fix(subtitles): keep overlapping lines that join an already active cue (#221)

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

@@ -218,3 +218,18 @@ test('removeLiveGlyphFragmentLines leaves ordinary short lines alone', () => {
   const text = 'え\nはい。\nそうだな';
   assert.equal(removeLiveGlyphFragmentLines(text), text);
 });
+
+test('normalizePlainSubtitleText folds cue-boundary blank lines for text consumers', () => {
+  // The display layer splits on the blank line before normalizing; everyone else --
+  // tokenizer, cache key, dedup gate, mined sentence -- wants the plain line form.
+  assert.equal(
+    normalizePlainSubtitleText('\u4e00\u884c\u76ee\n\n\u4e8c\u884c\u76ee'),
+    '\u4e00\u884c\u76ee\n\u4e8c\u884c\u76ee',
+  );
+  assert.equal(
+    normalizePlainSubtitleText('\u4e00\u884c\u76ee\n\n\u4e8c\u884c\u76ee', {
+      collapseLineBreaks: true,
+    }),
+    '\u4e00\u884c\u76ee \u4e8c\u884c\u76ee',
+  );
+});

src/core/services/ass-text.ts

@@ -153,6 +153,10 @@ export function normalizePlainSubtitleText(
   );
   if (collapseLineBreaks) {
     normalized = normalized.replace(/\n/g, ' ').replace(/\s+/g, ' ');
+  } else {
+    // Simultaneous cues reach the display layer separated by a blank line; every other
+    // consumer wants the plain one-break-per-line form.
+    normalized = normalized.replace(/\n{2,}/g, '\n');
   }
 
   return trim ? normalized.trim() : normalized;

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

@@ -1927,9 +1927,7 @@ test('parseSubtitleCues does not double a line rendered whole beside its glyph s
     ['わ', 1022],
     ['ね', 1064],
   ] as const;
-  const wholeLine = glyphs
-    .map(([glyph]) => `{\\an5\\fad(300,500)\\pos(960,50)}${glyph}`)
-    .join('');
+  const wholeLine = glyphs.map(([glyph]) => `{\\an5\\fad(300,500)\\pos(960,50)}${glyph}`).join('');
   const content = [
     ...eventsHeader,
     `Dialogue: 1,0:00:17.29,0:00:18.99,OP - JP,,0,0,0,,${wholeLine}`,
@@ -1952,7 +1950,7 @@ test('parseSubtitleCues drops a wall of near-invisible positioned texture string
   // faint translation is one or two events and stays published.
   const content = [
     ...eventsHeader,
-    'Dialogue: 90,0:00:12.66,0:00:14.91,Default,,0,0,0,,We\'ll play as a band, and then...',
+    "Dialogue: 90,0:00:12.66,0:00:14.91,Default,,0,0,0,,We'll play as a band, and then...",
     ...Array.from(
       { length: 12 },
       (_, index) =>
@@ -1997,3 +1995,25 @@ test('parseSubtitleCues keeps hidden events hidden when a transform animates an
     ['grows into view', 'wipes into view'],
   );
 });
+
+test('parseAssCues records the vertical band from style alignment, overrides, and \\pos', () => {
+  const ass = [
+    '[Script Info]',
+    'PlayResY: 720',
+    '',
+    '[V4+ Styles]',
+    'Format: Name, Fontname, Fontsize, PrimaryColour, Bold, Alignment, MarginV, Encoding',
+    'Style: Bottom,Arial,54,&H00FFFFFF,0,2,30,1',
+    'Style: TopSong,Arial,54,&H00FFFFFF,0,9,12,1',
+    '',
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:03.00,Bottom,,0,0,0,,\u4e0b\u306e\u30bb\u30ea\u30d5',
+    'Dialogue: 0,0:00:01.00,0:00:03.00,TopSong,,0,0,0,,\u6b4c\u8a5e\u306e\u884c',
+    'Dialogue: 0,0:00:01.00,0:00:03.00,Bottom,,0,0,0,,{\\an8}\u4e0a\u66f8\u304d\u306e\u884c',
+    'Dialogue: 0,0:00:01.00,0:00:03.00,Bottom,,0,0,0,,{\\pos(640,20)}\u770b\u677f\u306e\u884c',
+  ].join('\n');
+
+  const bands = parseAssCues(ass).map((cue) => cue.assLayout?.verticalBand);
+  assert.deepEqual(bands, ['bottom', 'top', 'top', 'top']);
+});

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

@@ -10,10 +10,13 @@ import {
 } from './ass-text';
 import { hasAssAnimationEvidence, mergeDuplicateCues } from './subtitle-cue-dedup';
 
