@torrent-tv/proxy 2.87.9 → 2.87.11

package/services/server/ViewerRequests.js

@@ -1,522 +1,703 @@
-/**
- * What a viewer asks of this proxy, one request at a time.
- *
- * Open an output of a file — the one that already exists when its parameters
- * agree — and be placed on it; state a new position; ask how far the material
- * in front of them has got. Each request states a fact about the viewer and
- * returns; where encoders go follows from those facts and is the plan's.
- */
-
-import { logger } from "../../utils/logger.js";
-import { isOutputName } from "../encode/output/index.js";
-
-function isWarmupTimeoutError(error) {
-  if (!(error instanceof Error)) {
-    return false;
-  }
-  return error.message === "HLS playlist is still warming up.";
-}
-
-/**
- * The segment number an assignment is recorded under for this file, −1 for the
- * init, or null for anything that is neither (a playlist).
- *
- * @param {object} output
- * @param {string} fileName
- * @returns {number | null}
- */
-function segmentAddressOf(output, fileName) {
-  const format = output.segmentFormat;
-  if (!format) {
-    return null;
-  }
-  if (format.initFileName !== null && fileName === format.initFileName) {
-    return -1;
-  }
-  if (format.isSegmentFileName(fileName)) {
-    return format.segmentIndexFromName(fileName);
-  }
-  return null;
-}
-
-export class ViewerRequests {
-  /** What this reads and asks of the rest of the proxy, and nothing else. @type {object} */
-  #host;
-
-  /**
-   * @param {object} host - `opening`, `activeOutputFor`, `viewerSecondsOn`, `minimumBufferSecondsFor`, `disposeSession`, `expectedFirstSegmentMs`, `expectedSessionCreateMs`, `planEncodersSoon`, `waitUntilReady`, `encodeRuns`, `outputTimes`, `outputs`, `quality`, `qualityOffer`, `segmentDurationSec`, `startupWaitMs`, `viewers`, `acceptsGeneration`, `generationOfRequest`, `noteGivenOutput`, `holdForResponse`, `noteServingVerdict`, `servingVerdictOf`
-   */
-  constructor(host) {
-    this.#host = host;
-  }
-
-  /**
-   * Open an output of a file for this viewer and put them on it.
-   *
-   * Which output answers is the encoding component's decision
-   * (`encode/OutputOpening.js`), made without knowing who asked. What is this
-   * operation's is the requester: they are registered on the output with the
-   * sound they chose and placed where their request says they are, and the
-   * answer waits for the output to be ready to play.
-   *
-   * Throws with `error.code === "TRANSCODE_DISABLED"` when transcoding is
-   * disabled on this proxy instance.
-   *
-   * @param {object} request - What `OutputOpening.open` takes, plus the viewer.
-   * @param {string} [request.consumerId] - Who asks. Empty for an output a
-   *   picture opens on its own behalf (a quality step, a soundtrack), which
-   *   places nobody on it.
-   * @returns {Promise<import("../encode/output/EncodedOutput.js").EncodedOutput>}
-   */
-  async createOrGetSession({ consumerId = "", ...request }) {
-    // What this viewer's link last measured, where the request does not say:
-    // a viewer who comes back to open another film has a link this proxy has
-    // already measured, and the output they are given is judged by it.
-    const known = consumerId ? this.#host.viewers.get(consumerId) : null;
-    const startPositionSeconds = request.startPositionSeconds ?? 0;
-    // Whether they were already on the output this request lands on, read at
-    // the moment they are put on it.
-    let alreadyOn = false;
-    // THE PLACE ON THIS MACHINE, AND THE VIEWER IN IT, in one synchronous call
-    // the opening makes the moment it has the output (roadmap item 97, step 14).
-    // Asked and placed together, so the next viewer to open anything is asked
-    // with this one counted: two openings cannot both be told there is room
-    // for one. Placed now for the second reason it always was — a viewer who
-    // has not been placed states no want, and an output all of whose viewers
-    // state nothing has every encoder on it stopped.
-    const claim = consumerId
-      ? (candidate) => {
-          alreadyOn = this.#host.viewers.forOutput(candidate).has(consumerId);
-          const answer = alreadyOn
-            ? { admitted: true, reason: "already watching it", speedX: null }
-            : this.#host.admitsWatching(candidate);
-          if (answer.admitted) {
-            this.placeViewer(candidate, this.#host.viewers.of(candidate, consumerId), startPositionSeconds);
-          }
-          return answer;
-        }
-      : null;
-    const { output, existed, audio, verdict } = await this.#host.opening.open({
-      ...request,
-      viewerLinkMbps: request.viewerLinkMbps ?? known?.linkReading()?.linkMbps ?? null,
-      claim
-    });
-    if (consumerId) {
-      const joined = existed && !alreadyOn;
-      // What THIS viewer wants of the sound, which an output they are joining
-      // knows nothing about: they may have chosen another language, and their
-      // browser may need a track re-encoded that the first viewer's could
-      // decode as it stands.
-      const viewer = this.#host.viewers.of(output, consumerId);
-      viewer.audio = audio;
-      // The picture as their page sees it, which bounds the height made for
-      // them from here on (roadmap item 98).
-      viewer.noteVisiblePicture(request.visiblePicture);
-      // On what their output was judged. Recorded where the output is opened
-      // for them; a step opened on their behalf is recorded by `Renditions`.
-      this.#host.noteServingVerdict(consumerId, verdict ? { ...verdict, outputKey: output.outputKey } : null);
-      // WHERE they are was stated when they were put on it, by `claim` above:
-      // a viewer joining an output already playing at 40:00 may be opening the
-      // film from a link that carries 05:00.
-      // Reuse said nothing at all once, so an output serving two viewers looked
-      // exactly like one serving a single viewer — and the whole question its
-      // key exists to answer is which of the two happened.
-      if (joined) {
-        logger.info(
-          `transcode ${output.id} joined by ${consumerId} ` +
-          `(${this.#host.viewers.forOutput(output).size} viewer(s)) key=${output.outputKey}`
-        );
-      }
-    }
-    this.#host.outputs.touch(output);
-    if (!existed) {
-      // Where the first encoder goes is the plan's, from where the viewers stand.
-      this.#host.planEncodersSoon();
-    }
-    try {
-      await this.#host.waitUntilReady(output);
-    } catch (error) {
-      if (!existed && this.#host.encodeRuns.hasFailed(output)) {
-        await this.#host.disposeSession(output.id);
-        throw error;
-      }
-      if (existed && !isWarmupTimeoutError(error)) {
-        throw error;
-      }
-      // Not a failure of the request: the synthetic playlist is already there
-      // and segments appear as the encoder produces them.
-    }
-    return output;
-  }
-
-  /**
-   * Put a viewer where their own request says they are.
-   *
-   * A viewer arrives by asking for a POSITION — zero, or the time an address
-   * bar carried — so "we do not know where they are" is not a state a viewer
-   * can be in. Before 2026-09-05 it was: position was written only by a segment
-   * request, so a viewer counted as placeless until they had asked for a
-   * segment, and an output whose viewers were all placeless had every encoder
-   * on it stopped for having nobody. The soundtrack that failed that day could
-   * not have asked: the segment it would have asked for needed an `init.mp4`
-   * that the stopped encoder was going to make.
