fix(subtitles): cap ASS \t nesting depth to prevent stack overflow

- guard override-block recursion in parseOverrideBlock at 8 levels; pathologically nested \t(...) tags no longer blow the call stack
- extract duplicate/animation-burst cue collapsing out of subtitle-cue-parser.ts into subtitle-cue-dedup.ts
- add regression tests for deep \t nesting, SRT/VTT brace stripping, and ASS content sniffed behind a lying .srt extension

changes/styled-subtitle-cue-parsing.md

@@ -0,0 +1,5 @@
+type: fixed
+area: subtitles
+
+- Heavily typeset ASS scripts (karaoke OP/ED, sign work) no longer fill the subtitle sidebar with garbage. Vector drawing runs (`\p1` … `\p0`) are no longer shown as subtitle text (e.g. `m 20 0 b 10 0 0 10 0 20 …`), and duplicate events in a parsed subtitle file collapse into one cue: identical text over an identical span (layered "shadow" copies), and per-frame animation bursts. An ASS burst has to prove itself with authoring evidence — a temporal tag (`\t`, `\move`, karaoke timing), an animated `Effect` column, or override values that change from event to event — plus one shared style and actor, so three rapid `えっ` reactions from three characters, or a sign repeated with the same static `\clip`, stay separate. SRT and VTT carry no such metadata, so there the run has to be at least five contiguous events all shorter than 0.1s, which is where ASS-to-SRT conversion leaves karaoke frames.
+- Subtitle text is now decoded from ASS exactly once, where it enters the app, and matches how mpv renders the same line (including `\N`, `\n`, `\h`, unclosed `{`, and the fact that `\{` is not an escape). Renderer, timing tracker, tokenizer and the tokenization cache take that decoded text as-is instead of each re-deriving it, so one authored line can no longer produce two different cache keys, and a cue that normalizes to nothing is no longer stored as subtitle text or cached under an empty key.

docs/architecture/2026-03-15-renderer-performance-design.md

@@ -64,18 +64,23 @@ External subtitle files only (SRT, VTT, ASS). Embedded subtitle tracks are out o
 A cue parser extracts both timing and text content from subtitle files for prefetching.
 
 **Parsed cue structure:**
+
 ```typescript
 interface SubtitleCue {
   startTime: number; // seconds
   endTime: number; // seconds
-  text: string;       // raw subtitle text
+  text: string; // plain text, decoded from the source format
 }
 ```
 
 **Supported formats:**
+
 - SRT/VTT: Regex-based parsing of timing lines + text content between timing blocks.
-- ASS: Parse `[Events]` section, extract `Dialogue:` lines, split on the first 9 commas only (ASS v4+ has 10 fields; the last field is Text which can itself contain commas). Strip ASS override tags (`{\...}`) from the text before storing.
+- ASS: Parse `[Events]` section, extract `Dialogue:` lines, read the field order from the `Format:` row, and take everything after the Text field index as the text (Text can itself contain commas).
-  ASS text fields contain inline override tags like `{\b1}`, `{\an8}`, `{\fad(200,300)}`. The cue parser strips these during extraction so the tokenizer receives clean text.
+
+**ASS decoding.** The parser is where ASS text is decoded, once, via `assToPlainText()` in `src/core/services/ass-text.ts`. That decoder mirrors mpv's `ass_to_plaintext` so a cue read from a file reads identically to the same line arriving live on `sub-text`: `{...}` override blocks are markup, `\pN … \p0` vector drawing runs are dropped rather than shown as text, `\N`/`\n`/`\h` are the only escapes (`\{`, `\}` and `\\` are not), and an unclosed `{` is rendered verbatim. Every layer downstream — renderer, timing tracker, tokenizer, tokenization cache keys — receives plain text and uses `normalizePlainSubtitleText()` for whitespace only, so nothing decodes the same string twice and one authored line always maps to one cache key.
+
+**Duplicate collapsing.** Typeset scripts emit one `Dialogue:` event per animation frame, plus layered copies of the same line. The parser collapses identical text over an identical span unconditionally, and collapses contiguous same-text runs of at least three events when the run looks like an animation. For ASS that means shared style and actor plus authoring evidence: a temporal tag (`\t`, `\move`, `\k`/`\kf`/`\ko`/`\K`, or anything wrapped in `\t(...)`), an animated `Effect` column (`Karaoke`, `Banner`, `Scroll`), or override values that change across the run. Static tags shared by every event (`\pos`, an identical `\clip`) are not evidence. SRT/VTT carry no such metadata, so there collapsing needs at least five contiguous events all under 0.1s — the frame timing left behind by ASS-to-SRT conversion. The parser keeps this authoring metadata (style, actor, layer, `Effect`, parsed override commands, source order) private; `parseSubtitleCues()` returns only `SubtitleCue`.
 
 #### Prefetch Service Lifecycle
 
@@ -153,6 +158,7 @@ tokens (already have frequencyRank values from parser-level applyFrequencyRanks)
 ### Dependency Analysis
 
 All annotations either depend on MeCab POS data or benefit from running after it:
+
 - **Known word marking:** Needs base tokens (surface/headword). No POS dependency, but no reason to run separately.
 - **Frequency filtering:** Uses `pos1Exclusions` and `pos2Exclusions` to clear frequency ranks on excluded tokens (particles, noise). Depends on MeCab POS data.
 - **JLPT marking:** Uses `shouldIgnoreJlptForMecabPos1` to filter. Depends on MeCab POS data.
@@ -169,18 +175,14 @@ function annotateTokens(tokens, deps, options): MergedToken[] {
 
   // Single pass: known word + frequency filtering + JLPT computed together
   const annotated = tokens.map((token) => {
-    const isKnown = nPlusOneEnabled
+    const isKnown = nPlusOneEnabled ? token.isKnown || computeIsKnown(token, deps) : false;
-      ? token.isKnown || computeIsKnown(token, deps)
-      : false;
 
     // Filter frequency rank using POS exclusions (rank values already set at parser level)
     const frequencyRank = frequencyEnabled
       ? filterFrequencyRank(token, pos1Exclusions, pos2Exclusions)
       : undefined;
 
-    const jlptLevel = jlptEnabled
+    const jlptLevel = jlptEnabled ? computeJlptLevel(token, deps.getJlptLevel) : undefined;
-      ? computeJlptLevel(token, deps.getJlptLevel)
-      : undefined;
 
     return { ...token, isKnown, frequencyRank, jlptLevel };
   });
@@ -221,6 +223,7 @@ Replace `document.createElement('span')` calls in the renderer with `templateSpa
 ### Current Behavior
 
 In `renderWithTokens` (`subtitle-render.ts`), each render cycle:
+
 1. Clears DOM with `innerHTML = ''`
 2. Creates a `DocumentFragment`
 3. Calls `document.createElement('span')` for each token (~10-15 per subtitle)
@@ -257,7 +260,7 @@ Full recycling (collecting old nodes, clearing attributes, reusing them) require
 ## Combined Impact Summary
 
 | Scenario                          | Before     | After      | Improvement |
-|----------|--------|-------|-------------|
+| --------------------------------- | ---------- | ---------- | ----------- |
 | Normal playback (prefetch-warmed) | ~200-320ms | ~30-50ms   | ~80-85%     |
 | Cache hit (repeated subtitle)     | ~72ms      | ~55-65ms   | ~10-20%     |
 | Cache miss (immediate seek)       | ~200-320ms | ~150-260ms | ~20-25%     |
@@ -267,16 +270,19 @@ Full recycling (collecting old nodes, clearing attributes, reusing them) require
 ## Files Summary
 
 ### New Files
+
 - `src/core/services/subtitle-prefetch.ts`
 - `src/core/services/subtitle-cue-parser.ts`
 
 ### Modified Files
+
 - `src/core/services/subtitle-processing-controller.ts` (expose `preCacheTokenization`)
 - `src/core/services/tokenizer/annotation-stage.ts` (batched single-pass)
 - `src/renderer/subtitle-render.ts` (template cloneNode)
 - `src/main.ts` (wire up prefetch service)
 