+/** Vertical third of the screen a cue is authored to occupy. */
+export type AssVerticalBand = 'top' | 'middle' | 'bottom';
+
 export type AssCueLayout =
-  | { kind: 'positioned'; sourceOrder: number; y: number }
-  | { kind: 'fragment-grid'; sourceOrder: number }
-  | { kind: 'source-order'; sourceOrder: number };
+  | { kind: 'positioned'; sourceOrder: number; y: number; verticalBand?: AssVerticalBand }
+  | { kind: 'fragment-grid'; sourceOrder: number; verticalBand?: AssVerticalBand }
+  | { kind: 'source-order'; sourceOrder: number; verticalBand?: AssVerticalBand };
 
 export interface SubtitleCue {
   startTime: number;
@@ -481,9 +484,7 @@ function structuralOverrideSignature(cue: AnnotatedSubtitleCue): string {
   let signature = structuralSignatureCache.get(cue);
   if (signature === undefined) {
     const names = new Set(
-      cue.overrides.map(
-        (command) => `${command.animated ? '~' : ''}${command.name.toLowerCase()}`,
-      ),
+      cue.overrides.map((command) => `${command.animated ? '~' : ''}${command.name.toLowerCase()}`),
     );
     signature = [...names].sort().join(',');
     structuralSignatureCache.set(cue, signature);
@@ -544,9 +545,7 @@ function buildCoalescedCopy(members: readonly AnnotatedSubtitleCue[]): Annotated
  * while the anchor says one glyph. Merging each stack into a single presence spanning
  * the union window lets timing clusters see the authored line instead of its phases.
  */
-function coalesceAssAnchorCopies(
-  events: readonly AnnotatedSubtitleCue[],
-): AnnotatedSubtitleCue[] {
+function coalesceAssAnchorCopies(events: readonly AnnotatedSubtitleCue[]): AnnotatedSubtitleCue[] {
   const buckets = new Map<string, number[]>();
   const anchorPoints: (AssFragmentPosition[] | null)[] = events.map(() => null);
   events.forEach((event, index) => {
@@ -1362,8 +1361,7 @@ function isRepeatedGlyphText(cue: AnnotatedSubtitleCue): boolean {
 
 function isClippedRepeatedGlyphFragment(cue: AnnotatedSubtitleCue): boolean {
   return (
-    isRepeatedGlyphText(cue) &&
-    (hasStaticOverride(cue, 'clip') || hasStaticOverride(cue, 'iclip'))
+    isRepeatedGlyphText(cue) && (hasStaticOverride(cue, 'clip') || hasStaticOverride(cue, 'iclip'))
   );
 }
 