-   *
-   * Only ever places a viewer who has none. A viewer already placed is being
-   * kept current by their own requests and seeks, and a fresh create request
-   * carries a default of zero that must not drag them back to the beginning.
-   *
-   * @param {HlsSession} session
-   * @param {import("../viewer/Viewer.js").Viewer} viewer
-   * @param {number} positionSeconds
-   * @returns {void}
-   */
-  /**
-   * Put a viewer named by id onto this output, where they have no position yet.
-   *
-   * The same operation as {@link placeViewer}, addressed the way another
-   * component may address it: by name. Encoding asks for this when a request
-   * for any file of an output is the first sign of the person behind it, and it
-   * must not be handed the viewer to do it — a viewer's record has one writer,
-   * and that is the viewer layer.
-   *
-   * @param {object} output
-   * @param {string} consumerId
-   * @param {number} positionSeconds
-   * @returns {void}
-   */
-  placeViewerOn(output, consumerId, positionSeconds) {
-    if (!output || !consumerId) {
-      return;
-    }
-    this.placeViewer(output, this.#host.viewers.of(output, consumerId), positionSeconds);
-  }
-
-  placeViewer(session, viewer, positionSeconds) {
-    if (viewer.position !== null) {
-      return;
-    }
-    const seconds = Number.isFinite(positionSeconds) && positionSeconds > 0 ? positionSeconds : 0;
-    let segment = 0;
-    try {
-      const index = this.#host.outputTimes.segmentIndexForTime(session, seconds);
-      if (Number.isInteger(index) && index >= 0) {
-        segment = index;
-      }
-    } catch {
-      // A session whose cut table is not built yet places its viewer at the
-      // beginning, which is where the run starts anyway.
-    }
-    viewer.position = { segment, seconds, at: Date.now(), seeked: null };
-    this.#host.planEncodersSoon();
-  }
-
-  /**
-   * The viewer seeked. Called from POST /api/transcode-sessions/:id/seek with
-   * the position the browser read off its own player once the scrub ended.
-   *
-   * This is the ONLY thing that repositions the encoder. It replaces inferring
-   * the target from segment requests, which cannot work: a single seek leaves
-   * ~25 concurrent requests outstanding across a wide span (measured), so no
-   * rule over them can recover which one the viewer meant.
-   *
-   * The existing settle/cooldown/first-segment guards still apply — they
-   * protect against restarting too eagerly, which is orthogonal to knowing
-   * WHERE to restart.
-   *
-   * @param {string} sessionId
-   * @param {number} positionSeconds - Absolute position on the source timeline.
-   * @param {string} [consumerId] - Who moved.
-   * @param {number} [generation] - Which viewing their requests now belong to,
-   *   as the page states it. Not a number this side counts: see below.
-   * @returns {boolean} False when the session is unknown or disposed.
-   */
-  requestSeek(sessionId, positionSeconds, consumerId = "", generation = Number.NaN) {
-    const named = this.#host.outputs.get(sessionId);
-    if (!named || !consumerId) {
-      return false;
-    }
-    // A SEEK DOES ONE THING: it puts the viewer where they now are.
-    //
-    // It used to do eleven, and wrote that position into five places: two
-    // fields on this session, two more on the soundtrack's, and the viewer. It
-    // also asked whether the jump would drag another viewer back, started an
-    // encoder itself, cancelled outstanding requests, backed off a segment to
-    // the preceding keyframe, and set a timer to restart ffmpeg. So it was a
-    // third authority over the encoders beside the plan and the start path, and
-    // not one of its branches ever asked what had already been made — a viewer
-    // jumping into a stretch that was finished and on disk got a fresh encoder
-    // for it.
-    //
-    // What follows from the move happens by itself: the priority map is built
-    // from where the viewers are, and both orchestrators read the map.
-    // A seek that names nobody moves nobody: the route refuses it before it
-    // gets here.
-    const viewer = this.#host.viewers.of(named, consumerId);
-    viewer.moveTo(positionSeconds);
-    // WHICH VIEWING THEIR REQUESTS NOW BELONG TO, as the page states it.
-    //
-    // Not counted here: the page raises its own number before it sends this,
-    // and stamps every request with it, so counting a second one on this side
-    // would make two owners of one fact — and they would disagree exactly when
-    // a request and a seek cross, which is the case the number exists for. A
-    // page that states nothing leaves the generation where it is, which is the
-    // behaviour every transport without a loader of ours has.
-    //
-    // The generation being left is not dropped: a request already in flight is
-    // still answered by the output it was made against — see the two deadlines
-    // in `Assignments.js`.
-    viewer.assignments.statedGeneration(generation);
-    this.#host.outputs.touch(named);
-    // EVERY REQUEST THIS VIEWER HAS OUTSTANDING WAS MADE FOR WHERE THEY WERE,
-    // and hls.js keeps one fragment load outstanding per track: a request held
-    // for a segment behind them blocks the one they need now. Measured
-    // 2026-08-04: 57 s of a 58 s backward seek was that wait, and the segment
-    // they wanted was served in 15 ms once it was asked for.
-    //
-    // WAKING IS NOT CANCELLING. It moves the epoch, and each held request then
-    // asks whether it is still wanted — for the viewer who made it, at that
-    // viewer's own position. So a request of somebody else's on an output
-    // shared with them is kept, and so is this viewer's own request for the
-    // segment they have just landed on, which races this notification.
-    //
-    // EVERY OUTPUT THEY WATCH, not the one the browser addressed: their picture,
-    // the quality step on their screen and their soundtrack are three outputs
-    // with three sets of held requests, and a seek leaves all three behind.
-    for (const outputId of viewer.outputs) {
-      const watched = this.#host.outputs.get(outputId);
-      if (watched) {
-        this.#host.invalidateWaits(watched);
-      }
-    }
-    this.#host.planEncodersSoon();
-    return true;
-  }
-
-  /**
-   * Whether this viewer still takes a request made in the generation it states.
-   *
-   * ASKED BEFORE ANYTHING THE REQUEST WOULD CAUSE. Resolving a step or a
-   * soundtrack can create an output and registers the viewer on it; a request
-   * made for a viewing the viewer has already left would do both and only then
-   * be refused. So every route asks this first.
-   *
-   * @param {string} consumerId
-   * @param {number} statedGeneration - What the request carried; NaN when
-   *   nothing, which is always taken.
-   * @returns {boolean}
-   */
-  acceptsRequest(consumerId, statedGeneration) {
-    return this.#host.acceptsGeneration(consumerId, statedGeneration);
-  }
-
-  /**
-   * Record that the output a request ADDRESSED answered it.
