the two remaining cross-boundary duplications (build scripts vs runtime) are documented as intentional with keep-in-sync notes.

[0.6.132] - 2026-06-12

Added

• Study page Search tab (user-requested '2D reviews'): pick a deck from the dropdown to browse its full word list in your actual review-queue order; typing then narrows within it, with words starting with your input ranked first. State chips are now multi-selectable, a compact sort row offers queue order / A→Z / frequency with an ascending/descending toggle, and bulk-select checkboxes only appear after tapping Select. Plain dictionary search stays the default when nothing is selected.

[0.6.131] - 2026-06-12

Added

• Study page (user-requested): jpdb-style combined kanji+word queue — locked words now serve their kanji card first, and the word unlocks once the kanji is learned. A new setting, 'Study kanji before unlocking words' (on by default), lets kanji-skippers turn this off and study locked words directly; the Kanji tab stays available for isolated kanji study, and toggling never affects your progression.
• The welcome screen now mentions the built-in study page.

[0.6.130] - 2026-06-12

Fixed

• Study page, JPDB (user-reported): 'No reviews ready — showing practice words' with due cards waiting, and endless 'searching for words', are fixed — the all-decks listing no longer scans your decks in tiny sequential chunks (it unions every deck in parallel and resolves them in one bulk request) and gets a longer time budget.

Changed

• Study page, JPDB: the review queue now sorts by jpdb's own due_at timestamp, so the order matches jpdb Learn exactly — the next word on jpdb is the next word here. Cards without a due time (new/locked) follow in deck order.

[0.6.129] - 2026-06-12

Added

• Study page (user-requested): a live session timer now ticks during study in every mode, and a daily study goal setting (default 60 minutes, set 0 to disable) shows your progress next to it ('12/60 min', with a checkmark when reached). Time only counts while the study tab is visible.

[0.6.128] - 2026-06-12

Fixed

• Study page (user-reported): the front of a card no longer spoils the answer — the target word's furigana in the example sentence stays hidden until you reveal (other words keep theirs, like jpdb's Learn).
• Study page (user-reported): Immersion Kit controls are compact — the 'IMMERSION KIT' label is gone (clip title is enough; provider shown on hover), and prev/audio/next sit inline with the title instead of on their own oversized row.
• Stats (user-reported): the last bar of the daily activity chart no longer looks permanently 'selected'; the outline now only appears when you actually pick a day.
• Study page (user-reported): without any API key the Show-only state filter dropdown is hidden — keyless dictionary study has no card states to filter.

[0.6.127] - 2026-06-12

Fixed

• YouTube channel shelf (user-reported): the one-line channel description no longer flashes into the full multi-line bio after parsing (newlines became
  and escaped the ellipsis clamp); this also removes the furigana-overlap seen inside expanded bios.
• YouTube feed (user-reported): community posts from non-Japanese channels are now filtered by their post text on desktop and mobile; a channel's own Posts page stays unfiltered, and the mobile 'read more' (続きを読む) button no longer counts as Japanese content.
• Subtitles (user-reported): foreign-language captions (e.g. Arabic) are no longer mirrored while the selected Japanese track is still loading — no more wrong-language flash before Japanese subs appear.

[0.6.126] - 2026-06-12

Fixed

• Mobile YouTube (user-reported): scroll position no longer jumps as videos load in — iOS Safari has no scroll anchoring, so filtering a card above the viewport shifted the page; the filter now keeps the element you are looking at stationary by compensating the scroll position.
• Subtitle player (user-reported): rerender loop fixed — identical render ticks no longer rebuild the subtitle DOM, which was wiping pitch-accent and word-state highlights moments after they appeared and causing flicker and page lag.

[0.6.125] - 2026-06-12

Fixed

• YouTube channel shelf (user-reported): Subscribe said 'Subscribed' but the account was never subscribed — InnerTube write calls now carry the signed-in SAPISIDHASH authorization (YouTube silently applied unauthorized writes to the anonymous visitor session). When signed out, the shelf now says 'Sign in to YouTube to subscribe to channels.' instead of faking success.
• Subtitle popover (user-reported): the track-status line ('2 subtitle tracks detected') could be appended to the sentence sent to translation when a cue ends without punctuation; subtitle player chrome is now excluded from sentence context.

[0.6.124] - 2026-06-11

Fixed

• AnkiMobile handoff (user-reported): adding a card from iOS failed with 'no such note type id よむ+Japanese' — the handoff URL encoded spaces as '+', which AnkiMobile does not decode. The x-callback URL now uses %20 encoding for the note type, deck, tags, and every field. AnkiDroid was unaffected.

[0.6.123] - 2026-06-11

Added

• When AnkiConnect can't be reached, the settings status now diagnoses WHICH step failed instead of a generic 'not connected': if AnkiConnect is running but rejecting this site (the usual Firefox case), the message names the exact origin to add to webCorsOriginList in the add-on's config — with the matching Japanese copy; only a genuine network failure keeps the 'open desktop Anki' guidance.

[0.6.122] - 2026-06-11

Fixed

• Reviews or edits done inside Anki itself no longer leave stale word colors when the collection's card count happens to stay the same: the status index's count check now also sweeps recently-edited cards by modification time (AnkiConnect cardsModTime) and refreshes when anything changed since the last sync. Older AnkiConnect versions without cardsModTime keep the previous behavior.

[0.6.121] - 2026-06-11

Changed

• JPDB review-order audit (last open P0): live-review mode mirrors jpdb.io's own session exactly — including locked-kanji interleave — and the default auto mode prefers it whenever a review tab is connected. The API cannot reproduce that order (it exposes deck order only, no due data), so the 'API vocabulary only' setting now says so in its label instead of implying parity.

[0.6.120] - 2026-06-11

Added

• Kana-run word identity (the last piece of the kana-run parity P0): when tapping part of a kana-only word like にほんご resolves the full word, every rendered fragment in the run now takes on the resolved word's identity — so grading, mining, or a cross-tab state change recolors the whole word everywhere, not just the fragment you tapped. Surface-mismatch fails closed; ruby annotations are handled.

[0.6.119] - 2026-06-11

Added

• Live JPDB reviews now complete the cross-tab card-state bus: after grading through the review bridge, Yomu reads the card's true post-grade state back via the API (the bridge card id carries the real vid/sid from the review URL) and broadcasts it, so the same word recolors in every open tab — exactly like API-graded and Anki-graded cards have since 0.6.82. Without an API key the live grade still lands; no state is guessed.

[0.6.118] - 2026-06-11

Changed

• Kanji study mode's extraction from Anki cards is now pinned by tests (word cards contribute their kanji with the Anki note linkage but stay ungradeable as kanji; standalone RTK-style kanji notes win dedup and keep their keyword) — closing the 'unverified/partial' backlog ticket.
• Removed the vite-plugin-monkey client runtime import: the document-mounted monkey windows it exposed are already discovered by the existing scan (P4 backlog cleanup).

[0.6.117] - 2026-06-11

Fixed

• Example sentences (Immersion Kit examples in popovers and study-card sentences) no longer leave words without pitch accent when you have a JPDB API key: local-dictionary pitch now fills in wherever the API returns none, in both surfaces — it was already fallback-only, so API pitch is never overridden. Closes the example-sentence pitch-gap ticket.

[0.6.116] - 2026-06-11

Changed

• Anki status lookups now search your mapped expression/reading fields first (e.g. Word:読む OR Kana:よむ), so text in sentence or definition fields can never create a false 'in Anki' match. The old whole-note search still runs, but only as a low-confidence final pass for words the field-scoped search couldn't find — keeping nonstandard and unmapped decks discoverable. Closes the last slice of the field-scoped lookup ticket (rest shipped 0.6.76).

[0.6.115] - 2026-06-11

Added

• My Cards browser: Anki rows now show a due-in bucket (Due / ≤1d / ≤7d / ≤30d) answered by Anki's own scheduler search — the last open study-hub parity ticket. Jiten and JPDB rows stay blank rather than guessing: their APIs expose no per-card due timestamps.

[0.6.114] - 2026-06-11

Fixed

• Mobile: on sites with their own bottom action dock — Jiten's study page with its grade bar and Blacklist/Master/More row — the よむ puck no longer overlaps the native controls; it rises above the dock (live-verified on an iPhone-size viewport with the split build).

[0.6.113] - 2026-06-11

Changed

• ADR-0003 phase 2 complete: the kanji drilldown surface (origin graphs, KanjiVG stroke data, RTK keywords, JPDB kanji pages, practice doodle, and their render layer) now lives entirely in the Yomu Kanji/Study companion. The core userscript drops from 1,961,982 to 1,825,206 bytes — 136,776 freed, headroom up from 38 KB to 175 KB — via a single registry seam in popup/render.ts, so no call sites changed. Extension/self-contained builds and the study page register the companion modules at build time and behave identically; on Greasy Fork the companion is @require'd with an SRI hash, and if it ever fails to load, kanji drilldowns degrade to dictionary-only sections instead of breaking.

[0.6.112] - 2026-06-11

Changed

• ADR-0003 phase 1: a third Greasy Fork companion library, Yomu Kanji/Study (kanji origin graphs, KanjiVG stroke data, RTK keywords, JPDB kanji pages), now builds and ships alongside the Settings Surface and Video companions, with its registry slot wired into the core. Phase 2 (severing the core's direct imports, freeing ~147 KB toward the 2 MB Greasy Fork limit) follows; behavior is unchanged in this release.

[0.6.111] - 2026-06-11

Added

• The in-page study-deck selector now works for the Anki source too: pick All decks or any single deck (subdecks included, disabled-deck toggles still honored) and the Anki review queue rescopes immediately — completing per-deck study for both providers that support it (JPDB shipped in 0.6.97; Jiten is parked until its API exposes study decks to scope).

[0.6.110] - 2026-06-11

Added

• New (unseen) Anki cards now show due-in previews too: Again/Hard/Good/Easy derive from your deck's actual learning steps via getDeckConfig ('<1m', '<6m', '<10m', '4d') — the same numbers Anki shows on a fresh card, one config fetch per distinct deck. With 0.6.108/0.6.109 this completes the due-in story across study queue and popover for every provider that has the data.

[0.6.109] - 2026-06-11

Added

• Due-in previews now appear on the popover grade row too — the buttons you grade page words with — matching the study page and Jiten's 1m/10m/16d/28d pattern. Anki existing-note grade rows get the same computed Hard/Good/Easy previews as the study queue.

[0.6.108] - 2026-06-11

Added

• Anki review cards in the study page now show due-in previews on the grade buttons (Hard/Good/Easy — '12d', '25d', '1.1mo'), computed the same way Anki's own answer buttons are: current interval x hard factor / x ease / x easy bonus. AnkiConnect only exposes the real strings inside its GUI reviewer, so queue cards never had them. Learning and new cards stay blank instead of guessing your deck's step config — no invented numbers on a trust surface.

[0.6.107] - 2026-06-11

Fixed

• jiten.moe/srs/study: Yomu's Immersion Kit and dictionary sections now appear inside the revealed study card — right after Jiten's own Kanji breakdown / Composed of — instead of detached at the top of the page. They also no longer mount during the question phase, where showing dictionary entries spoiled the answer. Live-verified on a signed-in study session.

[0.6.106] - 2026-06-11

Added

• Stats: the Anki Due-now metric now carries a real upcoming-review forecast — 'Next 7d / Next 30d' — answered by Anki's own scheduler (prop:due search), scoped to your enabled decks, suspended cards excluded. No client-side guessing at Anki's queue encodings.
• Study card backs now show the 'Part of the X deck' line for every provider: JPDB (deck-scoped queues and the live bridge's scraped line), Anki (the card's owning deck), and Jiten (the study batch's source deck).

[0.6.105] - 2026-06-11

Added

• When the study queue is scoped to a specific JPDB deck, every card's back now carries jpdb.io's 'Part of the X deck' membership line (localized). Live-bridge cards keep the line scraped from jpdb.io itself when present.

[0.6.104] - 2026-06-11

Added

• iPad/tablet landscape: the study card now uses the width — word and sentence on the left, revealed answer (reading, meanings, Composed-of, Immersion Kit) on the right, with session and grade controls spanning below. Applies on coarse-pointer landscape viewports 1000px and wider; verified against a real 1180x820 render.

[0.6.103] - 2026-06-11

Added

• The study-deck selector now shows each JPDB deck's progress like jpdb.io's Learn page: vocabulary count and known-coverage percentage ('誕生日 · 39 · 65%'), fetched through the same list-user-decks call (no extra requests).

[0.6.102] - 2026-06-11

Added

• My Cards browser now has Jiten-style bulk actions: a select-page checkbox plus per-row checkboxes, with Blacklist and Never forget acting on everything selected. Each card goes through the same action path as the popover buttons, so provider behavior is identical (JPDB deck moves, Jiten local workaround, Anki suspend/tag) and the list reloads with the new states afterwards.

Fixed

• Stats learning-progress total now reads 'Total known non-redundant vocabulary', matching jpdb.io's Learn page word for word (the count already excluded redundant/blacklisted cards).

[0.6.101] - 2026-06-11

Added

• The Word tab now has jpdb.io's 'Show only' filter controls as a compact select beside the study-deck scope: Study (the scheduled queue), All, or any single state — New, Learning, Due, Failed, Known, Never forget, Suspended, Locked, Blacklisted, Redundant. Picking a settled state merges the full card pool in (the scheduled loader intentionally drops known/blacklisted cards), so filters like Known and Blacklisted really browse everything — matching the deck-browse filters on jpdb.io. The choice persists per device.

[0.6.100] - 2026-06-11

Added

• The My Cards browser on the idle Search tab now includes your Anki cards alongside Jiten and JPDB — state chips, search, and rows span all three providers (study-hub parity SH-3 v2). Anki joins only the browser pool; the stats page keeps its dedicated Anki source so nothing double-counts.