@@ -2055,6 +2053,111 @@ function recoverCanonicalAssEvents({
   );
 }
 
+function bandFromNumpadAlignment(alignment: number): AssVerticalBand | null {
+  if (alignment >= 7 && alignment <= 9) return 'top';
+  if (alignment >= 4 && alignment <= 6) return 'middle';
+  if (alignment >= 1 && alignment <= 3) return 'bottom';
+  return null;
+}
+
+// SSA v4 alignment reuses the legacy `\a` codes: 1-3 bottom, +4 top, +8 middle.
+function bandFromLegacyAlignment(alignment: number): AssVerticalBand | null {
+  if (alignment >= 9 && alignment <= 11) return 'middle';
+  if (alignment >= 5 && alignment <= 7) return 'top';
+  if (alignment >= 1 && alignment <= 3) return 'bottom';
+  return null;
+}
+
+interface AssPlacementContext {
+  playResY: number | null;
+  /** Lowercased style name -> vertical band from the style's Alignment column. */
+  styleBands: Map<string, AssVerticalBand>;
+}
+
+const EMPTY_PLACEMENT_CONTEXT: AssPlacementContext = { playResY: null, styleBands: new Map() };
+
+function parseAssPlacementContext(content: string): AssPlacementContext {
+  const styleBands = new Map<string, AssVerticalBand>();
+  let playResY: number | null = null;
+  let section: 'info' | 'v4plus' | 'v4' | null = null;
+  let alignmentIndex = -1;
+  let nameIndex = -1;
+
+  for (const line of content.split(/\r?\n/)) {
+    const trimmed = line.trim();
+    if (trimmed.startsWith('[') && trimmed.endsWith(']')) {
+      const sectionName = trimmed.toLowerCase();
+      section =
+        sectionName === '[script info]'
+          ? 'info'
+          : sectionName === '[v4+ styles]'
+            ? 'v4plus'
+            : sectionName === '[v4 styles]'
+              ? 'v4'
+              : null;
+      alignmentIndex = -1;
+      nameIndex = -1;
+      continue;
+    }
+    if (section === 'info') {
+      const resMatch = trimmed.match(/^playresy\s*:\s*(\d+(?:\.\d+)?)\s*$/i);
+      if (resMatch) playResY = Number(resMatch[1]);
+      continue;
+    }
+    if (section !== 'v4plus' && section !== 'v4') continue;
+    const separator = trimmed.indexOf(':');
+    if (separator < 0) continue;
+    const key = trimmed.slice(0, separator).trim().toLowerCase();
+    const fields = trimmed.slice(separator + 1).split(',');
+    if (key === 'format') {
+      const names = fields.map((field) => field.trim().toLowerCase());
+      alignmentIndex = names.indexOf('alignment');
+      nameIndex = names.indexOf('name');
+      continue;
+    }
+    if (key !== 'style' || alignmentIndex < 0 || nameIndex < 0) continue;
+    const styleName = fields[nameIndex]?.trim().toLowerCase();
+    const alignment = Number(fields[alignmentIndex]?.trim());
+    if (!styleName || !Number.isFinite(alignment)) continue;
+    const band =
+      section === 'v4plus'
+        ? bandFromNumpadAlignment(alignment)
+        : bandFromLegacyAlignment(alignment);
+    if (band) styleBands.set(styleName, band);
+  }
+
+  return { playResY, styleBands };
+}
+
+/**
+ * Where on screen mpv will draw this event: an explicit `\pos`/`\move` coordinate when
+ * the script declares its coordinate space, else an `\an`/`\a` override, else the
+ * style's Alignment. Constant for the life of the event, which is what lets simultaneous
+ * lines keep a stable stacking order in the overlay.
+ */
+function resolveVerticalBand(
+  overrides: readonly AssOverrideCommand[],
+  y: number | null,
+  style: string,
+  context: AssPlacementContext,
+): AssVerticalBand | undefined {
+  if (y !== null && context.playResY && context.playResY > 0) {
+    const ratio = y / context.playResY;
+    return ratio < 1 / 3 ? 'top' : ratio < 2 / 3 ? 'middle' : 'bottom';
+  }
+  for (const command of overrides) {
+    if (command.animated) continue;
+    const name = command.name.toLowerCase();
+    if (name !== 'an' && name !== 'a') continue;
+    const band =
+      name === 'an'
+        ? bandFromNumpadAlignment(Number(command.args))
+        : bandFromLegacyAlignment(Number(command.args));
+    if (band) return band;
+  }
+  return context.styleBands.get(style.trim().toLowerCase());
+}
+
 function parseAssCoordinate(value: string | undefined): number | null {
   if (!value?.trim()) return null;
   const coordinate = Number(value.trim());
@@ -2064,6 +2167,8 @@ function parseAssCoordinate(value: string | undefined): number | null {
 function buildAssCueLayout(
   overrides: readonly AssOverrideCommand[],
   sourceOrder: number,
+  style: string,
+  placement: AssPlacementContext,
 ): AssCueLayout {
   let y: number | null = null;
   for (const command of overrides) {
@@ -2081,14 +2186,18 @@ function buildAssCueLayout(
       y = (startY + endY) / 2;
     }
   }
-  return y === null
-    ? { kind: 'source-order', sourceOrder }
-    : { kind: 'positioned', sourceOrder, y };
+  const verticalBand = resolveVerticalBand(overrides, y, style, placement);
+  const base: AssCueLayout =
+    y === null ? { kind: 'source-order', sourceOrder } : { kind: 'positioned', sourceOrder, y };
+  return verticalBand ? { ...base, verticalBand } : base;
 }
 
