@torrent-tv/proxy 2.75.0 → 2.76.1

package/test/subtitle-language.test.js

@@ -1,252 +1,252 @@
-/**
- * @file What the language detector is FED.
- *
- * Field 2026-09-01: a Russian subtitle file was offered to the viewer as
- * `English (Stan WarHammer & Nesitach)`. The detector was not at fault — it was
- * handed the raw `.ass` file, which is half Latin markup, while the markup-free
- * WebVTT it was about to serve sat in the variable beside it. Measured on that
- * file: 5040 Latin characters against 5983 Cyrillic, `franc(the file) = eng`,
- * `franc(the dialogue) = rus`.
- * `research/subtitle-language-ass-markup-2026-09-01.md`.
- *
- * These checks pin the input at each of the three places a language is read.
- *
- * **On the size of the fixtures.** franc's answer among the Cyrillic languages
- * is not stable on a small sample: measured 2026-09-01 by growing one Russian
- * text line by line, it answered `bul` at 129 characters, `srp` at 158, 241 and
- * `bul` again at 292, then `rus` at every length from 337 to 881. So the
- * dialogue here is ~880 Cyrillic characters — past that boundary by a factor of
- * about 2.6, and still an order of magnitude below a real episode's 5983. The
- * fixtures are NOT sized to make these checks pass; they are sized to resemble
- * the thing. That instability is a separate defect and is recorded as one.
- */
-
-import test from "node:test";
-import assert from "node:assert/strict";
-
-import { detectLanguage } from "../services/language-detect.js";
-import { SubtitleFileContainer } from "../services/container/SubtitleFileContainer.js";
-import { TextSubtitleTrack } from "../services/tracks/TextSubtitleTrack.js";
-import { MatroskaContainer } from "../services/container/MatroskaContainer.js";
-
-/** Varied Russian dialogue, the length a few minutes of an episode carries. */
-const RUSSIAN_DIALOGUE = [
-  "Ты видишь их?",
-  "Да.",
-  "Но не столько вижу, сколько чувствую.",
-  "Скорее всего, они будут здесь с минуты на минуту.",
-  "Почему господа октябристы решили напасть именно сейчас?",
-  "Им никогда не преодолеть эти стены.",
-  "Мы держали эту крепость три года и продержим ещё столько же.",
-  "Готовьте лучников на восточной стороне.",
-  "Если они подойдут ближе, мы откроем огонь без предупреждения.",
-  "Я не собираюсь умирать здесь, в этой богом забытой дыре.",
-  "Тогда возьми меч и вставай рядом со мной.",
-  "Сколько у нас осталось воды и хлеба?",
-  "На неделю, если не считать раненых.",
-  "Раненых считать придётся, они тоже люди.",
-  "Отправь гонца к южным воротам и передай приказ отступать.",
-  "Он не успеет, дорога перерезана ещё вчера вечером.",
-  "Значит пойду сам, а ты останешься за меня командовать.",
-  "Это безумие, и ты прекрасно об этом знаешь.",
-  "Безумие — это сидеть и ждать, пока нас перебьют по одному.",
-  "Хорошо. Но возьми с собой хотя бы двоих.",
-  "Двоих я возьму. Больше не могу себе позволить.",
-  "Береги себя. Мы будем держать стену до последнего.",
-  "Я знаю. Именно поэтому я и ухожу спокойно.",
-  "Смотри, дым над лесом. Они уже жгут деревни.",
-  "Тогда времени у нас меньше, чем мы думали."
-];
-
-/** `hh:mm:ss.cc` for the ASS `Start`/`End` fields. */
-function assTime(seconds) {
-  const mm = String(Math.floor(seconds / 60)).padStart(2, "0");
-  const ss = String(seconds % 60).padStart(2, "0");
-  return `0:${mm}:${ss}.00`;
-}
-
-/**
- * An ASS file in the shape a fansub release ships: an Aegisub header, a styles
- * section with English font and style names, and Russian dialogue carrying
- * override groups. Modelled on the field file.
