@torrent-tv/proxy 2.89.7 → 2.89.9

package/services/media/container/matroska-clusters.js

@@ -1,1165 +1,1165 @@
-/**
- * @file Where a Matroska file's clusters are, found from the bytes already
- * downloaded, and the subtitle blocks inside them — RFC 9559 §5.1.3.
- *
- * **Why the Cues table is not enough.** It used to be the only way in: the walk
- * read the clusters the Cues table names for a subtitle track and nothing else.
- * RFC 9559 §22.1 says each subtitle frame SHOULD be referenced there — a
- * recommendation, not a guarantee — and a file may have no Cues at all. Such a
- * file showed no subtitles, ever, while a comment promised the walk would read
- * "the clusters the viewer's own playback brings in" instead. Nothing did.
- *
- * **Where a cluster begins is established three ways, and only the first two
- * are proof:**
- *
- * 1. the file says so — the first cluster after the head, and every
- *    `CueClusterPosition` of every track (the video's name nearly every
- *    cluster, because §22.1 has each keyframe referenced);
- * 2. a chain from a position already established: a level-1 element begins
- *    exactly where the one before it ends, so from a cluster whose header is
- *    downloaded the next one is known without reading anything else;
- * 3. a search of downloaded bytes for the Cluster id, used only where the file
- *    has no Cues. What it finds is a CANDIDATE, checked against the structure a
- *    cluster must have before anything is read out of it, and taken back —
- *    together with every cue it gave — if a position established by the first
- *    two ways later lands inside it.
- *
- * **What is kept between passes, and what is not.** A cluster read whole is
- * DONE and never read again. One whose bytes are not all downloaded is left as
- * it is: nothing about it is recorded, because "not here yet" is not a fact
- * about the cluster (`unavailable.js`). One whose downloaded, hash-verified
- * bytes do not parse is UNREADABLE, with its reason, and not tried again —
- * reading the same bytes again gives the same answer.
- *
- * **Order, and when a pass ends.** The cluster a viewer is standing in is read
- * first, then the one before it — a line on screen now may have started there.
- * A pass that had such clusters to read ends with them, so what is on screen is
- * pushed after two clusters rather than after the whole file; it says `more`,
- * and the caller runs the next pass at once. A pass with nothing of that kind
- * reads onward from the viewers, one viewer's next cluster after another's, and
- * then the rest — and stops, saying `more`, the moment a viewer is somewhere
- * else, so a seek is answered by the next pass and not after the whole film.
- */
-
-import { crc32 } from "node:zlib";
-import { readUint, readVint } from "./ebml-reader.js";
-import { ElementReader, MAX_HEADER_BYTES } from "./ebml-stream.js";
-import { isUnavailable } from "./unavailable.js";
-
-const ID_CLUSTER = 0x1f43b675;
-const CLUSTER_ID_BYTES = Buffer.from([0x1f, 0x43, 0xb6, 0x75]);
-
-/** The top-level elements of a Segment (RFC 9559 §5.1), Void and CRC-32 aside. */
-const LEVEL_ONE = new Set([
-  ID_CLUSTER,
-  0x1c53bb6b, // Cues
-  0x114d9b74, // SeekHead
-  0x1549a966, // Info
-  0x1654ae6b, // Tracks
-  0x1043a770, // Chapters
-  0x1941a469, // Attachments
-  0x1254c367 // Tags
-]);
-
-const ID_CRC32 = 0xbf;
-const ID_VOID = 0xec;
-const ID_TIMESTAMP = 0xe7;
-const ID_SILENT_TRACKS = 0x5854;
-const ID_POSITION = 0xa7;
-const ID_PREV_SIZE = 0xab;
-const ID_SIMPLE_BLOCK = 0xa3;
-const ID_BLOCK_GROUP = 0xa0;
-const ID_ENCRYPTED_BLOCK = 0xaf;
-const ID_BLOCK = 0xa1;
-const ID_BLOCK_DURATION = 0x9b;
-
-/** The children RFC 9559 §5.1.3 allows inside a Cluster, the global two included. */
-const CLUSTER_CHILDREN = new Set([
-  ID_CRC32,
-  ID_VOID,
-  ID_TIMESTAMP,
-  ID_SILENT_TRACKS,
-  ID_POSITION,
-  ID_PREV_SIZE,
-  ID_SIMPLE_BLOCK,
-  ID_BLOCK_GROUP,
-  ID_ENCRYPTED_BLOCK
-]);
-
-/**
- * Enough of a cluster's start to hold its header, a CRC-32 and the Timestamp
- * that follows it (§5.1.3.1: first, or second after a CRC-32).
- */
-const CLUSTER_START_PROBE = 64;
-
-/**
- * The bytes one read of a cluster's structure takes: an element header and the
- * start of a block's data, which names the block's track — a track number is a
- * vint of at most 8 bytes (RFC 9559 §10.1). Reading a cluster this way moves
- * the headers of its video and audio blocks across the thread boundary and
- * none of their frames.
- */
-const STRUCTURE_READ_BYTES = MAX_HEADER_BYTES + 8;
-
-/** Bits 1-2 of a block's flags byte say how it is laced, or that it is not. */
-const LACING_MASK = 0x06;
-const LACING_NONE = 0x00;
-const LACING_XIPH = 0x02;
-const LACING_FIXED = 0x04;
-const LACING_EBML = 0x06;
-
-/**
- * A block's own header: its track, its time relative to the cluster, its flags.
- *
- * @param {Buffer} buffer
- * @param {number} start - First byte of the block's data.
- * @param {number} end - One past its last byte.
- * @returns {{ trackNumber: number, relativeTicks: number, flags: number, dataOffset: number } | null}
- */
-export function readBlockHeader(buffer, start, end) {
-  const track = readVint(buffer, start, false);
-  if (!track || track.value === null) {
-    return null;
-  }
-  const timestampAt = start + track.length;
-  // Signed, and it can be negative: a block may belong slightly before the
-  // cluster it is stored in.
-  if (timestampAt + 3 > end) {
-    return null;
-  }
-  return {
-    trackNumber: Number(track.value),
-    relativeTicks: buffer.readInt16BE(timestampAt),
-    flags: buffer[timestampAt + 2],
-    dataOffset: timestampAt + 3
-  };
-}
-
-/**
- * Where a laced block's first frame begins.
- *
- * @param {Buffer} buffer
- * @param {number} dataOffset - First byte after the block header.
- * @param {number} end
- * @param {number} flags
- * @returns {number | null}
- */
-export function firstFrameOffset(buffer, dataOffset, end, flags) {
-  const lacing = flags & LACING_MASK;
-  if (lacing === LACING_NONE) {
-    return dataOffset;
-  }
-  if (dataOffset >= end) {
-    return null;
-  }
-  const frames = buffer[dataOffset] + 1;
-  let at = dataOffset + 1;
-  if (lacing === LACING_FIXED) {
-    return at;
-  }
-  if (lacing === LACING_XIPH) {
-    for (let frame = 0; frame < frames - 1; frame += 1) {
-      while (at < end && buffer[at] === 0xff) {
-        at += 1;
-      }
-      at += 1;
-    }
-    return at <= end ? at : null;
-  }
-  if (lacing === LACING_EBML) {
-    for (let frame = 0; frame < frames - 1; frame += 1) {
-      const size = readVint(buffer, at, false);
-      if (!size) {
-        return null;
-      }
-      at += size.length;
-    }
-    return at <= end ? at : null;
-  }
-  return null;
-}
-
-/**
- * The walk's own record of one file, kept by the caller between passes.
- *
- * @typedef {object} ClusterStart
- * @property {"first" | "cues" | "chain" | "search"} source
- * @property {number | null} parent - The start this one was chained from.
- * @property {number | null} seconds - The cluster's Timestamp, once read.
- * @property {"cluster" | "other" | null} kind - Null until its header is read.
- * @property {number | null} end - One past its last byte, null while unknown.
- * @property {boolean} stepped - Whether the next start has been taken from it.
- */
-
-/**
- * @param {object} progress
- * @returns {object}
- */
-function stateOf(progress) {
-  if (!progress.matroska) {
-    progress.matroska = {
-      seeded: false,
-      /** @type {Map<number, ClusterStart>} */
-      starts: new Map(),
-      done: new Set(),
-      /** @type {Map<number, string>} */
-      unreadable: new Map(),
-      /** Candidates whose bytes are not all here yet. @type {Set<number>} */
-      candidates: new Set(),
-      /** Candidates the structure check refused. @type {Set<number>} */
-      rejected: new Set(),
-      /** Inclusive intervals the search has finished with. @type {Array<[number, number]>} */
-      scanned: [],
-      stats: { refusedElements: 0, rejectedCandidates: 0, contradictions: 0, bytesWalked: 0, fromSearch: 0 }
-    };
-  }
-  return progress.matroska;
-}
-
-/**
- * The start a chain of starts began from.
- *
- * @param {Map<number, ClusterStart>} starts
- * @param {number} at
- * @returns {ClusterStart | undefined}
- */
-function rootOf(starts, at) {
-  let current = starts.get(at);
-  const seen = new Set();
-  while (current && current.parent !== null && !seen.has(current.parent)) {
-    seen.add(current.parent);
-    const parent = starts.get(current.parent);
-    if (!parent) {
-      break;
-    }
-    current = parent;
-  }
-  return current;
-}
-
-/**
- * Whether every byte of an inclusive range lies in one downloaded run.
- *
- * @param {Array<[number, number]>} ranges
- * @param {number} start
- * @returns {number} One past the end of the run holding `start`, or `start`.
- */
-function runEndFrom(ranges, start) {
-  for (const [from, to] of ranges) {
-    if (start >= from && start <= to) {
-      return to + 1;
-    }
-  }
-  return start;
-}
-
-/**
- * The Timestamp of a cluster, read from its first children.
- *
- * Only the children inside the probe are looked at: the Timestamp is the first
- * child or the second after a CRC-32, and a cluster that puts it later is read
- * whole when it is processed.
- *
- * @param {ElementReader} reader
- * @param {{ dataOffset: number, end: number | null }} header
- * @param {number} limit
- * @param {number} secondsPerTick
- * @returns {Promise<number | null>}
- */
-async function leadingTimestamp(reader, header, limit, secondsPerTick) {
-  let at = header.dataOffset;
-  const stop = Math.min(limit, header.end ?? limit, header.dataOffset + CLUSTER_START_PROBE);
-  while (at < stop) {
-    const child = await reader.header(at, header.end ?? limit);
-    if (!child || child.size === null) {
-      return null;
-    }
-    if (child.id === ID_TIMESTAMP) {
-      const bytes = await reader.bytes(child.dataOffset, child.size);
-      return readUint(bytes, 0, child.size) * secondsPerTick;
-    }
-    if (child.id !== ID_CRC32 && child.id !== ID_VOID) {
-      return null;
-    }
-    at = child.end;
-  }
-  return null;
-}
-
-/**
- * Take the next start from every start whose header is downloaded.