 function parseAnnotatedAssEvents(content: string): ParsedAssEvents {
   const cues: AnnotatedSubtitleCue[] = [];
   const comments: AnnotatedSubtitleCue[] = [];
+  const placement = content.includes('[')
+    ? parseAssPlacementContext(content)
+    : EMPTY_PLACEMENT_CONTEXT;
   const lines = content.split(/\r?\n/);
   let inEventsSection = false;
   let eventOrder = 0;
@@ -2185,12 +2294,13 @@ function parseAnnotatedAssEvents(content: string): ParsedAssEvents {
     const effect = readField(fields, fieldIndex.effect);
     const layer = Number(readField(fields, fieldIndex.layer));
     const overrides = collectAssOverrideCommands(rawText);
+    const style = readField(fields, fieldIndex.style);
     const cue: AnnotatedSubtitleCue = {
       startTime,
       endTime,
       text,
       rawText,
-      style: readField(fields, fieldIndex.style),
+      style,
       layer: Number.isFinite(layer) ? layer : 0,
       name: readField(fields, fieldIndex.name),
       effect,
@@ -2198,7 +2308,7 @@ function parseAnnotatedAssEvents(content: string): ParsedAssEvents {
       overrides,
       overrideSignature: assOverrideSignature(overrides),
       order: eventOrder,
-      assLayout: buildAssCueLayout(overrides, eventOrder),
+      assLayout: buildAssCueLayout(overrides, eventOrder, style, placement),
     };
     eventOrder += 1;
     if (eventPrefix === ASS_COMMENT_PREFIX) {

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

@@ -134,7 +134,7 @@ export function createSubtitleProcessingController(
         try {
           const cachedTokenized = getCachedTokenization(text);
           if (cachedTokenized) {
-            output = cachedTokenized;
+            output = { ...cachedTokenized, text };
           } else {
             // Cache miss: show the plain line on time; the tokenized payload
             // upgrades it once ready. Skipped on refreshes of an already
@@ -266,7 +266,7 @@ export function createSubtitleProcessingController(
       lastEmittedText = text;
       lastEmittedGeneration = cacheGeneration;
       lastPlainEmittedText = null;
-      return cached;
+      return { ...cached, text };
     },
     hasCachedSubtitle: (text: string) => {
       const cacheKey = normalizeSubtitleCacheKey(text);

src/core/services/tokenizer.test.ts

@@ -84,6 +84,17 @@ function createDeferred<T>() {
   };
 }
 
+test('tokenizeSubtitle keeps the blank line separating simultaneous cues', async () => {
+  // The tokenized payload's text drives display; folding the cue boundary would merge
+  // two speakers back onto one line the moment tokenization upgrades the plain emit.
+  const result = await tokenizeSubtitle(
+    '\u4e00\u884c\u76ee\n\n\u4e8c\u884c\u76ee',
+    makeDeps({ getYomitanExt: () => null }),
+  );
+
+  assert.equal(result.text, '\u4e00\u884c\u76ee\n\n\u4e8c\u884c\u76ee');
+});
+
 test('tokenizeSubtitle splits same-line grammar endings before applying annotations', async () => {
   const result = await tokenizeSubtitle(
     '猫です',
@@ -1682,6 +1693,12 @@ test('tokenizeSubtitle normalizes newlines before Yomitan parse request', async
   assert.equal(result.tokens, null);
 });
 
+test('tokenizeSubtitle preserves CRLF boundaries between simultaneous cues', async () => {
+  const result = await tokenizeSubtitle('a\r\n\r\nb', makeDeps());
+
+  assert.deepEqual(result, { text: 'a\n\nb', tokens: null });
+});
+
 test('tokenizeSubtitle collapses zero-width separators before Yomitan parse request', async () => {
   let parseInput = '';
   const result = await tokenizeSubtitle(

src/core/services/tokenizer.ts

@@ -887,7 +887,15 @@ export async function tokenizeSubtitle(
   text: string,
   deps: TokenizerServiceDeps,
 ): Promise<SubtitleData> {
-  const displayText = normalizePlainSubtitleText(text);
+  // Normalize per cue group: the blank line separating simultaneous cues is display
+  // structure the payload text must keep, or the tokenized upgrade re-merges lines the
+  // provisional plain emit already showed apart.
+  const displayText = text
+    .replace(/\r\n/g, '\n')
+    .split(/\n{2,}/)
+    .map((part) => normalizePlainSubtitleText(part))
+    .filter(Boolean)
+    .join('\n\n');
 
   // 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.

src/main/runtime/primary-subtitle-text.test.ts

@@ -60,7 +60,7 @@ test('resolvePrimarySubtitleText combines unique simultaneous parsed cues', () =
         { startTime: 1, endTime: 3, text: '二行目' },
       ],
     }),
-    '一行目\n二行目',
+    '一行目\n\n二行目',
   );
 });
 
@@ -195,7 +195,7 @@ test('resolvePrimarySubtitleText combines parsed dialogue with a reconstructed l
     ],
   });
 
-  assert.equal(text, '普通のセリフ\n今　手にある');
+  assert.equal(text, '普通のセリフ\n\n今　手にある');
 });
 
 test('resolvePrimarySubtitleText uses fragment grids only to account for live sign pieces', () => {
@@ -300,7 +300,32 @@ test('resolvePrimarySubtitleText combines simultaneous canonical cues in source
     ],
   });
 
