@torrent-tv/proxy 2.89.1 → 2.89.2

package/services/server/SegmentServing.js

@@ -1,1150 +1,1150 @@
-/**
- * Answering one request for a file of an output: a playlist, the init segment
- * or a segment.
- *
- * A segment is served from the store when it is made and whole; otherwise the
- * request is held until it is, or until nobody wants it any more, and a hold
- * that lasts says why. A request is an operation on material that exists; where
- * encoders work is the plan's, and a request does not move one.
- */
-
-import { createReadStream } from "node:fs";
-import { access, readFile, stat, unlink } from "node:fs/promises";
-import { Readable } from "node:stream";
-import path from "node:path";
-import { logger } from "../../utils/logger.js";
-import { isOutputName, PLAYLIST_FILE_NAME } from "../encode/output/index.js";
-// The index of variants. Served from the same route as the media playlist, so
-// it needs no path of its own.
-export const MASTER_PLAYLIST_FILE_NAME = "master.m3u8";
-// Read segment files in large blocks so the body is delivered to the data
-// channel in few, big chunks. On a busy ARM host the in-process WebTorrent
-// hashing starves the event loop in bursts, so fewer read iterations means
-// far less time lost between chunks while serving the first segments.
-const SEGMENT_READ_HIGH_WATER_MARK = 4 * 1024 * 1024;
-/**
- * Resolve after a given number of milliseconds.
- *
- * @param {number} ms
- * @returns {Promise<void>}
- */
-function delay(ms) {
-  return new Promise((resolve) => {
-    setTimeout(resolve, ms);
-  });
-}
-/**
- * Guard against path traversal by restricting file names to the known
- * playlist and segment patterns produced by ffmpeg. Which segment names are
- * legal depends on the active container, so the format decides.
- *
- * @param {string} fileName
- * @param {import("../encode/segment-formats/index.js").SegmentFormat} segmentFormat
- * @returns {boolean}
- */
-function isSafeFileName(fileName, segmentFormat) {
-  return (
-    fileName === PLAYLIST_FILE_NAME ||
-    fileName === MASTER_PLAYLIST_FILE_NAME ||
-    (segmentFormat.initFileName !== null && fileName === segmentFormat.initFileName) ||
-    segmentFormat.isSegmentFileName(fileName)
-  );
-}
-/**
- * Whether this output is cut at times we hand the muxer, rather than at a
- * duration it chooses for itself.
- *
- * A property of the output and not of a run: the cut grid and the branch decide
- * it, so every run of one output answers alike. It decides how a segment is
- * judged finished — see getFileStream.
- *
- * @param {object} session
- * @returns {boolean}
- */
-function cutsAtGivenTimes(session) {
-  const explicit = session?.segmentFormat?.explicitTimesMuxerArgs?.() ?? null;
-  if (!explicit) {
-    return false;
-  }
-  return !session.spec.transcodesVideo || session.timeline?.cutGrid === "keyframe";
-}
-
-export class SegmentServing {
-  /** Serving state keyed by the output object it belongs to. */
-  #states = new WeakMap();
-  /** What this reads and asks of the rest of the proxy, and nothing else. @type {object} */
-  #host;
-
-  /**
-   * @param {object} host - `buildMasterPlaylist`, `declaredTracks`, `publishedGridFor`, `runStartTimeFor`, `cushion`, `encodeOrchestrator`, `encodeRuns`, `hostTimings`, `lookaheadSeconds`, `outputTimes`, `outputs`, `segmentStore`, `startupWaitMs`, `viewers`
-   */
-  constructor(host) {
-    this.#host = host;
-  }
-
-  #stateFor(output) {
-    let state = this.#states.get(output);
-    if (!state) {
-      state = {
-        holdExplainedAt: new Map(),
-        waitEpoch: 0,
-        waitListeners: new Set()
-      };
-      this.#states.set(output, state);
-    }
-    return state;
-  }
-
-  invalidateWaits(output) {
-    const state = this.#stateFor(output);
-    state.waitEpoch += 1;
-    for (const wake of state.waitListeners) wake(true);
-    state.waitListeners.clear();
-  }
-
-  /**
-   * Open a read stream for an HLS segment or playlist file from a session.
-   *
-   * @param {string} sessionId
-   * @param {string} fileName - Must match the playlist or segment name pattern.
-   * @param {{ consumerId?: string }} [options] -
-   *   `consumerId` says WHICH viewer is asking, so a session
-   *   shared by several of them can tell their positions apart; absent from a
-   *   browser or a transport that does not carry it, and then everything falls
-   *   back to the one shared position.
-   * @returns {Promise<
-   *   | { kind: "not-found" }
-   *   | { kind: "warming-up" }
-   *   | { kind: "failed"; message: string }
-   *   | { kind: "file"; stream: import("node:fs").ReadStream; contentType: string; isPlaylist: boolean }
-   * >}
-   */
-  async getFileStream(sessionId, fileName, options = {}) {
-    const consumerId = typeof options.consumerId === "string" ? options.consumerId : "";
-    if (!isOutputName(sessionId)) {
-      return { kind: "not-found" };
-    }
-    const session = this.#host.outputs.get(sessionId);
-    // The session is looked up BEFORE the name is validated, because what
-    // counts as a valid segment name depends on the container this session
-    // chose — `.mp4` for fMP4, `.ts` for MPEG-TS.
-    if (!session || !isSafeFileName(fileName, session.segmentFormat)) {
-      return { kind: "not-found" };
-    }
-    // WHETHER PRODUCTION FAILED IS NOT ASKED HERE. It explains a piece that is
-    // ABSENT and says nothing about one that is on disk: a segment under its
-    // served name is whole by construction, whoever wrote it and whatever has
-    // become of the encoder since. Asked at the door, as it was, a failed run
-    // answered 500 to every request for material it had already finished — the
-    // viewer lost what was made as well as what was not, and a seek back into
-    // it could not be served either. It is asked where the answer is needed:
-    // where the file is not there (`#holdForProduction`, and the init's own
-    // absent branch).
-    this.#host.outputs.touch(session);
-
-    // The index of variants. Served from here rather than a route of its own,
-    // because to a player it is simply another playlist under the session.
-    if (fileName === MASTER_PLAYLIST_FILE_NAME) {
-      const masterText = this.#host.buildMasterPlaylist(sessionId, consumerId);
-      if (!masterText) {
-        return { kind: "not-found" };
-      }
-      return {
-        kind: "file",
-        stream: Readable.from([masterText]),
-        contentType: "application/vnd.apple.mpegurl",
-        isPlaylist: true
-      };
-    }
-
-    // Serve the synthetic VOD playlist (full duration, terminated with
-    // #EXT-X-ENDLIST) so the player gets the correct total length and a fully
-    // seekable timeline up-front, independent of how far ffmpeg has encoded.
-    if (fileName === PLAYLIST_FILE_NAME && session.useSyntheticPlaylist) {
-      return {
-        kind: "file",
-        stream: Readable.from([session.playlistText]),
-        contentType: "application/vnd.apple.mpegurl",
-        isPlaylist: true
-      };
-    }
-
-    // The init segment (fMP4 only; referenced by #EXT-X-MAP). Each seek-restart
-    // run REWRITES it, so cache the FIRST one and always serve that — the
-    // player fetches it once and never re-fetches, so it must stay stable for
-    // the session's lifetime. (What that costs, and why segments must therefore
-    // carry their own position, is documented in `encode/segment-formats/mp4-boxes.js`
-    // `stampSegmentStartTime`.)
-    //
-    // ffmpeg creates init.mp4 before it has finished writing the fMP4 header
-    // boxes into it (unlike segments, its write is not gated behind an atomic
-    // rename), so a read can race a moment where the file EXISTS but is still
-    // EMPTY. Root cause of a real incident: that empty read used to be cached
-    // as the output's init — a zero-length Buffer is still a truthy object,
-    // so a check on its presence treated it as "already resolved" and served
-    // the empty file for the rest of the session's life, permanently breaking
-    // playback (hls.js can never initialize its SourceBuffer from an empty
-    // init segment) while the transcode itself kept encoding normally. Guard
-    // on non-empty content on both the cache check and the fresh read, so an
-    // empty read is treated as not-yet-ready and the caller's long-poll keeps
-    // retrying until ffmpeg has actually written the header.
-    const { initFileName } = session.segmentFormat;
-    if (initFileName !== null && fileName === initFileName) {
-      const kept = this.#host.segmentStore.initOf(session.outputKey ?? "");
-      if (kept) {
-        return {
-          kind: "file",
-          stream: Readable.from([kept]),
-          contentType: session.segmentFormat.initContentType,
-          isPlaylist: false
-        };
-      }
-      try {
-        // With explicit cut times there is no init file: that muxer writes each
-        // piece self-contained, header and all. The header is identical in every
-        // piece, so the first one to exist supplies it.
-        const bytes = cutsAtGivenTimes(session)
-          ? await this.#initFromFirstSegment(session)
-          : await readFile(path.join(this.#host.segmentStore.pathFor(session.outputKey ?? ""), initFileName));
-        const standing = this.#host.segmentStore.keepInit(session.outputKey ?? "", bytes);
-        if (!standing) {
-          // Nothing usable yet. On the branch that lifts the header out of the
-          // first piece there is no file to be missing, so "not produced yet"
-          // arrives HERE and not in the catch below: `#initFromFirstSegment`
-          // answers null rather than throwing `ENOENT`. A production that has
-          // ended in an error is not going to produce one either, and saying
-          // "still warming up" to that holds the request for its whole deadline.
-          return this.#failedOrWarming(session);
-        }
-        return {
-          kind: "file",
-          stream: Readable.from([standing]),
-          contentType: session.segmentFormat.initContentType,
-          isPlaylist: false
-        };
-      } catch (error) {
-        if (error?.code === "ENOENT") {
-          // Not produced yet — the encode run started at session creation
-          // writes it early; the caller long-polls until it appears. Unless
-          // production has ended in an error, in which case nothing is going to
-          // write it and holding the request only spends the viewer's patience.
-          return this.#failedOrWarming(session);
-        }
-        logger.error(
-          `transcode ${session.id} could not serve ${initFileName}: ${error?.message ?? error}` +
-          (error?.stack ? `\n${error.stack}` : "")
-        );
-        return {
-          kind: "failed",
-          message: `Could not serve ${initFileName}: ${error?.message ?? String(error)}`
-        };
-      }
-    }
-
-    // Which run's copy answers, when several have written this name. Chosen by
-    // what the copies CARRY, not by which run is newest — see #chooseProducedCopy.
-    const filePath = this.#host.segmentStore.pathOfName(session.outputKey ?? "", fileName) ??
-      path.join(this.#host.segmentStore.pathFor(session.outputKey ?? ""), fileName);
-    const isPlaylist = fileName === PLAYLIST_FILE_NAME;
-    if (!isPlaylist) {
-      // A REQUEST SAYS THE VIEWER IS HERE, AND NOTHING ELSE. It does not say
-      // where they are — that is what they state themselves — and it steers no
-      // encoder: the segment either exists and is served, or does not and is
-      // waited for.
-      const requested = session.segmentFormat.segmentIndexFromName(fileName);
-      if (requested >= 0) {
-        this.#noteViewerSeen(session, consumerId);
-        // A viewer who has caught up must not wait out the monitor's interval —
-        // but only if they HAVE caught up, which is why this re-evaluates the
-        // same condition instead of resuming outright.
-        this.#host.cushion.reportCushionFor(session);
-      }
-    }
-    // Whether the file is there is asked on its own, and nothing else shares
-    // this catch. Everything below is PREPARATION of a file that exists, and a
-    // failure there means something entirely different from "not produced yet"
-    // — but for one release the two were caught together, so an undeclared name
-    // in the fMP4 path read as "the segment is not ready". Every poll threw the
-    // same ReferenceError, every poll answered "wait", and playback never began
-    // on any file cut at keyframes (2.9.124; measured 2026-08-08: segment #0
-    // held for 45 281 ms with twelve finished segments on disk).
-    try {
-      await access(filePath);
-    } catch {
-      // Not produced yet.
-      return this.#holdForProduction(session, fileName, isPlaylist, options);
-    }
-    try {
-      // Existing is not the same as finished. The `hls` muxer wrote each
-      // segment to a temporary name and renamed it once complete, so a file
-      // appearing WAS a finished segment. The `segment` muxer has no such
-      // option: the file appears when writing begins. Serving it then hands the
-      // player a truncated segment, which it rejects and then simply stops —
-      // observed as playback dying a few seconds in with the encoder still
-      // running happily ahead. A segment is finished once the NEXT one has been
-      // started, or once the run producing it has ended.
-      if (!isPlaylist && cutsAtGivenTimes(session)) {
-        // WHAT PROVES A PIECE IS WHOLE is the encoder's own word for it: it
-        // names each file on a channel of its own the moment it closes it, and
-        // the store keeps those names.