 ### Test Files
+
 - New tests for subtitle cue parser (SRT, VTT, ASS formats)
 - New tests for subtitle prefetch service (priority window, seek, pause/resume)
 - Updated tests for annotation stage (same behavior, new implementation)

src/core/services/ass-text.test.ts

@@ -0,0 +1,193 @@
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import {
+  assOverrideSignature,
+  assToPlainText,
+  collectAssOverrideCommands,
+  extractAssOverrideBlocks,
+  hasAssTemporalOverride,
+  isAnimatedAssEffectKind,
+  isAssTemporalCommand,
+  normalizePlainSubtitleText,
+  parseAssEffectField,
+} from './ass-text';
+
+test('assToPlainText drops vector drawing runs', () => {
+  assert.equal(
+    assToPlainText(
+      '{\\an5\\pos(730,1042)\\p1\\blur1}m 20 0 b 10 0 0 10 0 20 b 0 31 10 40 20 40 {\\p0}',
+    ),
+    '',
+  );
+});
+
+test('assToPlainText keeps text around drawing runs on the same event', () => {
+  assert.equal(
+    assToPlainText('{\\p1}m 0 0 l 10 10{\\p0}本文{\\p1}m 5 5 l 6 6{\\p0}続き'),
+    '本文続き',
+  );
+});
+
+test('assToPlainText leaves \\pos alone when no drawing mode is active', () => {
+  assert.equal(assToPlainText('{\\pos(960,1068)\\bord3}位置指定'), '位置指定');
+});
+
+test('assToPlainText does not read \\pos as a drawing tag', () => {
+  assert.equal(assToPlainText('{\\p1\\pos(1,2)}m 0 0 l 5 5'), '');
+});
+
+test('assToPlainText resolves line-break and space escapes', () => {
+  assert.equal(assToPlainText('一行目\\N二行目'), '一行目\n二行目');
+  assert.equal(assToPlainText('一行目\\n二行目'), '一行目\n二行目');
+  assert.equal(assToPlainText('一行目\\N二行目', ' '), '一行目 二行目');
+  assert.equal(assToPlainText('間\\h隔'), '間 隔');
+});
+
+test('assToPlainText matches mpv on brace and backslash sequences', () => {
+  // mpv has no `\{` / `\}` / `\\` escapes: the backslashes are literal text and the
+  // braces still open and close an override block.
+  assert.equal(assToPlainText('\\{注\\}'), '\\');
+  assert.equal(assToPlainText('\\\\N'), '\\\n');
+});
+
+test('assToPlainText renders an unclosed override block verbatim', () => {
+  // mpv shows the stray brace; guessing where the block ended can eat a whole line.
+  assert.equal(assToPlainText('本文{\\pos(1,2)'), '本文{\\pos(1,2)');
+});
+
+test('assToPlainText is idempotent', () => {
+  const samples = [
+    '{\\an5\\p1}m 0 0 l 5 5{\\p0}本文',
+    '\\{注\\}',
+    '\\\\N',
+    '本文{\\pos(1,2)',
+    '一行目\\N二行目\\h終わり',
+  ];
+
+  for (const sample of samples) {
+    const once = assToPlainText(sample);
+    assert.equal(assToPlainText(once), once, sample);
+  }
+});
+
+test('assToPlainText normalizes CRLF before converting', () => {
+  assert.equal(assToPlainText('一行目\r\n二行目'), '一行目\n二行目');
+});
+
+test('normalizePlainSubtitleText settles whitespace without decoding ASS', () => {
+  // A brace reaching this layer is literal text mpv chose to show, not markup.
+  assert.equal(normalizePlainSubtitleText('本文{\\pos(1,2)'), '本文{\\pos(1,2)');
+  assert.equal(normalizePlainSubtitleText('一行目\\N二行目'), '一行目\n二行目');
+  assert.equal(
+    normalizePlainSubtitleText('一行目\\N二行目', { collapseLineBreaks: true }),
+    '一行目 二行目',
+  );
+  assert.equal(normalizePlainSubtitleText('  余白  ', { trim: false }), '  余白  ');
+});
+
+test('normalizePlainSubtitleText is idempotent', () => {
+  for (const sample of ['一行目\\N二行目', '間\\h隔', '本文{\\pos(1,2)', '  余白  ']) {
+    const once = normalizePlainSubtitleText(sample);
+    assert.equal(normalizePlainSubtitleText(once), once, sample);
+  }
+});
+
+test('extractAssOverrideBlocks returns block contents', () => {
+  assert.deepEqual(extractAssOverrideBlocks('{\\an8}上{\\fad(200,200)}下'), [
+    '\\an8',
+    '\\fad(200,200)',
+  ]);
+  assert.deepEqual(extractAssOverrideBlocks('括弧なし'), []);
+});
+
+test('collectAssOverrideCommands captures names and arguments from blocks only', () => {
+  const commands = collectAssOverrideCommands('{\\pos(1,2)\\1c&HFFFFFF&\\kf30}歌詞');
+
+  assert.deepEqual(commands, [
+    { name: 'pos', args: '1,2', animated: false },
+    { name: '1c', args: '&HFFFFFF&', animated: false },
+    { name: 'kf', args: '30', animated: false },
+  ]);
+
+  // A `\pos(...)` sitting in visible text is not typesetting markup.
+  assert.deepEqual(collectAssOverrideCommands('\\pos(730,1042) と書いてある'), []);
+});
+
+test('collectAssOverrideCommands marks tags animated by a wrapping \\t', () => {
+  const commands = collectAssOverrideCommands('{\\clip(0,0,10,10)\\t(0,500,\\frz30)}文字');
+
+  assert.deepEqual(
+    commands.map((command) => [command.name, command.animated]),
+    [
+      ['clip', false],
+      ['t', false],
+      ['frz', true],
+    ],
+  );
+  assert.equal(hasAssTemporalOverride(commands), true);
+});
+
+test('collectAssOverrideCommands survives pathologically nested \\t tags', () => {
+  const depth = 200000;
+  const block = `{${'\\t(0,500,'.repeat(depth)}\\frz30${')'.repeat(depth)}}文字`;
+
+  const commands = collectAssOverrideCommands(block);
+
+  // Recursion stops at the nesting cap; the outer tags are still reported, and nothing
+  // blows the call stack.
+  assert.equal(commands[0]!.name, 't');
+  assert.equal(hasAssTemporalOverride(commands), true);
+});
+
+test('hasAssTemporalOverride ignores static placement and shape tags', () => {
+  assert.equal(
+    hasAssTemporalOverride(collectAssOverrideCommands('{\\pos(1,2)\\clip(m 1 1)\\blur2}文字')),
+    false,
+  );
+  assert.equal(hasAssTemporalOverride(collectAssOverrideCommands('{\\move(1,2,3,4)}文字')), true);
+});
+
+test('isAssTemporalCommand covers only intrinsically animated tags', () => {
+  for (const command of ['t', 'move', 'k', 'kf', 'ko', 'K']) {
+    assert.equal(isAssTemporalCommand(command), true, command);
+  }
+  for (const command of ['clip', 'iclip', 'frz', 'fscx', 'blur', 'be', 'pos', 'fad']) {
+    assert.equal(isAssTemporalCommand(command), false, command);
+  }
+});
+
+test('assOverrideSignature distinguishes events by their override values', () => {
+  const first = assOverrideSignature(collectAssOverrideCommands('{\\clip(m 1 1)}歌詞'));
+  const second = assOverrideSignature(collectAssOverrideCommands('{\\clip(m 2 2)}歌詞'));
+  const repeat = assOverrideSignature(collectAssOverrideCommands('{\\clip(m 1 1)}別の行'));
+
+  assert.notEqual(first, second);
+  assert.equal(first, repeat);
+});
+
+test('parseAssEffectField classifies the event-level Effect column', () => {
+  assert.equal(parseAssEffectField(''), 'none');
+  assert.equal(parseAssEffectField('   '), 'none');
+  assert.equal(parseAssEffectField('Banner;20;1;0'), 'banner');
+  assert.equal(parseAssEffectField('Scroll up;0;0;30;10'), 'scroll');
+  assert.equal(parseAssEffectField('Scroll down;0;0;30;10'), 'scroll');
+  assert.equal(parseAssEffectField('Karaoke'), 'karaoke');
+  assert.equal(parseAssEffectField('fx-template'), 'other');
+});
+
+test('parseAssEffectField matches stock effect names exactly', () => {
+  // Custom effect names that merely start with a stock name are not stock effects.
+  assert.equal(parseAssEffectField('scrolling-credit'), 'other');
+  assert.equal(parseAssEffectField('bannerfx;1'), 'other');
+  assert.equal(parseAssEffectField('karaoke-template'), 'other');
+  assert.equal(parseAssEffectField('Scroll'), 'other');
+});
+
+test('isAnimatedAssEffectKind covers the stock animated effects only', () => {
+  assert.equal(isAnimatedAssEffectKind('karaoke'), true);
+  assert.equal(isAnimatedAssEffectKind('banner'), true);
+  assert.equal(isAnimatedAssEffectKind('scroll'), true);
+  // Typesetting groups put static template names in this column too.
+  assert.equal(isAnimatedAssEffectKind('other'), false);
+  assert.equal(isAnimatedAssEffectKind('none'), false);
+});