-  assert.equal(text, 'first\nsecond');
+  assert.equal(text, 'first\n\nsecond');
+});
+
+test('resolveCanonicalPrimarySubtitle orders active cues from top to bottom', () => {
+  const resolved = resolveCanonicalPrimarySubtitle({
+    liveText: 'bottom\ntop',
+    currentTimeSec: 2,
+    cues: [
+      {
+        startTime: 1,
+        endTime: 3,
+        text: 'bottom',
+        source: 'canonical-ass',
+        assLayout: { kind: 'source-order', sourceOrder: 1, verticalBand: 'bottom' },
+      },
+      {
+        startTime: 1,
+        endTime: 3,
+        text: 'top',
+        source: 'canonical-ass',
+        assLayout: { kind: 'source-order', sourceOrder: 0, verticalBand: 'top' },
+      },
+    ],
+  });
+
+  assert.equal(resolved?.text, 'top\n\nbottom');
 });
 
 test('resolvePrimarySubtitleText collapses whitespace variants of a canonical lyric', () => {
@@ -498,6 +523,23 @@ test('stripCanonicalFragmentLines drops a live glyph wall with no nearby canonic
   );
 });
 
+test('resolvePrimarySubtitleText keeps a line joining an active cue despite stale time-pos', () => {
+  // Issue #220: mpv publishes the combined sub-text the moment a joining line's first
+  // frame renders, while the observed time-pos still sits just before that line's
+  // start. The joining cue must not be filtered out as inactive.
+  assert.equal(
+    resolvePrimarySubtitleText({
+      liveText: 'Балда! Балда, балда, балда!\nСестренка не может остановиться',
+      currentTimeSec: 767.78,
+      cues: [
+        { startTime: 767.19, endTime: 772.78, text: 'Балда! Балда, балда, балда!' },
+        { startTime: 767.79, endTime: 771.15, text: 'Сестренка не может остановиться' },
+      ],
+    }),
+    'Балда! Балда, балда, балда!\n\nСестренка не может остановиться',
+  );
+});
+
 test('resolvePrimarySubtitleText drops a finished lyric whose exit ghosts outlive it beside a raw line', () => {
   // The reconstructed lyric ended at 6.0 but its exit ghost glyphs stay in the live
   // text until 7.0, while the next authored line is a plain raw event. The retired cue
@@ -524,3 +566,85 @@ test('resolvePrimarySubtitleText drops a finished lyric whose exit ghosts outliv
     '象徴的なパレード',
   );
 });
+
+test('resolvePrimarySubtitleText stacks simultaneous cues by screen position, not start order', () => {
+  // A top-anchored lyric and bottom dialogue: mpv draws the lyric above the dialogue for
+  // the whole overlap. Whichever event started first must not decide the row, or the
+  // pair swaps every time one side is replaced mid-overlap.
+  const lyricLayout = { kind: 'source-order', sourceOrder: 0, verticalBand: 'top' } as const;
+  const dialogueLayout = { kind: 'source-order', sourceOrder: 1, verticalBand: 'bottom' } as const;
+  const dialogue = {
+    startTime: 632.2,
+    endTime: 634.8,
+    text: '\u30e9\u30a4\u30d6\u3000\u3084\u3081\u3088\u3063\u304b',
+    assLayout: dialogueLayout,
+  };
+
+  // Lyric started before the dialogue...
+  assert.equal(
+    resolvePrimarySubtitleText({
+      liveText: '\u30e9\u30a4\u30d6\u3000\u3084\u3081\u3088\u3063\u304b\n\u6b4c\u8a5e\uff21',
+      currentTimeSec: 632.5,
+      cues: [
+        { startTime: 629.5, endTime: 633.5, text: '\u6b4c\u8a5e\uff21', assLayout: lyricLayout },
+        dialogue,
+      ],
+    }),
+    '\u6b4c\u8a5e\uff21\n\n\u30e9\u30a4\u30d6\u3000\u3084\u3081\u3088\u3063\u304b',
+  );
+  // ...and the next lyric starts after it: the rows must not swap.
+  assert.equal(
+    resolvePrimarySubtitleText({
+      liveText: '\u30e9\u30a4\u30d6\u3000\u3084\u3081\u3088\u3063\u304b\n\u6b4c\u8a5e\uff22',
+      currentTimeSec: 633.8,
+      cues: [
+        dialogue,
+        { startTime: 633.5, endTime: 637.0, text: '\u6b4c\u8a5e\uff22', assLayout: lyricLayout },
+      ],
+    }),
+    '\u6b4c\u8a5e\uff22\n\n\u30e9\u30a4\u30d6\u3000\u3084\u3081\u3088\u3063\u304b',
+  );
+});
+
+test('resolvePrimarySubtitleText puts an unreadable placement above bottom dialogue', () => {
+  // Dialogue is the case that reliably declares a bottom alignment, so a cue whose
+  // placement could not be read is more often a sign or song line. Keeping dialogue on
+  // the bottom row means the line worth reading stays where the eye already is.
+  assert.equal(
+    resolvePrimarySubtitleText({
+      liveText: '\u4e0b\u306e\u30bb\u30ea\u30d5\n\u4e0d\u660e\u306a\u884c',
+      currentTimeSec: 2,
+      cues: [
+        {
+          startTime: 1,
+          endTime: 3,
+          text: '\u4e0b\u306e\u30bb\u30ea\u30d5',
+          assLayout: { kind: 'source-order', sourceOrder: 0, verticalBand: 'bottom' },
+        },
+        {
+          startTime: 1.5,
+          endTime: 3,
+          text: '\u4e0d\u660e\u306a\u884c',
+          assLayout: { kind: 'source-order', sourceOrder: 1 },
+        },
+      ],
+    }),
+    '\u4e0d\u660e\u306a\u884c\n\n\u4e0b\u306e\u30bb\u30ea\u30d5',
+  );
+});
+
+test('resolvePrimarySubtitleText keeps source order when no cue declares a placement', () => {
+  // SRT and websocket cues carry no layout at all: every cue ties, so the stable sort
+  // must leave them exactly as the cue list had them.
+  assert.equal(
+    resolvePrimarySubtitleText({
+      liveText: 'First line\nSecond line',
+      currentTimeSec: 2,
+      cues: [
+        { startTime: 1, endTime: 3, text: 'First line' },
+        { startTime: 1.5, endTime: 3, text: 'Second line' },
+      ],
+    }),
+    'First line\n\nSecond line',
+  );
+});