- */
-const RUSSIAN_ASS = `[Script Info]
-Title: Default Aegisub file
-ScriptType: v4.00+
-PlayResX: 1280
-PlayResY: 720
-Original Translation: Nesitach
-Original Editing: Stan WarHammer
-WrapStyle: 0
-ScaledBorderAndShadow: no
-Video Aspect Ratio: c1.77778
-YCbCr Matrix: TV.601
-Aegisub Video Aspect Ratio: c1.777778
-
-[Aegisub Project Garbage]
-Last Style Storage: Default
-Audio File: [HorribleSubs] Drifters - 03 [720p].mkv
-Video File: [HorribleSubs] Drifters - 03 [720p].mkv
-Video AR Mode: 4
-Video AR Value: 1.777778
-Video Zoom Percent: 0.600000
-Scroll Position: 256
-Active Line: 257
-Video Position: 33159
-
-[V4+ Styles]
-Format: Name, Fontname, Fontsize, PrimaryColour, SecondaryColour, OutlineColour, BackColour, Bold, Italic, Underline, StrikeOut, ScaleX, ScaleY, Spacing, Angle, BorderStyle, Outline, Shadow, Alignment, MarginL, MarginR, MarginV, Encoding
-Style: Default,Trebuchet MS,54,&H00FFFFFF,&H000000FF,&H00000000,&H00000000,0,0,0,0,100,100,0,0,1,2.5,1.5,2,20,20,25,1
-Style: Signs,Times New Roman,48,&H00FFFF00,&H000000FF,&H00202020,&H00000000,-1,0,0,0,100,100,0,0,1,2,1,8,20,20,20,1
-Style: Italics,Trebuchet MS Italic,54,&H00FFFFFF,&H000000FF,&H00000000,&H00000000,0,-1,0,0,100,100,0,0,1,2.5,1.5,2,20,20,25,1
-
-[Events]
-Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text
-${RUSSIAN_DIALOGUE.map((line, index) => {
-    const style = index % 7 === 6 ? "Signs" : "Default";
-    const tag = index % 5 === 4 ? "{\\pos(640,620)}" : "";
-    return `Dialogue: 0,${assTime(index * 4)},${assTime(index * 4 + 3)},${style},,0,0,0,,${tag}${line}`;
-  }).join("\n")}
-`;
-
-test("a Russian .ass is reported as Russian, and its markup does not reach the detector", () => {
-  const vtt = SubtitleFileContainer.toVtt(RUSSIAN_ASS, ".ass");
-  assert.ok(vtt.startsWith("WEBVTT"), "the conversion produced a WebVTT document");
-
-  const spoken = TextSubtitleTrack.cueTextOfVtt(vtt);
-  // Every one of these is a piece of the file the viewer never reads, and each
-  // was competing with the dialogue for the detector's answer.
-  for (const markup of [
-    "Aegisub", "Script Info", "V4+ Styles", "Format:", "Dialogue:",
-    "Trebuchet MS", "Times New Roman", "Default", "Signs", "pos(",
-    "-->", "0:00:12"
-  ]) {
-    assert.ok(!spoken.includes(markup), `cue text still carries "${markup}"`);
-  }
-  assert.ok(spoken.includes("Ты видишь их?"), "cue text keeps the dialogue");
-
-  assert.deepEqual(TextSubtitleTrack.detectLanguageFromVtt(vtt), { code: "ru", name: "Russian" });
-});
-
-test("the fixture is the hard case: the file itself is nearly half Latin", () => {
-  // Not an assertion about franc — an assertion that this fixture reproduces the
-  // field file's proportions. A fixture whose markup were negligible would pass
-  // the check above with or without the fix.
