@torrent-tv/proxy 2.88.3 → 2.88.4

package/services/server/wire-outputs.js

@@ -1,992 +1,995 @@
-/**
- * @file Wiring of the outputs, the viewers and the encoding.
- *
- * Builds each component and hands it the narrow host it reads, and nothing
- * else. It holds no behaviour: every member that used to be a method of the
- * session manager is a method of the component that owns it now. What it
- * returns is the set of components and the settings they share, for
- * `server.js` to give each route the ones that route asks.
- */
-
-import { logger } from "../../utils/logger.js";
-import { KeyframeTables } from "../media/KeyframeTables.js";
-import { waits } from "./WaitLedger.js";
-import { readMachineState, readProcessCpuSeconds, readProxyCpuSeconds, readSystemCpu, shareOfMachine } from "../encode/host-load.js";
-import { minimumBufferFrom } from "../torrent/supply-margin.js";
-import { PriorityOrchestrator } from "../viewer/PriorityOrchestrator.js";
-import { softwareDescriptor } from "../encode/hwaccel.js";
-import { resolveSegmentFormat } from "../encode/segment-formats/index.js";
-import { Timelines } from "../encode/output/Timeline.js";
-import { SourceFiles } from "../media/SourceFile.js";
-import { SegmentStore } from "../storage/segment-store/SegmentStore.js";
-import { EncodeCost } from "../encode/quality/EncodeCost.js";
-import { QualityOffer } from "../encode/quality/QualityOffer.js";
-import { activeOutputFor } from "../viewer/active-output.js";
-import {
-  acceptsGeneration,
-  askQualityOf,
-  audioBeingWarmedOf,
-  audioChoiceOf,
-  chooseAudioTrack,
-  chooseOutput,
-  chosenOutputOf,
-  consumersOn,
-  dropAskOf,
-  generationOfRequest,
-  givenOutputOf,
-  heightsChosenAs,
-  highestGivenSegmentOf,
-  holdForResponse,
-  linkMbpsOf,
-  linkReportOf,
-  bufferOf,
-  bufferedSecondsOf,
-  visiblePictureOf,
-  linkReportsOn,
-  noteAudioBeingWarmed,
-  noteGivenOutput,
-  noteSameHeightSwitch,
-  noteServingVerdict,
-  noteStepBeingWarmed,
-  noteStepOnScreen,
-  outputsBeingPrepared,
-  placeOn,
-  presentOn,
-  qualityModeOf,
-  sameHeightSwitchOf,
-  servingVerdictOf,
-  standingAskOf,
-  stepBeingWarmedOf,
-  stepBeingWarmedSinceOf,
-  stepOnScreenOf,
-  switchingOnto,
-  watches
-} from "../viewer/choices.js";
-import { viewerSecondsOn, viewerSegmentsOn } from "../viewer/positions.js";
-import { Viewers } from "../viewer/Viewers.js";
-import { OutputCatalog } from "../encode/output/OutputCatalog.js";
-import { ViewerRequests } from "./ViewerRequests.js";
-import { OutputLifecycle } from "./OutputLifecycle.js";
-import { SegmentServing } from "./SegmentServing.js";
-import { audioStartSecondsFor } from "../viewer/audio-start.js";
-import { audioRenditionName } from "../media/audio-inventory.js";
-import { Renditions } from "../encode/Renditions.js";
-import { CushionReport, LOOKAHEAD_PAUSE_SECONDS } from "../encode/CushionReport.js";
-import { EncodeRuns } from "../encode/EncodeRuns.js";
-import { OutputTimes } from "../encode/OutputTimes.js";
-import { HostLoad } from "../encode/quality/HostLoad.js";
-import { HostTimings } from "../encode/quality/HostTimings.js";
-import { BUDGET_CHECK_INTERVAL_MS, QualityController } from "../encode/quality/QualityController.js";
-import { EncodeOrchestrator } from "../encode/EncodeOrchestrator.js";
-import { EncodeAdmission } from "../encode/EncodeAdmission.js";
-import { EncoderSelection } from "../encode/EncoderSelection.js";
-import { OutputOpening } from "../encode/OutputOpening.js";
-import { segmentFormatOfKey } from "../encode/output-key-format.js";
-import { OutputSpec } from "../encode/output/OutputSpec.js";
-import { compareInits } from "../encode/segment-formats/init-compat.js";
-import { nominalKbpsFor } from "../encode/args.js";
-import { probeInputMediaInfo } from "../media/media-info-probe.js";
-import { probeVideoKeyframeTimes } from "../media/keyframe-probe.js";
-import { wireMachineBudget } from "../storage/wire.js";
-import { Returns } from "../storage/returns.js";
-import { freeBytesFor } from "../storage/free.js";
-import path from "node:path";
-import { contentOf, LocalObservations } from "../encode/LocalObservations.js";
-import { configurationKeyOf } from "../encode/fingerprint.js";
-import { PROXY_ROOT } from "../encode/quality/HostTimings.js";
-import { qualityStateOf } from "../encode/quality/OutputQualityState.js";
-
-// Where a variant and an audio rendition live under a session — `v/<height>/…`
-// and `a/<track>/…` — is stated in `encode/output/playlists.js`, beside the lines that
-// write those addresses into a master playlist. The routes that parse them back
-// are in `server.js`.
-
-const CLEANUP_INTERVAL_MS = 30_000;
-const DEFAULT_SEGMENT_DURATION_SEC = 4;
-// Idle TTL: a session is disposed this long after the last segment/playlist
-// access. Long enough that a viewer who pauses, backgrounds the tab, or briefly
-// turns the phone off can resume WITHOUT a cold ffmpeg restart (the warm session
-// also backs the seamless auto-reconnect). ffmpeg stops producing at the
-// look-ahead cap when idle, so a lingering session costs retained segments on
-// disk, not sustained CPU. Active playback refreshes the timer on every segment
-// fetch, so it never expires mid-watch.
-// How long a session outlives the BROWSER, not the viewing. Since server
-// 0.8.103 a browser that holds a session re-asserts it every 30 s, so an open
-// tab never consumes this at all — not while paused, not across a three-hour
-// film. What is left is the case where the browser has genuinely gone: the tab
-// was killed without releasing, the phone slept, the network dropped. Keeping
-// the session means such a viewer comes back to a warm encoder instead of
-// waiting out a cold start; the cost while nobody is there is disk for the
-// produced segments, since the encoder is suspended and burns no CPU, and that
-// disk is already bounded by the pool's 10 GB cap with eviction. Thirty minutes
-// covers a meal, a phone call or a lift ride.
-const DEFAULT_SESSION_TTL_MS = 30 * 60 * 1000;
-const DEFAULT_STARTUP_WAIT_MS = 5_000;
-
-/**
- * Convert a bind-all host address to the loopback address so that
- * the HLS input URL is always reachable from the same machine.
- *
- * @param {string} host
- * @returns {string}
- */
-function toLoopbackHost(host) {
-  if (host === "0.0.0.0" || host === "::") {
-    return "127.0.0.1";
-  }
-  return host;
-}
-
-/**
- * Build the HTTP base URL (scheme + host + port) for the local proxy server.
- *
- * @param {string} host - Bind host (may be "0.0.0.0" or "::").
- * @param {number} port
- * @returns {string} e.g. "http://127.0.0.1:9090"
- */
-function buildHttpBaseUrl(host, port) {
-  const url = new URL("http://localhost");
-  url.hostname = toLoopbackHost(host);
-  url.port = String(port);
-  return url.origin;
-}
-
-/**
- * @typedef {Object} WireOutputsOptions
- * @property {boolean} enabled              - Whether HLS transcoding is enabled.
- * @property {string}  ffmpegBin            - Path to the ffmpeg executable.
- * @property {string}  localBindHost        - Host the proxy HTTP server is bound to.
- * @property {number}  localPort            - Port the proxy HTTP server is listening on.
- * @property {number}  [segmentDurationSec] - HLS segment length in seconds.
- * @property {number}  [sessionTtlMs]       - Session idle TTL in milliseconds.
- * @property {number}  [startupWaitMs]      - Max time to wait for the first playlist file.
- * @property {string}  [segmentFormatId]    - Output container: "fmp4" (default)
- *   or "mpegts". See `../segment-formats/index.js`.
- */
-
-/**
- * @typedef {Object} HlsSession
- * @property {string}  id            - The name of the output it produces.
- * @property {string}  fileName      - Display name of the file being transcoded.
- * @property {import("../media/container/KeyframeTable.js").KeyframeTable} keyframes -
- *   Where this file's keyframes are. One object per file, held by every session
- *   of it, so a table read late still reaches them. Used to snap a source seek
- *   onto a known-valid position (see #startEncodeRun).
- */
-
-/**
- * Build the components and connect them.
- *
- * @param {WireOutputsOptions} options
- * @returns {object} The components, and the settings they read.
- */
-export function wireOutputs({
-    enabled,
-    ffmpegBin,
-    localBindHost,
-    localPort,
-    segmentDurationSec = DEFAULT_SEGMENT_DURATION_SEC,
-    sessionTtlMs = DEFAULT_SESSION_TTL_MS,
-    startupWaitMs = DEFAULT_STARTUP_WAIT_MS,
-    // Where every file's keyframe table lives, and the only thing that reads
-    // one. A session asks for its file's table and is handed the object every
-    // other reader of that file holds, so a table that arrives late reaches the
-    // sessions created before it. Built here when nothing supplied one, which
-    // is a registry with no reader: it answers "not read" about every file, and
-    // every picture is then re-encoded — the correct behaviour for a proxy
-    // wired without a container reader, and what every unit test gets.
-    keyframeTables = new KeyframeTables(),
-    videoEncoder = null,
-    softwarePresetBenchmark = null,
-    // What every encoder this host may use was qualified and measured to do at
-    // startup, by kind (`encode/calibration.js`). Takes the place of the one
-    // array above, which a test may still hand in on its own.
-    calibration = null,
-    decodeCostModel = null,
-    getSourceStats = null,
-    setPriorityMap = null,
-    // What a second job costs on this host, measured at startup. Null when it
-    // could not be measured, and then nothing is corrected — the alternative
-    // is inventing a penalty, which is the same fault as inventing a fill rate.
-    contentionPenalties = null,
-    copySpeedX = null,
-    tonemapSupported = false,
-    getCachedMediaInfo = null,
-    getCachedAudioTracks = null,
-    getContainerMediaInfo = null,
-    fetchWholeFile = null,
-    segmentFormatId = undefined,
-    stateDir = "",
-    segmentStore = null,
-    getTorrentTotals,
-    startStopCost = null,
-    spillDisk = null,
-    wholeFiles = null,
-    diagnostics = null,
-    diagnosticsRoot = "",
-    memoryClaimant = null,
-    budgetPolicy = null}) {
-  const parts = {};
-  // Encode components receive only the storage operations they use. The
-  // concrete SegmentStore remains owned by this assembly and is never handed
-  // to the encode layer.
