@torrent-tv/proxy 2.88.3 → 2.88.4

package/services/media/container/Mp4Container.js

@@ -1,1244 +1,1344 @@
-/**
- * @file MP4/MOV container — ISO/IEC 14496-12.
- *
- * Parses moov for all track types in one walk:
- *  - tkhd: track_ID, flags track_enabled (0x000001), alternate_group, width/height
- *  - mdhd: timescale, language (packed 5-bit), version handling
- *  - hdlr: handler_type (vide/soun/text/sbtl/subt/subp/clcp)
- *  - elng: extendedLanguage BCP47 (when present, replaces mdhd language per spec)
- *  - stsd: sample entry format (avc1/hev1/mp4a/tx3g/wvtt/stpp)
- *  - stbl tables for subtitle cue ranges (stts/stsz/stsc/stco/co64) and video keyframes (stss/stts/ctts/elst)
- *
- * Keyframe reading is delegated to mp4.js. The subtitle sample table is read in
- * this module, because every rule in it is a statement of ISO/IEC 14496-12 about
- * this container. Track creation is centralised here so that track_enabled,
- * `elng` and alternate_group are handled once for every media type.
- */
-
-import { Container } from "./Container.js";
-import { VideoTrack } from "../tracks/VideoTrack.js";
-import { AudioTrack } from "../tracks/AudioTrack.js";
-import { TextSubtitleTrack, TEXT_FORMATS_MP4 } from "../tracks/TextSubtitleTrack.js";
-import { ImageSubtitleTrack } from "../tracks/ImageSubtitleTrack.js";
-
-/** A box header is eight bytes, or sixteen when the size field says 1 (§4.2). */
-const HEADER_BYTES = 8;
-const LARGE_SIZE_MARKER = 1;
-const LARGE_HEADER_BYTES = 16;
-
-/**
- * One box header, per ISO/IEC 14496-12 §4.2.
- *
- * @param {Buffer} buf
- * @param {number} off
- * @returns {{ type: string, size: number, dataOffset: number, end: number } | null}
- */
-function readBox(buffer, offset) {
-  if (offset + HEADER_BYTES > buffer.length) {
-    return null;
-  }
-  let size = buffer.readUInt32BE(offset);
-  const type = buffer.toString("latin1", offset + 4, offset + 8);
-  let headerBytes = HEADER_BYTES;
-  // A size of 1 means the real size is the 64-bit value after the type.
-  if (size === LARGE_SIZE_MARKER) {
-    if (offset + LARGE_HEADER_BYTES > buffer.length) {
-      return null;
-    }
-    size = Number(buffer.readBigUInt64BE(offset + HEADER_BYTES));
-    headerBytes = LARGE_HEADER_BYTES;
-  }
-  if (size < headerBytes) {
-    return null;
-  }
-  return { type, size, dataOffset: offset + headerBytes, end: offset + size };
-}
-
-/**
- * Every direct child box of the given type.
- *
- * @param {Buffer} buf
- * @param {number} start
- * @param {number} end
- * @param {string} type
- * @returns {Array<{ type: string, size: number, dataOffset: number, end: number }>}
- */
-function childrenOf(buf, start, end, type) {
-  const out = [];
-  let p = start;
-  while (p + 8 <= end) {
-    const b = readBox(buf, p);
-    if (!b) break;
-    if (b.type === type) out.push(b);
-    p = b.end;
-  }
-  return out;
-}
-
-/**
- * The first direct child box of the given type, or null.
- *
- * @param {Buffer} buf
- * @param {number} s
- * @param {number} e
- * @param {string} t
- * @returns {{ type: string, size: number, dataOffset: number, end: number } | null}
- */
-function childOf(buf, s, e, t) {
-  return childrenOf(buf, s, e, t)[0] ?? null;
-}
-
-export class Mp4Container extends Container {
-  get formatName() {
-    return "mp4";
-  }
-
-  static detect(head) {
-    return isMp4(head);
-  }
-
-  /**
-   * The keyframe times this container's own index states, in ascending seconds.
-   *
-   * Static so a caller that has bytes and no container can ask; the instance
-   * form is {@link Container#readKeyframeIndex}.
-   *
-   * @param {(start:number,end:number)=>Promise<Buffer|null>} readRange
-   * @param {number} fileSize
-   * @returns {Promise<number[]|null>} Null where the container has no index.
-   */
-  static readKeyframeTimes(readRange, fileSize) {
-    return readMp4KeyframeTimes(readRange, fileSize);
-  }
-
-  /**
-   * This container's text subtitle tracks, with every cue's time and byte
-   * range — ISO/IEC 14496-12 §8.5 and §8.7.
-   *
-   * @param {(start:number,end:number)=>Promise<Buffer|null>} readRange
-   * @param {number} fileSize
-   * @returns {Promise<{ tracks: object[] } | null>}
-   */
-  static readSubtitlePlan(readRange, fileSize) {
-    return readMp4SubtitlePlan(readRange, fileSize);
-  }
-
-  /**
-   * The same reading, over the file this container was built on.
-   *
-   * The static form exists for a caller that has bytes and no container; this
-   * is the one to use otherwise, because the reader is already here.
-   *
-   * @returns {Promise<object|null>}
-   */
-  async readSubtitlePlan() {
-    const plan = await Mp4Container.readSubtitlePlan(this.readRange, this.fileSize);
-    if (!plan) {
-      return null;
-    }
-    // An MP4 states every sample's byte range in its own table, so a cue costs
-    // its own few dozen bytes rather than the cluster around it — which is why
-    // there are `samples` here and no `clusterPositions`.
-    //
-    // An MP4 has no element meaning "show this subtitle track by default", so
-    // `declared` is empty: the container states nothing, and nothing is shown
-    // unasked.
-    return {
-      tracks: plan.tracks.map((track, order) => ({
-        trackNumber: track.trackId,
-        declaredIndex: Number.isInteger(track.declaredIndex) ? track.declaredIndex : order,
-        codecId: track.format,
-        language: track.language,
-        name: "",
-        isDefault: order === 0,
-        codecPrivate: "",
-        clusterPositions: [],
-        samples: track.samples
-      })),
-      declared: [],
-      secondsPerTick: 0.001,
-      segmentDataOffset: 0
-    };
-  }
-
-
-  /**
-   * Read the samples of one subtitle track that are HELD now.
-   *
-   * An MP4 states every sample's byte range in its own table, so a cue costs its
-   * own few dozen bytes rather than the cluster around it — which is why this
-   * reads per sample where Matroska reads per cluster.
-   *
-   * Nothing is fetched: a sample whose bytes are not downloaded is left for the
-   * next call.
-   *
-   * @param {{ codecId: string, samples: {offset: number, size: number, startSeconds: number, endSeconds: number}[] }} track
-   * @param {Set<number>} harvested - Sample offsets already read; added to.
-   * @returns {Promise<{startSeconds: number, endSeconds: number, text: string}[]>}
-   *   The cues found in THIS pass.
-   */
-  async readHeldCues(_plan, track, progress) {
-    let harvested = progress.harvested.get(track.trackNumber);
-    if (!harvested) {
-      harvested = new Set();
-      progress.harvested.set(track.trackNumber, harvested);
-    }
-    const cues = await this.readHeldSamples(track, harvested);
-    return {
-      found: cues.length > 0 ? new Map([[track.trackNumber, cues]]) : new Map(),
-      covered: harvested.size,
-      indexed: track?.samples?.length ?? 0
-    };
-  }
-
-  async readHeldSamples(track, harvested) {
-    const found = [];
-    for (const sample of track?.samples ?? []) {
-      if (harvested.has(sample.offset)) {
-        continue;
-      }
-      const last = Math.min(this.fileSize - 1, sample.offset + sample.size - 1);
-      if (!this.isHeld(sample.offset, last)) {
-        continue;
-      }
-      const bytes = await this.readHeld(sample.offset, last);
-      if (!bytes) {
-        continue;
-      }
-      harvested.add(sample.offset);
-      // The MP4 has framed this cue and is the one that unframes it.
-      const text = Mp4Container.cueTextOf(bytes, track.codecId);
-      if (text) {
-        found.push({ startSeconds: sample.startSeconds, endSeconds: sample.endSeconds, text });
-      }
-    }
-    return found;
-  }
-
-  async readTracks() {
-    const head = await this.readRange(0, Math.min(64 - 1, this.fileSize - 1));
-    if (!head || !isMp4(head)) return [];
-
-    // Use the subtitle plan reader's moov parsing as source for subtitle tracks,
-    // and a direct moov walk for video/audio to collect tkhd/mdhd/hdlr/elng uniformly.
-    // For simplicity, delegate entirely to readMp4SubtitlePlan for subtitles and
-    // do a lightweight moov walk for video/audio here — then merge.
-    const subtitlePlan = await readMp4SubtitlePlan(this.readRange, this.fileSize).catch(() => null);
-    const subtitleByDecl = new Map();
-    if (subtitlePlan?.tracks) {
-      for (const t of subtitlePlan.tracks) subtitleByDecl.set(t.declaredIndex, t);
-    }
-
-    // Minimal moov walk for video/audio: reuse isMp4 + findMoov logic by reading via existing helper
-    // Instead of duplicating, parse tracks via a second full moov read that collects vide/soun.
-    // We read moov box directly to extract video/audio tracks.
-    const tracks = await this.#readVideoAudioTracks();
-    // Append subtitle tracks from plan, converting to domain objects
-    if (subtitlePlan?.tracks) {
-      for (const s of subtitlePlan.tracks) {
-        const isText = TEXT_FORMATS_MP4.has(s.format);
-        const Cls = isText ? TextSubtitleTrack : ImageSubtitleTrack;
-        tracks.push(new Cls({
-          trackNumber: s.trackId,
-          declaredIndex: s.declaredIndex,
-          codecId: s.format,
-          language: s.language,
-          languageBcp47: "",
-          name: "",
-          isEnabled: true,
-          isDefault: s.declaredIndex === 0,
-          declaresDefault: false,
-          codecPrivateB64: "",
-          // The sample entry's own words, read below. Either
-          // bit is enough: a file that sets only "all samples are forced" is
-          // saying what a well-formed one says twice.
-          isForced: s.someSamplesForced === true || s.allSamplesForced === true,
-          isHearingImpaired: false,
-          clusterPositions: [],
-          samples: s.samples
-        }));
-      }
-      // Count non-text subtitle handlers (subp/clcp/stpp) for declaredIndex correctness — they are already
-      // accounted for in subtitlePlan's declaredIndex via SUBTITLE_HANDLERS, but we didn't create objects for
-      // them above when they were stpp (non-text not in plan's tracks). The plan already excludes stpp from tracks
-      // but increments declaredIndex, so alignment holds: we don't need extra placeholders.
-    }
-    return tracks;
-  }
-
-  /**
-   * The `moov` box, read whole.
-   *
-   * Held on the instance because every question this class answers is inside
-   * it, and the box can be tens of megabytes off a torrent — reading it once
-   * per file is the difference between one fetch and one per question.
-   *
-   * @returns {Promise<{ moov: Buffer, header: number } | null>}
-   */
-  async #moovBuffer() {
-    if (this.moovHeld !== undefined) {
-      return this.moovHeld;
-    }
-    const PROBE = 64;
-    const MAX_MOOV = 32 * 1024 * 1024;
-    let at = 0;
-    let moovBox = null;
-    while (at < this.fileSize) {
-      const probe = await this.readRange(at, Math.min(this.fileSize - 1, at + PROBE - 1));
-      if (!probe || probe.length < 8) break;
-      let size = probe.readUInt32BE(0);
-      const type = probe.toString("latin1", 4, 8);
-      let header = 8;
-      if (size === 1) {
-        if (probe.length < 16) break;
-        size = Number(probe.readBigUInt64BE(8));
-        header = 16;
-      }
-      if (size <= 0) break;
-      if (type === "moov") { moovBox = { offset: at, size, header }; break; }
-      at += size;
-    }
-    if (!moovBox || moovBox.size > MAX_MOOV) {
-      this.moovHeld = null;
-      return null;
-    }
-    const moov = await this.readRange(moovBox.offset, Math.min(this.fileSize - 1, moovBox.offset + moovBox.size - 1));
-    this.moovHeld = moov ? { moov, header: moovBox.header } : null;
-    return this.moovHeld;
-  }
-
-  /**
-   * Duration from `mvhd` and the presentation offset from the first track's
-   * edit list, per ISO/IEC 14496-12 §8.2.2 and §8.6.6.
-   *
-   * An edit entry whose `media_time` is -1 is an EMPTY edit: it presents
-   * nothing for `segment_duration`, which shifts everything after it later by
-   * that much. That shift is what a player reports as the file's start, and it
-   * is the only way an MP4 states one — a file without such an edit begins at
-   * zero, which is a declaration, not an absence.
