@torrent-tv/proxy 2.87.11 → 2.87.12

package/services/torrent/worker/client.js

@@ -1,765 +1,765 @@
-/**
- * @file Main-thread face of the torrent worker.
- *
- * Presents the same operations the routes and session manager already use, so
- * moving the torrent to its own thread does not ripple through calling code.
- * The one unavoidable change is that torrents are named by `sourceKey` instead
- * of passed around as objects — objects cannot cross a thread boundary, and
- * pretending otherwise would mean copying them on every call.
- *
- * Reads come back as an ordinary `ReadableStream`, so `/stream` and the codec
- * probe consume them exactly as they consume WebTorrent's own streams today.
- * What that hides is the part that matters: chunks arrive as transferred
- * buffers, never copied — 5.3 ms per 10 MB against 37 ms if cloned and 104 ms
- * through a transferable stream (measured 2026-08-02, see `protocol.js`).
- */
-
-import { Worker } from "node:worker_threads";
-import { Readable } from "node:stream";
-import { fileURLToPath } from "node:url";
-import { logger, writeAlreadyDecided } from "../../../utils/logger.js";
-import { createCaller, createReceiveStream } from "./channel.js";
-import { Command, Event } from "./protocol.js";
-
-const WORKER_URL = new URL("./worker.js", import.meta.url);
-/**
- * How long a shutdown waits for the thread to end by itself before forcing it.
- *
- * Ending by itself is what lets libuv drain the handle callbacks it is holding;
- * forcing it is what ran one of those against a freed isolate. Five seconds is
- * long enough for a destroyed torrent client to release its sockets and short
- * enough that a shutdown never appears to hang.
- */
-const WORKER_EXIT_GRACE_MS = 5_000;
-
-/**
- * Runs the torrent client on its own thread and exposes it to the main thread.
- *
- * Why this exists at all: profiling during a live seek found the main thread
- * ~85% busy with WebTorrent (buffer concatenation ~15%, wire updates ~9%,
- * garbage collection ~5%), while three of four cores idled. Serving a segment
- * queued behind that work, so reading a finished 10 MB file took 12-23 s where
- * handing it to the channel took 125 ms.
- */
-/**
- * Piece buffers this thread has been handed, and how many the collector has
- * taken back.
- *
- * A `SharedArrayBuffer`'s memory belongs to neither isolate: it lives until
- * BOTH have let go. So the worker's own count answers only half the question,
- * and this is the other half — 700 MB grew in the field on 2026-08-31 with the
- * store's own memory falling, and the two candidates are "a reference of ours
- * outlives the piece" and "the allocator keeps what we free". These two
- * counters separate them (roadmap item 2).
- *
- * A fragment arrives as a fresh handle onto the same shared memory every time,
- * so what is counted is handles, not pieces. That is the right quantity anyway:
- * one handle retained keeps the whole piece alive.
- */
-const fragmentBuffers = { seen: 0, collected: 0 };
-const fragmentCollector = typeof FinalizationRegistry === "function"
-  ? new FinalizationRegistry(() => {
-    fragmentBuffers.collected += 1;
-  })
-  : null;
-
-/**
- * What the collector has taken back against what this thread was handed.
- *
- * @returns {{ seen: number, collected: number }}
- */
-export function fragmentBufferCollection() {
-  return { seen: fragmentBuffers.seen, collected: fragmentBuffers.collected };
-}
-
-export class TorrentWorkerClient {
-  #worker;
-  /**
-   * Whether this shutdown was asked for, so the thread ending can be told from
-   * the thread dying. Without it both look identical from outside, which is how
-   * five crashes produced no line in the log.
-   */
-  #stopping = false;
-  #caller;
-  /** Receive-side handles for in-flight reads, keyed by request id. */
-  #reads = new Map();
-
-  /**
-   * The last piece served to each open read, so a fragment arriving out of
-   * order can be named. Cleared when the read ends.
-   *
-   * @type {Map<number, number>}
-   */
-   #lastPieceByRead = new Map();
-  /** Reads consuming fragments in place, keyed by request id. */
-  #fragmentReaders = new Map();
-
-  /**
-   * Files downloaded whole, by `${infoHash}/${fileIndex}`.
-   *
-   * Held on THIS thread because it is this thread that serves them: a file that
-   * is whole is read with an ordinary file read, and asking the torrent thread
-   * for anything about it would put back the very hop this removes. The torrent
-   * thread writes the files and says so; this is what it says.
-   *
-   * @type {Map<string, { path: string, length: number, name: string }>}
-   */
-  wholeFiles = new Map();
-
-  /** @type {(event: { sourceKey: string, fileIndex: number, trackIndex: number, cues: object[], language: string }) => void} */
-  /** @type {(event: { sourceKey: string, fileIndexes: number[] }) => void} */
-  #onPiecesArrived;
-
-  /**
-   * @param {{ memoryBytes?: number, stateDir?: string, onPiecesArrived?: (event: object) => void }} [options]
-   */
-  constructor({ memoryBytes, stateDir, onPiecesArrived } = {}) {
-    this.#worker = new Worker(fileURLToPath(WORKER_URL), {
-      // `stateDir` travels because the worker writes heap snapshots of its own
-      // isolate there. It cannot choose a directory any other way: a worker may
-      // not change the process's working directory, and the isolate that has
-      // died three times is the one no snapshot has ever been taken of.
-      workerData: { memoryBytes, stateDir }
-    });
-    this.#caller = createCaller(this.#worker);
-    this.#onPiecesArrived = onPiecesArrived ?? (() => undefined);
-
-    this.#worker.on("message", (message) => {
-      // A failed read must fail its stream. This is checked BEFORE the caller
-      // sees the message: until 2.9.76 nothing here handled a read error at
-      // all, so the worker's report was dropped as unknown, and because the
-      // worker sent the end-of-read marker from its `finally` even when the
-      // read had thrown, the reader saw a clean end of file instead. A read
-      // that failed before it produced anything simply hung forever.
-      if (message?.type === Event.ERROR && this.#reads.has(message.id)) {
-        const read = this.#reads.get(message.id);
-        this.#reads.delete(message.id);
-        read.fail(new Error(message.error ?? "Torrent worker read failed."));
-        return;
-      }
-      if (message?.type === Event.ERROR && this.#fragmentReaders.has(message.id)) {
-        const reader = this.#fragmentReaders.get(message.id);
-        this.#fragmentReaders.delete(message.id);
-        reader.fail(new Error(message.error ?? "Torrent worker read failed."));
-        return;
-      }
-      if (this.#caller.handleReply(message)) {
-        return;
-      }
-      switch (message?.type) {
-        case Event.FRAGMENT: {
-          const buffer = message.buffer;
-          if (!buffer) {
-            // No buffer means no way to read the fragment; confirm it so the
-            // worker is not left waiting, and let the read end short.
-            this.#worker.postMessage({ type: Event.FRAGMENT_DONE, id: message.id });
-            break;
-          }
-          if (message.offset + message.length > buffer.byteLength) {
-            logger.warn(
-              `torrent-worker: fragment ${message.offset}+${message.length} lies outside ` +
-              `its buffer of ${buffer.byteLength}B (read ${message.id}) — ending the read short`
-            );
-            this.#worker.postMessage({ type: Event.FRAGMENT_DONE, id: message.id });
-            break;
-          }
-          // A sequential read walks the file forwards, so each fragment either
-          // continues the piece before it or moves to the very next one.