src/core/services/ass-text.ts

@@ -0,0 +1,280 @@
+/*
+ * ASS/SSA text handling, split into two deliberately distinct contracts:
+ *
+ *   assToPlainText()             raw ASS event text -> plain text. Ingestion only.
+ *   normalizePlainSubtitleText() already-decoded text -> display/lookup form.
+ *
+ * Subtitle text is decoded from ASS exactly once, at the point it enters the app: the
+ * file cue parser does it for sidecar/embedded scripts, and mpv does it for live text
+ * (`sub-text` is already run through mpv's own `ass_to_plaintext`). Everything
+ * downstream -- renderer, timing tracker, tokenizer, tokenization cache keys -- gets
+ * plain text and only normalizes whitespace, so no layer decodes the same string twice.
+ *
+ * assToPlainText mirrors mpv's `ass_to_plaintext` rather than inventing its own rules,
+ * so a cue parsed from a file reads the same as the same line arriving live:
+ *   - `{...}` override blocks are markup
+ *   - `\pN ... \p0` runs are vector paths, not dialogue
+ *   - `\N`, `\n` and `\h` are the only escapes; `\{`, `\}` and `\\` are NOT escapes,
+ *     so `\{注\}` decodes to a lone backslash exactly as mpv renders it
+ *   - an unclosed `{` is rendered verbatim instead of swallowing the rest of the line
+ * Because the decoder never emits an escape or a closed brace, running it twice is a
+ * no-op -- but downstream code should still use normalizePlainSubtitleText.
+ */
+
+/** What `\N` and `\n` become. */
+export type AssLineBreak = '\n' | ' ';
+
+// `\p<n>` with n > 0 switches libass into vector-drawing mode: everything until the
+// next `\p0` is a path (`m 20 0 b 10 0 ...`), not dialogue. The negative lookahead keeps
+// `\pos(...)` from being read as a drawing tag.
+const ASS_DRAWING_SCALE_PATTERN = /\\p(?![a-zA-Z])(\d*)/g;
+
+function readDrawingScale(block: string): number | null {
+  ASS_DRAWING_SCALE_PATTERN.lastIndex = 0;
+  let scale: number | null = null;
+  let match: RegExpExecArray | null;
+  // Drawing mode is whatever the last `\p` tag in this block set it to.
+  while ((match = ASS_DRAWING_SCALE_PATTERN.exec(block)) !== null) {
+    scale = match[1] ? Number(match[1]) : 0;
+  }
+  return scale;
+}
+
+/** Resolve `\N`, `\n` and `\h`. The only text-level escapes libass recognises. */
+function resolveWhitespaceEscapes(text: string, lineBreak: AssLineBreak): string {
+  return text.replace(/\\([Nnh])/g, (_match, escaped: string) =>
+    escaped === 'h' ? ' ' : lineBreak,
+  );
+}
+
+/** Strip `{...}` override blocks and the drawing runs they enable. */
+function stripAssMarkup(raw: string): string {
+  let out = '';
+  let cursor = 0;
+  let drawing = false;
+
+  while (cursor < raw.length) {
+    if (raw[cursor] !== '{') {
+      if (!drawing) {
+        out += raw[cursor];
+      }
+      cursor += 1;
+      continue;
+    }
+
+    const close = raw.indexOf('}', cursor + 1);
+    if (close === -1) {
+      // mpv shows an unclosed `{` and everything after it. Guessing where the block was
+      // meant to end can eat a whole line of dialogue.
+      if (!drawing) {
+        out += raw.slice(cursor);
+      }
+      break;
+    }
+
+    const scale = readDrawingScale(raw.slice(cursor, close + 1));
+    if (scale !== null) {
+      drawing = scale > 0;
+    }
+    cursor = close + 1;
+  }
+
+  return out;
+}
+
+/**
+ * Decode a raw ASS/SSA event text field. Call this once, where the text enters the app;
+ * downstream layers take the result as plain text.
+ */
+export function assToPlainText(text: string, lineBreak: AssLineBreak = '\n'): string {
+  if (!text) return '';
+  return resolveWhitespaceEscapes(stripAssMarkup(text.replace(/\r\n/g, '\n')), lineBreak);
+}
+
+export interface NormalizePlainSubtitleTextOptions {
+  /** Fold every line break into a single space. */
+  collapseLineBreaks?: boolean;
+  trim?: boolean;
+}
+
+/**
+ * Whitespace normalization for text that has already been decoded -- by mpv for live
+ * subtitles, by the cue parser for files. Override blocks and drawing runs are none of
+ * this function's business; a `{` that reaches here is literal text mpv chose to show.
+ *
+ * `\N`/`\n`/`\h` are still folded, because subtitle sources outside the ASS path (asbplayer
+ * and other websocket clients) forward them raw and the display layer has to cope.
+ */
+export function normalizePlainSubtitleText(
+  text: string,
+  options: NormalizePlainSubtitleTextOptions = {},
+): string {
+  if (!text) return '';
+  const { collapseLineBreaks = false, trim = true } = options;
+
+  let normalized = resolveWhitespaceEscapes(
+    text.replace(/\r\n/g, '\n'),
+    collapseLineBreaks ? ' ' : '\n',
+  );
+  if (collapseLineBreaks) {
+    normalized = normalized.replace(/\n/g, ' ').replace(/\s+/g, ' ');
+  }
+
+  return trim ? normalized.trim() : normalized;
+}
+
+/** The contents of each `{...}` block, without the braces. */
+export function extractAssOverrideBlocks(text: string): string[] {
+  const blocks: string[] = [];
+  let cursor = 0;
+
+  while (cursor < text.length) {
+    const open = text.indexOf('{', cursor);
+    if (open === -1) {
+      break;
+    }
+    const close = text.indexOf('}', open + 1);
+    if (close === -1) {
+      break;
+    }
+    blocks.push(text.slice(open + 1, close));
+    cursor = close + 1;
+  }
+
+  return blocks;
+}
+
+export interface AssOverrideCommand {
+  /** Tag name without the backslash, e.g. `pos`, `kf`, `1c`. */
+  name: string;
+  /** Everything the tag was given, e.g. `960,1068` for `\pos(960,1068)`. */
+  args: string;
+  /** Nested inside a `\t(...)` argument, so its value is animated over the event. */
+  animated: boolean;
+}
+
+const ASS_OVERRIDE_NAME_PATTERN = /[1-4]?[a-zA-Z]+/y;
+
+function readCommandArgs(block: string, start: number): { args: string; next: number } {
+  if (block[start] === '(') {
+    let depth = 0;
+    for (let i = start; i < block.length; i += 1) {
+      if (block[i] === '(') depth += 1;
+      else if (block[i] === ')') {
+        depth -= 1;
+        if (depth === 0) {
+          return { args: block.slice(start + 1, i), next: i + 1 };
+        }
+      }
+    }
+    return { args: block.slice(start + 1), next: block.length };
+  }
+
+  const nextTag = block.indexOf('\\', start);
+  const end = nextTag === -1 ? block.length : nextTag;
+  return { args: block.slice(start, end), next: end };
+}
+
+// `\t(...)` can wrap another `\t(...)`, and nothing in the format stops an author (or a
+// malformed file) from nesting them thousands deep. Real typesetting never goes past one
+// or two levels, so stop recursing well before the call stack is at risk.
+const MAX_ANIMATION_NESTING_DEPTH = 8;
+
+function parseOverrideBlock(
+  block: string,
+  animated: boolean,
+  into: AssOverrideCommand[],
+  depth = 0,
+): void {
+  let cursor = 0;
+
+  while (cursor < block.length) {
+    if (block[cursor] !== '\\') {
+      cursor += 1;
+      continue;
+    }
+
+    ASS_OVERRIDE_NAME_PATTERN.lastIndex = cursor + 1;
+    const nameMatch = ASS_OVERRIDE_NAME_PATTERN.exec(block);
+    if (!nameMatch) {
+      cursor += 1;
+      continue;
+    }
+
+    const name = nameMatch[0];
+    const { args, next } = readCommandArgs(block, cursor + 1 + name.length);
+    into.push({ name, args: args.trim(), animated });
+    // `\t(0,500,\frz30)` animates whatever it wraps, so record the inner tags too.
+    if (name === 't' && args.includes('\\') && depth < MAX_ANIMATION_NESTING_DEPTH) {
+      parseOverrideBlock(args, true, into, depth + 1);
+    }
+    cursor = next;
+  }
+}
+
+/**
+ * Override commands with their arguments, in source order. Only `{...}` blocks are
+ * inspected, so a `\pos(...)` sitting in visible text is never mistaken for markup.
+ */
+export function collectAssOverrideCommands(text: string): AssOverrideCommand[] {
+  const commands: AssOverrideCommand[] = [];
+  for (const block of extractAssOverrideBlocks(text)) {
+    parseOverrideBlock(block, false, commands);
+  }
+  return commands;
+}
+
+// Tags that are animated by definition: `\t` interpolates, `\move` travels, and the
+// karaoke tags advance a highlight across the event's own duration. Everything else --
+// `\pos`, `\clip`, `\frz`, `\blur`, `\fad` -- is a static value for the event, so its
+// presence says nothing about whether neighbouring events form one animation.
+const ASS_TEMPORAL_COMMANDS = new Set(['t', 'move', 'k', 'kf', 'ko', 'K']);
+
+export function isAssTemporalCommand(name: string): boolean {
+  return ASS_TEMPORAL_COMMANDS.has(name);
+}
+
+/** True when the event animates on its own, or animates a static tag through `\t(...)`. */
+export function hasAssTemporalOverride(commands: readonly AssOverrideCommand[]): boolean {
+  return commands.some((command) => command.animated || isAssTemporalCommand(command.name));
+}
+
+/**
+ * Canonical form of an event's override values, for comparing consecutive events. Two
+ * events with the same signature were typeset identically, so neither is a frame of an
+ * animation the other belongs to.
+ */
+export function assOverrideSignature(commands: readonly AssOverrideCommand[]): string {
+  return commands.map((command) => `${command.name}(${command.args})`).join('|');
+}
+
+export type AssEffectKind = 'none' | 'banner' | 'scroll' | 'karaoke' | 'other';
+
+// The stock effects, matched exactly. Typesetting groups put their own template names in
+// this column -- `scrolling-credit` is a static sign, not libass's `Scroll up` -- so a
+// prefix match would hand out animation evidence to arbitrary custom effects.
+const STOCK_ASS_EFFECTS = new Map<string, AssEffectKind>([
+  ['banner', 'banner'],
+  ['scroll up', 'scroll'],
+  ['scroll down', 'scroll'],
+  ['karaoke', 'karaoke'],
+]);
+
+/**
+ * The event-level `Effect` column. The stock values (`Banner;...`, `Scroll up;...`,
+ * `Scroll down;...`, `Karaoke`) all animate; anything else is a custom name and lands in
+ * `other`.
+ */
+export function parseAssEffectField(raw: string): AssEffectKind {
+  const value = raw.trim().toLowerCase();
+  if (!value) return 'none';
+
+  const name = value.split(';', 1)[0]!.trim();
+  return STOCK_ASS_EFFECTS.get(name) ?? 'other';
+}
+
+const ANIMATED_ASS_EFFECT_KINDS = new Set<AssEffectKind>(['banner', 'scroll', 'karaoke']);
+
+export function isAnimatedAssEffectKind(kind: AssEffectKind): boolean {
+  return ANIMATED_ASS_EFFECT_KINDS.has(kind);
+}