-        //
-        // What stood here instead was the existence of the NEXT file, with two
-        // exceptions bolted on because it is not true. It is never true of the
-        // last piece of a run — nothing is producing a next one — so the first
-        // segment of every run was held: measured 2026-08-09, #807 held while
-        // it lay on disk, and in August the same shape held #317 for 46 seconds
-        // and then answered 404 to a browser that had given up.
-        const index = session.segmentFormat.segmentIndexFromName(fileName);
-        if (!this.#host.segmentStore.isClosed(session.outputKey ?? "", index)) {
-          this.#explainHold(session, fileName, "the encoder has not closed it yet");
-          return { kind: "warming-up" };
-        }
-      }
-
-      // Cold-start: the time from the create request to a playable first
-      // segment, measured once per output by the owner of host timings.
-      const coldStartMs = isPlaylist ? null : this.#host.hostTimings.noteSegmentServed(session);
-      if (coldStartMs !== null) {
-        // Data is flowing again, so the next loss starts its backoff afresh
-        // rather than inheriting the delay of the last one.
-        this.#host.encodeRuns.resetInputRetry(session);
-        logger.info(`cold-start ${sessionId.slice(0, 8)}: first-segment ready +${coldStartMs}ms`);
-      }
-      // Formats whose segments need correcting before they are valid against
-      // the session's cached init are read whole and passed through the format
-      // module; the rest stream straight off disk.
-      if (!isPlaylist && session.segmentFormat.needsSegmentRewrite) {
-        const index = session.segmentFormat.segmentIndexFromName(fileName);
-        const raw = await readFile(filePath);
-        // Self-contained pieces carry the init header; a media segment must not.
-        const bytes = cutsAtGivenTimes(session) && session.segmentFormat.stripInit
-          ? session.segmentFormat.stripInit(raw)
-          : raw;
-        // WHOLE OR SHORT IS A JUDGEMENT ABOUT BYTES, and the format that
-        // knows how to read them makes it (`judgeTracks`). What is done about
-        // the answer is this path's business and stays here: a whole piece is
-        // served, a short one is removed so it can be made again.
-        const verdict = session.segmentFormat.judgeTracks?.(raw, bytes, this.#host.segmentStore.initOf(session.outputKey ?? "")) ?? null;
-        if (verdict && !verdict.whole) {
-          logger.warn(
-            `transcode ${session.id} segment #${index} is short of a track — ` +
-            `${filePath}, ${raw.length} bytes on disk, ${bytes.length} of body, ` +
-            `${verdict.fragmentTracks} track(s) in its fragments against ` +
-            `${verdict.sessionTracks} the session's header declares and ` +
-            `${verdict.ownTracks} its own declares — ${verdict.because}`
-          );
-          if (!verdict.serve) {
-            try {
-              await unlink(filePath);
-            } catch {
-              // Already gone: either way nothing to do.
-            }
-            // What the store remembers of this directory is stale the moment a
-            // file is taken out of it.
-            this.#host.segmentStore.forget(session.outputKey ?? "");
-            return { kind: "warming-up" };
-          }
-        }
-        // Where this segment REALLY begins, taken from the piece itself, and
-        // only from the playlist when the piece does not say.
-        //
-        // The playlist's own answer is built from the container's keyframe
-        // index, and an index can be wrong: measured 2026-08-06 on a Matroska
-        // file whose index claimed a keyframe at 157.99 s where the real ones
-        // were 153.82 and 164.247. ffmpeg cut at 153.82, and stamping that
-        // picture with 157.99 told the player it belonged four seconds later
-        // than it did — while subtitles, extracted straight from the source,
-        // kept the true times. Speech and text drifted apart by 4.17 s.
-        //
-        // Read from `raw`, before the header is stripped: the position lives
-        // in an empty edit in the piece's own `moov`, which `stripInit`
-        // removes. Identical to the playlist's figure whenever the index is
-        // honest, so nothing changes for a well-formed file.
-        const trueStart = cutsAtGivenTimes(session)
-          ? session.segmentFormat.readSegmentStartSeconds?.(raw) ?? null
-          : null;
-        const declaredStart = this.#host.outputTimes.segmentStartTime(session, index);
-        if (trueStart !== null) {
-          this.#host.outputTimes.noteRunLanding(session, index, trueStart);
-          this.#host.outputTimes.noteIndexAccuracy(session, index, trueStart, declaredStart);
-        }
-        // WHERE THE PLAYER WAS TOLD THIS SEGMENT BEGINS, which is the playlist
-        // it holds and nothing else. The published text is fixed when the
-        // session is created; `#segmentStartTime` reads a table that a
-        // correction may since have moved, and a stamp taken from the moved
-        // table describes a timeline the player has never seen.
-        const publishedStart = this.#host.outputTimes.publishedStartTime(session, index);
-        // A player places a fragment by the playlist. If the bytes claim a
-        // different position, the fragment does not land where the fragment was
-        // expected, hls.js finds the range still unbuffered and asks for the
-        // same fragment again — for ever. Measured 2026-08-17: a seek to
-        // 1590.4 s produced audio segments #292/#293 whose own timeline said
-        // 1587.892 and 1592.692 against a playlist saying 1585.376 and
-        // 1590.585, and the browser fetched those two segments 1908 times each
-        // over ten minutes, every one of them served in 4 ms. The film was dead
-        // and no line said why.
-        //
-        // So the stamp follows the playlist whenever the two disagree by more
-        // than a player will bridge. hls.js bridges up to `maxBufferHole`,
-        // which it defaults to 0.5 s — that is the player's own published
-        // figure, not a number chosen here. Within it the file's own position
-        // is kept, because it is the honest one and it is what keeps speech and
-        // subtitles together on a file whose index is slightly out (2026-08-06,
-        // 4.17 s of drift on a Matroska index that lied).
-        // STAMPED WITH ITS OWN TRUE START, always. Two attempts at moving it
-        // toward the playlist both made things worse, and the reason is in what
-        // the first segment of a run is: it is not CUT at all — it begins where
-        // ffmpeg's seek landed. The picture must land on a keyframe; the sound
-        // needs none and starts at the instant asked for. So after every
-        // restart the two runs genuinely begin at different real times, and the
-        // whole run carries that difference (field 2026-08-17: the sound's
-        // #292 began at 1587.892 s and #293 at 1592.692 s — exactly one segment
-        // apart, the whole run shifted 2.5 s from the grid).
-        //
-        // Labelling each track with its own true time is therefore what keeps
-        // picture and sound together in real time. Moving them onto the
-        // published grid — separately (2.24.1) or by one family offset (2.25.0)
-        // — closes a gap that is real and opens one that is not: it desynced
-        // playback in the field within the hour, twice.
-        //
-        // What that leaves unsolved is the reason those attempts were made: a
-        // playlist that disagrees with the media by more than a player bridges
-        // makes hls.js refetch the same fragment for ever (1908 times each for
-        // two segments, measured). The answer to THAT is to make the published
-        // grid agree with where the runs really begin — not to relabel the
-        // media. Recorded as its own roadmap item rather than guessed at here.
-        const stampStart = trueStart ?? publishedStart;
-        if (trueStart !== null) {
-          this.#host.outputTimes.notePlaylistDisagreement(session, index, trueStart, publishedStart);
-        }
-        const prepared = session.segmentFormat.prepareSegmentBytes(bytes, {
-          startSeconds: stampStart,
-          rawBytes: raw,
-          initBytes: this.#host.segmentStore.initOf(session.outputKey ?? "")
-        });
-        this.#host.encodeRuns.noteRunProducedSegment(session, filePath);
-        return {
-          kind: "file",
-          stream: Readable.from([prepared]),
-          contentType: session.segmentFormat.segmentContentType,
-          isPlaylist: false
-        };
-      }
-      if (!isPlaylist) {
-        this.#host.encodeRuns.noteRunProducedSegment(session, filePath);
-      }
-      return {
-        kind: "file",
-        stream: isPlaylist
-          ? createReadStream(filePath)
-          : createReadStream(filePath, { highWaterMark: SEGMENT_READ_HIGH_WATER_MARK }),
-        contentType: isPlaylist
-          ? "application/vnd.apple.mpegurl"
-          : session.segmentFormat.segmentContentType,
-        isPlaylist
-      };
-    } catch (error) {
-      if (error?.code === "ENOENT") {
-        // The file went away between the check and the read — a leftover being
-        // removed so it can be produced again. Means exactly what never having
-        // existed means.
-        return this.#holdForProduction(session, fileName, isPlaylist, options);
-      }
-      // Anything else is a fault in producing the answer. Name it: a request
-      // answered "wait" for ever tells the viewer nothing and leaves no trace
-      // of what actually happened.
-      logger.error(
-        `transcode ${session.id} could not serve ${fileName}: ${error?.message ?? error}` +
-        (error?.stack ? `\n${error.stack}` : "")
-      );
-      return {
-        kind: "failed",
-        message: `Could not serve ${fileName}: ${error?.message ?? String(error)}`
-      };
-    }
-  }
-
-  /**
-   * Whether a piece already given to a viewer can still be served from what is
-   * stored under the key of the output that made it, after that output has
-   * gone.
-   *
-   * WHY THIS EXISTS (roadmap item 97, step 11). A repeat of an address must be
-   * answered by what answered it the first time: the player holds that
-   * output's header, and a piece of another output under the same address may
-   * not decode under it. When the output itself has been disposed, its pieces
-   * and its header stay in the store until the disk needs the room — so the
-   * very bytes that were given can be given again.
-   *
-   * @param {string} goneKey
-   * @param {string} fileName
-   * @returns {boolean}
-   */
-  hasStoredPiece(goneKey, fileName) {
-    const format = this.#host.segmentFormatOfKey(goneKey);
-    if (!format) {
-      return false;
-    }
-    if (format.initFileName !== null && fileName === format.initFileName) {
-      return Boolean(this.#host.segmentStore.initOf(goneKey));
-    }
-    if (!format.isSegmentFileName(fileName)) {
-      return false;
-    }
-    const index = format.segmentIndexFromName(fileName);
-    return this.#host.segmentStore.isClosed(goneKey, index) &&
-      Boolean(this.#host.segmentStore.pathOfName(goneKey, fileName));
-  }
-
-  /**
-   * That stored piece, prepared exactly as it would have been served while its
-   * output lived: its header stripped where the piece carries one, and stamped
-   * with where it truly begins or, where it does not say, where the playlist
-   * the player holds puts it.
-   *
-   * `like` is the picture the address belongs to. Every output interchangeable
-   * with it publishes one timeline, so its playlist is the gone output's too.
-   *
-   * @param {string} goneKey
-   * @param {string} likeId - The picture the address belongs to.
-   * @param {string} fileName
-   * @returns {Promise<{ kind: "file", stream: Readable, contentType: string, isPlaylist: false } | null>}
-   */
-  async storedPieceOf(goneKey, likeId, fileName) {
-    const like = isOutputName(likeId) ? this.#host.outputs.get(likeId) : null;
-    const format = this.#host.segmentFormatOfKey(goneKey);
-    const spec = this.#host.specOfKey(goneKey);
-    if (!format || !spec || !like || !this.hasStoredPiece(goneKey, fileName)) {
-      return null;
-    }
-    const answer = (bytes, contentType) => ({
-      kind: "file",
-      stream: Readable.from([bytes]),
-      contentType,
-      isPlaylist: false
-    });
-    if (format.initFileName !== null && fileName === format.initFileName) {
-      return answer(this.#host.segmentStore.initOf(goneKey), format.initContentType);
-    }
-    const index = format.segmentIndexFromName(fileName);
-    let raw;
-    try {
-      raw = await readFile(this.#host.segmentStore.pathOfName(goneKey, fileName));
-    } catch {
-      return null;
-    }
-    if (!format.needsSegmentRewrite) {
-      return answer(raw, format.segmentContentType);
-    }
-    const view = { segmentFormat: format, spec, timeline: like.timeline };
-    const selfContained = cutsAtGivenTimes(view);
-    const bytes = selfContained && format.stripInit ? format.stripInit(raw) : raw;
-    const trueStart = selfContained ? format.readSegmentStartSeconds?.(raw) ?? null : null;
-    const prepared = format.prepareSegmentBytes(bytes, {
-      startSeconds: trueStart ?? this.#host.outputTimes.publishedStartTime(like, index),
-      rawBytes: raw,
-      initBytes: this.#host.segmentStore.initOf(goneKey)
-    });
-    logger.info(`[hold] ${fileName} served from the stored pieces of ${goneKey}, the output that first answered it`);
-    return answer(prepared, format.segmentContentType);
-  }
-
-  /**
-   * Whether a piece of `liveKey` may stand where a piece of `goneKey` was
-   * given: the two headers compared by the product's own rule
-   * (`init-compat.js`), which refuses anything it cannot show a decoder
-   * ignores. A header not yet made is no proof, and so is a refusal.