[0.6.99] - 2026-06-11

Added

• Study controls now advertise their keyboard shortcuts the way jpdb.io and Jiten do: the reveal button shows Space and each grade button shows its digit (1..5 in rendered order). Hints are decorative (aria-hidden) and disappear on touch devices. Discovered by running the live Jiten study journey with the userscript injected — Jiten's own 'Show Answer — Space' hint set the bar.

[0.6.98] - 2026-06-11

Added

• Revealed word cards in the study page now show a 'Composed of' line — the word's component kanji as chips with their RTK/JPDB keywords, tappable to open the full kanji drilldown — matching the back of jpdb.io's review cards (study-hub parity SH-4). Kana-only words skip the line.

[0.6.97] - 2026-06-11

Added

• Word tab now has an in-page JPDB study-deck selector (study-hub parity SH-6): pick All vocabulary or any of your JPDB decks right where you study — mirroring jpdb.io's per-deck Learn — and the review queue rescopes immediately through the same scheduled-cards loader. The choice persists per device alongside the other study-page state; the settings default still applies until you pick something. Mobile-safe (16px select, no iOS zoom).

[0.6.96] - 2026-06-11

Fixed

• Mobile (m.youtube.com) Shorts: swiping no longer lands on non-Japanese shorts. The 2026 mobile player is a JS carousel (shorts-page > shorts-carousel) with no per-item card elements, so the card filter could never classify the active reel — the active short's title is now read from the player overlay, classified through the same rules as feed cards, and non-Japanese actives are skipped by clicking the carousel's hidden accessibility 'next video' button (locale-independent, verified against the live mobile player).

[0.6.95] - 2026-06-11

Added

• Stats tab Today parity (Jiten shape): a 'Due now' tile with the estimated clear time, and the reviews-today tile now shows '+N new' for cards introduced today. (SH-7 core.)
• My Cards browser: with a state chip active, typing in the search box now searches your own cards (spelling or reading) instead of the dictionaries — pick the All chip to return to dictionary search. (SH-3 v2.)

[0.6.94] - 2026-06-11

Added

• jpdb.io keyboard parity in the study tab: with the answer revealed, the 1–5 digit keys press the grade buttons in order (1=Nothing … 5=Easy on JPDB-style bars; 1=Fail 2=Pass on two-button bars). Space/Enter reveal and arrow-key navigation were already there. Digits never fire on the card front or while typing in a field. (Study-hub parity SH-8.)

[0.6.93] - 2026-06-11

Added

• Live JPDB review cards in the study tab now show jpdb.io's own deck-membership line ('Part of the Persona 5 deck (3x)') on the back, scraped from the review page through the bridge — the back of the card reads the same as on jpdb.io. (Study-hub parity SH-4, first slice.)

[0.6.92] - 2026-06-11

Added

• The study page's Search tab now opens as a 'My Cards' browser when Jiten or JPDB is connected: your whole SRS pool with state filter chips in JPDB's Show-only order (All/New/Learning/Due/Failed/Known/Never forget/Blacklisted/…) including live counts, 50-per-page rows showing the word, reading, first meaning, state badge, and frequency rank — and tapping a row opens Yomu's full lookup (dictionaries, Immersion Kit, mining) for that card. Typing still searches dictionaries as before. Touch-sized rows; meanings column folds away on narrow screens. (Study-hub parity SH-3 v1.)

[0.6.91] - 2026-06-11

Fixed

• Study cards backed by JPDB now front JPDB's own example sentence — exactly what jpdb.io shows on its review front — with Immersion Kit sentences as the fallback for cards the provider gives no sentence, instead of replacing JPDB's content. Non-JPDB cards keep the Immersion-Kit-first behavior. (Study-hub parity SH-5.)

[0.6.90] - 2026-06-11

Added

• Study session label now mirrors JPDB's Learn summary: when the due pile mixes vocabulary and kanji it shows the split ('20 words · 92 kanji'), and unseen items add an 'N new' count — alongside the existing Done/Left/Due session progress. Second slice of the study-hub parity plan.

[0.6.89] - 2026-06-11

Added

• Stats tab now opens with a JPDB-style learning-progress table — 'Learning | You know' columns with the Words row (total, learning count, known count and percentage) and a 'Total known vocabulary' line — for whichever provider is selected (Jiten, JPDB, Anki, or combined).

[0.6.88] - 2026-06-11

Added

• New mobile option: 'Mobile sheet: close button on the left' — parks the lookup sheet's close button on the left edge for one-handed reach (Jiten Reader parity). Off by default; localized.

[0.6.87] - 2026-06-11

Added

• New reading option: 'Hide styling on JPDB-redundant words' (Jiten Reader parity). When enabled, words JPDB marks redundant keep tap/hover lookup but render as plain page text — no state color, underline, or highlight. Off by default; localized.

[0.6.86] - 2026-06-11

Fixed

• Mining a card with a captured context image or audio into a Jiten or JPDB deck no longer drops the media silently: when no Anki note is created alongside (Anki co-mining off), the success toast now adds 'Captured image/audio stays in Yomu — this service has no media API.' (localized). Jiten and JPDB deck APIs cannot store media; now the UI says so.

[0.6.85] - 2026-06-11

Fixed

• Live JPDB review bridge: reveal and grade clicks now target jpdb.io's stable control ids (#show-answer, #grade-1…#grade-5) first, falling back to text matching — verified against a signed-in jpdb.io/review session, whose buttons carry ✘/✔ prefixes that pure text matching could trip over. Investigated next-review intervals for the live bridge and closed it as not possible: jpdb.io renders no interval data on the review page and has no setting to enable it.

[0.6.84] - 2026-06-11

Fixed

• Faster first colorization on busy pages: local pitch-accent lookups (IndexedDB, no network) now run 8 at a time instead of 2 during the cold-start backlog, and the first Anki status scan sends its findNotes batches as 50-term requests with three in flight instead of one big 120-term request — Anki stays responsive and page words color sooner.

[0.6.83] - 2026-06-11

Fixed

• YouTube channel suggestions shelf: being subscribed to every curated channel is now an explicit, celebratory state — the subscribe-all button reads "All 100 subscribed ✓" and disables, the subscribe-visible button hides when the compact view is empty, and the status line distinguishes "all shown channels subscribed — browse all for more" from "subscribed to all 100". The shelf no longer offers Subscribe all 100 against an empty list.

[0.6.82] - 2026-06-11

Added

• Cross-tab card-state propagation: grading, mining, or changing a card's deck state in any Yomu surface — the study new tab or a page popover — now recolors the same word in every other open tab immediately (GM storage change signals across origins, BroadcastChannel within an origin). Previously only the acting tab updated and other tabs waited for a rescan.

[0.6.81] - 2026-06-11

Fixed

• Anki notes tagged yomu-never-forget (the popover's never-forget action) now rank as never-forget everywhere Yomu colors words — popover lookups and the background status index — instead of flipping back to due/new whenever the card's queue state changed. JPDB parity: a word you marked as always-known stays that color.

[0.6.80] - 2026-06-11

Fixed

• New-tab Kanji tab: when the same kanji appears both as a real JPDB locked kanji card and as a candidate derived from a word, the study queue now keeps the locked card instead of whichever duplicate was seen first — locked kanji are scheduled SRS items in JPDB's own review flow, so the Kanji tab now reflects JPDB's selection.

[0.6.79] - 2026-06-11

Fixed

• asbplayer subtitle lines now appear already colorized instead of visibly recoloring after display: asbplayer pre-renders the whole track's cue HTML into its offscreen cache and moves the same DOM node onscreen when the cue is current, so Yomu now drains that cache in paced 12-cue batches (80ms apart) the moment it appears — every cue is parsed and colorized before it is shown, with or without a JPDB API key. The currently visible cue is prioritized into the first batch so it can never be starved by a long unprocessed backlog.

[0.6.78] - 2026-06-11

Fixed

• The video subtitle rail (previous/next subtitle, panel toggle) now appears and disappears in lockstep with the video player's own controls: when YouTube's chrome fades out during playback the rail hides with it, and it returns when the controls come back. On touch screens the old "always discoverable" overrides kept the rail pinned at 72% opacity through playback — those are gone, and sticky tap hover/focus can no longer hold the rail open (the stuck button is blurred when the chrome hides). Verified live on a real watch page.
• (Shipped in 0.6.77's artifacts, documenting here.) Logged-in YouTube home feed: filtered rows no longer leave gaps where videos could fit — the row-start margin compensation predated YouTube's 2026 lockup CSS (the gutter moved onto the grid container), so it overflowed and wrapped rows early; the grid rebalance also re-runs when YouTube re-asserts its row flags, and the loading skeleton (ghost cards) now takes a full-width row at true card size instead of squeezing into the leftover space of a partial row. Verified live on the signed-in feed.

[0.6.77] - 2026-06-11

Fixed

• Live JPDB reviews no longer trust a vanished review tab: the jpdb.io/review page now heartbeats over the bridge and announces when it closes, and the study tab marks a silent bridge as stale within 30 seconds — clearing the lingering card so jpdb-live drops out of the source list instead of letting grades post into the void. Returning to the study tab re-requests the current card automatically.

[0.6.76] - 2026-06-11

Fixed

• The popover's new-card Anki preview now shows the exact fields a mining write will target when the configured note type is an existing non-Yomu model (e.g. Kaishi/Core decks): the preview runs the same field retargeting as the write path (saved mappings, then field-name aliases against the live model), so values are listed under the real field names — Word/Kana/Definition — instead of Yomu field names that would have been silently remapped at write time.
• Anki searches now match literally: deck names and lookup terms are escaped through one shared helper (* and _ act as wildcards even inside Anki's double quotes, so decks like Core_2k previously single-char-wildcard-matched and terms containing * over-matched). Nested-deck :: stays unescaped so subdecks remain included. This replaces two divergent escaping copies used by status probes, model queries, and new-tab deck queries.

[0.6.75] - 2026-06-10

Fixed

• New-tab Jiten reviews now refresh the graded card's state from Jiten after submitting (the JPDB path already did this internally), so the review summary and any later renders of the card reflect the real post-review state.

[0.6.74] - 2026-06-10

Fixed

• Grading, mining, or blacklisting a Jiten-backed word now recolors that word everywhere on the page immediately, matching JPDB and Anki: Jiten card state is refreshed from a self-parse after each SRS action (Jiten only exposes state through /parse), and a new action-controller hook re-styles all rendered occurrences of the card for every API provider — no rescan or reload needed.

[0.6.73] - 2026-06-10

Added

• Multi-word expressions (気合いを入れる) now show pitch accent in the popover as one labelled mini graph per component: the expression is segmented against your local dictionaries (greedy longest match, particles skipped) and each component's own accent is rendered side by side — a single component's pitch is never presented as the whole expression's. Appears when the expression itself has no pitch entry, pitch accent display is on, and local dictionaries are enabled.

[0.6.72] - 2026-06-10

Fixed

• CI dead-code check: JpdbClient.ping is reached through the settings dialog's dependency interface, which the static analyzer cannot trace — marked with the same used-via-interface pragma as parse() (the 0.6.71 CI run flagged it).

[0.6.71] - 2026-06-10

Added

• The settings dialog now shows a live JPDB connection status: the static "JPDB key set" line upgrades to "Connected to JPDB" or "JPDB did not accept the key" via the jpdb /ping endpoint, probed when the dialog opens and whenever the key changes — matching the live probes Anki and Jiten already had.

[0.6.70] - 2026-06-10

Fixed

• youtube.com/feed/channels no longer renders empty with the feed filter on: the page wraps every subscribed channel in one shelf, which the filter classified as a hideable "non-video container" and removed as a single item. Containers holding channel listings are never treated as feed noise now (applies to channel shelves anywhere, not just /feed/channels).
• YouTube home feed gaps after filtering (the user-visible holes and full-width empty bands): the rowless lockup grid keys row-start margins off per-item is-in-first-column flags that YouTube computes for the unfiltered feed and never recomputes. After each filter pass the scan now strips the stale flags and re-marks the first visible item of each row with a margin-compensation class (the NihongoTube technique), and rich sections whose entire content is filtered are hidden with their wrapper so they stop leaving full-width bands. Scrolling stays natural — filtered cards still leave the flow with zero content-shift animation.
• Firefox: "Not allowed to define cross-origin object as property" console errors from the site-language preference are gone, and the Japanese-language spoof now actually applies on Firefox — property descriptors (getters and values) are cloned into the page compartment before being defined on Xray-wrapped objects.
• Restored the scan-word CSS regression assertion that broke the 0.6.69 CI run (the chip nowrap fix changed the selector; the older test still expected the unscoped rule).

[0.6.69] - 2026-06-10

Fixed

• m.youtube action chips (共有/保存/報告/チャンネル登録) no longer collapse into stacked, overlapping characters that spilled into the description area: scanned words inside passive UI controls now respect the host's nowrap contract instead of re-enabling per-character wrapping (the prose-wrapping rule from 0.4.47 now excludes .jpdb-reader-passive-word).
• New-tab auto review no longer shows the same word twice when providers disagree on kana script (JPDB ベッド vs Anki べっど): the cross-provider dedupe key is now kana-insensitive.

[0.6.68] - 2026-06-10

Fixed

• Hovering Japanese text on buttons, chips, tabs, and other clickable controls now shows the dictionary popover. Hover lookups treat interactive elements as readable text (a hover popover does not steal the click); click-driven lookups still leave controls alone so clicking a button performs its action. Structural content (form fields, furigana annotations, SVG, editable areas) stays excluded everywhere.