- *
- * @param {object} state
- * @param {object} plan
- * @param {import("./Container.js").HeldReader} held
- * @param {ElementReader} reader
- * @returns {Promise<void>}
- */
-async function extendChains(state, plan, held, reader) {
-  const queue = [...state.starts.entries()].filter(([, start]) => !start.stepped).map(([at]) => at);
-  while (queue.length > 0) {
-    const at = queue.shift();
-    const start = state.starts.get(at);
-    if (!start || start.stepped || state.unreadable.has(at)) {
-      continue;
-    }
-    const probeEnd = Math.min(plan.fileSize, at + CLUSTER_START_PROBE) - 1;
-    if (!held.isHeld(at, Math.min(probeEnd, at + MAX_HEADER_BYTES - 1))) {
-      continue;
-    }
-    let header;
-    try {
-      header = await reader.header(at, plan.segmentEnd);
-    } catch (error) {
-      if (isUnavailable(error)) {
-        continue;
-      }
-      throw error;
-    }
-    start.stepped = true;
-    if (!header || !(LEVEL_ONE.has(header.id) || header.id === ID_VOID || header.id === ID_CRC32)) {
-      // A position the file itself names (or one a chain led to) that holds no
-      // top-level element: the file is damaged here, and reading it again gives
-      // the same bytes.
-      state.unreadable.set(at, "no top-level element at a position established as a cluster start");
-      continue;
-    }
-    start.kind = header.id === ID_CLUSTER ? "cluster" : "other";
-    start.end = header.end;
-    if (start.kind === "cluster" && start.seconds === null && held.isHeld(at, probeEnd)) {
-      try {
-        start.seconds = await leadingTimestamp(reader, header, plan.segmentEnd, plan.secondsPerTick);
-      } catch (error) {
-        if (!isUnavailable(error)) {
-          throw error;
-        }
-      }
-    }
-    if (header.end === null) {
-      // An unknown size is legal for a cluster (RFC 9559 §5.1.3); where it ends
-      // is learned when it is read.
-      continue;
-    }
-    addStart(state, header.end, plan, { source: "chain", parent: at }, queue);
-  }
-}
-
-/**
- * @param {object} state
- * @param {number} at
- * @param {object} plan
- * @param {{ source: ClusterStart["source"], parent: number | null, seconds?: number | null }} origin
- * @param {number[]} [queue]
- * @returns {void}
- */
-function addStart(state, at, plan, origin, queue) {
-  if (!(at < plan.segmentEnd) || at < plan.firstClusterAt) {
-    return;
-  }
-  const known = state.starts.get(at);
-  if (known) {
-    // The same position reached a second way. A start the file established
-    // outranks one the search found, and keeps nothing of its doubt.
-    if (known.source === "search" && origin.source !== "search") {
-      known.source = origin.source;
-      known.parent = origin.parent;
-    }
-    if (known.seconds === null && Number.isFinite(origin.seconds)) {
-      known.seconds = origin.seconds;
-    }
-    return;
-  }
-  state.starts.set(at, {
-    source: origin.source,
-    parent: origin.parent,
-    seconds: Number.isFinite(origin.seconds) ? origin.seconds : null,
-    kind: null,
-    end: null,
-    stepped: false
-  });
-  state.candidates.delete(at);
-  queue?.push(at);
-}
-
-/**
- * The spans of the clusters already known, as inclusive intervals.
- *
- * @param {object} state
- * @returns {Array<[number, number]>}
- */
-function knownSpans(state) {
-  const spans = [];
-  for (const [at, start] of state.starts) {
-    if (start.kind !== null && start.end !== null) {
-      spans.push([at, start.end - 1]);
-    }
-  }
-  return spans.sort((left, right) => left[0] - right[0]);
-}
-
-/**
- * `[from, to]` with every interval in `taken` removed.
- *
- * @param {number} from
- * @param {number} to
- * @param {Array<[number, number]>} taken - Sorted.
- * @returns {Array<[number, number]>}
- */
-function subtract(from, to, taken) {
-  const left = [];
-  let cursor = from;
-  for (const [start, end] of taken) {
-    if (end < cursor || start > to) {
-      continue;
-    }
-    if (start > cursor) {
-      left.push([cursor, start - 1]);
-    }
-    cursor = Math.max(cursor, end + 1);
-    if (cursor > to) {
-      break;
-    }
-  }
-  if (cursor <= to) {
-    left.push([cursor, to]);
-  }
-  return left;
-}
-
-/**
- * Search downloaded bytes the search has not finished with for the Cluster id.
- *
- * The last `MAX_HEADER_BYTES - 1` bytes of every downloaded run are left
- * unfinished: an id and a size cut by the end of the run are found when the run
- * grows, by searching those bytes again.
- *
- * Each position found is decided at once ({@link settleCandidate}); one that is
- * accepted is followed by its chain, and the search goes on after the chain
- * ends. A file whose clusters follow one another is therefore searched only up
- * to its first cluster after each gap, and the rest is known by its headers.
- *
- * @param {object} state
- * @param {object} plan
- * @param {import("./Container.js").HeldReader} held
- * @param {{ whole: ElementReader, structure: ElementReader }} readers
- * @returns {Promise<void>}
- */
-async function searchForClusters(state, plan, held, readers) {
-  const portion = Number.isFinite(held.portionBytes) && held.portionBytes > 0 ? held.portionBytes : plan.fileSize;
-  const idLength = CLUSTER_ID_BYTES.length;
-  for (const [runFrom, runTo] of held.ranges) {
-    const from = Math.max(runFrom, plan.firstClusterAt);
-    const to = Math.min(runTo, plan.segmentEnd - 1);
-    let at = from;
-    while (at + idLength - 1 <= to) {
-      // What is left to look at is worked out afresh at every step: a cluster
-      // accepted below starts a chain, and the bytes the chain covers are
-      // stepped over rather than searched. They are not marked as searched, so
-      // if the chain is ever taken back they are searched then.
-      const taken = [...state.scanned, ...knownSpans(state)].sort((left, right) => left[0] - right[0]);
-      const next = subtract(at, to, taken)[0];
-      if (!next || next[1] - next[0] + 1 < idLength) {
-        break;
-      }
-      const [start, end] = next;
-      const last = Math.min(end, start + portion - 1);
-      const bytes = await held.read(start, last);
-      state.stats.bytesWalked += bytes.length;
-      let accepted = null;
-      let found = bytes.indexOf(CLUSTER_ID_BYTES);
-      while (found >= 0) {
-        const position = start + found;
-        if (!state.starts.has(position) && !state.rejected.has(position)) {
-          const outcome = await settleCandidate(state, plan, held, readers, position);
-          if (outcome === "accepted") {
-            accepted = position;
-            break;
-          }
-        }
-        found = bytes.indexOf(CLUSTER_ID_BYTES, found + 1);
-      }
-      if (accepted !== null) {
-        if (accepted - 1 >= start) {
-          state.scanned = mergeIntervals([...state.scanned, [start, accepted - 1]]);
-        }
-        at = accepted;
-        continue;
-      }
-      if (last < end) {
-        // Every position whose whole id lies in this portion has been looked
-        // at; the next portion overlaps by the id's length less one, so an id
-        // cut by the boundary is found there.
-        state.scanned = mergeIntervals([...state.scanned, [start, last - idLength + 1]]);
-        at = last - idLength + 2;
-        continue;
-      }
-      // Held back only where the RUN ends here: a span that ends at a known
-      // element's boundary has nothing after it that could complete an id.
-      const runEndsHere = end === to && runTo < plan.segmentEnd - 1;
-      const finished = runEndsHere ? end - (MAX_HEADER_BYTES - 1) : end;
-      if (finished >= start) {
-        state.scanned = mergeIntervals([...state.scanned, [start, finished]]);
-      }
-      at = end + 1;
-    }
-  }
-}
-
-/**
- * Decide one position the search found the Cluster ID at.
- *
- * Accepted, it becomes a start and its chain is followed at once, so the
- * clusters after it are known by their headers and never searched for.
- *
- * @param {object} state
- * @param {object} plan
- * @param {import("./Container.js").HeldReader} held
- * @param {{ whole: ElementReader, structure: ElementReader }} readers
- * @param {number} at
- * @returns {Promise<"accepted" | "rejected" | "pending">}
- */
-async function settleCandidate(state, plan, held, readers, at) {
-  if (knownSpans(state).some(([from, to]) => at > from && at <= to)) {
-    // Inside an element whose bounds are known: not a cluster start.
-    state.candidates.delete(at);
-    state.rejected.add(at);
-    return "rejected";
-  }
-  const verdict = await verifyCandidate(readers.whole, at, plan, state, held);
-  if (verdict.status === "pending") {
-    state.candidates.add(at);
-    return "pending";
-  }
-  state.candidates.delete(at);
-  if (verdict.status === "rejected") {
-    state.rejected.add(at);
-    state.stats.rejectedCandidates += 1;
-    return "rejected";
-  }
-  addStart(state, at, plan, { source: "search", parent: null, seconds: verdict.seconds });
-  const start = state.starts.get(at);
-  if (start) {
-    start.kind = "cluster";
-    start.end = verdict.end;
-  }
-  await extendChains(state, plan, held, readers.structure);
-  return "accepted";
-}
-
-/**
- * @param {Array<[number, number]>} intervals
- * @returns {Array<[number, number]>}
- */
-function mergeIntervals(intervals) {
-  const sorted = [...intervals].sort((left, right) => left[0] - right[0]);
-  const merged = [];
-  for (const [start, end] of sorted) {
-    const last = merged[merged.length - 1];
-    if (last && start <= last[1] + 1) {
-      last[1] = Math.max(last[1], end);
-    } else {
-      merged.push([start, end]);
-    }
-  }
-  return merged;
-}
-
-/**
- * The confirmed starts nearest a position on either side, with their times.
- *
- * @param {object} state
- * @param {number} at
- * @returns {{ before: { at: number, start: ClusterStart } | null, after: { at: number, start: ClusterStart } | null }}
- */
-function neighboursOf(state, at) {
-  let before = null;
-  let after = null;
-  for (const [position, start] of state.starts) {
-    if (start.kind !== "cluster") {
-      continue;
-    }
-    if (position < at && (!before || position > before.at)) {
-      before = { at: position, start };
-    } else if (position > at && (!after || position < after.at)) {
-      after = { at: position, start };
-    }
-  }
-  return { before, after };
-}
-
-/**
- * Check a candidate against everything a cluster must be.