src/core/services/subtitle-cue-dedup.ts

@@ -0,0 +1,185 @@
+/*
+ * Duplicate/animation-burst collapsing for parsed subtitle cues.
+ *
+ * Split out of the cue parser so the parsing rules and the "is this run one animation?"
+ * heuristics can be read -- and tested -- on their own. The parser owns the cue shape;
+ * this module only decides which cues survive.
+ */
+
+import { hasAssTemporalOverride, isAnimatedAssEffectKind } from './ass-text';
+import type {
+  AnnotatedSubtitleCue,
+  SubtitleCue,
+  SubtitleSourceFormat,
+} from './subtitle-cue-parser';
+
+// Back-to-back frames of the same animation are authored flush against each other; a
+// tiny tolerance absorbs the centisecond rounding of the ASS timestamp format.
+const DUPLICATE_CUE_GAP_TOLERANCE_SECONDS = 0.05;
+// A burst is a *sequence*. Two adjacent events are two events, not an animation --
+// characters do repeat each other, and a repeated line can legitimately be short.
+const MIN_BURST_EVENTS = 3;
+// Real dialogue holds on screen for about a second, so a run with a couple of much
+// shorter events among them looks like frames. Used only alongside authoring evidence.
+const ANIMATION_FRAME_MAX_SECONDS = 0.3;
+// A karaoke run usually ends on a long "hold" frame, so not every event is short.
+const MIN_TAGGED_BURST_FRAMES = 2;
+// SRT and VTT carry no authoring metadata at all, so timing is the only signal available
+// -- which makes it the easiest one to get wrong. ASS->SRT conversion leaves frames at
+// ~0.04s, well under any real utterance, and a burst leaves many of them behind. Both
+// bounds are deliberately far stricter than the ASS path: a run of ordinary short lines
+// (`えっ` traded between characters) must not clear them.
+const TIMING_ONLY_FRAME_MAX_SECONDS = 0.1;
+const MIN_TIMING_ONLY_FRAMES = 5;
+
+function cueKey(cue: SubtitleCue): string {
+  return `${cue.startTime}|${cue.endTime}|${cue.text}`;
+}
+
+/**
+ * Identical text over an identical span is redundant however it was authored -- most
+ * often a layered ASS event stacking a shadow copy under the visible one.
+ */
+function collapseExactDuplicates(cues: AnnotatedSubtitleCue[]): AnnotatedSubtitleCue[] {
+  const seen = new Set<string>();
+  return cues.filter((cue) => {
+    const key = cueKey(cue);
+    if (seen.has(key)) {
+      return false;
+    }
+    seen.add(key);
+    return true;
+  });
+}
+
+function countFramesShorterThan(run: AnnotatedSubtitleCue[], maxSeconds: number): number {
+  return run.filter((cue) => cue.endTime - cue.startTime < maxSeconds).length;
+}
+
+/**
+ * Evidence that a run of ASS events is one animation rather than several authored lines.
+ * A static tag says nothing on its own -- three events sharing one `\clip(...)` are three
+ * signs -- so the tag has to be temporal by nature (`\t`, `\move`, karaoke timing, or
+ * anything wrapped in `\t(...)`), an animated `Effect` column, or a value that actually
+ * changes from event to event, which is how per-frame typesetting is authored.
+ */
+export function hasAssAnimationEvidence(run: AnnotatedSubtitleCue[]): boolean {
+  if (run.every((cue) => hasAssTemporalOverride(cue.overrides))) {
+    return true;
+  }
+  if (run.every((cue) => isAnimatedAssEffectKind(cue.effectKind))) {
+    return true;
+  }
+
+  const [first] = run;
+  const everyEventTypeset = run.every((cue) => cue.overrides.length > 0);
+  const signatureChanges = run.some((cue) => cue.overrideSignature !== first!.overrideSignature);
+  return everyEventTypeset && signatureChanges;
+}
+
+export function isAnimationBurst(
+  run: AnnotatedSubtitleCue[],
+  format: SubtitleSourceFormat,
+): boolean {
+  if (run.length < MIN_BURST_EVENTS) {
+    return false;
+  }
+
+  if (format === 'srt') {
+    return (
+      run.length >= MIN_TIMING_ONLY_FRAMES &&
+      countFramesShorterThan(run, TIMING_ONLY_FRAME_MAX_SECONDS) === run.length
+    );
+  }
+
+  if (countFramesShorterThan(run, ANIMATION_FRAME_MAX_SECONDS) < MIN_TAGGED_BURST_FRAMES) {
+    return false;
+  }
+
+  // One animation belongs to one styled, one named source line. Two characters trading
+  // the same short word are two styles or two actors, and never merge.
+  const [first] = run;
+  if (run.some((cue) => cue.style !== first!.style || cue.name !== first!.name)) {
+    return false;
+  }
+
+  return hasAssAnimationEvidence(run);
+}
+
+/**
+ * Karaoke and sign typesetting emits one Dialogue event per animation frame, all carrying
+ * the same visible text over a contiguous span. Collapse each such run into a single cue.
+ *
+ * Only runs that look like animation collapse. Two ordinary lines that happen to repeat
+ * -- several characters each saying `おはよう` in turn, a positioned sign redrawn with a
+ * different fade -- stay separate, because merging them would destroy real mineable lines.
+ */
+function collapseAnimationBursts(
+  cues: AnnotatedSubtitleCue[],
+  format: SubtitleSourceFormat,
+): AnnotatedSubtitleCue[] {
+  const indicesByText = new Map<string, number[]>();
+  cues.forEach((cue, index) => {
+    const bucket = indicesByText.get(cue.text);
+    if (bucket) {
+      bucket.push(index);
+    } else {
+      indicesByText.set(cue.text, [index]);
+    }
+  });
+
+  const dropped = new Set<number>();
+  const extendedEnd = new Map<number, number>();
+
+  for (const indices of indicesByText.values()) {
+    if (indices.length < MIN_BURST_EVENTS) {
+      continue;
+    }
+
+    let runStart = 0;
+    while (runStart < indices.length) {
+      let runEnd = runStart;
+      let chainEnd = cues[indices[runStart]!]!.endTime;
+
+      while (runEnd + 1 < indices.length) {
+        const next = cues[indices[runEnd + 1]!]!;
+        if (next.startTime > chainEnd + DUPLICATE_CUE_GAP_TOLERANCE_SECONDS) {
+          break;
+        }
+        chainEnd = Math.max(chainEnd, next.endTime);
+        runEnd += 1;
+      }
+
+      const run = indices.slice(runStart, runEnd + 1).map((index) => cues[index]!);
+      if (isAnimationBurst(run, format)) {
+        for (let i = runStart + 1; i <= runEnd; i += 1) {
+          dropped.add(indices[i]!);
+        }
+        extendedEnd.set(indices[runStart]!, chainEnd);
+      }
+
+      runStart = runEnd + 1;
+    }
+  }
+
+  if (dropped.size === 0) {
+    return cues;
+  }
+
+  const merged: AnnotatedSubtitleCue[] = [];
+  cues.forEach((cue, index) => {
+    if (dropped.has(index)) {
+      return;
+    }
+    const end = extendedEnd.get(index);
+    merged.push(end !== undefined && end > cue.endTime ? { ...cue, endTime: end } : cue);
+  });
+  return merged;
+}
+
+export function mergeDuplicateCues(
+  cues: AnnotatedSubtitleCue[],
+  format: SubtitleSourceFormat,
+): AnnotatedSubtitleCue[] {
+  return collapseAnimationBursts(collapseExactDuplicates(cues), format);
+}

src/core/services/subtitle-cue-parser.test.ts

@@ -91,6 +91,17 @@ test('parseSrtCues skips malformed timing lines gracefully', () => {
   assert.equal(cues[0]!.text, '有効');
 });
 
+test('parseSubtitleCues strips complete brace blocks from SRT and VTT text', () => {
+  const content = ['1', '00:00:01,000 --> 00:00:02,000', '彼は{謎}と言った', ''].join('\n');
+
+  for (const filename of ['test.srt', 'test.vtt']) {
+    const cues = parseSubtitleCues(content, filename);
+
+    assert.equal(cues.length, 1, filename);
+    assert.equal(cues[0]!.text, '彼はと言った', filename);
+  }
+});
+
 test('parseAssCues parses basic ASS dialogue lines', () => {
   const content = [
     '[Script Info]',
@@ -137,7 +148,9 @@ test('parseAssCues handles text containing commas', () => {
   assert.equal(cues[0]!.text, 'はい、そうです、ね');
 });
 