[0.6.67] - 2026-06-10

Fixed

• YouTube native-caption fallback (the "YouTube native captions" track) now parses each caption during its 180ms stability window instead of after it, so captions appear colorized with furigana the moment they render instead of flashing plain first.
• Immersion Kit translations no longer render ALL-CAPS on pages whose styles set text-transform on ancestors: example sentences and translations now pin text-transform: none.

Added

• Playback-simulation regression tests pinning the just-in-time subtitle contract: continuous playback never displays a cue that is not already parsed and cached (40-cue walk under realistic parse latency), and a long seek re-warms the active cue plus its 10-cue lookahead within one warmup turn. These tests reproduce display-time parse misses if the warmup pipeline ever regresses.

[0.6.66] - 2026-06-10

Fixed

• Videos with only English subtitles now reliably offer (and auto-select) the auto-generated Japanese translation everywhere, not just on YouTube: native <track> subtitles (the Sky News-style player) now synthesize the translated track too, and the synthetic option no longer vanishes from the subtitle panel after the next track rescan — stale-track culling previously mistook it for a dead native/page track because it carries no TextTrack or page source of its own. The translation is removed together with its source track (orphan cascade) instead.
• A real Japanese track appearing after the machine translation was auto-selected now takes over as primary on all paths (native, page files, YouTube); the synthetic stays listed but never outranks the real thing.

[0.6.65] - 2026-06-10

Added

• Per-kanji furigana on all-kanji compounds: when the user's imported kanji dictionaries allow exactly one alignment of the whole-word reading (琉球藍 → 琉=りゅう・球=きゅう・藍=あい, including rendaku ぐみ and sokuon がっ surface forms), local-dictionary tokens now carry per-kanji ruby segments instead of one ruby spanning the compound. Ambiguous or unalignable readings keep the whole-word ruby — no guessing.

Fixed

• m.youtube rows (video titles, channel bylines) no longer show only furigana with the base text clipped away: single-line ellipsis rows (overflow hidden + text-overflow: ellipsis on nowrap text) are now treated as layout-sensitive on every site regardless of measured height, so words stay lookupable and colored but ruby that would grow the line is suppressed. Wrapping prose with an inert text-overflow declaration keeps its furigana.

Notes

• Restored a subtitle-CSS regression test that had drifted from the intentional 0.6.62 change (pitch underlines stay transparent until enrichment): the assertion now covers the transparent-by-default declaration, closing a gap where the full jpdb test shard had not been re-run to completion since.

[0.6.64] - 2026-06-10

Changed

• YouTube feed filtering now uses offscreen absolute positioning to hide non-Japanese cards (the technique used by NihongoTube) instead of display:none with max-height collapse animations. Filtered cards leave the layout instantly with no content-shift transition, YouTube's virtualized grid bookkeeping stays intact, and scrolling mostly-English feeds is dramatically smoother with fewer gaps.

Notes

• Expression pitch (e.g. 気合いを入れる): showing a single component's pitch as the whole expression would be misleading, so the correct fix is a per-component multi-graph popover — designed and queued in the backlog rather than faked.

[0.6.63] - 2026-06-10

Fixed

• Pitch underline colors now show during video playback instead of arriving "just in time" (or only when pausing): local pitch accent is included at subtitle parse time, so the colors are baked into each cue's HTML before it is displayed; the per-word pitch lookup is cached.
• Audio now plays even on pages whose CSP blocks blob/data media entirely (claude.ai): when the media element refuses the source, playback falls back to Web Audio decoding (AudioContext), which media-src does not govern.
• jiten.moe direct loads (e.g. /parse?text=…) no longer leave the Immersion Kit / dictionary addons stranded in the wrong place with no examples: when the SPA hands us only a coarse pre-hydration anchor, the mount is marked and automatically re-anchored once the real content exists (previously only a manual refresh fixed it).

Added

• smoke:late-content regression harness proving late-streamed SPA text (the Google Maps / claude.ai pattern) is parsed by the auto-scan observer; the live-site gap is narrowed to shadow-DOM/injection-timing suspects recorded in the backlog.

[0.6.62] - 2026-06-10

Fixed

• Audio on CSP-strict sites (e.g. claude.ai): fetched audio blobs were created without a MIME type, so media elements refused them ("Content-Type application/octet-stream is not supported", "No decoders"). Blob types are now inferred from the source URL (mp3/ogg/m4a/…), fixing word/sentence/Immersion Kit playback wherever blob URLs are permitted. (A Web Audio fallback for pages whose CSP blocks blob media entirely is tracked in the backlog.)
• Subtitles no longer render punctuation-only cues (a lone 。 or ?) as their own rows: sentence splitting merges trailing punctuation into the preceding line and contentless cues are dropped.
• Pitch-underline mode no longer flashes a colorless underline before the pitch class arrives — undetermined words keep a transparent underline.
• The channel-suggestions shelf now says "You are subscribed to all of these channels" when there is nothing left to suggest, instead of showing an empty list (or appearing broken).

[0.6.61] - 2026-06-10

Added

• Auto-mine on review (Jiten Reader parity, off by default — settings > Anki & mining): grading a word that isn't in your deck adds it to the mining deck automatically, so reviewing doubles as collecting.

Changed

• Battery: the subtitle player's internal clock now adapts — 250ms only while a video is actually playing, slower while paused, and 1.5s (with no per-tick work) when the tab is hidden or the page has no video; it snaps back instantly when the tab becomes visible.

Notes

• Jiten's API exposes no per-word deck-membership lookup (only per-deck word lists), so the "decks this word belongs to" popup list can't be built without N expensive requests; recorded as an upstream limitation in the backlog. Anki deck names already show on existing-note titles; JPDB membership already feeds the in-deck state dot.

[0.6.60] - 2026-06-10

Fixed

• Fixed "Subscribe all 100" (and "Subscribe visible") doing nothing: the already-subscribed detection scanned the whole channel payload, where unrelated shelves can carry subscribed flags, so every suggestion could be wrongly marked subscribed and the action list came out empty. Detection is now scoped to the channel's own header, and when everything genuinely is subscribed the shelf says so instead of silently ignoring the click.

Changed

• Further fallow code-health work (high-complexity findings 17 → 16): the reader-word click handler's guard chain extracted into readerWordClickSurfaces, and the per-word contrast variable application decomposed out of one large closure (word-contrast CRAP 71.3 → 26.5).

[0.6.59] - 2026-06-10

Added

• Anki deck-state parity: blacklisting an Anki-backed word now suspends its cards natively (suspended cards never come up for review and already have their own state color), and never-forget tags the note yomu-never-forget — both toggle on re-press and can be managed from inside Anki too.
• Jiten review history: Jiten exposes only per-day counters, so each study-batch load now snapshots them locally (up to ~400 days) and merges them into the stats page's activity/heatmap like the Anki history source.

Fixed

• Anki status colors now load in parallel with page rendering: the cached status lookup starts as soon as a parse finishes (overlapping the DOM work) instead of after it, removing most of the gray→color pop-in after a scan.

[0.6.58] - 2026-06-10

Fixed

• Furigana and word colors now appear much closer to instantly instead of loading in waves, from a root-cause audit of the latency chain:
  • Subtitle transcript warmup no longer paces the priority head of the queue — the visible rows and the lookahead window parse immediately, with the 120ms YouTube pacing applied only to the background tail (was ~1s before lookahead rows were ready).
  • Subtitle word-state colorisation retries every 1s instead of every 5s, so JPDB/Anki state colors land on cues during playback instead of lagging multiple lines behind.
  • Page scans apply parsed ruby in chunks of 48 targets (was 16), so the first paint covers the whole parsed batch instead of arriving in three visible waves.
  • The word-contrast pass runs once per container after a scan completes instead of once per 16-item chunk, removing repeated style recalcs (layout thrash) during page parsing.

Notes

• Remaining items from the latency audit (pre-warmed Anki status during render, persistent IntersectionObserver re-prioritisation, parallel cold-start pitch lookups) are recorded in the Snow Leopard backlog task with file:line evidence.

[0.6.57] - 2026-06-10

Changed

• Code-health pass driving the worst fallow CRAP findings down (high-complexity findings 26 → 17): the subtitle video-inset measurement now uses small style-snapshot helpers instead of one 100-line save/clear/restore function (was CRAP 299.6), and the Jiten kanji "words using this kanji" reading-filter/paging logic — previously triplicated across the popover, new tab, and hosted runtime — is now one shared module (jiten-kanji-words-actions.ts) so the three surfaces cannot drift.

Notes

• Live-verified with the corrected Jiten test credentials that the study-page addon mounts (Immersion Kit + imported dictionaries) work on jiten.moe/srs/study.
• Added the "Snow Leopard" quality-release task to the backlog: latent re-render/battery bugs, instant ruby/colorisation (page + subtitles), YouTube seamlessness, fallow health to zero.

[0.6.56] - 2026-06-10

Fixed

• Fixed furigana ruby disappearing from ordinary prose on sites whose page shell clips overflow (e.g. example sentences on jiten.moe): layout-sensitivity is now judged by the box's actual height — only short fixed boxes (under ~3 lines) that would really clip the taller ruby line suppress furigana, while tall clipped containers (page shells, scroll regions) keep ruby. One uniform rule across all sites, no per-site parsing.

Added

• Yomu enhancements (Immersion Kit examples, imported dictionaries, kanji tools) now also mount on the Jiten SRS study page (jiten.moe/srs/study), like the vocabulary/kanji/parse pages. Live verification pending — the Jiten test account is temporarily locked.
• Verified and pinned with a test that Jiten per-grade intervals from study-batch responses reach the new-tab grade buttons (data-grade-interval, aria/title labels).

[0.6.55] - 2026-06-10

Added

• New smoke:study-personas Playwright harness that simulates users on the study page across environments — keyless beginner (desktop + 390x844 mobile) and a reviewer whose Jiten API is unreachable — gathering timing/labeling/console feedback. All personas pass; the degraded-reviewer run verifies end-to-end that the "No reviews ready — showing practice words" notice appears exactly when configured review sources fail and never for keyless users.

Fixed

• Built-in starter study words are now labeled "Starter words" (入門単語) instead of "Dictionary", which confused keyless users who had never imported a dictionary.

[0.6.54] - 2026-06-10

Fixed

• Fixed Firefox extension errors "Not allowed to define cross-origin object as property" (XrayWrapper, content-script.js): restoring a page-shadowed window.dispatchEvent/addEventListener now clones the property descriptor into the page compartment via cloneInto. This also left window events broken afterwards, which is why popover sections such as Immersion Kit examples could hang on "Loading examples..." in the Firefox extension.

[0.6.53] - 2026-06-10

Fixed

• YouTube Shorts: with the immersion filter active, landing on a non-Japanese short now steps the player forward automatically so scrolling continues to the next Japanese short instead of parking on an English one.
• YouTube feed: filtered-out cards no longer leave the feed starved — the continuation loader is triggered earlier (within reach of the viewport, not only at the absolute page bottom, with the visible-video target raised and the throttle halved), and when triggering it requires a scroll the page position is restored, fixing the blank "scrolling in" gaps.

[0.6.52] - 2026-06-10

Fixed

• The subtitle rail no longer appears next to decorative or ad videos (e.g. Discord promos): our controls now follow the same rule as the player's own — they only show for videos that offer playback controls (native or a known player chrome) or that have subtitle data loaded.
• Removed the half-viewport cap on the bottom transcript drawer: it can now be pulled up to nearly the full window height.
• JPDB-redundant words can no longer be reviewed (parity with Jiten Reader); attempting it explains why instead of silently grading a card JPDB considers covered elsewhere.

Added

• New smoke:reader-sites Playwright smoke that injects the built userscript into live Ttsu Reader, Yatsu Reader, and a YouTube watch page, asserting clean install (no console errors), the reader FAB, and no stray subtitle rail. All three pass.
• Jiten stats coverage: new unit tests verify the stats page renders Jiten card-state breakdowns across SRS states and shows an explicit error (not a blank page) when the Jiten API fails.

[0.6.51] - 2026-06-10

Fixed

• The new tab now says so when it substitutes practice words for an empty or unreachable review queue ("No reviews ready — showing practice words") instead of silently showing non-review words as if they were your JPDB/Anki feed.
• Anki word colors no longer stay stale after reviewing in Anki itself: returning to the tab expires the count-validated status index (throttled to once per 2 minutes), so state changes that don't alter deck counts get picked up immediately instead of after up to 30 minutes.

Notes

• Verified the new-tab grade control already exposes an explicit JPDB/Anki/Both target whose label propagates onto the grade buttons — no silent dual-grading. Updated the trust-audit backlog accordingly.

[0.6.50] - 2026-06-10

Fixed

• Fixed clipped/invisible text in YouTube chrome: line-clamped and fixed-height boxes (video metadata rows, subscribe buttons) no longer receive furigana ruby — the taller ruby line was clipping the base text so only the furigana stayed visible, or wrapping fixed-height buttons. These words stay colorised and lookupable, just without ruby.
• Fixed subtitle word colorisation disappearing when stepping back to a previous line: cache-hit renders now re-apply JPDB/Anki state colors to the freshly inserted nodes, and the parse warm-up window keeps more history behind the active cue (6 behind / 10 ahead) so back-navigation always hits the cache.
• Fixed the subtitle rail and panel opener staying permanently visible on mobile: the controls now auto-hide after the idle delay (touch devices get no mouse movement to trigger it) and follow m.youtube.com's own player-controls overlay, so tapping the video reveals them together with YouTube's controls.

[0.6.49] - 2026-06-10

Fixed

• Fixed unreadable channel-suggestion text on mobile YouTube dark mode: m.youtube.com does not set the desktop html[dark] attribute, so channel names and Subscribe buttons fell back to light-theme colors (invisible on the dark background). The shelf now detects the page theme from the rendered background and applies the matching palette.
• Channel-suggestion descriptions are now parsed by Yomu like the rest of the page, so Japanese text in the shelf gets furigana ruby and dictionary lookups instead of being plain text.