- *
- * Every child is one RFC 9559 §5.1.3 allows; there is exactly one Timestamp,
- * wherever it stands; the children fill the cluster exactly; every block names a
- * track the file declares; a CRC-32, where there is one, matches; a PrevSize,
- * where there is one and the cluster before is known, matches; and the time
- * lies between the times of the known clusters either side.
- *
- * @param {ElementReader} reader
- * @param {number} at
- * @param {object} plan
- * @param {object} state
- * @param {import("./Container.js").HeldReader} held
- * @returns {Promise<{ status: "pending" } | { status: "rejected", reason: string } | { status: "confirmed", end: number, seconds: number }>}
- */
-async function verifyCandidate(reader, at, plan, state, held) {
-  try {
-    const header = await reader.header(at, plan.segmentEnd);
-    if (!header || header.id !== ID_CLUSTER) {
-      return { status: "rejected", reason: "no cluster header" };
-    }
-    if (header.end !== null && header.end > plan.segmentEnd) {
-      return { status: "rejected", reason: "runs past the segment" };
-    }
-    if (header.end !== null && !held.isHeld(at, header.end - 1)) {
-      return { status: "pending" };
-    }
-    const limit = header.end ?? plan.segmentEnd;
-    let cursor = header.dataOffset;
-    let timestamps = 0;
-    let ticks = null;
-    let crc = null;
-    let crcEnd = null;
-    let prevSize = null;
-    let end = header.end;
-    let first = true;
-    while (cursor < limit) {
-      if (!held.isHeld(cursor, Math.min(plan.fileSize, cursor + MAX_HEADER_BYTES) - 1)) {
-        return { status: "pending" };
-      }
-      const child = await reader.header(cursor, limit);
-      if (!child) {
-        return { status: "rejected", reason: "a child header does not parse" };
-      }
-      if (header.end === null && LEVEL_ONE.has(child.id)) {
-        end = cursor;
-        break;
-      }
-      if (!CLUSTER_CHILDREN.has(child.id)) {
-        return { status: "rejected", reason: `0x${child.id.toString(16)} is not a cluster child` };
-      }
-      if (child.size === null || child.end > limit) {
-        return { status: "rejected", reason: "a child runs past the cluster" };
-      }
-      if (child.id === ID_CRC32) {
-        if (!first || child.size !== 4) {
-          return { status: "rejected", reason: "a CRC-32 that is not the first child" };
-        }
-        crc = (await reader.bytes(child.dataOffset, 4)).readUInt32LE(0);
-        crcEnd = child.end;
-      } else if (child.id === ID_TIMESTAMP) {
-        timestamps += 1;
-        ticks = readUint(await reader.bytes(child.dataOffset, child.size), 0, child.size);
-      } else if (child.id === ID_PREV_SIZE) {
-        prevSize = readUint(await reader.bytes(child.dataOffset, child.size), 0, child.size);
-      } else if (child.id === ID_SIMPLE_BLOCK || child.id === ID_BLOCK_GROUP) {
-        const track = await blockTrackOf(reader, child);
-        if (track === null || !plan.allTrackNumbers.includes(track)) {
-          return { status: "rejected", reason: "a block names no declared track" };
-        }
-      }
-      first = false;
-      cursor = child.end;
-    }
-    if (end === null) {
-      end = limit;
-    }
-    if (cursor !== end) {
-      return { status: "rejected", reason: "the children do not fill the cluster" };
-    }
-    if (timestamps !== 1) {
-      return { status: "rejected", reason: `${timestamps} Timestamp elements` };
-    }
-    if (crc !== null) {
-      // RFC 8794 §11.3.1: the CRC covers every byte of the parent's data after
-      // the CRC-32 element itself, stored little-endian.
-      const computed = await crcOf(reader, crcEnd, end, held.portionBytes);
-      if (computed !== crc) {
-        return { status: "rejected", reason: "CRC-32 does not match" };
-      }
-    }
-    const seconds = ticks * plan.secondsPerTick;
-    const { before, after } = neighboursOf(state, at);
-    if (prevSize !== null && before && before.start.end === at && prevSize !== at - before.at) {
-      return { status: "rejected", reason: "PrevSize does not match the cluster before" };
-    }
-    if (before && before.start.seconds !== null && seconds < before.start.seconds) {
-      return { status: "rejected", reason: "earlier than the cluster before it" };
-    }
-    if (after && after.start.seconds !== null && seconds > after.start.seconds) {
-      return { status: "rejected", reason: "later than the cluster after it" };
-    }
-    return { status: "confirmed", end, seconds };
-  } catch (error) {
-    if (isUnavailable(error)) {
-      return { status: "pending" };
-    }
-    throw error;
-  }
-}
-
-/**
- * CRC-32 (IEEE 802.3, RFC 8794 §11.3.1) of a span, a portion at a time.
- *
- * @param {ElementReader} reader
- * @param {number} from
- * @param {number} to - One past the last byte.
- * @param {number} portionBytes
- * @returns {Promise<number>}
- */
-async function crcOf(reader, from, to, portionBytes) {
-  const step = Number.isFinite(portionBytes) && portionBytes > 0 ? portionBytes : to - from;
-  let value = 0;
-  for (let at = from; at < to; at += step) {
-    value = crc32(await reader.bytes(at, Math.min(step, to - at)), value);
-  }
-  return value >>> 0;
-}
-
-/**
- * The track a SimpleBlock or a BlockGroup's Block belongs to, read from its
- * first bytes only.
- *
- * @param {ElementReader} reader
- * @param {import("./ebml-stream.js").ElementHeader} child
- * @returns {Promise<number | null>}
- */
-async function blockTrackOf(reader, child) {
-  if (child.id === ID_SIMPLE_BLOCK) {
-    const head = await reader.peek(child, 8);
-    const track = readVint(head, 0, false);
-    return track ? track.value : null;
-  }
-  let at = child.dataOffset;
-  while (at < child.end) {
-    const inner = await reader.header(at, child.end);
-    if (!inner || inner.size === null) {
-      return null;
-    }
-    if (inner.id === ID_BLOCK) {
-      const head = await reader.peek(inner, 8);
-      const track = readVint(head, 0, false);
-      return track ? track.value : null;
-    }
-    at = inner.end;
-  }
-  return null;
-}
-
-/**
- * Take back a start the search found, and everything chained from it, when a
- * start the file established lands inside it — or it inside one.
- *
- * @param {object} state
- * @returns {number[]} The cluster starts withdrawn, whose cues must be taken back.
- */
-function withdrawContradicted(state) {
-  const established = [];
-  const searched = [];
-  for (const [at, start] of state.starts) {
-    if (start.kind !== "cluster") {
-      continue;
-    }
-    const root = rootOf(state.starts, at);
-    (root?.source === "search" ? searched : established).push([at, start]);
-  }
-  const contradicted = new Set();
-  for (const [at, start] of searched) {
-    for (const [other, otherStart] of established) {
-      const insideIt = start.end !== null && other > at && other < start.end;
-      const itInside = otherStart.end !== null && at > other && at < otherStart.end;
-      if (insideIt || itInside) {
-        contradicted.add(at);
-        break;
-      }
-    }
-  }
-  if (contradicted.size === 0) {
-    return [];
-  }
-  // Everything chained from a withdrawn start goes with it.
-  let grew = true;
-  while (grew) {
-    grew = false;
-    for (const [at, start] of state.starts) {
-      if (!contradicted.has(at) && start.parent !== null && contradicted.has(start.parent)) {
-        contradicted.add(at);
-        grew = true;
-      }
-    }
-  }
-  const withdrawn = [];
-  for (const at of contradicted) {
-    const start = state.starts.get(at);
-    state.starts.delete(at);
-    state.rejected.add(at);
-    if (state.done.delete(at) && start?.kind === "cluster") {
-      withdrawn.push(at);
-    }
-  }
-  state.stats.contradictions += contradicted.size;
-  return withdrawn;
-}
-
-/**
- * The order to read pending clusters in: the ones viewers stand in now, and the
- * rest.
- *
- * Which cluster a viewer stands in is decided over EVERY cluster whose time is
- * known, read or not: deciding it over the pending ones only would name an
- * earlier cluster whenever the viewer's own one had already been read.
- *
- * @param {Array<[number, ClusterStart]>} pending
- * @param {Array<[number, ClusterStart]>} clusters - Every known cluster start.
- * @param {number[]} wantedSeconds
- * @returns {{ urgent: number[], rest: number[] }}
- */
-function readingOrder(pending, clusters, wantedSeconds) {
-  const isPending = new Set(pending.map(([at]) => at));
-  const timed = clusters.filter(([, start]) => start.seconds !== null).sort((left, right) => left[1].seconds - right[1].seconds);
-  const lists = [];
-  const urgent = [];
-  for (const seconds of wantedSeconds ?? []) {
-    if (!Number.isFinite(seconds)) {
-      continue;
-    }
-    let containing = -1;
-    for (let index = 0; index < timed.length; index += 1) {
-      if (timed[index][1].seconds <= seconds) {
-        containing = index;
-      } else {
-        break;
-      }
-    }
-    if (containing >= 0) {
-      urgent.push(timed[containing][0]);
-      if (containing > 0) {
-        urgent.push(timed[containing - 1][0]);
-      }
-    }
-    const list = [];
-    for (let index = containing + 1; index < timed.length; index += 1) {
-      list.push(timed[index][0]);
-    }
-    lists.push(list.filter((at) => isPending.has(at)));
-  }
-  const now = [...new Set(urgent.filter((at) => isPending.has(at)))];
-  const order = [];
-  const seen = new Set(now);
-  // One viewer's next cluster after another's, so nobody waits for another
-  // viewer's whole film.
-  for (let step = 0; lists.some((list) => step < list.length); step += 1) {
-    for (const list of lists) {
-      const at = list[step];
-      if (at !== undefined && !seen.has(at)) {
-        seen.add(at);
-        order.push(at);
-      }
-    }
-  }
-  for (const [at] of [...pending].sort((left, right) => left[0] - right[0])) {
-    if (!seen.has(at)) {
-      order.push(at);
-    }
-  }
-  return { urgent: now, rest: order };
-}
-
-/**
- * Where the viewers stand right now, however the reader states it.
- *
- * @param {import("./Container.js").HeldReader} held
- * @returns {number[]}
- */
-function wantedNow(held) {
-  const stated = typeof held.wantedSeconds === "function" ? held.wantedSeconds() : held.wantedSeconds;
-  return Array.isArray(stated) ? stated.filter((value) => Number.isFinite(value)) : [];
-}
-
-/**
- * Whether a cluster's bytes are all here to be read now — known only for a
- * cluster whose size is known.
- *
- * @param {Map<number, ClusterStart>} starts
- * @param {number} at
- * @param {import("./Container.js").HeldReader} held
- * @returns {boolean}
- */
-function readableNow(starts, at, held) {
-  const start = starts.get(at);
-  return Boolean(start && start.end !== null && held.isHeld(at, start.end - 1));
-}
-
-/**
- * Read one cluster's subtitle blocks.