src/main/runtime/primary-subtitle-text.ts

@@ -1,4 +1,4 @@
-import type { SubtitleCue } from '../../types';
+import type { AssVerticalBand, SubtitleCue } from '../../types';
 import {
   removeAssControlDebrisLines,
   removeLiveGlyphFragmentLines,
@@ -57,21 +57,52 @@ function compactWhitespace(text: string): string {
   return text.normalize('NFKC').replace(/\s+/gu, '');
 }
 
+/**
+ * Distinct simultaneous cues are separated by a blank line so the display layer can tell
+ * a wrap inside one utterance from the boundary between two of them. Consumers that read
+ * the text rather than display it fold these back to single breaks.
+ */
+const CUE_BOUNDARY = '\n\n';
+
+const VERTICAL_BAND_RANK: Record<AssVerticalBand, number> = { top: 0, middle: 1, bottom: 2 };
+
+/**
+ * Stack simultaneous cues the way they sit on screen: mpv keeps a top-anchored lyric or
+ * sign above bottom dialogue for its whole run, while cue-list order follows start time
+ * and would swap the pair whenever one side is replaced mid-overlap. The band is
+ * constant per event, so a line never changes rows while it is displayed.
+ *
+ * A cue whose placement could not be read -- an unknown style, a script with no styles
+ * section -- sorts to the top. Dialogue is the case that reliably declares a bottom
+ * alignment, so what is left unresolved is more often a sign or a song line, and keeping
+ * the dialogue on the bottom row means the line worth reading stays where the eye
+ * already is. Sort is stable, so cues sharing a rank keep their existing order.
+ */
+function orderCuesForDisplay(cues: readonly SubtitleCue[]): SubtitleCue[] {
+  const rank = (cue: SubtitleCue): number =>
+    VERTICAL_BAND_RANK[cue.assLayout?.verticalBand ?? 'top'];
+  return [...cues].sort((a, b) => rank(a) - rank(b));
+}
+
 // ASS layers can encode the same visible spacing with ordinary, hard, or
 // ideographic spaces. Matching and emission must use the same identity or each
 // layer reappears as a copy.
-function uniqueCueTexts(cues: readonly SubtitleCue[]): string[] {
-  const texts: string[] = [];
+function uniqueCueTextGroups(cues: readonly SubtitleCue[]): string[] {
+  const groups: string[] = [];
   const seen = new Set<string>();
   for (const cue of cues) {
+    const lines: string[] = [];
     for (const line of cue.text.split('\n')) {
       const compactText = compactWhitespace(line);
       if (!compactText || seen.has(compactText)) continue;
       seen.add(compactText);
-      texts.push(line);
+      lines.push(line);
+    }
+    if (lines.length > 0) {
+      groups.push(lines.join('\n'));
     }
   }
-  return texts;
+  return groups;
 }
 
 function compactLineSegments(text: string): string[] {
@@ -134,23 +165,24 @@ function resolveActiveParsedPrimarySubtitle(options: {
     return null;
   }
 