-   *
-   * @param {string} goneKey
-   * @param {string} liveKey
-   * @returns {{ compatible: boolean, differences: string[] }}
-   */
-  headersCompatible(goneKey, liveKey) {
-    const verdict = this.#host.compareInits(
-      this.#host.segmentStore.initOf(goneKey),
-      this.#host.segmentStore.initOf(liveKey)
-    );
-    return { compatible: verdict.compatible === true, differences: verdict.differences ?? [] };
-  }
-
-  /**
-   * Say WHY a segment is being held, at most once every few seconds per file.
-   *
-   * A hold is silent today, and that silence has now cost three releases: a
-   * file that exists, a route that answers "not yet", and nothing anywhere
-   * saying which of the several reasons applied. Measured 2026-08-09: a run
-   * begun mid-file at segment #317 produced two minutes of video from #317
-   * upwards at 10.5x, and #317 itself was held 46 s and then answered 404 once
-   * the browser had given up — with not one line about the cause.
-   *
-   * @param {HlsSession} session
-   * @param {string} fileName
-   * @param {string} reason
-   * @returns {void}
-   */
-  #explainHold(session, fileName, reason) {
-    const now = Date.now();
-    const state = this.#stateFor(session);
-    const last = state.holdExplainedAt.get(fileName) ?? 0;
-    if (now - last < 5_000) {
-      return;
-    }
-    state.holdExplainedAt.set(fileName, now);
-    const index = session.segmentFormat.segmentIndexFromName(fileName);
-    // What the encoder has actually DONE since it restarted. "Alive at the right
-    // index" was as far as the old line went, and it left the two possible
-    // causes indistinguishable: an encoder waiting for torrent pieces looks
-    // exactly like one that is encoding and simply has not finished. The
-    // difference is whether its position has moved at all.
-    // Where this run began, from the run — the same reckoning
-    // `processedSeconds` is counted in. A table lookup here can disagree with
-    // it by the distance between the two grids, which is enough to print a
-    // negative "produced" and send the reader after the torrent when the
-    // encoder is the subject.
-    const progress = this.#host.encodeRuns.progressOf(session, index);
-    const runStartSeconds = Number.isFinite(progress?.startPositionSeconds)
-      ? progress.startPositionSeconds
-      : this.#host.runStartTimeFor(session, this.#host.encodeRuns.earliestStartOf(session) ?? 0);
-    const position = Number(progress?.processedSeconds);
-    const produced = Number.isFinite(position) ? position - runStartSeconds : null;
-    const speed = progress?.speed || "n/a";
-    logger.warn(
-      `transcode ${session.id} holding ${fileName}: ${reason} ` +
-      `(runs from #${this.#host.encodeRuns.earliestStartOf(session) ?? "?"}, viewer at #${this.#host.outputTimes.segmentIndexForTime(session, this.#host.viewerSecondsOn(session))}, ` +
-      `encoder ${this.#host.encodeRuns.liveRunsOf(session).length > 0 ? "alive" : "stopped"}, index #${index}, ` +
-      `produced ${produced === null ? "nothing yet — no position reported" : `${produced.toFixed(1)}s`} ` +
-      `at ${speed}${produced !== null && produced <= 0 ? " — the encoder has not moved, so it is waiting on its input" : ""})`
-    );
-  }
-
-  /**
-   * What an absent file answers when production has failed.
-   *
-   * @param {object} session
-   * @returns {{ kind: "failed", message: string } | { kind: "warming-up" }}
-   */
-  #failedOrWarming(session) {
-    if (!this.#host.encodeRuns.hasFailed(session)) {
-      return { kind: "warming-up" };
-    }
-    return {
-      kind: "failed",
-      message: this.#host.encodeRuns.lastErrorOf(session) || "ffmpeg failed for this transcode session."
-    };
-  }
-
-  #holdForProduction(session, fileName, isPlaylist, options) {
-    // The file is not there and production has ended in an error: nothing is
-    // going to write it, so the viewer is told instead of held.
-    if (this.#host.encodeRuns.hasFailed(session)) {
-      return this.#failedOrWarming(session);
-    }
-    if (this.#host.encodeRuns.isWaitingForInput(session)) {
-      // The data went away and is being fetched again. Holding the request is
-      // the truthful answer: nothing is broken and there is nothing for the
-      // viewer to retry. Only what is NOT on disk is held for it — a piece that
-      // exists is whole, and it is served whatever the encoder is doing.
-      return { kind: "warming-up" };
-    }
-    /** @type {{ address: string, rank: number, topRank: number } | null} */
-    let ranked = null;
-    if (!isPlaylist) {
-      this.#explainHold(session, fileName, "the file is not on disk");
-    }
-    // A segment was requested that ffmpeg has not produced yet.  Decide whether
-    // to wait for the current encode run to reach it or to restart the encoder
-    // at this position (server-side seeking).  The caller long-polls.
-    if (!isPlaylist) {
-      const requestedIndex = session.segmentFormat.segmentIndexFromName(fileName);
-      // Unanswerable, and known to be: behind a run that only moves forward,
-      // too far behind for the repair to fetch it, and no seek on its way to
-      // move the encoder there. Holding it changes nothing about whether it can
-      // be produced — it only spends the player's patience.
-      //
-      // This is what a track change costs when it is held instead: measured
-      // 2026-08-15, hls.js asked the new track for segment #0 while the run was
-      // at #354, the request was held for the full minute, and only when it
-      // failed did the player move to the segment it actually needed — 63 s of
-      // spinner after a track that had been made ready in 7.
-      //
-      // WHETHER ANYBODY IS COMING FOR IT, asked of the encoding (`rankAt`,
-      // which also tells "in nobody's zone" from "no map yet"). The same walk
-      // was written out here and could answer only yes or no, so the RANK was
-      // discarded at the one point where a viewer measurably waits for a named
-      // segment; it goes back out with the answer, because the wait is measured
-      // by whoever holds the request.
-      const address = session.outputKey ?? "";
-      const { rank, topRank } = this.#host.encodeOrchestrator.rankAt(address, requestedIndex);
-      ranked = { address, rank, topRank };
-      const nobodyIsComing = topRank > 0 && rank === 0;
-      if (
-        Number.isFinite(requestedIndex) &&
-        requestedIndex < (this.#host.encodeRuns.earliestStartOf(session) ?? 0) &&
-        nobodyIsComing &&
-        this.#host.encodeRuns.liveRunsOf(session).length > 0
-      ) {
-        logger.info(
-          `transcode ${session.id} segment #${requestedIndex} is ${(this.#host.encodeRuns.earliestStartOf(session) ?? 0) - requestedIndex} ` +
-          "segments behind the run and in nobody's zone; answered as absent rather than held"
-        );
-        return { kind: "not-found", ranked };
-      }
-      this.#noteWanted(session, requestedIndex);
-    }
-    return { kind: "warming-up", ranked };
-  }
-
-  /**
-   * Record that a segment nobody has made yet is wanted, and say so when it lies
-   * behind every encoder of this output.
-   *
-   * A REQUEST STEERS NO ENCODER. This used to be where one did: a request far
-   * from the encode head restarted ffmpeg there, and in the field one seek
-   * produced nine restarts in a minute, because a player holds a couple of
-   * dozen requests open at once and none of them is "the one the viewer ended
-   * on". What is missing in front of a viewer is stated by the priority map,
-   * and placing encoders on it is the plan's work. What is left here is a
-   * record for the restart accounting and a line for whoever reads the log.
-   *
-   * @param {HlsSession} session
-   * @param {number} index
-   * @returns {void}
-   */
-  #noteWanted(session, index) {
-    // When this segment was FIRST asked for and nobody was producing it. The
-    // restart itself costs 0.7-1.3 s (measured 2.9.132), while a seek costs
-    // 5-8 s end to end — so most of the wait happens before a restart is even
-    // decided on, and that is what this records.
-    this.#host.encodeRuns.noteWanted(session, index);
-    if (!this.#host.encodeRuns.isLive(session) || index < 0) {
-      return;
-    }
-    const head = this.#host.encodeRuns.earliestStartOf(session) ?? 0;
-    if (index < head) {
-      // Nothing running goes backwards, so nothing running will make it.
-      this.#explainHold(
-        session,
-        session.segmentFormat.segmentFileName(index),
-        `it is behind the run (#${head}); where the viewers are is what places encoders`
-      );
-    }
-  }
-
-  /**
-   * The init header, lifted out of the first segment that exists.
-   *
-   * Needed only on the explicit-cut path, where the muxer produces no init file
-   * of its own. Scans rather than assuming segment 0: a run started by a seek
-   * begins at whatever index the viewer asked for.
-   *
-   * @param {HlsSession} session
-   * @returns {Promise<Buffer | null>}
-   */
-
-  async #initFromFirstSegment(session) {
-    if (typeof session.segmentFormat.extractInit !== "function") {
-      return null;
-    }
-    // How many tracks a complete header must declare is ANSWERED, not assumed.
-    //
-    // The probe already knows the source's stream list, and the output maps at
-    // most one of each (`-map 0:v:0? -map 0:a:0?`), so the count follows from
-    // what the source actually has. A film with no soundtrack expects one; an
-    // ordinary file expects two; neither is a convention.
-    //
-    // Deriving it from the produced pieces instead — the first version of this
-    // — reads correctly only once a piece carrying every track exists, and the
-    // whole point is the moment BEFORE that: early pieces written before the
-    // video was muxed would set the requirement to one and wave through exactly
-    // the header this exists to reject. The pieces are still consulted, but
-    // only as a floor: a piece carrying more than the probe led us to expect is
-    // evidence, and evidence outranks the probe.
-    const declared = this.#host.declaredTracks(session);
-    let expectedTracks = (declared.video ? 1 : 0) + (declared.audio ? 1 : 0);
-    if (expectedTracks === 0) {
-      // Nothing to consult. Fall back to the evidence, with its known lag.
-      expectedTracks = 1;
-    }
-    let best = null;
-    let bestTracks = 0;
-    /** @type {Map<string, Buffer>} Pieces read once and used for both passes. */
-    const pieces = new Map();
-    let names;
-    try {
-      names = this.producedNumbers(session).map((index) => session.segmentFormat.segmentFileName(index));
-    } catch {
-      return null;
-    }
-    // First pass: what do the produced pieces actually carry? The answer is the
-    // requirement — no assumption about the source is involved.
-    if (typeof session.segmentFormat.countSegmentTracks === "function") {
-      for (const name of names) {
-        try {
-          // The first copy WITH BYTES IN IT, not the first name: a run stopped
-          // with a piece open leaves an empty file under the same name, and
-          // taking that one skips a number whose header is sitting in the run
-          // before it.
-          const found = await this.#firstCopyWithBytes(session, name);
-          if (!found) {
-            continue;
-          }
-          const bytes = await readFile(found);
-          pieces.set(name, bytes);
-          expectedTracks = Math.max(expectedTracks, session.segmentFormat.countSegmentTracks(bytes));
-        } catch {
-          // Being written right now — it says nothing about the others.
-        }
-      }
-    }
-
-    for (const name of names) {
-      try {
-        const cached = pieces.get(name);
-        const found = cached ? name : await this.#firstCopyWithBytes(session, name);
-        if (!found) {
-          continue;
-        }
-        const ownInit = session.segmentFormat.extractInit(cached ?? await readFile(found));
-        const init = ownInit && session.segmentFormat.prepareSharedInit ?
-          session.segmentFormat.prepareSharedInit(ownInit) : ownInit;
-        if (!init || init.length === 0) {
-          continue;
-        }
-        // The requirement computed above is APPLIED here. It was computed and
-        // then ignored: this loop returned the first header it found, so a
-        // piece written before the video was muxed supplied an audio-only
-        // header — and that header is cached for the session's whole life,
-        // because the player fetches `#EXT-X-MAP` once. Measured 2026-08-11:
-        // `videoWidth=0`, `totalVideoFrames=0`, `readyState=4` — sound playing
-        // and no picture, for as long as the session lasted.
-        const tracks = typeof session.segmentFormat.countInitTracks === "function"
-          ? session.segmentFormat.countInitTracks(init)
-          : expectedTracks;
-        if (tracks >= expectedTracks) {
-          return init;
-        }
-        if (tracks > bestTracks) {
-          best = init;
-          bestTracks = tracks;
-        }
-      } catch {
-        // Being written right now — try the next one.
-      }
-    }
-    if (best !== null) {
-      // Nothing carried the full set. The source is probably missing a stream;
-      // serving the richest header found is right, and saying so makes the
-      // other possibility — every piece so far written before the video was
-      // muxed — visible rather than silent.
-      logger.warn(
-        `transcode ${session.id} no piece declared ${expectedTracks} tracks; ` +
-        `serving an init with ${bestTracks}`
-      );
-      return best;
-    }
-    return best;
-  }
-
-  /**
-   * A produced file that has anything in it, or null.