- *
- * @param {{ structure: ElementReader, exact: ElementReader }} readers - Headers
- *   and track numbers through `structure`; a subtitle block's data through
- *   `exact`, which reads that element and nothing more.
- * @param {number} at
- * @param {ClusterStart} start
- * @param {object} plan
- * @param {object} state
- * @param {import("./Container.js").HeldReader} held
- * @returns {Promise<{ status: "pending" } | { status: "unreadable", reason: string } | { status: "done", blocks: object[], end: number }>}
- */
-async function readCluster({ structure: reader, exact }, at, start, plan, state, held) {
-  try {
-    if (start.end !== null && !held.isHeld(at, start.end - 1)) {
-      return { status: "pending" };
-    }
-    const header = await reader.header(at, plan.segmentEnd);
-    if (!header || header.id !== ID_CLUSTER) {
-      return { status: "unreadable", reason: "no cluster header" };
-    }
-    const limit = header.end ?? plan.segmentEnd;
-    const wanted = new Set(plan.subtitleTrackNumbers);
-    const raw = [];
-    let ticks = null;
-    let cursor = header.dataOffset;
-    let end = header.end;
-    while (cursor < limit) {
-      if (!held.isHeld(cursor, Math.min(plan.fileSize, cursor + MAX_HEADER_BYTES) - 1)) {
-        return { status: "pending" };
-      }
-      const child = await reader.header(cursor, limit);
-      if (!child) {
-        return { status: "unreadable", reason: "a child header does not parse" };
-      }
-      if (header.end === null && LEVEL_ONE.has(child.id)) {
-        end = cursor;
-        break;
-      }
-      if (child.size === null || child.end > limit) {
-        return { status: "unreadable", reason: "a child runs past the cluster" };
-      }
-      if (header.end === null && !held.isHeld(child.at, child.end - 1)) {
-        return { status: "pending" };
-      }
-      if (child.id === ID_TIMESTAMP) {
-        ticks = readUint(await reader.bytes(child.dataOffset, child.size), 0, child.size);
-      } else if (child.id === ID_SIMPLE_BLOCK || child.id === ID_BLOCK_GROUP) {
-        const track = await blockTrackOf(reader, child);
-        if (track !== null && wanted.has(track)) {
-          const data = await exact.data(child);
-          if (data === null) {
-            state.stats.refusedElements += 1;
-          } else {
-            const block = child.id === ID_SIMPLE_BLOCK ? simpleBlockOf(data) : blockGroupOf(data);
-            if (block) {
-              raw.push(block);
-            }
-          }
-        }
-      }
-      cursor = child.end;
-    }
-    if (end === null) {
-      end = limit;
-    }
-    if (ticks === null) {
-      return { status: "unreadable", reason: "a cluster with no Timestamp" };
-    }
-    state.stats.bytesWalked += end - at;
-    const blocks = raw.map((block) => {
-      const startSeconds = (ticks + block.relativeTicks) * plan.secondsPerTick;
-      return {
-        trackNumber: block.trackNumber,
-        startSeconds,
-        endSeconds: block.durationTicks === null ? null : startSeconds + block.durationTicks * plan.secondsPerTick,
-        payload: block.payload,
-        source: at
-      };
-    });
-    return { status: "done", blocks, end, seconds: ticks * plan.secondsPerTick };
-  } catch (error) {
-    if (isUnavailable(error)) {
-      return { status: "pending" };
-    }
-    throw error;
-  }
-}
-
-/**
- * @param {Buffer} data - A SimpleBlock's data.
- * @returns {{ trackNumber: number, relativeTicks: number, durationTicks: null, payload: Buffer } | null}
- */
-function simpleBlockOf(data) {
-  const header = readBlockHeader(data, 0, data.length);
-  if (!header) {
-    return null;
-  }
-  const payloadAt = firstFrameOffset(data, header.dataOffset, data.length, header.flags);
-  if (payloadAt === null || payloadAt >= data.length) {
-    return null;
-  }
-  return { trackNumber: header.trackNumber, relativeTicks: header.relativeTicks, durationTicks: null, payload: data.subarray(payloadAt) };
-}
-
-/**
- * @param {Buffer} data - A BlockGroup's data.
- * @returns {{ trackNumber: number, relativeTicks: number, durationTicks: number | null, payload: Buffer } | null}
- */
-function blockGroupOf(data) {
-  let block = null;
-  let durationTicks = null;
-  let at = 0;
-  while (at < data.length) {
-    const id = readVint(data, at, true);
-    const size = id && readVint(data, at + id.length, false);
-    if (!id || !size) {
-      return null;
-    }
-    const dataOffset = at + id.length + size.length;
-    const end = dataOffset + size.value;
-    if (end > data.length) {
-      return null;
-    }
-    if (id.value === ID_BLOCK) {
-      block = { start: dataOffset, end };
-    } else if (id.value === ID_BLOCK_DURATION) {
-      durationTicks = readUint(data, dataOffset, size.value);
-    }
-    at = end;
-  }
-  if (!block) {
-    return null;
-  }
-  const header = readBlockHeader(data, block.start, block.end);
-  if (!header) {
-    return null;
-  }
-  const payloadAt = firstFrameOffset(data, header.dataOffset, block.end, header.flags);
-  if (payloadAt === null || payloadAt >= block.end) {
-    return null;
-  }
-  return { trackNumber: header.trackNumber, relativeTicks: header.relativeTicks, durationTicks, payload: data.subarray(payloadAt, block.end) };
-}
-
-/**
- * One pass over what a Matroska file holds now.
- *
- * @param {object} params
- * @param {object} params.plan - From `MatroskaContainer.readSubtitlePlan`.
- * @param {object} params.progress - Kept by the caller between passes.
- * @param {import("./Container.js").HeldReader} params.held
- * @returns {Promise<{ found: Map<number, object[]>, covered: number, indexed: number, withdrawn: number[], more: boolean, stats: object }>}
- *   `more` says clusters are readable now that this pass left for the next.
- */
-export async function walkHeldClusters({ plan, progress, held }) {
-  const state = stateOf(progress);
-  const reader = new ElementReader({
-    read: held.read,
-    fileSize: plan.fileSize,
-    portionBytes: held.portionBytes,
-    prefetch: true,
-    readableUntil: (start) => runEndFrom(held.ranges, start)
-  });
-  if (!state.seeded) {
-    state.seeded = true;
-    if (Number.isFinite(plan.firstClusterAt)) {
-      addStart(state, plan.firstClusterAt, plan, { source: "first", parent: null });
-    }
-    for (const entry of plan.entryPoints ?? []) {
-      addStart(state, entry.at, plan, { source: "cues", parent: null, seconds: entry.seconds });
-    }
-  }
-
-  // The structure of clusters is read a header at a time, and a subtitle
-  // block's data by its own size. The reader above takes a whole portion per
-  // read and is for checking a candidate the search found, whose every byte is
-  // needed; used for structure, it moved every frame of the film into this
-  // thread — measured on the addon host 2026-10-01, the collector then spent
-  // three times as long as the walk itself, and the loop's 99th-percentile
-  // delay doubled.
-  const structure = new ElementReader({
-    read: held.read,
-    fileSize: plan.fileSize,
-    portionBytes: STRUCTURE_READ_BYTES,
-    prefetch: true,
-    readableUntil: (start) => runEndFrom(held.ranges, start)
-  });
-  const exact = new ElementReader({ read: held.read, fileSize: plan.fileSize, portionBytes: held.portionBytes });
-  await extendChains(state, plan, held, structure);
-
-  if (plan.cuesState === "absent") {
-    const readers = { whole: reader, structure };
-    // Candidates left waiting for bytes by an earlier pass, then the bytes not
-    // yet looked at.
-    for (const at of [...state.candidates].sort((left, right) => left - right)) {
-      if (state.starts.has(at)) {
-        state.candidates.delete(at);
-        continue;
-      }
-      await settleCandidate(state, plan, held, readers, at);
-    }
-    await searchForClusters(state, plan, held, readers);
-  }
-
-  const withdrawn = withdrawContradicted(state);
-
-  /** @type {Map<number, object[]>} */
-  const found = new Map();
-  const clusters = [...state.starts.entries()].filter(([, start]) => start.kind === "cluster");
-  const pending = clusters.filter(([at]) => !state.done.has(at) && !state.unreadable.has(at));
-  const wanted = wantedNow(held);
-  const { urgent, rest } = readingOrder(pending, clusters, wanted);
-  const urgentReadable = urgent.filter((at) => readableNow(state.starts, at, held));
-  let more = false;
-  /** @type {number[]} */
-  let order;
-  if (urgentReadable.length > 0) {
-    order = urgentReadable;
-    more = rest.some((at) => readableNow(state.starts, at, held));
-  } else {
-    order = [...urgent, ...rest];
-  }
-  const asked = JSON.stringify(wanted);
-  for (const at of order) {
-    if (urgentReadable.length === 0 && JSON.stringify(wantedNow(held)) !== asked) {
-      // A viewer is somewhere else now: the next pass starts from there.
-      more = true;
-      break;
-    }
-    const start = state.starts.get(at);
-    if (!start) {
-      continue;
-    }
-    const result = await readCluster({ structure, exact }, at, start, plan, state, held);
-    if (result.status === "pending") {
-      continue;
-    }
-    if (result.status === "unreadable") {
-      state.unreadable.set(at, result.reason);
-      continue;
-    }
-    state.done.add(at);
-    if (start.end === null) {
-      start.end = result.end;
-      addStart(state, result.end, plan, { source: "chain", parent: at });
-    }
-    if (start.seconds === null) {
-      start.seconds = result.seconds;
-    }
-    if (rootOf(state.starts, at)?.source === "search") {
-      state.stats.fromSearch += 1;
-    }
-    for (const block of result.blocks) {
-      const into = found.get(block.trackNumber) ?? [];
-      into.push(block);
-      found.set(block.trackNumber, into);
-    }
-  }
-  // A cluster read whole may have made the start after an unknown-sized one
-  // known; its header is taken on the next pass, with everything else new.
-
-  return {
-    found,
-    covered: state.done.size,
-    indexed: [...state.starts.values()].filter((start) => start.kind === "cluster").length,
-    withdrawn,
-    more,
-    stats: { ...state.stats, unreadable: state.unreadable.size, pendingCandidates: state.candidates.size }
-  };
-}
+/**
+ * @file Where a Matroska file's clusters are, found from the bytes already
+ * downloaded, and the subtitle blocks inside them — RFC 9559 §5.1.3.