-  // A cue selected only through the edge tolerance has already ended (or not yet
-  // started) by its published timing: a finished lyric whose exit ghosts linger into
-  // the next line. It still explains those live fragments above, but while any cue is
-  // strictly active, only the active cues supply the displayed text. With no strictly
-  // active cue, the edge cues remain the display fallback for stale time-pos readings.
-  const strictlyActive = selected.filter(
-    (cue) => cue.startTime <= options.currentTimeSec && cue.endTime > options.currentTimeSec,
-  );
-  const displayCues = strictlyActive.length > 0 ? strictlyActive : selected;
+  // A cue selected only through the edge tolerance on its end has already finished by
+  // its published timing: a lyric whose exit ghosts linger into the next line. It still
+  // explains those live fragments above, but must not re-surface beside cues that are
+  // still running. The start side keeps the tolerance: mpv publishes the combined
+  // sub-text the moment a joining line's first frame renders, while the observed
+  // time-pos still sits just before that line's start, and the selection above already
+  // required the cue's text to be on screen (#220). With every selected cue finished,
+  // the edge cues remain the display fallback for stale time-pos readings.
+  const unfinished = selected.filter((cue) => cue.endTime > options.currentTimeSec);
+  const displayCues = unfinished.length > 0 ? unfinished : selected;
 
   // Dense sign grids still explain their raw mpv fragments, but are visual
   // typesetting rather than a publishable subtitle line.
-  const texts = uniqueCueTexts(
-    displayCues.filter((cue) => cue.assLayout?.kind !== 'fragment-grid'),
+  const groups = uniqueCueTextGroups(
+    orderCuesForDisplay(displayCues.filter((cue) => cue.assLayout?.kind !== 'fragment-grid')),
   );
   return {
-    text: texts.join('\n'),
+    text: groups.join(CUE_BOUNDARY),
     startTime: Math.min(...displayCues.map((cue) => cue.startTime)),
     endTime: Math.max(...displayCues.map((cue) => cue.endTime)),
     cues: displayCues,
@@ -217,9 +249,9 @@ export function resolveCanonicalPrimarySubtitle(options: {
     return null;
   }
 
-  const texts = uniqueCueTexts(selected);
+  const groups = uniqueCueTextGroups(orderCuesForDisplay(selected));
   return {
-    text: texts.join('\n'),
+    text: groups.join(CUE_BOUNDARY),
     startTime: Math.min(...selected.map((cue) => cue.startTime)),
     endTime: Math.max(...selected.map((cue) => cue.endTime)),
     cues: selected,

src/renderer/subtitle-render.test.ts

@@ -11,6 +11,7 @@ import {
   getFrequencyRankLabelForToken,
   getJlptLevelLabelForToken,
   normalizeSubtitle,
+  normalizeSubtitleForDisplay,
   prepareSecondarySubtitleLines,
   sanitizeSubtitleHoverTokenColor,
   shouldRenderTokenizedSubtitle,
@@ -1004,6 +1005,34 @@ test('normalizeSubtitle collapses explicit line breaks when collapseLineBreaks i
   );
 });
 