-   *
-   * For callers that need a file's CONTENTS and cannot judge them — deriving
-   * the session's header is the case, since the header is what judging would
-   * need. An empty file answers no question, so it is passed over.
-   *
-   * @param {HlsSession} session
-   * @param {string} fileName
-   * @returns {Promise<string | null>}
-   */
-  async #firstCopyWithBytes(session, fileName) {
-    const held = this.#host.segmentStore.pathOfName(session.outputKey ?? "", fileName);
-    if (held === null) {
-      return null;
-    }
-    try {
-      const info = await stat(held);
-      return info.size > 0 ? held : null;
-    } catch {
-      return null;
-    }
-  }
-
-  /**
-   * Record where one viewer of this session is, and answer with the furthest
-   * any of them has reached.
-   *
-   * The furthest is what the single encoder is steered by: it has to serve
-   * everyone, and what lies behind the leader has already been produced and is
-   * served from disk without a wait. The individual positions exist for the
-   * opposite question — whether a particular held request is still wanted —
-   * which cannot be answered from a shared field.
-   *
-   * A REQUEST IS NOT A POSITION. It says the viewer is still here and nothing
-   * more: where they are is what they themselves state, on the viewer, and a
-   * request cannot reach it. Asking for a segment used to write the position,
-   * so two writers filled one field in turn and the priority map jumped
-   * backwards several times a second — measured 2026-09-13, 77 encoder starts
-   * and 141 stops in six minutes while both viewers sat frozen.
-   *
-   * A viewer is forgotten once nothing has been heard from them for longer than
-   * any silence a watching viewer can produce. The figure is the proxy's own
-   * look-ahead, not a chosen interval.
-   *
-   * @param {HlsSession} session
-   * @param {string} consumerId
-   * @returns {void}
-   */
-  #noteViewerSeen(session, consumerId) {
-    // A request that names nobody is evidence about nobody.
-    if (!consumerId) {
-      return;
-    }
-    this.#host.viewers.of(session, consumerId).seen();
-  }
-
-  /**
-   * Whether a held request is for a segment the viewer STILL needs.
-   *
-   * The epoch alone says a seek happened; it cannot say whether this particular
-   * request was made for the position left behind or for the one just arrived
-   * at. That distinction is the whole of the failure measured 2026-08-18: the
-   * viewer seeked to 1061.0 s, the request for `segment-00101` — the segment AT
-   * that position — raced the seek notification, the epoch moved underneath it,
-   * and it was answered 503 twice within 80 ms. The player then hunted at
-   * sn=105-107, never came back to 101, and looped two audio segments 1473
-   * times over 149 s while the picture stood still.
-   *
-   * A request is stale when its segment lies behind where the viewer now is, or
-   * so far ahead that the running encode will not reach it. Anything between is
-   * exactly what the viewer is waiting for, and holding it is the point.
-   *
-   * "So far ahead" is the encoder's own look-ahead, measured on this session's
-   * own cut grid — the same figure the browser sizes its forward buffer from.
-   * It used to be `MAX_LOOKAHEAD_SEGMENTS`, which is eight segments ahead of
-   * the ENCODE HEAD
-   * and has nothing to do with how far ahead of the VIEWER a request may
-   * legitimately sit; it happened to match a browser holding 30 s, and would
-   * have refused three quarters of the requests of one holding the whole
-   * cushion (roadmap item 4).
-   *
-   * Judged against the position of the viewer who MADE the request, when the
-   * transport carries who that is. A session is shared by everyone watching a
-   * copied picture and the epoch is per session, so a seek by the viewer in
-   * front used to release every request being held for the viewer behind them.
-   *
-   * @param {string} sessionId
-   * @param {string} fileName
-   * @param {string} [consumerId] - Who is asking. Without it the one shared
-   *   position decides, which is what a single viewer means anyway.
-   * @returns {boolean} True when the request should keep waiting.
-   */
-  requestStillWanted(sessionId, fileName, consumerId = "") {
-    const session = isOutputName(sessionId) ? this.#host.outputs.get(sessionId) : null;
-    if (!session) {
-      return false;
-    }
-    const index = session.segmentFormat?.segmentIndexFromName?.(fileName) ?? -1;
-    if (!(index >= 0)) {
-      return true; // a playlist or an init segment belongs to no position
-    }
-    const position = this.#host.viewerSecondsOn(session, consumerId);
-    const at = this.#host.outputTimes.segmentIndexForTime(session, position);
-    // The far edge on THIS session's own grid rather than a count of nominal
-    // segments: a copied picture is cut at the source's keyframes, so its
-    // segments are not four seconds long and dividing by that figure would put
-    // the edge somewhere else entirely. The segment CONTAINING the edge is
-    // wanted — it is the one the deepest allowed request lands in, and it
-    // already reaches past the cushion by whatever is left of its own duration.
-    const edge = this.#host.outputTimes.segmentIndexForTime(session, position + this.#host.lookaheadSeconds);
-    return index >= at && index <= edge;
-  }
-
-  /**
-   * How many times the viewer has moved since this session started.
-   *
-   * A request being held for a segment answers "retry" as soon as this changes,
-   * because it was made for a position the viewer has left — see `requestSeek`.
-   *
-   * @param {string} sessionId
-   * @returns {number}
-   */
-  seekEpoch(sessionId) {
-    const session = isOutputName(sessionId) ? this.#host.outputs.get(sessionId) : null;
-    return session ? this.#stateFor(session).waitEpoch : 0;
-  }
-
-  /**
-   * Wait for the requested segment to be published, without polling the disk.
-   *
-   * Ends on the publication, on the waits of this output being invalidated, on
-   * the deadline, or when the requester goes — and whichever ends it, the
-   * waiter is taken out of the store, so a wait with no deadline cannot outlive
-   * the request it was for. Returns false at once for playlists and init files,
-   * whose readiness has a different owner.
-   *
-   * @param {string} sessionId
-   * @param {string} fileName
-   * @param {number} timeoutMs - May be infinite.
-   * @param {Promise<unknown> | null} [requesterGone]
-   * @returns {Promise<boolean>} Whether it is worth asking for the file again.
-   */
-  async waitForSegment(sessionId, fileName, timeoutMs, requesterGone = null) {
-    const session = isOutputName(sessionId) ? this.#host.outputs.get(sessionId) : null;
-    const index = session?.segmentFormat?.segmentIndexFromName?.(fileName) ?? -1;
-    if (!session || index < 0) return false;
-    const state = this.#stateFor(session);
-    let wake;
-    const invalidated = new Promise((resolve) => {
-      wake = resolve;
-      state.waitListeners.add(resolve);
-    });
-    let finished = () => {};
-    const over = new Promise((resolve) => {
-      finished = resolve;
-    });
-    try {
-      return await Promise.race([
-        this.#host.segmentStore.waitFor(session.outputKey ?? "", index, timeoutMs, over),
-        invalidated,
-        ...(requesterGone ? [requesterGone.then(() => false)] : [])
-      ]);
-    } finally {
-      state.waitListeners.delete(wake);
-      finished();
-    }
-  }
-
-  /**
-   * Poll until the HLS playlist file exists and contains a valid `#EXTM3U`
-   * header, or until the session fails, or until the startup timeout elapses.
-   * Throws with message `"HLS playlist is still warming up."` on timeout.
-   *
-   * @param {HlsSession} session
-   * @returns {Promise<void>}
-   */
-  async waitUntilReady(session) {
-    // With a synthetic VOD playlist there is nothing to wait for: the playlist
-    // is generated from the probed duration and is available immediately.
-    // Individual segments are long-polled by the segment route as ffmpeg
-    // produces them.
-    if (session.useSyntheticPlaylist) {
-      if (this.#host.encodeRuns.hasFailed(session)) {
-        throw new Error(this.#host.encodeRuns.lastErrorOf(session) || "ffmpeg failed to start HLS session.");
-      }
-      return;
-    }
-
-    const playlistPath = path.join(this.#host.segmentStore.pathFor(session.outputKey ?? ""), PLAYLIST_FILE_NAME);
-    const deadline = Date.now() + this.#host.startupWaitMs;
-
-    while (Date.now() < deadline) {
-      if (this.#host.encodeRuns.hasFailed(session)) {
-        throw new Error(this.#host.encodeRuns.lastErrorOf(session) || "ffmpeg failed to start HLS session.");
-      }
-      try {
-        await access(playlistPath);
-        const text = await readFile(playlistPath, "utf8");
-        if (text.includes("#EXTM3U")) {
-            return;
-        }
-      } catch (_error) {
-        // Playlist is not ready yet.
-      }
-      await delay(250);
-    }
-
-    throw new Error("HLS playlist is still warming up.");
-  }
-
-  /**
-   * Answer the player's report that a delivered fragment sits far from the edge
-   * of its buffer, with the one fact only this side holds: which boundary the
-   * segment of that number really begins at.
-   *
-   * The player can say the gap; it cannot say whether the cause is its own
-   * loading or a run whose output no longer matches its numbering. Here both
-   * are in hand — the time the playlist gave that segment, and, when the
-   * segment has been served, the time it truly began at — so the line either
-   * names a shifted run or clears this side of it.
-   *
-   * Diagnostic only: nothing is repositioned on the strength of a browser's
-   * reading, deliberately, because a wrong answer here would restart an encoder
-   * the viewer is waiting on.
-   *
-   * @param {string} sessionId
-   * @param {{ sn: number, track?: string, fragStartSec: number, bufferEndSec: number, currentTimeSec: number }} report
-   * @returns {boolean} False when no such session exists.
-   */
-  recordFragmentFar(sessionId, { sn, track, fragStartSec, bufferEndSec, currentTimeSec }) {
-    const named = this.#host.outputs.get(sessionId);
-    if (!named) {
-      return false;
-    }
-    // Which of the two streams the report is about. The browser addresses
-    // everything to the video session's id — the soundtrack is served under
-    // `/a/<n>/` on that same id — but it is a session of its own, with its own
-    // run and its own position, and that is exactly the pair this report exists
-    // to tell apart. Answering an audio report from the picture's records would
-    // state, confidently, something about the wrong stream.
-    const onScreen = this.#host.activeOutputFor({ base: named, outputs: this.#host.outputs });
-    const session = track === "audio"
-      ? ([...this.#host.outputs.familyOf(onScreen)].find((member) => member.spec.carries === "audio-only") ?? onScreen)
-      : onScreen;
-    const gap = fragStartSec - bufferEndSec;
-    const declared = this.#host.outputTimes.publishedStartTime(session, sn);
-    const trueStart = this.#host.outputTimes.trueStartAt(session, sn);
-    const verdict = trueStart === undefined
-      // Where a segment truly began is only ever read off one that was cut on
-      // an explicit list — a uniform grid has nothing to read back — so this is
-      // "not recorded", which is not the same as "not produced", and the line
-      // must not claim the second.
-      ? "where that segment began is not recorded on this side, so the gap cannot be attributed here"
-      : (() => {
-        const at = this.#host.outputTimes.boundaryIndexAt(session, trueStart, this.#host.publishedGridFor(session));
-        if (at === null) {
-          return `it really began at ${trueStart.toFixed(3)}s, which is no boundary of this grid`;
-        }
-        if (at === sn) {
-          return `it really began at boundary #${sn}, where it should — the gap is not this run's`;
-        }
-        return `it really began at boundary #${at}, ${sn - at} place(s) before its own number — ` +
-          "this run's output does not match its numbering";
-      })();
-    logger.warn(
-      `transcode ${session.id} the player is stuck: ${session.spec.carries === "audio-only" ? "sound" : "picture"} ` +
-      `fragment #${sn} starts ${gap.toFixed(1)}s past ` +
-      `the end of its buffer (${bufferEndSec.toFixed(1)}s, viewer at ${currentTimeSec.toFixed(1)}s, ` +
-      `the playlist puts it at ${declared.toFixed(3)}s) — ${verdict}`
-    );
-    return true;
-  }
-
-  /**
-   * Every segment number this session's OUTPUT holds, whoever produced it.
-   *
-   * Public because it is the one thing worth asserting about the address
-   * change: two sessions of one output answer with the same list, including
-   * segments the other one's encoder made.
-   *
-   * @param {HlsSession} session
-   * @returns {number[]}
-   */
-  producedSegmentNumbers(session) {
-    return new Set(this.producedNumbers(session));
-  }
-
-  /**
-   * What this session's OUTPUT holds, asked of the one thing that owns it.
-   *
-   * There were two owners of this fact over one directory: the store, addressed
-   * by the output's own key, and a `ProducedIndex` built per SESSION over a
-   * path the store had handed out. Two viewers of one output therefore built
-   * two indexes over one directory, each with its own idea of what was in it.
-   *
-   * @param {HlsSession} session
-   * @returns {number[]} Every segment number it holds, in order.