-          // Anything else means the bytes handed to the decoder are not the
-          // file's bytes in order — which is what a decoder complaining about
-          // its input has twice turned out to mean (2.9.126, and the AC-3
-          // failure of 2026-08-09: the encoder ran at 9.3x, the piece store
-          // reported no spills and 100% of reads from memory, and the decoder
-          // still saw "new coupling strategy must be present in block 0"). The
-          // bounds check catches a fragment outside the pool; this catches one
-          // inside it that belongs somewhere else.
-          const lastPiece = this.#lastPieceByRead.get(message.id);
-          if (lastPiece !== undefined && message.pieceIndex !== lastPiece && message.pieceIndex !== lastPiece + 1) {
-            logger.warn(
-              `torrent-worker: read ${message.id} jumped from piece ${lastPiece} to ` +
-              `${message.pieceIndex} (${message.length}B at pool offset ${message.offset}) — ` +
-              "the consumer is being handed the file out of order"
-            );
-          }
-          this.#lastPieceByRead.set(message.id, message.pieceIndex);
-          fragmentBuffers.seen += 1;
-          fragmentCollector?.register(buffer, null);
-          const view = new Uint8Array(buffer, message.offset, message.length);
-
-          const reader = this.#fragmentReaders.get(message.id);
-          if (reader) {
-            // Handed on as a view into the pool — no copy anywhere. The piece
-            // stays pinned until the consumer says it is done with these exact
-            // bytes, which for a response body means the socket write has
-            // completed.
-            reader.push(view, () => {
-              this.#worker.postMessage({ type: Event.FRAGMENT_DONE, id: message.id });
-            });
-            break;
-          }
-
-          // Plain-stream consumers keep what they are given while the slot may
-          // be reused, so they get a copy and the piece is released at once.
-          this.#reads.get(message.id)?.push(Uint8Array.prototype.slice.call(view));
-          this.#worker.postMessage({ type: Event.FRAGMENT_DONE, id: message.id });
-          break;
-        }
-        case Event.CHUNK: {
-          const bytes = message.bytes;
-          this.#reads.get(message.id)?.push(
-            new Uint8Array(bytes.buffer, bytes.byteOffset, bytes.length)
-          );
-          break;
-        }
-        case Event.READ_END:
-          this.#lastPieceByRead.delete(message.id);
-          this.#reads.get(message.id)?.close();
-          this.#reads.delete(message.id);
-          this.#fragmentReaders.get(message.id)?.close();
-          this.#fragmentReaders.delete(message.id);
-          break;
-        case Event.LOG:
-          // Written as it is: the worker holds the same repeat rule and has
-          // already applied it, and deciding again here would be one decision
-          // taken twice on two different histories.
-          writeAlreadyDecided(message.level ?? "info", `torrent-worker: ${message.message}`);
-          break;
-        case Event.FILE_COMPLETE:
-          // Recorded here so the stream route can answer from the file without
-          // going near the torrent thread at all.
-          this.wholeFiles.set(`${message.infoHash}/${message.fileIndex}`, {
-            path: message.path,
-            length: message.length,
-            name: message.name
-          });
-          break;
-        case Event.PIECES_ARRIVED:
-          // ANNOUNCED by the thread that owns the swarm, acted on here: the
-          // subtitle walk is this thread's and needs to know a piece arrived.
-          this.#onPiecesArrived({
-            sourceKey: message.sourceKey,
-            fileIndexes: Array.isArray(message.fileIndexes) ? message.fileIndexes : []
-          });
-          break;
-        default:
-          break;
-      }
-    });
-
-    this.#worker.on("error", (error) => {
-      logger.error(`torrent-worker crashed: ${error?.message ?? error}`);
-      // Fail everything outstanding rather than leaving callers hanging: a dead
-      // worker will never answer, and a stalled request is worse than an error
-      // the loading flow can retry.
-      this.#failEverythingOutstanding(new Error("Torrent worker stopped unexpectedly."));
-    });
-
-    // A worker that ENDS was, until now, not noticed at all: only `message` and
-    // `error` were listened for. So when the thread went away the proxy simply
-    // stopped, the log ended mid-sentence, and nothing said whether we had
-    // asked for it — which is precisely the reading that was missing on
-    // 2026-08-21, when a core dump showed the thread faulting inside
-    // `Environment::CleanupHandles` and there was no way to tell our own
-    // shutdown from the thread ending on its own.
-    this.#worker.on("exit", (code) => {
-      if (this.#stopping) {
-        logger.info(`torrent-worker: thread ended as asked (code ${code})`);
-        return;
-      }
-      logger.error(
-        `torrent-worker: thread ended on its own with code ${code} — nobody asked it to. ` +
-        "Everything waiting on it is failed; the proxy has no torrent client until it is rebuilt."
-      );
-      this.#failEverythingOutstanding(new Error("Torrent worker ended unexpectedly."));
-    });
-  }
-
-  /**
-   * A dead thread will never answer, so everything waiting on it is failed.
-   *
-   * INCLUDING THE FRAGMENT READS, which is what ffmpeg's input is. Both death
-   * handlers walked `#reads` alone and left `#fragmentReaders` untouched, so an
-   * encoder's input neither ended nor errored — it went quiet. Field 2026-08-31,
-   * three times: the thread died, both ffmpeg runs stayed ALIVE and stopped
-   * producing (167 `holding segment-00085.mp4 … encoder alive` lines), the
-   * browser drained its cushion and then retried a segment that would never
-   * exist, for ever, with nothing shown to the viewer.
-   *
-   * @param {Error} reason
-   */
-  #failEverythingOutstanding(reason) {
-    this.#caller.rejectAll(reason);
-    for (const [, read] of this.#reads) {
-      read.fail(reason);
-    }
-    this.#reads.clear();
-    for (const [, reader] of this.#fragmentReaders) {
-      reader.fail(reason);
-    }
-    this.#fragmentReaders.clear();
-  }
-
-  /**
-   * Add (or join) a torrent and register it under `sourceKey`.
-   *
-   * @param {{ sourceKey: string, sourceType: "magnet" | "torrent", source: string }} params
-   * @returns {Promise<{ infoHash: string, name: string, files: { index: number, name: string, path: string, length: number }[] }>}
-   */
-  async addSource({ sourceKey, sourceType, source }) {
-    return this.#caller.call(Command.ADD_SOURCE, { sourceKey, sourceType, source });
-  }
-
-  /**
-   * The torrent's files, as plain data.
-   *
-   * @param {string} sourceKey
-   * @returns {Promise<{ index: number, name: string, path: string, length: number }[]>}
-   */
-  async listFiles(sourceKey) {
-    return this.#caller.call(Command.LIST_FILES, { sourceKey });
-  }
-
-  /**
-   * Live download figures for the progress display.
-   *
-   * @param {{ sourceKey: string, fileIndex: number, resumeAnchorByteStart?: number | null }} params
-   * @returns {Promise<object>}
-   */
-  async getFileStats({ sourceKey, fileIndex, resumeAnchorByteStart = null }) {
-    return this.#caller.call(Command.FILE_STATS, { sourceKey, fileIndex, resumeAnchorByteStart });
-  }
-
-  /**
-   * The text subtitle tracks a file carries.
-   *
-   * @param {{ sourceKey: string, fileIndex: number }} params
-   * @returns {Promise<{ tracks: object[] }>}
-   */
-
-
-
-
-  /**
-   * Start fetching the region a viewer is about to resume at.
-   *
-   * @param {{ sourceKey: string, fileIndex: number, positionSeconds: number }} params
-   * @returns {Promise<{ started: boolean }>}
-   */
-  /**
-   * Fetch one whole file in the room the viewer's own reading leaves.