-test('parseAssCues handles \\N line breaks', () => {
+test('parseAssCues decodes \\N line breaks into real newlines', () => {
+  // ASS is decoded once, here at ingestion, so cue text matches what mpv hands over for
+  // the same line played live.
   const content = [
     '[Events]',
     'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
@@ -146,7 +159,7 @@ test('parseAssCues handles \\N line breaks', () => {
 
   const cues = parseAssCues(content);
 
-  assert.equal(cues[0]!.text, '一行目\\N二行目');
+  assert.equal(cues[0]!.text, '一行目\n二行目');
 });
 
 test('parseAssCues strips HTML-like markup while preserving ASS line breaks', () => {
@@ -158,7 +171,46 @@ test('parseAssCues strips HTML-like markup while preserving ASS line breaks', ()
 
   const cues = parseAssCues(content);
 
-  assert.equal(cues[0]!.text, '一行目\\N二行目');
+  assert.equal(cues[0]!.text, '一行目\n二行目');
+});
+
+test('parseAssCues drops vector drawing runs enabled by \\p', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 1,0:00:01.00,0:00:04.00,Default,,0,0,0,,{\\an5\\pos(730,1042)\\p1\\blur1}m 20 0 b 10 0 0 10 0 20 b 0 31 10 40 20 40 {\\p0}',
+    'Dialogue: 0,0:00:05.00,0:00:08.00,Default,,0,0,0,,これは字幕',
+  ].join('\n');
+
+  const cues = parseAssCues(content);
+
+  assert.equal(cues.length, 1);
+  assert.equal(cues[0]!.text, 'これは字幕');
+});
+
+test('parseAssCues keeps text that follows a \\p0 reset on the same line', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:04.00,Default,,0,0,0,,{\\p1}m 0 0 l 10 10{\\p0}本文{\\p1}m 5 5 l 6 6{\\p0}続き',
+  ].join('\n');
+
+  const cues = parseAssCues(content);
+
+  assert.equal(cues.length, 1);
+  assert.equal(cues[0]!.text, '本文続き');
+});
+
+test('parseAssCues leaves \\pos untouched when no drawing mode is active', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:04.00,Default,,0,0,0,,{\\pos(960,1068)\\bord3}位置指定',
+  ].join('\n');
+
+  const cues = parseAssCues(content);
+
+  assert.equal(cues[0]!.text, '位置指定');
 });
 
 test('parseAssCues returns empty for content without Events section', () => {
@@ -258,6 +310,344 @@ test('parseSubtitleCues returns cues sorted by start time', () => {
   assert.equal(cues[1]!.text, '二番目');
 });
 
+test('parseSubtitleCues collapses per-frame karaoke duplicates into one cue', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.05,OP_JP,,0,0,0,,{\\clip(m 1 1)}過ぎ去ってしまう瞬間を',
+    'Dialogue: 0,0:00:01.05,0:00:01.09,OP_JP,,0,0,0,,{\\clip(m 2 2)}過ぎ去ってしまう瞬間を',
+    'Dialogue: 0,0:00:01.09,0:00:03.55,OP_JP,,0,0,0,,{\\clip(m 3 3)}過ぎ去ってしまう瞬間を',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 1);
+  assert.equal(cues[0]!.startTime, 1.0);
+  assert.equal(cues[0]!.endTime, 3.55);
+  assert.equal(cues[0]!.text, '過ぎ去ってしまう瞬間を');
+});
+
+test('parseSubtitleCues keeps back-to-back plain dialogue repeats separate', () => {
+  // Several characters greeting in turn: distinct utterances that happen to abut.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:04:05.67,0:04:06.82,Dial_JP,,0,0,0,,おはよう',
+    'Dialogue: 0,0:04:06.82,0:04:07.56,Dial_JP,,0,0,0,,おはよう',
+    'Dialogue: 0,0:04:07.56,0:04:08.78,Dial_JP,,0,0,0,,おはよう',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 3);
+  assert.equal(cues[0]!.endTime, 246.82);
+  assert.equal(cues[2]!.startTime, 247.56);
+});
+
+test('parseSubtitleCues collapses exact duplicate cues even without effect tags', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:04.00,Default,,0,0,0,,重なった行',
+    'Dialogue: 1,0:00:01.00,0:00:04.00,Default,,0,0,0,,重なった行',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 1);
+});
+
+test('parseSubtitleCues collapses tag-less animation frames in converted SRT', () => {
+  // ASS -> SRT conversion drops override tags, so only the ~0.04s frame timing remains.
+  const lines = ['1', '00:00:07,870 --> 00:00:07,910', 'Kaguya Wants to be Confessed to', ''];
+  for (let i = 1; i < 8; i++) {
+    const start = 7910 + (i - 1) * 40;
+    const end = start + 40;
+    const at = (ms: number) =>
+      `00:00:0${Math.floor(ms / 1000)},${String(ms % 1000).padStart(3, '0')}`;
+    lines.push(String(i + 1), `${at(start)} --> ${at(end)}`, 'Kaguya Wants to be Confessed to', '');
+  }
+
+  const cues = parseSubtitleCues(lines.join('\n'), 'test.srt');
+
+  assert.equal(cues.length, 1);
+  assert.equal(cues[0]!.startTime, 7.87);
+});
+
+test('parseSubtitleCues keeps identical lines that recur far apart', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:02.00,Default,,0,0,0,,なんで',
+    'Dialogue: 0,0:05:00.00,0:05:01.00,Default,,0,0,0,,なんで',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 2);
+  assert.equal(cues[0]!.startTime, 1.0);
+  assert.equal(cues[1]!.startTime, 300.0);
+});
+
+test('parseSubtitleCues keeps two positioned signs that repeat the same text', () => {
+  // Both carry override tags, but `\pos` and `\fad` are static placement, not animation.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:01:00.00,0:01:03.00,Sign,,0,0,0,,{\\pos(960,120)\\fad(200,200)}第一話',
+    'Dialogue: 0,0:01:03.00,0:01:06.00,Sign,,0,0,0,,{\\pos(960,900)\\fad(200,200)}第一話',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 2);
+  assert.equal(cues[1]!.startTime, 63.0);
+});
+
+test('parseSubtitleCues keeps a run of ordinary positioned lines separate', () => {
+  // Three events is a sequence, but none of them runs at animation-frame speed.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:01:00.00,0:01:02.00,Sign,,0,0,0,,{\\pos(960,120)\\fad(100,100)}止まれ',
+    'Dialogue: 0,0:01:02.00,0:01:04.00,Sign,,0,0,0,,{\\pos(960,120)\\fad(100,100)}止まれ',
+    'Dialogue: 0,0:01:04.00,0:01:06.00,Sign,,0,0,0,,{\\pos(960,120)\\fad(100,100)}止まれ',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 3);
+});
+
+test('parseSubtitleCues keeps a short repeated SRT pair without burst evidence', () => {
+  const content = [
+    '1',
+    '00:00:01,000 --> 00:00:01,200',
+    'えっ',
+    '',
+    '2',
+    '00:00:01,200 --> 00:00:01,400',
+    'えっ',
+    '',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.srt');
+
+  assert.equal(cues.length, 2);
+});
+
+test('parseSubtitleCues collapses a burst marked only by the Effect column', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.05,OP_JP,,0,0,0,Karaoke,歌詞',
+    'Dialogue: 0,0:00:01.05,0:00:01.09,OP_JP,,0,0,0,Karaoke,歌詞',
+    'Dialogue: 0,0:00:01.09,0:00:03.55,OP_JP,,0,0,0,Karaoke,歌詞',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 1);
+  assert.equal(cues[0]!.endTime, 3.55);
+});
+
+test('parseSubtitleCues keeps a second karaoke burst that starts after a gap', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.05,OP_JP,,0,0,0,,{\\clip(m 1 1)}リフレイン',
+    'Dialogue: 0,0:00:01.05,0:00:01.09,OP_JP,,0,0,0,,{\\clip(m 2 2)}リフレイン',
+    'Dialogue: 0,0:00:01.09,0:00:03.00,OP_JP,,0,0,0,,{\\clip(m 3 3)}リフレイン',
+    'Dialogue: 0,0:00:20.00,0:00:20.05,OP_JP,,0,0,0,,{\\clip(m 1 1)}リフレイン',
+    'Dialogue: 0,0:00:20.05,0:00:20.09,OP_JP,,0,0,0,,{\\clip(m 2 2)}リフレイン',
+    'Dialogue: 0,0:00:20.09,0:00:22.00,OP_JP,,0,0,0,,{\\clip(m 3 3)}リフレイン',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 2);
+  assert.equal(cues[0]!.endTime, 3.0);
+  assert.equal(cues[1]!.startTime, 20.0);
+  assert.equal(cues[1]!.endTime, 22.0);
+});
+
+test('parseSubtitleCues does not merge a burst into unrelated dialogue between frames', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.05,OP_JP,,0,0,0,,{\\clip(m 1 1)}歌詞',
+    'Dialogue: 0,0:00:01.02,0:00:03.00,Dial_JP,,0,0,0,,別のセリフ',
+    'Dialogue: 0,0:00:01.05,0:00:01.09,OP_JP,,0,0,0,,{\\clip(m 2 2)}歌詞',
+    'Dialogue: 0,0:00:01.09,0:00:03.55,OP_JP,,0,0,0,,{\\clip(m 3 3)}歌詞',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 2);
+  assert.deepEqual(
+    cues.map((cue) => cue.text),
+    ['歌詞', '別のセリフ'],
+  );
+  assert.equal(cues[0]!.endTime, 3.55);
+});
+
+test('parseSubtitleCues keeps rapid ASS lines from different actors separate', () => {
+  // Three 200ms `えっ` reactions traded between characters. Fast, adjacent and identical,
+  // but authored as three lines: different styles and different actors.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.20,Dial_A,アリス,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.20,0:00:01.40,Dial_B,ボブ,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.40,0:00:01.60,Dial_C,キャロル,0,0,0,,えっ',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 3);
+});
+
+test('parseSubtitleCues reads the speaker column when it is spelled Actor', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Actor, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.20,Dial_JP,アリス,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.20,0:00:01.40,Dial_JP,ボブ,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.40,0:00:01.60,Dial_JP,キャロル,0,0,0,,えっ',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 3);
+});
+
+test('parseSubtitleCues does not treat a custom Effect name as animation', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.20,Sign,,0,0,0,scrolling-credit,制作',
+    'Dialogue: 0,0:00:01.20,0:00:01.40,Sign,,0,0,0,scrolling-credit,制作',
+    'Dialogue: 0,0:00:01.40,0:00:01.60,Sign,,0,0,0,scrolling-credit,制作',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 3);
+});
+
+test('parseSubtitleCues keeps rapid ASS lines that share a style but not an actor', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.20,Dial_JP,アリス,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.20,0:00:01.40,Dial_JP,ボブ,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.40,0:00:01.60,Dial_JP,キャロル,0,0,0,,えっ',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 3);
+});
+
+test('parseSubtitleCues keeps untagged rapid ASS repeats separate', () => {
+  // No overrides at all: timing-only evidence is an SRT/VTT fallback and must not apply
+  // to ASS, where the absence of typesetting is itself evidence of plain dialogue.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.05,Dial_JP,,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.05,0:00:01.10,Dial_JP,,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.10,0:00:01.15,Dial_JP,,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.15,0:00:01.20,Dial_JP,,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.20,0:00:01.25,Dial_JP,,0,0,0,,えっ',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 5);
+});
+
+test('parseSubtitleCues keeps repeated signs sharing one static clip', () => {
+  // `\clip` is a static shape for the event. Three events with the identical clip were
+  // typeset the same way, so none of them is a frame of the others.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.20,Sign,,0,0,0,,{\\clip(0,0,100,100)}注意',
+    'Dialogue: 0,0:00:01.20,0:00:01.40,Sign,,0,0,0,,{\\clip(0,0,100,100)}注意',
+    'Dialogue: 0,0:00:01.40,0:00:01.60,Sign,,0,0,0,,{\\clip(0,0,100,100)}注意',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 3);
+});
+
+test('parseSubtitleCues collapses a sign animated through \\t', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.20,Sign,,0,0,0,,{\\pos(10,10)\\t(0,200,\\frz30)}回る',
+    'Dialogue: 0,0:00:01.20,0:00:01.40,Sign,,0,0,0,,{\\pos(10,10)\\t(0,200,\\frz30)}回る',
+    'Dialogue: 0,0:00:01.40,0:00:03.00,Sign,,0,0,0,,{\\pos(10,10)\\t(0,200,\\frz30)}回る',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 1);
+  assert.equal(cues[0]!.endTime, 3.0);
+});
+
+test('parseSubtitleCues keeps a short repeated SRT run above the frame threshold', () => {
+  // Five contiguous 200ms cues: a sequence, but nowhere near animation-frame speed.
+  const lines: string[] = [];
+  for (let i = 0; i < 5; i++) {
+    const start = 1000 + i * 200;
+    const at = (ms: number) =>
+      `00:00:0${Math.floor(ms / 1000)},${String(ms % 1000).padStart(3, '0')}`;
+    lines.push(String(i + 1), `${at(start)} --> ${at(start + 200)}`, 'えっ', '');
+  }
+
+  const cues = parseSubtitleCues(lines.join('\n'), 'test.srt');
+
+  assert.equal(cues.length, 5);
+});
+
+test('parseSubtitleCues keeps a short SRT frame run below the minimum length', () => {
+  // Four 40ms frames: frame-speed, but too few to tell an animation from an artefact.
+  const lines: string[] = [];
+  for (let i = 0; i < 4; i++) {
+    const start = 7870 + i * 40;
+    const at = (ms: number) =>
+      `00:00:0${Math.floor(ms / 1000)},${String(ms % 1000).padStart(3, '0')}`;
+    lines.push(String(i + 1), `${at(start)} --> ${at(start + 40)}`, 'タイトル', '');
+  }
+
+  const cues = parseSubtitleCues(lines.join('\n'), 'test.srt');
+
+  assert.equal(cues.length, 4);
+});
+
+test('parseSubtitleCues applies ASS burst rules to ASS content behind an .srt filename', () => {
+  // The extension lies, so the SRT parser finds nothing and the content-sniffing fallback
+  // takes over -- which has to carry the `ass` source format with it, or the far stricter
+  // timing-only thresholds would let this karaoke burst through as three cues.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.20,Karaoke,,0,0,0,,{\\k20}歌詞',
+    'Dialogue: 0,0:00:01.20,0:00:01.40,Karaoke,,0,0,0,,{\\k20}歌詞',
+    'Dialogue: 0,0:00:01.40,0:00:03.00,Karaoke,,0,0,0,,{\\k20}歌詞',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.srt');
+
+  assert.equal(cues.length, 1);
+  assert.equal(cues[0]!.startTime, 1.0);
+  assert.equal(cues[0]!.endTime, 3.0);
+  assert.equal(cues[0]!.text, '歌詞');
+});
+
 test('parseSubtitleCues detects subtitle formats from remote URLs', () => {
   const assContent = [
     '[Events]',