-   */
-  producedNumbers(session) {
-    return this.#host.segmentStore.provenNumbers(session.outputKey ?? "");
-  }
-}
+/**
+ * Answering one request for a file of an output: a playlist, the init segment
+ * or a segment.
+ *
+ * A segment is served from the store when it is made and whole; otherwise the
+ * request is held until it is, or until nobody wants it any more, and a hold
+ * that lasts says why. A request is an operation on material that exists; where
+ * encoders work is the plan's, and a request does not move one.
+ */
+
+import { createReadStream } from "node:fs";
+import { access, readFile, stat, unlink } from "node:fs/promises";
+import { Readable } from "node:stream";
+import path from "node:path";
+import { logger } from "../../utils/logger.js";
+import { isOutputName, PLAYLIST_FILE_NAME } from "../encode/output/index.js";
+// The index of variants. Served from the same route as the media playlist, so
+// it needs no path of its own.
+export const MASTER_PLAYLIST_FILE_NAME = "master.m3u8";
+// Read segment files in large blocks so the body is delivered to the data
+// channel in few, big chunks. On a busy ARM host the in-process WebTorrent
+// hashing starves the event loop in bursts, so fewer read iterations means
+// far less time lost between chunks while serving the first segments.
+const SEGMENT_READ_HIGH_WATER_MARK = 4 * 1024 * 1024;
+/**
+ * Resolve after a given number of milliseconds.
+ *
+ * @param {number} ms
+ * @returns {Promise<void>}
+ */
+function delay(ms) {
+  return new Promise((resolve) => {
+    setTimeout(resolve, ms);
+  });
+}
+/**
+ * Guard against path traversal by restricting file names to the known
+ * playlist and segment patterns produced by ffmpeg. Which segment names are
+ * legal depends on the active container, so the format decides.
+ *
+ * @param {string} fileName
+ * @param {import("../encode/segment-formats/index.js").SegmentFormat} segmentFormat
+ * @returns {boolean}
+ */
+function isSafeFileName(fileName, segmentFormat) {
+  return (
+    fileName === PLAYLIST_FILE_NAME ||
+    fileName === MASTER_PLAYLIST_FILE_NAME ||
+    (segmentFormat.initFileName !== null && fileName === segmentFormat.initFileName) ||
+    segmentFormat.isSegmentFileName(fileName)
+  );
+}
+/**
+ * Whether this output is cut at times we hand the muxer, rather than at a
+ * duration it chooses for itself.
+ *
+ * A property of the output and not of a run: the cut grid and the branch decide
+ * it, so every run of one output answers alike. It decides how a segment is
+ * judged finished — see getFileStream.
+ *
+ * @param {object} session
+ * @returns {boolean}
+ */
+function cutsAtGivenTimes(session) {
+  const explicit = session?.segmentFormat?.explicitTimesMuxerArgs?.() ?? null;
+  if (!explicit) {
+    return false;
+  }
+  return !session.spec.transcodesVideo || session.timeline?.cutGrid === "keyframe";
+}
+
+export class SegmentServing {
+  /** Serving state keyed by the output object it belongs to. */
+  #states = new WeakMap();
+  /** What this reads and asks of the rest of the proxy, and nothing else. @type {object} */
+  #host;
+
+  /**
+   * @param {object} host - `buildMasterPlaylist`, `declaredTracks`, `publishedGridFor`, `runStartTimeFor`, `cushion`, `encodeOrchestrator`, `encodeRuns`, `hostTimings`, `lookaheadSeconds`, `outputTimes`, `outputs`, `segmentStore`, `startupWaitMs`, `viewers`
+   */
+  constructor(host) {
+    this.#host = host;
+  }
+
+  #stateFor(output) {
+    let state = this.#states.get(output);
+    if (!state) {
+      state = {
+        holdExplainedAt: new Map(),
+        waitEpoch: 0,
+        waitListeners: new Set()
+      };
+      this.#states.set(output, state);
+    }
+    return state;
+  }
+
+  invalidateWaits(output) {
+    const state = this.#stateFor(output);
+    state.waitEpoch += 1;
+    for (const wake of state.waitListeners) wake(true);
+    state.waitListeners.clear();
+  }
+
+  /**
+   * Open a read stream for an HLS segment or playlist file from a session.
+   *
+   * @param {string} sessionId
+   * @param {string} fileName - Must match the playlist or segment name pattern.
+   * @param {{ consumerId?: string }} [options] -
+   *   `consumerId` says WHICH viewer is asking, so a session
+   *   shared by several of them can tell their positions apart; absent from a
+   *   browser or a transport that does not carry it, and then everything falls
+   *   back to the one shared position.
+   * @returns {Promise<
+   *   | { kind: "not-found" }
+   *   | { kind: "warming-up" }
+   *   | { kind: "failed"; message: string }
+   *   | { kind: "file"; stream: import("node:fs").ReadStream; contentType: string; isPlaylist: boolean }
+   * >}
+   */
+  async getFileStream(sessionId, fileName, options = {}) {
+    const consumerId = typeof options.consumerId === "string" ? options.consumerId : "";
+    if (!isOutputName(sessionId)) {
+      return { kind: "not-found" };
+    }
+    const session = this.#host.outputs.get(sessionId);
+    // The session is looked up BEFORE the name is validated, because what
+    // counts as a valid segment name depends on the container this session
+    // chose — `.mp4` for fMP4, `.ts` for MPEG-TS.
+    if (!session || !isSafeFileName(fileName, session.segmentFormat)) {
+      return { kind: "not-found" };
+    }
+    // WHETHER PRODUCTION FAILED IS NOT ASKED HERE. It explains a piece that is
+    // ABSENT and says nothing about one that is on disk: a segment under its
+    // served name is whole by construction, whoever wrote it and whatever has
+    // become of the encoder since. Asked at the door, as it was, a failed run
+    // answered 500 to every request for material it had already finished — the
+    // viewer lost what was made as well as what was not, and a seek back into
+    // it could not be served either. It is asked where the answer is needed:
+    // where the file is not there (`#holdForProduction`, and the init's own
+    // absent branch).
+    this.#host.outputs.touch(session);
+
+    // The index of variants. Served from here rather than a route of its own,
+    // because to a player it is simply another playlist under the session.
+    if (fileName === MASTER_PLAYLIST_FILE_NAME) {
+      const masterText = this.#host.buildMasterPlaylist(sessionId, consumerId);
+      if (!masterText) {
+        return { kind: "not-found" };
+      }
+      return {
+        kind: "file",
+        stream: Readable.from([masterText]),
+        contentType: "application/vnd.apple.mpegurl",
+        isPlaylist: true
+      };
+    }
+
+    // Serve the synthetic VOD playlist (full duration, terminated with
+    // #EXT-X-ENDLIST) so the player gets the correct total length and a fully
+    // seekable timeline up-front, independent of how far ffmpeg has encoded.
+    if (fileName === PLAYLIST_FILE_NAME && session.useSyntheticPlaylist) {
+      return {
+        kind: "file",
+        stream: Readable.from([session.playlistText]),
+        contentType: "application/vnd.apple.mpegurl",
+        isPlaylist: true
+      };
+    }
+
+    // The init segment (fMP4 only; referenced by #EXT-X-MAP). Each seek-restart
+    // run REWRITES it, so cache the FIRST one and always serve that — the
+    // player fetches it once and never re-fetches, so it must stay stable for
+    // the session's lifetime. (What that costs, and why segments must therefore
+    // carry their own position, is documented in `encode/segment-formats/mp4-boxes.js`
+    // `stampSegmentStartTime`.)
+    //
+    // ffmpeg creates init.mp4 before it has finished writing the fMP4 header
+    // boxes into it (unlike segments, its write is not gated behind an atomic
+    // rename), so a read can race a moment where the file EXISTS but is still
+    // EMPTY. Root cause of a real incident: that empty read used to be cached
+    // as the output's init — a zero-length Buffer is still a truthy object,
+    // so a check on its presence treated it as "already resolved" and served
+    // the empty file for the rest of the session's life, permanently breaking
+    // playback (hls.js can never initialize its SourceBuffer from an empty
+    // init segment) while the transcode itself kept encoding normally. Guard
+    // on non-empty content on both the cache check and the fresh read, so an
+    // empty read is treated as not-yet-ready and the caller's long-poll keeps
+    // retrying until ffmpeg has actually written the header.
+    const { initFileName } = session.segmentFormat;
+    if (initFileName !== null && fileName === initFileName) {
+      const kept = this.#host.segmentStore.initOf(session.outputKey ?? "");
+      if (kept) {
+        return {
+          kind: "file",
+          stream: Readable.from([kept]),
+          contentType: session.segmentFormat.initContentType,
+          isPlaylist: false
+        };
+      }
+      try {
+        // With explicit cut times there is no init file: that muxer writes each
+        // piece self-contained, header and all. The header is identical in every
+        // piece, so the first one to exist supplies it.
+        const bytes = cutsAtGivenTimes(session)
+          ? await this.#initFromFirstSegment(session)
+          : await readFile(path.join(this.#host.segmentStore.pathFor(session.outputKey ?? ""), initFileName));
+        const standing = this.#host.segmentStore.keepInit(session.outputKey ?? "", bytes);
+        if (!standing) {
+          // Nothing usable yet. On the branch that lifts the header out of the
+          // first piece there is no file to be missing, so "not produced yet"
+          // arrives HERE and not in the catch below: `#initFromFirstSegment`
+          // answers null rather than throwing `ENOENT`. A production that has
+          // ended in an error is not going to produce one either, and saying
+          // "still warming up" to that holds the request for its whole deadline.
+          return this.#failedOrWarming(session);
+        }
+        return {
+          kind: "file",
+          stream: Readable.from([standing]),
+          contentType: session.segmentFormat.initContentType,
+          isPlaylist: false
+        };
+      } catch (error) {
+        if (error?.code === "ENOENT") {
+          // Not produced yet — the encode run started at session creation
+          // writes it early; the caller long-polls until it appears. Unless
+          // production has ended in an error, in which case nothing is going to
+          // write it and holding the request only spends the viewer's patience.
+          return this.#failedOrWarming(session);
+        }
+        logger.error(
+          `transcode ${session.id} could not serve ${initFileName}: ${error?.message ?? error}` +
+          (error?.stack ? `\n${error.stack}` : "")
+        );
+        return {
+          kind: "failed",
+          message: `Could not serve ${initFileName}: ${error?.message ?? String(error)}`
+        };
+      }
+    }
+
+    // Which run's copy answers, when several have written this name. Chosen by
+    // what the copies CARRY, not by which run is newest — see #chooseProducedCopy.
+    const filePath = this.#host.segmentStore.pathOfName(session.outputKey ?? "", fileName) ??
+      path.join(this.#host.segmentStore.pathFor(session.outputKey ?? ""), fileName);
+    const isPlaylist = fileName === PLAYLIST_FILE_NAME;
+    if (!isPlaylist) {
+      // A REQUEST SAYS THE VIEWER IS HERE, AND NOTHING ELSE. It does not say
+      // where they are — that is what they state themselves — and it steers no
+      // encoder: the segment either exists and is served, or does not and is
+      // waited for.
+      const requested = session.segmentFormat.segmentIndexFromName(fileName);
+      if (requested >= 0) {
+        this.#noteViewerSeen(session, consumerId);
+        // A viewer who has caught up must not wait out the monitor's interval —
+        // but only if they HAVE caught up, which is why this re-evaluates the
+        // same condition instead of resuming outright.
+        this.#host.cushion.reportCushionFor(session);
+      }
+    }
+    // Whether the file is there is asked on its own, and nothing else shares
+    // this catch. Everything below is PREPARATION of a file that exists, and a
+    // failure there means something entirely different from "not produced yet"
+    // — but for one release the two were caught together, so an undeclared name
+    // in the fMP4 path read as "the segment is not ready". Every poll threw the
+    // same ReferenceError, every poll answered "wait", and playback never began
+    // on any file cut at keyframes (2.9.124; measured 2026-08-08: segment #0
+    // held for 45 281 ms with twelve finished segments on disk).
+    try {
+      await access(filePath);
+    } catch {
+      // Not produced yet.
+      return this.#holdForProduction(session, fileName, isPlaylist, options);
+    }
+    try {
+      // Existing is not the same as finished. The `hls` muxer wrote each
+      // segment to a temporary name and renamed it once complete, so a file
+      // appearing WAS a finished segment. The `segment` muxer has no such
+      // option: the file appears when writing begins. Serving it then hands the
+      // player a truncated segment, which it rejects and then simply stops —
+      // observed as playback dying a few seconds in with the encoder still
+      // running happily ahead. A segment is finished once the NEXT one has been
+      // started, or once the run producing it has ended.
+      if (!isPlaylist && cutsAtGivenTimes(session)) {
+        // WHAT PROVES A PIECE IS WHOLE is the encoder's own word for it: it
+        // names each file on a channel of its own the moment it closes it, and
+        // the store keeps those names.