-  const segmentCoverage = {
-    provenNumbers: (address) => parts.segmentStore.provenNumbers(address),
-    announce: (address, index) => parts.segmentStore.announce(address, index),
-    filesHeld: (address) => parts.segmentStore.filesHeld(address),
-    clearUpAfter: (address, startedAt) => parts.segmentStore.clearUpAfter(address, startedAt)
-  };
-  const segmentFiles = {
-    pathFor: (address) => parts.segmentStore.pathFor(address),
-    initOf: (address) => parts.segmentStore.initOf(address),
-    directoryFor: (address) => parts.segmentStore.directoryFor(address),
-    publish: (address, makingName, format) => parts.segmentStore.publish(address, makingName, format)
-  };
-  const segmentOutputFiles = {
-    addresses: () => parts.segmentStore.addresses(),
-    isClosed: (address, index) => parts.segmentStore.isClosed(address, index),
-    lastReadAt: (address) => parts.segmentStore.lastReadAt(address),
-    directoryFor: (address) => parts.segmentStore.directoryFor(address),
-    useFormat: (address, format) => parts.segmentStore.useFormat(address, format)
-  };
-  const segmentPaths = {
-    pathOf: (address, index) => parts.segmentStore.pathOf(address, index)
-  };
-  /**
-   * Whether everything this output will ever serve is made: every segment of
-   * its timeline closed in the store. Such an output needs no encoder and
-   * holds no place, however many watch it.
-   *
-   * @param {object} session
-   * @returns {boolean}
-   */
-  const finishedOutput = (session) => {
-    const count = Number(session?.timeline?.segmentCount) || 0;
-    return count > 0 && parts.segmentStore.provenNumbers(session.outputKey).length >= count;
-  };
-  // The nominal limits a size may be produced at, highest first. ONE list, read
-  // by the opening of an output and by the move between limits of a height, so
-  // the two cannot disagree about which limits exist. Only the size's own until
-  // a set of lower limits is decided (roadmap item 97, step 14): a lower limit
-  // handed to a viewer must come with the page saying so. Asked with the whole
-  // frame, because the row is chosen by area (`limitRowFor`).
-  parts.limitsFor = (frame) => [nominalKbpsFor(frame)];
-  // What the torrent and the proxy itself spend on this host, and how fast each watched torrent moves.
-  parts.hostLoad = new HostLoad({
-    readMachineState,
-    readProcessCpuSeconds,
-    readProxyCpuSeconds,
-    readSystemCpu,
-    shareOfMachine,
-    liveRunsOf: (...args) => parts.encodeRuns.liveRunsOf(...args),
-    runStateOf: (...args) => parts.encodeRuns.runStateOf(...args),
-    get getSourceStats() { return parts.getSourceStats; },
-    get getTorrentTotals() { return parts.getTorrentTotals; },
-    get outputs() { return parts.outputs; },
-  });
-  // Where each segment of an output begins, and how that table is corrected from what the encoder produced.
-  parts.outputTimes = new OutputTimes({
-    logger,
-    runsOf: (...args) => parts.encodeRuns.runsOf(...args),
-    stopEncodeRun: (...args) => parts.encodeRuns.stopEncodeRun(...args),
-    planEncodersSoon: (...args) => parts.encodeRuns.planEncodersSoon(...args),
-    get outputs() { return parts.outputs; },
-    get segmentDurationSec() { return parts.segmentDurationSec; },
-  });
-  // The encoders of this proxy: built where the plan places them, followed while they run, accounted when they end.
-  parts.encodeRuns = new EncodeRuns({
-    logger,
-    // What an admitted encode was seen to do, filed under its configuration.
-    observeEncodeEnded: (session) => {
-      if (!parts.localObservations || !session?.spec?.video?.encode) {
-        return;
-      }
-      const alone = Number(qualityStateOf(session).lastAloneSpeed);
-      parts.localObservations.noteEncode({
-        spec: session.spec,
-        content: contentOf(session.file),
-        aloneSpeedX: Number.isFinite(alone) && alone > 0 ? alone : null,
-        segmentKbps: segmentRatesOf(session)
-      });
-    },
-    viewerSecondsOn: (output, consumerId, now) => viewerSecondsOn(parts.viewers, output, consumerId, now),
-    noteRunStarting: (output) => parts.quality.noteRunStarting(output),
-    inputOf: (...args) => parts.renditions.inputOf(...args),
-    producedNumbers: (...args) => parts.serving.producedNumbers(...args),
-    servesAudioSeparately: (...args) => parts.renditions.servesAudioSeparately(...args),
-    disposeSession: (...args) => parts.lifecycle.disposeSession(...args),
-    get contentionPenalties() { return parts.contentionPenalties; },
-    get encodeCost() { return parts.encodeCost; },
-    get encodeOrchestrator() { return parts.encodeOrchestrator; },
-    get ffmpegBin() { return parts.ffmpegBin; },
-    get outputTimes() { return parts.outputTimes; },
-    get outputs() { return parts.outputs; },
-    get priority() { return parts.priority; },
-    get segmentDurationSec() { return parts.segmentDurationSec; },
-    segmentFiles,
-    get videoEncoder() { return parts.encoders.current; },
-    get encoders() { return parts.encoders; },
-    invalidateWaits: (output) => parts.serving.invalidateWaits(output),
-    productionFailed: (output) => parts.renditions.noteProductionFailed(output),
-  });
-  // How much film is ready in front of the viewers, said; and the spare soundtracks fetched once it is full.
-  parts.cushion = new CushionReport({
-    viewerSecondsOn: (output, consumerId, now) => viewerSecondsOn(parts.viewers, output, consumerId, now),
-    linkReportsOn: (output) => linkReportsOn(parts.viewers, output),
-    SourceFiles,
-    logger,
-    producedNumbers: (...args) => parts.serving.producedNumbers(...args),
-    get encodeRuns() { return parts.encodeRuns; },
-    get fetchWholeFile() { return parts.fetchWholeFile; },
-    get getCachedAudioTracks() { return parts.getCachedAudioTracks; },
-    get hostLoad() { return parts.hostLoad; },
-    get lookaheadSeconds() { return parts.lookaheadSeconds; },
-    get outputTimes() { return parts.outputTimes; },
-    get outputs() { return parts.outputs; },
-  });
-  // The steps of a picture and its soundtracks: which output answers each, made when first asked for, and the master playlist listing them.
-  // The cushion this file needs, from its segment length and the worst
-  // interruption its supply has shown — one statement read by every operation
-  // that asks it.
-  const minimumBufferSecondsOf = (output) => minimumBufferFrom({
-    segmentSeconds: parts.segmentDurationSec,
-    worstSupplyWaitSec: parts.hostLoad.supplyFor(output.file)?.worstWaitSec
-  })?.seconds ?? null;
-  parts.renditions = new Renditions({
-    observedPeakMbps: (spec) => parts.localObservations?.peakMbps(spec) ?? null,
-    // A viewer has been moved onto an output prepared for them: how long it
-    // took, and what they held.
-    notePreparation: (output, seconds, bufferedSec) =>
-      parts.localObservations?.notePreparation({ spec: output?.spec, seconds, bufferedSec }),
-    viewerSecondsOn: (output, consumerId, now) => viewerSecondsOn(parts.viewers, output, consumerId, now),
-    audioStartSecondsFor: (args) => audioStartSecondsFor({ ...args, viewers: parts.viewers }),
-    activeOutputFor: (args) => activeOutputFor({ ...args, viewers: parts.viewers }),
-    // WHAT ENCODING IS TOLD ABOUT A PERSON, and what it may tell the viewer
-    // layer about them: values and names, never the viewer itself. Steps and
-    // soundtracks are chosen, warmed and left per person, so this is the widest
-    // of the three sets — and every one of them is a call on the owner, which
-    // is what keeps one writer per fact.
-    viewerCountOn: (output) => parts.viewers.forOutput(output).size,
-    // Whether anything still stands on it — somebody watching, or an
-    // assignment not yet released. See `Viewers.stillNeeded`.
-    outputStillNeeded: (output) => parts.viewers.stillNeeded(output),
-    consumersOn: (output) => consumersOn(parts.viewers, output),
-    presentOn: (output) => presentOn(parts.viewers, output),
-    qualityModeOf: (output, consumerId) => qualityModeOf(parts.viewers, output, consumerId),
-    linkMbpsOf: (output, consumerId) => linkMbpsOf(parts.viewers, output, consumerId),
-    audioChoiceOf: (output, consumerId) => audioChoiceOf(parts.viewers, output, consumerId),
-    chooseAudioTrack: (output, consumerId, choice) =>
-      chooseAudioTrack(parts.viewers, output, consumerId, choice),
-    stepOnScreenOf: (output, consumerId) => stepOnScreenOf(parts.viewers, output, consumerId),
-    noteStepOnScreen: (output, consumerId, stepId) =>
-      noteStepOnScreen(parts.viewers, output, consumerId, stepId),
-    stepBeingWarmedOf: (output, consumerId) => stepBeingWarmedOf(parts.viewers, output, consumerId),
-    stepBeingWarmedSinceOf: (output, consumerId) => stepBeingWarmedSinceOf(parts.viewers, output, consumerId),
-    noteStepBeingWarmed: (output, consumerId, stepId) =>
-      noteStepBeingWarmed(parts.viewers, output, consumerId, stepId),
-    audioBeingWarmedOf: (output, consumerId) => audioBeingWarmedOf(parts.viewers, output, consumerId),
-    noteAudioBeingWarmed: (output, consumerId, renditionId) =>
-      noteAudioBeingWarmed(parts.viewers, output, consumerId, renditionId),
-    watches: (output, consumerId) => watches(parts.viewers, output, consumerId),
-    placeOn: (output, consumerId, seconds) => placeOn(parts.viewers, output, consumerId, seconds),
-    // A viewer's own assignments: which output answered a height and segment in
-    // one generation of their viewing, so a repeat is answered the same way.
-    // Values in and out, like everything above.
-    generationOfRequest: (consumerId, stated) => generationOfRequest(parts.viewers, consumerId, stated),
-    givenOutputOf: (consumerId, generation, askedHeight, segmentIndex) =>
-      givenOutputOf(parts.viewers, consumerId, generation, askedHeight, segmentIndex),
-    noteGivenOutput: (consumerId, generation, askedHeight, segmentIndex, outputKey) =>
-      noteGivenOutput(parts.viewers, consumerId, generation, askedHeight, segmentIndex, outputKey),
-    // The output this viewer was chosen at a height, and the rule's choice of
-    // it — the viewer's own record, never one per file.
-    chosenOutputOf: (consumerId, askedHeight) => chosenOutputOf(parts.viewers, consumerId, askedHeight),
-    chooseOutput: (consumerId, askedHeight, outputKey) => chooseOutput(parts.viewers, consumerId, askedHeight, outputKey),
-    noteServingVerdict: (consumerId, verdict) => noteServingVerdict(parts.viewers, consumerId, verdict),
-    // A move between two limits of the height on a viewer's screen (roadmap
-    // item 97, step 12): which heights their choice is the output on screen
-    // under, the segment they were last given, the move being prepared, and
-    // whether a segment is closed on the output being prepared.
-    limitsFor: (frame) => parts.limitsFor(frame),
-    heightsChosenAs: (consumerId, outputKey) => heightsChosenAs(parts.viewers, consumerId, outputKey),
-    highestGivenSegmentOf: (consumerId, askedHeight) => highestGivenSegmentOf(parts.viewers, consumerId, askedHeight),
-    sameHeightSwitchOf: (consumerId) => sameHeightSwitchOf(parts.viewers, consumerId),
-    noteSameHeightSwitch: (consumerId, value) => noteSameHeightSwitch(parts.viewers, consumerId, value),
-    switchingOnto: (output) => switchingOnto(parts.viewers, output),
-    // What a viewer holds, and the cushion this file needs before a move that
-    // is not urgent is made (roadmap item 98).