-   *
-   * The picture's own route names its output outright, so there is nothing to
-   * choose; what is recorded is the answer, under the height that output is
-   * named after (`variantHeightOf`). That is the same height under which the
-   * step route answers with the picture itself, so the two addresses of one
-   * height lead to one output and one record. An init is recorded as segment
-   * −1. A soundtrack is not a height and records nothing: its repeat is not
-   * kept, while the check of the generation and the hold of a response still
-   * apply to it.
-   *
-   * @param {string} sessionId
-   * @param {string} consumerId
-   * @param {number} statedGeneration
-   * @param {string} fileName
-   * @returns {void}
-   */
-  noteAnsweredDirectly(sessionId, consumerId, statedGeneration, fileName) {
-    const output = isOutputName(sessionId) ? this.#host.outputs.get(sessionId) : null;
-    if (!output || !consumerId || !output.spec?.video) {
-      return;
-    }
-    const segmentIndex = segmentAddressOf(output, fileName);
-    if (segmentIndex === null) {
-      return;
-    }
-    const height = this.#host.outputs.variantHeightOf(output);
-    const generation = this.#host.generationOfRequest(consumerId, statedGeneration);
-    this.#host.noteGivenOutput(consumerId, generation, height, segmentIndex, output.outputKey ?? "");
-  }
-
-  /**
-   * A response from this output to this viewer has begun: the output is held
-   * until the function returned is called, whatever happens to the generation
-   * meanwhile. Safe to call more than once.
-   *
-   * @param {string} sessionId
-   * @param {string} consumerId
-   * @returns {() => void}
-   */
-  holdResponse(sessionId, consumerId) {
-    const output = isOutputName(sessionId) ? this.#host.outputs.get(sessionId) : null;
-    return this.#host.holdForResponse(consumerId, output?.outputKey ?? "");
-  }
-
-  /**
-   * The same hold, named by an output's KEY: a piece served from what a gone
-   * output left in the store has no live output to name it by.
-   *
-   * @param {string} outputKey
-   * @param {string} consumerId
-   * @returns {() => void}
-   */
-  holdResponseForKey(outputKey, consumerId) {
-    return this.#host.holdForResponse(consumerId, outputKey ?? "");
-  }
-
-  /**
-   * Where the viewer of this session is, in seconds.
-   *
-   * Exists so a refusal can name it. A log line that says only "superseded"
-   * cannot be read afterwards: it does not say what was refused or against what
-   * position, which is exactly what the 2026-08-18 investigation lacked.
-   *
-   * Named per viewer, because that is what the refusal is about: a request is
-   * refused for being behind where THAT viewer is, and a line naming the
-   * furthest viewer of a shared session would explain a refusal by somebody
-   * else's position.
-   *
-   * @param {string} sessionId
-   * @param {string} [consumerId]
-   * @returns {number} Zero when the session is gone or nothing has been reported.
-   */
-  viewerPositionOf(sessionId, consumerId = "") {
-    const session = isOutputName(sessionId) ? this.#host.outputs.get(sessionId) : null;
-    if (!session) {
-      return 0;
-    }
-    return this.#host.viewerSecondsOn(session, consumerId);
-  }
-
-  async getSessionProgress(sessionId, consumerId = "") {
-    if (!isOutputName(sessionId)) {
-      return null;
-    }
-    const named = this.#host.outputs.get(sessionId);
-    if (!named) {
-      return null;
-    }
-    this.#host.outputs.touch(named);
-    // Progress is asked about the stream on screen, which after a quality
-    // change is another session. Touching the named one as well is what keeps
-    // the family alive: only the ACTIVE variant gets segment requests, so
-    // without this the base session would idle out from under its own variants.
-    const session = this.#host.activeOutputFor({ base: named, consumerId, outputs: this.#host.outputs });
-    this.#host.outputs.touch(session);
-    const warmupTotalSeconds = this.#host.startupWaitMs / 1000;
-    const warmupElapsedSeconds = Math.max(
-      0,
-      (Date.now() - (this.#host.outputs.startedAt(session) ?? Date.now())) / 1000
-    );
-    // One question, one answer. Run state comes only from the encoding layer.
-    const isWarmupPhase = this.#host.encodeRuns.wireStateOf(session) === "starting";
-    const warmupPercent = isWarmupPhase
-      ? Math.max(0, Math.min(100, (warmupElapsedSeconds / warmupTotalSeconds) * 100))
-      : null;
-    const warmupRemainingSeconds = isWarmupPhase
-      ? Math.max(0, warmupTotalSeconds - warmupElapsedSeconds)
-      : null;
-    // Observed OUTPUT bitrate (Mbit/s) from recently completed segment sizes —
-    // already computed for the viewer-link budget check (#checkLinkBudget); also
-    // exposed here so the browser can turn its OWN measured link throughput into
-    // a "content-seconds delivered per wall-clock second" rate for the unified
-    // three-stage ETA (download / transcode / delivery), the same way the
-    // transcode's own `speed` already is one. Null when not enough segments yet.
-    const outputMbps = await this.#host.quality.observedStreamMbps(session);
-    const progress = this.#host.encodeRuns.progressOf(session);
-    return {
-      // The id the caller asked about, not the variant it was answered from —
-      // the browser tracks its sessions by the id it was given.
-      sessionId: named.id,
-      state: this.#host.encodeRuns.wireStateOf(session),
-      // The smallest buffer at which no interruption reaches the viewer, from
-      // THIS file's own recent interruptions: one whole segment — the one being
-      // played — plus the worst wait that can arrive before the buffer refills.
-      // On the field torrent that is 7-9 s where the browser waits for a
-      // hand-chosen 25, which is sixteen seconds of staring at a spinner that
-      // nothing had shown to be necessary. Null until the reader has seen two
-      // interruptions; the browser keeps its own figure until then.
-      minimumBufferSeconds: this.#host.minimumBufferSecondsFor(session),
-      processedSeconds: progress.processedSeconds,
-      // Bytes this session's own reads have received from the swarm.
-      //
-      // The second proof that a session is alive, and the only one available
-      // before its first frame exists: `processedSeconds` cannot move until the
-      // decoder has a frame, so on a cold start it stands at the start position
-      // for as long as the first piece takes to arrive. Field 2026-09-03 — one
-      // piece took 46.3 s while the swarm delivered 55.9 MB across the torrent,
-      // `processedSeconds` frozen at 171.3 throughout, and the browser declared
-      // the proxy dead 0.4 s before the piece landed.
-      //
-      // Counted per SESSION and not per torrent, deliberately: in that same
-      // episode the torrent received 55.9 MB while the picture's own reads
-      // received 4.5 MB of it, so a torrent-wide figure would have called a
-      // starved session healthy.
-      inputBytes: this.#host.encodeRuns.inputBytesOf(session),
-      startPositionSeconds: progress.startPositionSeconds ?? 0,
-      totalSeconds: progress.totalSeconds,
-      percent: progress.percent,
-      remainingSeconds: progress.remainingSeconds,
-      warmupPercent,
-      warmupRemainingSeconds,
-      // Segment length, so the browser can show progress toward the FIRST
-      // segment (the only thing it waits for before playback starts) instead
-      // of a percentage of the whole-file transcode.