+        //
+        // What stood here instead was the existence of the NEXT file, with two
+        // exceptions bolted on because it is not true. It is never true of the
+        // last piece of a run — nothing is producing a next one — so the first
+        // segment of every run was held: measured 2026-08-09, #807 held while
+        // it lay on disk, and in August the same shape held #317 for 46 seconds
+        // and then answered 404 to a browser that had given up.
+        const index = session.segmentFormat.segmentIndexFromName(fileName);
+        if (!this.#host.segmentStore.isClosed(session.outputKey ?? "", index)) {
+          this.#explainHold(session, fileName, "the encoder has not closed it yet");
+          return { kind: "warming-up" };
+        }
+      }
+
+      // Cold-start: the time from the create request to a playable first
+      // segment, measured once per output by the owner of host timings.
+      const coldStartMs = isPlaylist ? null : this.#host.hostTimings.noteSegmentServed(session);
+      if (coldStartMs !== null) {
+        // Data is flowing again, so the next loss starts its backoff afresh
+        // rather than inheriting the delay of the last one.
+        this.#host.encodeRuns.resetInputRetry(session);
+        logger.info(`cold-start ${sessionId.slice(0, 8)}: first-segment ready +${coldStartMs}ms`);
+      }
+      // Formats whose segments need correcting before they are valid against
+      // the session's cached init are read whole and passed through the format
+      // module; the rest stream straight off disk.
+      if (!isPlaylist && session.segmentFormat.needsSegmentRewrite) {
+        const index = session.segmentFormat.segmentIndexFromName(fileName);
+        const raw = await readFile(filePath);
+        // Self-contained pieces carry the init header; a media segment must not.
+        const bytes = cutsAtGivenTimes(session) && session.segmentFormat.stripInit
+          ? session.segmentFormat.stripInit(raw)
+          : raw;
+        // WHOLE OR SHORT IS A JUDGEMENT ABOUT BYTES, and the format that
+        // knows how to read them makes it (`judgeTracks`). What is done about
+        // the answer is this path's business and stays here: a whole piece is
+        // served, a short one is removed so it can be made again.
+        const verdict = session.segmentFormat.judgeTracks?.(raw, bytes, this.#host.segmentStore.initOf(session.outputKey ?? "")) ?? null;
+        if (verdict && !verdict.whole) {
+          logger.warn(
+            `transcode ${session.id} segment #${index} is short of a track — ` +
+            `${filePath}, ${raw.length} bytes on disk, ${bytes.length} of body, ` +
+            `${verdict.fragmentTracks} track(s) in its fragments against ` +
+            `${verdict.sessionTracks} the session's header declares and ` +
+            `${verdict.ownTracks} its own declares — ${verdict.because}`
+          );
+          if (!verdict.serve) {
+            try {
+              await unlink(filePath);
+            } catch {
+              // Already gone: either way nothing to do.
+            }
+            // What the store remembers of this directory is stale the moment a
+            // file is taken out of it.
+            this.#host.segmentStore.forget(session.outputKey ?? "");
+            return { kind: "warming-up" };
+          }
+        }
+        // Where this segment REALLY begins, taken from the piece itself, and
+        // only from the playlist when the piece does not say.
+        //
+        // The playlist's own answer is built from the container's keyframe
+        // index, and an index can be wrong: measured 2026-08-06 on a Matroska
+        // file whose index claimed a keyframe at 157.99 s where the real ones
+        // were 153.82 and 164.247. ffmpeg cut at 153.82, and stamping that
+        // picture with 157.99 told the player it belonged four seconds later
+        // than it did — while subtitles, extracted straight from the source,
+        // kept the true times. Speech and text drifted apart by 4.17 s.
+        //
+        // Read from `raw`, before the header is stripped: the position lives
+        // in an empty edit in the piece's own `moov`, which `stripInit`
+        // removes. Identical to the playlist's figure whenever the index is
+        // honest, so nothing changes for a well-formed file.
+        const trueStart = cutsAtGivenTimes(session)
+          ? session.segmentFormat.readSegmentStartSeconds?.(raw) ?? null
+          : null;
+        const declaredStart = this.#host.outputTimes.segmentStartTime(session, index);
+        if (trueStart !== null) {
+          this.#host.outputTimes.noteRunLanding(session, index, trueStart);
+          this.#host.outputTimes.noteIndexAccuracy(session, index, trueStart, declaredStart);
+        }
+        // WHERE THE PLAYER WAS TOLD THIS SEGMENT BEGINS, which is the playlist
+        // it holds and nothing else. The published text is fixed when the
+        // session is created; `#segmentStartTime` reads a table that a
+        // correction may since have moved, and a stamp taken from the moved
+        // table describes a timeline the player has never seen.
+        const publishedStart = this.#host.outputTimes.publishedStartTime(session, index);
+        // A player places a fragment by the playlist. If the bytes claim a
+        // different position, the fragment does not land where the fragment was
+        // expected, hls.js finds the range still unbuffered and asks for the
+        // same fragment again — for ever. Measured 2026-08-17: a seek to
+        // 1590.4 s produced audio segments #292/#293 whose own timeline said
+        // 1587.892 and 1592.692 against a playlist saying 1585.376 and
+        // 1590.585, and the browser fetched those two segments 1908 times each
+        // over ten minutes, every one of them served in 4 ms. The film was dead
+        // and no line said why.
+        //
+        // So the stamp follows the playlist whenever the two disagree by more
+        // than a player will bridge. hls.js bridges up to `maxBufferHole`,
+        // which it defaults to 0.5 s — that is the player's own published
+        // figure, not a number chosen here. Within it the file's own position
+        // is kept, because it is the honest one and it is what keeps speech and
+        // subtitles together on a file whose index is slightly out (2026-08-06,
+        // 4.17 s of drift on a Matroska index that lied).
+        // STAMPED WITH ITS OWN TRUE START, always. Two attempts at moving it
+        // toward the playlist both made things worse, and the reason is in what
+        // the first segment of a run is: it is not CUT at all — it begins where
+        // ffmpeg's seek landed. The picture must land on a keyframe; the sound
+        // needs none and starts at the instant asked for. So after every
+        // restart the two runs genuinely begin at different real times, and the
+        // whole run carries that difference (field 2026-08-17: the sound's
+        // #292 began at 1587.892 s and #293 at 1592.692 s — exactly one segment
+        // apart, the whole run shifted 2.5 s from the grid).
+        //
+        // Labelling each track with its own true time is therefore what keeps
+        // picture and sound together in real time. Moving them onto the
+        // published grid — separately (2.24.1) or by one family offset (2.25.0)
+        // — closes a gap that is real and opens one that is not: it desynced
+        // playback in the field within the hour, twice.
+        //
+        // What that leaves unsolved is the reason those attempts were made: a
+        // playlist that disagrees with the media by more than a player bridges
+        // makes hls.js refetch the same fragment for ever (1908 times each for
+        // two segments, measured). The answer to THAT is to make the published
+        // grid agree with where the runs really begin — not to relabel the
+        // media. Recorded as its own roadmap item rather than guessed at here.
+        const stampStart = trueStart ?? publishedStart;
+        if (trueStart !== null) {
+          this.#host.outputTimes.notePlaylistDisagreement(session, index, trueStart, publishedStart);
+        }
+        const prepared = session.segmentFormat.prepareSegmentBytes(bytes, {
+          startSeconds: stampStart,
+          rawBytes: raw,
+          initBytes: this.#host.segmentStore.initOf(session.outputKey ?? "")
+        });
+        this.#host.encodeRuns.noteRunProducedSegment(session, filePath);
+        return {
+          kind: "file",
+          stream: Readable.from([prepared]),
+          contentType: session.segmentFormat.segmentContentType,
+          isPlaylist: false
+        };
+      }
+      if (!isPlaylist) {
+        this.#host.encodeRuns.noteRunProducedSegment(session, filePath);
+      }
+      return {
+        kind: "file",
+        stream: isPlaylist
+          ? createReadStream(filePath)
+          : createReadStream(filePath, { highWaterMark: SEGMENT_READ_HIGH_WATER_MARK }),
+        contentType: isPlaylist
+          ? "application/vnd.apple.mpegurl"
+          : session.segmentFormat.segmentContentType,
+        isPlaylist
+      };
+    } catch (error) {
+      if (error?.code === "ENOENT") {
+        // The file went away between the check and the read — a leftover being
+        // removed so it can be produced again. Means exactly what never having
+        // existed means.
+        return this.#holdForProduction(session, fileName, isPlaylist, options);
+      }
+      // Anything else is a fault in producing the answer. Name it: a request
+      // answered "wait" for ever tells the viewer nothing and leaves no trace
+      // of what actually happened.
+      logger.error(
+        `transcode ${session.id} could not serve ${fileName}: ${error?.message ?? error}` +
+        (error?.stack ? `\n${error.stack}` : "")
+      );
+      return {
+        kind: "failed",
+        message: `Could not serve ${fileName}: ${error?.message ?? String(error)}`
+      };
+    }
+  }
+
+  /**
+   * Whether a piece already given to a viewer can still be served from what is
+   * stored under the key of the output that made it, after that output has
+   * gone.
+   *
+   * WHY THIS EXISTS (roadmap item 97, step 11). A repeat of an address must be
+   * answered by what answered it the first time: the player holds that
+   * output's header, and a piece of another output under the same address may
+   * not decode under it. When the output itself has been disposed, its pieces
+   * and its header stay in the store until the disk needs the room — so the
+   * very bytes that were given can be given again.
+   *
+   * @param {string} goneKey
+   * @param {string} fileName
+   * @returns {boolean}
+   */
+  hasStoredPiece(goneKey, fileName) {
+    const format = this.#host.segmentFormatOfKey(goneKey);
+    if (!format) {
+      return false;
+    }
+    if (format.initFileName !== null && fileName === format.initFileName) {
+      return Boolean(this.#host.segmentStore.initOf(goneKey));
+    }
+    if (!format.isSegmentFileName(fileName)) {
+      return false;
+    }
+    const index = format.segmentIndexFromName(fileName);
+    return this.#host.segmentStore.isClosed(goneKey, index) &&
+      Boolean(this.#host.segmentStore.pathOfName(goneKey, fileName));
+  }
+
+  /**
+   * That stored piece, prepared exactly as it would have been served while its
+   * output lived: its header stripped where the piece carries one, and stamped
+   * with where it truly begins or, where it does not say, where the playlist
+   * the player holds puts it.
+   *
+   * `like` is the picture the address belongs to. Every output interchangeable
+   * with it publishes one timeline, so its playlist is the gone output's too.
+   *
+   * @param {string} goneKey
+   * @param {string} likeId - The picture the address belongs to.
+   * @param {string} fileName
+   * @returns {Promise<{ kind: "file", stream: Readable, contentType: string, isPlaylist: false } | null>}
+   */
+  async storedPieceOf(goneKey, likeId, fileName) {
+    const like = isOutputName(likeId) ? this.#host.outputs.get(likeId) : null;
+    const format = this.#host.segmentFormatOfKey(goneKey);
+    const spec = this.#host.specOfKey(goneKey);
+    if (!format || !spec || !like || !this.hasStoredPiece(goneKey, fileName)) {
+      return null;
+    }
+    const answer = (bytes, contentType) => ({
+      kind: "file",
+      stream: Readable.from([bytes]),
+      contentType,
+      isPlaylist: false
+    });
+    if (format.initFileName !== null && fileName === format.initFileName) {
+      return answer(this.#host.segmentStore.initOf(goneKey), format.initContentType);
+    }
+    const index = format.segmentIndexFromName(fileName);
+    let raw;
+    try {
+      raw = await readFile(this.#host.segmentStore.pathOfName(goneKey, fileName));
+    } catch {
+      return null;
+    }
+    if (!format.needsSegmentRewrite) {
+      return answer(raw, format.segmentContentType);
+    }
+    const view = { segmentFormat: format, spec, timeline: like.timeline };
+    const selfContained = cutsAtGivenTimes(view);
+    const bytes = selfContained && format.stripInit ? format.stripInit(raw) : raw;
+    const trueStart = selfContained ? format.readSegmentStartSeconds?.(raw) ?? null : null;
+    const prepared = format.prepareSegmentBytes(bytes, {
+      startSeconds: trueStart ?? this.#host.outputTimes.publishedStartTime(like, index),
+      rawBytes: raw,
+      initBytes: this.#host.segmentStore.initOf(goneKey)
+    });
+    logger.info(`[hold] ${fileName} served from the stored pieces of ${goneKey}, the output that first answered it`);
+    return answer(prepared, format.segmentContentType);
+  }
+
+  /**
+   * Whether a piece of `liveKey` may stand where a piece of `goneKey` was
+   * given: the two headers compared by the product's own rule
+   * (`init-compat.js`), which refuses anything it cannot show a decoder
+   * ignores. A header not yet made is no proof, and so is a refusal.