-   *
-   * @returns {Promise<import("./Container.js").ContainerMediaInfo>}
-   */
-  async readMediaInfo() {
-    if (this.mediaInfo) {
-      return this.mediaInfo;
-    }
-    /** @type {import("./Container.js").ContainerMediaInfo} */
-    const info = { format: this.formatName, durationSeconds: null, startTimeSeconds: null };
-    this.mediaInfo = info;
-    const held = await this.#moovBuffer();
-    if (!held) {
-      return info;
-    }
-    const { moov, header } = held;
-    const mvhd = childOf(moov, header, moov.length, "mvhd");
-    let movieTimescale = 0;
-    if (mvhd) {
-      const version = moov[mvhd.dataOffset];
-      // version 0: creation(4) modification(4) timescale(4) duration(4)
-      // version 1: creation(8) modification(8) timescale(4) duration(8)
-      const at = version === 1 ? mvhd.dataOffset + 20 : mvhd.dataOffset + 12;
-      if (at + 8 <= moov.length) {
-        movieTimescale = moov.readUInt32BE(at);
-        const duration = version === 1 ? Number(moov.readBigUInt64BE(at + 4)) : moov.readUInt32BE(at + 4);
-        if (movieTimescale > 0 && duration > 0) {
-          info.durationSeconds = duration / movieTimescale;
-        }
-      }
-    }
-    info.startTimeSeconds = movieTimescale > 0
-      ? Mp4Container.#emptyEditSeconds(moov, header, movieTimescale)
-      : null;
-    return info;
-  }
-
-  /**
-   * The presentation shift of the first empty edit, in seconds; 0 when no track
-   * declares one.
-   *
-   * @param {Buffer} moov
-   * @param {number} header
-   * @param {number} movieTimescale
-   * @returns {number}
-   */
-  static #emptyEditSeconds(moov, header, movieTimescale) {
-    let shift = 0;
-    for (const trak of childrenOf(moov, header, moov.length, "trak")) {
-      const edts = childOf(moov, trak.dataOffset, trak.end, "edts");
-      const elst = edts && childOf(moov, edts.dataOffset, edts.end, "elst");
-      if (!elst) {
-        continue;
-      }
-      const version = moov[elst.dataOffset];
-      const count = moov.readUInt32BE(elst.dataOffset + 4);
-      if (count < 1) {
-        continue;
-      }
-      const entry = elst.dataOffset + 8;
-      const segmentDuration = version === 1
-        ? Number(moov.readBigUInt64BE(entry))
-        : moov.readUInt32BE(entry);
-      const mediaTime = version === 1
-        ? Number(moov.readBigInt64BE(entry + 8))
-        : moov.readInt32BE(entry + 4);
-      if (mediaTime === -1 && segmentDuration > 0) {
-        shift = Math.max(shift, segmentDuration / movieTimescale);
-      }
-    }
-    return shift;
-  }
-
-  async #readVideoAudioTracks() {
-    const held = await this.#moovBuffer();
-    if (!held) return [];
-    const { moov, header: moovHeader } = held;
-
-    const result = [];
-    let videoIdx = -1;
-    let audioIdx = -1;
-
-    const moovContentStart = moovHeader;
-    const moovEnd = moov.length;
-    for (const trak of childrenOf(moov, moovContentStart, moovEnd, "trak")) {
-      const mdia = childOf(moov, trak.dataOffset, trak.end, "mdia");
-      if (!mdia) continue;
-      const hdlr = childOf(moov, mdia.dataOffset, mdia.end, "hdlr");
-      const handler = hdlr ? moov.toString("latin1", hdlr.dataOffset + 8, hdlr.dataOffset + 12) : "";
-      const mdhd = childOf(moov, mdia.dataOffset, mdia.end, "mdhd");
-      const tkhd = childOf(moov, trak.dataOffset, trak.end, "tkhd");
-      let language = "";
-      if (mdhd) {
-        const ver = moov[mdhd.dataOffset];
-        const langAt = ver === 1 ? mdhd.dataOffset + 32 : mdhd.dataOffset + 20;
-        if (langAt + 2 <= mdhd.end) {
-          const packed = moov.readUInt16BE(langAt);
-          language = [10, 5, 0].map((s) => String.fromCharCode(((packed >> s) & 0x1f) + 0x60)).join("").replace(/[^a-z]/g, "");
-        }
-      }
-      // elng overrides mdhd language per spec §8.4.6
-      let languageBcp47 = "";
-      const elng = childOf(moov, mdia.dataOffset, mdia.end, "elng");
-      if (elng && elng.end - elng.dataOffset >= 4) {
-        languageBcp47 = moov.toString("utf8", elng.dataOffset + 4, elng.end).replace(/\0+$/, "");
-      }
-      let trackId = 0;
-      let isEnabled = true;
-      let alternateGroup = 0;
-      let width = null;
-      let height = null;
-      if (tkhd) {
-        const ver = moov[tkhd.dataOffset];
-        const flags = moov.readUInt32BE(tkhd.dataOffset + 1) & 0xffffff; // 3 bytes after version
-        isEnabled = (flags & 0x000001) !== 0;
-        trackId = moov.readUInt32BE(ver === 1 ? tkhd.dataOffset + 20 : tkhd.dataOffset + 12);
-        alternateGroup = moov.readUInt16BE(ver === 1 ? tkhd.dataOffset + 26 : tkhd.dataOffset + 18);
-        // width/height are 16.16 fixed point at end of tkhd
-        if (tkhd.end - tkhd.dataOffset >= 84) {
-          const w = moov.readUInt32BE(ver === 1 ? tkhd.dataOffset + 76 : tkhd.dataOffset + 68);
-          const h = moov.readUInt32BE(ver === 1 ? tkhd.dataOffset + 80 : tkhd.dataOffset + 72);
-          width = w / 65536;
-          height = h / 65536;
-        }
-      }
-      const resolvedLang = languageBcp47 || language;
-      if (handler === "vide") {
-        videoIdx += 1;
-        // stsd format for codecId
-        let codecId = "";
-        const minf = childOf(moov, mdia.dataOffset, mdia.end, "minf");
-        const stbl = minf && childOf(moov, minf.dataOffset, minf.end, "stbl");
-        const stsd = stbl && childOf(moov, stbl.dataOffset, stbl.end, "stsd");
-        if (stsd) {
-          const first = readBox(moov, stsd.dataOffset + 8);
-          if (first) codecId = first.type;
-        }
-        result.push(new VideoTrack({
-          trackNumber: trackId,
-          declaredIndex: videoIdx,
-          codecId,
-          language: resolvedLang,
-          languageBcp47,
-          name: "",
-          isEnabled,
-          isDefault: true,
-          declaresDefault: false,
-          codecPrivateB64: "",
-          alternateGroup,
-          width,
-          height
-        }));
-      } else if (handler === "soun") {
-        audioIdx += 1;
-        let codecId = "";
-        const minf = childOf(moov, mdia.dataOffset, mdia.end, "minf");
-        const stbl = minf && childOf(moov, minf.dataOffset, minf.end, "stbl");
-        const stsd = stbl && childOf(moov, stbl.dataOffset, stbl.end, "stsd");
-        if (stsd) {
-          const first = readBox(moov, stsd.dataOffset + 8);
-          if (first) codecId = first.type;
-        }
-        result.push(new AudioTrack({
-          trackNumber: trackId,
-          declaredIndex: audioIdx,
-          codecId,
-          language: resolvedLang,
-          languageBcp47,
-          name: "",
-          isEnabled,
-          isDefault: true,
-          declaresDefault: false,
-          codecPrivateB64: "",
-          alternateGroup,
-          isOriginal: false,
-          isCommentary: false,
-          isVisualImpaired: false
-        }));
-      }
-    }
-    return result;
-  }
-
-  /**
-   * The text field of one cue as MP4 frames it.
-   *
-   * A `tx3g`/`text` sample is a 16-bit big-endian length followed by that many
-   * bytes of UTF-8 (ISO/IEC 14496-12 §12.6.3 and Apple's text sample format); a
-   * `wvtt` sample is a sequence of boxes whose `vttc`/`payl` holds the cue text
-   * (§12.6.3.2). Neither carries the subtitle format's own markup, so the
-   * markup step that follows has nothing to take off — it is applied all the
-   * same, because which step applies is decided by the codec and not here.
-   *
-   * The byte reading itself is this module's,
-   * alongside the sample-table walk that found the range.
-   *
-   * @param {Buffer} payload - The sample's own bytes.
-   * @param {string} codecId - Sample entry type: `tx3g`, `text` or `wvtt`.
-   * @returns {string}
-   */
-  static cueTextOf(payload, codecId) {
-    return decodeSubtitleSample(payload, codecId);
-  }
-
-  async parseKeyframeIndex() {
-    const r = await readMp4KeyframeTimes(this.readRange, this.fileSize);
-    if (!r) return null;
-    if (Array.isArray(r)) return { times: r, tolerance: 0 };
-    return r;
-  }
-}
-
-// ---------------------------------------------------------------------------
-// The MP4's own reading of its subtitle sample table. Here because every rule
-// in it is ISO/IEC 14496-12 speaking about this container, and the class is
-// the only way in.
-// ---------------------------------------------------------------------------
-/**
- * @file The text subtitle tracks of an MP4, and where each cue's bytes are.
- *
- * The same rule as the Matroska side: nothing is extracted with ffmpeg and
- * nothing is fetched for its own sake. Here it is cheaper still. Matroska hides
- * its subtitle blocks inside clusters shared with the picture, so a cue costs
- * whatever cluster holds it; an MP4 states every sample's offset and length in
- * the sample table, so a cue costs its own bytes and nothing more — usually a
- * few dozen of them.
- *
- * The tables, from ISO/IEC 14496-12:
- *
- *   stsd — what the samples are (`tx3g` timed text, `wvtt` WebVTT, `stpp` TTML)
- *   stts — how long each sample lasts, run-length encoded (§8.6.1.2)
- *   stsz — how long each sample is, in bytes (§8.7.3)
- *   stsc — how samples are grouped into chunks (§8.7.4)
- *   stco / co64 — where each chunk begins in the file (§8.7.5)
- *
- * Together they give, for sample N: when it starts, how long it stays, and the
- * exact byte range holding it. That is everything needed to show a cue without
- * reading anything else.
- */
-
-const PROBE_BYTES = 64;
-const MAX_MOOV_BYTES = 32 * 1024 * 1024;
-
-/** Handlers that mean "this track is text on screen". */
-const TEXT_HANDLERS = new Set(["text", "sbtl", "subt"]);
-/**
- * Every handler ffmpeg's mov demuxer turns into a SUBTITLE stream, whether or
- * not this file can read it — `subp` is a DVD subpicture and `clcp` closed
- * captions, both pictures or caption data rather than text. They are counted
- * because `declaredIndex` has to equal ffmpeg's `0:s:N`, and a track left out
- * of the count shifts every text track after it, which is the very defect
- * `declaredIndex` exists to remove.
- */
-const SUBTITLE_HANDLERS = new Set([...TEXT_HANDLERS, "subp", "clcp"]);
-/** Sample formats this can turn into cues. `stpp` (TTML) is XML and is not one. */
-const TEXT_FORMATS = new Set(["tx3g", "text", "wvtt"]);
-
-
-
-
-
-/**
- * Sample durations, expanded from the run-length table.
- *
- * @param {Buffer} moov
- * @param {{ dataOffset: number, end: number }} stts
- * @param {number} total - How many samples the size table declares.
- * @returns {number[]} Ticks each sample lasts.
- */
-function sampleDurations(moov, stts, total) {
-  const durations = new Array(total).fill(0);
-  const entries = moov.readUInt32BE(stts.dataOffset + 4);
-  let at = stts.dataOffset + 8;
-  let sample = 0;
-  for (let entry = 0; entry < entries && at + 8 <= stts.end && sample < total; entry += 1, at += 8) {
-    const count = moov.readUInt32BE(at);
-    const delta = moov.readUInt32BE(at + 4);
-    for (let index = 0; index < count && sample < total; index += 1, sample += 1) {
-      durations[sample] = delta;
-    }
-  }
-  return durations;
-}
-
-/**
- * Sample sizes, whether the table states one for all or one for each.
- *
- * @param {Buffer} moov
- * @param {{ dataOffset: number, end: number }} stsz
- * @returns {number[]}
- */
-function sampleSizes(moov, stsz) {
-  const uniform = moov.readUInt32BE(stsz.dataOffset + 4);
-  const count = moov.readUInt32BE(stsz.dataOffset + 8);
-  if (uniform > 0) {
-    return new Array(count).fill(uniform);
-  }
-  const sizes = new Array(count).fill(0);
-  let at = stsz.dataOffset + 12;
-  for (let index = 0; index < count && at + 4 <= stsz.end; index += 1, at += 4) {
-    sizes[index] = moov.readUInt32BE(at);
-  }
-  return sizes;
-}
-
-/**
- * Where every sample of a track begins in the file.
- *
- * The sample-to-chunk table says how many samples each run of chunks holds, and
- * the chunk-offset table says where each chunk starts; a sample's own offset is
- * its chunk's start plus the sizes of the samples before it in that chunk.
- *
- * @param {Buffer} moov
- * @param {{ dataOffset: number, end: number }} stsc
- * @param {number[]} chunkOffsets
- * @param {number[]} sizes
- * @returns {number[]}
- */
-function sampleOffsets(moov, stsc, chunkOffsets, sizes) {
-  const offsets = new Array(sizes.length).fill(0);
-  const entries = moov.readUInt32BE(stsc.dataOffset + 4);
-  /** @type {{ firstChunk: number, perChunk: number }[]} */
-  const runs = [];
-  let at = stsc.dataOffset + 8;
-  for (let entry = 0; entry < entries && at + 12 <= stsc.end; entry += 1, at += 12) {
-    runs.push({ firstChunk: moov.readUInt32BE(at), perChunk: moov.readUInt32BE(at + 4) });
-  }
-  let sample = 0;
-  for (let run = 0; run < runs.length && sample < sizes.length; run += 1) {
-    const from = runs[run].firstChunk;
-    const to = run + 1 < runs.length ? runs[run + 1].firstChunk - 1 : chunkOffsets.length;
-    for (let chunk = from; chunk <= to && sample < sizes.length; chunk += 1) {
-      let inChunk = chunkOffsets[chunk - 1];
-      if (inChunk === undefined) {
-        break;
-      }
-      for (let index = 0; index < runs[run].perChunk && sample < sizes.length; index += 1, sample += 1) {
-        offsets[sample] = inChunk;
-        inChunk += sizes[sample];
-      }
-    }
-  }
-  return offsets;
-}
-
-/**
- * @typedef {object} Mp4SubtitleSample
- * @property {number} startSeconds
- * @property {number} endSeconds
- * @property {number} offset - Where the sample's bytes are in the file.