-   *
-   * @param {{ sourceKey: string, fileIndex: number }} params
-   * @returns {Promise<{ started: boolean }>}
-   */
-  async fillFile({ sourceKey, fileIndex }) {
-    return this.#caller.call(Command.FILL_FILE, { sourceKey, fileIndex });
-  }
-
-  /**
-   * The byte ranges of one file the torrent holds WHOLE.
-   *
-   * @param {{ sourceKey: string, fileIndex: number }} params
-   * @returns {Promise<Array<[number, number]>>}
-   */
-  async heldRanges({ sourceKey, fileIndex }) {
-    const answer = await this.#caller.call(Command.HELD_RANGES, { sourceKey, fileIndex });
-    return Array.isArray(answer?.ranges) ? answer.ranges : [];
-  }
-
-  /**
-   * Bytes of a range the torrent already holds. Never fetched: null where they
-   * are not there.
-   *
-   * @param {{ sourceKey: string, fileIndex: number, start: number, end: number }} params
-   * @returns {Promise<Buffer | null>}
-   */
-  async readHeld({ sourceKey, fileIndex, start, end }) {
-    const answer = await this.#caller.call(Command.READ_HELD, { sourceKey, fileIndex, start, end });
-    return answer?.bytes ? Buffer.from(answer.bytes) : null;
-  }
-
-  async warmResumePosition({ sourceKey, fileIndex, positionSeconds, durationSeconds }) {
-    return this.#caller.call(Command.WARM_POSITION, { sourceKey, fileIndex, positionSeconds, durationSeconds });
-  }
-
-
-  /**
-   * Bytes every torrent on the worker has moved, downloaded and uploaded apart.
-   *
-   * The client lives on the worker thread, so this cannot be read as a property
-   * from here — which is exactly the mistake 2.19.0 shipped, leaving the figure
-   * always zero and the whole feature inert.
-   *
-   * @returns {Promise<{ downloaded: number, uploaded: number }>}
-   */
-  async getTorrentTotals() {
-    return this.#caller.call(Command.TORRENT_TOTALS, {});
-  }
-
-  /**
-   * Which films this proxy holds right now, and how much of each.
-   *
-   * Asked of the worker rather than of the main thread's map of stand-ins: that
-   * map is only cleared on shutdown, so it names films this proxy let go of
-   * hours ago.
-   *
-   * @returns {Promise<{ held: { infoHash: string, progress: number, bytes: number }[] }>}
-   */
-  async heldTorrents() {
-    return this.#caller.call(Command.HELD_TORRENTS, {});
-  }
-
-  /** State conditional demand for every file in a selected torrent. */
-  async fillTorrent(sourceKey) {
-    return this.#caller.call(Command.FILL_TORRENT, { sourceKey });
-  }
-
-  /**
-   * Reorder piece selection around a read position (seek prioritisation).
-   *
-   * @param {{ sourceKey: string, fileIndex: number, byteStart: number, windowBytes?: number, wholeFileRead?: boolean }} params
-   * @returns {Promise<void>}
-   */
-  async prioritizeByteRange({ sourceKey, fileIndex, byteStart, windowBytes, wholeFileRead }) {
-    await this.#caller.call(Command.PRIORITIZE, {
-      sourceKey,
-      fileIndex,
-      byteStart,
-      windowBytes,
-      wholeFileRead
-    });
-  }
-
-  /**
-   * Hand the download the priority map for one file.
-   *
-   * Seconds of film against a number, and nothing else: what is wanted and in
-   * what order. Turning that into bytes, into pieces, and into what the swarm
-   * is asked for is the download's own business, and so is what it keeps in
-   * memory.
-   *
-   * @param {{ sourceKey: string, fileIndex: number, durationSeconds: number, zones: { from: number, to: number, priority: number }[] }} params
-   * @returns {Promise<void>}
-   */
-  async setPriorityMap({ sourceKey, fileIndex, zones, durationSeconds }) {
-    await this.#caller.call(Command.PRIORITY_MAP, {
-      sourceKey,
-      fileIndex,
-      zones,
-      durationSeconds
-    });
-  }
-
-  /**
-   * Pre-fetch the head and tail the codec probe needs.
-   *
-   * @param {{ sourceKey: string, fileIndex: number, options?: { headBytes?: number, tailBytes?: number, timeoutMs?: number, awaited?: boolean } }} params
-   * @returns {Promise<unknown>}
-   */
-  async prefetchFileEdges({ sourceKey, fileIndex, options = {} }) {
-    return this.#caller.call(Command.PREFETCH_EDGES, { sourceKey, fileIndex, options });
-  }
-
-  /**
-   * Read a byte range as a stream.
-   *
-   * Returns immediately with a stream that fills as chunks arrive; cancelling it
-   * (viewer gone, seek superseded) stops the worker reading, so pieces are not
-   * fetched for a stream nobody will drain.
-   *
-   * @param {{ sourceKey: string, fileIndex: number, start?: number | null, end?: number | null, windowBytes?: number }} params
-   * @returns {ReadableStream<Uint8Array>}
-   */
-  createReadStream({ sourceKey, fileIndex, start = null, end = null, windowBytes }) {
-    // Same id sequence as commands — see `nextId` in `channel.js`.
-    const readId = this.#caller.nextId();
-    const receive = createReceiveStream({
-      port: this.#worker,
-      requestId: readId,
-      onCancel: () => {
-        void this.#caller.call(Command.CANCEL_READ, { readId }).catch(() => undefined);
-        this.#reads.delete(readId);
-        this.#lastPieceByRead.delete(readId);
-      }
-    });
-    this.#reads.set(readId, receive);
-
-    // The worker replies to READ_RANGE only once the body is fully sent; a
-    // failure before that must surface on the stream, not vanish.
-    this.#worker.postMessage({
-      command: Command.READ_RANGE,
-      id: readId,
-      params: { sourceKey, fileIndex, start, end, windowBytes }
-    });
-
-    return receive.stream;
-  }
-
-  /**
-   * Read a byte range as fragments of shared memory, without copying.
-   *
-   * Each fragment is a view straight into the torrent's piece pool, and the
-   * piece behind it stays pinned until `release()` is called — so the consumer
-   * must call it once it is genuinely finished with those bytes. For a response
-   * body that means after the socket write has completed, not when it was
-   * queued: verified that writing a shared-memory view and then overwriting the
-   * pool from the write callback leaves the client's copy intact, and that
-   * overwriting it earlier corrupts it silently.
-   *
-   * Returns `null` when this source has no shared pool, so the caller can fall
-   * back to {@link createReadStream}.
-   *
-   * @param {{ sourceKey: string, fileIndex: number, start?: number | null, end?: number | null, windowBytes?: number }} params
-   * @returns {{ [Symbol.asyncIterator]: () => AsyncGenerator<{ bytes: Uint8Array, release: () => void }>, cancel: () => void } | null}
-   */
-  createFragmentReader({ sourceKey, fileIndex, start = null, end = null, windowBytes }) {
-    const readId = this.#caller.nextId();
-    /** @type {{ bytes: Uint8Array, release: () => void }[]} */
-    const queue = [];
-    let wake = null;
-    let finished = false;
-    let failure = null;
-
-    const notify = () => {
-      const resume = wake;
-      wake = null;
-      resume?.();
-    };
-
-    this.#fragmentReaders.set(readId, {
-      push(bytes, confirm) {
-        queue.push({ bytes, release: confirm });
-        notify();
-      },
-      close() {
-        finished = true;
-        notify();
-      },
-      fail(error) {
-        failure = error;
-        finished = true;
-        notify();
-      }
-    });
-
-    const cancel = () => {
-      if (this.#fragmentReaders.delete(readId)) {
-        void this.#caller.call(Command.CANCEL_READ, { readId }).catch(() => undefined);
-      }
-      finished = true;
-      notify();
-    };
-
-    this.#worker.postMessage({
-      command: Command.READ_RANGE,
-      id: readId,
-      params: { sourceKey, fileIndex, start, end, windowBytes }
-    });
-
-    return {
-      cancel,
-      async *[Symbol.asyncIterator]() {
-        try {
-          for (;;) {
-            while (queue.length > 0) {
-              yield queue.shift();
-            }
-            if (failure) {
-              throw failure;
-            }
-            if (finished) {
-              return;
-            }
-            await new Promise((resolve) => {
-              wake = resolve;
-            });
-          }
-        } finally {
-          // Covers the consumer breaking out early — a client that hung up, a
-          // superseded seek — which must stop the read rather than leave it
-          // fetching pieces nobody will take.