-      segmentDurationSec: this.#host.segmentDurationSec,
-      speed: progress.speed,
-      outputMbps,
-      // The height the viewer is WATCHING right now, which is what the menu
-      // has to say next to "Auto". When the video is re-encoded that is the
-      // rung the proxy has settled on — it steps down when the host cannot keep
-      // up or the link cannot carry the stream. When the video is COPIED it is
-      // the source's own height, and reporting zero there was simply wrong:
-      // most sessions copy the video, so the menu read a bare "Auto" almost
-      // always, which is exactly the question it was supposed to answer.
-      currentHeight: session.spec.transcodesVideo
-        ? (session.output.encodeHeight ?? session.file.height ?? 0)
-        : (session.file.height ?? 0),
-      // The rungs still worth offering, as they stand NOW. The list the browser
-      // was given when the file opened came from the startup benchmarks; this
-      // one is corrected by what the encoder has since been seen to do with
-      // this very source, so a rung that turns out to be beyond the host
-      // disappears from the menu instead of being discovered by switching to it.
-      offeredHeights: this.#host.qualityOffer.offeredHeights(session),
-      // The variant this proxy would rather serve, or 0 when it is content.
-      //
-      // A REQUEST, not an instruction — this side cannot move a player between
-      // variants and must not pretend to. The browser honours it only in
-      // automatic mode: a height the viewer picked by hand is theirs, and the
-      // rule that automatic quality changes belong to automatic mode alone is
-      // enforced where the viewer's choice actually lives.
-      //
-      // This is what replaced rewriting the picture's size underneath a running
-      // session. Every height is published in the master with its own init, so
-      // asking the player to move is the only form of the act that a decoder
-      // can follow.
-      requestedHeight: this.#host.quality.standingAskFor(named, consumerId),
-      // Whether that request is URGENT: this viewer's buffer would run dry
-      // before anything else could arrive, so their page switches as soon as
-      // the rung is ready instead of waiting for a cushion (roadmap item 98).
-      requestedUrgent: this.#host.quality.standingAskIsUrgent(named, consumerId),
-      // On what the output this viewer is given was judged against THEIR link:
-      // `fits` (confirmed), `estimated to fit` (admitted by an average, not
-      // confirmed), or `no measurement`. The page says which, so an estimate
-      // is never presented as a guarantee.
-      servingVerdict: this.#host.servingVerdictOf(consumerId),
-      // What this host takes to create a session and to make a first segment.
-      // Also on the playback plan, but the browser reads that once per file:
-      // measured 2026-08-06 across four seeks, a proxy that had just restarted
-      // answered null for both, and every later seek then computed its estimate
-      // with one term of four — the figure hit zero after 3.5 s of an 11.8 s
-      // wait and read "starting now" for the remaining 8.4 s. This response is
-      // polled about every 1.5 s, so carrying them here keeps them current.
-      expectedSessionCreateMs: this.#host.expectedSessionCreateMs(),
-      expectedFirstSegmentMs: this.#host.expectedFirstSegmentMs(),
-      updatedAt: progress.updatedAt,
-      error: this.#host.encodeRuns.failureOf(session)
-    };
-  }
-}
+/**
+ * What a viewer asks of this proxy, one request at a time.
+ *
+ * Open an output of a file — the one that already exists when its parameters
+ * agree — and be placed on it; state a new position; ask how far the material
+ * in front of them has got. Each request states a fact about the viewer and
+ * returns; where encoders go follows from those facts and is the plan's.
+ */
+
+import { logger } from "../../utils/logger.js";
+import { isOutputName } from "../encode/output/index.js";
+import { predictPlaybackReadiness, RateTrend } from "./playback-readiness.js";
+
+function isWarmupTimeoutError(error) {
+  if (!(error instanceof Error)) {
+    return false;
+  }
+  return error.message === "HLS playlist is still warming up.";
+}
+
+/**
+ * The segment number an assignment is recorded under for this file, −1 for the
+ * init, or null for anything that is neither (a playlist).
+ *
+ * @param {object} output
+ * @param {string} fileName
+ * @returns {number | null}
+ */
+function segmentAddressOf(output, fileName) {
+  const format = output.segmentFormat;
+  if (!format) {
+    return null;
+  }
+  if (format.initFileName !== null && fileName === format.initFileName) {
+    return -1;
+  }
+  if (format.isSegmentFileName(fileName)) {
+    return format.segmentIndexFromName(fileName);
+  }
+  return null;
+}
+
+export class ViewerRequests {
+  /** What this reads and asks of the rest of the proxy, and nothing else. @type {object} */
+  #host;
+  /** @type {WeakMap<object, Map<string, object>>} */
+  #rateStateBySession = new WeakMap();
+
+  /**
+   * @param {object} host - `opening`, `activeOutputFor`, `viewerSecondsOn`, `minimumBufferSecondsFor`, `getSourceStats`, `disposeSession`, `expectedFirstSegmentMs`, `expectedSessionCreateMs`, `planEncodersSoon`, `waitUntilReady`, `encodeRuns`, `encodeSpeedReadingOf`, `outputTimes`, `outputs`, `lookaheadSeconds`, `quality`, `qualityOffer`, `renditions`, `segmentDurationSec`, `segmentStore`, `sourceFiles`, `startupWaitMs`, `viewers`, `acceptsGeneration`, `generationOfRequest`, `noteGivenOutput`, `holdForResponse`, `noteServingVerdict`, `servingVerdictOf`
+   */
+  constructor(host) {
+    this.#host = host;
+  }
+
+  /**
+   * Open an output of a file for this viewer and put them on it.
+   *
+   * Which output answers is the encoding component's decision
+   * (`encode/OutputOpening.js`), made without knowing who asked. What is this
+   * operation's is the requester: they are registered on the output with the
+   * sound they chose and placed where their request says they are, and the
+   * answer waits for the output to be ready to play.
+   *
+   * Throws with `error.code === "TRANSCODE_DISABLED"` when transcoding is
+   * disabled on this proxy instance.
+   *
+   * @param {object} request - What `OutputOpening.open` takes, plus the viewer.
+   * @param {string} [request.consumerId] - Who asks. Empty for an output a
+   *   picture opens on its own behalf (a quality step, a soundtrack), which
+   *   places nobody on it.
+   * @returns {Promise<import("../encode/output/EncodedOutput.js").EncodedOutput>}
+   */
+  async createOrGetSession({ consumerId = "", ...request }) {
+    // What this viewer's link last measured, where the request does not say:
+    // a viewer who comes back to open another film has a link this proxy has
+    // already measured, and the output they are given is judged by it.
+    const known = consumerId ? this.#host.viewers.get(consumerId) : null;
+    const startPositionSeconds = request.startPositionSeconds ?? 0;
+    // Whether they were already on the output this request lands on, read at
+    // the moment they are put on it.
+    let alreadyOn = false;
+    // THE PLACE ON THIS MACHINE, AND THE VIEWER IN IT, in one synchronous call
+    // the opening makes the moment it has the output (roadmap item 97, step 14).