-    bufferedSecondsOf: (consumerId) => bufferedSecondsOf(parts.viewers, consumerId),
-    minimumBufferSecondsFor: (output) => minimumBufferSecondsOf(output),
-    // Whether the whole machine has a place for one more encoder on this output,
-    // asked before a preparation is recorded.
-    admitsPreparation: (output) => parts.admission.admitsPreparation(output?.outputKey ?? ""),
-    segmentClosed: (key, index) => parts.segmentStore.isClosed(key, index),
-    // What a gone output left in the store, and whether another output's header
-    // may stand in for its own.
-    storedPieceReady: (key, fileName) => parts.serving.hasStoredPiece(key, fileName),
-    headersCompatible: (goneKey, liveKey) => parts.serving.headersCompatible(goneKey, liveKey),
-    placeViewerOn: (...args) => parts.viewerRequests.placeViewerOn(...args),
-    audioRenditionName,
-    logger,
-    viewerLeaves: (...args) => parts.lifecycle.viewerLeaves(...args),
-    createOrGetSession: (...args) => parts.viewerRequests.createOrGetSession(...args),
-    planEncodersSoon: (...args) => parts.encodeRuns.planEncodersSoon(...args),
-    disposeSession: (...args) => parts.lifecycle.disposeSession(...args),
-    viewerPositionOf: (...args) => parts.viewerRequests.viewerPositionOf(...args),
-    get encodeRuns() { return parts.encodeRuns; },
-    get fileStartTimeReads() { return parts.fileStartTimeReads; },
-    set fileStartTimeReads(value) { parts.fileStartTimeReads = value; },
-    get getCachedAudioTracks() { return parts.getCachedAudioTracks; },
-    get getCachedMediaInfo() { return parts.getCachedMediaInfo; },
-    get getContainerMediaInfo() { return parts.getContainerMediaInfo; },
-    get localBaseUrl() { return parts.localBaseUrl; },
-    get outputTimes() { return parts.outputTimes; },
-    get outputs() { return parts.outputs; },
-    get quality() { return parts.quality; },
-    get qualityOffer() { return parts.qualityOffer; },
-    get segmentDurationSec() { return parts.segmentDurationSec; },
-    get sourceFiles() { return parts.sourceFiles; },
-    invalidateWaits: (output) => parts.serving.invalidateWaits(output),
-  });
-  // Answering a request for a playlist, an init segment or a segment: from the store when the piece is made and whole, otherwise held until it is.
-  parts.serving = new SegmentServing({
-    // What a key names, read back from the key: how to find a gone output's
-    // stored pieces and whether they are cut at given times.
-    segmentFormatOfKey,
-    specOfKey: (key) => OutputSpec.fromKey(key),
-    compareInits,
-    activeOutputFor: (args) => activeOutputFor({ ...args, viewers: parts.viewers }),
-    viewerSecondsOn: (output, consumerId, now) => viewerSecondsOn(parts.viewers, output, consumerId, now),
-    buildMasterPlaylist: (...args) => parts.renditions.buildMasterPlaylist(...args),
-    declaredTracks: (...args) => parts.renditions.declaredTracks(...args),
-    publishedGridFor: (...args) => parts.outputTimes.publishedGridFor(...args),
-    runStartTimeFor: (...args) => parts.outputTimes.runStartTimeFor(...args),
-    get cushion() { return parts.cushion; },
-    get encodeOrchestrator() { return parts.encodeOrchestrator; },
-    get encodeRuns() { return parts.encodeRuns; },
-    get hostTimings() { return parts.hostTimings; },
-    get lookaheadSeconds() { return parts.lookaheadSeconds; },
-    get outputTimes() { return parts.outputTimes; },
-    get outputs() { return parts.outputs; },
-    get segmentStore() { return parts.segmentStore; },
-    get startupWaitMs() { return parts.startupWaitMs; },
-    get viewers() { return parts.viewers; },
-  });
-  // How an output ends: disposed when nobody is left on it and it has been idle, or all at once on shutdown; and the segments an earlier process left behind adopted at startup.
-  parts.lifecycle = new OutputLifecycle({
-    invalidateWaits: (output) => parts.serving.invalidateWaits(output),
-    segmentFormatOfKey,
-    // Where the viewers of one output stand, in its segments: what the store
-    // reads to give disk back in the right order.
-    viewerSegmentsOn: (key) => viewerSegmentsOn({
-      outputs: parts.outputs.values(),
-      outputKey: key,
-      segmentAt: (output, seconds) => parts.outputTimes.segmentIndexForTime(output, seconds),
-      now: Date.now(),
-      viewers: parts.viewers
-    }),
-    get budgetTimer() { return parts.budgetTimer; },
-    get cleanupTimer() { return parts.cleanupTimer; },
-    get encodeRuns() { return parts.encodeRuns; },
-    get keyframeTables() { return parts.keyframeTables; },
-    get machineBudget() { return parts.machineBudget; },
-    get outputTimes() { return parts.outputTimes; },
-    get outputs() { return parts.outputs; },
-    get returns() { return parts.returns; },
-    get segmentStore() { return parts.segmentStore; },
-    get sessionTtlMs() { return parts.sessionTtlMs; },
-    get sourceFiles() { return parts.sourceFiles; },
-    get timelines() { return parts.timelines; },
-    get viewers() { return parts.viewers; },
-  });
-  // Which output answers a request for a file, made if it is not here yet: the encoding component's decision, made without knowing who asked.
-  parts.opening = new OutputOpening({
-    logger,
-    // The largest peak an output of this mode and rate control has been seen
-    // carrying here: the only figure an encoder with no bound of its own has.
-    observedPeakMbps: (spec) => parts.localObservations?.peakMbps(spec) ?? null,
-    limitsFor: (frame) => parts.limitsFor(frame),
-    probeMediaInfo: (url) => probeInputMediaInfo(parts.ffmpegBin, url),
-    probeKeyframeTimes: (url, budgetMs) => probeVideoKeyframeTimes(parts.ffmpegBin, url, budgetMs),
-    get decodeCostModel() { return parts.decodeCostModel; },
-    get enabled() { return parts.enabled; },
-    get encodeCost() { return parts.encodeCost; },
-    get encodeRuns() { return parts.encodeRuns; },
-    get admission() { return parts.admission; },
-    get getCachedMediaInfo() { return parts.getCachedMediaInfo; },
-    get hostLoad() { return parts.hostLoad; },
-    get hostTimings() { return parts.hostTimings; },
-    get keyframeTables() { return parts.keyframeTables; },
-    get localBaseUrl() { return parts.localBaseUrl; },
-    get outputs() { return parts.outputs; },
-    get renditions() { return parts.renditions; },
-    get returns() { return parts.returns; },
-    get segmentDurationSec() { return parts.segmentDurationSec; },
-    get segmentFormat() { return parts.segmentFormat; },
-    segmentOutputFiles,
-    get softwarePresetBenchmark() { return parts.softwarePresetBenchmark; },
-    get sourceFiles() { return parts.sourceFiles; },
-    get timelines() { return parts.timelines; },
-    get tonemapSupported() { return parts.tonemapSupported; },
-    get videoEncoder() { return parts.encoders.current; },
-  });
-  // What a viewer asks for: an output of a file, a position, a report of progress.
-  parts.viewerRequests = new ViewerRequests({
-    // Whether this machine has a place for one more viewer on this output —
-    // asked in the same synchronous stretch as they are put on it.
-    admitsWatching: (output) => parts.admission.admitsWatching(output?.outputKey ?? ""),
-    activeOutputFor: (args) => activeOutputFor({ ...args, viewers: parts.viewers }),
-    viewerSecondsOn: (output, consumerId, now) => viewerSecondsOn(parts.viewers, output, consumerId, now),
-    minimumBufferSecondsFor: (output) => minimumBufferSecondsOf(output),
-    getSourceStats: (...args) => parts.getSourceStats?.(...args) ?? null,
-    disposeSession: (...args) => parts.lifecycle.disposeSession(...args),
-    expectedFirstSegmentMs: (...args) => parts.hostTimings.expectedFirstSegmentMs(...args),
-    expectedSessionCreateMs: (...args) => parts.hostTimings.expectedSessionCreateMs(...args),
-    planEncodersSoon: (...args) => parts.encodeRuns.planEncodersSoon(...args),
-    waitUntilReady: (...args) => parts.serving.waitUntilReady(...args),
-    get encodeRuns() { return parts.encodeRuns; },
-    encodeSpeedReadingOf: (output) => parts.encodeCost.latestSpeedReadingOf(output),
-    projectedEncodeSpeedOf: () => parts.admission.projectedSpeedX(),
-    get opening() { return parts.opening; },
-    get outputTimes() { return parts.outputTimes; },
-    get outputs() { return parts.outputs; },
-    get lookaheadSeconds() { return parts.lookaheadSeconds; },
-    get quality() { return parts.quality; },
-    get qualityOffer() { return parts.qualityOffer; },
-    get renditions() { return parts.renditions; },
-    get segmentDurationSec() { return parts.segmentDurationSec; },
-    get segmentStore() { return parts.segmentStore; },
-    get sourceFiles() { return parts.sourceFiles; },
-    get startupWaitMs() { return parts.startupWaitMs; },
-    get viewers() { return parts.viewers; },
-    invalidateWaits: (output) => parts.serving.invalidateWaits(output),
-    // Whether a request of this generation is still taken, what it belongs to,
-    // what answered it, and the hold a response keeps while it is being sent.
-    acceptsGeneration: (consumerId, stated) => acceptsGeneration(parts.viewers, consumerId, stated),
-    generationOfRequest: (consumerId, stated) => generationOfRequest(parts.viewers, consumerId, stated),
-    noteGivenOutput: (consumerId, generation, askedHeight, segmentIndex, outputKey) =>
-      noteGivenOutput(parts.viewers, consumerId, generation, askedHeight, segmentIndex, outputKey),
-    holdForResponse: (consumerId, outputKey) => holdForResponse(parts.viewers, consumerId, outputKey),
-    noteServingVerdict: (consumerId, verdict) => noteServingVerdict(parts.viewers, consumerId, verdict),
-    servingVerdictOf: (consumerId) => servingVerdictOf(parts.viewers, consumerId),
-  });
-  // What this host takes to create an output and to produce its first segment, kept across restarts until the synthetic figure can replace it (CLAUDE.md, host timings).
-  parts.hostTimings = new HostTimings({
-    get segmentDurationSec() { return parts.segmentDurationSec; },
-    get softwarePresetBenchmark() { return parts.softwarePresetBenchmark; },
-    get stateDir() { return parts.stateDir; },
-  });
-  parts.enabled = Boolean(enabled);
-  parts.ffmpegBin = ffmpegBin;
-  parts.keyframeTables = keyframeTables;
-  // Where measurements about this host are kept between runs. Empty means
-  // beside the installed proxy; a deployment with somewhere persistent to
-  // write names it (--state-dir).