-          if (!finished) {
-            cancel();
-          }
-        }
-      }
-    };
-  }
-
-  /**
-   * A stand-in for the WebTorrent torrent object, backed by the worker.
-   *
-   * Callers already hold a torrent and reach into `torrent.files[i]` — for the
-   * length, the name, or a read stream. Handing back an object of the same
-   * shape keeps every one of those call sites working unchanged, which matters:
-   * they are spread across the stream route, the subtitle route, the playback
-   * planner and the health report, and rewriting all of them to thread a
-   * `sourceKey` through would be a large change with nothing to show for it.
-   *
-   * Only what is actually used is provided. Anything else would be a promise we
-   * cannot keep — the real object lives on the other thread and its methods are
-   * not reachable from here.
-   *
-   * @param {{ sourceKey: string, sourceType: "magnet" | "torrent", source: string }} params
-   * @returns {Promise<{ infoHash: string, name: string, sourceKey: string, files: object[] }>}
-   */
-  async getTorrent({ sourceKey, sourceType, source }) {
-    const info = await this.addSource({ sourceKey, sourceType, source });
-    const client = this;
-    return {
-      infoHash: info.infoHash,
-      name: info.name,
-      // Carried so helpers that receive only the torrent can still name it to
-      // the worker.
-      sourceKey,
-      files: info.files.map((file) => ({
-        ...file,
-        /**
-         * @param {{ start?: number, end?: number, windowBytes?: number }} [options]
-         * @returns {ReadableStream<Uint8Array>}
-         */
-        createReadStream(options = {}) {
-          // Node stream, not a web one: Fastify replies and the ffmpeg pipe
-          // both expect that shape, and every existing call site passes the
-          // result straight to one of them. `Readable.fromWeb` adds no copy —
-          // it wraps the same buffers.
-          return Readable.fromWeb(
-            client.createReadStream({
-              sourceKey,
-              fileIndex: file.index,
-              start: options.start ?? null,
-              end: options.end ?? null,
-              windowBytes: options.windowBytes
-            })
-          );
-        },
-
-        /**
-         * Fragments of shared memory, for a caller that can say when it has
-         * finished with each one. `null` when this source has no shared pool.
-         *
-         * @param {{ start?: number, end?: number, windowBytes?: number }} [options]
-         * @returns {ReturnType<TorrentWorkerClient["createFragmentReader"]>}
-         */
-        createFragmentReader(options = {}) {
-          return client.createFragmentReader({
-            sourceKey,
-            fileIndex: file.index,
-            start: options.start ?? null,
-            end: options.end ?? null,
-            windowBytes: options.windowBytes
-          });
-        }
-      }))
-    };
-  }
-
-  /**
-   * Shut the torrent client down and stop the thread.
-   *
-   * @returns {Promise<void>}
-   */
-  /**
-   * What the spilled pieces weigh, as of the last revision. Zero until one has
-   * happened, which is what "we have not asked yet" means.
-   *
-   * @type {number}
-   */
-  spilledBytes = 0;
-
-  /**
-   * Tell the piece stores on the torrent thread their share of MEMORY.
-   *
-   * The same owner divides memory and disk, because the two trade: pieces that
-   * do not fit in memory are spilled to disk, so what the stores are given here
-   * decides how much disk they need there.
-   *
-   * @param {number} bytes
-   * @returns {Promise<{ held: number, wanted: number }>}
-   */
-  async allowMemoryBytes(bytes) {
-    try {
-      const claim = await this.#caller.call(Command.MEMORY_ALLOWANCE, { bytes });
-      this.memoryClaim = {
-        held: Number(claim?.held) || 0,
-        wanted: Number(claim?.wanted) || 0
-      };
-    } catch {
-      // The thread is gone or busy; the next revision says it again.
-    }
-    return this.memoryClaim ?? { held: 0, wanted: 0 };
-  }
-
-  /**
-   * Tell the whole files on the torrent thread their share of the disk.
-   *
-   * Same shape as the spilled pieces, and for the same reason: they hold bytes
-   * on this machine and the owner of the disk is on the other thread.
-   *
-   * @param {number} bytes
-   * @returns {Promise<number>} What they hold after it.
-   */
-  async allowWholeFileBytes(bytes) {
-    try {
-      this.wholeFileBytes = Number(await this.#caller.call(Command.WHOLE_FILES_ALLOWANCE, { bytes })) || 0;
-    } catch {
-      // The thread is gone or busy; the next revision says it again.
-    }
-    return this.wholeFileBytes ?? 0;
-  }
-
-  /**
-   * Say how much disk the spilled pieces on this thread may take between them.
-   *
-   * Best effort and never awaited by anything that matters: a share that does
-   * not arrive leaves the ceiling where it was, and the next revision is a
-   * minute away.
-   *
-   * @param {number} bytes
-   * @returns {Promise<void>}
-   */
-  async allowSpillBytes(bytes) {
-    try {
-      const revised = await this.#caller.call(Command.SPILL_ALLOWANCE, { bytes });
-      // The reply says what those stores actually hold, which is what the owner
-      // of the disk needs for the next division. One exchange, both directions.
-      this.spilledBytes = Array.isArray(revised)
-        ? revised.reduce((sum, store) => sum + (Number(store?.bytes) || 0), 0)
-        : this.spilledBytes;
-    } catch {
-      // The thread is gone or busy; the next revision says it again.
-    }
-    return this.spilledBytes;
-  }
-
-  async destroyAll() {
-    this.#stopping = true;
-    try {
-      await this.#caller.call(Command.DESTROY_ALL, {});
-    } catch {
-      // Already gone — the wait below settles either way.
-    }
-    // Let the thread END rather than tearing it down under itself.
-    //
-    // `terminate()` frees the environment immediately, with the handle
-    // callbacks libuv still holds queued. One of those is utp-native's UDP
-    // read, and running it against a freed isolate is what the core dump of
-    // 2026-08-21 caught: `on_utp_accept` → `napi_get_buffer_info` →
-    // `v8::Value::IsArrayBufferView` inside `Environment::CleanupHandles`.
-    // Once the client inside is destroyed the thread has nothing left holding
-    // its loop open, so it exits by itself and the callbacks drain first.
-    //
-    // `terminate()` stays as the bounded fallback, because a shutdown that
-    // hangs is worse than one that is forced.
-    await new Promise((resolve) => {
-      const timer = setTimeout(() => {
-        logger.warn("torrent-worker: thread did not end in 5s; terminating it");
-        void this.#worker.terminate().finally(resolve);
-      }, WORKER_EXIT_GRACE_MS);
-      this.#worker.once("exit", () => {
-        clearTimeout(timer);
-        resolve();
-      });
-    });
-  }
-}
+/**
+ * @file Main-thread face of the torrent worker.