Added

• Frequency dictionaries now have their own settings group with enable/disable, drag/button reorder, rename, and remove — the order controls which frequency badge shows first on cards.
• Added a keyless word-popover Playwright smoke (smoke:keyless-popover).

Fixed

• Fixed subtitle overlays and transcript rows blocking dictionary popover lookups on Japanese words (such as katakana ハグ) and particles (と/を).
• Fixed transcript panel drawer layout: prevented jpdb-subtitle-drawer-title and jpdb-subtitle-drawer-meta from being cut off on narrow screens by wrapping them and adding ellipsis truncation.
• Fixed an infinite requestAnimationFrame loop when aligning layout insets by checking if the inset metrics actually changed.
• Pitch accent now picks the variant that fits the word's contextual reading instead of always the first stored pattern, and falls back to the public JPDB pitch lookup when a Jiten/local pattern belongs to a different reading (e.g. dictionary form).
• Keyless mode: player-overlay subtitles are parsed and colorised through the same fast provisional tier as the transcript panel — previously cues without an API key waited on a slow JPDB-timeout path and loaded visibly late. Both surfaces share one cache, so no parse work is repeated.
• Keyless mode: tapping a kana word that exactly matches its card (e.g. タップ) opens the popover again instead of being swallowed by the kana-fragment suppression.
• Anki popover preview: dictionary definitions are spaced and sectioned again (the mined-note styles only existed in the Anki model CSS), and Anki bracket furigana (漢字[かんじ]) no longer shows raw brackets — the base text is re-parsed so ruby matches Yomu's kanji-only furigana everywhere else.
• The subtitle overlay rail (panel toggle) no longer appears on pages whose video is hidden via visibility or opacity.
• JPDB vocabulary payload parsing is now defensive against malformed rows.
• YouTube channel suggestions: already-subscribed channels are excluded (detected from YouTube page data), subscribing flips the row to an accessible "Subscribed ✓" state announced by the live status and then backfills it with a fresh suggestion, the compact shelf fills its rows, and the shelf follows YouTube's own dark/light theme so a light extension theme on dark YouTube stays readable.

Changed

• The default interface language for fresh installs is now Japanese; existing users' stored language preference is unchanged.
• The settings dictionary panel shows an expandable select-options summary instead of clipped text.

[0.6.47] - 2026-06-09

Fixed

• Fixed the player-overlay subtitle lagging behind the transcript panel: the overlay's parse warm-up refused the fast provisional parse tier the panel already used, so the overlay showed unparsed text until the slow authoritative batch returned. Both surfaces now share the same two-tier parse pipeline (provisional immediately, authoritative upgrade in place) with deduplicated in-flight requests.
• Fixed furigana covering whole words instead of just the kanji (e.g. 話す showed はなす over the full word); explicit readings over mixed kanji-kana bases are now trimmed so ruby only sits above the kanji.
• Fixed the transcript panel docked to the left sitting flush against the player; left docking now keeps the same gap as right docking.
• Removed the phantom "YouTube native captions" track from the track selector on videos that have no captions at all.
• Pressing the placement button for the currently active dock position now closes the transcript panel instead of doing nothing.
• Hid the left/right placement buttons on small screens, where the panel always uses the bottom drawer layout and the side buttons were inert.

Changed

• Pausing and resuming the video no longer rebuilds the pause transcript panel synchronously inside the media event; the work is deferred past the next paint so play/pause stays responsive on desktop and iPhone.

[0.6.46] - 2026-06-09

Fixed

• Fixed the browser-extension / self-contained build throwing ReferenceError: System is not defined (a code-split SystemJS chunk with no loader) which broke subtitle track loading; the subtitle translator is now statically bundled.
• Fixed the transcript panel's left/right placement toggle doing nothing on smaller screens: docking left was measured against the player's pre-shift right edge (so it always fell back to the bottom layout). Left and right now use the same available-width calculation, and the side layout stays available on narrower screens.

Changed

• The whole YouTube transcript is now parsed in the background ahead of playback (prioritising the lines around the current position) so furigana/ruby is ready in advance instead of appearing line-by-line as each cue becomes active.

[0.6.45] - 2026-06-09

Fixed

• Fixed YouTube watch page elements (such as the video player, description, owner, actions, comments, and playlist header) failing to display (appearing as blank spaces) by explicitly ignoring them in the YouTube immersion filter so they are never hidden or marked as pending.
• Fixed a TypeError (window.dispatchEvent is not a function) that occurred when closing or cancelling the settings dialog by routing all settings-related theme and dialog events through the unshadowed, hardened event dispatcher (dispatchWindowEvent).

[0.6.44] - 2026-06-09

Fixed

• Fixed YouTube page elements (such as the video player and description) failing to load by implementing a Trusted Types policy (yomu-reader-script) for injecting global language preferences, resolving CSP script blocks (This document requires 'TrustedScript' assignment).

[0.6.43] - 2026-06-09

Changed

• YouTube text scanning no longer excludes the page chrome, masthead, sidebar, and recommendation feeds. Interactive elements (the subscribe button, video-title links, chips, menu items, etc.) are now parsed with the passive-interaction pattern: hovering their Japanese text opens the dictionary while the click still triggers the native control. The video player chrome and the SPA watch title remain untouched.

[0.6.42] - 2026-06-09

Added

• Auto-generate a Japanese subtitle track by translating the English track when a video has no Japanese track available.

Fixed

• Lookup pills (dictionary links, copy, and frequency badges) now wrap onto multiple lines instead of overflowing/clipping off the edge of the popover.

[0.6.41] - 2026-06-09

Fixed

• Fixed the YouTube (and other video sites') player not resizing the video to match its frame after the subtitle panel closed or switched sides: the video kept the old inset size until a manual fullscreen toggle. The relayout nudge is now dispatched when the inset is cleared as well as when it is applied.
• Fixed the subtitle panel toggle being unable to close the panel when "Open side panel when paused" was enabled and the video was paused: an explicit close now sticks until the video plays again.
• The closed subtitle rail toggle now shows the remembered transcript placement icon instead of always showing the right-dock icon.
• Anki deck collapse state now persists across word lookups (the Anki section and each deck card remember whether you collapsed them).

[0.6.40] - 2026-06-09

Fixed

• Stopped uncaught errors from keyboard events that arrive without a key (seen on YouTube), which were aborting reader and subtitle keydown handling.
• Routed cross-origin requests (sentence translation, grammar data, jiten lookups/audio, and kanji radical images) through the userscript bridge so a strict page Content-Security-Policy such as jpdb.io's no longer blocks them.
• Moved jiten.moe kanji-page enhancements to the top of the page (just below the kanji header) instead of the bottom.
• Made Anki card furigana match the reader's ruby everywhere by re-parsing from base text; this also fixes single-kanji (RTK-style) cards that were rendering the character several times.
• Clicking a word inside the lookup popover now dives into that word so the Back button returns to the previous card.
• Hardened the YouTube subtitle side-panel layout: switching the panel between left and right no longer leaves a gap, exiting fullscreen restores the video, and a transient zero-size player no longer shrinks the video away.

Changed

• Improved test stability by cancelling the audio source-race timer when a source wins, resetting fake timers between tests, and retrying the few timing-sensitive audio/bridge tests.

[0.6.39] - 2026-06-09

Added

• Brought jiten.moe to parity with jpdb.io page enhancements: your dictionaries, Immersion Kit, kanji stroke practice, RTK/Heisig/Koohii, the component graph, uchisen, and the other sources now inject onto jiten vocabulary, kanji, and parse pages, including across in-app (SPA) navigation.

Fixed

• Stopped browsers and password managers from autofilling saved site credentials over the API-key field (and spilling the username into the settings search box) with a decoy credential trap, per-field opt-outs, and readonly-until-focus inputs.
• Synced the reader theme with jpdb.io and jiten.moe light/dark mode: an explicit theme is pushed to the host, auto follows the site, and toggling the site's own theme updates the reader to match.
• Let a click on a scanned word inside a native link (such as jpdb's clickable kanji) follow the link; hover and modifier+click still open the lookup popup.
• Kept the centered settings dialog in place when an Immersion Kit image finishes loading behind it, instead of flinging the panel into the top-left corner.

[0.6.38] - 2026-06-08

Fixed

• Removed the hosted userscript-bridge requirement for AnkiConnect so hosted study pages can use local AnkiConnect directly when CORS is configured.
• Kept the optional userscript bridge as a preferred transport when present, without showing bridge-only setup guidance.
• Added expanded/collapsed state to the new-tab mining target drawer handle so review controls expose the same state as the popover drawer.

[0.6.37] - 2026-06-08

Fixed

• Collapsed a word lookup stacked over the settings dialog back to settings when the settings panel is tapped, instead of trapping the lookup open (notably on iPad/touch).
• Kept the settings dialog footer pinned to the bottom of the drawer after an iPad rotation that grows the viewport; the scroll area now fills the drawer so the Save/Factory-reset row no longer detaches in Safari/WebKit.
• Stopped the YouTube immersion filter from hiding YouTube's virtualized active Shorts player, and added a script-tag fallback for reading the InnerTube client config when ytcfg is unavailable (Firefox userscript isolation).

Changed

• Showed six stable-randomized starter channels in the compact YouTube channel shelf instead of the same first five.
• Added a Google Search site-parser profile and excluded furigana ruby from hosted-reader surface text extraction.

[0.6.36] - 2026-06-07

Changed

• Moved verbose Anki/JPDB grade target details into the mining drawer gutter so review controls stay on one row without repeated labels or target pills.
• Loaded hosted docs companion userscript surfaces alongside the main runtime so settings/video features stay available on the docs install surface.

Fixed

• Kept generated API text-to-speech sources behind recorded audio in fallback mode and out of fallback preloads.

[0.6.35] - 2026-06-07

Fixed

• Kept JPDB and Anki new-tab fallback tests aligned with the current Jiten-first lookup order and smaller Immersion Kit sentence limits, restoring the GitHub CI shards.
• Preserved all playable Jisho audio sources from the exact matching audio element instead of dropping alternate formats.
• Let JPDB audio voice selection target explicit voices such as Female 1, Female 2, Male 1, and Male 2 without falling back to the wrong voice.

[0.6.34] - 2026-06-07

Fixed

• Refreshed stale Anki status-index hits through exact AnkiConnect lookups after review, add, merge, or update actions, avoiding broad deck rescans while keeping word colors and popover status current.
• Preserved custom Anki field mappings during automatic library scans when the live model still has those fields, replacing only stale roles with scanned suggestions.
• Fixed smoke-test path resolution after the script helper refactor, restoring hosted Chromium and Firefox Anki bridge verification.
• Rebuilt the docs, new-tab, and userscript release artifacts for the current Anki/JPDB integration fixes.

[0.6.33] - 2026-06-07

Changed

• Reused shared smoke-test cleanup and dictionary ranking helpers, keeping live/mobile smoke scripts and local dictionary lookup paths smaller and easier to maintain.
• Shared kanji source mount rendering and progressive similar-kanji loading between the reader and new tab, reducing duplicated popover code.
• Reused JPDB vocabulary URL generation and Anki lookup hydration helpers across dictionary and new-tab popovers.

Fixed

• Kept ruby visible inside Yomu-owned popovers, settings, and docs even when reader CSS is already installed, while still suppressing ruby on external clipped titles that would otherwise shift page layout.

[0.6.32] - 2026-06-07

Changed

• Deduplicated JPDB vocabulary URL parsing, pointer-rectangle helpers, popover height stabilization, Anki review targets, audio cache pruning, and rendered-word expansion display paths without changing lookup behavior.
• Kept the userscript release gate on Vite builds with readable Greasy Fork output under the 2 MB script limit.

[0.6.30] - 2026-06-07

Fixed

• Rechecked hosted AnkiConnect automatically when the userscript bridge becomes ready, avoiding stale "needs setup" status after the page bridge loads.
• Kept the new-tab JPDB/Anki source switch from reusing unreachable Anki cache entries, restoring fallback study words instead of "No review cards ready."
• Sped up the CI suite by sharding long generated settings, new-tab, subtitles, and JPDB tests with stable generated imports.

[0.6.29] - 2026-06-07

Fixed

• Restored full-word kana lookup on mobile and split inline text, so taps inside words like にほんご resolve the full Jiten/JPDB candidate instead of fragment entries.
• Kept hosted AnkiConnect on the userscript bridge for local Anki while allowing detailed clicked-word card hydration to retry past stale availability cooldowns.
• Restored new-tab fallback study words when no JPDB or Anki review cards are ready, and made JPDB/Anki source switching deterministic.
• Refined mobile new-tab layout, Anki card audio controls, and Anki opt-in defaults so fresh installs do not show loud mobile handoff actions by default.
• Added one compact popover review target selector for mixed JPDB plus Anki and duplicate Anki-card grading, with exact deck/card targets and one clear grade row.

[0.6.28] - 2026-06-06

Changed

• Simplified the settings API panel to one JPDB-or-Jiten key field, routing ak_ keys to Jiten and other keys to JPDB.
• Removed the separate Jiten connection test button and split API-key fields from the settings UI.

[0.6.27] - 2026-06-06

Fixed

• Kept stale Anki new-tab selections from showing an empty review queue after Anki is turned off, falling back to study words instead.
• Rechecked local AnkiConnect transport, exact Anki status hydration, no-key kana lookup, new-tab source toggling, and Jiten/Anki queue smokes against the current package.
• Kept the generated userscript self-contained and readable while preserving Greasy Fork size headroom.