-  parts.stateDir = typeof stateDir === "string" ? stateDir : "";
-  // Output container (fMP4/CMAF or MPEG-TS). Everything container-specific —
-  // muxer args, file naming, playlist header, per-segment correction — lives
-  // in this module; nothing here branches on the format.
-  parts.segmentFormat = resolveSegmentFormat(segmentFormatId);
-  // Optional accessor for media info the playback planner already probed for
-  // (sourceKey, fileIndex), so session create can skip its own ffmpeg scan.
-  parts.getCachedMediaInfo = typeof getCachedMediaInfo === "function" ? getCachedMediaInfo : null;
-  // The file's audio tracks, for the master playlist's rendition group. Same
-  // inventory the browser's audio menu is built from.
-  parts.getCachedAudioTracks = typeof getCachedAudioTracks === "function" ? getCachedAudioTracks : null;
-  // What a file declares about itself, read by the container layer from the
-  // same header its track table comes from: format, duration, and where its
-  // own timeline begins. The last of those is why this exists — a soundtrack
-  // shipped as its own file has a timeline of its own, and asking ffmpeg for
-  // it meant reading a header this proxy had already read.
-  parts.getContainerMediaInfo =
-    typeof getContainerMediaInfo === "function" ? getContainerMediaInfo : null;
-  // Fetch one whole file of a source, as a bounded read rather than a
-  // selection. Used to pull a soundtrack that ships beside the picture onto
-  // the disk while the swarm has capacity to spare — see
-  // `#fetchSpareSoundtracks`. Optional: a proxy wired without it simply reads
-  // such a soundtrack when it is played.
-  parts.fetchWholeFile = typeof fetchWholeFile === "function" ? fetchWholeFile : null;
-  // Optional async accessor for a source's live download stats, used by the
-  // realtime budget to tell a CPU limit from a download-starved input:
-  // (sourceKey, fileIndex) => Promise<{ downloadSpeed, fileLength, fileProgress } | null>.
-  parts.getSourceStats = typeof getSourceStats === "function" ? getSourceStats : null;
-  parts.setPriorityMap = typeof setPriorityMap === "function" ? setPriorityMap : null;
-  parts.contentionPenalties = contentionPenalties instanceof Map ? contentionPenalties : null;
-  // Seconds of film per second when the picture is COPIED, measured at
-  // startup. Nothing else prices that branch: the other benchmarks measure
-  // encoding and decoding, and a copy does neither.
-  parts.copySpeedX = Number.isFinite(copySpeedX) && copySpeedX > 0 ? copySpeedX : null;
-  // Totals across every torrent this proxy holds, used to price what the
-  // torrent itself costs the machine (item 7). Optional: a proxy wired
-  // without it simply never learns that figure.
-  parts.getTorrentTotals = typeof getTorrentTotals === "function" ? getTorrentTotals : null;
-  // The spilled pieces, as a pair of closures over the torrent thread: what
-  // they weigh and how to tell them their share. Not the pool — the pool is
-  // on the other side of the thread boundary and this side holds none of it.
-  parts.spillDisk = spillDisk;
-  parts.wholeFiles = wholeFiles;
-  parts.diagnostics = diagnostics;
-  parts.diagnosticsRoot = diagnosticsRoot;
-  parts.memoryClaimant = memoryClaimant;
-  parts.budgetPolicy = budgetPolicy;
-  // Detected H.264 encoder descriptor (hardware or software). Defaults to
-  // software libx264 when no detection result is supplied. May be downgraded
-  // to software at runtime if a hardware encode fails.
-  parts.encoders = new EncoderSelection({ detected: videoEncoder, softwareDescriptor });
-  // WHAT THIS HOST HAS SEEN ITS OWN ADMITTED ENCODES DO, kept between runs and
-  // used only on the configuration it was seen on (roadmap item 97, step 14).
-  // A refinement of figures that exist without it: nothing is selected,
-  // admitted or served differently when it is empty. Absent where the startup
-  // calibration was not run, which is only a wiring made without it.
-  parts.localObservations = null;
-  if (calibration?.fingerprint) {
-    parts.localObservations = new LocalObservations({
-      fingerprint: calibration.fingerprint,
-      filePath: path.join(typeof stateDir === "string" && stateDir.length > 0 ? stateDir : PROXY_ROOT, "local-observations.json"),
-      logger
-    });
-    parts.localObservations.load(calibration.kinds().map((kind) => configurationKeyOf(calibration.fingerprint, {
-      kind,
-      name: kind === "software" ? softwareDescriptor().name : (videoEncoder?.kind === kind ? videoEncoder.name : kind)
-    })));
-  }
-  /**
-   * What the segments an output has closed carried, from their sizes and the
-   * spans its timeline gives them: the average over all of them and the
-   * largest one.
-   *
-   * @param {object} session
-   * @returns {{ averageKbps: number, peakKbps: number } | null}
-   */
-  const segmentRatesOf = (session) => {
-    const sizes = parts.segmentStore.sizesOf(session.outputKey);
-    const timeline = session.timeline;
-    if (!timeline || sizes.size === 0) {
-      return null;
-    }
-    let bits = 0;
-    let seconds = 0;
-    let peakKbps = 0;
-    for (const [index, size] of sizes) {
-      const start = timeline.publishedStartOf(index);
-      const end = index < timeline.segmentCount ? timeline.publishedStartOf(index + 1) : null;
-      if (!(Number.isFinite(start) && Number.isFinite(end) && end > start) || !(size > 0)) {
-        continue;
-      }
-      bits += size * 8;
-      seconds += end - start;
-      peakKbps = Math.max(peakKbps, (size * 8) / (end - start) / 1000);
-    }
-    return seconds > 0 ? { averageKbps: bits / seconds / 1000, peakKbps } : null;
-  };
-  // The modes of the encoder IN USE NOW, as measured at startup: slowest
-  // first, each with its throughput by size. Read through the selection rather
-  // than copied, because the encoder in use can change — a failing device
-  // moves this host to software for the rest of the process, and from then on
-  // it is software's own modes that price and choose, not the device's.
-  // Null where nothing was measured, and then nothing is re-encoded here.
-  parts.calibration = calibration ?? null;
-  const givenModes = Array.isArray(softwarePresetBenchmark) ? softwarePresetBenchmark : null;
-  Object.defineProperty(parts, "softwarePresetBenchmark", {
-    get: () => (parts.calibration
-      ? parts.calibration.modesFor(parts.encoders.current?.kind)
-      : givenModes),
-    enumerable: true
-  });
-  // Host decode cost solved at startup from the calibration clips:
-  // `a × Mpixel/s + b × Mbit/s + c` seconds of decoding per second of video.
-  // A re-encode pays for this as well as for the encoder, and leaving it out
-  // is what made the budget offer rungs this host ran at a third of realtime.
-  // Null when the clips are missing or the fit was rejected — the budget then
-  // prices the encoder alone, as it did before.
-  parts.decodeCostModel = decodeCostModel ?? null;
-  // Whether this ffmpeg build can tone-map HDR→SDR (zscale + tonemap filters).
-  // Gates the tonemap chain for HDR sources on the software path.
-  parts.tonemapSupported = Boolean(tonemapSupported);
-  parts.segmentDurationSec = segmentDurationSec;
-  // How far ahead of the viewer this proxy lets an encoder run, in seconds of
-  // playback. Stated rather than kept private, because the browser's forward
-  // buffer is bounded by the same quantity and used to carry a copy of its
-  // own: a hand-written 30 s, justified in a comment by a DIFFERENT constant
-  // (the eight segments that bound a request ahead of the ENCODE HEAD), so
-  // three quarters of what the encoder had already produced was thrown away.
-  // One figure, said by the side that owns it (roadmap item 4).
-  parts.lookaheadSeconds = LOOKAHEAD_PAUSE_SECONDS;
-  parts.sessionTtlMs = sessionTtlMs;
-  parts.startupWaitMs = startupWaitMs;
-  parts.localBaseUrl = buildHttpBaseUrl(localBindHost, localPort);
-  // Everyone watching anything, one object per person rather than one per
-  // person per output. What a viewer chose, where they are and which outputs
-  // they are watching are facts about the person; kept per output they were
-  // three copies of which two were always stale.
-  // Every change to who is watching what re-decides which encoders should
-  // exist, because that decision reads nothing else about viewers. It used to
-  // be re-taken on a five-second timer instead, which made a just-created
-  // output wait up to five seconds before anything noticed it had a viewer.
-  parts.viewers = new Viewers({ onChange: () => parts.encodeRuns.planEncodersSoon() });
-  // The outputs that exist, and every question about them: the picture a step
-  // belongs to, the steps, the soundtracks, the height a step is named by.
-  parts.outputs = new OutputCatalog({
-    fileLengthOf: (session) => parts.hostLoad.fileLengthByKey.get(session.file.key) ?? 0,
-    largestPieceOf: (address) => parts.segmentStore.largestPiece(address)
-  });
-  // What this host learned last time it ran. Without it every restart shows
-  // the first viewer a figure with no measurement behind it.
-  parts.hostTimings.loadHostTimings();
-  // Where produced segments live, addressed by WHAT they are rather than by
-  // which session's encoder wrote them. Two sessions of one output — two
-  // viewers who opened the same film at different places — write into one
-  // directory and each serves what the other has already made. Injectable so
-  // a test can give it a root of its own.
-  parts.segmentStore = segmentStore instanceof SegmentStore
-    ? segmentStore
-    : new SegmentStore({ logger });
-  // The store has grown: whether it still fits its share is asked now rather
-  // than on the next cleanup pass.
-  parts.segmentStore.onPublished(() => parts.lifecycle.keepWithinRoomSoon());
-  // A segment closed may be the one a viewer's move between two limits of a
-  // height is waiting for.
-  parts.segmentStore.onPublished((key, index) => parts.renditions.noteSegmentPublished(key, index));
-  // What encoders there should be on each output, and where. Given the two
-  // things only this class can answer: how many this machine can afford, and
-  // how to make one.
-  // What encoding costs this machine, and which heights follow from it. Given
-  // what it cannot work out for itself: which sessions belong to one file, the
-  // host's own readings AT THE MOMENT OF THE QUESTION rather than copied, how
-  // a soundtrack is keyed, how many encoders are running, and what a file
-  // costs merely by being fetched.
-  parts.encodeCost = new EncodeCost({
-    outputs: parts.outputs,
-    host: () => ({
-      benchmark: parts.softwarePresetBenchmark,
-      decodeModel: parts.decodeCostModel,
-      contentionPenalties: parts.contentionPenalties,
-      copySpeedX: parts.copySpeedX,
-      availability: parts.hostLoad.hostAvailability,
-      // WHICH encoder this host settled on. Only the software ladder is
-      // benchmarked, so a reading taken off a hardware encoder cannot be
-      // split into its decode and encode halves and is not filed as one.
-      encoderKind: parts.encoders.current?.kind ?? null,
-      // The slowest this output's mode has been seen running alone on
-      // comparable material on this configuration, or null.
-      observedAloneSpeed: (session) => parts.localObservations?.slowestAloneSpeed(session.spec, contentOf(session.file)) ?? null
-    }),
-    // HOW MANY ENCODER PROCESSES ARE RUNNING, asked of the one thing that
-    // makes and unmakes them. This used to be two counts of one fact, taken
-    // by walking the session registry — one per session, one per run — and
-    // they agreed only while a session held at most one run.