+ *
+ * Presents the same operations the routes and session manager already use, so
+ * moving the torrent to its own thread does not ripple through calling code.
+ * The one unavoidable change is that torrents are named by `sourceKey` instead
+ * of passed around as objects — objects cannot cross a thread boundary, and
+ * pretending otherwise would mean copying them on every call.
+ *
+ * Reads come back as an ordinary `ReadableStream`, so `/stream` and the codec
+ * probe consume them exactly as they consume WebTorrent's own streams today.
+ * What that hides is the part that matters: chunks arrive as transferred
+ * buffers, never copied — 5.3 ms per 10 MB against 37 ms if cloned and 104 ms
+ * through a transferable stream (measured 2026-08-02, see `protocol.js`).
+ */
+
+import { Worker } from "node:worker_threads";
+import { Readable } from "node:stream";
+import { fileURLToPath } from "node:url";
+import { logger, writeAlreadyDecided } from "../../../utils/logger.js";
+import { createCaller, createReceiveStream } from "./channel.js";
+import { Command, Event } from "./protocol.js";
+
+const WORKER_URL = new URL("./worker.js", import.meta.url);
+/**
+ * How long a shutdown waits for the thread to end by itself before forcing it.
+ *
+ * Ending by itself is what lets libuv drain the handle callbacks it is holding;
+ * forcing it is what ran one of those against a freed isolate. Five seconds is
+ * long enough for a destroyed torrent client to release its sockets and short
+ * enough that a shutdown never appears to hang.
+ */
+const WORKER_EXIT_GRACE_MS = 5_000;
+
+/**
+ * Runs the torrent client on its own thread and exposes it to the main thread.
+ *
+ * Why this exists at all: profiling during a live seek found the main thread
+ * ~85% busy with WebTorrent (buffer concatenation ~15%, wire updates ~9%,
+ * garbage collection ~5%), while three of four cores idled. Serving a segment
+ * queued behind that work, so reading a finished 10 MB file took 12-23 s where
+ * handing it to the channel took 125 ms.
+ */
+/**
+ * Piece buffers this thread has been handed, and how many the collector has
+ * taken back.
+ *
+ * A `SharedArrayBuffer`'s memory belongs to neither isolate: it lives until
+ * BOTH have let go. So the worker's own count answers only half the question,
+ * and this is the other half — 700 MB grew in the field on 2026-08-31 with the
+ * store's own memory falling, and the two candidates are "a reference of ours
+ * outlives the piece" and "the allocator keeps what we free". These two
+ * counters separate them (roadmap item 2).
+ *
+ * A fragment arrives as a fresh handle onto the same shared memory every time,
+ * so what is counted is handles, not pieces. That is the right quantity anyway:
+ * one handle retained keeps the whole piece alive.
+ */
+const fragmentBuffers = { seen: 0, collected: 0 };
+const fragmentCollector = typeof FinalizationRegistry === "function"
+  ? new FinalizationRegistry(() => {
+    fragmentBuffers.collected += 1;
+  })
+  : null;
+
+/**
+ * What the collector has taken back against what this thread was handed.
+ *
+ * @returns {{ seen: number, collected: number }}
+ */
+export function fragmentBufferCollection() {
+  return { seen: fragmentBuffers.seen, collected: fragmentBuffers.collected };
+}
+
+export class TorrentWorkerClient {
+  #worker;
+  /**
+   * Whether this shutdown was asked for, so the thread ending can be told from
+   * the thread dying. Without it both look identical from outside, which is how
+   * five crashes produced no line in the log.
+   */
+  #stopping = false;
+  #caller;
+  /** Receive-side handles for in-flight reads, keyed by request id. */
+  #reads = new Map();
+
+  /**
+   * The last piece served to each open read, so a fragment arriving out of
+   * order can be named. Cleared when the read ends.
+   *
+   * @type {Map<number, number>}
+   */
+   #lastPieceByRead = new Map();
+  /** Reads consuming fragments in place, keyed by request id. */
+  #fragmentReaders = new Map();
+
+  /**
+   * Files downloaded whole, by `${infoHash}/${fileIndex}`.
+   *
+   * Held on THIS thread because it is this thread that serves them: a file that
+   * is whole is read with an ordinary file read, and asking the torrent thread
+   * for anything about it would put back the very hop this removes. The torrent
+   * thread writes the files and says so; this is what it says.
+   *
+   * @type {Map<string, { path: string, length: number, name: string }>}
+   */
+  wholeFiles = new Map();
+
+  /** @type {(event: { sourceKey: string, fileIndex: number, trackIndex: number, cues: object[], language: string }) => void} */
+  /** @type {(event: { sourceKey: string, fileIndexes: number[] }) => void} */
+  #onPiecesArrived;
+
+  /**
+   * @param {{ memoryBytes?: number, stateDir?: string, onPiecesArrived?: (event: object) => void }} [options]
+   */
+  constructor({ memoryBytes, stateDir, onPiecesArrived } = {}) {
+    this.#worker = new Worker(fileURLToPath(WORKER_URL), {
+      // `stateDir` travels because the worker writes heap snapshots of its own
+      // isolate there. It cannot choose a directory any other way: a worker may
+      // not change the process's working directory, and the isolate that has
+      // died three times is the one no snapshot has ever been taken of.
+      workerData: { memoryBytes, stateDir }
+    });
+    this.#caller = createCaller(this.#worker);
+    this.#onPiecesArrived = onPiecesArrived ?? (() => undefined);
+
+    this.#worker.on("message", (message) => {
+      // A failed read must fail its stream. This is checked BEFORE the caller
+      // sees the message: until 2.9.76 nothing here handled a read error at
+      // all, so the worker's report was dropped as unknown, and because the
+      // worker sent the end-of-read marker from its `finally` even when the
+      // read had thrown, the reader saw a clean end of file instead. A read
+      // that failed before it produced anything simply hung forever.
+      if (message?.type === Event.ERROR && this.#reads.has(message.id)) {
+        const read = this.#reads.get(message.id);
+        this.#reads.delete(message.id);
+        read.fail(new Error(message.error ?? "Torrent worker read failed."));
+        return;
+      }
+      if (message?.type === Event.ERROR && this.#fragmentReaders.has(message.id)) {
+        const reader = this.#fragmentReaders.get(message.id);
+        this.#fragmentReaders.delete(message.id);
+        reader.fail(new Error(message.error ?? "Torrent worker read failed."));
+        return;
+      }
+      if (this.#caller.handleReply(message)) {
+        return;
+      }
+      switch (message?.type) {
+        case Event.FRAGMENT: {
+          const buffer = message.buffer;
+          if (!buffer) {
+            // No buffer means no way to read the fragment; confirm it so the
+            // worker is not left waiting, and let the read end short.
+            this.#worker.postMessage({ type: Event.FRAGMENT_DONE, id: message.id });
+            break;
+          }
+          if (message.offset + message.length > buffer.byteLength) {
+            logger.warn(
+              `torrent-worker: fragment ${message.offset}+${message.length} lies outside ` +
+              `its buffer of ${buffer.byteLength}B (read ${message.id}) — ending the read short`
+            );
+            this.#worker.postMessage({ type: Event.FRAGMENT_DONE, id: message.id });
+            break;
+          }
+          // A sequential read walks the file forwards, so each fragment either
+          // continues the piece before it or moves to the very next one.