[0.6.26] - 2026-06-06

Fixed

• Removed the transcript drawer header close button so the player rail icon is the single open/close control.
• Added in-drawer left, below, and right docking controls for transcript and track panels.
• Kept oversized side-panel resizing from falling back below when the player can remain usable, and capped below-panel height so wide layouts do not shrink the video too far.
• Kept Japanese-learning YouTube searches usable by preserving English-titled comprehensible-input videos and Shorts, while moving the channel starter guide to a gentler home-feed-only trigger.
• Kept iPad kana taps falling back to the full tapped surface when JPDB/pitch lookup is disabled, instead of replacing kana-only lookups with kanji spellings.

[0.6.25] - 2026-06-06

Fixed

• Keep Send to Anki off after fresh installs and factory resets so Anki stays opt-in and mobile popups do not reserve space for an app the user may not have.
• Show Anki setup guidance on the new-tab page when AnkiConnect is unavailable, instead of making the JPDB/Anki source toggle look like it did nothing.
• Refresh new-tab popup Anki status and cached card details after Anki add, merge, update, or grade actions.
• Prefer Kaishi-style word audio fields over sentence audio when automatically mapping existing Anki note types.
• Check Anki duplicates before attaching audio or images to a new desktop Anki note, keeping duplicate mining failures cleaner and avoiding unnecessary media writes.

[0.6.24] - 2026-06-06

Fixed

• Kept Anki truly off on fresh installs and factory resets by ignoring Anki/status color channels unless Anki mining or the Anki dictionary section is explicitly enabled.

[0.6.23] - 2026-06-06

Fixed

• Fixed rendered kana fragment lookups without a JPDB API key, so tapping split text like に, ほん, or ご in にほんご resolves the full JPDB word instead of opening misleading shorter entries.
• Hydrated Anki card details in the popup when AnkiConnect returns matching notes without a single primary note, and kept Anki status cache entries fresh after card updates or grading.
• Kept new-tab Anki card audio on the same icon-button pattern as dictionary audio and updated the Anki smoke test to cover the current JPDB locked/due review flow.

[0.6.22] - 2026-06-06

Fixed

• Made raw page taps build lookup context across split inline kana nodes, so tapping any character in text like にほんご opens the full JPDB dictionary entry instead of fragment lookups like ほん.
• Removed unused refactor leftovers and private-only exports so the CI dead-code gate passes again.

[0.6.21] - 2026-06-06

Fixed

• Kept Anki fully opt-in on fresh installs and factory resets, including Anki mining, the Anki dictionary section, new-tab Anki reviews, Anki+JPDB dual mining, and mobile Anki handoff.
• Made the mobile AnkiMobile/AnkiDroid handoff respect the main Anki toggle so turning Anki off removes the send-to-Anki route.
• Restored the mobile YouTube subtitle sidebar control when compact video controls idle, keeping the panel button visible and tappable.
• Rendered cached provisional subtitle ruby and pitch styling on the first primary subtitle paint while full JPDB parsing finishes.
• Tightened JPDB page Immersion Kit spacing so image controls sit with the title row, captions use the same subtitle styling as dictionary cards, and ruby captions are not clipped.
• Kept JPDB page alternate forms and compounds atomic after Yomu ruby is injected, so forms like おつかれさま and 疲れ stay full-word lookup targets without duplicate furigana loops.
• Fixed kana-only lookup without a JPDB API key by trying JPDB span candidates before fragment fallbacks, so tapping fragments like に, ほん, or ご resolves the full word にほんご.

[0.6.19] - 2026-06-05

Fixed

• Made Mokuro scans parse each manga text box as one target, preserving full words across OCR line fragments without adding ruby that changes the page layout.
• Fixed Mokuro vertical-text clicks so rendered Yomu word geometry is used before raw text fallback, preventing neighboring suffixes from opening instead of the clicked word.
• Allowed locked JPDB cards to be graded in the popup and new-tab reviewer while keeping blacklisted and never-forget cards blocked.

[0.6.18] - 2026-06-05

Fixed

• Kept scanned card titles, clipped text, and fixed-size overlay text lookupable without injecting furigana that can change the host page layout and cause blinking or swapping.

[0.6.17] - 2026-06-05

Changed

• Refactored Anki status/detail, JPDB, audio, settings-form, and userscript build helpers to reduce duplication while keeping the userscript self-contained for offline installs and extension packaging.
• Kept the generated userscript release checks shared and stricter so bundle-size, readable-code, and no-remote-executable-code constraints are enforced from one path.

Fixed

• Fixed Anki rendered-card audio on the new-tab page so card media buttons use the same icon control pattern as dictionary audio and route through AnkiConnect media playback instead of doing nothing.
• Fixed the mobile new-tab header so Word/Kanji/Search/Stats stays in a compact single row; the two-column rule now applies only to Stats mode.

[0.6.16] - 2026-06-05

Changed

• Made Anki status lookups scale by using cached status hits plus exact lazy lookups for visible words instead of broad routine collection scans.
• Refined the new-tab Anki/JPDB review flow with clearer Anki setup states, compact grade-target controls, and safer multi-card Anki grading.
• Improved Anki card rendering in the popup and new tab with original card content, clearer multi-entry separation, capped card typography, and separate lookup-vs-card audio handling.

Fixed

• Fixed Anki-disabled coloring so page highlights stay untrusted and silent when the Anki section is off.
• Fixed Jisho audio matching for exact term/reading audio and ambiguous homophones.
• Normalized saved settings so stale pitch highlight/underline combinations cannot leak across page scans or subtitles.
• Made the CI suite faster and less flaky by sharding generated settings tests and assigning deterministic Vitest API ports.

[0.6.15] - 2026-06-04

Fixed

• Restored first-run Anki mining to enabled by default so Anki status and the Anki dictionary section appear without extra setup when Anki is available.

[0.6.14] - 2026-06-04

Fixed

• Made AnkiMobile handoff use Anki's built-in Default deck for Yomu's built-in deck names when AnkiConnect is unavailable, avoiding missing-deck errors on iPad.
• Kept Default available in the Anki deck picker even before the desktop Anki library can be scanned.

[0.6.13] - 2026-06-04

Fixed

• Let settings saves close the dialog immediately instead of waiting on dictionary-style refreshes, avoiding apparent freezes after changing settings.
• Prevented pitch accent underline styling from leaking onto OCR and subtitle words that do not have pitch accent classes, including stale mobile settings states.

[0.6.12] - 2026-06-04

Added

• Added a live browser smoke check for hosted cache-busting, userscript bridge Jisho audio, local AnkiConnect, and hosted Anki bridge requests.
• Added compact new-tab grade target controls so cards present in both JPDB and Anki can grade both by default, or just JPDB / a specific Anki card when needed.

Changed

• New-tab Anki reviews now render the selected Anki card's original front and back HTML, with sanitized media/audio controls and capped card styling instead of oversized generated labels.
• Mobile Anki setup guidance now lives in the docs, including beginner-friendly Tailscale instructions for reaching desktop AnkiConnect from a phone or tablet.
• Local CI now prepares generated JPDB shards once and reuses them across bounded parallel shard runs, with deterministic Vitest API ports.

Fixed

• Preserved existing Anki word color on hover while allowing trusted modal misses and post-grade refreshes to repaint stale status.
• Made Jisho audio bridge QA deterministic and verified blob playback without relying on live Jisho or CloudFront during the smoke.
• Kept mobile settings inputs at no-zoom size after shared input styling, and removed the crowded mobile-handoff disclaimer from the Mining drawer.
• Kept first-run Anki mining and the Anki popover section enabled by default while removing the manual scan button from settings.

[0.6.11] - 2026-06-04

Changed

• Split the Anki, audio, settings, new-tab review, subtitle, and reader-runtime code paths into smaller modules so the mining and review experience has room to grow without one-file bottlenecks.
• Kept the userscript offline/self-contained for executed code by bundling ZIP support locally, removing remote @require JavaScript, and loading reader CSS through a userscript @resource.
• Added readable generated-whitespace compaction plus stricter release verification so the userscript stays under Greasy Fork's 2 MB limit without identifier or syntax minification.

Fixed

• Fixed Jisho audio fallback so hosted/no-bridge pages do not try the broken default public-proxy Jisho path before browser speech.
• Fixed dictionary preference saving after the settings-form split by routing dictionary priorities through the shared form reader.
• Preserved hover/status coloring, stale nested-parse guards, and generic page scanning across the parser/CSS refactor.

[0.6.10] - 2026-06-04

Fixed

• Fixed stale hosted userscript bridge markers so the live page retries AnkiConnect bridge setup instead of staying in a confusing disconnected state.
• Fixed hosted settings guidance so both automatic and manual AnkiConnect checks show the userscript setup path when the live page cannot reach local Anki.
• Fixed kana-only page terms such as よむ matching existing kanji Anki notes by reading, while keeping homophone guards for kanji terms.
• Fixed partial stale pitch-highlight settings that could leave mobile page words both highlighted and underlined by pitch after updating.
• Made Anki status-cache warmup prompt and scalable by indexing note fields plus review-state card sets instead of hydrating every card detail up front.
• Kept the settings puck reachable on coarse-pointer mobile devices even when stale saved settings had hidden it.
• Registered settings menu commands through both classic GM_registerMenuCommand and modern GM.registerMenuCommand userscript APIs.
• Narrowed new-tab service-worker cache cleanup to Yomu newtab caches only.
• Pointed docs install links at the stable hosted userscript URL so userscript managers keep normal install/update behavior.
• Rebuilt generated userscript/docs assets so mobile cache-busting metadata and the shorter mobile Anki handoff copy ship with the release.

[0.6.9] - 2026-06-04

Fixed

• Cleaned up stale saved color-channel settings from earlier builds that could leave mobile users seeing pitch accent as both a highlight and an underline after updating.

[0.6.8] - 2026-06-04

Added

• Added beginner-friendly desktop AnkiConnect/Tailscale setup guidance for using a home computer's Anki library from mobile devices.
• Added Anki tag chips in settings, with inline add/remove controls instead of a confusing free-form tags field.
• Added beginner-facing update and stale-cache guidance for the hosted new-tab page, mobile shortcuts, and userscript updates.

Changed

• Made the release test path use the same sharded JPDB runner as CI, avoiding the long single-file JPDB test stall.
• Shortened AnkiConnect setup messages in settings and moved advanced hosted/CORS guidance out of the crowded drawer copy.
• Let new-tab Anki reviews work independently from the reader's Anki mining toggle.

Fixed

• Fixed hosted Firefox userscript bridge event details so live AnkiConnect status can cross the page/userscript boundary.
• Fixed kana-only Anki status cache hits such as よむ matching existing cards whose expression is kanji and reading is kana.
• Fixed explicit popover Anki card hydration being skipped by an overly short background availability probe.
• Fixed Jisho audio lookup through the public proxy and matched Jisho audio by the requested reading instead of the first same-spelling source.
• Fixed automatic Anki library scans in settings, removed the manual scan button, and guarded stale scans from overwriting newer connection status.
• Fixed confusing new-tab JPDB to Anki source toggles when Anki is unavailable by showing Anki connection guidance.
• Fixed mobile/settings polish around no-zoom text inputs, full-width color swatches, donate accent styling, puck restoration, legacy pitch-highlight migration, and stale Yomitan import test isolation.

[0.6.7] - 2026-06-04

Fixed

• Retried hosted-page Anki bridge installation when userscript request APIs appear shortly after document start, so local AnkiConnect status can show up without a manual refresh.
• Guarded hosted Anki bridge refresh events so settings and popovers do not recheck Anki repeatedly after a successful bridge install.

[0.6.6] - 2026-06-04

Fixed

• Show hosted-page AnkiConnect setup as a normal setup state when the よむ userscript bridge is missing, instead of surfacing raw request-bridge errors.
• Exclude generated JPDB shard files from default Vitest discovery while keeping explicit CI JPDB shards runnable.

[0.6.5] - 2026-06-04

Fixed

• Re-render cached Anki matches as "details unavailable" when full AnkiConnect card-detail hydration fails, instead of leaving popovers stuck on a loading message.
• Added Jlab-style rendered-card QA coverage for Anki template HTML, media controls, font caps, and fallback-field hiding.

[0.6.4] - 2026-06-04

Added

• Added Anki review deck toggles in New tab settings after an existing-deck scan, with newly scanned decks included automatically and saved exclusions preserved.

Fixed

• Made settings relocalization faster by using direct form-control lookups and avoiding repeated select metadata rebuilds in large settings forms.

[0.6.3] - 2026-06-04

Fixed

• Kept failed AnkiConnect checks in the settings setup tone instead of presenting normal hosted-page connection setup as a hard error.
• Clarified desktop AnkiConnect versus mobile Anki handoff behavior across settings, docs, and smoke checks so users know which features require local Anki access.

[0.6.2] - 2026-06-03

Fixed

• Made the AnkiConnect settings error shorter and more actionable, moving hosted userscript bridge/CORS guidance into the normal setup help with the AnkiConnect add-on link.
• Reworded mobile Anki handoff limitations around the current desktop AnkiConnect requirement instead of speculative future bridge copy.
• Split CI tests across eight Vitest shards with an explicit timeout so large JPDB/new-tab coverage no longer leaves one long-running shard looking stuck.

[0.6.1] - 2026-06-03

Fixed

• Kept the hosted app on the userscript request bridge for local AnkiConnect, with clearer Anki setup guidance and a direct AnkiConnect add-on link in settings.
• Restored a readable Greasy Fork build by using a pinned fflate @require instead of post-build compaction/minification.
• Opened mobile Anki handoff immediately from card actions instead of waiting on hosted AnkiConnect/detail-provider probes.
• Improved existing Anki card rendering for RRTK/Core/Yomu-style templates by removing nested card scroll traps, capping imported template fonts, separating Anki media controls, and avoiding duplicated Anki fronts when the answer already contains the question.
• Added realistic Anki template QA fixtures and a Chromium smoke check for dark/light popover rendering.
• Split CI into typecheck, sharded Vitest, and build/docs jobs, and stopped the generated-userscript workflow from rerunning the full test suite after CI.