+   *
+   * @param {string} goneKey
+   * @param {string} liveKey
+   * @returns {{ compatible: boolean, differences: string[] }}
+   */
+  headersCompatible(goneKey, liveKey) {
+    const verdict = this.#host.compareInits(
+      this.#host.segmentStore.initOf(goneKey),
+      this.#host.segmentStore.initOf(liveKey)
+    );
+    return { compatible: verdict.compatible === true, differences: verdict.differences ?? [] };
+  }
+
+  /**
+   * Say WHY a segment is being held, at most once every few seconds per file.
+   *
+   * A hold is silent today, and that silence has now cost three releases: a
+   * file that exists, a route that answers "not yet", and nothing anywhere
+   * saying which of the several reasons applied. Measured 2026-08-09: a run
+   * begun mid-file at segment #317 produced two minutes of video from #317
+   * upwards at 10.5x, and #317 itself was held 46 s and then answered 404 once
+   * the browser had given up — with not one line about the cause.
+   *
+   * @param {HlsSession} session
+   * @param {string} fileName
+   * @param {string} reason
+   * @returns {void}
+   */
+  #explainHold(session, fileName, reason) {
+    const now = Date.now();
+    const state = this.#stateFor(session);
+    const last = state.holdExplainedAt.get(fileName) ?? 0;
+    if (now - last < 5_000) {
+      return;
+    }
+    state.holdExplainedAt.set(fileName, now);
+    const index = session.segmentFormat.segmentIndexFromName(fileName);
+    // What the encoder has actually DONE since it restarted. "Alive at the right
+    // index" was as far as the old line went, and it left the two possible
+    // causes indistinguishable: an encoder waiting for torrent pieces looks
+    // exactly like one that is encoding and simply has not finished. The
+    // difference is whether its position has moved at all.
+    // Where this run began, from the run — the same reckoning
+    // `processedSeconds` is counted in. A table lookup here can disagree with
+    // it by the distance between the two grids, which is enough to print a
+    // negative "produced" and send the reader after the torrent when the
+    // encoder is the subject.
+    const progress = this.#host.encodeRuns.progressOf(session, index);
+    const runStartSeconds = Number.isFinite(progress?.startPositionSeconds)
+      ? progress.startPositionSeconds
+      : this.#host.runStartTimeFor(session, this.#host.encodeRuns.earliestStartOf(session) ?? 0);
+    const position = Number(progress?.processedSeconds);
+    const produced = Number.isFinite(position) ? position - runStartSeconds : null;
+    const speed = progress?.speed || "n/a";
+    logger.warn(
+      `transcode ${session.id} holding ${fileName}: ${reason} ` +
+      `(runs from #${this.#host.encodeRuns.earliestStartOf(session) ?? "?"}, viewer at #${this.#host.outputTimes.segmentIndexForTime(session, this.#host.viewerSecondsOn(session))}, ` +
+      `encoder ${this.#host.encodeRuns.liveRunsOf(session).length > 0 ? "alive" : "stopped"}, index #${index}, ` +
+      `produced ${produced === null ? "nothing yet — no position reported" : `${produced.toFixed(1)}s`} ` +
+      `at ${speed}${produced !== null && produced <= 0 ? " — the encoder has not moved, so it is waiting on its input" : ""})`
+    );
+  }
+
+  /**
+   * What an absent file answers when production has failed.
+   *
+   * @param {object} session
+   * @returns {{ kind: "failed", message: string } | { kind: "warming-up" }}
+   */
+  #failedOrWarming(session) {
+    if (!this.#host.encodeRuns.hasFailed(session)) {
+      return { kind: "warming-up" };
+    }
+    return {
+      kind: "failed",
+      message: this.#host.encodeRuns.lastErrorOf(session) || "ffmpeg failed for this transcode session."
+    };
+  }
+
+  #holdForProduction(session, fileName, isPlaylist, options) {
+    // The file is not there and production has ended in an error: nothing is
+    // going to write it, so the viewer is told instead of held.
+    if (this.#host.encodeRuns.hasFailed(session)) {
+      return this.#failedOrWarming(session);
+    }
+    if (this.#host.encodeRuns.isWaitingForInput(session)) {
+      // The data went away and is being fetched again. Holding the request is
+      // the truthful answer: nothing is broken and there is nothing for the
+      // viewer to retry. Only what is NOT on disk is held for it — a piece that
+      // exists is whole, and it is served whatever the encoder is doing.
+      return { kind: "warming-up" };
+    }
+    /** @type {{ address: string, rank: number, topRank: number } | null} */
+    let ranked = null;
+    if (!isPlaylist) {
+      this.#explainHold(session, fileName, "the file is not on disk");
+    }
+    // A segment was requested that ffmpeg has not produced yet.  Decide whether
+    // to wait for the current encode run to reach it or to restart the encoder
+    // at this position (server-side seeking).  The caller long-polls.
+    if (!isPlaylist) {
+      const requestedIndex = session.segmentFormat.segmentIndexFromName(fileName);
+      // Unanswerable, and known to be: behind a run that only moves forward,
+      // too far behind for the repair to fetch it, and no seek on its way to
+      // move the encoder there. Holding it changes nothing about whether it can
+      // be produced — it only spends the player's patience.
+      //
+      // This is what a track change costs when it is held instead: measured
+      // 2026-08-15, hls.js asked the new track for segment #0 while the run was
+      // at #354, the request was held for the full minute, and only when it
+      // failed did the player move to the segment it actually needed — 63 s of
+      // spinner after a track that had been made ready in 7.
+      //
+      // WHETHER ANYBODY IS COMING FOR IT, asked of the encoding (`rankAt`,
+      // which also tells "in nobody's zone" from "no map yet"). The same walk
+      // was written out here and could answer only yes or no, so the RANK was
+      // discarded at the one point where a viewer measurably waits for a named
+      // segment; it goes back out with the answer, because the wait is measured
+      // by whoever holds the request.
+      const address = session.outputKey ?? "";
+      const { rank, topRank } = this.#host.encodeOrchestrator.rankAt(address, requestedIndex);
+      ranked = { address, rank, topRank };
+      const nobodyIsComing = topRank > 0 && rank === 0;
+      if (
+        Number.isFinite(requestedIndex) &&
+        requestedIndex < (this.#host.encodeRuns.earliestStartOf(session) ?? 0) &&
+        nobodyIsComing &&
+        this.#host.encodeRuns.liveRunsOf(session).length > 0
+      ) {
+        logger.info(
+          `transcode ${session.id} segment #${requestedIndex} is ${(this.#host.encodeRuns.earliestStartOf(session) ?? 0) - requestedIndex} ` +
+          "segments behind the run and in nobody's zone; answered as absent rather than held"
+        );
+        return { kind: "not-found", ranked };
+      }
+      this.#noteWanted(session, requestedIndex);
+    }
+    return { kind: "warming-up", ranked };
+  }
+
+  /**
+   * Record that a segment nobody has made yet is wanted, and say so when it lies
+   * behind every encoder of this output.
+   *
+   * A REQUEST STEERS NO ENCODER. This used to be where one did: a request far
+   * from the encode head restarted ffmpeg there, and in the field one seek
+   * produced nine restarts in a minute, because a player holds a couple of
+   * dozen requests open at once and none of them is "the one the viewer ended
+   * on". What is missing in front of a viewer is stated by the priority map,
+   * and placing encoders on it is the plan's work. What is left here is a
+   * record for the restart accounting and a line for whoever reads the log.
+   *
+   * @param {HlsSession} session
+   * @param {number} index
+   * @returns {void}
+   */
+  #noteWanted(session, index) {
+    // When this segment was FIRST asked for and nobody was producing it. The
+    // restart itself costs 0.7-1.3 s (measured 2.9.132), while a seek costs
+    // 5-8 s end to end — so most of the wait happens before a restart is even
+    // decided on, and that is what this records.
+    this.#host.encodeRuns.noteWanted(session, index);
+    if (!this.#host.encodeRuns.isLive(session) || index < 0) {
+      return;
+    }
+    const head = this.#host.encodeRuns.earliestStartOf(session) ?? 0;
+    if (index < head) {
+      // Nothing running goes backwards, so nothing running will make it.
+      this.#explainHold(
+        session,
+        session.segmentFormat.segmentFileName(index),
+        `it is behind the run (#${head}); where the viewers are is what places encoders`
+      );
+    }
+  }
+
+  /**
+   * The init header, lifted out of the first segment that exists.
+   *
+   * Needed only on the explicit-cut path, where the muxer produces no init file
+   * of its own. Scans rather than assuming segment 0: a run started by a seek
+   * begins at whatever index the viewer asked for.
+   *
+   * @param {HlsSession} session
+   * @returns {Promise<Buffer | null>}
+   */
+
+  async #initFromFirstSegment(session) {
+    if (typeof session.segmentFormat.extractInit !== "function") {
+      return null;
+    }
+    // How many tracks a complete header must declare is ANSWERED, not assumed.
+    //
+    // The probe already knows the source's stream list, and the output maps at
+    // most one of each (`-map 0:v:0? -map 0:a:0?`), so the count follows from
+    // what the source actually has. A film with no soundtrack expects one; an
+    // ordinary file expects two; neither is a convention.
+    //
+    // Deriving it from the produced pieces instead — the first version of this
+    // — reads correctly only once a piece carrying every track exists, and the
+    // whole point is the moment BEFORE that: early pieces written before the
+    // video was muxed would set the requirement to one and wave through exactly
+    // the header this exists to reject. The pieces are still consulted, but
+    // only as a floor: a piece carrying more than the probe led us to expect is
+    // evidence, and evidence outranks the probe.
+    const declared = this.#host.declaredTracks(session);
+    let expectedTracks = (declared.video ? 1 : 0) + (declared.audio ? 1 : 0);
+    if (expectedTracks === 0) {
+      // Nothing to consult. Fall back to the evidence, with its known lag.
+      expectedTracks = 1;
+    }
+    let best = null;
+    let bestTracks = 0;
+    /** @type {Map<string, Buffer>} Pieces read once and used for both passes. */
+    const pieces = new Map();
+    let names;
+    try {
+      names = this.producedNumbers(session).map((index) => session.segmentFormat.segmentFileName(index));
+    } catch {
+      return null;
+    }
+    // First pass: what do the produced pieces actually carry? The answer is the
+    // requirement — no assumption about the source is involved.
+    if (typeof session.segmentFormat.countSegmentTracks === "function") {
+      for (const name of names) {
+        try {
+          // The first copy WITH BYTES IN IT, not the first name: a run stopped
+          // with a piece open leaves an empty file under the same name, and
+          // taking that one skips a number whose header is sitting in the run
+          // before it.
+          const found = await this.#firstCopyWithBytes(session, name);
+          if (!found) {
+            continue;
+          }
+          const bytes = await readFile(found);
+          pieces.set(name, bytes);
+          expectedTracks = Math.max(expectedTracks, session.segmentFormat.countSegmentTracks(bytes));
+        } catch {
+          // Being written right now — it says nothing about the others.
+        }
+      }
+    }
+
+    for (const name of names) {
+      try {
+        const cached = pieces.get(name);
+        const found = cached ? name : await this.#firstCopyWithBytes(session, name);
+        if (!found) {
+          continue;
+        }
+        const ownInit = session.segmentFormat.extractInit(cached ?? await readFile(found));
+        const init = ownInit && session.segmentFormat.prepareSharedInit ?
+          session.segmentFormat.prepareSharedInit(ownInit) : ownInit;
+        if (!init || init.length === 0) {
+          continue;
+        }
+        // The requirement computed above is APPLIED here. It was computed and
+        // then ignored: this loop returned the first header it found, so a
+        // piece written before the video was muxed supplied an audio-only
+        // header — and that header is cached for the session's whole life,
+        // because the player fetches `#EXT-X-MAP` once. Measured 2026-08-11:
+        // `videoWidth=0`, `totalVideoFrames=0`, `readyState=4` — sound playing
+        // and no picture, for as long as the session lasted.
+        const tracks = typeof session.segmentFormat.countInitTracks === "function"
+          ? session.segmentFormat.countInitTracks(init)
+          : expectedTracks;
+        if (tracks >= expectedTracks) {
+          return init;
+        }
+        if (tracks > bestTracks) {
+          best = init;
+          bestTracks = tracks;
+        }
+      } catch {
+        // Being written right now — try the next one.
+      }
+    }
+    if (best !== null) {
+      // Nothing carried the full set. The source is probably missing a stream;
+      // serving the richest header found is right, and saying so makes the
+      // other possibility — every piece so far written before the video was
+      // muxed — visible rather than silent.
+      logger.warn(
+        `transcode ${session.id} no piece declared ${expectedTracks} tracks; ` +
+        `serving an init with ${bestTracks}`
+      );
+      return best;
+    }
+    return best;
+  }
+
+  /**
+   * A produced file that has anything in it, or null.