-  const cyrillic = (RUSSIAN_ASS.match(/[Ѐ-ӿ]/gu) ?? []).length;
-  const latin = (RUSSIAN_ASS.match(/[A-Za-z]/g) ?? []).length;
-  assert.ok(cyrillic > 800, `too little dialogue: ${cyrillic} Cyrillic characters`);
-  assert.ok(latin > cyrillic * 0.5, `markup too small to reproduce the fault: ${latin} vs ${cyrillic}`);
-});
-
-test("cue identifiers, NOTE, STYLE and REGION blocks are not text", () => {
-  const vtt = [
-    "WEBVTT - This file has cues.",
-    "Kind: captions",
-    "Language: en",
-    "",
-    "NOTE",
-    "Translated by an English speaking volunteer, all rights reserved.",
-    "",
-    "STYLE",
-    "::cue { background-image: linear-gradient(to bottom, dimgray, lightgray); }",
-    "",
-    "REGION",
-    "id:speaker width:40% lines:3 regionanchor:0%,100%",
-    "",
-    "opening-line",
-    "00:00:12.060 --> 00:00:13.270 align:start position:0%",
-    "<v Тоёхиса>Ты видишь их?</v>",
-    "",
-    "2",
-    "00:00:15.400 --> 00:00:16.650",
-    "<i>Но не столько вижу,</i>",
-    "сколько чувствую &amp; ощущаю.",
-    ""
-  ].join("\n");
-
-  const spoken = TextSubtitleTrack.cueTextOfVtt(vtt);
-  assert.equal(
-    spoken,
-    "Ты видишь их?\nНо не столько вижу,\nсколько чувствую   ощущаю."
-  );
-});
-
-test("a document with no cues yields no language rather than a guess", () => {
-  assert.equal(TextSubtitleTrack.cueTextOfVtt("WEBVTT\n"), "");
-  assert.equal(TextSubtitleTrack.detectLanguageFromVtt("WEBVTT\n"), null);
-  assert.equal(TextSubtitleTrack.detectLanguageFromVtt(null), null);
-});
-
-test("an embedded ASS cue is unframed by the container, then read", () => {
-  // What a Matroska block actually holds, per
-  // `matroska.org/technical/subtitles.html`: "Events are stored in the Block in
-  // this order: ReadOrder, Layer, Style, Name, MarginL, MarginR, MarginV,
-  // Effect, Text", with Start and End taken out into the block's own timing.
-  // Eight fields, and the earlier fixture here wrongly carried ten — it kept
-  // the two timestamps a FILE has, which is why it passed while the field
-  // subtitles of a real `.mkv` were shown to the viewer whole.
-  const blocks = RUSSIAN_DIALOGUE.map((line, index) => Buffer.from(
-    `${index + 1},0,Default,,0000,0000,0000,,` +
-    `${index % 5 === 4 ? "{\\pos(640,620)}" : ""}${line}`,
-    "utf8"
-  ));
-
-  const cues = blocks.map((payload, index) => ({
-    startSeconds: index * 4,
-    endSeconds: index * 4 + 3,
-    text: MatroskaContainer.cueTextOf(payload, "S_TEXT/ASS")
-  }));
-
-  const spoken = TextSubtitleTrack.finalizeCues(cues, "S_TEXT/ASS").map((cue) => cue.text).join("\n");
-  assert.ok(!spoken.includes("Default"), "the style field survived into the text");
-  assert.ok(!spoken.includes("0000"), "a margin field survived into the text");
-  assert.ok(!spoken.includes("pos("), "an override group survived into the text");
-  assert.deepEqual(detectLanguage(spoken), { code: "ru", name: "Russian" });
-});
-
-test("too little text is answered with nothing, not with a guess", () => {
-  // Measured 2026-09-02 (`research/franc-boundary-2026-09-02.md`): franc's
-  // answer for Russian walks between Bulgarian, Serbian and Russian until about
-  // 650 characters. Below that the honest answer is none — the browser then
-  // shows "Unknown", and the label moves when enough of the file has arrived.