+test('normalizeSubtitleForDisplay always breaks between simultaneous cues', () => {
+  // The blank line marks two distinct cues on screen at once. Flattening it would run a
+  // sign or a second speaker into the line beside it as one sentence.
+  const twoCues =
+    '\u6b21\u306f\u9b3c\u5b50\u6bcd\u795e\u524d\u3000\u9b3c\u5b50\u6bcd\u795e\u524d\n\n\u611b\u97f3\u3061\u3083\u3093\u3000\u3082\u3046\u5199\u771f\u4e0a\u3052\u3066\u308b';
+
+  assert.equal(
+    normalizeSubtitleForDisplay(twoCues, false),
+    '\u6b21\u306f\u9b3c\u5b50\u6bcd\u795e\u524d \u9b3c\u5b50\u6bcd\u795e\u524d\n\u611b\u97f3\u3061\u3083\u3093 \u3082\u3046\u5199\u771f\u4e0a\u3052\u3066\u308b',
+  );
+  assert.equal(normalizeSubtitleForDisplay(twoCues, true), twoCues.replace('\n\n', '\n'));
+});
+
+test('normalizeSubtitleForDisplay preserves CRLF boundaries between simultaneous cues', () => {
+  assert.equal(normalizeSubtitleForDisplay('a\r\n\r\nb', false), 'a\nb');
+});
+
+test('normalizeSubtitleForDisplay still flattens a wrap inside one cue', () => {
+  // A typesetter's \\N inside a single utterance is what preserveLineBreaks governs.
+  assert.equal(
+    normalizeSubtitleForDisplay(
+      '\u5e38\u4eba\u304c\u4f7f\u3048\u3070\\N\u305d\u306e\u5727\u5012\u7684\u306a\u529b\u306b',
+      false,
+    ),
+    '\u5e38\u4eba\u304c\u4f7f\u3048\u3070 \u305d\u306e\u5727\u5012\u7684\u306a\u529b\u306b',
+  );
+});
+
 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.

src/renderer/subtitle-render.ts

@@ -50,6 +50,21 @@ export function normalizeSubtitle(text: string, trim = true, collapseLineBreaks
   return normalizePlainSubtitleText(text, { trim, collapseLineBreaks });
 }
 
+/**
+ * Display form of a resolved subtitle. `preserveLineBreaks` governs wrapping inside one
+ * utterance, which is what a typesetter's `\N` means. The blank line the resolver puts
+ * between two simultaneous cues is a different thing and always breaks, so a sign or a
+ * second speaker never runs into the line beside it.
+ */
+export function normalizeSubtitleForDisplay(text: string, preserveLineBreaks: boolean): string {
+  return text
+    .replace(/\r\n/g, '\n')
+    .split(/\n{2,}/)
+    .map((cueText) => normalizeSubtitle(cueText, true, !preserveLineBreaks))
+    .filter((cueText) => cueText.length > 0)
+    .join('\n');
+}
+
 const HEX_COLOR_PATTERN = /^#(?:[0-9a-fA-F]{3}|[0-9a-fA-F]{4}|[0-9a-fA-F]{6}|[0-9a-fA-F]{8})$/;
 const SAFE_CSS_COLOR_PATTERN =
   /^(?:#(?:[0-9a-fA-F]{3}|[0-9a-fA-F]{4}|[0-9a-fA-F]{6}|[0-9a-fA-F]{8})|(?:rgba?|hsla?)\([^)]*\)|var\([^)]*\)|[a-zA-Z]+)$/;
@@ -414,16 +429,13 @@ function renderWithTokens(
   const fragment = document.createDocumentFragment();
 
   if (sourceText) {
-    const normalizedSource = normalizeSubtitle(sourceText, true, !preserveLineBreaks);
+    const normalizedSource = normalizeSubtitleForDisplay(sourceText, preserveLineBreaks);
     const segments = alignTokensToSourceText(tokens, normalizedSource);
 
     for (const segment of segments) {
       if (segment.kind === 'text') {
-        if (preserveLineBreaks) {
-          renderPlainTextPreserveLineBreaks(fragment, segment.text);
-        } else {
-          fragment.appendChild(document.createTextNode(segment.text));
-        }
+        // Normalization already resolved which breaks survive; every one left is real.
+        renderPlainTextPreserveLineBreaks(fragment, segment.text);
         continue;
       }
 
@@ -748,7 +760,7 @@ export function createSubtitleRenderer(ctx: RendererContext) {
       return;
     }
 
-    const normalized = normalizeSubtitle(text, true, !ctx.state.preserveSubtitleLineBreaks);
+    const normalized = normalizeSubtitleForDisplay(text, ctx.state.preserveSubtitleLineBreaks);
     const hasRenderableTokens =
       shouldRenderTokenizedSubtitle(tokens?.length ?? 0) && Boolean(tokens);
     if (

src/types/subtitle.ts

@@ -1,4 +1,4 @@
-import type { SubtitleCue } from '../core/services/subtitle-cue-parser';
+import type { AssVerticalBand, SubtitleCue } from '../core/services/subtitle-cue-parser';
 
 export enum PartOfSpeech {
   noun = 'noun',
@@ -187,7 +187,7 @@ export interface ResolvedTokenPos2ExclusionConfig {
 
 export type FrequencyDictionaryMode = 'single' | 'banded';
 
-export type { SubtitleCue };
+export type { AssVerticalBand, SubtitleCue };
 
 export type SubtitleSidebarLayout = 'overlay' | 'embedded';