+   *
+   * For callers that need a file's CONTENTS and cannot judge them — deriving
+   * the session's header is the case, since the header is what judging would
+   * need. An empty file answers no question, so it is passed over.
+   *
+   * @param {HlsSession} session
+   * @param {string} fileName
+   * @returns {Promise<string | null>}
+   */
+  async #firstCopyWithBytes(session, fileName) {
+    const held = this.#host.segmentStore.pathOfName(session.outputKey ?? "", fileName);
+    if (held === null) {
+      return null;
+    }
+    try {
+      const info = await stat(held);
+      return info.size > 0 ? held : null;
+    } catch {
+      return null;
+    }
+  }
+
+  /**
+   * Record where one viewer of this session is, and answer with the furthest
+   * any of them has reached.
+   *
+   * The furthest is what the single encoder is steered by: it has to serve
+   * everyone, and what lies behind the leader has already been produced and is
+   * served from disk without a wait. The individual positions exist for the
+   * opposite question — whether a particular held request is still wanted —
+   * which cannot be answered from a shared field.
+   *
+   * A REQUEST IS NOT A POSITION. It says the viewer is still here and nothing
+   * more: where they are is what they themselves state, on the viewer, and a
+   * request cannot reach it. Asking for a segment used to write the position,
+   * so two writers filled one field in turn and the priority map jumped
+   * backwards several times a second — measured 2026-09-13, 77 encoder starts
+   * and 141 stops in six minutes while both viewers sat frozen.
+   *
+   * A viewer is forgotten once nothing has been heard from them for longer than
+   * any silence a watching viewer can produce. The figure is the proxy's own
+   * look-ahead, not a chosen interval.
+   *
+   * @param {HlsSession} session
+   * @param {string} consumerId
+   * @returns {void}
+   */
+  #noteViewerSeen(session, consumerId) {
+    // A request that names nobody is evidence about nobody.
+    if (!consumerId) {
+      return;
+    }
+    this.#host.viewers.of(session, consumerId).seen();
+  }
+
+  /**
+   * Whether a held request is for a segment the viewer STILL needs.
+   *
+   * The epoch alone says a seek happened; it cannot say whether this particular
+   * request was made for the position left behind or for the one just arrived
+   * at. That distinction is the whole of the failure measured 2026-08-18: the
+   * viewer seeked to 1061.0 s, the request for `segment-00101` — the segment AT
+   * that position — raced the seek notification, the epoch moved underneath it,
+   * and it was answered 503 twice within 80 ms. The player then hunted at
+   * sn=105-107, never came back to 101, and looped two audio segments 1473
+   * times over 149 s while the picture stood still.
+   *
+   * A request is stale when its segment lies behind where the viewer now is, or
+   * so far ahead that the running encode will not reach it. Anything between is
+   * exactly what the viewer is waiting for, and holding it is the point.
+   *
+   * "So far ahead" is the encoder's own look-ahead, measured on this session's
+   * own cut grid — the same figure the browser sizes its forward buffer from.
+   * It used to be `MAX_LOOKAHEAD_SEGMENTS`, which is eight segments ahead of
+   * the ENCODE HEAD
+   * and has nothing to do with how far ahead of the VIEWER a request may
+   * legitimately sit; it happened to match a browser holding 30 s, and would
+   * have refused three quarters of the requests of one holding the whole
+   * cushion (roadmap item 4).
+   *
+   * Judged against the position of the viewer who MADE the request, when the
+   * transport carries who that is. A session is shared by everyone watching a
+   * copied picture and the epoch is per session, so a seek by the viewer in
+   * front used to release every request being held for the viewer behind them.
+   *
+   * @param {string} sessionId
+   * @param {string} fileName
+   * @param {string} [consumerId] - Who is asking. Without it the one shared
+   *   position decides, which is what a single viewer means anyway.
+   * @returns {boolean} True when the request should keep waiting.
+   */
+  requestStillWanted(sessionId, fileName, consumerId = "") {
+    const session = isOutputName(sessionId) ? this.#host.outputs.get(sessionId) : null;
+    if (!session) {
+      return false;
+    }
+    const index = session.segmentFormat?.segmentIndexFromName?.(fileName) ?? -1;
+    if (!(index >= 0)) {
+      return true; // a playlist or an init segment belongs to no position
+    }
+    const position = this.#host.viewerSecondsOn(session, consumerId);
+    const at = this.#host.outputTimes.segmentIndexForTime(session, position);
+    // The far edge on THIS session's own grid rather than a count of nominal
+    // segments: a copied picture is cut at the source's keyframes, so its
+    // segments are not four seconds long and dividing by that figure would put
+    // the edge somewhere else entirely. The segment CONTAINING the edge is
+    // wanted — it is the one the deepest allowed request lands in, and it
+    // already reaches past the cushion by whatever is left of its own duration.
+    const edge = this.#host.outputTimes.segmentIndexForTime(session, position + this.#host.lookaheadSeconds);
+    return index >= at && index <= edge;
+  }
+
+  /**
+   * How many times the viewer has moved since this session started.
+   *
+   * A request being held for a segment answers "retry" as soon as this changes,
+   * because it was made for a position the viewer has left — see `requestSeek`.
+   *
+   * @param {string} sessionId
+   * @returns {number}
+   */
+  seekEpoch(sessionId) {
+    const session = isOutputName(sessionId) ? this.#host.outputs.get(sessionId) : null;
+    return session ? this.#stateFor(session).waitEpoch : 0;
+  }
+
+  /**
+   * Wait for the requested segment to be published, without polling the disk.
+   *
+   * Ends on the publication, on the waits of this output being invalidated, on
+   * the deadline, or when the requester goes — and whichever ends it, the
+   * waiter is taken out of the store, so a wait with no deadline cannot outlive
+   * the request it was for. Returns false at once for playlists and init files,
+   * whose readiness has a different owner.
+   *
+   * @param {string} sessionId
+   * @param {string} fileName
+   * @param {number} timeoutMs - May be infinite.
+   * @param {Promise<unknown> | null} [requesterGone]
+   * @returns {Promise<boolean>} Whether it is worth asking for the file again.
+   */
+  async waitForSegment(sessionId, fileName, timeoutMs, requesterGone = null) {
+    const session = isOutputName(sessionId) ? this.#host.outputs.get(sessionId) : null;
+    const index = session?.segmentFormat?.segmentIndexFromName?.(fileName) ?? -1;
+    if (!session || index < 0) return false;
+    const state = this.#stateFor(session);
+    let wake;
+    const invalidated = new Promise((resolve) => {
+      wake = resolve;
+      state.waitListeners.add(resolve);
+    });
+    let finished = () => {};
+    const over = new Promise((resolve) => {
+      finished = resolve;
+    });
+    try {
+      return await Promise.race([
+        this.#host.segmentStore.waitFor(session.outputKey ?? "", index, timeoutMs, over),
+        invalidated,
+        ...(requesterGone ? [requesterGone.then(() => false)] : [])
+      ]);
+    } finally {
+      state.waitListeners.delete(wake);
+      finished();
+    }
+  }
+
+  /**
+   * Poll until the HLS playlist file exists and contains a valid `#EXTM3U`
+   * header, or until the session fails, or until the startup timeout elapses.
+   * Throws with message `"HLS playlist is still warming up."` on timeout.
+   *
+   * @param {HlsSession} session
+   * @returns {Promise<void>}
+   */
+  async waitUntilReady(session) {
+    // With a synthetic VOD playlist there is nothing to wait for: the playlist
+    // is generated from the probed duration and is available immediately.
+    // Individual segments are long-polled by the segment route as ffmpeg
+    // produces them.
+    if (session.useSyntheticPlaylist) {
+      if (this.#host.encodeRuns.hasFailed(session)) {
+        throw new Error(this.#host.encodeRuns.lastErrorOf(session) || "ffmpeg failed to start HLS session.");
+      }
+      return;
+    }
+
+    const playlistPath = path.join(this.#host.segmentStore.pathFor(session.outputKey ?? ""), PLAYLIST_FILE_NAME);
+    const deadline = Date.now() + this.#host.startupWaitMs;
+
+    while (Date.now() < deadline) {
+      if (this.#host.encodeRuns.hasFailed(session)) {
+        throw new Error(this.#host.encodeRuns.lastErrorOf(session) || "ffmpeg failed to start HLS session.");
+      }
+      try {
+        await access(playlistPath);
+        const text = await readFile(playlistPath, "utf8");
+        if (text.includes("#EXTM3U")) {
+            return;
+        }
+      } catch (_error) {
+        // Playlist is not ready yet.
+      }
+      await delay(250);
+    }
+
+    throw new Error("HLS playlist is still warming up.");
+  }
+
+  /**
+   * Answer the player's report that a delivered fragment sits far from the edge
+   * of its buffer, with the one fact only this side holds: which boundary the
+   * segment of that number really begins at.
+   *
+   * The player can say the gap; it cannot say whether the cause is its own
+   * loading or a run whose output no longer matches its numbering. Here both
+   * are in hand — the time the playlist gave that segment, and, when the
+   * segment has been served, the time it truly began at — so the line either
+   * names a shifted run or clears this side of it.
+   *
+   * Diagnostic only: nothing is repositioned on the strength of a browser's
+   * reading, deliberately, because a wrong answer here would restart an encoder
+   * the viewer is waiting on.
+   *
+   * @param {string} sessionId
+   * @param {{ sn: number, track?: string, fragStartSec: number, bufferEndSec: number, currentTimeSec: number }} report
+   * @returns {boolean} False when no such session exists.
+   */
+  recordFragmentFar(sessionId, { sn, track, fragStartSec, bufferEndSec, currentTimeSec }) {
+    const named = this.#host.outputs.get(sessionId);
+    if (!named) {
+      return false;
+    }
+    // Which of the two streams the report is about. The browser addresses
+    // everything to the video session's id — the soundtrack is served under
+    // `/a/<n>/` on that same id — but it is a session of its own, with its own
+    // run and its own position, and that is exactly the pair this report exists
+    // to tell apart. Answering an audio report from the picture's records would
+    // state, confidently, something about the wrong stream.
+    const onScreen = this.#host.activeOutputFor({ base: named, outputs: this.#host.outputs });
+    const session = track === "audio"
+      ? ([...this.#host.outputs.familyOf(onScreen)].find((member) => member.spec.carries === "audio-only") ?? onScreen)
+      : onScreen;
+    const gap = fragStartSec - bufferEndSec;
+    const declared = this.#host.outputTimes.publishedStartTime(session, sn);
+    const trueStart = this.#host.outputTimes.trueStartAt(session, sn);
+    const verdict = trueStart === undefined
+      // Where a segment truly began is only ever read off one that was cut on
+      // an explicit list — a uniform grid has nothing to read back — so this is
+      // "not recorded", which is not the same as "not produced", and the line
+      // must not claim the second.
+      ? "where that segment began is not recorded on this side, so the gap cannot be attributed here"
+      : (() => {
+        const at = this.#host.outputTimes.boundaryIndexAt(session, trueStart, this.#host.publishedGridFor(session));
+        if (at === null) {
+          return `it really began at ${trueStart.toFixed(3)}s, which is no boundary of this grid`;
+        }
+        if (at === sn) {
+          return `it really began at boundary #${sn}, where it should — the gap is not this run's`;
+        }
+        return `it really began at boundary #${at}, ${sn - at} place(s) before its own number — ` +
+          "this run's output does not match its numbering";
+      })();
+    logger.warn(
+      `transcode ${session.id} the player is stuck: ${session.spec.carries === "audio-only" ? "sound" : "picture"} ` +
+      `fragment #${sn} starts ${gap.toFixed(1)}s past ` +
+      `the end of its buffer (${bufferEndSec.toFixed(1)}s, viewer at ${currentTimeSec.toFixed(1)}s, ` +
+      `the playlist puts it at ${declared.toFixed(3)}s) — ${verdict}`
+    );
+    return true;
+  }
+
+  /**
+   * Every segment number this session's OUTPUT holds, whoever produced it.
+   *
+   * Public because it is the one thing worth asserting about the address
+   * change: two sessions of one output answer with the same list, including
+   * segments the other one's encoder made.
+   *
+   * @param {HlsSession} session
+   * @returns {number[]}
+   */
+  producedSegmentNumbers(session) {
+    return new Set(this.producedNumbers(session));
+  }
+
+  /**
+   * What this session's OUTPUT holds, asked of the one thing that owns it.
+   *
+   * There were two owners of this fact over one directory: the store, addressed
+   * by the output's own key, and a `ProducedIndex` built per SESSION over a
+   * path the store had handed out. Two viewers of one output therefore built
+   * two indexes over one directory, each with its own idea of what was in it.
+   *
+   * @param {HlsSession} session
+   * @returns {number[]} Every segment number it holds, in order.
+   */
+  producedNumbers(session) {
+    return this.#host.segmentStore.provenNumbers(session.outputKey ?? "");
+  }
+}