- * @property {number} size
- */
-
-/**
- * @typedef {object} Mp4SubtitleTrack
- * @property {number} trackId
- * @property {number} declaredIndex - Its position among ALL of the file's
- *   subtitle tracks, including the ones whose sample format this cannot turn
- *   into cues (`stpp` TTML). That is the number ffmpeg gives the same stream in
- *   `0:s:N`, which is the number the browser names; counting only the readable
- *   ones would shift every track after a TTML one.
- * @property {string} format - `tx3g`, `text` or `wvtt`.
- * @property {string} language - Three letters, as the file declares them.
- * @property {boolean} someSamplesForced - The sample entry says at least one
- *   cue carries a forced (`frcd`) atom.
- * @property {boolean} allSamplesForced - The sample entry says every cue is to
- *   be treated as forced, whether or not it carries that atom.
- * @property {Mp4SubtitleSample[]} samples - In time order.
- */
-
-/**
- * The text subtitle tracks of an MP4, with every cue's time and byte range.
- *
- * @param {(start: number, end: number) => Promise<Buffer | null>} readRange
- * @param {number} fileSize
- * @returns {Promise<{ tracks: Mp4SubtitleTrack[] } | null>}
- */
-async function readMp4SubtitlePlan(readRange, fileSize) {
-  const found = await findMoov(readRange, fileSize);
-  if (!found || found.size > MAX_MOOV_BYTES) {
-    return null;
-  }
-  const moov = await readRange(found.offset, Math.min(fileSize - 1, found.offset + found.size - 1));
-  if (!moov || moov.length < HEADER_BYTES) {
-    return null;
-  }
-  const moovBox = readBox(moov, 0);
-  if (!moovBox) {
-    return null;
-  }
-
-  /** @type {Mp4SubtitleTrack[]} */
-  const tracks = [];
-  // Counts every subtitle track the file has, whether or not this can read it,
-  // so the number handed out matches ffmpeg's `0:s:N`. See `declaredIndex`.
-  let declaredIndex = -1;
-  for (const trak of childrenOf(moov, moovBox.dataOffset, moov.length, "trak")) {
-    const mdia = childOf(moov, trak.dataOffset, trak.end, "mdia");
-    if (!mdia) {
-      continue;
-    }
-    const hdlr = childOf(moov, mdia.dataOffset, mdia.end, "hdlr");
-    const handler = hdlr ? moov.toString("latin1", hdlr.dataOffset + 8, hdlr.dataOffset + 12) : "";
-    if (!SUBTITLE_HANDLERS.has(handler)) {
-      continue;
-    }
-    // Counted before the readability checks below, and before the handler is
-    // narrowed to the text ones: this number is the track's place in the file,
-    // not its place among the tracks this code can turn into cues.
-    declaredIndex += 1;
-    if (!TEXT_HANDLERS.has(handler)) {
-      continue;
-    }
-    const mdhd = childOf(moov, mdia.dataOffset, mdia.end, "mdhd");
-    if (!mdhd) {
-      continue;
-    }
-    const version = moov[mdhd.dataOffset];
-    const timescale = moov.readUInt32BE(version === 1 ? mdhd.dataOffset + 20 : mdhd.dataOffset + 12);
-    if (!timescale) {
-      continue;
-    }
-    // The language is five bits per letter, offset from 0x60, packed into two
-    // bytes after the times (ISO/IEC 14496-12 §8.4.2.3).
-    const languageAt = version === 1 ? mdhd.dataOffset + 32 : mdhd.dataOffset + 20;
-    let language = "";
-    if (languageAt + 2 <= mdhd.end) {
-      const packed = moov.readUInt16BE(languageAt);
-      language = [10, 5, 0]
-        .map((shift) => String.fromCharCode(((packed >> shift) & 0x1f) + 0x60))
-        .join("")
-        .replace(/[^a-z]/g, "");
-    }
-
-    const tkhd = childOf(moov, trak.dataOffset, trak.end, "tkhd");
-    const trackId = tkhd
-      ? moov.readUInt32BE(moov[tkhd.dataOffset] === 1 ? tkhd.dataOffset + 20 : tkhd.dataOffset + 12)
-      : tracks.length + 1;
-
-    const minf = childOf(moov, mdia.dataOffset, mdia.end, "minf");
-    const stbl = minf && childOf(moov, minf.dataOffset, minf.end, "stbl");
-    if (!stbl) {
-      continue;
-    }
-    const stsd = childOf(moov, stbl.dataOffset, stbl.end, "stsd");
-    const first = stsd && readBox(moov, stsd.dataOffset + 8);
-    const format = first ? first.type : "";
-    if (!TEXT_FORMATS.has(format)) {
-      continue;
-    }
-    // Whether the file itself says this track is forced — subtitles shown even
-    // to a viewer who did not ask for subtitles, because the dialogue on screen
-    // is in a language the soundtrack is not.
-    //
-    // Apple's QuickTime File Format, "Display flags" under Subtitle sample
-    // description, defines the two bits read here: `0x40000000` "Some samples
-    // are forced" ("at least one sample contains a forced (`frcd`) atom") and
-    // `0x80000000` "All samples are forced" ("the subtitle media handler treats
-    // all samples as forced subtitles, regardless of the presence or absence of
-    // a `frcd` atom"), with the note that setting the second requires the first
-    // — the pair together being `0xC0000000`. We honour that requirement rather
-    // than trusting a writer to have met it: either bit alone is enough to call
-    // the track forced, because a file that sets only `0x80000000` is saying
-    // exactly what a well-formed one would say twice.
-    //
-    // The field sits at a fixed place in the sample entry. `dataOffset` is
-    // already past the box header, and every sample entry opens with 6 reserved
-    // bytes and a 2-byte data reference index (ISO/IEC 14496-12 §8.5.2.2), so
-    // `displayFlags` is the 32 bits eight bytes in.
-    let someSamplesForced = false;
-    let allSamplesForced = false;
-    if (format === "tx3g" && first && first.dataOffset + 8 + 4 <= first.end) {
-      const displayFlags = moov.readUInt32BE(first.dataOffset + 8);
-      someSamplesForced = (displayFlags & 0x40000000) !== 0;
-      allSamplesForced = (displayFlags & 0x80000000) !== 0;
-    }
-    const stts = childOf(moov, stbl.dataOffset, stbl.end, "stts");
-    const stsz = childOf(moov, stbl.dataOffset, stbl.end, "stsz");
-    const stsc = childOf(moov, stbl.dataOffset, stbl.end, "stsc");
-    const stco = childOf(moov, stbl.dataOffset, stbl.end, "stco");
-    const co64 = childOf(moov, stbl.dataOffset, stbl.end, "co64");
-    if (!stts || !stsz || !stsc || (!stco && !co64)) {
-      continue;
-    }
-
-    const sizes = sampleSizes(moov, stsz);
-    const durations = sampleDurations(moov, stts, sizes.length);
-    const chunkOffsets = [];
-    if (stco) {
-      const count = moov.readUInt32BE(stco.dataOffset + 4);
-      let at = stco.dataOffset + 8;
-      for (let index = 0; index < count && at + 4 <= stco.end; index += 1, at += 4) {
-        chunkOffsets.push(moov.readUInt32BE(at));
-      }
-    } else {
-      const count = moov.readUInt32BE(co64.dataOffset + 4);
-      let at = co64.dataOffset + 8;
-      for (let index = 0; index < count && at + 8 <= co64.end; index += 1, at += 8) {
-        chunkOffsets.push(Number(moov.readBigUInt64BE(at)));
-      }
-    }
-    const offsets = sampleOffsets(moov, stsc, chunkOffsets, sizes);
-
-    /** @type {Mp4SubtitleSample[]} */
-    const samples = [];
-    let ticks = 0;
-    for (let index = 0; index < sizes.length; index += 1) {
-      const start = ticks / timescale;
-      ticks += durations[index];
-      // An empty sample is a gap between cues, which the format uses to say
-      // "nothing on screen"; it is not a cue and would show as a blank line.
-      if (sizes[index] > 2) {
-        samples.push({
-          startSeconds: start,
-          endSeconds: ticks / timescale,
-          offset: offsets[index],
-          size: sizes[index]
-        });
-      }
-    }
-    tracks.push({
-      trackId,
-      declaredIndex,
-      format,
-      language,
-      someSamplesForced,
-      allSamplesForced,
-      samples
-    });
-  }
-  return { tracks };
-}
-
-/**
- * The text of one sample.
- *
- * `tx3g` is a two-byte length followed by UTF-8; anything after that is styling
- * boxes, which this deliberately drops. `wvtt` is a sequence of boxes, and the
- * text lives in the `payl` inside a `vttc`.
- *
- * @param {Buffer} bytes
- * @param {string} format
- * @returns {string}
- */
-function decodeSubtitleSample(bytes, format) {
-  if (format === "wvtt") {
-    let at = 0;
-    const parts = [];
-    while (at + HEADER_BYTES <= bytes.length) {
-      const box = readBox(bytes, at);
-      if (!box) {
-        break;
-      }
-      if (box.type === "vttc") {
-        const payl = childOf(bytes, box.dataOffset, Math.min(bytes.length, box.end), "payl");
-        if (payl) {
-          parts.push(bytes.toString("utf8", payl.dataOffset, Math.min(bytes.length, payl.end)));
-        }
-      }
-      at = box.end;
-    }
-    return parts.join("\n").trim();
-  }
-  if (bytes.length < 2) {
-    return "";
-  }
-  const length = bytes.readUInt16BE(0);
-  return bytes.toString("utf8", 2, Math.min(bytes.length, 2 + length)).trim();
-}
-
-// ---------------------------------------------------------------------------
-// ISO/IEC 14496-12 speaking about MP4: stss, stts, ctts and the edit list.
-// Here because the class is the only way in.
-// ---------------------------------------------------------------------------
-/**
- * @file Keyframe index for MP4/MOV, read without downloading the file.
- *
- * MP4 keeps its tables in a `moov` box: `stss` lists which samples are sync
- * samples (keyframes) by number, and `stts` gives each sample's duration, so
- * the two together turn "sample #N" into "second T". `moov` sits either at the
- * start (files written for streaming) or at the end (the common case for a
- * plain mux); box headers state their own size, so it is found by stepping over
- * top-level boxes rather than scanning bytes — a couple of 64-byte reads even
- * when `mdat` is gigabytes.
- *
- * Same purpose as the Matroska reader: on the video-COPY path the segment
- * boundaries ARE the source's keyframes, and inventing an even grid instead
- * makes players walk the whole file or present audio with no picture.
- */
-
-// A 64-bit box size is signalled by a 32-bit size of 1, the real size following
-// in the next 8 bytes.
-// Enough to read any box header while walking the top level.
-// Cap on the moov read. A feature-length file indexes to a few hundred KB;
-// beyond this is not a normal index and not worth pulling over a torrent.
-
-/**
- * Whether this looks like MP4/MOV — every real file opens with an `ftyp` box.
- *
- * @param {Buffer} head
- * @returns {boolean}
- */
-function isMp4(head) {
-  return head.length >= 12 && head.toString("latin1", 4, 8) === "ftyp";
-}
-
-/**
- * Read a box header at `offset`.
- *
- * @param {Buffer} buffer
- * @param {number} offset
- * @returns {{ type: string, size: number, headerBytes: number } | null}
- */
-function readBoxHeader(buffer, offset) {
-  if (offset + HEADER_BYTES > buffer.length) {
-    return null;
-  }
-  const size32 = buffer.readUInt32BE(offset);
-  const type = buffer.toString("latin1", offset + 4, offset + 8);
-  if (size32 === LARGE_SIZE_MARKER) {
-    if (offset + LARGE_HEADER_BYTES > buffer.length) {
-      return null;
-    }
-    // High word is zero for any file we can practically handle.
-    const high = buffer.readUInt32BE(offset + 8);
-    const low = buffer.readUInt32BE(offset + 12);
-    return { type, size: high * 4294967296 + low, headerBytes: LARGE_HEADER_BYTES };
-  }
-  return { type, size: size32, headerBytes: HEADER_BYTES };
-}
-
-/**
- * Walk the top-level boxes to find `moov`, reading only each box header.
- *
- * @param {(start: number, end: number) => Promise<Buffer | null>} readRange
- * @param {number} fileSize
- * @returns {Promise<{ offset: number, size: number, headerBytes: number } | null>}
- */
-async function findMoov(readRange, fileSize) {
-  let offset = 0;
-  while (offset < fileSize) {
-    const probe = await readRange(offset, Math.min(fileSize - 1, offset + PROBE_BYTES - 1));
-    if (!probe || probe.length < HEADER_BYTES) {
-      return null;
-    }
-    const header = readBoxHeader(probe, 0);
-    // Size 0 means "extends to end of file" — legal only for the last box, and
-    // never for one we would step over.
-    if (!header || header.size <= 0) {
-      return null;
-    }
-    if (header.type === "moov") {
-      return { offset, size: header.size, headerBytes: header.headerBytes };
-    }
-    offset += header.size;
-  }
-  return null;
-}
-
-/**
- * Find the first box of `type` directly inside a range of an already-read buffer.