[0.6.0] - 2026-06-03

Added

• Added explicit lookup grading targets for JPDB and individual Anki cards so words that exist in both systems, or in multiple Anki notes/cards, can be graded deliberately.
• Added Anki scan confidence chips in settings field mapping so nonstandard decks make it clearer which fields were auto-detected with high, medium, or low confidence.
• Added many more local grammar rules, with rule explanations and examples now loaded from hosted grammar-rule data (English and Japanese) instead of being embedded in the script.
• Added passive scan coverage for dictionary hyperlink text, UI chrome labels after prose, compact onclick controls in whole-page fallback scans, and hosted video-player empty-state and control text, keeping links clickable while words stay tappable.

Changed

• Existing Anki rendered cards now keep multiple card sides separated behind collapsible Anki card headers, preserve Anki card bodies without extra Yomu labels, and cap oversized template fonts.
• The userscript now ships as a single self-contained file: fflate is bundled inline, no code is downloaded at install time, and compact-readable formatting keeps the build inside Greasy Fork's 2 MB limit with identifiers intact.
• The sticky bottom-sheet option now only renders enabled in sheet-capable popup modes, and saved popover-mode settings stay usable with Japanese interface copy.
• Explicit furigana over kanji-containing words is preserved verbatim during scans, while kana-only words no longer repeat their reading as ruby.

Fixed

• Fixed word lookup popover grade buttons so selected JPDB/Anki/card targets submit from ordinary word popovers as well as kanji popovers.
• Fixed the bundled script failing to load in pages without a userscript manager (browser extension builds and test harnesses) when fflate was externalized.

[0.5.0] - 2026-06-02

Added

• Added a scalable Anki status index backed by browser storage so large Anki libraries can color parsed page words from cached lookup keys, with detailed AnkiConnect hydration deferred until interaction.
• Added Anki review support on the new-tab study page, including merged due/new queues across enabled decks and deck-disable settings for users with multiple decks.
• Added existing-library adaptation for Anki notes, including scanned note fields, per-note-type field mappings, existing-card content in popovers, and merge/update actions when a word is already in Anki.
• Added automatic Anki deck-shape handling for Core 2k/6k, Jlab, Kaishi, RRTK, Vocab 2k, Yomu, and よむ-style note fields, including RTK keyword-only kanji cards.
• Added Anki smoke coverage for reader mining, Japanese Wikipedia coloring, new-tab source toggling, multi-deck review order, and JPDB/Anki popover hydration.
• Added expanded hosted-docs setup guidance for desktop, iPhone, iPad, mobile handoff, local services, reading-site recommendations, and localized hosted navigation.

Changed

• New tab no longer shows the "Start with a dictionary" setup screen. When no local dictionary is installed, both the Dictionary source and the Auto fallback skip straight to public JPDB lookup, which works without an API key. Add Yomitan dictionaries any time from Settings → Dictionaries.
• Hosted docs now use the same generic page parser as normal websites, including the Try Me sample, current VitePress hero text, and route-mounted docs content.
• Anki and JPDB status labels in popovers and new-tab lookup metadata now show actual review state more consistently, while JPDB status is hidden when no JPDB API key is configured.
• Existing Anki cards in dictionary popovers and new-tab lookup now use Anki-rendered card HTML without Yomu-added front/back or raw field labels, keep multiple matching notes collapsible, and expose Anki as a reorderable popover source.
• Review controls now show whether they will grade JPDB, Anki, or a specific Anki card so same-reading words and multiple Anki entries are less ambiguous.
• Reader page scanning is now core behavior rather than a pair of user-facing toggles; obsolete saved scan settings are ignored and stripped on settings save.
• The settings Anki area now emphasizes connection/status, deck/model choices, field mapping, and library scanning instead of free-form fields and duplicated setup prose.
• Hosted docs localization now preserves research links and updates visible route content without leaving stale parsed spans behind.

Removed

• Removed the dedicated new-tab dictionary-setup screen along with its now-unused rendering, cached-setup state, load-dictionary action, and setup-only copy.
• Removed the old "auto-scan Japanese" and "scan visible page on load" settings because parsing Japanese text is the point of the app and should not be presented as optional duplicate behavior.
• Removed demo-specific reader paths, including isDemo/Try Me special cases and ruby suppression hooks, so docs samples are treated like ordinary page text.

Fixed

• Fixed Anki-colored words losing their Anki state/color on hover or click.
• Fixed slow Anki parsing on large pages by avoiding eager detailed card hydration during initial coloring and relying on cached status data first.
• Fixed known Anki words incorrectly showing "Add to Anki" in popovers when an existing card is present in a non-standard deck or note shape.
• Fixed Anki rendered-card previews trapping scroll, oversized card fonts, confusing raw all-caps fallback labels, and card audio playback buttons that did not clearly use Anki media.
• Fixed existing Anki detail hydration so slow or empty AnkiConnect card-detail responses fall back to cached status instead of leaving the popover stuck on a loading message.
• Fixed long rendered-card Anki audio files blocking card details by hydrating image media immediately while leaving Anki audio lazy and playable from its own Anki media button.
• Fixed public JPDB vocabulary furigana on hosted docs when automatic furigana mode resolves through JPDB/API settings.
• Fixed hosted docs text that failed to receive ruby, pitch/status coloring, or click targets after VitePress route changes.
• Fixed settings help/status rows being parsed into awkward ruby text, including the JPDB API status message.
• Fixed new-tab source toggling so JPDB and Anki review modes switch promptly and preserve the intended target.

[0.4.62] - 2026-05-31

Added

• Added previous/next word lookup shortcuts that can move through parsed words without mouse hover and stay inside selected text when a selection is active.
• Added popup Japanese font family and weight settings with a jpdb.io-matched default font stack.
• Added a pause-only subtitle side panel option and a visible Subtitles button on the hosted local video player for adding tracks.

Changed

• Routed popup, kanji, new-tab, example sentence, grammar, and local dictionary Japanese text surfaces through the popup Japanese font settings.
• Renamed the hosted subtitle track action from primary subtitles to Japanese subtitles so new users know where to load external Japanese subtitle files.

Fixed

• Sent Jisho search-page lookups through the Jina text fallback instead of the hosted worker, avoiding visible Cloudflare 525 failures during audio discovery.
• Cloned userscript bridge event payloads in Firefox so hosted-app requests do not trip XrayWrapper cross-origin object errors.
• Pointed local VitePress new-tab guidance at /yomu-reader/newtab/index.html so local docs do not accidentally load the VitePress shell at /yomu-reader/newtab/.
• Kept the keyboard-selected lookup word visibly highlighted while its popup finishes loading.

[0.4.61] - 2026-05-31

Changed

• Removed stale demo, new-tab, JPDB-page, popup, settings, OCR, and kanji-graph helper code to keep the userscript under the hosted size limit.
• Reused resolved new-tab navigation sources when loading boundary batches, avoiding duplicated source checks during card navigation.

[0.4.60] - 2026-05-31

Changed

• Shared the public-proxy route classification used by direct-fetch skipping and fallback ordering, keeping Jisho's special fallback order explicit while reducing userscript size.
• Simplified legacy lookup-link migration matching and removed a dead KanjiVG cache branch.

[0.4.59] - 2026-05-30

Fixed

• Restored hosted Try Me word targets when JPDB/local dictionary data is unavailable, including one-character words such as 下.
• Kept hosted Try Me scanning scoped to the demo text so surrounding docs copy is not accidentally turned into lookup targets.
• Restored YouTube subtitle sidebar controls while the transcript panel is open.
• Unwrapped YouTube watch titles before filtering so SPA navigation cannot leave stale reader spans in the title.
• Prevented low-value short particles in Immersion Kit/example sentences from opening noisy dictionary cards while preserving them when they are the actual target word.
• Added back-navigation context for nested example-sentence lookups opened from a word card.
• Reverted unrelated Uchisen prompt wording drift from the regression branch.

[0.4.58] - 2026-05-26

Fixed

• Let explicit Anki/status underline and color settings read existing Anki card status without enabling Anki mining.
• Batched Anki existing-card lookups across all decks and recognized common imported vocabulary fields, so cards in non-mining decks can show status, edit, and grading controls instead of “Add to Anki”.
• Kept JPDB vocabulary details, pitch accent, and Anki status loading independently so popovers settle faster and do not wait sequentially on slow Anki responses.
• Rebound translation, grammar, and Immersion Kit loaders after deferred popup rerenders so cards cannot get stuck on “Finding grammar...” or “Loading dictionary details...”.
• Kept hosted runtime and QA audio enabled by default while preserving explicit disabled-audio behavior for tests and user settings.
• Ignored VitePress's check-only home --vp-offset hydration warning in the docs audit while continuing to fail real console and page errors.

[0.4.57] - 2026-05-26

Fixed

• Replanned partially parsed Immersion Kit and example sentences from their full visible text so existing highlighted words no longer fragment later parsing passes.
• Routed pointer, selection, Immersion Kit, and new-tab parsing through the same JPDB-first path so inflected words keep JPDB readings and pitch instead of falling back to single-character cards.
• Made subtitle and image-backed Immersion Kit highlights more legible and prevented target highlights from stacking with stray underline decoration.
• Avoided caching all-fallback sentence parses as final Immersion Kit/new-tab results so transient JPDB timeouts do not poison later renders.
• Sent Jisho lookup fallbacks through alternate public proxies before the hosted worker so transient worker 525s no longer block dictionary audio discovery.

[0.4.56] - 2026-05-26

Fixed

• Re-enabled manual audio on the hosted docs demo while keeping autoplay suppressed.
• Routed term and Immersion Kit example audio through blob/proxy playback before media elements touch remote URLs, avoiding page CSP media-src blocks on sites such as Wiktionary.
• Kept dictionary glossary hyperlinks clickable after nested parsing while still allowing their visible Japanese text to be inspected with hover lookup.

[0.4.55] - 2026-05-26

Fixed

• Highlighted the word under review consistently in popover dictionary sources, JPDB examples, and JPDB page Immersion Kit examples.

[0.4.54] - 2026-05-26

Fixed

• Kept playing videos running after transcript timestamp seeks, including sites that briefly pause the media while currentTime changes.
• Added a default-on video-safe autoplay guard so automatic lookup/example audio does not interrupt visible video playback, with a setting to turn it off.
• Added a popover setting to disable the dimmed page backdrop for sticky click-opened popovers.
• Timed out hung popup and new-tab Immersion Kit example loads so iPad/Safari requests no longer leave the section stuck on “Loading examples...”.
• Paused repeated local OCR requests briefly after an unreachable local endpoint so iPad and desktop sessions do not keep hammering a down OCR server.
• Narrowed JPDB-page enhancement refreshes around dynamic Immersion Kit content while still detecting real text and anchor changes.

[0.4.53] - 2026-05-26

Fixed

• Kept scanned native page controls clickable across websites by rendering Japanese text inside links, buttons, summaries, and compact click handlers as passive lookup spans.
• Prevented press-drag lookup and injected furigana from hijacking passive control text while preserving hover lookup on those words.

[0.4.52] - 2026-05-26

Fixed

• Normalized JPDB pitch accent patterns against Japanese morae before cards render, fixing small-kana readings such as 今日/きょう whose pitch graph could appear flat.
• Avoided recreating JPDB page add-ons for unrelated JPDB-page mutations, preventing noisy refreshes from dynamic sections such as Immersion Kit examples.
• Replaced the GitHub Pages social preview image with a full-size PNG card so WhatsApp no longer enlarges the small touch icon.

[0.4.51] - 2026-05-26

Changed

• Subtitle parsing now batches the active cue warmup and uses two background transcript parse workers, so YouTube lines are prepared sooner while playback advances.

Fixed

• Kept YouTube comment and description controls such as "続きを読む", links, and buttons clickable while preserving passive hover lookup on their Japanese text.
• Unwrapped object-shaped YouTube caption labels for localized and auto-translated tracks, preventing [object Object] from appearing in the subtitle drawer title.

[0.4.50] - 2026-05-25

Changed

• Subtitle track selection now stays on Tracks and tells users to click Lines next, with updated English and Japanese copy.

Fixed

• Restored full timed transcript loading for YouTube videos whose page caption URL is empty by merging Android player caption tracks and matching equivalent Japanese streams across localized labels.
• Prevented YouTube hover-preview videos from becoming Yomu subtitle sources or leaking stale .jpdb-subtitle-primary overlays onto feed pages.
• Tightened YouTube feed filtering so playlist/mix cards and filtered videos collapse without blank grid gaps, while kanji/katakana Japanese titles remain visible and sparse feeds nudge YouTube continuation loading.

[0.4.49] - 2026-05-25

Added

• Added YouTube auto-translated caption choices for preferred languages such as Japanese, so English videos can be watched with YouTube-provided Japanese translated subtitles when available.

Changed

• After choosing a subtitle track, the sidebar now moves directly back to Lines when transcript lines are available, with a small Tracks hint explaining the flow.
• YouTube caption discovery now keeps creator captions, auto-generated captions, and auto-translated captions as separate track choices instead of collapsing same-language options together.

Fixed

• Made transcript row clicks and keyboard activation seek to the exact subtitle timestamp, while keeping the configured seek padding for previous/next controls.
• Reduced the YouTube DOM-caption fallback delay and preserved full transcript rendering when timed YouTube tracks are available, avoiding the one-line-only sidebar state.
• Kept the transcript drawer visible when subtitle controls are hidden and tightened current YouTube player insets so the side drawer does not overlap the video after layout settles.
• Refreshed transcript JPDB parsing when parse-affecting settings change, keeping sidebar subtitle words colorized, hoverable, and clickable like player subtitles.

