The greedy name pre-pass asked the Yomitan backend at every Japanese
position, because a character name can begin mid-token. With the character
dictionary enabled that roughly doubled the round trips per line (measured
10 -> 21 on a 23-char line).
SubMiner generates the character dictionary, so the cached snapshots already
list every form a character entry can be matched by (term and reading). Those
forms are installed into the scan runtime once per media and the pre-pass now
probes only positions where one of them starts, compared after kana
normalization so a katakana name still matches a hiragana reading form. The
overhead drops to zero (21 -> 10, the same as with the dictionary disabled).
Fail-safe: with no candidate list (no media id, no cached snapshot, failed
install) the pre-pass keeps its exhaustive behavior, so stale character data
costs speed rather than a missing name. Halfwidth katakana positions bypass
the filter since kana normalization does not fold them.
The candidate lookup is consulted per subtitle line, so it caches its snapshot
directory signature for 5s; dictionary writes still call invalidate().
- extract the injected scan runtime (helpers, install script, call-script
builder) into tokenizer/yomitan-scan-runtime-script.ts; the host module drops
from ~2700 to ~1900 lines
- append the kana run to the reading as well as the surface when an unparsed
run extends the previous token, so the reading keeps covering the surface and
the known-word reading fallback stays enabled (bumps scan runtime version)
- stop annotateMs before character-image resolution so the stage timing
measures the annotation stage only
- drop the duplicate parseText full parse per line; the termsFind scanner walk
is now authoritative and emits its own unparsed filler runs (parseText kept
only as error fallback)
- install scan helpers once per parser window (__subminerYomitanScan) instead
of re-shipping ~500 lines of script per subtitle line
- persist termsFind results across lines in a window-scoped LRU keyed by
substring, invalidated via a cache epoch on dictionary/settings changes
- skip lookups at punctuation/whitespace positions and cap the shrinking-window
retry ladder at 4 lookups per position
- build tokenizer runtime deps once (JLPT lookup cache never hit before; mecab
availability check ran per line)
- stop restarting the prefetch run on every subtitle change; resume prefetch
only after the tokenized payload lands, not on provisional raw emits
- add per-stage debug timings (scanMs/mecabMs/frequencyMs/annotateMs)
- Add allowReadingOnlyMatch:false to kanji token known-word lookups so 渓谷/けいこく no longer matches a mined 警告/けいこく card via reading-only index
- Greedy name pre-pass yields when a strictly longer generic word starts at the same position (空 no longer splits 空気; ties still go to the name)
* feat(tokenizer): use Yomitan word classes for subtitle POS filtering
- Carry matched headword wordClasses from termsFind into YomitanScanToken
- Map recognized Yomitan wordClasses to SubMiner coarse POS before annotation
- MeCab enrichment now fills only missing POS fields, preserving existing coarse pos1
- Exclude standalone grammar particles, して helper fragments, and single-kana surfaces from annotations
- Respect source-text punctuation gaps when counting N+1 sentence words
- Preserve known-word highlight on excluded kanji-containing tokens
- Add backlog tasks 304 (N+1 boundary bug) and 305 (wordClasses POS, done)
* fix(tokenizer): preserve annotation and enrichment behavior
* fix: restore jlpt subtitle underlines
* fix: exclude kana-only n+1 targets
* fix: refresh overlay on Hyprland fullscreen
* fix: address fullscreen and n-plus-one review notes
* fix: address CodeRabbit review comments
* fix: accept modified digits for multi-line sentence mining
* Cancel pending Linux MPV fullscreen overlay refresh bursts
- return a cancel handle from the Linux refresh burst scheduler
- clear pending refresh bursts when overlays hide or windows close
- tighten the burst test polling to wait for the async refresh
* fix: suppress N+1 for kana-only candidates and fix minSentenceWords coun
- Treat kana-only tokens with surrounding subtitle punctuation (…, ―, etc.) as kana-only so they are not promoted to N+1 targets
- Exclude unknown tokens filtered from N+1 targeting from the minSentenceWords count so filtered kana-only unknowns cannot satisfy sentence length threshold
- Add regression tests for kana-only candidate suppression and filtered-unknown padding cases
* Suppress subtitle annotations for grammar fragments
- Hide annotation metadata for auxiliary inflection and ja-nai endings
- Preserve lexical `くれる` forms and add regression coverage
* Fix kana-only N+1 tokenizer regression test
- Use a pure-kana fixture for the subtitle token N+1 case
- Update task notes for the latest CodeRabbit follow-up
* Fix managed playback exit and tokenizer grammar splits
- Ignore background stats daemons during regular app startup
- Split standalone grammar endings before applying annotations
- Clear helper-span annotations for auxiliary-only tokens
* fix: refresh current subtitle after known-word mining
* fix: suppress sigh interjection annotations
* fix: preserve jlpt underline color after lookup
* Replace grammar-ending permutations with shared matcher; preserve word a
- Extract `grammar-ending.ts` with `isStandaloneGrammarEndingText` / `isSubtitleGrammarEndingText` pattern matchers
- Replace `STANDALONE_GRAMMAR_ENDINGS` set in parser-selection-stage with shared matcher
- Replace generated phrase sets in subtitle-annotation-filter with shared matcher