src/core/services/subtitle-cue-parser.ts

@@ -1,9 +1,46 @@
+import {
+  assOverrideSignature,
+  assToPlainText,
+  collectAssOverrideCommands,
+  parseAssEffectField,
+  type AssEffectKind,
+  type AssOverrideCommand,
+} from './ass-text';
+import { mergeDuplicateCues } from './subtitle-cue-dedup';
+
 export interface SubtitleCue {
   startTime: number;
   endTime: number;
   text: string;
 }
 
+/**
+ * Everything the parser knows about a source event, shared only with the dedup engine.
+ * Deduplication needs the authoring context -- which style the line belongs to, which
+ * override commands it carries, whether the `Effect` column was set -- to tell a karaoke
+ * burst apart from two characters saying the same word in turn. None of it is meaningful
+ * outside the parser, so the public API stays `{startTime, endTime, text}`.
+ */
+export interface AnnotatedSubtitleCue extends SubtitleCue {
+  /** Text exactly as authored, override blocks and all. */
+  rawText: string;
+  style: string;
+  layer: number;
+  /** ASS `Name`/`Actor` column. */
+  name: string;
+  /** ASS `Effect` column, verbatim. */
+  effect: string;
+  effectKind: AssEffectKind;
+  /** Override commands found in `{...}` blocks, with their arguments. */
+  overrides: readonly AssOverrideCommand[];
+  /** Canonical form of `overrides`, for spotting values that change across a run. */
+  overrideSignature: string;
+  /** Position in the source file, so sorting by time stays deterministic across layers. */
+  order: number;
+}
+
+export type SubtitleSourceFormat = 'ass' | 'srt';
+
 const HTML_SUBTITLE_TAG_PATTERN = /<\/?[A-Za-z][^>\n]*>/g;
 
 const SRT_TIMING_PATTERN =
@@ -23,12 +60,21 @@ function parseTimestamp(
   );
 }
 
+/**
+ * The single ASS decode for the file path: cues leave the parser as plain text with real
+ * line breaks, matching what mpv hands over for the same line played live. No layer
+ * downstream decodes ASS again.
+ */
 function sanitizeSubtitleCueText(text: string): string {
-  return text.replace(ASS_OVERRIDE_TAG_PATTERN, '').replace(HTML_SUBTITLE_TAG_PATTERN, '').trim();
+  return assToPlainText(text, '\n').replace(HTML_SUBTITLE_TAG_PATTERN, '').trim();
 }
 
-export function parseSrtCues(content: string): SubtitleCue[] {
+function toPublicCues(cues: AnnotatedSubtitleCue[]): SubtitleCue[] {
-  const cues: SubtitleCue[] = [];
+  return cues.map(({ startTime, endTime, text }) => ({ startTime, endTime, text }));
+}
+
+function parseAnnotatedSrtCues(content: string): AnnotatedSubtitleCue[] {
+  const cues: AnnotatedSubtitleCue[] = [];
   const lines = content.split(/\r?\n/);
   let i = 0;
 
@@ -60,20 +106,39 @@ export function parseSrtCues(content: string): SubtitleCue[] {
       i += 1;
     }
 