+ *
+ * **Why the Cues table is not enough.** It used to be the only way in: the walk
+ * read the clusters the Cues table names for a subtitle track and nothing else.
+ * RFC 9559 §22.1 says each subtitle frame SHOULD be referenced there — a
+ * recommendation, not a guarantee — and a file may have no Cues at all. Such a
+ * file showed no subtitles, ever, while a comment promised the walk would read
+ * "the clusters the viewer's own playback brings in" instead. Nothing did.
+ *
+ * **Where a cluster begins is established three ways, and only the first two
+ * are proof:**
+ *
+ * 1. the file says so — the first cluster after the head, and every
+ *    `CueClusterPosition` of every track (the video's name nearly every
+ *    cluster, because §22.1 has each keyframe referenced);
+ * 2. a chain from a position already established: a level-1 element begins
+ *    exactly where the one before it ends, so from a cluster whose header is
+ *    downloaded the next one is known without reading anything else;
+ * 3. a search of downloaded bytes for the Cluster id, used only where the file
+ *    has no Cues. What it finds is a CANDIDATE, checked against the structure a
+ *    cluster must have before anything is read out of it, and taken back —
+ *    together with every cue it gave — if a position established by the first
+ *    two ways later lands inside it.
+ *
+ * **What is kept between passes, and what is not.** A cluster read whole is
+ * DONE and never read again. One whose bytes are not all downloaded is left as
+ * it is: nothing about it is recorded, because "not here yet" is not a fact
+ * about the cluster (`unavailable.js`). One whose downloaded, hash-verified
+ * bytes do not parse is UNREADABLE, with its reason, and not tried again —
+ * reading the same bytes again gives the same answer.
+ *
+ * **Order, and when a pass ends.** The cluster a viewer is standing in is read
+ * first, then the one before it — a line on screen now may have started there.
+ * A pass that had such clusters to read ends with them, so what is on screen is
+ * pushed after two clusters rather than after the whole file; it says `more`,
+ * and the caller runs the next pass at once. A pass with nothing of that kind
+ * reads onward from the viewers, one viewer's next cluster after another's, and
+ * then the rest — and stops, saying `more`, the moment a viewer is somewhere
+ * else, so a seek is answered by the next pass and not after the whole film.
+ */
+
+import { crc32 } from "node:zlib";
+import { readUint, readVint } from "./ebml-reader.js";
+import { ElementReader, MAX_HEADER_BYTES } from "./ebml-stream.js";
+import { isUnavailable } from "./unavailable.js";
+
+const ID_CLUSTER = 0x1f43b675;
+const CLUSTER_ID_BYTES = Buffer.from([0x1f, 0x43, 0xb6, 0x75]);
+
+/** The top-level elements of a Segment (RFC 9559 §5.1), Void and CRC-32 aside. */
+const LEVEL_ONE = new Set([
+  ID_CLUSTER,
+  0x1c53bb6b, // Cues
+  0x114d9b74, // SeekHead
+  0x1549a966, // Info
+  0x1654ae6b, // Tracks
+  0x1043a770, // Chapters
+  0x1941a469, // Attachments
+  0x1254c367 // Tags
+]);
+
+const ID_CRC32 = 0xbf;
+const ID_VOID = 0xec;
+const ID_TIMESTAMP = 0xe7;
+const ID_SILENT_TRACKS = 0x5854;
+const ID_POSITION = 0xa7;
+const ID_PREV_SIZE = 0xab;
+const ID_SIMPLE_BLOCK = 0xa3;
+const ID_BLOCK_GROUP = 0xa0;
+const ID_ENCRYPTED_BLOCK = 0xaf;
+const ID_BLOCK = 0xa1;
+const ID_BLOCK_DURATION = 0x9b;
+
+/** The children RFC 9559 §5.1.3 allows inside a Cluster, the global two included. */
+const CLUSTER_CHILDREN = new Set([
+  ID_CRC32,
+  ID_VOID,
+  ID_TIMESTAMP,
+  ID_SILENT_TRACKS,
+  ID_POSITION,
+  ID_PREV_SIZE,
+  ID_SIMPLE_BLOCK,
+  ID_BLOCK_GROUP,
+  ID_ENCRYPTED_BLOCK
+]);
+
+/**
+ * Enough of a cluster's start to hold its header, a CRC-32 and the Timestamp
+ * that follows it (§5.1.3.1: first, or second after a CRC-32).
+ */
+const CLUSTER_START_PROBE = 64;
+
+/**
+ * The bytes one read of a cluster's structure takes: an element header and the
+ * start of a block's data, which names the block's track — a track number is a
+ * vint of at most 8 bytes (RFC 9559 §10.1). Reading a cluster this way moves
+ * the headers of its video and audio blocks across the thread boundary and
+ * none of their frames.
+ */
+const STRUCTURE_READ_BYTES = MAX_HEADER_BYTES + 8;
+
+/** Bits 1-2 of a block's flags byte say how it is laced, or that it is not. */
+const LACING_MASK = 0x06;
+const LACING_NONE = 0x00;
+const LACING_XIPH = 0x02;
+const LACING_FIXED = 0x04;
+const LACING_EBML = 0x06;
+
+/**
+ * A block's own header: its track, its time relative to the cluster, its flags.
+ *
+ * @param {Buffer} buffer
+ * @param {number} start - First byte of the block's data.
+ * @param {number} end - One past its last byte.
+ * @returns {{ trackNumber: number, relativeTicks: number, flags: number, dataOffset: number } | null}
+ */
+export function readBlockHeader(buffer, start, end) {
+  const track = readVint(buffer, start, false);
+  if (!track || track.value === null) {
+    return null;
+  }
+  const timestampAt = start + track.length;
+  // Signed, and it can be negative: a block may belong slightly before the
+  // cluster it is stored in.
+  if (timestampAt + 3 > end) {
+    return null;
+  }
+  return {
+    trackNumber: Number(track.value),
+    relativeTicks: buffer.readInt16BE(timestampAt),
+    flags: buffer[timestampAt + 2],
+    dataOffset: timestampAt + 3
+  };
+}
+
+/**
+ * Where a laced block's first frame begins.
+ *
+ * @param {Buffer} buffer
+ * @param {number} dataOffset - First byte after the block header.
+ * @param {number} end
+ * @param {number} flags
+ * @returns {number | null}
+ */
+export function firstFrameOffset(buffer, dataOffset, end, flags) {
+  const lacing = flags & LACING_MASK;
+  if (lacing === LACING_NONE) {
+    return dataOffset;
+  }
+  if (dataOffset >= end) {
+    return null;
+  }
+  const frames = buffer[dataOffset] + 1;
+  let at = dataOffset + 1;
+  if (lacing === LACING_FIXED) {
+    return at;
+  }
+  if (lacing === LACING_XIPH) {
+    for (let frame = 0; frame < frames - 1; frame += 1) {
+      while (at < end && buffer[at] === 0xff) {
+        at += 1;
+      }
+      at += 1;
+    }
+    return at <= end ? at : null;
+  }
+  if (lacing === LACING_EBML) {
+    for (let frame = 0; frame < frames - 1; frame += 1) {
+      const size = readVint(buffer, at, false);
+      if (!size) {
+        return null;
+      }
+      at += size.length;
+    }
+    return at <= end ? at : null;
+  }
+  return null;
+}
+
+/**
+ * The walk's own record of one file, kept by the caller between passes.
+ *
+ * @typedef {object} ClusterStart
+ * @property {"first" | "cues" | "chain" | "search"} source
+ * @property {number | null} parent - The start this one was chained from.
+ * @property {number | null} seconds - The cluster's Timestamp, once read.
+ * @property {"cluster" | "other" | null} kind - Null until its header is read.
+ * @property {number | null} end - One past its last byte, null while unknown.
+ * @property {boolean} stepped - Whether the next start has been taken from it.
+ */
+
+/**
+ * @param {object} progress
+ * @returns {object}
+ */
+function stateOf(progress) {
+  if (!progress.matroska) {
+    progress.matroska = {
+      seeded: false,
+      /** @type {Map<number, ClusterStart>} */
+      starts: new Map(),
+      done: new Set(),
+      /** @type {Map<number, string>} */
+      unreadable: new Map(),
+      /** Candidates whose bytes are not all here yet. @type {Set<number>} */
+      candidates: new Set(),
+      /** Candidates the structure check refused. @type {Set<number>} */
+      rejected: new Set(),
+      /** Inclusive intervals the search has finished with. @type {Array<[number, number]>} */
+      scanned: [],
+      stats: { refusedElements: 0, rejectedCandidates: 0, contradictions: 0, bytesWalked: 0, fromSearch: 0 }
+    };
+  }
+  return progress.matroska;
+}
+
+/**
+ * The start a chain of starts began from.
+ *
+ * @param {Map<number, ClusterStart>} starts
+ * @param {number} at
+ * @returns {ClusterStart | undefined}
+ */
+function rootOf(starts, at) {
+  let current = starts.get(at);
+  const seen = new Set();
+  while (current && current.parent !== null && !seen.has(current.parent)) {
+    seen.add(current.parent);
+    const parent = starts.get(current.parent);
+    if (!parent) {
+      break;
+    }
+    current = parent;
+  }
+  return current;
+}
+
+/**
+ * Whether every byte of an inclusive range lies in one downloaded run.
+ *
+ * @param {Array<[number, number]>} ranges
+ * @param {number} start
+ * @returns {number} One past the end of the run holding `start`, or `start`.
+ */
+function runEndFrom(ranges, start) {
+  for (const [from, to] of ranges) {
+    if (start >= from && start <= to) {
+      return to + 1;
+    }
+  }
+  return start;
+}
+
+/**
+ * The Timestamp of a cluster, read from its first children.
+ *
+ * Only the children inside the probe are looked at: the Timestamp is the first
+ * child or the second after a CRC-32, and a cluster that puts it later is read
+ * whole when it is processed.
+ *
+ * @param {ElementReader} reader
+ * @param {{ dataOffset: number, end: number | null }} header
+ * @param {number} limit
+ * @param {number} secondsPerTick
+ * @returns {Promise<number | null>}
+ */
+async function leadingTimestamp(reader, header, limit, secondsPerTick) {
+  let at = header.dataOffset;
+  const stop = Math.min(limit, header.end ?? limit, header.dataOffset + CLUSTER_START_PROBE);
+  while (at < stop) {
+    const child = await reader.header(at, header.end ?? limit);
+    if (!child || child.size === null) {
+      return null;
+    }
+    if (child.id === ID_TIMESTAMP) {
+      const bytes = await reader.bytes(child.dataOffset, child.size);
+      return readUint(bytes, 0, child.size) * secondsPerTick;
+    }
+    if (child.id !== ID_CRC32 && child.id !== ID_VOID) {
+      return null;
+    }
+    at = child.end;
+  }
+  return null;
+}
+
+/**
+ * Take the next start from every start whose header is downloaded.