- Remove stale duplicate subtitle-exclusion constants and helpers from annotation-stage
- Manual clipboard card updates now write only to the sentence audio field, leaving word/expression audio untouched
* fix: CI changelog, annotation options threading, and Jellyfin quit
- Add `type: fixed` / `area:` frontmatter to `changes/319` to pass `changelog:lint`
- Thread `TokenizerAnnotationOptions` through `stripSubtitleAnnotationMetadata` so `sourceText` is honored
- Include `jellyfinPlay` in `shouldQuitOnDisconnectWhenOverlayRuntimeInitialized` predicate
- Make mouse test `elementFromPoint` stubs coordinate-sensitive
- Make Lua test `.tmp` mkdir portable on Windows
* Preserve overlay across macOS flaps and mpv playlist changes
- keep visible overlays alive during transient macOS tracker loss
- reuse the running mpv overlay path on playlist navigation
- update regression coverage and changelog fragments
* fix: restore stats daemon deferral
* fix: keep subtitle prefetch alive after cache hits
* Fix JLPT underline color drift and AniList skipped-threshold sync
- Replace JLPT `text-decoration` underlines with `border-bottom` so Chromium selection/hover cannot repaint them to another annotation's color
- Lock JLPT underline color for combined annotation selectors (known, n+1, frequency) and character hover/selection states
- Trigger AniList post-watch check on every mpv time-position update to catch skipped completion thresholds
- Fall back to filename-parser season/episode when guessit omits them
* fix: address coderabbit feedback
* fix: sync AniList after seeked completion
* fix: preserve ordinal frequency annotations
* fix: preserve known highlighting for filtered tokens
* fix: address PR #57 CodeRabbit feedback
- Acquire AniList post-watch in-flight lock before async gating to prevent duplicate writes
- Isolate manual watched mark result from AniList post-watch callback failures
- Report known-word cache clears as mutations during immediate append when state existed
- Add regression tests for each fix
* fix: stop AniList setup reopening on Linux when keyring token exists
- Gate setup success on token persistence: `saveToken` now returns `boolean`; on failure, keeps the setup window open instead of reporting success
- Config reload passes `allowSetupPrompt: false` so playback reloads don't re-open the setup window
- Add regression test for persistence-failure path
* fix: suppress known highlights for subtitle particles
* fix: retry transient AniList safeStorage failures
* fix: hide overlay focus ring
* fix: align Hyprland fullscreen overlays
* fix: restore subtitle playback keybindings
* fix: align Hyprland overlay windows to mpv and stop pinning them
- Force-apply exact Hyprland move/resize/setprop dispatches when bounds are provided
- Stop pinning overlay windows; toggle pin off when Hyprland reports pinned=true
- Compensate stats overlay outer placement for Electron/Wayland content insets
- Make stats overlay window and page opaque so mpv cannot show through transparent insets
- Constrain stats app to h-screen with internal scroll so content covers mpv from y=0
- Lock overlay/stats window titles against page-title-updated events
- Add regression coverage for placement dispatches, inset compensation, and CSS overlay mode
* fix: retain frequency rank for honorific prefix-noun tokens
- Add `shouldAllowHonorificPrefixNounFrequency` to exempt お/ご/御 + noun merged tokens from frequency exclusion
- Add regression test for `ご機嫌` asserting rank 5484 is preserved after MeCab enrichment and annotation
- Close TASK-341
* fix: map openCharacterDictionary session action to --open-character-dict
- Add missing Lua CLI dispatch entry for openCharacterDictionary
- Add regression test for Alt+Meta+A binding and CLI flag forwarding
* fix: keep macOS overlay interactive while mpv remains active
- Overlay no longer hides or becomes click-through during tracker refreshes when mpv is the focused window
- Preserve already-visible overlay when tracker is temporarily not ready but mpv target signal is active
- Add regression tests for active-mpv tracker refresh and transient tracker-not-ready paths
* fix: address coderabbit subtitle follow-ups
* fix: resolve media detail from sessions when lifetime summary is absent
- Change `getMediaDetail` JOIN to LEFT JOIN on `imm_lifetime_media` and fall back to aggregated session metrics when no lifetime row exists
- Add filter `AND (lm.video_id IS NOT NULL OR s.session_id IS NOT NULL)` to keep results valid
- Add regression test covering the session-visible / media-detail-missing mismatch
* fix: address PR-57 CodeRabbit findings and CI failures
- use filtered word counts in media detail session token aggregation
- cancel fullscreen refresh burst on exit via updateLinuxMpvFullscreenOverlayRefreshBurst
- guard Hyprland JSON.parse in try/catch; exclude windowtitle from geometry events
- narrow focus suppression from :focus to :focus-visible
- apply JLPT lock selectors to word-name-match tokens (N1–N5)
* fix: macOS overlay z-order and Yomitan compound token known highlighting
- Release always-on-top when tracked mpv loses foreground on macOS
- Skip visible overlay blur restacking on macOS to avoid covering unrelated windows
- Prefer Yomitan internal parse tokens over fragmented scanner output for known-word decisions
- Add regression tests for both behaviors
* fix: macOS visible-overlay blur no longer invokes Windows-only blur call
- Split win32/darwin branches in handleOverlayWindowBlurred so darwin visible blur returns early without calling onWindowsVisibleOverlayBlur
- Add regression test asserting Windows callback stays inactive on macOS visible overlay blur
- Close TASK-347
sudacode