+          // Anything else means the bytes handed to the decoder are not the
+          // file's bytes in order — which is what a decoder complaining about
+          // its input has twice turned out to mean (2.9.126, and the AC-3
+          // failure of 2026-08-09: the encoder ran at 9.3x, the piece store
+          // reported no spills and 100% of reads from memory, and the decoder
+          // still saw "new coupling strategy must be present in block 0"). The
+          // bounds check catches a fragment outside the pool; this catches one
+          // inside it that belongs somewhere else.
+          const lastPiece = this.#lastPieceByRead.get(message.id);
+          if (lastPiece !== undefined && message.pieceIndex !== lastPiece && message.pieceIndex !== lastPiece + 1) {
+            logger.warn(
+              `torrent-worker: read ${message.id} jumped from piece ${lastPiece} to ` +
+              `${message.pieceIndex} (${message.length}B at pool offset ${message.offset}) — ` +
+              "the consumer is being handed the file out of order"
+            );
+          }
+          this.#lastPieceByRead.set(message.id, message.pieceIndex);
+          fragmentBuffers.seen += 1;
+          fragmentCollector?.register(buffer, null);
+          const view = new Uint8Array(buffer, message.offset, message.length);
+
+          const reader = this.#fragmentReaders.get(message.id);
+          if (reader) {
+            // Handed on as a view into the pool — no copy anywhere. The piece
+            // stays pinned until the consumer says it is done with these exact
+            // bytes, which for a response body means the socket write has
+            // completed.
+            reader.push(view, () => {
+              this.#worker.postMessage({ type: Event.FRAGMENT_DONE, id: message.id });
+            });
+            break;
+          }
+
+          // Plain-stream consumers keep what they are given while the slot may
+          // be reused, so they get a copy and the piece is released at once.
+          this.#reads.get(message.id)?.push(Uint8Array.prototype.slice.call(view));
+          this.#worker.postMessage({ type: Event.FRAGMENT_DONE, id: message.id });
+          break;
+        }
+        case Event.CHUNK: {
+          const bytes = message.bytes;
+          this.#reads.get(message.id)?.push(
+            new Uint8Array(bytes.buffer, bytes.byteOffset, bytes.length)
+          );
+          break;
+        }
+        case Event.READ_END:
+          this.#lastPieceByRead.delete(message.id);
+          this.#reads.get(message.id)?.close();
+          this.#reads.delete(message.id);
+          this.#fragmentReaders.get(message.id)?.close();
+          this.#fragmentReaders.delete(message.id);
+          break;
+        case Event.LOG:
+          // Written as it is: the worker holds the same repeat rule and has
+          // already applied it, and deciding again here would be one decision
+          // taken twice on two different histories.
+          writeAlreadyDecided(message.level ?? "info", `torrent-worker: ${message.message}`);
+          break;
+        case Event.FILE_COMPLETE:
+          // Recorded here so the stream route can answer from the file without
+          // going near the torrent thread at all.
+          this.wholeFiles.set(`${message.infoHash}/${message.fileIndex}`, {
+            path: message.path,
+            length: message.length,
+            name: message.name
+          });
+          break;
+        case Event.PIECES_ARRIVED:
+          // ANNOUNCED by the thread that owns the swarm, acted on here: the
+          // subtitle walk is this thread's and needs to know a piece arrived.
+          this.#onPiecesArrived({
+            sourceKey: message.sourceKey,
+            fileIndexes: Array.isArray(message.fileIndexes) ? message.fileIndexes : []
+          });
+          break;
+        default:
+          break;
+      }
+    });
+
+    this.#worker.on("error", (error) => {
+      logger.error(`torrent-worker crashed: ${error?.message ?? error}`);
+      // Fail everything outstanding rather than leaving callers hanging: a dead
+      // worker will never answer, and a stalled request is worse than an error
+      // the loading flow can retry.
+      this.#failEverythingOutstanding(new Error("Torrent worker stopped unexpectedly."));
+    });
+
+    // A worker that ENDS was, until now, not noticed at all: only `message` and
+    // `error` were listened for. So when the thread went away the proxy simply
+    // stopped, the log ended mid-sentence, and nothing said whether we had
+    // asked for it — which is precisely the reading that was missing on
+    // 2026-08-21, when a core dump showed the thread faulting inside
+    // `Environment::CleanupHandles` and there was no way to tell our own
+    // shutdown from the thread ending on its own.
+    this.#worker.on("exit", (code) => {
+      if (this.#stopping) {
+        logger.info(`torrent-worker: thread ended as asked (code ${code})`);
+        return;
+      }
+      logger.error(
+        `torrent-worker: thread ended on its own with code ${code} — nobody asked it to. ` +
+        "Everything waiting on it is failed; the proxy has no torrent client until it is rebuilt."
+      );
+      this.#failEverythingOutstanding(new Error("Torrent worker ended unexpectedly."));
+    });
+  }
+
+  /**
+   * A dead thread will never answer, so everything waiting on it is failed.
+   *
+   * INCLUDING THE FRAGMENT READS, which is what ffmpeg's input is. Both death
+   * handlers walked `#reads` alone and left `#fragmentReaders` untouched, so an
+   * encoder's input neither ended nor errored — it went quiet. Field 2026-08-31,
+   * three times: the thread died, both ffmpeg runs stayed ALIVE and stopped
+   * producing (167 `holding segment-00085.mp4 … encoder alive` lines), the
+   * browser drained its cushion and then retried a segment that would never
+   * exist, for ever, with nothing shown to the viewer.
+   *
+   * @param {Error} reason
+   */
+  #failEverythingOutstanding(reason) {
+    this.#caller.rejectAll(reason);
+    for (const [, read] of this.#reads) {
+      read.fail(reason);
+    }
+    this.#reads.clear();
+    for (const [, reader] of this.#fragmentReaders) {
+      reader.fail(reason);
+    }
+    this.#fragmentReaders.clear();
+  }
+
+  /**
+   * Add (or join) a torrent and register it under `sourceKey`.
+   *
+   * @param {{ sourceKey: string, sourceType: "magnet" | "torrent", source: string }} params
+   * @returns {Promise<{ infoHash: string, name: string, files: { index: number, name: string, path: string, length: number }[] }>}
+   */
+  async addSource({ sourceKey, sourceType, source }) {
+    return this.#caller.call(Command.ADD_SOURCE, { sourceKey, sourceType, source });
+  }
+
+  /**
+   * The torrent's files, as plain data.
+   *
+   * @param {string} sourceKey
+   * @returns {Promise<{ index: number, name: string, path: string, length: number }[]>}
+   */
+  async listFiles(sourceKey) {
+    return this.#caller.call(Command.LIST_FILES, { sourceKey });
+  }
+
+  /**
+   * Live download figures for the progress display.
+   *
+   * @param {{ sourceKey: string, fileIndex: number, resumeAnchorByteStart?: number | null }} params
+   * @returns {Promise<object>}
+   */
+  async getFileStats({ sourceKey, fileIndex, resumeAnchorByteStart = null }) {
+    return this.#caller.call(Command.FILE_STATS, { sourceKey, fileIndex, resumeAnchorByteStart });
+  }
+
+  /**
+   * The text subtitle tracks a file carries.
+   *
+   * @param {{ sourceKey: string, fileIndex: number }} params
+   * @returns {Promise<{ tracks: object[] }>}
+   */
+
+
+
+
+  /**
+   * Start fetching the region a viewer is about to resume at.
+   *
+   * @param {{ sourceKey: string, fileIndex: number, positionSeconds: number }} params
+   * @returns {Promise<{ started: boolean }>}
+   */
+  /**
+   * Fetch one whole file in the room the viewer's own reading leaves.