-  const short = RUSSIAN_DIALOGUE.slice(0, 6).join("\n");
-  assert.ok(short.length < 300, `the fixture must be short: ${short.length}`);
-  assert.equal(detectLanguage(short), null);
-
-  const long = RUSSIAN_DIALOGUE.join("\n");
-  assert.ok(long.length > 650, `the fixture must be long enough: ${long.length}`);
-  assert.deepEqual(detectLanguage(long), { code: "ru", name: "Russian" });
-});
-
-test("an answer that does not survive losing half the text is not an answer", () => {
-  // A file carrying two languages — a bilingual release, or a track that turns
-  // into signs-only English partway. franc answers something for the whole; the
-  // halves disagree with it, and that disagreement is the text saying the
-  // answer rests on where it happened to be cut.
-  const english = [
-    "Do you see them out there beyond the wall?",
-    "I do, though not as clearly as I would like to.",
-    "They have been gathering since the morning came.",
-    "We should send word to the southern gate at once.",
-    "The road was cut yesterday and no rider will pass.",
-    "Then we hold what we have and wait for the dawn."
-  ].join("\n");
-  const mixed = `${RUSSIAN_DIALOGUE.slice(0, 13).join("\n")}\n${english}`;
-  assert.equal(detectLanguage(mixed), null);
-});
-
-test("a WebVTT body is decoded whole, so no character is cut in half", () => {
-  // The ffmpeg extraction path used to detect on `String(body.subarray(0, 4096))`.
-  // Two faults: a byte cut lands mid-character on any non-Latin track, and most
-  // of those bytes are timestamps rather than words. Cyrillic is two bytes per
-  // character in UTF-8, so 4096 bytes of this document is a few hundred
-  // characters of dialogue — inside the unstable region measured above.
-  const cues = RUSSIAN_DIALOGUE.map((line, index) => {
-    const start = `00:00:${String(index * 2).padStart(2, "0")}.000`;
-    const end = `00:00:${String(index * 2 + 1).padStart(2, "0")}.000`;
-    return `${start} --> ${end}\n${line}`;
-  });
-  const body = Buffer.from(`WEBVTT\n\n${cues.join("\n\n")}\n`, "utf8");
-
-  assert.deepEqual(TextSubtitleTrack.detectLanguageFromVtt(body.toString("utf8")), { code: "ru", name: "Russian" });
-});
+/**
+ * @file What the language detector is FED.
+ *
+ * Field 2026-09-01: a Russian subtitle file was offered to the viewer as
+ * `English (Stan WarHammer & Nesitach)`. The detector was not at fault — it was
+ * handed the raw `.ass` file, which is half Latin markup, while the markup-free
+ * WebVTT it was about to serve sat in the variable beside it. Measured on that
+ * file: 5040 Latin characters against 5983 Cyrillic, `franc(the file) = eng`,
+ * `franc(the dialogue) = rus`.
+ * `research/subtitle-language-ass-markup-2026-09-01.md`.
+ *
+ * These checks pin the input at each of the three places a language is read.
+ *
+ * **On the size of the fixtures.** franc's answer among the Cyrillic languages
+ * is not stable on a small sample: measured 2026-09-01 by growing one Russian
+ * text line by line, it answered `bul` at 129 characters, `srp` at 158, 241 and
+ * `bul` again at 292, then `rus` at every length from 337 to 881. So the
+ * dialogue here is ~880 Cyrillic characters — past that boundary by a factor of
+ * about 2.6, and still an order of magnitude below a real episode's 5983. The
+ * fixtures are NOT sized to make these checks pass; they are sized to resemble
+ * the thing. That instability is a separate defect and is recorded as one.