-    runningEncoders: () => parts.encodeOrchestrator.runningCount(),
-    encodersRunningNow: () => parts.encodeOrchestrator.runningCount(),
-    torrentCostSecFor: (session) => parts.hostLoad.torrentCostSecFor(session),
-    boundBy: (session) => parts.quality.classifyTranscodeBound(session),
-    runsFor: (session) => parts.encodeRuns.runsOf(session),
-    stateFor: (session) => parts.encodeRuns.runStateOf(session),
-    // The run's OWN progress: a speed is a pair of readings of one process.
-    progressFor: (_session, run) => run?.progress?.snapshot?.() ?? null
-  });
-  // WHICH HEIGHTS ARE ON THE MENU, which is the arithmetic above plus three
-  // things that are nothing to do with it: whose answer it is, what may never
-  // be withdrawn, and when the answer may be reused.
-  parts.qualityOffer = new QualityOffer({
-    encodeCost: parts.encodeCost,
-    occupiedCostSec: (fileKey) => parts.admission?.occupiedCost({ exceptFileKey: fileKey }).costSec ?? 0,
-    occupancyKnownFor: (fileKey) => (parts.admission?.occupiedCost({ exceptFileKey: fileKey }).unpriced ?? 0) === 0,
-    outputs: parts.outputs,
-    stateFor: (session) => parts.encodeRuns.runStateOf(session),
-    // WHICH HEIGHTS A LIVE VIEWER HAS ON SCREEN, as numbers. Who is watching
-    // what is the viewer layer's, and which session is which height is the
-    // film's shape; neither travels — what crosses is the list of heights.
-    heightsOnScreen: (owner) => parts.quality.heightsOnScreen(owner),
-    // WHAT THE SWARM IS DOING WITH THIS FILE. Three readings, taken by
-    // whoever reads it and handed over as three numbers: the speed this
-    // file's own interruptions demand, the megabytes a second a viewer draws
-    // through it, and what a megabyte costs this process.
-    supplyFor: (file) => ({
-      requiredSpeed: parts.hostLoad.requiredSpeedFor(file.sourceKey, file.fileIndex),
-      megabytesPerSecond: parts.hostLoad.torrentMegabytesPerSecond(
-        file.sourceKey,
-        file.fileIndex,
-        file.lengthBytes ?? parts.hostLoad.fileLengthByKey.get(file.key) ?? null,
-        file.durationSeconds
-      ),
-      costPerMegabyte: parts.hostLoad.observedTorrentCostPerMegabyte
-    })
-  });
-  // The priority map, built from where the viewers are and handed to both
-  // sides that act on it. The downloading lives in another thread, so its
-  // copy travels over the worker channel.
-  parts.priority = new PriorityOrchestrator({
-    publish: ({ sourceKey, fileIndex, durationSeconds, zones }) => {
-      void Promise.resolve(
-        parts.setPriorityMap?.({ sourceKey, fileIndex, durationSeconds, zones })
-      ).catch(() => {});
-    },
-    get viewers() { return parts.viewers; },
-    // WHERE THE TWO FACTS MEET, and this is the only place that holds both.
-    // Which step is on somebody's screen belongs to the person; which output
-    // a step supersedes belongs to the film's shape. Neither layer is handed
-    // the other — one gets a plain id, the other is read for one field.
-    watchedBy: (session, viewer) => !parts.outputs.supersededBy(session, viewer.activeVariantId ?? null),
-    allowanceFor: (session) => minimumBufferFrom({
-      segmentSeconds: parts.segmentDurationSec,
-      worstSupplyWaitSec: parts.hostLoad.supplyFor(session.file)?.worstWaitSec
-    })?.seconds ?? parts.segmentDurationSec
-  });
-  // WHETHER THE WHOLE MACHINE HAS A PLACE for one more encoder, across outputs.
-  // Given what it cannot hold itself: what runs (the orchestrator's), what is
-  // being prepared (the viewers' records), what an output costs (the encoding
-  // cost) and how much of the machine nobody has priced (the host's reading).
-  parts.admission = new EncodeAdmission({
-    liveRunsByAddress: () => parts.encodeOrchestrator.liveRunsByAddress(),
-    preparedAddresses: () => {
-      const addresses = new Set();
-      for (const id of outputsBeingPrepared(parts.viewers)) {
-        const key = parts.outputs.get(id)?.outputKey;
-        if (key) {
-          addresses.add(key);
-        }
-      }
-      return addresses;
-    },
-    // Every output a present viewer RECEIVES — by the same rule the per-output
-    // priority map is built by — that still has something left to make. The
-    // place an output is opened on is held from the moment its viewer is put
-    // on it, not from when the plan gets round to its first encoder.
-    watchedAddresses: () => {
-      const addresses = new Set();
-      for (const session of parts.outputs.values()) {
-        const address = session.outputKey;
-        if (!address || addresses.has(address) || finishedOutput(session)) {
-          continue;
-        }
-        for (const viewer of parts.viewers.forOutput(session).values()) {
-          if (viewer.isPresent() && !parts.outputs.supersededBy(session, viewer.activeVariantId ?? null)) {
-            addresses.add(address);
-            break;
-          }
-        }
-      }
-      return addresses;
-    },
-    finished: (address) => parts.outputs.outputsOn(address).some((session) => finishedOutput(session)),
-    loadOf: (address) => parts.encodeCost.loadOfOutput(address),
-    loadForCandidate: (spec, file) => parts.encodeCost.loadForCandidate(spec, file),
-    availability: () => parts.hostLoad.hostAvailability ?? null
-  });
-  parts.encodeOrchestrator = new EncodeOrchestrator({
-    admission: parts.admission,
-    // What the viewers actually waited for, by band. The ledger is the
-    // priority layer's; the encoding is handed a way to ask it.
-    describeWaits: (address) => waits.describe(address),
-    maxRunsFor: (address) => parts.encodeRuns.maxRunsForOutput(address),
-    makeRun: ({ address, from, to, because }) => parts.encodeRuns.makeRunAt(address, from, to, because),
-    segmentSeconds: parts.segmentDurationSec,
-    contentionPenalties: parts.contentionPenalties,
-    startingSpeedFor: (address) => parts.encodeCost.speedForOutput(address),
-    segmentCoverage,
-    // HOW IT ASKS TO DECIDE AGAIN. A plan that refuses to place anything
-    // because an output's input is away needs something to bring it back:
-    // nothing about the state changes while the data is missing, so no event
-    // arrives on its own.
-    planSoon: () => parts.encodeRuns.planEncodersSoon(),
-    logger
-  });
-  // What a start and a stop were measured to cost here, before any viewer
-  // existed. Without it both read zero at a cold open, and zero is not
-  // "unmeasured" — it is "free", which is what moved an encoder between two
-  // adjacent numbers every half second in the field.
-  parts.encodeOrchestrator.noteStartupCosts(startStopCost);
-  // Where each file is cut, held once per file and grid rather than once per
-  // session. Two sessions of one film MUST agree about this to the
-  // millisecond — a segment made by either has to be appendable where the
-  // other's would have gone — and until now they agreed by copying, which is
-  // a thing somebody has to remember to do and which drifted twice in the
-  // field. They share the table now.
-  parts.timelines = new Timelines();
-  // The files this proxy is serving, one object per file however many
-  // sessions are of it. It holds the file's key — which every cache about a
-  // file is keyed by — its name, and the facts a probe returned.
-  parts.sourceFiles = new SourceFiles();
-  parts.cleanupTimer = setInterval(() => {
-    void parts.lifecycle.cleanupExpired();
-  }, CLEANUP_INTERVAL_MS);
-  parts.cleanupTimer.unref();
-  // How long after material stops being read somebody asks for it again — the
-  // one term of the keeping period that is guessed rather than measured, and
-  // the only place it can be measured from.
-  parts.returns = new Returns();
-  // One owner of the disk, and the list of what takes it lives with the owner.
-  parts.machineBudget = wireMachineBudget({
-    segmentStore: parts.segmentStore,
-    spill: parts.spillDisk,
-    wholeFiles: parts.wholeFiles,
-    diagnostics: parts.diagnostics,
-    diagnosticsRoot: parts.diagnosticsRoot,
-    memory: parts.memoryClaimant,
-    policy: parts.budgetPolicy ?? undefined,
-    readFree: freeBytesFor,
-    logger
-  });
-  // Realtime-budget monitor: only meaningful for the software encoder with a
-  // benchmark (the only path that can pick/step resolution). Cheap no-op scan
-  // otherwise.
-  // The quality budget: what each output is asked to step to, and the bitrate ceiling a viewer's link sets. Everything it reads of the rest of the proxy is listed here.
-  parts.quality = new QualityController({
-    // The longest an output of this mode has been seen taking to be ready for
-    // a viewer moving onto it, in milliseconds, or null.
-    observedPreparationMs: (output) => {
-      const seconds = parts.localObservations?.longestPreparationSec(output?.spec) ?? null;
-      return seconds === null ? null : seconds * 1000;
-    },
-    // WHAT ENCODING IS TOLD ABOUT A PERSON: values, by name. Not the viewer,
-    // whose record belongs to the viewer layer and is written only there.
-    consumersOn: (output) => consumersOn(parts.viewers, output),
-    qualityModeOf: (output, consumerId) => qualityModeOf(parts.viewers, output, consumerId),
-    stepOnScreenOf: (output, consumerId) => stepOnScreenOf(parts.viewers, output, consumerId),
-    standingAskOf: (output, consumerId) => standingAskOf(parts.viewers, output, consumerId),
-    askQualityOf: (output, consumerId, height, reason, now, urgent) =>
-      askQualityOf(parts.viewers, output, consumerId, height, reason, now, urgent),
-    // What a viewer's page said about their buffer and the picture they see
-    // (roadmap item 98), as values.
-    bufferOf: (output, consumerId, spanSec) => bufferOf(parts.viewers, output, consumerId, spanSec),
-    visiblePictureOf: (output, consumerId) => visiblePictureOf(parts.viewers, output, consumerId),
-    // How long this host takes to close a first segment: how soon another
-    // output could have the piece a viewer needs.
-    expectedFirstSegmentMs: () => parts.hostTimings.expectedFirstSegmentMs(),
-    dropAskOf: (output, consumerId) => dropAskOf(parts.viewers, output, consumerId),
-    // EACH viewer's own link and who is present: a thin link decides for the
-    // person on it and for nobody else (roadmap item 97, step 11), so the
-    // budget asks per viewer instead of taking the worst over all of them.
-    presentOn: (output) => presentOn(parts.viewers, output),
-    linkReportOf: (output, consumerId) => linkReportOf(parts.viewers, output, consumerId),
-    // The soundtrack a viewer receives, as part of the load their link carries.
-    viewerAudioLoadOf: (base, consumerId) => parts.renditions.viewerAudioLoadOf(base, consumerId),
-    // The first lever on a viewer's link: another limit of the height on their
-    // screen, prepared and moved to without their player being told.