[0.4.48] - 2026-05-25

Fixed

• Smoothed YouTube immersion filtering by avoiding repeated oEmbed retries, batching rescans, and keeping translated-looking titles visible until the original title check finishes.
• Hid playlist, mix, radio, and shelf items without treating normal videos whose title contains "mix" as playlists.
• Stopped the YouTube hidden-items notice from reappearing on every count change, left Shorts scrolling untouched, and prevented よむ spans from leaving stale YouTube watch titles after SPA navigation.

[0.4.47] - 2026-05-25

Fixed

• Allowed scanned page words with furigana to wrap normally, preventing long annotated Japanese lines from overflowing into adjacent page columns.
• Added passive no-furigana scanning for safe page UI labels such as navigation links and buttons, so labels can be recognized without hijacking the controls' normal clicks.

[0.4.46] - 2026-05-25

Fixed

• Restored iOS tap/click lookup behavior for OCR text regions so tapping an OCR line opens the same sticky/modal lookup state as tapping normal scanned text.

[0.4.45] - 2026-05-25

Changed

• Refined the iPhone/iPad install guide for current Userscripts behavior, including the automatic scripts folder and Safari AA/extensions menu flow.

Fixed

• Kept image OCR quiet when a scanned image has no Japanese text, instead of showing a "No Japanese text found" banner on ordinary non-Japanese images.
• Changed Immersion Kit rate-limit backoff to start at 1 s (was 100 ms) and double on each consecutive 429, capping at 30 s, then reset after a successful search.

[0.4.44] - 2026-05-25

Fixed

• Fixed OCR pronunciation cleanup and Japanese translation quote parsing in popup results.

[0.4.43] - 2026-05-25

Fixed

• Fixed hosted Try Me furigana alignment by keeping parsed words on native inline ruby layout instead of flex layout.

[0.4.42] - 2026-05-25

Added

• OCR now triggers automatically when hovering or moving the pointer over a manga/image panel, so you no longer need to click to start a scan on a new image.
• Immersion Kit searches now respect a 2-minute rate-limit cooldown after receiving a 429 response, preventing repeated failed requests from hammering the API.

Fixed

• Fixed a bug where ImmersionPopoverController.loadExamples resolved to false instead of resolving cleanly when a popover load was aborted mid-flight, which could incorrectly suppress a follow-up load on a new popover.
• Fixed the Immersion Kit <details> mount starting open when it should start closed on both the reader and new-tab pages.
• Fixed the docs navigation "more" button icon rendering incorrectly in some browsers.

Internal

• Removed an empty else {} branch from LruCache.get that had no effect.
• Added dedicated test suites for LruCache, runLimited, card-state normalization, and immersion query utilities (60 new tests).

[0.4.41] - 2026-05-25

Changed

• Made the YouTube immersion filter enabled by default and updated its settings/docs copy to match the new default.
• Aligned YouTube card filtering with NihongoTube by checking original YouTube oEmbed titles when video ids are available and using kana-based Japanese detection instead of localized title text alone.

Fixed

• Replaced the persistent YouTube hidden-video bar with a temporary toast-style notice and clearer controls for showing hidden videos or hiding future notices without turning off the filter.

[0.4.40] - 2026-05-25

Fixed

• Restored the known-good Yomu icon and regenerated every shipped icon asset from the same source visual.
• Fixed hosted/iOS reader styling when userscript GM_* resource APIs are unavailable, restoring lookup drawers, word colors, and OCR overlays on the main site.
• Published the reader CSS asset with the hosted docs so the main site can load the full stylesheet without userscript manager APIs.

[0.4.39] - 2026-05-24

Fixed

• Restored the pre-0.4.38 raster Yomu app icons while keeping the hosted share metadata pointed at the SVG icon.

[0.4.38] - 2026-05-24

Fixed

• Made hosted social metadata point at the live Yomu SVG icon and regenerated fallback raster assets from that SVG.

[0.4.37] - 2026-05-24

Added

• Added a development userscript launcher and shipped readable reader CSS as a Greasy Fork @resource, with release verification for the external CSS asset and userscript metadata.

Changed

• Refined settings, hosted new-tab, subtitle, and nested lookup behavior, including stacked settings popups and subtitle-player click lookup handling.
• Refreshed the Yomu icon assets and tightened hosted docs/new-tab styling.

Fixed

• Hardened subtitle parsing, native-track handling, JPDB source parsing, performance caches, and settings form behavior with broader regression coverage.

[0.4.36] - 2026-05-24

Fixed

• Added an iOS/Safari-safe idle scheduling fallback so OCR and deferred dictionary lookups no longer crash when window.requestIdleCallback is missing or not callable.

[0.4.35] - 2026-05-24

Fixed

• Removed the duplicate over-video subtitle file loader buttons; local subtitle files are now chosen from the subtitle tracks side panel.

[0.4.34] - 2026-05-24

Added

• Restored JPDB.io page enhancements for JPDB word/search and kanji surfaces, using the same ordered source-card styling as the popup and hosted new-tab dictionary.
• Added Immersion Kit as a reorderable kanji source for popup, hosted new-tab, and JPDB kanji pages.

Fixed

• Suppressed native browser title tooltips while a dictionary popup is open, preventing OCR sentence/status tooltips from covering the lookup card.

[0.4.33] - 2026-05-24

Fixed

• Disabled hover lookup and hover-close behavior while a click/tap-opened popup is active, so iPad mouse, trackpad, and Apple Pencil hover surfaces no longer replace or close the sticky lookup when the pointer moves away.

[0.4.32] - 2026-05-24

Changed

• Moved the hosted new-tab Stats and Settings actions into a compact more menu, refreshed new-tab source switching, and tightened responsive/new-tab styling.
• Improved Japanese UI parsing for settings and nested reader text while preserving click-first lookup behavior inside Yomu-controlled surfaces.

Fixed

• Hardened pointer lookup edge cases around low-value kana tokens, JPDB reading matches, theme/settings controls, and new-tab review flows.

[0.4.30] - 2026-05-24

Fixed

• Fixed iPhone/touch OCR image interaction so tapping an OCR line reveals and pins the sentence first, while tapping parsed words inside the revealed line still opens the normal lookup popup.

[0.4.29] - 2026-05-24

Added

• Added richer hosted new-tab review flows for study cards, including expanded lookup/review behavior and stronger regression coverage.
• Added media activation handling and broader Uchisen publishing/validation utilities for kanji study assets.

Changed

• Improved JPDB, Anki, Immersion Kit, audio, settings, stats, and study-tool runtime behavior across popup and hosted new-tab surfaces.
• Refined responsive styling for new-tab, stats, settings, kanji, and immersion study views.

Fixed

• Hardened dictionary source ordering, JPDB parsing, proxy fetches, performance caches, study grammar handling, and stats tests around the latest study flows.

[0.4.28] - 2026-05-23

Added

• Added beginner-facing reading-site recommendations to the getting started guide, including graded readers, easy news, ebook readers, web novels, and subtitle-based YouTube practice.
• Added hosted-page copy that explains よむ as a reading-first immersion system across JPDB, Yomitan dictionaries, Anki, OCR, subtitles, and example sources.

Changed

• Made shuffled audio behave like a deck, trying available clips for a word before reshuffling, and updated the README and audio docs to describe the new behavior.
• Improved hosted new-tab and stats behavior around JPDB/Anki study cards, dictionary-backed study words, and cached runtime data.

Fixed

• Hardened lookup, audio, OCR, proxy fetches, and local dictionary paths with broader regression coverage for JPDB pages, new-tab review, performance caches, and stats.

[0.4.27] - 2026-05-23

Fixed

• Kept the back arrow available when clicking parsed study-source words inside an already-open popup, including hover-opened popovers and fallback/local parsed sentence spans.

[0.4.26] - 2026-05-23

Added

• Added a hosted New Tab Stats view that combines JPDB and Anki progress, including daily review activity, study minutes, new-card charts, retention, streaks, card distribution, average review speed, and due-time estimates.
• Added JPDB reviews.json import support for historical review graphs while still using the JPDB API key for current card-state stats.
• Added AnkiConnect stats loading for daily review counts, retention sampling, and deck card breakdowns.

Changed

• Added a Stats nav mode alongside Word, Kanji, and Search, with a matching docs nav link and Yomu-themed responsive styling.
• Let the Stats card distribution jump back into the study deck focused on due or failed cards.

[0.4.25] - 2026-05-22

Changed

• Made popup, subtitle, YouTube, and hosted new-tab work prioritize the current word more aggressively while moving background parsing, card data, grammar, translation, audio, and Immersion Kit preloads onto shorter waits and bounded caches.
• Batched visible-page DOM updates and transcript hydration work so long pages and YouTube transcripts stay responsive on iPhone-class devices.
• Added outbound-link icons to hosted Settings help links so external destinations match the Video Player affordance.

Fixed

• Prevented disabled YouTube filtering from attaching observers or scanning video cards until the filter is enabled.
• Kept subtitle transcript cue parsing from dropping rows that were already pending in another hydration batch.
• Reused fallback Japanese word segmentation and cleared in-flight JPDB parses on reset to avoid stale parser work.
• Stopped userscript HTTP timeouts from starting a second fetch fallback, avoiding extra waits on slow requests.

[0.4.24] - 2026-05-21

Fixed

• Made hosted new-tab search include already-loaded JPDB and Anki review cards, so English glossary searches can find study cards even before a local dictionary is installed.
• Limited the userscript HTTP bridge to Yomu-hosted app pages while keeping direct GM requests available inside the userscript on normal websites.

[0.4.23] - 2026-05-19

Changed

• Extension releases now use the generated popup menu instead of an options/reviewer page, and release assets include one consolidated submission guide instead of separate review markdown files.
• Extension new-tab builds now load the cache/version helper from an external script so Firefox and other strict extension pages do not block startup with inline-script policy errors.

Fixed

• Fixed extension new-tab branding so the よむ icon resolves from the packaged extension instead of a broken hosted-page relative path.
• Hardened Firefox/userscript page injection against cross-origin property descriptors and Trusted Types/CSP-protected parsing, restoring raw userscript installs on strict pages such as Google, YouTube, and NHK.
• Restored YouTube subtitle discovery fallback behavior when native caption tracks are not exposed immediately, while keeping DOM caption fallback available for currently visible captions.
• Let NHK Easy pages fall back to whole-page parsing when the site-specific parser finds no article targets, so visible Japanese can still be colored and looked up.
• Made OCR lookup hit targets larger, kept pinned OCR lines stable during hover, and kept lookup popovers from opening directly under the pointer.
• Kept words inside an open popup click-driven instead of hover-driven, and preserved pinned/modal popup mode while moving through nested lookup content.

[0.4.15] - 2026-05-19

Added

• Added GitHub Release assets for the compiled userscript, Chrome extension ZIP, Firefox XPI, Safari Web Extension ZIP, compiler project bundle, and review notes.
• Added an Audio setting for whether JPDB/browser text-to-speech is fallback-only or participates in the configured source order/random pool.
• Replaced the Help-panel Help link with a Factory Reset action that clears settings, API keys, preferences, cached cards, local dictionary storage, and other よむ local data before reloading defaults.
• Hardened Factory Reset so settings storage is cleared even if dictionary database cleanup is blocked, including userscript managers that expose modern GM.* storage APIs.
• Added capped JPDB "used in vocabulary" and public example rows to popup JPDB definitions, including compact buttons for JPDB-provided example audio.
• Added Anki card front controls for hiding the reading, sentence, or image on word-first cards.
• Added Immersion Kit audio to Anki-mined notes when the selected Immersion Kit example is used as the card context.

Changed

• Made GitHub Pages deploys rebuild and sync the hosted userscript and new-tab assets before publishing.
• Made /newtab keep JPDB and Anki SRS queue order, alternate JPDB/Anki cards in Auto mode, and only fall back to random dictionary words after both review queues are empty.
• Made the hosted new tab paint cached cards immediately while JPDB, Anki, or dictionary sources refresh in the background, and sped up deck/dictionary refresh work so a card appears sooner after reload.
• Moved the shared cross-origin proxy URL control into Audio settings so proxy-dependent audio sources are easier to diagnose.
• Moved existing Anki note actions into the Anki preview, simplified the header, removed the default Status row from Yomu Anki cards, and refreshed Anki lookup state immediately after adding a note.

Fixed

• Fixed raw userscript installs on pages that shadow window event methods by routing the userscript HTTP bridge through hardened event helpers.
• Let the hosted new-tab page attempt direct AnkiConnect requests and show the CORS/userscript bridge setup hint when standalone Anki tests fail.
• Made recommended dictionary installs visibly queue/import on their buttons and kept Settings Save unavailable until dictionary imports finish, so new-tab setup no longer looks broken while dictionaries are still loading.
• Restored LanguagePod101 lookups through the configured/public proxy path in browser-fetch contexts.
• Made Escape close the Settings dialog even when focus is inside a normal settings field.
• Made hosted new-tab bottom-sheet lookups modeless so sentence-mining taps can update the open drawer, and hid the sticky bottom-sheet option when Popover mode is forced.
• Restored JPDB-status word highlights and pitch underlines across hosted new-tab prompt sentences and Immersion Kit example cards.

[0.4.14] - 2026-05-16

Added

• Added hosted new-tab search as a first-class Word / Kanji / Search mode, with JPDB/local word results, kanji drilldown results, external lookup links, and autocomplete suggestions.

Changed

• Renamed the homepage hero CTA to Try Out and removed the duplicate See Features hero pill while keeping the nav Features link.
• Widened mobile sheet and mining-drawer drag targets so the handle bands are easier to grab on touch devices.
• Spaced kanji origin graph nodes more aggressively to reduce overlap in crowded component maps.