+ */
+
+import test from "node:test";
+import assert from "node:assert/strict";
+
+import { detectLanguage } from "../services/tracks/language-detect.js";
+import { SubtitleFileContainer } from "../services/container/SubtitleFileContainer.js";
+import { TextSubtitleTrack } from "../services/tracks/TextSubtitleTrack.js";
+import { MatroskaContainer } from "../services/container/MatroskaContainer.js";
+
+/** Varied Russian dialogue, the length a few minutes of an episode carries. */
+const RUSSIAN_DIALOGUE = [
+  "Ты видишь их?",
+  "Да.",
+  "Но не столько вижу, сколько чувствую.",
+  "Скорее всего, они будут здесь с минуты на минуту.",
+  "Почему господа октябристы решили напасть именно сейчас?",
+  "Им никогда не преодолеть эти стены.",
+  "Мы держали эту крепость три года и продержим ещё столько же.",
+  "Готовьте лучников на восточной стороне.",
+  "Если они подойдут ближе, мы откроем огонь без предупреждения.",
+  "Я не собираюсь умирать здесь, в этой богом забытой дыре.",
+  "Тогда возьми меч и вставай рядом со мной.",
+  "Сколько у нас осталось воды и хлеба?",
+  "На неделю, если не считать раненых.",
+  "Раненых считать придётся, они тоже люди.",
+  "Отправь гонца к южным воротам и передай приказ отступать.",
+  "Он не успеет, дорога перерезана ещё вчера вечером.",
+  "Значит пойду сам, а ты останешься за меня командовать.",
+  "Это безумие, и ты прекрасно об этом знаешь.",
+  "Безумие — это сидеть и ждать, пока нас перебьют по одному.",
+  "Хорошо. Но возьми с собой хотя бы двоих.",
+  "Двоих я возьму. Больше не могу себе позволить.",
+  "Береги себя. Мы будем держать стену до последнего.",
+  "Я знаю. Именно поэтому я и ухожу спокойно.",
+  "Смотри, дым над лесом. Они уже жгут деревни.",
+  "Тогда времени у нас меньше, чем мы думали."
+];
+
+/** `hh:mm:ss.cc` for the ASS `Start`/`End` fields. */
+function assTime(seconds) {
+  const mm = String(Math.floor(seconds / 60)).padStart(2, "0");
+  const ss = String(seconds % 60).padStart(2, "0");
+  return `0:${mm}:${ss}.00`;
+}
+
+/**
+ * An ASS file in the shape a fansub release ships: an Aegisub header, a styles
+ * section with English font and style names, and Russian dialogue carrying
+ * override groups. Modelled on the field file.
+ */
+const RUSSIAN_ASS = `[Script Info]
+Title: Default Aegisub file
+ScriptType: v4.00+
+PlayResX: 1280
+PlayResY: 720
+Original Translation: Nesitach
+Original Editing: Stan WarHammer
+WrapStyle: 0
+ScaledBorderAndShadow: no
+Video Aspect Ratio: c1.77778
+YCbCr Matrix: TV.601
+Aegisub Video Aspect Ratio: c1.777778
+
+[Aegisub Project Garbage]
+Last Style Storage: Default
+Audio File: [HorribleSubs] Drifters - 03 [720p].mkv
+Video File: [HorribleSubs] Drifters - 03 [720p].mkv
+Video AR Mode: 4
+Video AR Value: 1.777778
+Video Zoom Percent: 0.600000
+Scroll Position: 256
+Active Line: 257
+Video Position: 33159
+
+[V4+ Styles]
+Format: Name, Fontname, Fontsize, PrimaryColour, SecondaryColour, OutlineColour, BackColour, Bold, Italic, Underline, StrikeOut, ScaleX, ScaleY, Spacing, Angle, BorderStyle, Outline, Shadow, Alignment, MarginL, MarginR, MarginV, Encoding
+Style: Default,Trebuchet MS,54,&H00FFFFFF,&H000000FF,&H00000000,&H00000000,0,0,0,0,100,100,0,0,1,2.5,1.5,2,20,20,25,1
+Style: Signs,Times New Roman,48,&H00FFFF00,&H000000FF,&H00202020,&H00000000,-1,0,0,0,100,100,0,0,1,2,1,8,20,20,20,1
+Style: Italics,Trebuchet MS Italic,54,&H00FFFFFF,&H000000FF,&H00000000,&H00000000,0,-1,0,0,100,100,0,0,1,2.5,1.5,2,20,20,25,1
+
+[Events]
+Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text
+${RUSSIAN_DIALOGUE.map((line, index) => {
+    const style = index % 7 === 6 ? "Signs" : "Default";
+    const tag = index % 5 === 4 ? "{\\pos(640,620)}" : "";
+    return `Dialogue: 0,${assTime(index * 4)},${assTime(index * 4 + 3)},${style},,0,0,0,,${tag}${line}`;
+  }).join("\n")}
+`;
+
+test("a Russian .ass is reported as Russian, and its markup does not reach the detector", () => {
+  const vtt = SubtitleFileContainer.toVtt(RUSSIAN_ASS, ".ass");
+  assert.ok(vtt.startsWith("WEBVTT"), "the conversion produced a WebVTT document");
+
+  const spoken = TextSubtitleTrack.cueTextOfVtt(vtt);
+  // Every one of these is a piece of the file the viewer never reads, and each
+  // was competing with the dialogue for the detector's answer.