+   *
+   * @param {{ sourceKey: string, fileIndex: number }} params
+   * @returns {Promise<{ started: boolean }>}
+   */
+  async fillFile({ sourceKey, fileIndex }) {
+    return this.#caller.call(Command.FILL_FILE, { sourceKey, fileIndex });
+  }
+
+  /**
+   * The byte ranges of one file the torrent holds WHOLE.
+   *
+   * @param {{ sourceKey: string, fileIndex: number }} params
+   * @returns {Promise<Array<[number, number]>>}
+   */
+  async heldRanges({ sourceKey, fileIndex }) {
+    const answer = await this.#caller.call(Command.HELD_RANGES, { sourceKey, fileIndex });
+    return Array.isArray(answer?.ranges) ? answer.ranges : [];
+  }
+
+  /**
+   * Bytes of a range the torrent already holds. Never fetched: null where they
+   * are not there.
+   *
+   * @param {{ sourceKey: string, fileIndex: number, start: number, end: number }} params
+   * @returns {Promise<Buffer | null>}
+   */
+  async readHeld({ sourceKey, fileIndex, start, end }) {
+    const answer = await this.#caller.call(Command.READ_HELD, { sourceKey, fileIndex, start, end });
+    return answer?.bytes ? Buffer.from(answer.bytes) : null;
+  }
+
+  async warmResumePosition({ sourceKey, fileIndex, positionSeconds, durationSeconds }) {
+    return this.#caller.call(Command.WARM_POSITION, { sourceKey, fileIndex, positionSeconds, durationSeconds });
+  }
+
+
+  /**
+   * Bytes every torrent on the worker has moved, downloaded and uploaded apart.
+   *
+   * The client lives on the worker thread, so this cannot be read as a property
+   * from here — which is exactly the mistake 2.19.0 shipped, leaving the figure
+   * always zero and the whole feature inert.
+   *
+   * @returns {Promise<{ downloaded: number, uploaded: number }>}
+   */
+  async getTorrentTotals() {
+    return this.#caller.call(Command.TORRENT_TOTALS, {});
+  }
+
+  /**
+   * Which films this proxy holds right now, and how much of each.
+   *
+   * Asked of the worker rather than of the main thread's map of stand-ins: that
+   * map is only cleared on shutdown, so it names films this proxy let go of
+   * hours ago.
+   *
+   * @returns {Promise<{ held: { infoHash: string, progress: number, bytes: number }[] }>}
+   */
+  async heldTorrents() {
+    return this.#caller.call(Command.HELD_TORRENTS, {});
+  }
+
+  /** State conditional demand for every file in a selected torrent. */
+  async fillTorrent(sourceKey) {
+    return this.#caller.call(Command.FILL_TORRENT, { sourceKey });
+  }
+
+  /**
+   * Reorder piece selection around a read position (seek prioritisation).
+   *
+   * @param {{ sourceKey: string, fileIndex: number, byteStart: number, windowBytes?: number, wholeFileRead?: boolean }} params
+   * @returns {Promise<void>}
+   */
+  async prioritizeByteRange({ sourceKey, fileIndex, byteStart, windowBytes, wholeFileRead }) {
+    await this.#caller.call(Command.PRIORITIZE, {
+      sourceKey,
+      fileIndex,
+      byteStart,
+      windowBytes,
+      wholeFileRead
+    });
+  }
+
+  /**
+   * Hand the download the priority map for one file.
+   *
+   * Seconds of film against a number, and nothing else: what is wanted and in
+   * what order. Turning that into bytes, into pieces, and into what the swarm
+   * is asked for is the download's own business, and so is what it keeps in
+   * memory.
+   *
+   * @param {{ sourceKey: string, fileIndex: number, durationSeconds: number, zones: { from: number, to: number, priority: number }[] }} params
+   * @returns {Promise<void>}
+   */
+  async setPriorityMap({ sourceKey, fileIndex, zones, durationSeconds }) {
+    await this.#caller.call(Command.PRIORITY_MAP, {
+      sourceKey,
+      fileIndex,
+      zones,
+      durationSeconds
+    });
+  }
+
+  /**
+   * Pre-fetch the head and tail the codec probe needs.
+   *
+   * @param {{ sourceKey: string, fileIndex: number, options?: { headBytes?: number, tailBytes?: number, timeoutMs?: number, awaited?: boolean } }} params
+   * @returns {Promise<unknown>}
+   */
+  async prefetchFileEdges({ sourceKey, fileIndex, options = {} }) {
+    return this.#caller.call(Command.PREFETCH_EDGES, { sourceKey, fileIndex, options });
+  }
+
+  /**
+   * Read a byte range as a stream.
+   *
+   * Returns immediately with a stream that fills as chunks arrive; cancelling it
+   * (viewer gone, seek superseded) stops the worker reading, so pieces are not
+   * fetched for a stream nobody will drain.
+   *
+   * @param {{ sourceKey: string, fileIndex: number, start?: number | null, end?: number | null, windowBytes?: number }} params
+   * @returns {ReadableStream<Uint8Array>}
+   */
+  createReadStream({ sourceKey, fileIndex, start = null, end = null, windowBytes }) {
+    // Same id sequence as commands — see `nextId` in `channel.js`.
+    const readId = this.#caller.nextId();
+    const receive = createReceiveStream({
+      port: this.#worker,
+      requestId: readId,
+      onCancel: () => {
+        void this.#caller.call(Command.CANCEL_READ, { readId }).catch(() => undefined);
+        this.#reads.delete(readId);
+        this.#lastPieceByRead.delete(readId);
+      }
+    });
+    this.#reads.set(readId, receive);
+
+    // The worker replies to READ_RANGE only once the body is fully sent; a
+    // failure before that must surface on the stream, not vanish.
+    this.#worker.postMessage({
+      command: Command.READ_RANGE,
+      id: readId,
+      params: { sourceKey, fileIndex, start, end, windowBytes }
+    });
+
+    return receive.stream;
+  }
+
+  /**
+   * Read a byte range as fragments of shared memory, without copying.
+   *
+   * Each fragment is a view straight into the torrent's piece pool, and the
+   * piece behind it stays pinned until `release()` is called — so the consumer
+   * must call it once it is genuinely finished with those bytes. For a response
+   * body that means after the socket write has completed, not when it was
+   * queued: verified that writing a shared-memory view and then overwriting the
+   * pool from the write callback leaves the client's copy intact, and that
+   * overwriting it earlier corrupts it silently.
+   *
+   * Returns `null` when this source has no shared pool, so the caller can fall
+   * back to {@link createReadStream}.
+   *
+   * @param {{ sourceKey: string, fileIndex: number, start?: number | null, end?: number | null, windowBytes?: number }} params
+   * @returns {{ [Symbol.asyncIterator]: () => AsyncGenerator<{ bytes: Uint8Array, release: () => void }>, cancel: () => void } | null}
+   */
+  createFragmentReader({ sourceKey, fileIndex, start = null, end = null, windowBytes }) {
+    const readId = this.#caller.nextId();
+    /** @type {{ bytes: Uint8Array, release: () => void }[]} */
+    const queue = [];
+    let wake = null;
+    let finished = false;
+    let failure = null;
+
+    const notify = () => {
+      const resume = wake;
+      wake = null;
+      resume?.();
+    };
+
+    this.#fragmentReaders.set(readId, {
+      push(bytes, confirm) {
+        queue.push({ bytes, release: confirm });
+        notify();
+      },
+      close() {
+        finished = true;
+        notify();
+      },
+      fail(error) {
+        failure = error;
+        finished = true;
+        notify();
+      }
+    });
+
+    const cancel = () => {
+      if (this.#fragmentReaders.delete(readId)) {
+        void this.#caller.call(Command.CANCEL_READ, { readId }).catch(() => undefined);
+      }
+      finished = true;
+      notify();
+    };
+
+    this.#worker.postMessage({
+      command: Command.READ_RANGE,
+      id: readId,
+      params: { sourceKey, fileIndex, start, end, windowBytes }
+    });
+
+    return {
+      cancel,
+      async *[Symbol.asyncIterator]() {
+        try {
+          for (;;) {
+            while (queue.length > 0) {
+              yield queue.shift();
+            }
+            if (failure) {
+              throw failure;
+            }
+            if (finished) {
+              return;
+            }
+            await new Promise((resolve) => {
+              wake = resolve;
+            });
+          }
+        } finally {
+          // Covers the consumer breaking out early — a client that hung up, a
+          // superseded seek — which must stop the read rather than leave it
+          // fetching pieces nobody will take.
