feat(subtitles): generate Jellyfin subtitles and widen halfwidth kana

- Whisper subtitle generation now works for Jellyfin playback. It reads the item's original file through toJellyfinOriginalStreamUrl and writes to a reused cache folder. For transcoded playback it uses Jellyfin's audio index
- generateJapaneseSubtitles accepts http(s) streams when an output path is set; subprocess errors redact URL tokens
- Add widenHalfwidthKana/narrowFullwidthKana and widen subtitle text at ingest (mpv sub-text, parsed cues, live primary text), so halfwidth and fullwidth duplicates merge into one cue
- Stats queries widen older stored lines, sentence search matches katakana in either width, and stats mining widens sentence/word
- Update Jellyfin, subtitle generation, and architecture docs; add changelog fragments

src/core/services/tokenizer/token-classification.ts

@@ -1,23 +1,12 @@
 import { MergedToken, PartOfSpeech } from '../../../types';
+import { widenHalfwidthKana } from '../../utils/halfwidth-kana';
 
 const KATAKANA_TO_HIRAGANA_OFFSET = 0x60;
 const KATAKANA_CODEPOINT_START = 0x30a1;
 const KATAKANA_CODEPOINT_END = 0x30f6;
 
-// No `u` flag: the range is entirely BMP so it changes nothing here, and
-// Bun's unicode-mode matcher mis-handles this class next to certain ligatures.
-const HALFWIDTH_KANA_RUN = /[\uff66-\uff9f]+/g;
-
-// NFKC over the halfwidth kana only, never the whole string: it composes the
-// voiced pairs (ｶ + ﾞ) into single characters so ｶﾞｸ compares equal to ガク
-// instead of counting one character longer than the word it spells, but run
-// over everything it would also rewrite unrelated text (① → 1, ㍑ → リットル).
-function composeHalfwidthKana(text: string): string {
-  return text.replace(HALFWIDTH_KANA_RUN, (run) => run.normalize('NFKC'));
-}
-
 export function normalizeKana(text: string): string {
-  const raw = composeHalfwidthKana(text).trim();
+  const raw = widenHalfwidthKana(text).trim();
   if (!raw) {
     return '';
   }