+  for (const markup of [
+    "Aegisub", "Script Info", "V4+ Styles", "Format:", "Dialogue:",
+    "Trebuchet MS", "Times New Roman", "Default", "Signs", "pos(",
+    "-->", "0:00:12"
+  ]) {
+    assert.ok(!spoken.includes(markup), `cue text still carries "${markup}"`);
+  }
+  assert.ok(spoken.includes("Ты видишь их?"), "cue text keeps the dialogue");
+
+  assert.deepEqual(TextSubtitleTrack.detectLanguageFromVtt(vtt), { code: "ru", name: "Russian" });
+});
+
+test("the fixture is the hard case: the file itself is nearly half Latin", () => {
+  // Not an assertion about franc — an assertion that this fixture reproduces the
+  // field file's proportions. A fixture whose markup were negligible would pass
+  // the check above with or without the fix.
+  const cyrillic = (RUSSIAN_ASS.match(/[Ѐ-ӿ]/gu) ?? []).length;
+  const latin = (RUSSIAN_ASS.match(/[A-Za-z]/g) ?? []).length;
+  assert.ok(cyrillic > 800, `too little dialogue: ${cyrillic} Cyrillic characters`);
+  assert.ok(latin > cyrillic * 0.5, `markup too small to reproduce the fault: ${latin} vs ${cyrillic}`);
+});
+
+test("cue identifiers, NOTE, STYLE and REGION blocks are not text", () => {
+  const vtt = [
+    "WEBVTT - This file has cues.",
+    "Kind: captions",
+    "Language: en",
+    "",
+    "NOTE",
+    "Translated by an English speaking volunteer, all rights reserved.",
+    "",
+    "STYLE",
+    "::cue { background-image: linear-gradient(to bottom, dimgray, lightgray); }",
+    "",
+    "REGION",
+    "id:speaker width:40% lines:3 regionanchor:0%,100%",
+    "",
+    "opening-line",
+    "00:00:12.060 --> 00:00:13.270 align:start position:0%",
+    "<v Тоёхиса>Ты видишь их?</v>",
+    "",
+    "2",
+    "00:00:15.400 --> 00:00:16.650",
+    "<i>Но не столько вижу,</i>",
+    "сколько чувствую &amp; ощущаю.",
+    ""
+  ].join("\n");
+
+  const spoken = TextSubtitleTrack.cueTextOfVtt(vtt);
+  assert.equal(
+    spoken,
+    "Ты видишь их?\nНо не столько вижу,\nсколько чувствую   ощущаю."
+  );
+});
+
+test("a document with no cues yields no language rather than a guess", () => {
+  assert.equal(TextSubtitleTrack.cueTextOfVtt("WEBVTT\n"), "");
+  assert.equal(TextSubtitleTrack.detectLanguageFromVtt("WEBVTT\n"), null);
+  assert.equal(TextSubtitleTrack.detectLanguageFromVtt(null), null);
+});
+
+test("an embedded ASS cue is unframed by the container, then read", () => {
+  // What a Matroska block actually holds, per
+  // `matroska.org/technical/subtitles.html`: "Events are stored in the Block in
+  // this order: ReadOrder, Layer, Style, Name, MarginL, MarginR, MarginV,
+  // Effect, Text", with Start and End taken out into the block's own timing.