- *
- * @param {Buffer} buffer
- * @param {number} start
- * @param {number} end
- * @param {string} type
- * @returns {{ dataOffset: number, end: number } | null}
- */
-function findBox(buffer, start, end, type) {
-  let offset = start;
-  while (offset + HEADER_BYTES <= end) {
-    const header = readBoxHeader(buffer, offset);
-    if (!header || header.size <= 0) {
-      return null;
-    }
-    if (header.type === type) {
-      return { dataOffset: offset + header.headerBytes, end: Math.min(end, offset + header.size) };
-    }
-    offset += header.size;
-  }
-  return null;
-}
-
-/**
- * All boxes of `type` directly inside a range.
- *
- * @param {Buffer} buffer
- * @param {number} start
- * @param {number} end
- * @param {string} type
- * @returns {{ dataOffset: number, end: number }[]}
- */
-function findAllBoxes(buffer, start, end, type) {
-  const found = [];
-  let offset = start;
-  while (offset + HEADER_BYTES <= end) {
-    const header = readBoxHeader(buffer, offset);
-    if (!header || header.size <= 0) {
-      break;
-    }
-    if (header.type === type) {
-      found.push({ dataOffset: offset + header.headerBytes, end: Math.min(end, offset + header.size) });
-    }
-    offset += header.size;
-  }
-  return found;
-}
-
-/**
- * Turn sample numbers into seconds using the time-to-sample table.
- *
- * `stts` is run-length encoded — pairs of (sample count, per-sample duration) —
- * so one walk yields every sample's start time without expanding the table.
- *
- * @param {Buffer} buffer
- * @param {{ dataOffset: number, end: number }} stts
- * @param {number} timescale - Ticks per second.
- * @param {Set<number>} wanted - Sample numbers (1-based).
- * @returns {number[]} Seconds, ascending.
- */
-function resolveSampleTimes(buffer, stts, timescale, wanted, offsets = null) {
-  const entryCount = buffer.readUInt32BE(stts.dataOffset + 4);
-  const times = [];
-  let sampleNumber = 1;
-  let ticks = 0;
-  let cursor = stts.dataOffset + 8;
-  for (let entry = 0; entry < entryCount && cursor + 8 <= stts.end; entry += 1) {
-    const count = buffer.readUInt32BE(cursor);
-    const delta = buffer.readUInt32BE(cursor + 4);
-    for (let index = 0; index < count; index += 1) {
-      if (wanted.has(sampleNumber)) {
-        // `CT(n) = DT(n) + CTTS(n)` — ISO/IEC 14496-12 §8.6.1.3. The offset is
-        // what turns decode order into the order frames are shown in, and it is
-        // the timeline ffmpeg cuts on.
-        times.push((ticks + (offsets?.get(sampleNumber) ?? 0)) / timescale);
-      }
-      ticks += delta;
-      sampleNumber += 1;
-    }
-    cursor += 8;
-  }
-  return times;
-}
-
-/**
- * Composition offsets for the sample numbers asked for.
- *
- * `ctts` is run-length encoded like `stts`, and version 1 carries SIGNED
- * offsets — which is what the version exists for: a frame may be shown before
- * it is decoded. Reading them as unsigned turns a small negative offset into
- * roughly four billion ticks.
- *
- * @param {Buffer} buffer
- * @param {{ dataOffset: number, end: number }} ctts
- * @param {Set<number>} wanted - Sample numbers (1-based).
- * @returns {Map<number, number>} Sample number to offset in media ticks.
- */
-function readCompositionOffsets(buffer, ctts, wanted) {
-  const version = buffer[ctts.dataOffset];
-  const entryCount = buffer.readUInt32BE(ctts.dataOffset + 4);
-  const offsets = new Map();
-  let sampleNumber = 1;
-  let cursor = ctts.dataOffset + 8;
-  for (let entry = 0; entry < entryCount && cursor + 8 <= ctts.end; entry += 1) {
-    const count = buffer.readUInt32BE(cursor);
-    const offset = version === 1 ? buffer.readInt32BE(cursor + 4) : buffer.readUInt32BE(cursor + 4);
-    for (let index = 0; index < count; index += 1) {
-      if (wanted.has(sampleNumber)) {
-        offsets.set(sampleNumber, offset);
-      }
-      sampleNumber += 1;
-    }
-    cursor += 8;
-  }
-  return offsets;
-}
-
-/**
- * How far the edit list shifts this track's composition timeline, in media
- * ticks.
- *
- * ISO/IEC 14496-12 §8.6.6.3: `media_time` is the start of the edit within the
- * media, in the MEDIA timescale and in composition time, while
- * `segment_duration` is in the MOVIE timescale — two different units in one
- * structure, which is why only the first is read here. `media_time = -1` is an
- * empty edit: it inserts blank presentation time and starts no media, so the
- * first real edit is the one that matters.
- *
- * Measured 2026-08-19: every LostFilm MP4 that carries a composition offset
- * also carries an edit list cancelling it exactly, which is why decode times
- * have been right on those files. `Firefly.S01E03` has the offset and NO edit
- * list, and its times were 62.1 ms early on all 34 keyframes checked.
- *
- * @param {Buffer} buffer
- * @param {{ dataOffset: number, end: number }} elst
- * @returns {number} Ticks to subtract; zero when nothing is shifted.
- */
-function readEditShift(buffer, elst) {
-  const version = buffer[elst.dataOffset];
-  const entryCount = buffer.readUInt32BE(elst.dataOffset + 4);
-  const wide = version === 1;
-  const entryBytes = wide ? 20 : 12;
-  let cursor = elst.dataOffset + 8;
-  for (let entry = 0; entry < entryCount && cursor + entryBytes <= elst.end; entry += 1) {
-    const mediaTime = wide
-      ? Number(buffer.readBigInt64BE(cursor + 8))
-      : buffer.readInt32BE(cursor + 4);
-    if (mediaTime >= 0) {
-      return mediaTime;
-    }
-    cursor += entryBytes;
-  }
-  return 0;
-}
-
-/**
- * Whether this track's handler says it carries video.
- *
- * The standard identifies a track by its `hdlr`, and nothing else does. Picking
- * "the first track that happens to carry sync samples" worked only because the
- * seven releases measured all put video first; a file whose audio track carries
- * them, or one that leads with a cover-art video track, would be read from the
- * wrong place. That is the same defect that was fixed in the Matroska reader on
- * 2026-08-18, arrived at from the other side.
- *
- * @param {Buffer} buffer
- * @param {{ dataOffset: number, end: number }} mdia
- * @returns {boolean}
- */
-function isVideoTrack(buffer, mdia) {
-  const hdlr = findBox(buffer, mdia.dataOffset, mdia.end, "hdlr");
-  if (!hdlr || hdlr.dataOffset + 12 > hdlr.end) {
-    return false;
-  }
-  // FullBox header (4) then a reserved pre_defined (4), then the handler type.
-  return buffer.toString("latin1", hdlr.dataOffset + 8, hdlr.dataOffset + 12) === "vide";
-}
-
-/**
- * Read the keyframe times of an MP4/MOV file.
- *
- * @param {(start: number, end: number) => Promise<Buffer | null>} readRange
- * @param {number} fileSize
- * @returns {Promise<number[] | null>} Ascending seconds, or null when the file
- *   carries no usable index (fragmented MP4, truncated or damaged `moov`).
- */
-async function readMp4KeyframeTimes(readRange, fileSize) {
-  const moovBox = await findMoov(readRange, fileSize);
-  if (!moovBox || moovBox.size > MAX_MOOV_BYTES) {
-    return null;
-  }
-
-  const moov = await readRange(moovBox.offset, Math.min(fileSize - 1, moovBox.offset + moovBox.size - 1));
-  if (!moov || moov.length < moovBox.headerBytes) {
-    return null;
-  }
-
-  // Examine every track, and take the one whose HANDLER says it is video. A
-  // track with no `stss` has every sample a keyframe, so it constrains nothing
-  // and is skipped even when it is the video one.
-  for (const trak of findAllBoxes(moov, moovBox.headerBytes, moov.length, "trak")) {
-    const mdia = findBox(moov, trak.dataOffset, trak.end, "mdia");
-    if (!mdia || !isVideoTrack(moov, mdia)) {
-      continue;
-    }
-    const mdhd = findBox(moov, mdia.dataOffset, mdia.end, "mdhd");
-    if (!mdhd) {
-      continue;
-    }
-    // mdhd layout: version(1) + flags(3), then creation/modification times —
-    // 32-bit each in version 0, 64-bit in version 1 — then the timescale.
-    const version = moov[mdhd.dataOffset];
-    const timescaleOffset = version === 1 ? mdhd.dataOffset + 20 : mdhd.dataOffset + 12;
-    if (timescaleOffset + 4 > mdhd.end) {
-      continue;
-    }
-    const timescale = moov.readUInt32BE(timescaleOffset);
-    if (!timescale) {
-      continue;
-    }
-
-    const minf = findBox(moov, mdia.dataOffset, mdia.end, "minf");
-    const stbl = minf && findBox(moov, minf.dataOffset, minf.end, "stbl");
-    if (!stbl) {
-      continue;
-    }
-    const stss = findBox(moov, stbl.dataOffset, stbl.end, "stss");
-    const stts = findBox(moov, stbl.dataOffset, stbl.end, "stts");
-    if (!stss || !stts) {
-      continue;
-    }
-
-    const syncCount = moov.readUInt32BE(stss.dataOffset + 4);
-    const wanted = new Set();
-    for (let index = 0; index < syncCount; index += 1) {
-      const at = stss.dataOffset + 8 + index * 4;
-      if (at + 4 > stss.end) {
-        break;
-      }
-      wanted.add(moov.readUInt32BE(at));
-    }
-    if (wanted.size === 0) {
-      continue;
-    }
-
-    // The two terms that turn decode times into the timeline ffmpeg cuts on.
-    // Both are optional: a file without them is one whose decode and
-    // composition orders already agree, and then nothing is added or taken.
-    const ctts = findBox(moov, stbl.dataOffset, stbl.end, "ctts");
-    const offsets = ctts ? readCompositionOffsets(moov, ctts, wanted) : null;
-    const edts = findBox(moov, trak.dataOffset, trak.end, "edts");
-    const elst = edts && findBox(moov, edts.dataOffset, edts.end, "elst");
-    const editShift = elst ? readEditShift(moov, elst) : 0;
-
-    const times = resolveSampleTimes(moov, stts, timescale, wanted, offsets);
-    if (times.length > 0) {
-      // A shift applied after the division would be in the wrong units: the
-      // edit's `media_time` is in MEDIA ticks, like everything else here.
-      const shifted = editShift === 0 ? times : times.map((time) => time - editShift / timescale);
-      // A negative time is not a position in the file. It happens when an edit
-      // starts later than a keyframe the table lists, and those frames are not
-      // presented at all.
-      return shifted.filter((time) => time >= 0);
-    }
-  }
-  return null;
-}
+/**
+ * @file MP4/MOV container — ISO/IEC 14496-12.
+ *
+ * Parses moov for all track types in one walk:
+ *  - tkhd: track_ID, flags track_enabled (0x000001), alternate_group, width/height
+ *  - mdhd: timescale, language (packed 5-bit), version handling
+ *  - hdlr: handler_type (vide/soun/text/sbtl/subt/subp/clcp)
+ *  - elng: extendedLanguage BCP47 (when present, replaces mdhd language per spec)
+ *  - stsd: sample entry format (avc1/hev1/mp4a/tx3g/wvtt/stpp)
+ *  - stbl tables for subtitle cue ranges (stts/stsz/stsc/stco/co64) and video keyframes (stss/stts/ctts/elst)
+ *
+ * Keyframe reading is delegated to mp4.js. The subtitle sample table is read in
+ * this module, because every rule in it is a statement of ISO/IEC 14496-12 about
+ * this container. Track creation is centralised here so that track_enabled,
+ * `elng` and alternate_group are handled once for every media type.
+ */
+
+import { Container } from "./Container.js";
+import { VideoTrack } from "../tracks/VideoTrack.js";
+import { AudioTrack } from "../tracks/AudioTrack.js";
+import { TextSubtitleTrack, TEXT_FORMATS_MP4 } from "../tracks/TextSubtitleTrack.js";
+import { ImageSubtitleTrack } from "../tracks/ImageSubtitleTrack.js";
+
+/** A box header is eight bytes, or sixteen when the size field says 1 (§4.2). */
+const HEADER_BYTES = 8;
+const LARGE_SIZE_MARKER = 1;
+const LARGE_HEADER_BYTES = 16;
+
+/**
+ * One box header, per ISO/IEC 14496-12 §4.2.
+ *
+ * @param {Buffer} buf
+ * @param {number} off
+ * @returns {{ type: string, size: number, dataOffset: number, end: number } | null}
+ */
+function readBox(buffer, offset) {
+  if (offset + HEADER_BYTES > buffer.length) {
+    return null;
+  }
+  let size = buffer.readUInt32BE(offset);
+  const type = buffer.toString("latin1", offset + 4, offset + 8);
+  let headerBytes = HEADER_BYTES;
+  // A size of 1 means the real size is the 64-bit value after the type.