+ *
+ * @param {object} state
+ * @param {object} plan
+ * @param {import("./Container.js").HeldReader} held
+ * @param {ElementReader} reader
+ * @returns {Promise<void>}
+ */
+async function extendChains(state, plan, held, reader) {
+  const queue = [...state.starts.entries()].filter(([, start]) => !start.stepped).map(([at]) => at);
+  while (queue.length > 0) {
+    const at = queue.shift();
+    const start = state.starts.get(at);
+    if (!start || start.stepped || state.unreadable.has(at)) {
+      continue;
+    }
+    const probeEnd = Math.min(plan.fileSize, at + CLUSTER_START_PROBE) - 1;
+    if (!held.isHeld(at, Math.min(probeEnd, at + MAX_HEADER_BYTES - 1))) {
+      continue;
+    }
+    let header;
+    try {
+      header = await reader.header(at, plan.segmentEnd);
+    } catch (error) {
+      if (isUnavailable(error)) {
+        continue;
+      }
+      throw error;
+    }
+    start.stepped = true;
+    if (!header || !(LEVEL_ONE.has(header.id) || header.id === ID_VOID || header.id === ID_CRC32)) {
+      // A position the file itself names (or one a chain led to) that holds no
+      // top-level element: the file is damaged here, and reading it again gives
+      // the same bytes.
+      state.unreadable.set(at, "no top-level element at a position established as a cluster start");
+      continue;
+    }
+    start.kind = header.id === ID_CLUSTER ? "cluster" : "other";
+    start.end = header.end;
+    if (start.kind === "cluster" && start.seconds === null && held.isHeld(at, probeEnd)) {
+      try {
+        start.seconds = await leadingTimestamp(reader, header, plan.segmentEnd, plan.secondsPerTick);
+      } catch (error) {
+        if (!isUnavailable(error)) {
+          throw error;
+        }
+      }
+    }
+    if (header.end === null) {
+      // An unknown size is legal for a cluster (RFC 9559 §5.1.3); where it ends
+      // is learned when it is read.
+      continue;
+    }
+    addStart(state, header.end, plan, { source: "chain", parent: at }, queue);
+  }
+}
+
+/**
+ * @param {object} state
+ * @param {number} at
+ * @param {object} plan
+ * @param {{ source: ClusterStart["source"], parent: number | null, seconds?: number | null }} origin
+ * @param {number[]} [queue]
+ * @returns {void}
+ */
+function addStart(state, at, plan, origin, queue) {
+  if (!(at < plan.segmentEnd) || at < plan.firstClusterAt) {
+    return;
+  }
+  const known = state.starts.get(at);
+  if (known) {
+    // The same position reached a second way. A start the file established
+    // outranks one the search found, and keeps nothing of its doubt.
+    if (known.source === "search" && origin.source !== "search") {
+      known.source = origin.source;
+      known.parent = origin.parent;
+    }
+    if (known.seconds === null && Number.isFinite(origin.seconds)) {
+      known.seconds = origin.seconds;
+    }
+    return;
+  }
+  state.starts.set(at, {
+    source: origin.source,
+    parent: origin.parent,
+    seconds: Number.isFinite(origin.seconds) ? origin.seconds : null,
+    kind: null,
+    end: null,
+    stepped: false
+  });
+  state.candidates.delete(at);
+  queue?.push(at);
+}
+
+/**
+ * The spans of the clusters already known, as inclusive intervals.
+ *
+ * @param {object} state
+ * @returns {Array<[number, number]>}
+ */
+function knownSpans(state) {
+  const spans = [];
+  for (const [at, start] of state.starts) {
+    if (start.kind !== null && start.end !== null) {
+      spans.push([at, start.end - 1]);
+    }
+  }
+  return spans.sort((left, right) => left[0] - right[0]);
+}
+
+/**
+ * `[from, to]` with every interval in `taken` removed.
+ *
+ * @param {number} from
+ * @param {number} to
+ * @param {Array<[number, number]>} taken - Sorted.
+ * @returns {Array<[number, number]>}
+ */
+function subtract(from, to, taken) {
+  const left = [];
+  let cursor = from;
+  for (const [start, end] of taken) {
+    if (end < cursor || start > to) {
+      continue;
+    }
+    if (start > cursor) {
+      left.push([cursor, start - 1]);
+    }
+    cursor = Math.max(cursor, end + 1);
+    if (cursor > to) {
+      break;
+    }
+  }
+  if (cursor <= to) {
+    left.push([cursor, to]);
+  }
+  return left;
+}
+
+/**
+ * Search downloaded bytes the search has not finished with for the Cluster id.
+ *
+ * The last `MAX_HEADER_BYTES - 1` bytes of every downloaded run are left
+ * unfinished: an id and a size cut by the end of the run are found when the run
+ * grows, by searching those bytes again.
+ *
+ * Each position found is decided at once ({@link settleCandidate}); one that is
+ * accepted is followed by its chain, and the search goes on after the chain
+ * ends. A file whose clusters follow one another is therefore searched only up
+ * to its first cluster after each gap, and the rest is known by its headers.
+ *
+ * @param {object} state
+ * @param {object} plan
+ * @param {import("./Container.js").HeldReader} held
+ * @param {{ whole: ElementReader, structure: ElementReader }} readers
+ * @returns {Promise<void>}
+ */
+async function searchForClusters(state, plan, held, readers) {
+  const portion = Number.isFinite(held.portionBytes) && held.portionBytes > 0 ? held.portionBytes : plan.fileSize;
+  const idLength = CLUSTER_ID_BYTES.length;
+  for (const [runFrom, runTo] of held.ranges) {
+    const from = Math.max(runFrom, plan.firstClusterAt);
+    const to = Math.min(runTo, plan.segmentEnd - 1);
+    let at = from;
+    while (at + idLength - 1 <= to) {
+      // What is left to look at is worked out afresh at every step: a cluster
+      // accepted below starts a chain, and the bytes the chain covers are
+      // stepped over rather than searched. They are not marked as searched, so
+      // if the chain is ever taken back they are searched then.
+      const taken = [...state.scanned, ...knownSpans(state)].sort((left, right) => left[0] - right[0]);
+      const next = subtract(at, to, taken)[0];
+      if (!next || next[1] - next[0] + 1 < idLength) {
+        break;
+      }
+      const [start, end] = next;
+      const last = Math.min(end, start + portion - 1);
+      const bytes = await held.read(start, last);
+      state.stats.bytesWalked += bytes.length;
+      let accepted = null;
+      let found = bytes.indexOf(CLUSTER_ID_BYTES);
+      while (found >= 0) {
+        const position = start + found;
+        if (!state.starts.has(position) && !state.rejected.has(position)) {
+          const outcome = await settleCandidate(state, plan, held, readers, position);
+          if (outcome === "accepted") {
+            accepted = position;
+            break;
+          }
+        }
+        found = bytes.indexOf(CLUSTER_ID_BYTES, found + 1);
+      }
+      if (accepted !== null) {
+        if (accepted - 1 >= start) {
+          state.scanned = mergeIntervals([...state.scanned, [start, accepted - 1]]);
+        }
+        at = accepted;
+        continue;
+      }
+      if (last < end) {
+        // Every position whose whole id lies in this portion has been looked
+        // at; the next portion overlaps by the id's length less one, so an id
+        // cut by the boundary is found there.
+        state.scanned = mergeIntervals([...state.scanned, [start, last - idLength + 1]]);
+        at = last - idLength + 2;
+        continue;
+      }
+      // Held back only where the RUN ends here: a span that ends at a known
+      // element's boundary has nothing after it that could complete an id.
+      const runEndsHere = end === to && runTo < plan.segmentEnd - 1;
+      const finished = runEndsHere ? end - (MAX_HEADER_BYTES - 1) : end;
+      if (finished >= start) {
+        state.scanned = mergeIntervals([...state.scanned, [start, finished]]);
+      }
+      at = end + 1;
+    }
+  }
+}
+
+/**
+ * Decide one position the search found the Cluster ID at.
+ *
+ * Accepted, it becomes a start and its chain is followed at once, so the
+ * clusters after it are known by their headers and never searched for.
+ *
+ * @param {object} state
+ * @param {object} plan
+ * @param {import("./Container.js").HeldReader} held
+ * @param {{ whole: ElementReader, structure: ElementReader }} readers
+ * @param {number} at
+ * @returns {Promise<"accepted" | "rejected" | "pending">}
+ */
+async function settleCandidate(state, plan, held, readers, at) {
+  if (knownSpans(state).some(([from, to]) => at > from && at <= to)) {
+    // Inside an element whose bounds are known: not a cluster start.
+    state.candidates.delete(at);
+    state.rejected.add(at);
+    return "rejected";
+  }
+  const verdict = await verifyCandidate(readers.whole, at, plan, state, held);
+  if (verdict.status === "pending") {
+    state.candidates.add(at);
+    return "pending";
+  }
+  state.candidates.delete(at);
+  if (verdict.status === "rejected") {
+    state.rejected.add(at);
+    state.stats.rejectedCandidates += 1;
+    return "rejected";
+  }
+  addStart(state, at, plan, { source: "search", parent: null, seconds: verdict.seconds });
+  const start = state.starts.get(at);
+  if (start) {
+    start.kind = "cluster";
+    start.end = verdict.end;
+  }
+  await extendChains(state, plan, held, readers.structure);
+  return "accepted";
+}
+
+/**
+ * @param {Array<[number, number]>} intervals
+ * @returns {Array<[number, number]>}
+ */
+function mergeIntervals(intervals) {
+  const sorted = [...intervals].sort((left, right) => left[0] - right[0]);
+  const merged = [];
+  for (const [start, end] of sorted) {
+    const last = merged[merged.length - 1];
+    if (last && start <= last[1] + 1) {
+      last[1] = Math.max(last[1], end);
+    } else {
+      merged.push([start, end]);
+    }
+  }
+  return merged;
+}
+
+/**
+ * The confirmed starts nearest a position on either side, with their times.
+ *
+ * @param {object} state
+ * @param {number} at
+ * @returns {{ before: { at: number, start: ClusterStart } | null, after: { at: number, start: ClusterStart } | null }}
+ */
+function neighboursOf(state, at) {
+  let before = null;
+  let after = null;
+  for (const [position, start] of state.starts) {
+    if (start.kind !== "cluster") {
+      continue;
+    }
+    if (position < at && (!before || position > before.at)) {
+      before = { at: position, start };
+    } else if (position > at && (!after || position < after.at)) {
+      after = { at: position, start };
+    }
+  }
+  return { before, after };
+}
+
+/**
+ * Check a candidate against everything a cluster must be.