+    // Asked and placed together, so the next viewer to open anything is asked
+    // with this one counted: two openings cannot both be told there is room
+    // for one. Placed now for the second reason it always was — a viewer who
+    // has not been placed states no want, and an output all of whose viewers
+    // state nothing has every encoder on it stopped.
+    const claim = consumerId
+      ? (candidate) => {
+          alreadyOn = this.#host.viewers.forOutput(candidate).has(consumerId);
+          const answer = alreadyOn
+            ? { admitted: true, reason: "already watching it", speedX: null }
+            : this.#host.admitsWatching(candidate);
+          if (answer.admitted) {
+            this.placeViewer(candidate, this.#host.viewers.of(candidate, consumerId), startPositionSeconds);
+          }
+          return answer;
+        }
+      : null;
+    const { output, existed, audio, verdict } = await this.#host.opening.open({
+      ...request,
+      viewerLinkMbps: request.viewerLinkMbps ?? known?.linkReading()?.linkMbps ?? null,
+      claim
+    });
+    if (consumerId) {
+      const joined = existed && !alreadyOn;
+      // What THIS viewer wants of the sound, which an output they are joining
+      // knows nothing about: they may have chosen another language, and their
+      // browser may need a track re-encoded that the first viewer's could
+      // decode as it stands.
+      const viewer = this.#host.viewers.of(output, consumerId);
+      viewer.audio = audio;
+      // The picture as their page sees it, which bounds the height made for
+      // them from here on (roadmap item 98).
+      viewer.noteVisiblePicture(request.visiblePicture);
+      // On what their output was judged. Recorded where the output is opened
+      // for them; a step opened on their behalf is recorded by `Renditions`.
+      this.#host.noteServingVerdict(consumerId, verdict ? { ...verdict, outputKey: output.outputKey } : null);
+      // WHERE they are was stated when they were put on it, by `claim` above:
+      // a viewer joining an output already playing at 40:00 may be opening the
+      // film from a link that carries 05:00.
+      // Reuse said nothing at all once, so an output serving two viewers looked
+      // exactly like one serving a single viewer — and the whole question its
+      // key exists to answer is which of the two happened.
+      if (joined) {
+        logger.info(
+          `transcode ${output.id} joined by ${consumerId} ` +
+          `(${this.#host.viewers.forOutput(output).size} viewer(s)) key=${output.outputKey}`
+        );
+      }
+    }
+    this.#host.outputs.touch(output);
+    if (!existed) {
+      // Where the first encoder goes is the plan's, from where the viewers stand.
+      this.#host.planEncodersSoon();
+    }
+    try {
+      await this.#host.waitUntilReady(output);
+    } catch (error) {
+      if (!existed && this.#host.encodeRuns.hasFailed(output)) {
+        await this.#host.disposeSession(output.id);
+        throw error;
+      }
+      if (existed && !isWarmupTimeoutError(error)) {
+        throw error;
+      }
+      // Not a failure of the request: the synthetic playlist is already there
+      // and segments appear as the encoder produces them.
+    }
+    return output;
+  }
+
+  /**
+   * Put a viewer where their own request says they are.
+   *
+   * A viewer arrives by asking for a POSITION — zero, or the time an address
+   * bar carried — so "we do not know where they are" is not a state a viewer
+   * can be in. Before 2026-09-05 it was: position was written only by a segment
+   * request, so a viewer counted as placeless until they had asked for a
+   * segment, and an output whose viewers were all placeless had every encoder
+   * on it stopped for having nobody. The soundtrack that failed that day could
+   * not have asked: the segment it would have asked for needed an `init.mp4`
+   * that the stopped encoder was going to make.
+   *
+   * Only ever places a viewer who has none. A viewer already placed is being
+   * kept current by their own requests and seeks, and a fresh create request
+   * carries a default of zero that must not drag them back to the beginning.
+   *
+   * @param {HlsSession} session
+   * @param {import("../viewer/Viewer.js").Viewer} viewer
+   * @param {number} positionSeconds
+   * @returns {void}
+   */
+  /**
+   * Put a viewer named by id onto this output, where they have no position yet.
+   *
+   * The same operation as {@link placeViewer}, addressed the way another
+   * component may address it: by name. Encoding asks for this when a request
+   * for any file of an output is the first sign of the person behind it, and it
+   * must not be handed the viewer to do it — a viewer's record has one writer,
+   * and that is the viewer layer.
+   *
+   * @param {object} output
+   * @param {string} consumerId
+   * @param {number} positionSeconds
+   * @returns {void}
+   */
+  placeViewerOn(output, consumerId, positionSeconds) {
+    if (!output || !consumerId) {
+      return;
+    }
+    this.placeViewer(output, this.#host.viewers.of(output, consumerId), positionSeconds);
+  }
+
+  placeViewer(session, viewer, positionSeconds) {
+    if (viewer.position !== null) {
+      return;
+    }
+    const seconds = Number.isFinite(positionSeconds) && positionSeconds > 0 ? positionSeconds : 0;
+    let segment = 0;
+    try {
+      const index = this.#host.outputTimes.segmentIndexForTime(session, seconds);
+      if (Number.isInteger(index) && index >= 0) {
+        segment = index;
+      }
+    } catch {
+      // A session whose cut table is not built yet places its viewer at the
+      // beginning, which is where the run starts anyway.
+    }
+    viewer.position = { segment, seconds, at: Date.now(), seeked: null };
+    this.#host.planEncodersSoon();
+  }
+
+  /**
+   * The viewer seeked. Called from POST /api/transcode-sessions/:id/seek with
+   * the position the browser read off its own player once the scrub ended.
+   *
+   * This is the ONLY thing that repositions the encoder. It replaces inferring
+   * the target from segment requests, which cannot work: a single seek leaves
+   * ~25 concurrent requests outstanding across a wide span (measured), so no
+   * rule over them can recover which one the viewer meant.
+   *
+   * The existing settle/cooldown/first-segment guards still apply — they
+   * protect against restarting too eagerly, which is orthogonal to knowing
+   * WHERE to restart.
+   *
+   * @param {string} sessionId
+   * @param {number} positionSeconds - Absolute position on the source timeline.
+   * @param {string} [consumerId] - Who moved.
+   * @param {number} [generation] - Which viewing their requests now belong to,
+   *   as the page states it. Not a number this side counts: see below.
+   * @returns {boolean} False when the session is unknown or disposed.
+   */
+  requestSeek(sessionId, positionSeconds, consumerId = "", generation = Number.NaN) {
+    const named = this.#host.outputs.get(sessionId);
+    if (!named || !consumerId) {
+      return false;
+    }
+    // A SEEK DOES ONE THING: it puts the viewer where they now are.
+    //
+    // It used to do eleven, and wrote that position into five places: two
+    // fields on this session, two more on the soundtrack's, and the viewer. It
+    // also asked whether the jump would drag another viewer back, started an
+    // encoder itself, cancelled outstanding requests, backed off a segment to
+    // the preceding keyframe, and set a timer to restart ffmpeg. So it was a
+    // third authority over the encoders beside the plan and the start path, and
+    // not one of its branches ever asked what had already been made — a viewer
+    // jumping into a stretch that was finished and on disk got a fresh encoder
+    // for it.