-    prepareSameHeightSwitch: (output, consumerId, direction, reason) =>
-      parts.renditions.prepareSameHeightSwitch(output, consumerId, direction, reason),
-    sameHeightSwitchPending: (consumerId) => parts.renditions.sameHeightSwitchPending(consumerId),
-    sameHeightSwitchDirection: (consumerId) => parts.renditions.sameHeightSwitchDirection(consumerId),
-    cancelSameHeightSwitch: (consumerId, reason) => parts.renditions.cancelSameHeightSwitch(consumerId, reason),
-    heightReadyFor: (base, consumerId, height) => parts.renditions.heightReadyFor(base, consumerId, height),
-    isLive: (...args) => parts.encodeRuns.isLive(...args),
-    liveConsumers: (...args) => parts.renditions.liveConsumers(...args),
-    liveRunsOf: (...args) => parts.encodeRuns.liveRunsOf(...args),
-    producedNumbers: (...args) => parts.serving.producedNumbers(...args),
-    reportHostLoad: (...args) => parts.hostLoad.reportHostLoad(...args),
-    runStateOf: (...args) => parts.encodeRuns.runStateOf(...args),
-    sampleDownloadRates: (...args) => parts.hostLoad.sampleDownloadRates(...args),
-    get encodeCost() { return parts.encodeCost; },
-    get getSourceStats() { return parts.getSourceStats; },
-    get outputs() { return parts.outputs; },
-    get qualityOffer() { return parts.qualityOffer; },
-    get segmentDurationSec() { return parts.segmentDurationSec; },
-    segmentPaths,
-    get videoEncoder() { return parts.encoders.current; },
-  });
-  parts.budgetTimer = setInterval(() => {
-    parts.cushion.reportCushions();
-    void parts.quality.runQualityBudgetOnce();
-  }, BUDGET_CHECK_INTERVAL_MS);
-  parts.budgetTimer.unref();
-  return parts;
-}
+/**
+ * @file Wiring of the outputs, the viewers and the encoding.
+ *
+ * Builds each component and hands it the narrow host it reads, and nothing
+ * else. It holds no behaviour: every member that used to be a method of the
+ * session manager is a method of the component that owns it now. What it
+ * returns is the set of components and the settings they share, for
+ * `server.js` to give each route the ones that route asks.
+ */
+
+import { logger } from "../../utils/logger.js";
+import { KeyframeTables } from "../media/KeyframeTables.js";
+import { waits } from "./WaitLedger.js";
+import { readMachineState, readProcessCpuSeconds, readProxyCpuSeconds, readSystemCpu, shareOfMachine } from "../encode/host-load.js";
+import { minimumBufferFrom } from "../torrent/supply-margin.js";
+import { PriorityOrchestrator } from "../viewer/PriorityOrchestrator.js";
+import { softwareDescriptor } from "../encode/hwaccel.js";
+import { resolveSegmentFormat } from "../encode/segment-formats/index.js";
+import { Timelines } from "../encode/output/Timeline.js";
+import { SourceFiles } from "../media/SourceFile.js";
+import { SegmentStore } from "../storage/segment-store/SegmentStore.js";
+import { EncodeCost } from "../encode/quality/EncodeCost.js";
+import { QualityOffer } from "../encode/quality/QualityOffer.js";
+import { activeOutputFor } from "../viewer/active-output.js";
+import {
+  acceptsGeneration,
+  askQualityOf,
+  audioBeingWarmedOf,
+  audioChoiceOf,
+  chooseAudioTrack,
+  chooseOutput,
+  chosenOutputOf,
+  consumersOn,
+  dropAskOf,
+  generationOfRequest,
+  givenOutputOf,
+  heightsChosenAs,
+  highestGivenSegmentOf,
+  holdForResponse,
+  linkMbpsOf,
+  linkReportOf,
+  bufferOf,
+  bufferedSecondsOf,
+  visiblePictureOf,
+  linkReportsOn,
+  noteAudioBeingWarmed,
+  noteGivenOutput,
+  noteSameHeightSwitch,
+  noteServingVerdict,
+  noteStepBeingWarmed,
+  noteStepOnScreen,
+  outputsBeingPrepared,
+  placeOn,
+  presentOn,
+  qualityModeOf,
+  sameHeightSwitchOf,
+  servingVerdictOf,
+  standingAskOf,
+  stepBeingWarmedOf,
+  stepBeingWarmedSinceOf,
+  stepOnScreenOf,
+  switchingOnto,
+  watches
+} from "../viewer/choices.js";
+import { viewerSecondsOn, viewerSegmentsOn } from "../viewer/positions.js";
+import { Viewers } from "../viewer/Viewers.js";
+import { OutputCatalog } from "../encode/output/OutputCatalog.js";
+import { ViewerRequests } from "./ViewerRequests.js";
+import { OutputLifecycle } from "./OutputLifecycle.js";
+import { SegmentServing } from "./SegmentServing.js";
+import { audioStartSecondsFor } from "../viewer/audio-start.js";
+import { audioRenditionName } from "../media/audio-inventory.js";
+import { Renditions } from "../encode/Renditions.js";
+import { CushionReport, LOOKAHEAD_PAUSE_SECONDS } from "../encode/CushionReport.js";
+import { EncodeRuns } from "../encode/EncodeRuns.js";
+import { OutputTimes } from "../encode/OutputTimes.js";
+import { HostLoad } from "../encode/quality/HostLoad.js";
+import { HostTimings } from "../encode/quality/HostTimings.js";
+import { BUDGET_CHECK_INTERVAL_MS, QualityController } from "../encode/quality/QualityController.js";
+import { EncodeOrchestrator } from "../encode/EncodeOrchestrator.js";
+import { EncodeAdmission } from "../encode/EncodeAdmission.js";
+import { EncoderSelection } from "../encode/EncoderSelection.js";
+import { OutputOpening } from "../encode/OutputOpening.js";
+import { segmentFormatOfKey } from "../encode/output-key-format.js";
+import { OutputSpec } from "../encode/output/OutputSpec.js";
+import { compareInits } from "../encode/segment-formats/init-compat.js";
+import { nominalKbpsFor } from "../encode/args.js";
+import { probeInputMediaInfo } from "../media/media-info-probe.js";
+import { probeVideoKeyframeTimes } from "../media/keyframe-probe.js";
+import { wireMachineBudget } from "../storage/wire.js";
+import { Returns } from "../storage/returns.js";
+import { freeBytesFor } from "../storage/free.js";
+import path from "node:path";
+import { contentOf, LocalObservations } from "../encode/LocalObservations.js";
+import { configurationKeyOf } from "../encode/fingerprint.js";
+import { PROXY_ROOT } from "../encode/quality/HostTimings.js";
+import { qualityStateOf } from "../encode/quality/OutputQualityState.js";
+
+// Where a variant and an audio rendition live under a session — `v/<height>/…`
+// and `a/<track>/…` — is stated in `encode/output/playlists.js`, beside the lines that
+// write those addresses into a master playlist. The routes that parse them back
+// are in `server.js`.
+
+const CLEANUP_INTERVAL_MS = 30_000;
+const DEFAULT_SEGMENT_DURATION_SEC = 4;
+// Idle TTL: a session is disposed this long after the last segment/playlist
+// access. Long enough that a viewer who pauses, backgrounds the tab, or briefly
+// turns the phone off can resume WITHOUT a cold ffmpeg restart (the warm session
+// also backs the seamless auto-reconnect). ffmpeg stops producing at the
+// look-ahead cap when idle, so a lingering session costs retained segments on
+// disk, not sustained CPU. Active playback refreshes the timer on every segment
+// fetch, so it never expires mid-watch.
+// How long a session outlives the BROWSER, not the viewing. Since server
+// 0.8.103 a browser that holds a session re-asserts it every 30 s, so an open
+// tab never consumes this at all — not while paused, not across a three-hour
+// film. What is left is the case where the browser has genuinely gone: the tab
+// was killed without releasing, the phone slept, the network dropped. Keeping
+// the session means such a viewer comes back to a warm encoder instead of
+// waiting out a cold start; the cost while nobody is there is disk for the
+// produced segments, since the encoder is suspended and burns no CPU, and that
+// disk is already bounded by the pool's 10 GB cap with eviction. Thirty minutes
+// covers a meal, a phone call or a lift ride.
+const DEFAULT_SESSION_TTL_MS = 30 * 60 * 1000;
+const DEFAULT_STARTUP_WAIT_MS = 5_000;
+
+/**
+ * Convert a bind-all host address to the loopback address so that
+ * the HLS input URL is always reachable from the same machine.
+ *
+ * @param {string} host
+ * @returns {string}
+ */
+function toLoopbackHost(host) {
+  if (host === "0.0.0.0" || host === "::") {
+    return "127.0.0.1";
+  }
+  return host;
+}
+
+/**
+ * Build the HTTP base URL (scheme + host + port) for the local proxy server.
+ *
+ * @param {string} host - Bind host (may be "0.0.0.0" or "::").
+ * @param {number} port
+ * @returns {string} e.g. "http://127.0.0.1:9090"
+ */
+function buildHttpBaseUrl(host, port) {
+  const url = new URL("http://localhost");
+  url.hostname = toLoopbackHost(host);
+  url.port = String(port);
+  return url.origin;
+}
+
+/**
+ * @typedef {Object} WireOutputsOptions
+ * @property {boolean} enabled              - Whether HLS transcoding is enabled.
+ * @property {string}  ffmpegBin            - Path to the ffmpeg executable.
+ * @property {string}  localBindHost        - Host the proxy HTTP server is bound to.
+ * @property {number}  localPort            - Port the proxy HTTP server is listening on.
+ * @property {number}  [segmentDurationSec] - HLS segment length in seconds.
+ * @property {number}  [sessionTtlMs]       - Session idle TTL in milliseconds.
+ * @property {number}  [startupWaitMs]      - Max time to wait for the first playlist file.
+ * @property {string}  [segmentFormatId]    - Output container: "fmp4" (default)
+ *   or "mpegts". See `../segment-formats/index.js`.
+ */
+
+/**
+ * @typedef {Object} HlsSession
+ * @property {string}  id            - The name of the output it produces.
+ * @property {string}  fileName      - Display name of the file being transcoded.
+ * @property {import("../media/container/KeyframeTable.js").KeyframeTable} keyframes -
+ *   Where this file's keyframes are. One object per file, held by every session
+ *   of it, so a table read late still reaches them. Used to snap a source seek
+ *   onto a known-valid position (see #startEncodeRun).
+ */
+
+/**
+ * Build the components and connect them.
+ *
+ * @param {WireOutputsOptions} options
+ * @returns {object} The components, and the settings they read.
+ */
+export function wireOutputs({
+    enabled,
+    ffmpegBin,
+    localBindHost,
+    localPort,
+    segmentDurationSec = DEFAULT_SEGMENT_DURATION_SEC,
+    sessionTtlMs = DEFAULT_SESSION_TTL_MS,
+    startupWaitMs = DEFAULT_STARTUP_WAIT_MS,
+    // Where every file's keyframe table lives, and the only thing that reads
+    // one. A session asks for its file's table and is handed the object every
+    // other reader of that file holds, so a table that arrives late reaches the
+    // sessions created before it. Built here when nothing supplied one, which
+    // is a registry with no reader: it answers "not read" about every file, and
+    // every picture is then re-encoded — the correct behaviour for a proxy
+    // wired without a container reader, and what every unit test gets.