+ *
+ * Every child is one RFC 9559 §5.1.3 allows; there is exactly one Timestamp,
+ * wherever it stands; the children fill the cluster exactly; every block names a
+ * track the file declares; a CRC-32, where there is one, matches; a PrevSize,
+ * where there is one and the cluster before is known, matches; and the time
+ * lies between the times of the known clusters either side.
+ *
+ * @param {ElementReader} reader
+ * @param {number} at
+ * @param {object} plan
+ * @param {object} state
+ * @param {import("./Container.js").HeldReader} held
+ * @returns {Promise<{ status: "pending" } | { status: "rejected", reason: string } | { status: "confirmed", end: number, seconds: number }>}
+ */
+async function verifyCandidate(reader, at, plan, state, held) {
+  try {
+    const header = await reader.header(at, plan.segmentEnd);
+    if (!header || header.id !== ID_CLUSTER) {
+      return { status: "rejected", reason: "no cluster header" };
+    }
+    if (header.end !== null && header.end > plan.segmentEnd) {
+      return { status: "rejected", reason: "runs past the segment" };
+    }
+    if (header.end !== null && !held.isHeld(at, header.end - 1)) {
+      return { status: "pending" };
+    }
+    const limit = header.end ?? plan.segmentEnd;
+    let cursor = header.dataOffset;
+    let timestamps = 0;
+    let ticks = null;
+    let crc = null;
+    let crcEnd = null;
+    let prevSize = null;
+    let end = header.end;
+    let first = true;
+    while (cursor < limit) {
+      if (!held.isHeld(cursor, Math.min(plan.fileSize, cursor + MAX_HEADER_BYTES) - 1)) {
+        return { status: "pending" };
+      }
+      const child = await reader.header(cursor, limit);
+      if (!child) {
+        return { status: "rejected", reason: "a child header does not parse" };
+      }
+      if (header.end === null && LEVEL_ONE.has(child.id)) {
+        end = cursor;
+        break;
+      }
+      if (!CLUSTER_CHILDREN.has(child.id)) {
+        return { status: "rejected", reason: `0x${child.id.toString(16)} is not a cluster child` };
+      }
+      if (child.size === null || child.end > limit) {
+        return { status: "rejected", reason: "a child runs past the cluster" };
+      }
+      if (child.id === ID_CRC32) {
+        if (!first || child.size !== 4) {
+          return { status: "rejected", reason: "a CRC-32 that is not the first child" };
+        }
+        crc = (await reader.bytes(child.dataOffset, 4)).readUInt32LE(0);
+        crcEnd = child.end;
+      } else if (child.id === ID_TIMESTAMP) {
+        timestamps += 1;
+        ticks = readUint(await reader.bytes(child.dataOffset, child.size), 0, child.size);
+      } else if (child.id === ID_PREV_SIZE) {
+        prevSize = readUint(await reader.bytes(child.dataOffset, child.size), 0, child.size);
+      } else if (child.id === ID_SIMPLE_BLOCK || child.id === ID_BLOCK_GROUP) {
+        const track = await blockTrackOf(reader, child);
+        if (track === null || !plan.allTrackNumbers.includes(track)) {
+          return { status: "rejected", reason: "a block names no declared track" };
+        }
+      }
+      first = false;
+      cursor = child.end;
+    }
+    if (end === null) {
+      end = limit;
+    }
+    if (cursor !== end) {
+      return { status: "rejected", reason: "the children do not fill the cluster" };
+    }
+    if (timestamps !== 1) {
+      return { status: "rejected", reason: `${timestamps} Timestamp elements` };
+    }
+    if (crc !== null) {
+      // RFC 8794 §11.3.1: the CRC covers every byte of the parent's data after
+      // the CRC-32 element itself, stored little-endian.
+      const computed = await crcOf(reader, crcEnd, end, held.portionBytes);
+      if (computed !== crc) {
+        return { status: "rejected", reason: "CRC-32 does not match" };
+      }
+    }
+    const seconds = ticks * plan.secondsPerTick;
+    const { before, after } = neighboursOf(state, at);
+    if (prevSize !== null && before && before.start.end === at && prevSize !== at - before.at) {
+      return { status: "rejected", reason: "PrevSize does not match the cluster before" };
+    }
+    if (before && before.start.seconds !== null && seconds < before.start.seconds) {
+      return { status: "rejected", reason: "earlier than the cluster before it" };
+    }
+    if (after && after.start.seconds !== null && seconds > after.start.seconds) {
+      return { status: "rejected", reason: "later than the cluster after it" };
+    }
+    return { status: "confirmed", end, seconds };
+  } catch (error) {
+    if (isUnavailable(error)) {
+      return { status: "pending" };
+    }
+    throw error;
+  }
+}
+
+/**
+ * CRC-32 (IEEE 802.3, RFC 8794 §11.3.1) of a span, a portion at a time.
+ *
+ * @param {ElementReader} reader
+ * @param {number} from
+ * @param {number} to - One past the last byte.
+ * @param {number} portionBytes
+ * @returns {Promise<number>}
+ */
+async function crcOf(reader, from, to, portionBytes) {
+  const step = Number.isFinite(portionBytes) && portionBytes > 0 ? portionBytes : to - from;
+  let value = 0;
+  for (let at = from; at < to; at += step) {
+    value = crc32(await reader.bytes(at, Math.min(step, to - at)), value);
+  }
+  return value >>> 0;
+}
+
+/**
+ * The track a SimpleBlock or a BlockGroup's Block belongs to, read from its
+ * first bytes only.
+ *
+ * @param {ElementReader} reader
+ * @param {import("./ebml-stream.js").ElementHeader} child
+ * @returns {Promise<number | null>}
+ */
+async function blockTrackOf(reader, child) {
+  if (child.id === ID_SIMPLE_BLOCK) {
+    const head = await reader.peek(child, 8);
+    const track = readVint(head, 0, false);
+    return track ? track.value : null;
+  }
+  let at = child.dataOffset;
+  while (at < child.end) {
+    const inner = await reader.header(at, child.end);
+    if (!inner || inner.size === null) {
+      return null;
+    }
+    if (inner.id === ID_BLOCK) {
+      const head = await reader.peek(inner, 8);
+      const track = readVint(head, 0, false);
+      return track ? track.value : null;
+    }
+    at = inner.end;
+  }
+  return null;
+}
+
+/**
+ * Take back a start the search found, and everything chained from it, when a
+ * start the file established lands inside it — or it inside one.
+ *
+ * @param {object} state
+ * @returns {number[]} The cluster starts withdrawn, whose cues must be taken back.
+ */
+function withdrawContradicted(state) {
+  const established = [];
+  const searched = [];
+  for (const [at, start] of state.starts) {
+    if (start.kind !== "cluster") {
+      continue;
+    }
+    const root = rootOf(state.starts, at);
+    (root?.source === "search" ? searched : established).push([at, start]);
+  }
+  const contradicted = new Set();
+  for (const [at, start] of searched) {
+    for (const [other, otherStart] of established) {
+      const insideIt = start.end !== null && other > at && other < start.end;
+      const itInside = otherStart.end !== null && at > other && at < otherStart.end;
+      if (insideIt || itInside) {
+        contradicted.add(at);
+        break;
+      }
+    }
+  }
+  if (contradicted.size === 0) {
+    return [];
+  }
+  // Everything chained from a withdrawn start goes with it.
+  let grew = true;
+  while (grew) {
+    grew = false;
+    for (const [at, start] of state.starts) {
+      if (!contradicted.has(at) && start.parent !== null && contradicted.has(start.parent)) {
+        contradicted.add(at);
+        grew = true;
+      }
+    }
+  }
+  const withdrawn = [];
+  for (const at of contradicted) {
+    const start = state.starts.get(at);
+    state.starts.delete(at);
+    state.rejected.add(at);
+    if (state.done.delete(at) && start?.kind === "cluster") {
+      withdrawn.push(at);
+    }
+  }
+  state.stats.contradictions += contradicted.size;
+  return withdrawn;
+}
+
+/**
+ * The order to read pending clusters in: the ones viewers stand in now, and the
+ * rest.
+ *
+ * Which cluster a viewer stands in is decided over EVERY cluster whose time is
+ * known, read or not: deciding it over the pending ones only would name an
+ * earlier cluster whenever the viewer's own one had already been read.
+ *
+ * @param {Array<[number, ClusterStart]>} pending
+ * @param {Array<[number, ClusterStart]>} clusters - Every known cluster start.
+ * @param {number[]} wantedSeconds
+ * @returns {{ urgent: number[], rest: number[] }}
+ */
+function readingOrder(pending, clusters, wantedSeconds) {
+  const isPending = new Set(pending.map(([at]) => at));
+  const timed = clusters.filter(([, start]) => start.seconds !== null).sort((left, right) => left[1].seconds - right[1].seconds);
+  const lists = [];
+  const urgent = [];
+  for (const seconds of wantedSeconds ?? []) {
+    if (!Number.isFinite(seconds)) {
+      continue;
+    }
+    let containing = -1;
+    for (let index = 0; index < timed.length; index += 1) {
+      if (timed[index][1].seconds <= seconds) {
+        containing = index;
+      } else {
+        break;
+      }
+    }
+    if (containing >= 0) {
+      urgent.push(timed[containing][0]);
+      if (containing > 0) {
+        urgent.push(timed[containing - 1][0]);
+      }
+    }
+    const list = [];
+    for (let index = containing + 1; index < timed.length; index += 1) {
+      list.push(timed[index][0]);
+    }
+    lists.push(list.filter((at) => isPending.has(at)));
+  }
+  const now = [...new Set(urgent.filter((at) => isPending.has(at)))];
+  const order = [];
+  const seen = new Set(now);
+  // One viewer's next cluster after another's, so nobody waits for another
+  // viewer's whole film.
+  for (let step = 0; lists.some((list) => step < list.length); step += 1) {
+    for (const list of lists) {
+      const at = list[step];
+      if (at !== undefined && !seen.has(at)) {
+        seen.add(at);
+        order.push(at);
+      }
+    }
+  }
+  for (const [at] of [...pending].sort((left, right) => left[0] - right[0])) {
+    if (!seen.has(at)) {
+      order.push(at);
+    }
+  }
+  return { urgent: now, rest: order };
+}
+
+/**
+ * Where the viewers stand right now, however the reader states it.
+ *
+ * @param {import("./Container.js").HeldReader} held
+ * @returns {number[]}
+ */
+function wantedNow(held) {
+  const stated = typeof held.wantedSeconds === "function" ? held.wantedSeconds() : held.wantedSeconds;
+  return Array.isArray(stated) ? stated.filter((value) => Number.isFinite(value)) : [];
+}
+
+/**
+ * Whether a cluster's bytes are all here to be read now — known only for a
+ * cluster whose size is known.