-    const text = sanitizeSubtitleCueText(textLines.join('\n'));
+    const rawText = textLines.join('\n');
+    const text = sanitizeSubtitleCueText(rawText);
     if (text) {
-      cues.push({ startTime, endTime, text });
+      cues.push({
+        startTime,
+        endTime,
+        text,
+        rawText,
+        style: '',
+        layer: 0,
+        name: '',
+        effect: '',
+        effectKind: 'none',
+        // SRT and VTT carry no authoring metadata, and the dedup engine never reads
+        // overrides for those formats -- collecting them would be parsing for nobody.
+        overrides: [],
+        overrideSignature: '',
+        order: cues.length,
+      });
     }
   }
 
   return cues;
 }
 
-const ASS_OVERRIDE_TAG_PATTERN = /\{[^}]*\}/g;
+export function parseSrtCues(content: string): SubtitleCue[] {
+  return toPublicCues(parseAnnotatedSrtCues(content));
+}
 
 const ASS_TIMING_PATTERN = /^(\d+):(\d{2}):(\d{2})\.(\d{1,2})$/;
 const ASS_FORMAT_PREFIX = 'Format:';
 const ASS_DIALOGUE_PREFIX = 'Dialogue:';
+const ASS_NAME_FIELD_ALIASES = ['name', 'actor'];
 
 function parseAssTimestamp(raw: string): number | null {
   const match = ASS_TIMING_PATTERN.exec(raw.trim());
@@ -87,13 +152,43 @@ function parseAssTimestamp(raw: string): number | null {
   return hours * 3600 + minutes * 60 + seconds + centiseconds / 100;
 }
 
-export function parseAssCues(content: string): SubtitleCue[] {
+function readField(fields: string[], index: number): string {
-  const cues: SubtitleCue[] = [];
+  return index >= 0 && index < fields.length ? fields[index]!.trim() : '';
+}
+
+function findFieldIndex(formatFields: string[], aliases: string[]): number {
+  for (const alias of aliases) {
+    const index = formatFields.indexOf(alias);
+    if (index >= 0) {
+      return index;
+    }
+  }
+  return -1;
+}
+
+function parseAnnotatedAssCues(content: string): AnnotatedSubtitleCue[] {
+  const cues: AnnotatedSubtitleCue[] = [];
   const lines = content.split(/\r?\n/);
   let inEventsSection = false;
-  let startFieldIndex = -1;
+  const fieldIndex = {
-  let endFieldIndex = -1;
+    start: -1,
-  let textFieldIndex = -1;
+    end: -1,
+    text: -1,
+    style: -1,
+    layer: -1,
+    name: -1,
+    effect: -1,
+  };
+
+  const resetFieldIndex = () => {
+    fieldIndex.start = -1;
+    fieldIndex.end = -1;
+    fieldIndex.text = -1;
+    fieldIndex.style = -1;
+    fieldIndex.layer = -1;
+    fieldIndex.name = -1;
+    fieldIndex.effect = -1;
+  };
 
   for (const line of lines) {
     const trimmed = line.trim();
@@ -101,9 +196,7 @@ export function parseAssCues(content: string): SubtitleCue[] {
     if (trimmed.startsWith('[') && trimmed.endsWith(']')) {
       inEventsSection = trimmed.toLowerCase() === '[events]';
       if (!inEventsSection) {
-        startFieldIndex = -1;
+        resetFieldIndex();
-        endFieldIndex = -1;
-        textFieldIndex = -1;
       }
       continue;
     }
@@ -117,9 +210,15 @@ export function parseAssCues(content: string): SubtitleCue[] {
         .slice(ASS_FORMAT_PREFIX.length)
         .split(',')
         .map((field) => field.trim().toLowerCase());
-      startFieldIndex = formatFields.indexOf('start');
+      fieldIndex.start = formatFields.indexOf('start');
-      endFieldIndex = formatFields.indexOf('end');
+      fieldIndex.end = formatFields.indexOf('end');
-      textFieldIndex = formatFields.indexOf('text');
+      fieldIndex.text = formatFields.indexOf('text');
+      fieldIndex.style = formatFields.indexOf('style');
+      fieldIndex.layer = formatFields.indexOf('layer');
+      // Aegisub writes the speaker column as `Actor`; the v4+ spec calls it `Name`.
+      // Missing it costs the burst check its speaker guard, so both spellings count.
+      fieldIndex.name = findFieldIndex(formatFields, ASS_NAME_FIELD_ALIASES);
+      fieldIndex.effect = formatFields.indexOf('effect');
       continue;
     }
 
@@ -127,34 +226,57 @@ export function parseAssCues(content: string): SubtitleCue[] {
       continue;
     }
 
-    if (startFieldIndex < 0 || endFieldIndex < 0 || textFieldIndex < 0) {
+    if (fieldIndex.start < 0 || fieldIndex.end < 0 || fieldIndex.text < 0) {
       continue;
     }
 
     const fields = trimmed.slice(ASS_DIALOGUE_PREFIX.length).split(',');
     if (
-      startFieldIndex >= fields.length ||
+      fieldIndex.start >= fields.length ||
-      endFieldIndex >= fields.length ||
+      fieldIndex.end >= fields.length ||
-      textFieldIndex >= fields.length
+      fieldIndex.text >= fields.length
     ) {
       continue;
     }
 
-    const startTime = parseAssTimestamp(fields[startFieldIndex]!);
+    const startTime = parseAssTimestamp(fields[fieldIndex.start]!);
-    const endTime = parseAssTimestamp(fields[endFieldIndex]!);
+    const endTime = parseAssTimestamp(fields[fieldIndex.end]!);
     if (startTime === null || endTime === null) {
       continue;
     }
 
-    const text = sanitizeSubtitleCueText(fields.slice(textFieldIndex).join(','));
+    const rawText = fields.slice(fieldIndex.text).join(',');
-    if (text) {
+    const text = sanitizeSubtitleCueText(rawText);
-      cues.push({ startTime, endTime, text });
+    if (!text) {
+      continue;
     }
+
+    const effect = readField(fields, fieldIndex.effect);
+    const layer = Number(readField(fields, fieldIndex.layer));
+    const overrides = collectAssOverrideCommands(rawText);
+    cues.push({
+      startTime,
+      endTime,
+      text,
+      rawText,
+      style: readField(fields, fieldIndex.style),
+      layer: Number.isFinite(layer) ? layer : 0,
+      name: readField(fields, fieldIndex.name),
+      effect,
+      effectKind: parseAssEffectField(effect),
+      overrides,
+      overrideSignature: assOverrideSignature(overrides),
+      order: cues.length,
+    });
   }
 
   return cues;
 }
 
+export function parseAssCues(content: string): SubtitleCue[] {
+  return toPublicCues(parseAnnotatedAssCues(content));
+}
+
 function detectSubtitleFormat(source: string): 'srt' | 'vtt' | 'ass' | 'ssa' | null {
   const [normalizedSource = source] =
     (() => {
@@ -173,27 +295,31 @@ function detectSubtitleFormat(source: string): 'srt' | 'vtt' | 'ass' | 'ssa' | n
 
 export function parseSubtitleCues(content: string, filename: string): SubtitleCue[] {
   const format = detectSubtitleFormat(filename);
-  let cues: SubtitleCue[];
+  let cues: AnnotatedSubtitleCue[];
+  let sourceFormat: SubtitleSourceFormat = 'srt';
 
   switch (format) {
     case 'srt':
     case 'vtt':
-      cues = parseSrtCues(content);
+      cues = parseAnnotatedSrtCues(content);
       break;
     case 'ass':
     case 'ssa':
-      cues = parseAssCues(content);
+      cues = parseAnnotatedAssCues(content);
+      sourceFormat = 'ass';
       break;
     default:
       cues = [];
   }
 
   if (cues.length === 0) {
-    const assCues = parseAssCues(content);
+    const assCues = parseAnnotatedAssCues(content);
-    const srtCues = parseSrtCues(content);
+    const srtCues = parseAnnotatedSrtCues(content);
-    cues = assCues.length >= srtCues.length ? assCues : srtCues;
+    const preferAss = assCues.length >= srtCues.length;
+    cues = preferAss ? assCues : srtCues;
+    sourceFormat = preferAss && assCues.length > 0 ? 'ass' : 'srt';
   }
 
-  cues.sort((a, b) => a.startTime - b.startTime);
+  cues.sort((a, b) => a.startTime - b.startTime || a.endTime - b.endTime || a.order - b.order);
-  return cues;
+  return toPublicCues(mergeDuplicateCues(cues, sourceFormat));
 }

src/core/services/subtitle-processing-controller.test.ts

@@ -115,6 +115,21 @@ test('subtitle processing does not emit plain payload for cached lines', async (
   assert.deepEqual(emitted, [{ text: '字幕', tokens: [] }]);
 });
 
+test('text that normalizes to nothing is never cached', () => {
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => ({ text, tokens: [] }),
+    emitSubtitle: () => {},
+  });
+
+  // Two different inputs both reduce to an empty key; sharing one entry would serve the
+  // first one's tokens for the second.
+  controller.preCacheTokenization('  ', { text: '  ', tokens: [] });
+
+  assert.equal(controller.hasCachedSubtitle('  '), false);
+  assert.equal(controller.hasCachedSubtitle('\\n'), false);
+  assert.equal(controller.consumeCachedSubtitle('\\n'), null);
+});
+
 test('subtitle processing shows plain line while tokenization is still pending', async () => {
   const emitted: SubtitleData[] = [];
   let resolveTokenization: ((value: SubtitleData | null) => void) | undefined;

src/core/services/subtitle-processing-controller.ts

@@ -1,4 +1,5 @@
 import type { SubtitleData } from '../../types';
+import { normalizePlainSubtitleText } from './ass-text';
 
 export interface SubtitleProcessingControllerDeps {
   tokenizeSubtitle: (text: string) => Promise<SubtitleData | null>;
@@ -25,8 +26,15 @@ export interface SubtitleProcessingController {
   hasCachedSubtitle: (text: string) => boolean;
 }
 