+  // Eight fields, and the earlier fixture here wrongly carried ten — it kept
+  // the two timestamps a FILE has, which is why it passed while the field
+  // subtitles of a real `.mkv` were shown to the viewer whole.
+  const blocks = RUSSIAN_DIALOGUE.map((line, index) => Buffer.from(
+    `${index + 1},0,Default,,0000,0000,0000,,` +
+    `${index % 5 === 4 ? "{\\pos(640,620)}" : ""}${line}`,
+    "utf8"
+  ));
+
+  const cues = blocks.map((payload, index) => ({
+    startSeconds: index * 4,
+    endSeconds: index * 4 + 3,
+    text: MatroskaContainer.cueTextOf(payload, "S_TEXT/ASS")
+  }));
+
+  const spoken = TextSubtitleTrack.finalizeCues(cues, "S_TEXT/ASS").map((cue) => cue.text).join("\n");
+  assert.ok(!spoken.includes("Default"), "the style field survived into the text");
+  assert.ok(!spoken.includes("0000"), "a margin field survived into the text");
+  assert.ok(!spoken.includes("pos("), "an override group survived into the text");
+  assert.deepEqual(detectLanguage(spoken), { code: "ru", name: "Russian" });
+});
+
+test("too little text is answered with nothing, not with a guess", () => {
+  // Measured 2026-09-02 (`research/franc-boundary-2026-09-02.md`): franc's
+  // answer for Russian walks between Bulgarian, Serbian and Russian until about
+  // 650 characters. Below that the honest answer is none — the browser then
+  // shows "Unknown", and the label moves when enough of the file has arrived.
+  const short = RUSSIAN_DIALOGUE.slice(0, 6).join("\n");
+  assert.ok(short.length < 300, `the fixture must be short: ${short.length}`);
+  assert.equal(detectLanguage(short), null);
+
+  const long = RUSSIAN_DIALOGUE.join("\n");
+  assert.ok(long.length > 650, `the fixture must be long enough: ${long.length}`);
+  assert.deepEqual(detectLanguage(long), { code: "ru", name: "Russian" });
+});
+
+test("an answer that does not survive losing half the text is not an answer", () => {
+  // A file carrying two languages — a bilingual release, or a track that turns
+  // into signs-only English partway. franc answers something for the whole; the
+  // halves disagree with it, and that disagreement is the text saying the
+  // answer rests on where it happened to be cut.
+  const english = [
+    "Do you see them out there beyond the wall?",
+    "I do, though not as clearly as I would like to.",
+    "They have been gathering since the morning came.",
+    "We should send word to the southern gate at once.",
+    "The road was cut yesterday and no rider will pass.",
+    "Then we hold what we have and wait for the dawn."
+  ].join("\n");
+  const mixed = `${RUSSIAN_DIALOGUE.slice(0, 13).join("\n")}\n${english}`;
+  assert.equal(detectLanguage(mixed), null);
+});
+
+test("a WebVTT body is decoded whole, so no character is cut in half", () => {
+  // The ffmpeg extraction path used to detect on `String(body.subarray(0, 4096))`.
+  // Two faults: a byte cut lands mid-character on any non-Latin track, and most
+  // of those bytes are timestamps rather than words. Cyrillic is two bytes per
+  // character in UTF-8, so 4096 bytes of this document is a few hundred
+  // characters of dialogue — inside the unstable region measured above.
+  const cues = RUSSIAN_DIALOGUE.map((line, index) => {
+    const start = `00:00:${String(index * 2).padStart(2, "0")}.000`;
+    const end = `00:00:${String(index * 2 + 1).padStart(2, "0")}.000`;
+    return `${start} --> ${end}\n${line}`;
+  });
+  const body = Buffer.from(`WEBVTT\n\n${cues.join("\n\n")}\n`, "utf8");
+
+  assert.deepEqual(TextSubtitleTrack.detectLanguageFromVtt(body.toString("utf8")), { code: "ru", name: "Russian" });
+});

package/test/timeline.test.js

@@ -0,0 +1,95 @@
+/**
+ * @file Where a file is cut is a fact about the FILE, held once.