+          if (!finished) {
+            cancel();
+          }
+        }
+      }
+    };
+  }
+
+  /**
+   * A stand-in for the WebTorrent torrent object, backed by the worker.
+   *
+   * Callers already hold a torrent and reach into `torrent.files[i]` — for the
+   * length, the name, or a read stream. Handing back an object of the same
+   * shape keeps every one of those call sites working unchanged, which matters:
+   * they are spread across the stream route, the subtitle route, the playback
+   * planner and the health report, and rewriting all of them to thread a
+   * `sourceKey` through would be a large change with nothing to show for it.
+   *
+   * Only what is actually used is provided. Anything else would be a promise we
+   * cannot keep — the real object lives on the other thread and its methods are
+   * not reachable from here.
+   *
+   * @param {{ sourceKey: string, sourceType: "magnet" | "torrent", source: string }} params
+   * @returns {Promise<{ infoHash: string, name: string, sourceKey: string, files: object[] }>}
+   */
+  async getTorrent({ sourceKey, sourceType, source }) {
+    const info = await this.addSource({ sourceKey, sourceType, source });
+    const client = this;
+    return {
+      infoHash: info.infoHash,
+      name: info.name,
+      // Carried so helpers that receive only the torrent can still name it to
+      // the worker.
+      sourceKey,
+      files: info.files.map((file) => ({
+        ...file,
+        /**
+         * @param {{ start?: number, end?: number, windowBytes?: number }} [options]
+         * @returns {ReadableStream<Uint8Array>}
+         */
+        createReadStream(options = {}) {
+          // Node stream, not a web one: Fastify replies and the ffmpeg pipe
+          // both expect that shape, and every existing call site passes the
+          // result straight to one of them. `Readable.fromWeb` adds no copy —
+          // it wraps the same buffers.
+          return Readable.fromWeb(
+            client.createReadStream({
+              sourceKey,
+              fileIndex: file.index,
+              start: options.start ?? null,
+              end: options.end ?? null,
+              windowBytes: options.windowBytes
+            })
+          );
+        },
+
+        /**
+         * Fragments of shared memory, for a caller that can say when it has
+         * finished with each one. `null` when this source has no shared pool.
+         *
+         * @param {{ start?: number, end?: number, windowBytes?: number }} [options]
+         * @returns {ReturnType<TorrentWorkerClient["createFragmentReader"]>}
+         */
+        createFragmentReader(options = {}) {
+          return client.createFragmentReader({
+            sourceKey,
+            fileIndex: file.index,
+            start: options.start ?? null,
+            end: options.end ?? null,
+            windowBytes: options.windowBytes
+          });
+        }
+      }))
+    };
+  }
+
+  /**
+   * Shut the torrent client down and stop the thread.
+   *
+   * @returns {Promise<void>}
+   */
+  /**
+   * What the spilled pieces weigh, as of the last revision. Zero until one has
+   * happened, which is what "we have not asked yet" means.
+   *
+   * @type {number}
+   */
+  spilledBytes = 0;
+
+  /**
+   * Tell the piece stores on the torrent thread their share of MEMORY.
+   *
+   * The same owner divides memory and disk, because the two trade: pieces that
+   * do not fit in memory are spilled to disk, so what the stores are given here
+   * decides how much disk they need there.
+   *
+   * @param {number} bytes
+   * @returns {Promise<{ held: number, wanted: number }>}
+   */
+  async allowMemoryBytes(bytes) {
+    try {
+      const claim = await this.#caller.call(Command.MEMORY_ALLOWANCE, { bytes });
+      this.memoryClaim = {
+        held: Number(claim?.held) || 0,
+        wanted: Number(claim?.wanted) || 0
+      };
+    } catch {
+      // The thread is gone or busy; the next revision says it again.
+    }
+    return this.memoryClaim ?? { held: 0, wanted: 0 };
+  }
+
+  /**
+   * Tell the whole files on the torrent thread their share of the disk.
+   *
+   * Same shape as the spilled pieces, and for the same reason: they hold bytes
+   * on this machine and the owner of the disk is on the other thread.
+   *
+   * @param {number} bytes
+   * @returns {Promise<number>} What they hold after it.
+   */
+  async allowWholeFileBytes(bytes) {
+    try {
+      this.wholeFileBytes = Number(await this.#caller.call(Command.WHOLE_FILES_ALLOWANCE, { bytes })) || 0;
+    } catch {
+      // The thread is gone or busy; the next revision says it again.
+    }
+    return this.wholeFileBytes ?? 0;
+  }
+
+  /**
+   * Say how much disk the spilled pieces on this thread may take between them.
+   *
+   * Best effort and never awaited by anything that matters: a share that does
+   * not arrive leaves the ceiling where it was, and the next revision is a
+   * minute away.
+   *
+   * @param {number} bytes
+   * @returns {Promise<void>}
+   */
+  async allowSpillBytes(bytes) {
+    try {
+      const revised = await this.#caller.call(Command.SPILL_ALLOWANCE, { bytes });
+      // The reply says what those stores actually hold, which is what the owner
+      // of the disk needs for the next division. One exchange, both directions.
+      this.spilledBytes = Array.isArray(revised)
+        ? revised.reduce((sum, store) => sum + (Number(store?.bytes) || 0), 0)
+        : this.spilledBytes;
+    } catch {
+      // The thread is gone or busy; the next revision says it again.
+    }
+    return this.spilledBytes;
+  }
+
+  async destroyAll() {
+    this.#stopping = true;
+    try {
+      await this.#caller.call(Command.DESTROY_ALL, {});
+    } catch {
+      // Already gone — the wait below settles either way.
+    }
+    // Let the thread END rather than tearing it down under itself.
+    //
+    // `terminate()` frees the environment immediately, with the handle
+    // callbacks libuv still holds queued. One of those is utp-native's UDP
+    // read, and running it against a freed isolate is what the core dump of
+    // 2026-08-21 caught: `on_utp_accept` → `napi_get_buffer_info` →
+    // `v8::Value::IsArrayBufferView` inside `Environment::CleanupHandles`.
+    // Once the client inside is destroyed the thread has nothing left holding
+    // its loop open, so it exits by itself and the callbacks drain first.
+    //
+    // `terminate()` stays as the bounded fallback, because a shutdown that
+    // hangs is worse than one that is forced.
+    await new Promise((resolve) => {
+      const timer = setTimeout(() => {
+        logger.warn("torrent-worker: thread did not end in 5s; terminating it");
+        void this.#worker.terminate().finally(resolve);
+      }, WORKER_EXIT_GRACE_MS);
+      this.#worker.once("exit", () => {
+        clearTimeout(timer);
+        resolve();
+      });
+    });
+  }
+}