+  if (size === LARGE_SIZE_MARKER) {
+    if (offset + LARGE_HEADER_BYTES > buffer.length) {
+      return null;
+    }
+    size = Number(buffer.readBigUInt64BE(offset + HEADER_BYTES));
+    headerBytes = LARGE_HEADER_BYTES;
+  }
+  if (size < headerBytes) {
+    return null;
+  }
+  return { type, size, dataOffset: offset + headerBytes, end: offset + size };
+}
+
+/**
+ * Every direct child box of the given type.
+ *
+ * @param {Buffer} buf
+ * @param {number} start
+ * @param {number} end
+ * @param {string} type
+ * @returns {Array<{ type: string, size: number, dataOffset: number, end: number }>}
+ */
+function childrenOf(buf, start, end, type) {
+  const out = [];
+  let p = start;
+  while (p + 8 <= end) {
+    const b = readBox(buf, p);
+    if (!b) break;
+    if (b.type === type) out.push(b);
+    p = b.end;
+  }
+  return out;
+}
+
+/**
+ * The first direct child box of the given type, or null.
+ *
+ * @param {Buffer} buf
+ * @param {number} s
+ * @param {number} e
+ * @param {string} t
+ * @returns {{ type: string, size: number, dataOffset: number, end: number } | null}
+ */
+function childOf(buf, s, e, t) {
+  return childrenOf(buf, s, e, t)[0] ?? null;
+}
+
+/**
+ * The channel count, sampling frequency and decoder configuration one audio
+ * sample entry states (ISO/IEC 14496-12 §12.2.3, AudioSampleEntry).
+ *
+ * After the 8 bytes every SampleEntry begins with come 8 reserved bytes —
+ * whose first two QuickTime uses as a version that lengthens the entry by 16
+ * (version 1) or 36 (version 2) bytes — then `channelcount`, `samplesize`, two
+ * reserved fields and `samplerate` as 16.16 fixed point. Child boxes follow.
+ * For `mp4a` the child `esds` holds the AudioSpecificConfig (ISO/IEC 14496-1
+ * §7.2.6: ES_Descriptor, then DecoderConfigDescriptor, then
+ * DecoderSpecificInfo).
+ *
+ * @param {Buffer} buf
+ * @param {{ type: string, dataOffset: number, end: number }} entry
+ * @returns {{ channels: number | null, sampleRate: number | null, decoderConfig: Buffer | null } | null}
+ */
+function readAudioSampleEntry(buf, entry) {
+  const base = entry.dataOffset;
+  if (base + 28 > entry.end || base + 28 > buf.length) {
+    return null;
+  }
+  const version = buf.readUInt16BE(base + 8);
+  const channels = buf.readUInt16BE(base + 16);
+  const sampleRate = buf.readUInt32BE(base + 24) / 65536;
+  const childrenStart = base + 28 + (version === 1 ? 16 : version === 2 ? 36 : 0);
+  let decoderConfig = null;
+  if (entry.type === "mp4a" && childrenStart < entry.end) {
+    const esds = childOf(buf, childrenStart, Math.min(entry.end, buf.length), "esds");
+    if (esds) {
+      decoderConfig = decoderSpecificInfoOf(buf, esds.dataOffset + 4, Math.min(esds.end, buf.length));
+    }
+  }
+  return {
+    channels: channels > 0 ? channels : null,
+    sampleRate: sampleRate > 0 ? sampleRate : null,
+    decoderConfig
+  };
+}
+
+/**
+ * The DecoderSpecificInfo inside an `esds` payload, or null.
+ *
+ * Descriptors are a tag byte and a size of one to four bytes, seven bits each,
+ * the high bit saying another follows (ISO/IEC 14496-1 §8.3.3).
+ *
+ * @param {Buffer} buf
+ * @param {number} start
+ * @param {number} end
+ * @returns {Buffer | null}
+ */
+function decoderSpecificInfoOf(buf, start, end) {
+  const descriptorAt = (offset) => {
+    if (offset >= end) return null;
+    const tag = buf[offset];
+    let size = 0;
+    let cursor = offset + 1;
+    for (let i = 0; i < 4 && cursor < end; i += 1) {
+      const byte = buf[cursor];
+      cursor += 1;
+      size = (size * 128) + (byte & 0x7f);
+      if ((byte & 0x80) === 0) break;
+    }
+    return { tag, dataOffset: cursor, end: Math.min(end, cursor + size) };
+  };
+  const es = descriptorAt(start);
+  if (!es || es.tag !== 0x03 || es.dataOffset + 3 > es.end) return null;
+  const flags = buf[es.dataOffset + 2];
+  let cursor = es.dataOffset + 3;
+  if (flags & 0x80) cursor += 2;
+  if (flags & 0x40) cursor += 1 + (buf[cursor] ?? 0);
+  if (flags & 0x20) cursor += 2;
+  while (cursor < es.end) {
+    const descriptor = descriptorAt(cursor);
+    if (!descriptor) return null;
+    if (descriptor.tag === 0x04) {
+      // objectTypeIndication, streamType, bufferSizeDB, maxBitrate, avgBitrate.
+      let inner = descriptor.dataOffset + 13;
+      while (inner < descriptor.end) {
+        const child = descriptorAt(inner);
+        if (!child) return null;
+        if (child.tag === 0x05) {
+          return child.end > child.dataOffset ? Buffer.from(buf.subarray(child.dataOffset, child.end)) : null;
+        }
+        inner = child.end;
+      }
+      return null;
+    }
+    cursor = descriptor.end;
+  }
+  return null;
+}
+
+export class Mp4Container extends Container {
+  get formatName() {
+    return "mp4";
+  }
+
+  static detect(head) {
+    return isMp4(head);
+  }
+
+  /**
+   * The keyframe times this container's own index states, in ascending seconds.
+   *
+   * Static so a caller that has bytes and no container can ask; the instance
+   * form is {@link Container#readKeyframeIndex}.
+   *
+   * @param {(start:number,end:number)=>Promise<Buffer|null>} readRange
+   * @param {number} fileSize
+   * @returns {Promise<number[]|null>} Null where the container has no index.
+   */
+  static readKeyframeTimes(readRange, fileSize) {
+    return readMp4KeyframeTimes(readRange, fileSize);
+  }
+
+  /**
+   * This container's text subtitle tracks, with every cue's time and byte
+   * range — ISO/IEC 14496-12 §8.5 and §8.7.
+   *
+   * @param {(start:number,end:number)=>Promise<Buffer|null>} readRange
+   * @param {number} fileSize
+   * @returns {Promise<{ tracks: object[] } | null>}
+   */
+  static readSubtitlePlan(readRange, fileSize) {
+    return readMp4SubtitlePlan(readRange, fileSize);
+  }
+
+  /**
+   * The same reading, over the file this container was built on.
+   *
+   * The static form exists for a caller that has bytes and no container; this
+   * is the one to use otherwise, because the reader is already here.
+   *
+   * @returns {Promise<object|null>}
+   */
+  async readSubtitlePlan() {
+    const plan = await Mp4Container.readSubtitlePlan(this.readRange, this.fileSize);
+    if (!plan) {
+      return null;
+    }
+    // An MP4 states every sample's byte range in its own table, so a cue costs
+    // its own few dozen bytes rather than the cluster around it — which is why
+    // there are `samples` here and no `clusterPositions`.
+    //
+    // An MP4 has no element meaning "show this subtitle track by default", so
+    // `declared` is empty: the container states nothing, and nothing is shown
+    // unasked.
+    return {
+      tracks: plan.tracks.map((track, order) => ({
+        trackNumber: track.trackId,
+        declaredIndex: Number.isInteger(track.declaredIndex) ? track.declaredIndex : order,
+        codecId: track.format,
+        language: track.language,
+        name: "",
+        isDefault: order === 0,
+        codecPrivate: "",
+        clusterPositions: [],
+        samples: track.samples
+      })),
+      declared: [],
+      secondsPerTick: 0.001,
+      segmentDataOffset: 0
+    };
+  }
+
+
+  /**
+   * Read the samples of one subtitle track that are HELD now.
+   *
+   * An MP4 states every sample's byte range in its own table, so a cue costs its
+   * own few dozen bytes rather than the cluster around it — which is why this
+   * reads per sample where Matroska reads per cluster.
+   *
+   * Nothing is fetched: a sample whose bytes are not downloaded is left for the
+   * next call.
+   *
+   * @param {{ codecId: string, samples: {offset: number, size: number, startSeconds: number, endSeconds: number}[] }} track
+   * @param {Set<number>} harvested - Sample offsets already read; added to.
+   * @returns {Promise<{startSeconds: number, endSeconds: number, text: string}[]>}
+   *   The cues found in THIS pass.
+   */
+  async readHeldCues(_plan, track, progress) {
+    let harvested = progress.harvested.get(track.trackNumber);
+    if (!harvested) {
+      harvested = new Set();
+      progress.harvested.set(track.trackNumber, harvested);
+    }
+    const cues = await this.readHeldSamples(track, harvested);
+    return {
+      found: cues.length > 0 ? new Map([[track.trackNumber, cues]]) : new Map(),
+      covered: harvested.size,
+      indexed: track?.samples?.length ?? 0
+    };
+  }
+
+  async readHeldSamples(track, harvested) {
+    const found = [];
+    for (const sample of track?.samples ?? []) {
+      if (harvested.has(sample.offset)) {
+        continue;
+      }
+      const last = Math.min(this.fileSize - 1, sample.offset + sample.size - 1);
+      if (!this.isHeld(sample.offset, last)) {
+        continue;
+      }
+      const bytes = await this.readHeld(sample.offset, last);
+      if (!bytes) {
+        continue;
+      }
+      harvested.add(sample.offset);
+      // The MP4 has framed this cue and is the one that unframes it.
+      const text = Mp4Container.cueTextOf(bytes, track.codecId);
+      if (text) {
+        found.push({ startSeconds: sample.startSeconds, endSeconds: sample.endSeconds, text });
+      }
+    }
+    return found;
+  }
+
+  async readTracks() {
+    const head = await this.readRange(0, Math.min(64 - 1, this.fileSize - 1));
+    if (!head || !isMp4(head)) return [];
+
+    // Use the subtitle plan reader's moov parsing as source for subtitle tracks,
+    // and a direct moov walk for video/audio to collect tkhd/mdhd/hdlr/elng uniformly.
+    // For simplicity, delegate entirely to readMp4SubtitlePlan for subtitles and
+    // do a lightweight moov walk for video/audio here — then merge.
+    const subtitlePlan = await readMp4SubtitlePlan(this.readRange, this.fileSize).catch(() => null);
+    const subtitleByDecl = new Map();
+    if (subtitlePlan?.tracks) {
+      for (const t of subtitlePlan.tracks) subtitleByDecl.set(t.declaredIndex, t);
+    }
+
+    // Minimal moov walk for video/audio: reuse isMp4 + findMoov logic by reading via existing helper
+    // Instead of duplicating, parse tracks via a second full moov read that collects vide/soun.
+    // We read moov box directly to extract video/audio tracks.
+    const tracks = await this.#readVideoAudioTracks();
+    // Append subtitle tracks from plan, converting to domain objects
+    if (subtitlePlan?.tracks) {
+      for (const s of subtitlePlan.tracks) {
+        const isText = TEXT_FORMATS_MP4.has(s.format);
+        const Cls = isText ? TextSubtitleTrack : ImageSubtitleTrack;
+        tracks.push(new Cls({
+          trackNumber: s.trackId,
+          declaredIndex: s.declaredIndex,
+          codecId: s.format,
+          language: s.language,
+          languageBcp47: "",
+          name: "",
+          isEnabled: true,
+          isDefault: s.declaredIndex === 0,
+          declaresDefault: false,
+          codecPrivateB64: "",
+          // The sample entry's own words, read below. Either
+          // bit is enough: a file that sets only "all samples are forced" is
+          // saying what a well-formed one says twice.
+          isForced: s.someSamplesForced === true || s.allSamplesForced === true,
+          isHearingImpaired: false,
+          clusterPositions: [],
+          samples: s.samples
+        }));
+      }
+      // Count non-text subtitle handlers (subp/clcp/stpp) for declaredIndex correctness — they are already
+      // accounted for in subtitlePlan's declaredIndex via SUBTITLE_HANDLERS, but we didn't create objects for
+      // them above when they were stpp (non-text not in plan's tracks). The plan already excludes stpp from tracks
+      // but increments declaredIndex, so alignment holds: we don't need extra placeholders.
+    }
+    return tracks;
+  }
+
+  /**
+   * The `moov` box, read whole.
+   *
+   * Held on the instance because every question this class answers is inside
+   * it, and the box can be tens of megabytes off a torrent — reading it once
+   * per file is the difference between one fetch and one per question.
+   *
+   * @returns {Promise<{ moov: Buffer, header: number } | null>}
+   */
+  async #moovBuffer() {
+    if (this.moovHeld !== undefined) {
+      return this.moovHeld;
+    }
+    const PROBE = 64;
+    const MAX_MOOV = 32 * 1024 * 1024;
+    let at = 0;
+    let moovBox = null;
+    while (at < this.fileSize) {
+      const probe = await this.readRange(at, Math.min(this.fileSize - 1, at + PROBE - 1));
+      if (!probe || probe.length < 8) break;
+      let size = probe.readUInt32BE(0);
+      const type = probe.toString("latin1", 4, 8);
+      let header = 8;
+      if (size === 1) {
+        if (probe.length < 16) break;
+        size = Number(probe.readBigUInt64BE(8));
+        header = 16;
+      }
+      if (size <= 0) break;
+      if (type === "moov") { moovBox = { offset: at, size, header }; break; }
+      at += size;
+    }
+    if (!moovBox || moovBox.size > MAX_MOOV) {
+      this.moovHeld = null;
+      return null;
+    }
+    const moov = await this.readRange(moovBox.offset, Math.min(this.fileSize - 1, moovBox.offset + moovBox.size - 1));
+    this.moovHeld = moov ? { moov, header: moovBox.header } : null;
+    return this.moovHeld;
+  }
+
+  /**
+   * Duration from `mvhd` and the presentation offset from the first track's
+   * edit list, per ISO/IEC 14496-12 §8.2.2 and §8.6.6.