+    //
+    // What follows from the move happens by itself: the priority map is built
+    // from where the viewers are, and both orchestrators read the map.
+    // A seek that names nobody moves nobody: the route refuses it before it
+    // gets here.
+    const viewer = this.#host.viewers.of(named, consumerId);
+    viewer.moveTo(positionSeconds);
+    // WHICH VIEWING THEIR REQUESTS NOW BELONG TO, as the page states it.
+    //
+    // Not counted here: the page raises its own number before it sends this,
+    // and stamps every request with it, so counting a second one on this side
+    // would make two owners of one fact — and they would disagree exactly when
+    // a request and a seek cross, which is the case the number exists for. A
+    // page that states nothing leaves the generation where it is, which is the
+    // behaviour every transport without a loader of ours has.
+    //
+    // The generation being left is not dropped: a request already in flight is
+    // still answered by the output it was made against — see the two deadlines
+    // in `Assignments.js`.
+    viewer.assignments.statedGeneration(generation);
+    this.#host.outputs.touch(named);
+    // EVERY REQUEST THIS VIEWER HAS OUTSTANDING WAS MADE FOR WHERE THEY WERE,
+    // and hls.js keeps one fragment load outstanding per track: a request held
+    // for a segment behind them blocks the one they need now. Measured
+    // 2026-08-04: 57 s of a 58 s backward seek was that wait, and the segment
+    // they wanted was served in 15 ms once it was asked for.
+    //
+    // WAKING IS NOT CANCELLING. It moves the epoch, and each held request then
+    // asks whether it is still wanted — for the viewer who made it, at that
+    // viewer's own position. So a request of somebody else's on an output
+    // shared with them is kept, and so is this viewer's own request for the
+    // segment they have just landed on, which races this notification.
+    //
+    // EVERY OUTPUT THEY WATCH, not the one the browser addressed: their picture,
+    // the quality step on their screen and their soundtrack are three outputs
+    // with three sets of held requests, and a seek leaves all three behind.
+    for (const outputId of viewer.outputs) {
+      const watched = this.#host.outputs.get(outputId);
+      if (watched) {
+        this.#host.invalidateWaits(watched);
+      }
+    }
+    this.#host.planEncodersSoon();
+    return true;
+  }
+
+  /**
+   * Whether this viewer still takes a request made in the generation it states.
+   *
+   * ASKED BEFORE ANYTHING THE REQUEST WOULD CAUSE. Resolving a step or a
+   * soundtrack can create an output and registers the viewer on it; a request
+   * made for a viewing the viewer has already left would do both and only then
+   * be refused. So every route asks this first.
+   *
+   * @param {string} consumerId
+   * @param {number} statedGeneration - What the request carried; NaN when
+   *   nothing, which is always taken.
+   * @returns {boolean}
+   */
+  acceptsRequest(consumerId, statedGeneration) {
+    return this.#host.acceptsGeneration(consumerId, statedGeneration);
+  }
+
+  /**
+   * Record that the output a request ADDRESSED answered it.
+   *
+   * The picture's own route names its output outright, so there is nothing to
+   * choose; what is recorded is the answer, under the height that output is
+   * named after (`variantHeightOf`). That is the same height under which the
+   * step route answers with the picture itself, so the two addresses of one
+   * height lead to one output and one record. An init is recorded as segment
+   * −1. A soundtrack is not a height and records nothing: its repeat is not
+   * kept, while the check of the generation and the hold of a response still
+   * apply to it.
+   *
+   * @param {string} sessionId
+   * @param {string} consumerId
+   * @param {number} statedGeneration
+   * @param {string} fileName
+   * @returns {void}
+   */
+  noteAnsweredDirectly(sessionId, consumerId, statedGeneration, fileName) {
+    const output = isOutputName(sessionId) ? this.#host.outputs.get(sessionId) : null;
+    if (!output || !consumerId || !output.spec?.video) {
+      return;
+    }
+    const segmentIndex = segmentAddressOf(output, fileName);
+    if (segmentIndex === null) {
+      return;
+    }
+    const height = this.#host.outputs.variantHeightOf(output);
+    const generation = this.#host.generationOfRequest(consumerId, statedGeneration);
+    this.#host.noteGivenOutput(consumerId, generation, height, segmentIndex, output.outputKey ?? "");
+  }
+
+  /**
+   * A response from this output to this viewer has begun: the output is held
+   * until the function returned is called, whatever happens to the generation
+   * meanwhile. Safe to call more than once.
+   *
+   * @param {string} sessionId
+   * @param {string} consumerId
+   * @returns {() => void}
+   */
+  holdResponse(sessionId, consumerId) {
+    const output = isOutputName(sessionId) ? this.#host.outputs.get(sessionId) : null;
+    return this.#host.holdForResponse(consumerId, output?.outputKey ?? "");
+  }
+
+  /**
+   * The same hold, named by an output's KEY: a piece served from what a gone
+   * output left in the store has no live output to name it by.
+   *
+   * @param {string} outputKey
+   * @param {string} consumerId
+   * @returns {() => void}
+   */
+  holdResponseForKey(outputKey, consumerId) {
+    return this.#host.holdForResponse(consumerId, outputKey ?? "");
+  }
+
+  /**
+   * Where the viewer of this session is, in seconds.
+   *
+   * Exists so a refusal can name it. A log line that says only "superseded"
+   * cannot be read afterwards: it does not say what was refused or against what
+   * position, which is exactly what the 2026-08-18 investigation lacked.
+   *
+   * Named per viewer, because that is what the refusal is about: a request is
+   * refused for being behind where THAT viewer is, and a line naming the
+   * furthest viewer of a shared session would explain a refusal by somebody
+   * else's position.
+   *
+   * @param {string} sessionId
+   * @param {string} [consumerId]
+   * @returns {number} Zero when the session is gone or nothing has been reported.
+   */
+  viewerPositionOf(sessionId, consumerId = "") {
+    const session = isOutputName(sessionId) ? this.#host.outputs.get(sessionId) : null;
+    if (!session) {
+      return 0;
+    }
+    return this.#host.viewerSecondsOn(session, consumerId);
+  }
+
+  async getSessionProgress(sessionId, consumerId = "") {
+    if (!isOutputName(sessionId)) {
+      return null;
+    }
+    const named = this.#host.outputs.get(sessionId);
+    if (!named) {
+      return null;
+    }
+    this.#host.outputs.touch(named);
+    // Progress is asked about the stream on screen, which after a quality
+    // change is another session. Touching the named one as well is what keeps
+    // the family alive: only the ACTIVE variant gets segment requests, so
+    // without this the base session would idle out from under its own variants.
+    const session = this.#host.activeOutputFor({ base: named, consumerId, outputs: this.#host.outputs });
+    this.#host.outputs.touch(session);
+    const warmupTotalSeconds = this.#host.startupWaitMs / 1000;
+    const warmupElapsedSeconds = Math.max(
+      0,
+      (Date.now() - (this.#host.outputs.startedAt(session) ?? Date.now())) / 1000
+    );
+    // One question, one answer. Run state comes only from the encoding layer.