+ *
+ * Every quality step of one film has to be cut at exactly the same times, and
+ * every session serving it has to publish the same playlist. That agreement was
+ * arranged by COPYING — a table handed to each new session at creation — and it
+ * drifted twice in the field: 0.6-2.9 s between two sessions of one film on
+ * 2026-08-17, and segments arriving a uniform 2.002 s before the times the
+ * playlist named for them on 2026-08-20, four times what a player bridges.
+ */
+
+import test from "node:test";
+import assert from "node:assert/strict";
+import { Timeline, Timelines } from "../services/output/Timeline.js";
+
+/**
+ * @returns {Timeline}
+ */
+function fourSecondGrid() {
+  // Only what is per grid. The container's own keyframe table, how exact it is,
+  // which container answered and how long the file runs are facts of the FILE
+  // and live on the source file — stating them here stated nothing.
+  return new Timeline({
+    boundaries: [0, 4, 8, 12, 16],
+    cutGrid: "keyframe"
+  });
+}
+
+test("the table the player was given does not move when the live one is corrected", () => {
+  const timeline = fourSecondGrid();
+
+  // A produced segment says where the file's cut really is. That corrects what
+  // a run will cut at; it must never correct what the player was told, because
+  // the player places a fragment by the playlist it holds and that text was
+  // written once.
+  timeline.boundaries[2] = 8.5;
+
+  assert.equal(timeline.liveStartOf(2), 8.5);
+  assert.equal(timeline.publishedStartOf(2), 8);
+});
+
+test("one file and grid answer with one table, whoever asks", () => {
+  const timelines = new Timelines();
+  const key = Timelines.keyFor("torrent:abc", 0, "keyframe");
+
+  const first = timelines.get(key, fourSecondGrid);
+  const second = timelines.get(key, () => {
+    throw new Error("a second table would be the drift this exists to remove");
+  });
+
+  assert.equal(second, first);
+  // A correction found by one session is seen by every other, because there is
+  // nothing to keep in step.
+  first.boundaries[1] = 4.5;
+  assert.equal(second.liveStartOf(1), 4.5);
+});
+
+test("the same file cut two ways is two tables", () => {
+  const timelines = new Timelines();
+  const onKeyframes = timelines.get(Timelines.keyFor("torrent:abc", 0, "keyframe"), fourSecondGrid);
+  const onTheEvenGrid = timelines.get(Timelines.keyFor("torrent:abc", 0, "uniform"), fourSecondGrid);
+
+  assert.notEqual(onTheEvenGrid, onKeyframes);
+});
+
+test("which segment holds a moment", () => {
+  const timeline = fourSecondGrid();
+
+  assert.equal(timeline.indexForTime(0), 0);
+  assert.equal(timeline.indexForTime(3.9), 0);
+  assert.equal(timeline.indexForTime(4), 1);
+  assert.equal(timeline.indexForTime(15.9), 3);
+  assert.equal(timeline.indexForTime(99), 3, "past the end is the last segment, not an error");
+});
+
+test("a fresh table starts with an empty tally of how well its index matched", () => {
+  const timeline = fourSecondGrid();
+
+  // The tally is a fact about the FILE and its index: asked per session it
+  // would be answered a different number of times for one film depending on
+  // how many people happened to watch it.
+  assert.equal(timeline.indexCheck.checked, 0);
+  assert.equal(timeline.indexCheck.disagreed, 0);
+  assert.equal(timeline.indexCheck.firstDisagreementIndex, -1);
+});
+
+test("a timeline nobody holds is dropped", () => {
+  const timelines = new Timelines();
+  const kept = timelines.get(Timelines.keyFor("torrent:abc", 0, "keyframe"), fourSecondGrid);
+  timelines.get(Timelines.keyFor("torrent:xyz", 0, "keyframe"), fourSecondGrid);
+
+  assert.equal(timelines.size, 2);
+  assert.equal(timelines.forgetUnused(new Set([kept])), 1);
+  assert.equal(timelines.size, 1);
+});