+   *
+   * An edit entry whose `media_time` is -1 is an EMPTY edit: it presents
+   * nothing for `segment_duration`, which shifts everything after it later by
+   * that much. That shift is what a player reports as the file's start, and it
+   * is the only way an MP4 states one — a file without such an edit begins at
+   * zero, which is a declaration, not an absence.
+   *
+   * @returns {Promise<import("./Container.js").ContainerMediaInfo>}
+   */
+  async readMediaInfo() {
+    if (this.mediaInfo) {
+      return this.mediaInfo;
+    }
+    /** @type {import("./Container.js").ContainerMediaInfo} */
+    const info = { format: this.formatName, durationSeconds: null, startTimeSeconds: null };
+    this.mediaInfo = info;
+    const held = await this.#moovBuffer();
+    if (!held) {
+      return info;
+    }
+    const { moov, header } = held;
+    const mvhd = childOf(moov, header, moov.length, "mvhd");
+    let movieTimescale = 0;
+    if (mvhd) {
+      const version = moov[mvhd.dataOffset];
+      // version 0: creation(4) modification(4) timescale(4) duration(4)
+      // version 1: creation(8) modification(8) timescale(4) duration(8)
+      const at = version === 1 ? mvhd.dataOffset + 20 : mvhd.dataOffset + 12;
+      if (at + 8 <= moov.length) {
+        movieTimescale = moov.readUInt32BE(at);
+        const duration = version === 1 ? Number(moov.readBigUInt64BE(at + 4)) : moov.readUInt32BE(at + 4);
+        if (movieTimescale > 0 && duration > 0) {
+          info.durationSeconds = duration / movieTimescale;
+        }
+      }
+    }
+    info.startTimeSeconds = movieTimescale > 0
+      ? Mp4Container.#emptyEditSeconds(moov, header, movieTimescale)
+      : null;
+    return info;
+  }
+
+  /**
+   * The presentation shift of the first empty edit, in seconds; 0 when no track
+   * declares one.
+   *
+   * @param {Buffer} moov
+   * @param {number} header
+   * @param {number} movieTimescale
+   * @returns {number}
+   */
+  static #emptyEditSeconds(moov, header, movieTimescale) {
+    let shift = 0;
+    for (const trak of childrenOf(moov, header, moov.length, "trak")) {
+      const edts = childOf(moov, trak.dataOffset, trak.end, "edts");
+      const elst = edts && childOf(moov, edts.dataOffset, edts.end, "elst");
+      if (!elst) {
+        continue;
+      }
+      const version = moov[elst.dataOffset];
+      const count = moov.readUInt32BE(elst.dataOffset + 4);
+      if (count < 1) {
+        continue;
+      }
+      const entry = elst.dataOffset + 8;
+      const segmentDuration = version === 1
+        ? Number(moov.readBigUInt64BE(entry))
+        : moov.readUInt32BE(entry);
+      const mediaTime = version === 1
+        ? Number(moov.readBigInt64BE(entry + 8))
+        : moov.readInt32BE(entry + 4);
+      if (mediaTime === -1 && segmentDuration > 0) {
+        shift = Math.max(shift, segmentDuration / movieTimescale);
+      }
+    }
+    return shift;
+  }
+
+  async #readVideoAudioTracks() {
+    const held = await this.#moovBuffer();
+    if (!held) return [];
+    const { moov, header: moovHeader } = held;
+
+    const result = [];
+    let videoIdx = -1;
+    let audioIdx = -1;
+
+    const moovContentStart = moovHeader;
+    const moovEnd = moov.length;
+    for (const trak of childrenOf(moov, moovContentStart, moovEnd, "trak")) {
+      const mdia = childOf(moov, trak.dataOffset, trak.end, "mdia");
+      if (!mdia) continue;
+      const hdlr = childOf(moov, mdia.dataOffset, mdia.end, "hdlr");
+      const handler = hdlr ? moov.toString("latin1", hdlr.dataOffset + 8, hdlr.dataOffset + 12) : "";
+      const mdhd = childOf(moov, mdia.dataOffset, mdia.end, "mdhd");
+      const tkhd = childOf(moov, trak.dataOffset, trak.end, "tkhd");
+      let language = "";
+      if (mdhd) {
+        const ver = moov[mdhd.dataOffset];
+        const langAt = ver === 1 ? mdhd.dataOffset + 32 : mdhd.dataOffset + 20;
+        if (langAt + 2 <= mdhd.end) {
+          const packed = moov.readUInt16BE(langAt);
+          language = [10, 5, 0].map((s) => String.fromCharCode(((packed >> s) & 0x1f) + 0x60)).join("").replace(/[^a-z]/g, "");
+        }
+      }
+      // elng overrides mdhd language per spec §8.4.6
+      let languageBcp47 = "";
+      const elng = childOf(moov, mdia.dataOffset, mdia.end, "elng");
+      if (elng && elng.end - elng.dataOffset >= 4) {
+        languageBcp47 = moov.toString("utf8", elng.dataOffset + 4, elng.end).replace(/\0+$/, "");
+      }
+      let trackId = 0;
+      let isEnabled = true;
+      let alternateGroup = 0;
+      let width = null;
+      let height = null;
+      if (tkhd) {
+        const ver = moov[tkhd.dataOffset];
+        const flags = moov.readUInt32BE(tkhd.dataOffset + 1) & 0xffffff; // 3 bytes after version
+        isEnabled = (flags & 0x000001) !== 0;
+        trackId = moov.readUInt32BE(ver === 1 ? tkhd.dataOffset + 20 : tkhd.dataOffset + 12);
+        alternateGroup = moov.readUInt16BE(ver === 1 ? tkhd.dataOffset + 26 : tkhd.dataOffset + 18);
+        // width/height are 16.16 fixed point at end of tkhd
+        if (tkhd.end - tkhd.dataOffset >= 84) {
+          const w = moov.readUInt32BE(ver === 1 ? tkhd.dataOffset + 76 : tkhd.dataOffset + 68);
+          const h = moov.readUInt32BE(ver === 1 ? tkhd.dataOffset + 80 : tkhd.dataOffset + 72);
+          width = w / 65536;
+          height = h / 65536;
+        }
+      }
+      const resolvedLang = languageBcp47 || language;
+      if (handler === "vide") {
+        videoIdx += 1;
+        // stsd format for codecId
+        let codecId = "";
+        const minf = childOf(moov, mdia.dataOffset, mdia.end, "minf");
+        const stbl = minf && childOf(moov, minf.dataOffset, minf.end, "stbl");
+        const stsd = stbl && childOf(moov, stbl.dataOffset, stbl.end, "stsd");
+        if (stsd) {
+          const first = readBox(moov, stsd.dataOffset + 8);
+          if (first) codecId = first.type;
+        }
+        result.push(new VideoTrack({
+          trackNumber: trackId,
+          declaredIndex: videoIdx,
+          codecId,
+          language: resolvedLang,
+          languageBcp47,
+          name: "",
+          isEnabled,
+          isDefault: true,
+          declaresDefault: false,
+          codecPrivateB64: "",
+          alternateGroup,
+          width,
+          height
+        }));
+      } else if (handler === "soun") {
+        audioIdx += 1;
+        let codecId = "";
+        let sampleEntry = null;
+        const minf = childOf(moov, mdia.dataOffset, mdia.end, "minf");
+        const stbl = minf && childOf(moov, minf.dataOffset, minf.end, "stbl");
+        const stsd = stbl && childOf(moov, stbl.dataOffset, stbl.end, "stsd");
+        if (stsd) {
+          const first = readBox(moov, stsd.dataOffset + 8);
+          if (first) {
+            codecId = first.type;
+            sampleEntry = readAudioSampleEntry(moov, first);
+          }
+        }
+        result.push(new AudioTrack({
+          trackNumber: trackId,
+          declaredIndex: audioIdx,
+          codecId,
+          language: resolvedLang,
+          languageBcp47,
+          name: "",
+          isEnabled,
+          isDefault: true,
+          declaresDefault: false,
+          // The AudioSpecificConfig an `mp4a` entry carries in its `esds`, which
+          // is what Matroska writes as CodecPrivate for the same codec.
+          codecPrivateB64: sampleEntry?.decoderConfig ? sampleEntry.decoderConfig.toString("base64") : "",
+          alternateGroup,
+          isOriginal: false,
+          isCommentary: false,
+          isVisualImpaired: false,
+          channels: sampleEntry?.channels ?? null,
+          samplingFrequency: sampleEntry?.sampleRate ?? null
+        }));
+      }
+    }
+    return result;
+  }
+
+  /**
+   * The text field of one cue as MP4 frames it.
+   *
+   * A `tx3g`/`text` sample is a 16-bit big-endian length followed by that many
+   * bytes of UTF-8 (ISO/IEC 14496-12 §12.6.3 and Apple's text sample format); a
+   * `wvtt` sample is a sequence of boxes whose `vttc`/`payl` holds the cue text
+   * (§12.6.3.2). Neither carries the subtitle format's own markup, so the
+   * markup step that follows has nothing to take off — it is applied all the
+   * same, because which step applies is decided by the codec and not here.
+   *
+   * The byte reading itself is this module's,
+   * alongside the sample-table walk that found the range.
+   *
+   * @param {Buffer} payload - The sample's own bytes.
+   * @param {string} codecId - Sample entry type: `tx3g`, `text` or `wvtt`.
+   * @returns {string}
+   */
+  static cueTextOf(payload, codecId) {
+    return decodeSubtitleSample(payload, codecId);
+  }
+
+  async parseKeyframeIndex() {
+    const r = await readMp4KeyframeTimes(this.readRange, this.fileSize);
+    if (!r) return null;
+    if (Array.isArray(r)) return { times: r, tolerance: 0 };
+    return r;
+  }
+}
+
+// ---------------------------------------------------------------------------
+// The MP4's own reading of its subtitle sample table. Here because every rule
+// in it is ISO/IEC 14496-12 speaking about this container, and the class is
+// the only way in.
+// ---------------------------------------------------------------------------
+/**
+ * @file The text subtitle tracks of an MP4, and where each cue's bytes are.
+ *
+ * The same rule as the Matroska side: nothing is extracted with ffmpeg and
+ * nothing is fetched for its own sake. Here it is cheaper still. Matroska hides
+ * its subtitle blocks inside clusters shared with the picture, so a cue costs
+ * whatever cluster holds it; an MP4 states every sample's offset and length in
+ * the sample table, so a cue costs its own bytes and nothing more — usually a
+ * few dozen of them.
+ *
+ * The tables, from ISO/IEC 14496-12:
+ *
+ *   stsd — what the samples are (`tx3g` timed text, `wvtt` WebVTT, `stpp` TTML)
+ *   stts — how long each sample lasts, run-length encoded (§8.6.1.2)
+ *   stsz — how long each sample is, in bytes (§8.7.3)
+ *   stsc — how samples are grouped into chunks (§8.7.4)
+ *   stco / co64 — where each chunk begins in the file (§8.7.5)
+ *
+ * Together they give, for sample N: when it starts, how long it stays, and the
+ * exact byte range holding it. That is everything needed to show a cue without
+ * reading anything else.
+ */
+
+const PROBE_BYTES = 64;
+const MAX_MOOV_BYTES = 32 * 1024 * 1024;
+
+/** Handlers that mean "this track is text on screen". */
+const TEXT_HANDLERS = new Set(["text", "sbtl", "subt"]);
+/**
+ * Every handler ffmpeg's mov demuxer turns into a SUBTITLE stream, whether or
+ * not this file can read it — `subp` is a DVD subpicture and `clcp` closed
+ * captions, both pictures or caption data rather than text. They are counted
+ * because `declaredIndex` has to equal ffmpeg's `0:s:N`, and a track left out
+ * of the count shifts every text track after it, which is the very defect
+ * `declaredIndex` exists to remove.
+ */
+const SUBTITLE_HANDLERS = new Set([...TEXT_HANDLERS, "subp", "clcp"]);
+/** Sample formats this can turn into cues. `stpp` (TTML) is XML and is not one. */
+const TEXT_FORMATS = new Set(["tx3g", "text", "wvtt"]);
+
+
+
+
+
+/**
+ * Sample durations, expanded from the run-length table.
+ *
+ * @param {Buffer} moov
+ * @param {{ dataOffset: number, end: number }} stts
+ * @param {number} total - How many samples the size table declares.
+ * @returns {number[]} Ticks each sample lasts.
+ */
+function sampleDurations(moov, stts, total) {
+  const durations = new Array(total).fill(0);
+  const entries = moov.readUInt32BE(stts.dataOffset + 4);
+  let at = stts.dataOffset + 8;
+  let sample = 0;
+  for (let entry = 0; entry < entries && at + 8 <= stts.end && sample < total; entry += 1, at += 8) {
+    const count = moov.readUInt32BE(at);
+    const delta = moov.readUInt32BE(at + 4);
+    for (let index = 0; index < count && sample < total; index += 1, sample += 1) {
+      durations[sample] = delta;
+    }
+  }
+  return durations;
+}
+
+/**
+ * Sample sizes, whether the table states one for all or one for each.