+    keyframeTables = new KeyframeTables(),
+    videoEncoder = null,
+    softwarePresetBenchmark = null,
+    // What every encoder this host may use was qualified and measured to do at
+    // startup, by kind (`encode/calibration.js`). Takes the place of the one
+    // array above, which a test may still hand in on its own.
+    calibration = null,
+    decodeCostModel = null,
+    getSourceStats = null,
+    setPriorityMap = null,
+    // What a second job costs on this host, measured at startup. Null when it
+    // could not be measured, and then nothing is corrected — the alternative
+    // is inventing a penalty, which is the same fault as inventing a fill rate.
+    contentionPenalties = null,
+    copySpeedX = null,
+    tonemapSupported = false,
+    getCachedMediaInfo = null,
+    getCachedAudioTracks = null,
+    getContainerMediaInfo = null,
+    fetchWholeFile = null,
+    segmentFormatId = undefined,
+    stateDir = "",
+    segmentStore = null,
+    getTorrentTotals,
+    startStopCost = null,
+    spillDisk = null,
+    wholeFiles = null,
+    diagnostics = null,
+    diagnosticsRoot = "",
+    memoryClaimant = null,
+    budgetPolicy = null}) {
+  const parts = {};
+  // Encode components receive only the storage operations they use. The
+  // concrete SegmentStore remains owned by this assembly and is never handed
+  // to the encode layer.
+  const segmentCoverage = {
+    provenNumbers: (address) => parts.segmentStore.provenNumbers(address),
+    announce: (address, index) => parts.segmentStore.announce(address, index),
+    filesHeld: (address) => parts.segmentStore.filesHeld(address),
+    clearUpAfter: (address, startedAt) => parts.segmentStore.clearUpAfter(address, startedAt)
+  };
+  const segmentFiles = {
+    pathFor: (address) => parts.segmentStore.pathFor(address),
+    initOf: (address) => parts.segmentStore.initOf(address),
+    directoryFor: (address) => parts.segmentStore.directoryFor(address),
+    publish: (address, makingName, format) => parts.segmentStore.publish(address, makingName, format)
+  };
+  const segmentOutputFiles = {
+    addresses: () => parts.segmentStore.addresses(),
+    isClosed: (address, index) => parts.segmentStore.isClosed(address, index),
+    lastReadAt: (address) => parts.segmentStore.lastReadAt(address),
+    directoryFor: (address) => parts.segmentStore.directoryFor(address),
+    useFormat: (address, format) => parts.segmentStore.useFormat(address, format)
+  };
+  const segmentPaths = {
+    pathOf: (address, index) => parts.segmentStore.pathOf(address, index)
+  };
+  /**
+   * Whether everything this output will ever serve is made: every segment of
+   * its timeline closed in the store. Such an output needs no encoder and
+   * holds no place, however many watch it.
+   *
+   * @param {object} session
+   * @returns {boolean}
+   */
+  const finishedOutput = (session) => {
+    const count = Number(session?.timeline?.segmentCount) || 0;
+    return count > 0 && parts.segmentStore.provenNumbers(session.outputKey).length >= count;
+  };
+  // The nominal limits a size may be produced at, highest first. ONE list, read
+  // by the opening of an output and by the move between limits of a height, so
+  // the two cannot disagree about which limits exist. Only the size's own until
+  // a set of lower limits is decided (roadmap item 97, step 14): a lower limit
+  // handed to a viewer must come with the page saying so. Asked with the whole
+  // frame, because the row is chosen by area (`limitRowFor`).
+  parts.limitsFor = (frame) => [nominalKbpsFor(frame)];
+  // What the torrent and the proxy itself spend on this host, and how fast each watched torrent moves.
+  parts.hostLoad = new HostLoad({
+    readMachineState,
+    readProcessCpuSeconds,
+    readProxyCpuSeconds,
+    readSystemCpu,
+    shareOfMachine,
+    liveRunsOf: (...args) => parts.encodeRuns.liveRunsOf(...args),
+    runStateOf: (...args) => parts.encodeRuns.runStateOf(...args),
+    get getSourceStats() { return parts.getSourceStats; },
+    get getTorrentTotals() { return parts.getTorrentTotals; },
+    get outputs() { return parts.outputs; },
+  });
+  // Where each segment of an output begins, and how that table is corrected from what the encoder produced.
+  parts.outputTimes = new OutputTimes({
+    logger,
+    runsOf: (...args) => parts.encodeRuns.runsOf(...args),
+    stopEncodeRun: (...args) => parts.encodeRuns.stopEncodeRun(...args),
+    planEncodersSoon: (...args) => parts.encodeRuns.planEncodersSoon(...args),
+    get outputs() { return parts.outputs; },
+    get segmentDurationSec() { return parts.segmentDurationSec; },
+  });
+  // The encoders of this proxy: built where the plan places them, followed while they run, accounted when they end.
+  parts.encodeRuns = new EncodeRuns({
+    logger,
+    // What an admitted encode was seen to do, filed under its configuration.
+    observeEncodeEnded: (session) => {
+      if (!parts.localObservations || !session?.spec?.video?.encode) {
+        return;
+      }
+      const alone = Number(qualityStateOf(session).lastAloneSpeed);
+      parts.localObservations.noteEncode({
+        spec: session.spec,
+        content: contentOf(session.file),
+        aloneSpeedX: Number.isFinite(alone) && alone > 0 ? alone : null,
+        segmentKbps: segmentRatesOf(session)
+      });
+    },
+    viewerSecondsOn: (output, consumerId, now) => viewerSecondsOn(parts.viewers, output, consumerId, now),
+    noteRunStarting: (output) => parts.quality.noteRunStarting(output),
+    inputOf: (...args) => parts.renditions.inputOf(...args),
+    producedNumbers: (...args) => parts.serving.producedNumbers(...args),
+    servesAudioSeparately: (...args) => parts.renditions.servesAudioSeparately(...args),
+    disposeSession: (...args) => parts.lifecycle.disposeSession(...args),
+    get contentionPenalties() { return parts.contentionPenalties; },
+    get encodeCost() { return parts.encodeCost; },
+    get encodeOrchestrator() { return parts.encodeOrchestrator; },
+    get ffmpegBin() { return parts.ffmpegBin; },
+    get outputTimes() { return parts.outputTimes; },
+    get outputs() { return parts.outputs; },
+    get priority() { return parts.priority; },
+    get segmentDurationSec() { return parts.segmentDurationSec; },
+    segmentFiles,
+    get videoEncoder() { return parts.encoders.current; },
+    get encoders() { return parts.encoders; },
+    invalidateWaits: (output) => parts.serving.invalidateWaits(output),
+    productionFailed: (output) => parts.renditions.noteProductionFailed(output),
+  });
+  // How much film is ready in front of the viewers, said; and the spare soundtracks fetched once it is full.
+  parts.cushion = new CushionReport({
+    viewerSecondsOn: (output, consumerId, now) => viewerSecondsOn(parts.viewers, output, consumerId, now),
+    linkReportsOn: (output) => linkReportsOn(parts.viewers, output),
+    SourceFiles,
+    logger,
+    producedNumbers: (...args) => parts.serving.producedNumbers(...args),
+    get encodeRuns() { return parts.encodeRuns; },
+    get fetchWholeFile() { return parts.fetchWholeFile; },
+    get getCachedAudioTracks() { return parts.getCachedAudioTracks; },
+    get hostLoad() { return parts.hostLoad; },
+    get lookaheadSeconds() { return parts.lookaheadSeconds; },
+    get outputTimes() { return parts.outputTimes; },
+    get outputs() { return parts.outputs; },
+  });
+  // The steps of a picture and its soundtracks: which output answers each, made when first asked for, and the master playlist listing them.
+  // The cushion this file needs, from its segment length and the worst
+  // interruption its supply has shown — one statement read by every operation
+  // that asks it.
+  const minimumBufferSecondsOf = (output) => minimumBufferFrom({
+    segmentSeconds: parts.segmentDurationSec,
+    worstSupplyWaitSec: parts.hostLoad.supplyFor(output.file)?.worstWaitSec
+  })?.seconds ?? null;
+  parts.renditions = new Renditions({
+    // Whether this proxy re-encodes at all: a soundtrack with no stated rate is
+    // re-encoded where it may be, and copied where it may not.
+    get transcodeEnabled() { return parts.enabled; },
+    observedPeakMbps: (spec) => parts.localObservations?.peakMbps(spec) ?? null,
+    // A viewer has been moved onto an output prepared for them: how long it
+    // took, and what they held.
+    notePreparation: (output, seconds, bufferedSec) =>
+      parts.localObservations?.notePreparation({ spec: output?.spec, seconds, bufferedSec }),
+    viewerSecondsOn: (output, consumerId, now) => viewerSecondsOn(parts.viewers, output, consumerId, now),
+    audioStartSecondsFor: (args) => audioStartSecondsFor({ ...args, viewers: parts.viewers }),
+    activeOutputFor: (args) => activeOutputFor({ ...args, viewers: parts.viewers }),
+    // WHAT ENCODING IS TOLD ABOUT A PERSON, and what it may tell the viewer
+    // layer about them: values and names, never the viewer itself. Steps and
+    // soundtracks are chosen, warmed and left per person, so this is the widest
+    // of the three sets — and every one of them is a call on the owner, which
+    // is what keeps one writer per fact.
+    viewerCountOn: (output) => parts.viewers.forOutput(output).size,
+    // Whether anything still stands on it — somebody watching, or an
+    // assignment not yet released. See `Viewers.stillNeeded`.
+    outputStillNeeded: (output) => parts.viewers.stillNeeded(output),
+    consumersOn: (output) => consumersOn(parts.viewers, output),
+    presentOn: (output) => presentOn(parts.viewers, output),
+    qualityModeOf: (output, consumerId) => qualityModeOf(parts.viewers, output, consumerId),
+    linkMbpsOf: (output, consumerId) => linkMbpsOf(parts.viewers, output, consumerId),
+    audioChoiceOf: (output, consumerId) => audioChoiceOf(parts.viewers, output, consumerId),
+    chooseAudioTrack: (output, consumerId, choice) =>
+      chooseAudioTrack(parts.viewers, output, consumerId, choice),
+    stepOnScreenOf: (output, consumerId) => stepOnScreenOf(parts.viewers, output, consumerId),
+    noteStepOnScreen: (output, consumerId, stepId) =>
+      noteStepOnScreen(parts.viewers, output, consumerId, stepId),
+    stepBeingWarmedOf: (output, consumerId) => stepBeingWarmedOf(parts.viewers, output, consumerId),
+    stepBeingWarmedSinceOf: (output, consumerId) => stepBeingWarmedSinceOf(parts.viewers, output, consumerId),
+    noteStepBeingWarmed: (output, consumerId, stepId) =>
+      noteStepBeingWarmed(parts.viewers, output, consumerId, stepId),
+    audioBeingWarmedOf: (output, consumerId) => audioBeingWarmedOf(parts.viewers, output, consumerId),
+    noteAudioBeingWarmed: (output, consumerId, renditionId) =>
+      noteAudioBeingWarmed(parts.viewers, output, consumerId, renditionId),
+    watches: (output, consumerId) => watches(parts.viewers, output, consumerId),
+    placeOn: (output, consumerId, seconds) => placeOn(parts.viewers, output, consumerId, seconds),
+    // A viewer's own assignments: which output answered a height and segment in
+    // one generation of their viewing, so a repeat is answered the same way.