Fixed

• Kept the new-tab word answer layout from bunching or overlapping on small screens.

Fixed

• Restored hosted new-tab kanji drill-down back navigation and kept the current lookup sheet height while moving through components.
• Made mobile lookup sheets resize continuously by dragging the handle, remember the chosen height for the next popup, and close when the handle is tapped.
• Added the same pull-to-resize affordance to the mobile Settings drawer, with an independent remembered height.
• Kept Translation and Grammar as separate popup sources that follow the Dictionaries settings order and remember collapsed state across rerenders.
• Removed duplicate Immersion Kit source labels and spaced parsed example captions so furigana, Japanese text, and translations do not overlap.

[0.4.11] - 2026-05-16

Fixed

• Routed public page/media requests through the shared CORS fallback stack, including hosted-page audio, Uchisen, JPDB public kanji/vocabulary pages, RTK, KanjiVG, pitch, and Immersion Kit media.
• Routed hosted JPDB API review/deck calls through the restricted Cloudflare Worker fallback on iPad and GitHub Pages, while keeping direct localhost calls available for development.
• Made JPDB API deck cards load deterministically in the hosted new tab so the review total no longer changes randomly on refresh.
• Matched new-tab kanji source ordering to Settings, with JPDB kanji first, RTK immediately after it, kanji facts/graph next, and every section open by default.
• Flattened nested new-tab kanji detail cards while keeping useful component and similar-word cards, and centered Uchisen controls, image, and story text.
• Tightened the new-tab kanji reveal layout so the reference and drawing panels align cleanly, and removed the repeated lower JPDB detail card stack.
• Kept JPDB kanji fact items on one line when space allows.
• Repaired prompt tapping in word mode so tapping the displayed word opens the dictionary popover instead of doing nothing.
• Cleared stale parse markers across progressive popup renders so JPDB, Immersion Kit, and local sentences remain tokenized and clickable after details finish loading.
• Made kanji taps inside the hosted new-tab popup open the full kanji panel with Settings-ordered sections instead of falling back to a plain one-character word lookup.
• Tightened iPhone word-review spacing so the prompt, answer, Immersion Kit media, and five grading buttons no longer overlap or wrap awkwardly.
• Reworked touch settings layout so phone and tablet use a bottom drawer, source rows do not overlap controls, and inactive tabs do not look selected after tapping.
• Restored hosted mobile popup behavior: Add to Anki now uses the iOS/Android handoff without AnkiConnect preflights, popup sections collapse/expand normally, and bottom sheets recover their size after rotation.

[0.4.10] - 2026-05-15

Fixed

• Added a Cloudflare-hosted public-resource proxy fallback for hosted-page JPDB kanji/vocabulary, pitch, audio, and dictionary downloads without sending logged-in JPDB actions through public proxies.
• Reworked the hosted new-tab iPad flow so missing dictionaries send users to Settings, dictionary downloads do not open surprise tabs, new-tab word taps open the lookup popover, kanji graph nodes drag, and the bottom controls stay fixed.
• Kept the Dictionaries settings source-order table compact on iPad when no imported dictionaries exist, removing the empty Display name and Remove columns until they are useful.
• Added hosted-page userscript coverage back for the new-tab page so JPDB kanji, Uchisen, RTK, and remote dictionary downloads can use the userscript request bridge on iPad.
• Improved the settings Help and tablet layout, including visible donation/support links and better wrapping for settings rows.
• Treat mobile Anki handoff as a valid iPad/Android path instead of reporting mobile-only setups as broken AnkiConnect.
• Let the standalone hosted new-tab page use normal browser CORS for Immersion Kit examples/media, so Chrome, Safari/WebKit, Firefox, and mobile browsers do not need an installed userscript just to load examples.
• Kept compressed Yomitan ZIP imports working when DecompressionStream is unavailable or unreliable.

[0.4.9] - 2026-05-14

Fixed

• Restored the hosted new-tab demo so local dictionary cards, nested popup dictionary links, kanji drilldowns, and similar-word lookups work even before the userscript is installed.
• Preserved Immersion Kit context when mining from nested example lookups, including the active example image in Anki notes.
• Restored explicit transcript and track-picker subtitle controls, kept the transcript drawer closed by default, and tightened transcript accessibility contrast/target sizing.
• Reworked release QA coverage for the hosted docs/new-tab page, JPDB add-ons, OCR, Immersion Kit, subtitles, and userscript bundle verification.

[0.4.8] - 2026-05-14

Added

• Added karaoke-style subtitle word timing, smarter transcript layout fallback, and a resizable transcript panel that stays usable across ordinary video pages, YouTube, and Comprehensible Japanese layouts.
• Added JPDB vocabulary compounds/examples inside popup JPDB sections and JPDB page side panels, with better Immersion Kit fallback queries for compound terms.
• Added userscript menu actions to open the hosted new-tab page and reset all local よむ data, plus subtitle smoke/e2e scripts for release QA.

Changed

• Improved subtitle sentence recovery, transcript hydration, OCR/site parser handling, and audio preview matching so mining and playback stay more reliable around transcript-heavy pages.
• Updated popup and kanji navigation so nested kanji drilldowns can return through prior kanji cards or back to the source word without losing position.
• Made the support/donation copy more transparent about the AI/API token costs behind よむ development.

Fixed

• Prevented side transcript layouts from shrinking videos too aggressively by falling back below the player when space gets too tight.
• Tightened JPDB page parsing and popup rendering so compounds, examples, and kana-backed audio behave more consistently on compound-heavy entries.
• Restored no-key JPDB public lookup data so popup JPDB definitions/examples, public pitch accent, and JPDB kanji details load from public pages while mining stays API-key gated.
• Made the reset-all command refresh back into first-run onboarding and tell other open よむ tabs to drop stale stores before reloading.

[0.4.7] - 2026-05-13

Changed

• Tuned Help-card spacing and Add button accent color so settings and kanji controls feel calmer and more consistent.

[0.4.6] - 2026-05-13

Changed

• Polished shared settings/action button styling so support and settings controls read as one quieter system.

[0.4.5] - 2026-05-13

Fixed

• Aligned the settings light/dark switch so it sits cleanly in the settings control row.

[0.4.4] - 2026-05-13

Changed

• Added a visible light/dark theme switch in settings and refreshed the new-tab theme switch without the old logo halo treatment.
• Removed the stale hosted screenshot gallery and old public reader/video/OCR test pages from the docs deployment.
• Tightened popup action styling so kanji-card lookup pills and mining/review buttons sit more quietly in the layout.

Fixed

• Locked mining controls to the bottom of fixed-height popups, with the expand/minimize control in a slim gutter instead of floating over action buttons.
• Replaced the cramped icon plus on Add to deck with the plain text label Add to deck +.
• Restored reliable JPDB kanji drilldown, review doodle preview carryover, OCR parsing, subtitle transcript layout, and new-tab fallback coverage in the Playwright QA pass.
• Stopped the docs site from self-injecting the userscript bundle, removing misleading CORS and lookup noise on GitHub Pages.

[0.4.3] - 2026-05-13

Changed

• Simplified the hosted video page into a single native video host so Yomu's normal subtitle overlay, track picker, transcript sidebar, and mining controls own the full subtitle workflow.
• Reworked YouTube/local subtitle sidebars around a transcript-first Lines surface, with left/bottom/right placement, auto-scroll, and fullscreen behavior that hides page chrome while keeping Yomu subtitles visible.
• Pre-warmed parsed subtitle styling and dynamically fit subtitle font size to the actual video bounds to avoid unstyled flashes and overflow.

Fixed

• Removed the hosted video page's custom file queue, layout sidebar, and Clear button so it no longer behaves like a second video player.
• Prevented Yomu and YouTube/native captions from displaying at the same time, with a hidden fallback for current-line YouTube captions when timedtext returns no cue list.
• Stopped idle subtitle/sidebar layout loops from repeatedly writing player styles, and kept the floating puck away from unfocused video surfaces.

[0.4.2] - 2026-05-12

Added

• Added richer learner grammar cues for common particles, polite forms, conditionals, negatives, and verb endings, with matched sentence context and guide links.
• Added local dictionary/new-tab and source-order improvements from the current workspace changes.
• Added a hosted video-player page for local browser-supported media and subtitle files.
• Added Uchisen as an ordered kanji-popup source and brought Uchisen, RTK, and stroke practice into JPDB kanji/review surfaces.
• Added JPDB review reveal word-audio autoplay using よむ's configured Yomitan-compatible audio sources.

Changed

• Bumped the package and userscript version to 0.4.2.
• Reworked translation and grammar study panels into compact learner rows instead of nested cards, with smaller typography and better ruby/furigana spacing.
• Moved reader styles into the bundled stylesheet path while keeping the userscript self-contained.
• Tuned default subtitle appearance to be smaller, lighter, and closer to ASB-style captions.
• Replaced the legacy subtitle-bridge menu action with an Open Video Player action that launches the GitHub Pages player.
• Renamed the JPDB review sentence toggle to "Auto-reveal review example sentences" and made Immersion Kit reveal audio mutually exclusive with よむ word reveal audio.

Fixed

• Removed the misleading “Pattern hints are best guesses from the full sentence shape.” note and the redundant grammar cue count.
• Tightened grammar matching so forms such as 読みました and 確認できます show cleaner “Found in” text and できます is not mistaken for the particle で.
• Prevented detected page captions from stacking artificial DOM line breaks, and expanded the subtitle backing/shadow so furigana stays visually covered.
• Cleared JPDB Immersion Kit reveal audio automatically when the global or JPDB Immersion Kit add-ons are disabled.

[0.4.1] - 2026-05-12

Fixed

• Completed the furigana/highlight settings UI so the existing furigana modes, hidden-known behavior, and highlight-off mode are exposed consistently in settings.
• Updated subtitle/transcript parsing cache keys so furigana and word-highlight mode changes refresh parsed subtitle lines correctly.

[0.4.0] - 2026-05-12

Added

• Added a transcript panel for video subtitle mining, with active-line highlighting, auto-scroll, responsive left/right/below placement, and tappable lookup on visible transcript lines.
• Added a Local Audio docs page covering hosted Ultimate Yomitan Audio, self-hosted audio files, local server setup, startup tasks, custom ports, and Tailscale access.
• Added optional one-time Immersion Kit hover audio on desktop, with a setting to turn it off and manual replay kept available on every device.
• Added deterministic Playwright plus axe/WCAG audits for product fixtures and GitHub Pages docs, a complexity audit, a local .env.example, and a live JPDB smoke test that reads secrets only from local .env.
• Added the GitHub Actions release workflow for tagged releases with dist/yomu.user.js attached.

Changed

• Bumped the package version to 0.4.0.
• Added automatic word highlight mode: status colors stay tied to JPDB/Anki mining, while pitch-accent colors become the default when mining status is not configured.
• Made JPDB mining actions independently configurable so users can keep a JPDB API key for popup lookup without showing add/Never Forget/blacklist actions.
• Reworked video subtitle controls into compact icon buttons; transcript and track panels now share the same side-panel surface instead of competing popovers.
• Reworked Immersion Kit example controls so navigation, audio, and count alignment stay compact inside the existing card instead of looking like separate panels.
• Expanded the docs and README with fuller feature descriptions, iPhone/iPad limitations, beginner-friendly local audio guidance, release links, and source credits.
• Refreshed screenshots, docs contrast, new-tab accent theming, donation/support copy, and the homepage/new-tab assets.
• Improved new-tab fallback behavior so the page can use Anki, JPDB, then top-ranked local dictionary words without showing a dead setup warning.

Fixed

• Improved page scanning on JPDB, Jisho, and ruby-heavy NHK-style pages so split Japanese text, one-kanji terms, and inline furigana are wrapped for lookup more reliably.
• Fixed the new-tab loading path so static placeholder markup is replaced by the live study UI.
• Fixed duplicate/old userscript menu-command wiring from earlier subtitle actions.
• Fixed cramped subtitle controls, oversized hide button styling, and broken/low-contrast docs imagery.
• Fixed native-page CSS collisions that could stretch よむ popup action buttons on JPDB search/review pages.
• Prevented local JPDB keys from leaking into source scans by documenting and ignoring local .env files.

[0.3.0] - 2026-05-11

Added

• Added the GitHub Pages documentation site with beginner-friendly install instructions, feature docs, support links, and Playwright screenshots.
• Added GitHub Actions deployment for the docs site.
• Added the optional よむ new-tab study page for browser home pages, new tabs, and iPad Home Screen shortcuts.
• Added clearer iPhone/iPad guidance, including Tampermonkey and the free open-source Userscripts app.
• Added documentation for upcoming native Chrome, Firefox, and Safari extensions.
• Added this changelog as the canonical release-notes source for the website.
• Added a copy button to word and kanji lookup pills.
• Added public JPDB pitch-accent fallback for words that do not include pitch data in the parsed card.

Changed

• Bumped the package version to 0.3.0.
• Updated the settings Help area to link to the documentation site.
• Expanded support documentation for GitHub issues, Discord, and optional donations.
• Improved JPDB add-on example audio handling so repeat taps do not stack duplicate playback or leak temporary blob URLs.

Fixed

• Made the first-run mobile onboarding choices clearer.
• Prevented mobile audio-source controls from clipping in settings.
• Fixed the kanji drilldown JPDB button so it opens the matching JPDB kanji page.
• Made the hover lookup QA screenshot deterministic by drilling into the seeded 今日 kanji fixture before continuing hover and press-drag checks.

[0.2.0] - 2026-05-10

Added

• Released the initial よむ userscript baseline with JPDB popup lookup, JPDB mining, Yomitan dictionary imports, OCR, subtitles, YouTube filtering, kanji drilldown, Anki support, and browser QA fixtures.