+ *
+ * @param {Buffer} moov
+ * @param {{ dataOffset: number, end: number }} stsz
+ * @returns {number[]}
+ */
+function sampleSizes(moov, stsz) {
+  const uniform = moov.readUInt32BE(stsz.dataOffset + 4);
+  const count = moov.readUInt32BE(stsz.dataOffset + 8);
+  if (uniform > 0) {
+    return new Array(count).fill(uniform);
+  }
+  const sizes = new Array(count).fill(0);
+  let at = stsz.dataOffset + 12;
+  for (let index = 0; index < count && at + 4 <= stsz.end; index += 1, at += 4) {
+    sizes[index] = moov.readUInt32BE(at);
+  }
+  return sizes;
+}
+
+/**
+ * Where every sample of a track begins in the file.
+ *
+ * The sample-to-chunk table says how many samples each run of chunks holds, and
+ * the chunk-offset table says where each chunk starts; a sample's own offset is
+ * its chunk's start plus the sizes of the samples before it in that chunk.
+ *
+ * @param {Buffer} moov
+ * @param {{ dataOffset: number, end: number }} stsc
+ * @param {number[]} chunkOffsets
+ * @param {number[]} sizes
+ * @returns {number[]}
+ */
+function sampleOffsets(moov, stsc, chunkOffsets, sizes) {
+  const offsets = new Array(sizes.length).fill(0);
+  const entries = moov.readUInt32BE(stsc.dataOffset + 4);
+  /** @type {{ firstChunk: number, perChunk: number }[]} */
+  const runs = [];
+  let at = stsc.dataOffset + 8;
+  for (let entry = 0; entry < entries && at + 12 <= stsc.end; entry += 1, at += 12) {
+    runs.push({ firstChunk: moov.readUInt32BE(at), perChunk: moov.readUInt32BE(at + 4) });
+  }
+  let sample = 0;
+  for (let run = 0; run < runs.length && sample < sizes.length; run += 1) {
+    const from = runs[run].firstChunk;
+    const to = run + 1 < runs.length ? runs[run + 1].firstChunk - 1 : chunkOffsets.length;
+    for (let chunk = from; chunk <= to && sample < sizes.length; chunk += 1) {
+      let inChunk = chunkOffsets[chunk - 1];
+      if (inChunk === undefined) {
+        break;
+      }
+      for (let index = 0; index < runs[run].perChunk && sample < sizes.length; index += 1, sample += 1) {
+        offsets[sample] = inChunk;
+        inChunk += sizes[sample];
+      }
+    }
+  }
+  return offsets;
+}
+
+/**
+ * @typedef {object} Mp4SubtitleSample
+ * @property {number} startSeconds
+ * @property {number} endSeconds
+ * @property {number} offset - Where the sample's bytes are in the file.
+ * @property {number} size
+ */
+
+/**
+ * @typedef {object} Mp4SubtitleTrack
+ * @property {number} trackId
+ * @property {number} declaredIndex - Its position among ALL of the file's
+ *   subtitle tracks, including the ones whose sample format this cannot turn
+ *   into cues (`stpp` TTML). That is the number ffmpeg gives the same stream in
+ *   `0:s:N`, which is the number the browser names; counting only the readable
+ *   ones would shift every track after a TTML one.
+ * @property {string} format - `tx3g`, `text` or `wvtt`.
+ * @property {string} language - Three letters, as the file declares them.
+ * @property {boolean} someSamplesForced - The sample entry says at least one
+ *   cue carries a forced (`frcd`) atom.
+ * @property {boolean} allSamplesForced - The sample entry says every cue is to
+ *   be treated as forced, whether or not it carries that atom.
+ * @property {Mp4SubtitleSample[]} samples - In time order.
+ */
+
+/**
+ * The text subtitle tracks of an MP4, with every cue's time and byte range.
+ *
+ * @param {(start: number, end: number) => Promise<Buffer | null>} readRange
+ * @param {number} fileSize
+ * @returns {Promise<{ tracks: Mp4SubtitleTrack[] } | null>}
+ */
+async function readMp4SubtitlePlan(readRange, fileSize) {
+  const found = await findMoov(readRange, fileSize);
+  if (!found || found.size > MAX_MOOV_BYTES) {
+    return null;
+  }
+  const moov = await readRange(found.offset, Math.min(fileSize - 1, found.offset + found.size - 1));
+  if (!moov || moov.length < HEADER_BYTES) {
+    return null;
+  }
+  const moovBox = readBox(moov, 0);
+  if (!moovBox) {
+    return null;
+  }
+
+  /** @type {Mp4SubtitleTrack[]} */
+  const tracks = [];
+  // Counts every subtitle track the file has, whether or not this can read it,
+  // so the number handed out matches ffmpeg's `0:s:N`. See `declaredIndex`.
+  let declaredIndex = -1;
+  for (const trak of childrenOf(moov, moovBox.dataOffset, moov.length, "trak")) {
+    const mdia = childOf(moov, trak.dataOffset, trak.end, "mdia");
+    if (!mdia) {
+      continue;
+    }
+    const hdlr = childOf(moov, mdia.dataOffset, mdia.end, "hdlr");
+    const handler = hdlr ? moov.toString("latin1", hdlr.dataOffset + 8, hdlr.dataOffset + 12) : "";
+    if (!SUBTITLE_HANDLERS.has(handler)) {
+      continue;
+    }
+    // Counted before the readability checks below, and before the handler is
+    // narrowed to the text ones: this number is the track's place in the file,
+    // not its place among the tracks this code can turn into cues.
+    declaredIndex += 1;
+    if (!TEXT_HANDLERS.has(handler)) {
+      continue;
+    }
+    const mdhd = childOf(moov, mdia.dataOffset, mdia.end, "mdhd");
+    if (!mdhd) {
+      continue;
+    }
+    const version = moov[mdhd.dataOffset];
+    const timescale = moov.readUInt32BE(version === 1 ? mdhd.dataOffset + 20 : mdhd.dataOffset + 12);
+    if (!timescale) {
+      continue;
+    }
+    // The language is five bits per letter, offset from 0x60, packed into two
+    // bytes after the times (ISO/IEC 14496-12 §8.4.2.3).
+    const languageAt = version === 1 ? mdhd.dataOffset + 32 : mdhd.dataOffset + 20;
+    let language = "";
+    if (languageAt + 2 <= mdhd.end) {
+      const packed = moov.readUInt16BE(languageAt);
+      language = [10, 5, 0]
+        .map((shift) => String.fromCharCode(((packed >> shift) & 0x1f) + 0x60))
+        .join("")
+        .replace(/[^a-z]/g, "");
+    }
+
+    const tkhd = childOf(moov, trak.dataOffset, trak.end, "tkhd");
+    const trackId = tkhd
+      ? moov.readUInt32BE(moov[tkhd.dataOffset] === 1 ? tkhd.dataOffset + 20 : tkhd.dataOffset + 12)
+      : tracks.length + 1;
+
+    const minf = childOf(moov, mdia.dataOffset, mdia.end, "minf");
+    const stbl = minf && childOf(moov, minf.dataOffset, minf.end, "stbl");
+    if (!stbl) {
+      continue;
+    }
+    const stsd = childOf(moov, stbl.dataOffset, stbl.end, "stsd");
+    const first = stsd && readBox(moov, stsd.dataOffset + 8);
+    const format = first ? first.type : "";
+    if (!TEXT_FORMATS.has(format)) {
+      continue;
+    }
+    // Whether the file itself says this track is forced — subtitles shown even
+    // to a viewer who did not ask for subtitles, because the dialogue on screen
+    // is in a language the soundtrack is not.
+    //
+    // Apple's QuickTime File Format, "Display flags" under Subtitle sample
+    // description, defines the two bits read here: `0x40000000` "Some samples
+    // are forced" ("at least one sample contains a forced (`frcd`) atom") and
+    // `0x80000000` "All samples are forced" ("the subtitle media handler treats
+    // all samples as forced subtitles, regardless of the presence or absence of
+    // a `frcd` atom"), with the note that setting the second requires the first
+    // — the pair together being `0xC0000000`. We honour that requirement rather
+    // than trusting a writer to have met it: either bit alone is enough to call
+    // the track forced, because a file that sets only `0x80000000` is saying
+    // exactly what a well-formed one would say twice.
+    //
+    // The field sits at a fixed place in the sample entry. `dataOffset` is
+    // already past the box header, and every sample entry opens with 6 reserved
+    // bytes and a 2-byte data reference index (ISO/IEC 14496-12 §8.5.2.2), so
+    // `displayFlags` is the 32 bits eight bytes in.
+    let someSamplesForced = false;
+    let allSamplesForced = false;
+    if (format === "tx3g" && first && first.dataOffset + 8 + 4 <= first.end) {
+      const displayFlags = moov.readUInt32BE(first.dataOffset + 8);
+      someSamplesForced = (displayFlags & 0x40000000) !== 0;
+      allSamplesForced = (displayFlags & 0x80000000) !== 0;
+    }
+    const stts = childOf(moov, stbl.dataOffset, stbl.end, "stts");
+    const stsz = childOf(moov, stbl.dataOffset, stbl.end, "stsz");
+    const stsc = childOf(moov, stbl.dataOffset, stbl.end, "stsc");
+    const stco = childOf(moov, stbl.dataOffset, stbl.end, "stco");
+    const co64 = childOf(moov, stbl.dataOffset, stbl.end, "co64");
+    if (!stts || !stsz || !stsc || (!stco && !co64)) {
+      continue;
+    }
+
+    const sizes = sampleSizes(moov, stsz);
+    const durations = sampleDurations(moov, stts, sizes.length);
+    const chunkOffsets = [];
+    if (stco) {
+      const count = moov.readUInt32BE(stco.dataOffset + 4);
+      let at = stco.dataOffset + 8;
+      for (let index = 0; index < count && at + 4 <= stco.end; index += 1, at += 4) {
+        chunkOffsets.push(moov.readUInt32BE(at));
+      }
+    } else {
+      const count = moov.readUInt32BE(co64.dataOffset + 4);
+      let at = co64.dataOffset + 8;
+      for (let index = 0; index < count && at + 8 <= co64.end; index += 1, at += 8) {
+        chunkOffsets.push(Number(moov.readBigUInt64BE(at)));
+      }
+    }
+    const offsets = sampleOffsets(moov, stsc, chunkOffsets, sizes);
+
+    /** @type {Mp4SubtitleSample[]} */
+    const samples = [];
+    let ticks = 0;
+    for (let index = 0; index < sizes.length; index += 1) {
+      const start = ticks / timescale;
+      ticks += durations[index];
+      // An empty sample is a gap between cues, which the format uses to say
+      // "nothing on screen"; it is not a cue and would show as a blank line.
+      if (sizes[index] > 2) {
+        samples.push({
+          startSeconds: start,
+          endSeconds: ticks / timescale,
+          offset: offsets[index],
+          size: sizes[index]
+        });
+      }
+    }
+    tracks.push({
+      trackId,
+      declaredIndex,
+      format,
+      language,
+      someSamplesForced,
+      allSamplesForced,
+      samples
+    });
+  }
+  return { tracks };
+}
+
+/**
+ * The text of one sample.
+ *
+ * `tx3g` is a two-byte length followed by UTF-8; anything after that is styling
+ * boxes, which this deliberately drops. `wvtt` is a sequence of boxes, and the
+ * text lives in the `payl` inside a `vttc`.
+ *
+ * @param {Buffer} bytes
+ * @param {string} format
+ * @returns {string}
+ */
+function decodeSubtitleSample(bytes, format) {
+  if (format === "wvtt") {
+    let at = 0;
+    const parts = [];
+    while (at + HEADER_BYTES <= bytes.length) {
+      const box = readBox(bytes, at);
+      if (!box) {
+        break;
+      }
+      if (box.type === "vttc") {
+        const payl = childOf(bytes, box.dataOffset, Math.min(bytes.length, box.end), "payl");
+        if (payl) {
+          parts.push(bytes.toString("utf8", payl.dataOffset, Math.min(bytes.length, payl.end)));
+        }
+      }
+      at = box.end;
+    }
+    return parts.join("\n").trim();
+  }
+  if (bytes.length < 2) {
+    return "";
+  }
+  const length = bytes.readUInt16BE(0);
+  return bytes.toString("utf8", 2, Math.min(bytes.length, 2 + length)).trim();
+}
+
+// ---------------------------------------------------------------------------
+// ISO/IEC 14496-12 speaking about MP4: stss, stts, ctts and the edit list.
+// Here because the class is the only way in.
+// ---------------------------------------------------------------------------
+/**
+ * @file Keyframe index for MP4/MOV, read without downloading the file.
+ *
+ * MP4 keeps its tables in a `moov` box: `stss` lists which samples are sync
+ * samples (keyframes) by number, and `stts` gives each sample's duration, so
+ * the two together turn "sample #N" into "second T". `moov` sits either at the
+ * start (files written for streaming) or at the end (the common case for a
+ * plain mux); box headers state their own size, so it is found by stepping over
+ * top-level boxes rather than scanning bytes — a couple of 64-byte reads even
+ * when `mdat` is gigabytes.
+ *
+ * Same purpose as the Matroska reader: on the video-COPY path the segment
+ * boundaries ARE the source's keyframes, and inventing an even grid instead
+ * makes players walk the whole file or present audio with no picture.
+ */
+
+// A 64-bit box size is signalled by a 32-bit size of 1, the real size following
+// in the next 8 bytes.
+// Enough to read any box header while walking the top level.