+    const isWarmupPhase = this.#host.encodeRuns.wireStateOf(session) === "starting";
+    const warmupPercent = isWarmupPhase
+      ? Math.max(0, Math.min(100, (warmupElapsedSeconds / warmupTotalSeconds) * 100))
+      : null;
+    const warmupRemainingSeconds = isWarmupPhase
+      ? Math.max(0, warmupTotalSeconds - warmupElapsedSeconds)
+      : null;
+    // Observed output bitrate (Mbit/s) from recently completed segments. The
+    // proxy uses it for the viewer-link budget and the readiness forecast.
+    const outputMbps = await this.#host.quality.observedStreamMbps(session);
+    const progress = this.#host.encodeRuns.progressOf(session);
+    const playbackReadiness = await this.#playbackReadinessFor(session, consumerId, progress, outputMbps);
+    return {
+      // The id the caller asked about, not the variant it was answered from —
+      // the browser tracks its sessions by the id it was given.
+      sessionId: named.id,
+      state: this.#host.encodeRuns.wireStateOf(session),
+      // The smallest buffer at which no interruption reaches the viewer, from
+      // THIS file's own recent interruptions: one whole segment — the one being
+      // played — plus the worst wait that can arrive before the buffer refills.
+      // On the field torrent that is 7-9 s where the browser waits for a
+      // hand-chosen 25, which is sixteen seconds of staring at a spinner that
+      // nothing had shown to be necessary. Null until the reader has seen two
+      // interruptions; the browser keeps its own figure until then.
+      minimumBufferSeconds: this.#host.minimumBufferSecondsFor(session),
+      processedSeconds: progress.processedSeconds,
+      // Bytes this session's own reads have received from the swarm.
+      //
+      // The second proof that a session is alive, and the only one available
+      // before its first frame exists: `processedSeconds` cannot move until the
+      // decoder has a frame, so on a cold start it stands at the start position
+      // for as long as the first piece takes to arrive. Field 2026-09-03 — one
+      // piece took 46.3 s while the swarm delivered 55.9 MB across the torrent,
+      // `processedSeconds` frozen at 171.3 throughout, and the browser declared
+      // the proxy dead 0.4 s before the piece landed.
+      //
+      // Counted per SESSION and not per torrent, deliberately: in that same
+      // episode the torrent received 55.9 MB while the picture's own reads
+      // received 4.5 MB of it, so a torrent-wide figure would have called a
+      // starved session healthy.
+      inputBytes: this.#host.encodeRuns.inputBytesOf(session),
+      startPositionSeconds: progress.startPositionSeconds ?? 0,
+      totalSeconds: progress.totalSeconds,
+      percent: progress.percent,
+      remainingSeconds: progress.remainingSeconds,
+      warmupPercent,
+      warmupRemainingSeconds,
+      // Segment length, so the browser can show progress toward the FIRST
+      // segment (the only thing it waits for before playback starts) instead
+      // of a percentage of the whole-file transcode.
+      segmentDurationSec: this.#host.segmentDurationSec,
+      speed: progress.speed,
+      outputMbps,
+      playbackReadiness,
+      // The height the viewer is WATCHING right now, which is what the menu
+      // has to say next to "Auto". When the video is re-encoded that is the
+      // rung the proxy has settled on — it steps down when the host cannot keep
+      // up or the link cannot carry the stream. When the video is COPIED it is
+      // the source's own height, and reporting zero there was simply wrong:
+      // most sessions copy the video, so the menu read a bare "Auto" almost
+      // always, which is exactly the question it was supposed to answer.
+      currentHeight: session.spec.transcodesVideo
+        ? (session.output.encodeHeight ?? session.file.height ?? 0)
+        : (session.file.height ?? 0),
+      // The rungs still worth offering, as they stand NOW. The list the browser
+      // was given when the file opened came from the startup benchmarks; this
+      // one is corrected by what the encoder has since been seen to do with
+      // this very source, so a rung that turns out to be beyond the host
+      // disappears from the menu instead of being discovered by switching to it.
+      offeredHeights: this.#host.qualityOffer.offeredHeights(session),
+      // The variant this proxy would rather serve, or 0 when it is content.
+      //
+      // A REQUEST, not an instruction — this side cannot move a player between
+      // variants and must not pretend to. The browser honours it only in
+      // automatic mode: a height the viewer picked by hand is theirs, and the
+      // rule that automatic quality changes belong to automatic mode alone is
+      // enforced where the viewer's choice actually lives.
+      //
+      // This is what replaced rewriting the picture's size underneath a running
+      // session. Every height is published in the master with its own init, so
+      // asking the player to move is the only form of the act that a decoder
+      // can follow.
+      requestedHeight: this.#host.quality.standingAskFor(named, consumerId),
+      // Whether that request is URGENT: this viewer's buffer would run dry
+      // before anything else could arrive, so their page switches as soon as
+      // the rung is ready instead of waiting for a cushion (roadmap item 98).
+      requestedUrgent: this.#host.quality.standingAskIsUrgent(named, consumerId),
+      // On what the output this viewer is given was judged against THEIR link:
+      // `fits` (confirmed), `estimated to fit` (admitted by an average, not
+      // confirmed), or `no measurement`. The page says which, so an estimate
+      // is never presented as a guarantee.
+      servingVerdict: this.#host.servingVerdictOf(consumerId),
+      // What this host takes to create a session and to make a first segment.
+      // Also on the playback plan, but the browser reads that once per file:
+      // measured 2026-08-06 across four seeks, a proxy that had just restarted
+      // answered null for both, and every later seek then computed its estimate
+      // with one term of four — the figure hit zero after 3.5 s of an 11.8 s
+      // wait and read "starting now" for the remaining 8.4 s. This response is
+      // polled about every 1.5 s, so carrying them here keeps them current.