+ *
+ * @param {Map<number, ClusterStart>} starts
+ * @param {number} at
+ * @param {import("./Container.js").HeldReader} held
+ * @returns {boolean}
+ */
+function readableNow(starts, at, held) {
+  const start = starts.get(at);
+  return Boolean(start && start.end !== null && held.isHeld(at, start.end - 1));
+}
+
+/**
+ * Read one cluster's subtitle blocks.
+ *
+ * @param {{ structure: ElementReader, exact: ElementReader }} readers - Headers
+ *   and track numbers through `structure`; a subtitle block's data through
+ *   `exact`, which reads that element and nothing more.
+ * @param {number} at
+ * @param {ClusterStart} start
+ * @param {object} plan
+ * @param {object} state
+ * @param {import("./Container.js").HeldReader} held
+ * @returns {Promise<{ status: "pending" } | { status: "unreadable", reason: string } | { status: "done", blocks: object[], end: number }>}
+ */
+async function readCluster({ structure: reader, exact }, at, start, plan, state, held) {
+  try {
+    if (start.end !== null && !held.isHeld(at, start.end - 1)) {
+      return { status: "pending" };
+    }
+    const header = await reader.header(at, plan.segmentEnd);
+    if (!header || header.id !== ID_CLUSTER) {
+      return { status: "unreadable", reason: "no cluster header" };
+    }
+    const limit = header.end ?? plan.segmentEnd;
+    const wanted = new Set(plan.subtitleTrackNumbers);
+    const raw = [];
+    let ticks = null;
+    let cursor = header.dataOffset;
+    let end = header.end;
+    while (cursor < limit) {
+      if (!held.isHeld(cursor, Math.min(plan.fileSize, cursor + MAX_HEADER_BYTES) - 1)) {
+        return { status: "pending" };
+      }
+      const child = await reader.header(cursor, limit);
+      if (!child) {
+        return { status: "unreadable", reason: "a child header does not parse" };
+      }
+      if (header.end === null && LEVEL_ONE.has(child.id)) {
+        end = cursor;
+        break;
+      }
+      if (child.size === null || child.end > limit) {
+        return { status: "unreadable", reason: "a child runs past the cluster" };
+      }
+      if (header.end === null && !held.isHeld(child.at, child.end - 1)) {
+        return { status: "pending" };
+      }
+      if (child.id === ID_TIMESTAMP) {
+        ticks = readUint(await reader.bytes(child.dataOffset, child.size), 0, child.size);
+      } else if (child.id === ID_SIMPLE_BLOCK || child.id === ID_BLOCK_GROUP) {
+        const track = await blockTrackOf(reader, child);
+        if (track !== null && wanted.has(track)) {
+          const data = await exact.data(child);
+          if (data === null) {
+            state.stats.refusedElements += 1;
+          } else {
+            const block = child.id === ID_SIMPLE_BLOCK ? simpleBlockOf(data) : blockGroupOf(data);
+            if (block) {
+              raw.push(block);
+            }
+          }
+        }
+      }
+      cursor = child.end;
+    }
+    if (end === null) {
+      end = limit;
+    }
+    if (ticks === null) {
+      return { status: "unreadable", reason: "a cluster with no Timestamp" };
+    }
+    state.stats.bytesWalked += end - at;
+    const blocks = raw.map((block) => {
+      const startSeconds = (ticks + block.relativeTicks) * plan.secondsPerTick;
+      return {
+        trackNumber: block.trackNumber,
+        startSeconds,
+        endSeconds: block.durationTicks === null ? null : startSeconds + block.durationTicks * plan.secondsPerTick,
+        payload: block.payload,
+        source: at
+      };
+    });
+    return { status: "done", blocks, end, seconds: ticks * plan.secondsPerTick };
+  } catch (error) {
+    if (isUnavailable(error)) {
+      return { status: "pending" };
+    }
+    throw error;
+  }
+}
+
+/**
+ * @param {Buffer} data - A SimpleBlock's data.
+ * @returns {{ trackNumber: number, relativeTicks: number, durationTicks: null, payload: Buffer } | null}
+ */
+function simpleBlockOf(data) {
+  const header = readBlockHeader(data, 0, data.length);
+  if (!header) {
+    return null;
+  }
+  const payloadAt = firstFrameOffset(data, header.dataOffset, data.length, header.flags);
+  if (payloadAt === null || payloadAt >= data.length) {
+    return null;
+  }
+  return { trackNumber: header.trackNumber, relativeTicks: header.relativeTicks, durationTicks: null, payload: data.subarray(payloadAt) };
+}
+
+/**
+ * @param {Buffer} data - A BlockGroup's data.
+ * @returns {{ trackNumber: number, relativeTicks: number, durationTicks: number | null, payload: Buffer } | null}
+ */
+function blockGroupOf(data) {
+  let block = null;
+  let durationTicks = null;
+  let at = 0;
+  while (at < data.length) {
+    const id = readVint(data, at, true);
+    const size = id && readVint(data, at + id.length, false);
+    if (!id || !size) {
+      return null;
+    }
+    const dataOffset = at + id.length + size.length;
+    const end = dataOffset + size.value;
+    if (end > data.length) {
+      return null;
+    }
+    if (id.value === ID_BLOCK) {
+      block = { start: dataOffset, end };
+    } else if (id.value === ID_BLOCK_DURATION) {
+      durationTicks = readUint(data, dataOffset, size.value);
+    }
+    at = end;
+  }
+  if (!block) {
+    return null;
+  }
+  const header = readBlockHeader(data, block.start, block.end);
+  if (!header) {
+    return null;
+  }
+  const payloadAt = firstFrameOffset(data, header.dataOffset, block.end, header.flags);
+  if (payloadAt === null || payloadAt >= block.end) {
+    return null;
+  }
+  return { trackNumber: header.trackNumber, relativeTicks: header.relativeTicks, durationTicks, payload: data.subarray(payloadAt, block.end) };
+}
+
+/**
+ * One pass over what a Matroska file holds now.
+ *
+ * @param {object} params
+ * @param {object} params.plan - From `MatroskaContainer.readSubtitlePlan`.
+ * @param {object} params.progress - Kept by the caller between passes.
+ * @param {import("./Container.js").HeldReader} params.held
+ * @returns {Promise<{ found: Map<number, object[]>, covered: number, indexed: number, withdrawn: number[], more: boolean, stats: object }>}
+ *   `more` says clusters are readable now that this pass left for the next.
+ */
+export async function walkHeldClusters({ plan, progress, held }) {
+  const state = stateOf(progress);
+  const reader = new ElementReader({
+    read: held.read,
+    fileSize: plan.fileSize,
+    portionBytes: held.portionBytes,
+    prefetch: true,
+    readableUntil: (start) => runEndFrom(held.ranges, start)
+  });
+  if (!state.seeded) {
+    state.seeded = true;
+    if (Number.isFinite(plan.firstClusterAt)) {
+      addStart(state, plan.firstClusterAt, plan, { source: "first", parent: null });
+    }
+    for (const entry of plan.entryPoints ?? []) {
+      addStart(state, entry.at, plan, { source: "cues", parent: null, seconds: entry.seconds });
+    }
+  }
+
+  // The structure of clusters is read a header at a time, and a subtitle
+  // block's data by its own size. The reader above takes a whole portion per
+  // read and is for checking a candidate the search found, whose every byte is
+  // needed; used for structure, it moved every frame of the film into this
+  // thread — measured on the addon host 2026-10-01, the collector then spent
+  // three times as long as the walk itself, and the loop's 99th-percentile
+  // delay doubled.
+  const structure = new ElementReader({
+    read: held.read,
+    fileSize: plan.fileSize,
+    portionBytes: STRUCTURE_READ_BYTES,
+    prefetch: true,
+    readableUntil: (start) => runEndFrom(held.ranges, start)
+  });
+  const exact = new ElementReader({ read: held.read, fileSize: plan.fileSize, portionBytes: held.portionBytes });
+  await extendChains(state, plan, held, structure);
+
+  if (plan.cuesState === "absent") {
+    const readers = { whole: reader, structure };
+    // Candidates left waiting for bytes by an earlier pass, then the bytes not
+    // yet looked at.
+    for (const at of [...state.candidates].sort((left, right) => left - right)) {
+      if (state.starts.has(at)) {
+        state.candidates.delete(at);
+        continue;
+      }
+      await settleCandidate(state, plan, held, readers, at);
+    }
+    await searchForClusters(state, plan, held, readers);
+  }
+
+  const withdrawn = withdrawContradicted(state);
+
+  /** @type {Map<number, object[]>} */
+  const found = new Map();
+  const clusters = [...state.starts.entries()].filter(([, start]) => start.kind === "cluster");
+  const pending = clusters.filter(([at]) => !state.done.has(at) && !state.unreadable.has(at));
+  const wanted = wantedNow(held);
+  const { urgent, rest } = readingOrder(pending, clusters, wanted);
+  const urgentReadable = urgent.filter((at) => readableNow(state.starts, at, held));
+  let more = false;
+  /** @type {number[]} */
+  let order;
+  if (urgentReadable.length > 0) {
+    order = urgentReadable;
+    more = rest.some((at) => readableNow(state.starts, at, held));
+  } else {
+    order = [...urgent, ...rest];
+  }
+  const asked = JSON.stringify(wanted);
+  for (const at of order) {
+    if (urgentReadable.length === 0 && JSON.stringify(wantedNow(held)) !== asked) {
+      // A viewer is somewhere else now: the next pass starts from there.
+      more = true;
+      break;
+    }
+    const start = state.starts.get(at);
+    if (!start) {
+      continue;
+    }
+    const result = await readCluster({ structure, exact }, at, start, plan, state, held);
+    if (result.status === "pending") {
+      continue;
+    }
+    if (result.status === "unreadable") {
+      state.unreadable.set(at, result.reason);
+      continue;
+    }
+    state.done.add(at);
+    if (start.end === null) {
+      start.end = result.end;
+      addStart(state, result.end, plan, { source: "chain", parent: at });
+    }
+    if (start.seconds === null) {
+      start.seconds = result.seconds;
+    }
+    if (rootOf(state.starts, at)?.source === "search") {
+      state.stats.fromSearch += 1;
+    }
+    for (const block of result.blocks) {
+      const into = found.get(block.trackNumber) ?? [];
+      into.push(block);
+      found.set(block.trackNumber, into);
+    }
+  }
+  // A cluster read whole may have made the start after an unknown-sized one
+  // known; its header is taken on the next pass, with everything else new.
+
+  return {
+    found,
+    covered: state.done.size,
+    indexed: [...state.starts.values()].filter((start) => start.kind === "cluster").length,
+    withdrawn,
+    more,
+    stats: { ...state.stats, unreadable: state.unreadable.size, pendingCandidates: state.candidates.size }
+  };
+}