+// Cap on the moov read. A feature-length file indexes to a few hundred KB;
+// beyond this is not a normal index and not worth pulling over a torrent.
+
+/**
+ * Whether this looks like MP4/MOV — every real file opens with an `ftyp` box.
+ *
+ * @param {Buffer} head
+ * @returns {boolean}
+ */
+function isMp4(head) {
+  return head.length >= 12 && head.toString("latin1", 4, 8) === "ftyp";
+}
+
+/**
+ * Read a box header at `offset`.
+ *
+ * @param {Buffer} buffer
+ * @param {number} offset
+ * @returns {{ type: string, size: number, headerBytes: number } | null}
+ */
+function readBoxHeader(buffer, offset) {
+  if (offset + HEADER_BYTES > buffer.length) {
+    return null;
+  }
+  const size32 = buffer.readUInt32BE(offset);
+  const type = buffer.toString("latin1", offset + 4, offset + 8);
+  if (size32 === LARGE_SIZE_MARKER) {
+    if (offset + LARGE_HEADER_BYTES > buffer.length) {
+      return null;
+    }
+    // High word is zero for any file we can practically handle.
+    const high = buffer.readUInt32BE(offset + 8);
+    const low = buffer.readUInt32BE(offset + 12);
+    return { type, size: high * 4294967296 + low, headerBytes: LARGE_HEADER_BYTES };
+  }
+  return { type, size: size32, headerBytes: HEADER_BYTES };
+}
+
+/**
+ * Walk the top-level boxes to find `moov`, reading only each box header.
+ *
+ * @param {(start: number, end: number) => Promise<Buffer | null>} readRange
+ * @param {number} fileSize
+ * @returns {Promise<{ offset: number, size: number, headerBytes: number } | null>}
+ */
+async function findMoov(readRange, fileSize) {
+  let offset = 0;
+  while (offset < fileSize) {
+    const probe = await readRange(offset, Math.min(fileSize - 1, offset + PROBE_BYTES - 1));
+    if (!probe || probe.length < HEADER_BYTES) {
+      return null;
+    }
+    const header = readBoxHeader(probe, 0);
+    // Size 0 means "extends to end of file" — legal only for the last box, and
+    // never for one we would step over.
+    if (!header || header.size <= 0) {
+      return null;
+    }
+    if (header.type === "moov") {
+      return { offset, size: header.size, headerBytes: header.headerBytes };
+    }
+    offset += header.size;
+  }
+  return null;
+}
+
+/**
+ * Find the first box of `type` directly inside a range of an already-read buffer.
+ *
+ * @param {Buffer} buffer
+ * @param {number} start
+ * @param {number} end
+ * @param {string} type
+ * @returns {{ dataOffset: number, end: number } | null}
+ */
+function findBox(buffer, start, end, type) {
+  let offset = start;
+  while (offset + HEADER_BYTES <= end) {
+    const header = readBoxHeader(buffer, offset);
+    if (!header || header.size <= 0) {
+      return null;
+    }
+    if (header.type === type) {
+      return { dataOffset: offset + header.headerBytes, end: Math.min(end, offset + header.size) };
+    }
+    offset += header.size;
+  }
+  return null;
+}
+
+/**
+ * All boxes of `type` directly inside a range.
+ *
+ * @param {Buffer} buffer
+ * @param {number} start
+ * @param {number} end
+ * @param {string} type
+ * @returns {{ dataOffset: number, end: number }[]}
+ */
+function findAllBoxes(buffer, start, end, type) {
+  const found = [];
+  let offset = start;
+  while (offset + HEADER_BYTES <= end) {
+    const header = readBoxHeader(buffer, offset);
+    if (!header || header.size <= 0) {
+      break;
+    }
+    if (header.type === type) {
+      found.push({ dataOffset: offset + header.headerBytes, end: Math.min(end, offset + header.size) });
+    }
+    offset += header.size;
+  }
+  return found;
+}
+
+/**
+ * Turn sample numbers into seconds using the time-to-sample table.
+ *
+ * `stts` is run-length encoded — pairs of (sample count, per-sample duration) —
+ * so one walk yields every sample's start time without expanding the table.
+ *
+ * @param {Buffer} buffer
+ * @param {{ dataOffset: number, end: number }} stts
+ * @param {number} timescale - Ticks per second.
+ * @param {Set<number>} wanted - Sample numbers (1-based).
+ * @returns {number[]} Seconds, ascending.
+ */
+function resolveSampleTimes(buffer, stts, timescale, wanted, offsets = null) {
+  const entryCount = buffer.readUInt32BE(stts.dataOffset + 4);
+  const times = [];
+  let sampleNumber = 1;
+  let ticks = 0;
+  let cursor = stts.dataOffset + 8;
+  for (let entry = 0; entry < entryCount && cursor + 8 <= stts.end; entry += 1) {
+    const count = buffer.readUInt32BE(cursor);
+    const delta = buffer.readUInt32BE(cursor + 4);
+    for (let index = 0; index < count; index += 1) {
+      if (wanted.has(sampleNumber)) {
+        // `CT(n) = DT(n) + CTTS(n)` — ISO/IEC 14496-12 §8.6.1.3. The offset is
+        // what turns decode order into the order frames are shown in, and it is
+        // the timeline ffmpeg cuts on.
+        times.push((ticks + (offsets?.get(sampleNumber) ?? 0)) / timescale);
+      }
+      ticks += delta;
+      sampleNumber += 1;
+    }
+    cursor += 8;
+  }
+  return times;
+}
+
+/**
+ * Composition offsets for the sample numbers asked for.
+ *
+ * `ctts` is run-length encoded like `stts`, and version 1 carries SIGNED
+ * offsets — which is what the version exists for: a frame may be shown before
+ * it is decoded. Reading them as unsigned turns a small negative offset into
+ * roughly four billion ticks.
+ *
+ * @param {Buffer} buffer
+ * @param {{ dataOffset: number, end: number }} ctts
+ * @param {Set<number>} wanted - Sample numbers (1-based).
+ * @returns {Map<number, number>} Sample number to offset in media ticks.
+ */
+function readCompositionOffsets(buffer, ctts, wanted) {
+  const version = buffer[ctts.dataOffset];
+  const entryCount = buffer.readUInt32BE(ctts.dataOffset + 4);
+  const offsets = new Map();
+  let sampleNumber = 1;
+  let cursor = ctts.dataOffset + 8;
+  for (let entry = 0; entry < entryCount && cursor + 8 <= ctts.end; entry += 1) {
+    const count = buffer.readUInt32BE(cursor);
+    const offset = version === 1 ? buffer.readInt32BE(cursor + 4) : buffer.readUInt32BE(cursor + 4);
+    for (let index = 0; index < count; index += 1) {
+      if (wanted.has(sampleNumber)) {
+        offsets.set(sampleNumber, offset);
+      }
+      sampleNumber += 1;
+    }
+    cursor += 8;
+  }
+  return offsets;
+}
+
+/**
+ * How far the edit list shifts this track's composition timeline, in media
+ * ticks.
+ *
+ * ISO/IEC 14496-12 §8.6.6.3: `media_time` is the start of the edit within the
+ * media, in the MEDIA timescale and in composition time, while
+ * `segment_duration` is in the MOVIE timescale — two different units in one
+ * structure, which is why only the first is read here. `media_time = -1` is an
+ * empty edit: it inserts blank presentation time and starts no media, so the
+ * first real edit is the one that matters.
+ *
+ * Measured 2026-08-19: every LostFilm MP4 that carries a composition offset
+ * also carries an edit list cancelling it exactly, which is why decode times
+ * have been right on those files. `Firefly.S01E03` has the offset and NO edit
+ * list, and its times were 62.1 ms early on all 34 keyframes checked.
+ *
+ * @param {Buffer} buffer
+ * @param {{ dataOffset: number, end: number }} elst
+ * @returns {number} Ticks to subtract; zero when nothing is shifted.
+ */
+function readEditShift(buffer, elst) {
+  const version = buffer[elst.dataOffset];
+  const entryCount = buffer.readUInt32BE(elst.dataOffset + 4);
+  const wide = version === 1;
+  const entryBytes = wide ? 20 : 12;
+  let cursor = elst.dataOffset + 8;
+  for (let entry = 0; entry < entryCount && cursor + entryBytes <= elst.end; entry += 1) {
+    const mediaTime = wide
+      ? Number(buffer.readBigInt64BE(cursor + 8))
+      : buffer.readInt32BE(cursor + 4);
+    if (mediaTime >= 0) {
+      return mediaTime;
+    }
+    cursor += entryBytes;
+  }
+  return 0;
+}
+
+/**
+ * Whether this track's handler says it carries video.
+ *
+ * The standard identifies a track by its `hdlr`, and nothing else does. Picking
+ * "the first track that happens to carry sync samples" worked only because the
+ * seven releases measured all put video first; a file whose audio track carries
+ * them, or one that leads with a cover-art video track, would be read from the
+ * wrong place. That is the same defect that was fixed in the Matroska reader on
+ * 2026-08-18, arrived at from the other side.
+ *
+ * @param {Buffer} buffer
+ * @param {{ dataOffset: number, end: number }} mdia
+ * @returns {boolean}
+ */
+function isVideoTrack(buffer, mdia) {
+  const hdlr = findBox(buffer, mdia.dataOffset, mdia.end, "hdlr");
+  if (!hdlr || hdlr.dataOffset + 12 > hdlr.end) {
+    return false;
+  }
+  // FullBox header (4) then a reserved pre_defined (4), then the handler type.
+  return buffer.toString("latin1", hdlr.dataOffset + 8, hdlr.dataOffset + 12) === "vide";
+}
+
+/**
+ * Read the keyframe times of an MP4/MOV file.
+ *
+ * @param {(start: number, end: number) => Promise<Buffer | null>} readRange
+ * @param {number} fileSize
+ * @returns {Promise<number[] | null>} Ascending seconds, or null when the file
+ *   carries no usable index (fragmented MP4, truncated or damaged `moov`).
+ */
+async function readMp4KeyframeTimes(readRange, fileSize) {
+  const moovBox = await findMoov(readRange, fileSize);
+  if (!moovBox || moovBox.size > MAX_MOOV_BYTES) {
+    return null;
+  }
+
+  const moov = await readRange(moovBox.offset, Math.min(fileSize - 1, moovBox.offset + moovBox.size - 1));
+  if (!moov || moov.length < moovBox.headerBytes) {
+    return null;
+  }
+
+  // Examine every track, and take the one whose HANDLER says it is video. A
+  // track with no `stss` has every sample a keyframe, so it constrains nothing
+  // and is skipped even when it is the video one.
+  for (const trak of findAllBoxes(moov, moovBox.headerBytes, moov.length, "trak")) {
+    const mdia = findBox(moov, trak.dataOffset, trak.end, "mdia");
+    if (!mdia || !isVideoTrack(moov, mdia)) {
+      continue;
+    }
+    const mdhd = findBox(moov, mdia.dataOffset, mdia.end, "mdhd");
+    if (!mdhd) {
+      continue;
+    }
+    // mdhd layout: version(1) + flags(3), then creation/modification times —
+    // 32-bit each in version 0, 64-bit in version 1 — then the timescale.
+    const version = moov[mdhd.dataOffset];
+    const timescaleOffset = version === 1 ? mdhd.dataOffset + 20 : mdhd.dataOffset + 12;
+    if (timescaleOffset + 4 > mdhd.end) {
+      continue;
+    }
+    const timescale = moov.readUInt32BE(timescaleOffset);
+    if (!timescale) {
+      continue;
+    }
+
+    const minf = findBox(moov, mdia.dataOffset, mdia.end, "minf");
+    const stbl = minf && findBox(moov, minf.dataOffset, minf.end, "stbl");
+    if (!stbl) {
+      continue;
+    }
+    const stss = findBox(moov, stbl.dataOffset, stbl.end, "stss");
+    const stts = findBox(moov, stbl.dataOffset, stbl.end, "stts");
+    if (!stss || !stts) {
+      continue;
+    }
+
+    const syncCount = moov.readUInt32BE(stss.dataOffset + 4);
+    const wanted = new Set();
+    for (let index = 0; index < syncCount; index += 1) {
+      const at = stss.dataOffset + 8 + index * 4;
+      if (at + 4 > stss.end) {
+        break;
+      }
+      wanted.add(moov.readUInt32BE(at));
+    }
+    if (wanted.size === 0) {
+      continue;
+    }
+
+    // The two terms that turn decode times into the timeline ffmpeg cuts on.
+    // Both are optional: a file without them is one whose decode and
+    // composition orders already agree, and then nothing is added or taken.
+    const ctts = findBox(moov, stbl.dataOffset, stbl.end, "ctts");
+    const offsets = ctts ? readCompositionOffsets(moov, ctts, wanted) : null;
+    const edts = findBox(moov, trak.dataOffset, trak.end, "edts");
+    const elst = edts && findBox(moov, edts.dataOffset, edts.end, "elst");
+    const editShift = elst ? readEditShift(moov, elst) : 0;
+
+    const times = resolveSampleTimes(moov, stts, timescale, wanted, offsets);
+    if (times.length > 0) {
+      // A shift applied after the division would be in the wrong units: the
+      // edit's `media_time` is in MEDIA ticks, like everything else here.
+      const shifted = editShift === 0 ? times : times.map((time) => time - editShift / timescale);
+      // A negative time is not a position in the file. It happens when an edit
+      // starts later than a keyframe the table lists, and those frames are not
+      // presented at all.
+      return shifted.filter((time) => time >= 0);
+    }
+  }
+  return null;
+}