+      expectedSessionCreateMs: this.#host.expectedSessionCreateMs(),
+      expectedFirstSegmentMs: this.#host.expectedFirstSegmentMs(),
+      updatedAt: progress.updatedAt,
+      error: this.#host.encodeRuns.failureOf(session)
+    };
+  }
+
+  async #playbackReadinessFor(session, consumerId, progress, outputMbps) {
+    const now = Date.now();
+    const viewer = consumerId ? this.#host.viewers.get(consumerId) : null;
+    const viewerReading = viewer?.linkReading() ?? null;
+    const measurement = this.#rateStateFor(session, consumerId);
+    const linkAt = Number(viewerReading?.linkSampleMeasuredAt);
+    if (Number.isFinite(viewerReading?.linkSampleMbps) && viewerReading.linkSampleMbps > 0 && Number.isFinite(linkAt)) {
+      measurement.link.add(linkAt, viewerReading.linkSampleMbps * 1_000_000);
+    }
+
+    const outputs = [session];
+    const requiresSeparateAudio = this.#host.renditions.servesAudioSeparately(session);
+    const audioOutput = requiresSeparateAudio
+      ? this.#host.renditions.playbackAudioOutputFor(session, consumerId)
+      : null;
+    if (audioOutput) {
+      outputs.push(audioOutput);
+    }
+
+    const sourceMeasurements = new Map();
+    const tracks = [];
+    for (const output of outputs) {
+      const trackProgress = output === session ? progress : this.#host.encodeRuns.progressOf(output);
+      const trackRates = this.#trackRateReadings(measurement, output, now);
+      const timeline = output.timeline;
+      const segments = Array.from({ length: timeline?.segmentCount ?? 0 }, (_, index) => ({
+        index,
+        startSeconds: timeline.publishedStartOf(index),
+        endSeconds: timeline.publishedStartOf(index + 1)
+      }));
+      const sourceIndexes = [output.spec.video?.fileIndex, output.spec.audio?.fileIndex]
+        .filter((fileIndex) => Number.isInteger(fileIndex) && fileIndex >= 0);
+      if (sourceIndexes.length === 0) {
+        sourceIndexes.push(output.file.fileIndex);
+      }
+      const sourceIds = [];
+      for (const fileIndex of new Set(sourceIndexes)) {
+        const sourceId = `${output.file.sourceKey}:${fileIndex}`;
+        sourceIds.push(sourceId);
+        if (sourceMeasurements.has(sourceId)) {
+          continue;
+        }
+        const sourceFile = fileIndex === output.file.fileIndex
+          ? output.file
+          : this.#host.sourceFiles.get(output.file.sourceKey, fileIndex);
+        let stats = null;
+        try {
+          stats = await this.#host.getSourceStats?.(output.file.sourceKey, fileIndex) ?? null;
+        } catch {
+          stats = null;
+        }
+        const sourceTrend = this.#downloadRateReadings(
+          measurement,
+          output.file.sourceKey,
+          stats?.downloadSpeed,
+          now
+        );
+        const fileLength = Number(stats?.fileLength);
+        const downloaded = Number(stats?.fileDownloaded);
+        const sourceDuration = Number(sourceFile?.durationSeconds);
+        sourceMeasurements.set(sourceId, {
+          id: sourceId,
+          serviceId: output.file.sourceKey,
+          complete: Number.isFinite(fileLength) && fileLength > 0 && downloaded >= fileLength,
+          bytesPerMediaSecond: Number.isFinite(fileLength) && fileLength > 0 &&
+            Number.isFinite(sourceDuration) && sourceDuration > 0
+            ? fileLength / sourceDuration
+            : 0,
+          readings: sourceTrend
+        });
+      }
+
+      const inventory = this.#host.segmentStore.sizesOf(output.outputKey);
+      const observedMbps = output === session
+        ? outputMbps
+        : await this.#host.quality.observedStreamMbps(output);
+      tracks.push({
+        id: output.outputKey,
+        sourceIds,
+        processedSeconds: trackProgress?.processedSeconds,
+        bitsPerMediaSecond: Number.isFinite(observedMbps) && observedMbps > 0 ? observedMbps * 1_000_000 : 0,
+        readings: trackRates,
+        segments,
+        readySegmentIndices: [...inventory.keys()],
+        segmentSizesBytes: inventory
+      });
+    }
+
+    const activeTrackIds = new Set(tracks.map(({ id }) => id));
+    for (const id of measurement.tracks.keys()) {
+      if (!activeTrackIds.has(id)) {
+        measurement.tracks.delete(id);
+      }
+    }
+    const activeDownloadServices = new Set([...sourceMeasurements.values()].map(({ serviceId }) => serviceId));
+    for (const id of measurement.downloads.keys()) {
+      if (!activeDownloadServices.has(id)) {
+        measurement.downloads.delete(id);
+      }
+    }
+    const bufferedAheadSeconds = Number(viewerReading?.bufferedAheadSec);
+    const positionSeconds = Number.isFinite(viewerReading?.positionSeconds)
+      ? viewerReading.positionSeconds
+      : this.#host.viewerSecondsOn(session, consumerId, now);
+    // Both separately delivered tracks can run short independently. Use the
+    // largest measured supply interruption across the tracks the viewer needs;
+    // when a track has no interruption history yet, its actual next segment
+    // duration is the observed floor for that track.
+    const reserveSeconds = Math.max(...outputs.map((output, index) => {
+      const measured = Number(this.#host.minimumBufferSecondsFor(output));
+      if (Number.isFinite(measured) && measured > 0) {
+        return measured;
+      }
+      const trackSegments = tracks[index]?.segments ?? [];
+      const nextSegment = trackSegments.find((segment) =>
+        segment.startSeconds <= positionSeconds && segment.endSeconds > positionSeconds
+      ) ?? trackSegments.find((segment) => segment.startSeconds >= positionSeconds);
+      const segmentSeconds = nextSegment
+        ? nextSegment.endSeconds - nextSegment.startSeconds
+        : Number(this.#host.segmentDurationSec);
+      return Number.isFinite(segmentSeconds) && segmentSeconds > 0 ? segmentSeconds : 0;
+    }));
+
+    return predictPlaybackReadiness({
+      now,
+      positionSeconds,
+      durationSeconds: session.file.durationSeconds,
+      bufferedAheadSeconds,
+      bufferLimitSeconds: viewerReading?.bufferLimitSeconds,
+      reserveSeconds,
+      lookaheadSeconds: this.#host.lookaheadSeconds,
+      requiredAudio: requiresSeparateAudio,
+      sources: [...sourceMeasurements.values()],
+      tracks,
+      linkReadings: measurement.link.snapshot()
+    });
+  }
+
+  #rateStateFor(session, consumerId) {
+    let byViewer = this.#rateStateBySession.get(session);
+    if (!byViewer) {
+      byViewer = new Map();
+      this.#rateStateBySession.set(session, byViewer);
+    }
+    const id = typeof consumerId === "string" ? consumerId : "";
+    let state = byViewer.get(id);
+    if (!state) {
+      state = { link: new RateTrend(), downloads: new Map(), tracks: new Map() };
+      byViewer.set(id, state);
+    }
+    return state;
+  }
+
+  #downloadRateReadings(state, serviceId, downloadSpeed, now) {
+    let trend = state.downloads.get(serviceId);
+    if (!trend) {
+      trend = new RateTrend();
+      state.downloads.set(serviceId, trend);
+    }
+    const speed = Number(downloadSpeed);
+    if (Number.isFinite(speed) && speed >= 0) {
+      trend.add(now, speed);
+    }
+    return trend.snapshot();
+  }
+
+  #trackRateReadings(state, output, now) {
+    const id = output.outputKey;
+    let trend = state.tracks.get(id);
+    if (!trend) {
+      trend = new RateTrend();
+      state.tracks.set(id, trend);
+    }
+    const reading = this.#host.encodeSpeedReadingOf?.(output, now) ?? null;
+    if (Number.isFinite(reading?.speed) && reading.speed > 0 && Number.isFinite(reading.at)) {
+      trend.add(reading.at, reading.speed);
+    }
+    return trend.snapshot();
+  }
+}