+/**
+ * Prefetched cues and live mpv text are both already decoded from ASS, so the key only
+ * has to settle whitespace for one authored line to resolve to one entry.
+ *
+ * An empty key is not a line: it is whatever normalization reduced to nothing. Callers
+ * must skip the cache for it rather than let every such input share one entry.
+ */
 export function normalizeSubtitleCacheKey(text: string): string {
-  return text.replace(/\r\n/g, '\n').replace(/\\N/g, '\n').replace(/\\n/g, '\n').trim();
+  return normalizePlainSubtitleText(text);
 }
 
 export function createSubtitleProcessingController(
@@ -50,6 +58,9 @@ export function createSubtitleProcessingController(
 
   const getCachedTokenization = (text: string): SubtitleData | null => {
     const cacheKey = normalizeSubtitleCacheKey(text);
+    if (!cacheKey) {
+      return null;
+    }
     const cached = tokenizationCache.get(cacheKey);
     if (!cached) {
       return null;
@@ -61,7 +72,11 @@ export function createSubtitleProcessingController(
   };
 
   const setCachedTokenization = (text: string, payload: SubtitleData): void => {
-    tokenizationCache.set(normalizeSubtitleCacheKey(text), payload);
+    const cacheKey = normalizeSubtitleCacheKey(text);
+    if (!cacheKey) {
+      return;
+    }
+    tokenizationCache.set(cacheKey, payload);
     while (tokenizationCache.size > SUBTITLE_TOKENIZATION_CACHE_LIMIT) {
       const firstKey = tokenizationCache.keys().next().value;
       if (firstKey !== undefined) {
@@ -219,7 +234,8 @@ export function createSubtitleProcessingController(
       return cached;
     },
     hasCachedSubtitle: (text: string) => {
-      return tokenizationCache.has(normalizeSubtitleCacheKey(text));
+      const cacheKey = normalizeSubtitleCacheKey(text);
+      return cacheKey.length > 0 && tokenizationCache.has(cacheKey);
     },
   };
 }

src/core/services/tokenizer.test.ts

@@ -1651,9 +1651,11 @@ test('tokenizeSubtitle clears JLPT level from standalone Yomitan particle token'
   assert.equal(result.tokens?.[0]?.jlptLevel, undefined);
 });
 
-test('tokenizeSubtitle returns null tokens for empty normalized text', async () => {
+test('tokenizeSubtitle returns the normalized text when it comes out empty', async () => {
+  // Handing back the original would push whatever normalization dropped into app state
+  // as if it were subtitle text.
   const result = await tokenizeSubtitle(' \\n  ', makeDeps());
-  assert.deepEqual(result, { text: ' \\n  ', tokens: null });
+  assert.deepEqual(result, { text: '', tokens: null });
 });
 
 test('tokenizeSubtitle normalizes newlines before Yomitan parse request', async () => {

src/core/services/tokenizer.ts

@@ -27,6 +27,7 @@ import {
 } from './tokenizer/yomitan-parser-runtime';
 import type { YomitanTermFrequency } from './tokenizer/yomitan-parser-runtime';
 import { isKanaChar } from './tokenizer/token-classification';
+import { normalizePlainSubtitleText } from './ass-text';
 
 const logger = createLogger('main:tokenizer');
 
@@ -858,14 +859,14 @@ export async function tokenizeSubtitle(
   text: string,
   deps: TokenizerServiceDeps,
 ): Promise<SubtitleData> {
-  const displayText = text
+  const displayText = normalizePlainSubtitleText(text);
-    .replace(/\r\n/g, '\n')
-    .replace(/\\N/g, '\n')
-    .replace(/\\n/g, '\n')
-    .trim();
 
+  // ASS decoding already happened upstream (cue parser for files, mpv for live text), so
+  // all this drops is whitespace -- but a whitespace-only line still normalizes to empty.
+  // Return the normalized form anyway: handing back the original would put a blank line
+  // into application state as if it were subtitle text.
   if (!displayText) {
-    return { text, tokens: null };
+    return { text: displayText, tokens: null };
   }
 
   const tokenizeText = displayText

src/renderer/subtitle-render.test.ts

@@ -1004,6 +1004,24 @@ test('normalizeSubtitle collapses explicit line breaks when collapseLineBreaks i
   );
 });
 
+test('normalizeSubtitle leaves already-decoded text alone', () => {
+  // Primary subtitle text is decoded from ASS once, upstream: by mpv for live lines and
+  // by the cue parser for prefetched ones. A brace that survives that is literal text.
+  assert.equal(normalizeSubtitle('本文{\\pos(1,2)'), '本文{\\pos(1,2)');
+  assert.equal(normalizeSubtitle('  余白  ', false), '  余白  ');
+});
+
+test('prepareSecondarySubtitleLines drops ASS vector drawing runs', () => {
+  assert.deepEqual(
+    prepareSecondarySubtitleLines(
+      '{\\an5\\pos(730,1042)\\p1\\blur1}m 20 0 b 10 0 0 10 0 20 b 0 31 10 40 20 40 {\\p0}',
+    ),
+    [],
+  );
+  assert.deepEqual(prepareSecondarySubtitleLines('{\\p1}m 0 0 l 10 10{\\p0}本文'), ['本文']);
+  assert.deepEqual(prepareSecondarySubtitleLines('{\\pos(960,1068)\\bord3}位置指定'), ['位置指定']);
+});
+
 test('shouldRenderTokenizedSubtitle enables token rendering when tokens exist', () => {
   assert.equal(shouldRenderTokenizedSubtitle(5), true);
   assert.equal(shouldRenderTokenizedSubtitle(0), false);

src/renderer/subtitle-render.ts

@@ -5,6 +5,7 @@ import type {
   SubtitleData,
   SubtitleRendererStyleConfig,
 } from '../types';
+import { assToPlainText, normalizePlainSubtitleText } from '../core/services/ass-text.js';
 import type { RendererContext } from './context';
 import { PRIMARY_SUB_VISIBLE_ON_YOMITAN_POPUP_CLASS } from './yomitan-popup.js';
 
@@ -42,17 +43,10 @@ function isWhitespaceOnly(value: string): boolean {
   return value.trim().length === 0;
 }
 
+// Text reaching the overlay has already been decoded from ASS -- by mpv for live lines,
+// by the cue parser for prefetched ones -- so this only settles line breaks.
 export function normalizeSubtitle(text: string, trim = true, collapseLineBreaks = false): string {
-  if (!text) return '';
+  return normalizePlainSubtitleText(text, { trim, collapseLineBreaks });
-
-  let normalized = text.replace(/\\N/g, '\n').replace(/\\n/g, '\n');
-  normalized = normalized.replace(/\{[^}]*\}/g, '');
-  if (collapseLineBreaks) {
-    normalized = normalized.replace(/\n/g, ' ');
-    normalized = normalized.replace(/\s+/g, ' ');
-  }
-
-  return trim ? normalized.trim() : normalized;
 }
 
 const HEX_COLOR_PATTERN = /^#(?:[0-9a-fA-F]{3}|[0-9a-fA-F]{4}|[0-9a-fA-F]{6}|[0-9a-fA-F]{8})$/;
@@ -672,7 +666,10 @@ function isKaraokeLikeLineSet(lines: string[]): boolean {
 }
 
 export function prepareSecondarySubtitleLines(text: string): string[] {
-  const normalized = normalizeSubtitle(text, true, false);
+  // The one display-side ASS decode: secondary text also reaches the overlay from
+  // websocket clients that forward their source line untouched, so unlike the primary
+  // path it cannot assume mpv already decoded it.
+  const normalized = assToPlainText(text).trim();
 
   if (!normalized) return [];
 

src/subtitle-timing-tracker.ts

@@ -16,6 +16,8 @@
  * along with this program.  If not, see <https://www.gnu.org/licenses/>.
  */
 
+import { normalizePlainSubtitleText } from './core/services/ass-text';
+
 interface TimingEntry {
   startTime: number;
   endTime: number;
@@ -191,23 +193,14 @@ export class SubtitleTimingTracker {
     return costs[shorter.length] || 0;
   }
 
+  // Both sides take text mpv has already decoded from ASS; only whitespace differs
+  // between the lookup key (single line) and the display form (line breaks kept).
   private normalizeText(text: string): string {
-    return text
+    return normalizePlainSubtitleText(text, { collapseLineBreaks: true });
-      .replace(/\\N/g, ' ')
-      .replace(/\\n/g, ' ')
-      .replace(/\n/g, ' ')
-      .replace(/{[^}]*}/g, '')
-      .replace(/\s+/g, ' ')
-      .trim();
   }
 
   private prepareDisplayText(text: string): string {
-    // Convert ASS/SSA newlines to real newlines, strip tags
+    return normalizePlainSubtitleText(text);
-    return text
-      .replace(/\\N/g, '\n')
-      .replace(/\\n/g, '\n')
-      .replace(/{[^}]*}/g, '')
-      .trim();
   }
 
   private startCleanup(): void {