chore(release): prepare 0.18.0-beta.3

package.json

@@ -2,7 +2,7 @@
   "name": "subminer",
   "productName": "SubMiner",
   "desktopName": "SubMiner.desktop",
-  "version": "0.18.0-beta.2",
+  "version": "0.18.0-beta.3",
   "description": "All-in-one sentence mining overlay with AnkiConnect and dictionary integration",
   "packageManager": "bun@1.3.5",
   "main": "dist/main-entry.js",

release/prerelease-notes.md

@@ -21,17 +21,24 @@
 ### Fixed
 - **Character Name Highlighting in Subtitles**
   - Fixed unspaced Japanese names (e.g. 東紫乃, 渡辺真奈美) being split at the wrong point, which left surnames like 東 and 渡辺 without their character portrait or hover lookup.
-  - Fixed names getting cut off or losing their highlight when caught by the subtitle scanner's punctuation handling, or stripped entirely when grammar tagging misclassified the name token.
+  - Fixed names getting cut off or losing their highlight when caught by the subtitle scanner's punctuation handling, misclassified by grammar tagging, or swallowed entirely by a longer generic dictionary match (e.g. ヨータ disappearing inside a false とヨー match).
+  - Fixed a single unrecognized word in a subtitle line (like a stray interjection) causing character-name highlighting to drop for the whole line instead of just that word.
   - No action needed — existing data upgrades automatically the next time a matching name is seen.
 - **Known-Word Highlighting**
   - Words are no longer marked "known" (green) just because they share spelling with a known Anki card that actually teaches a different reading (e.g. 床 read as とこ no longer falsely matches a known 床/ゆか card).
+  - Kanji words are also no longer marked known just because a different mined word happens to share their reading (e.g. 渓谷/けいこく no longer falsely matches a known 警告/けいこく card).
   - Single-kana grammar tokens (particles like よ, え) no longer borrow an unrelated card's reading and get falsely painted as known.
   - Stats sessions now correctly reflect known-word counts again after the reading-aware matching upgrade, instead of showing 0 everywhere.
+- **Annotation Highlighting Refinements**
+  - Restored frequency/JLPT highlighting and vocabulary-stat counting for words like 確かに and やはり, which were wrongly treated as grammar noise.
+  - Kanji nouns that MeCab tags as "non-independent" (e.g. 日, 点, 以外) also keep their highlighting and stats counting again.
+  - Suffix-only tokens (e.g. さん, れる) are now excluded from JLPT/frequency highlighting by default to match how particles and interjections are treated; known-word highlighting for them still works, and this is configurable.
 - **Unparsed Subtitle Text**
   - Subtitle text the dictionary can't recognize (like a truncated verb form) is now still hoverable for lookup and correctly counted toward a sentence's difficulty, instead of showing as dead, non-interactive text.
 - **Kiku Manual Field Grouping**
   - Fixed the field-grouping dialog getting stuck invisible behind fullscreen video on Hyprland/Wayland, and failing silently on repeated attempts after the first use.
-  - Fixed a duplicate "Field grouping cancelled" notification appearing when grouping was cancelled via the trigger shortcut.
+  - Fixed a timed-out or failed grouping request leaving an invisible, stuck dialog covering the video; it now closes automatically so the overlay recovers.
+  - Fixed a duplicate "Field grouping cancelled" notification appearing when grouping was cancelled via the trigger shortcut, and added a proper error message for the previously-silent case where the original card can no longer be loaded.
 - **Secondary Subtitles**
   - Karaoke-style secondary subtitles (common in opening/ending songs) no longer spam dozens of lines down the screen; repeated lines are now collapsed and the subtitle area is capped to a strip at the top.
 - **YouTube Extraction**
@@ -59,7 +66,8 @@
 - fix(overlay): resolve unspaced Japanese name splits and scan recovery by @ksyasuda in #146
 - fix(tokenizer): prevent grammar tokens from borrowing known-word highlight via unrelated readings by @ksyasuda in #147
 - fix(stats): fetch cover art eagerly at session start instead of on series page visit by @ksyasuda in #148
-- fix(stats): parse v3 reading-aware known-word cache in stats server by @ksyasuda in #149
+- fix(overlay): keep frequency/JLPT highlight for kanji non-independent nouns by @ksyasuda in #150
+- fix(tokenizer): greedy name pre-pass to prevent generic matches swallowing character names by @ksyasuda in #151
 
 ## Installation