+    // Values in and out, like everything above.
+    generationOfRequest: (consumerId, stated) => generationOfRequest(parts.viewers, consumerId, stated),
+    givenOutputOf: (consumerId, generation, askedHeight, segmentIndex) =>
+      givenOutputOf(parts.viewers, consumerId, generation, askedHeight, segmentIndex),
+    noteGivenOutput: (consumerId, generation, askedHeight, segmentIndex, outputKey) =>
+      noteGivenOutput(parts.viewers, consumerId, generation, askedHeight, segmentIndex, outputKey),
+    // The output this viewer was chosen at a height, and the rule's choice of
+    // it — the viewer's own record, never one per file.
+    chosenOutputOf: (consumerId, askedHeight) => chosenOutputOf(parts.viewers, consumerId, askedHeight),
+    chooseOutput: (consumerId, askedHeight, outputKey) => chooseOutput(parts.viewers, consumerId, askedHeight, outputKey),
+    noteServingVerdict: (consumerId, verdict) => noteServingVerdict(parts.viewers, consumerId, verdict),
+    // A move between two limits of the height on a viewer's screen (roadmap
+    // item 97, step 12): which heights their choice is the output on screen
+    // under, the segment they were last given, the move being prepared, and
+    // whether a segment is closed on the output being prepared.
+    limitsFor: (frame) => parts.limitsFor(frame),
+    heightsChosenAs: (consumerId, outputKey) => heightsChosenAs(parts.viewers, consumerId, outputKey),
+    highestGivenSegmentOf: (consumerId, askedHeight) => highestGivenSegmentOf(parts.viewers, consumerId, askedHeight),
+    sameHeightSwitchOf: (consumerId) => sameHeightSwitchOf(parts.viewers, consumerId),
+    noteSameHeightSwitch: (consumerId, value) => noteSameHeightSwitch(parts.viewers, consumerId, value),
+    switchingOnto: (output) => switchingOnto(parts.viewers, output),
+    // What a viewer holds, and the cushion this file needs before a move that
+    // is not urgent is made (roadmap item 98).
+    bufferedSecondsOf: (consumerId) => bufferedSecondsOf(parts.viewers, consumerId),
+    minimumBufferSecondsFor: (output) => minimumBufferSecondsOf(output),
+    // Whether the whole machine has a place for one more encoder on this output,
+    // asked before a preparation is recorded.
+    admitsPreparation: (output) => parts.admission.admitsPreparation(output?.outputKey ?? ""),
+    segmentClosed: (key, index) => parts.segmentStore.isClosed(key, index),
+    // What a gone output left in the store, and whether another output's header
+    // may stand in for its own.
+    storedPieceReady: (key, fileName) => parts.serving.hasStoredPiece(key, fileName),
+    headersCompatible: (goneKey, liveKey) => parts.serving.headersCompatible(goneKey, liveKey),
+    placeViewerOn: (...args) => parts.viewerRequests.placeViewerOn(...args),
+    audioRenditionName,
+    logger,
+    viewerLeaves: (...args) => parts.lifecycle.viewerLeaves(...args),
+    createOrGetSession: (...args) => parts.viewerRequests.createOrGetSession(...args),
+    planEncodersSoon: (...args) => parts.encodeRuns.planEncodersSoon(...args),
+    disposeSession: (...args) => parts.lifecycle.disposeSession(...args),
+    viewerPositionOf: (...args) => parts.viewerRequests.viewerPositionOf(...args),
+    get encodeRuns() { return parts.encodeRuns; },
+    get fileStartTimeReads() { return parts.fileStartTimeReads; },
+    set fileStartTimeReads(value) { parts.fileStartTimeReads = value; },
+    get getCachedAudioTracks() { return parts.getCachedAudioTracks; },
+    get getCachedMediaInfo() { return parts.getCachedMediaInfo; },
+    get getContainerMediaInfo() { return parts.getContainerMediaInfo; },
+    get localBaseUrl() { return parts.localBaseUrl; },
+    get outputTimes() { return parts.outputTimes; },
+    get outputs() { return parts.outputs; },
+    get quality() { return parts.quality; },
+    get qualityOffer() { return parts.qualityOffer; },
+    get segmentDurationSec() { return parts.segmentDurationSec; },
+    get sourceFiles() { return parts.sourceFiles; },
+    invalidateWaits: (output) => parts.serving.invalidateWaits(output),
+  });
+  // Answering a request for a playlist, an init segment or a segment: from the store when the piece is made and whole, otherwise held until it is.
+  parts.serving = new SegmentServing({
+    // What a key names, read back from the key: how to find a gone output's
+    // stored pieces and whether they are cut at given times.
+    segmentFormatOfKey,
+    specOfKey: (key) => OutputSpec.fromKey(key),
+    compareInits,
+    activeOutputFor: (args) => activeOutputFor({ ...args, viewers: parts.viewers }),
+    viewerSecondsOn: (output, consumerId, now) => viewerSecondsOn(parts.viewers, output, consumerId, now),
+    buildMasterPlaylist: (...args) => parts.renditions.buildMasterPlaylist(...args),
+    declaredTracks: (...args) => parts.renditions.declaredTracks(...args),
+    publishedGridFor: (...args) => parts.outputTimes.publishedGridFor(...args),
+    runStartTimeFor: (...args) => parts.outputTimes.runStartTimeFor(...args),
+    get cushion() { return parts.cushion; },
+    get encodeOrchestrator() { return parts.encodeOrchestrator; },
+    get encodeRuns() { return parts.encodeRuns; },
+    get hostTimings() { return parts.hostTimings; },
+    get lookaheadSeconds() { return parts.lookaheadSeconds; },
+    get outputTimes() { return parts.outputTimes; },
+    get outputs() { return parts.outputs; },
+    get segmentStore() { return parts.segmentStore; },
+    get startupWaitMs() { return parts.startupWaitMs; },
+    get viewers() { return parts.viewers; },
+  });
+  // How an output ends: disposed when nobody is left on it and it has been idle, or all at once on shutdown; and the segments an earlier process left behind adopted at startup.
+  parts.lifecycle = new OutputLifecycle({
+    invalidateWaits: (output) => parts.serving.invalidateWaits(output),
+    segmentFormatOfKey,
+    // Where the viewers of one output stand, in its segments: what the store
+    // reads to give disk back in the right order.
+    viewerSegmentsOn: (key) => viewerSegmentsOn({
+      outputs: parts.outputs.values(),
+      outputKey: key,
+      segmentAt: (output, seconds) => parts.outputTimes.segmentIndexForTime(output, seconds),
+      now: Date.now(),
+      viewers: parts.viewers
+    }),
+    get budgetTimer() { return parts.budgetTimer; },
+    get cleanupTimer() { return parts.cleanupTimer; },
+    get encodeRuns() { return parts.encodeRuns; },
+    get keyframeTables() { return parts.keyframeTables; },
+    get machineBudget() { return parts.machineBudget; },
+    get outputTimes() { return parts.outputTimes; },
+    get outputs() { return parts.outputs; },
+    get returns() { return parts.returns; },
+    get segmentStore() { return parts.segmentStore; },
+    get sessionTtlMs() { return parts.sessionTtlMs; },
+    get sourceFiles() { return parts.sourceFiles; },
+    get timelines() { return parts.timelines; },
+    get viewers() { return parts.viewers; },
+  });
+  // Which output answers a request for a file, made if it is not here yet: the encoding component's decision, made without knowing who asked.
+  parts.opening = new OutputOpening({
+    logger,
+    // The largest peak an output of this mode and rate control has been seen
+    // carrying here: the only figure an encoder with no bound of its own has.
+    observedPeakMbps: (spec) => parts.localObservations?.peakMbps(spec) ?? null,
+    limitsFor: (frame) => parts.limitsFor(frame),
+    probeMediaInfo: (url) => probeInputMediaInfo(parts.ffmpegBin, url),
+    probeKeyframeTimes: (url, budgetMs) => probeVideoKeyframeTimes(parts.ffmpegBin, url, budgetMs),
+    get decodeCostModel() { return parts.decodeCostModel; },
+    get enabled() { return parts.enabled; },
+    get encodeCost() { return parts.encodeCost; },
+    get encodeRuns() { return parts.encodeRuns; },
+    get admission() { return parts.admission; },
+    get getCachedMediaInfo() { return parts.getCachedMediaInfo; },
+    get hostLoad() { return parts.hostLoad; },
+    get hostTimings() { return parts.hostTimings; },
+    get keyframeTables() { return parts.keyframeTables; },
+    get localBaseUrl() { return parts.localBaseUrl; },
+    get outputs() { return parts.outputs; },
+    get renditions() { return parts.renditions; },
+    get returns() { return parts.returns; },
+    get segmentDurationSec() { return parts.segmentDurationSec; },
+    get segmentFormat() { return parts.segmentFormat; },
+    segmentOutputFiles,
+    get softwarePresetBenchmark() { return parts.softwarePresetBenchmark; },
+    get sourceFiles() { return parts.sourceFiles; },
+    get timelines() { return parts.timelines; },
+    get tonemapSupported() { return parts.tonemapSupported; },
+    get videoEncoder() { return parts.encoders.current; },
+  });
+  // What a viewer asks for: an output of a file, a position, a report of progress.
+  parts.viewerRequests = new ViewerRequests({
+    // Whether this machine has a place for one more viewer on this output —
+    // asked in the same synchronous stretch as they are put on it.
+    admitsWatching: (output) => parts.admission.admitsWatching(output?.outputKey ?? ""),
+    activeOutputFor: (args) => activeOutputFor({ ...args, viewers: parts.viewers }),
+    viewerSecondsOn: (output, consumerId, now) => viewerSecondsOn(parts.viewers, output, consumerId, now),
+    minimumBufferSecondsFor: (output) => minimumBufferSecondsOf(output),
+    getSourceStats: (...args) => parts.getSourceStats?.(...args) ?? null,
+    disposeSession: (...args) => parts.lifecycle.disposeSession(...args),
+    expectedFirstSegmentMs: (...args) => parts.hostTimings.expectedFirstSegmentMs(...args),
+    expectedSessionCreateMs: (...args) => parts.hostTimings.expectedSessionCreateMs(...args),
+    planEncodersSoon: (...args) => parts.encodeRuns.planEncodersSoon(...args),
+    waitUntilReady: (...args) => parts.serving.waitUntilReady(...args),
+    get encodeRuns() { return parts.encodeRuns; },
+    encodeSpeedReadingOf: (output) => parts.encodeCost.latestSpeedReadingOf(output),
+    projectedEncodeSpeedOf: () => parts.admission.projectedSpeedX(),
+    get opening() { return parts.opening; },
+    get outputTimes() { return parts.outputTimes; },
+    get outputs() { return parts.outputs; },
+    get lookaheadSeconds() { return parts.lookaheadSeconds; },
+    get quality() { return parts.quality; },
+    get qualityOffer() { return parts.qualityOffer; },
+    get renditions() { return parts.renditions; },
+    get segmentDurationSec() { return parts.segmentDurationSec; },
+    get segmentStore() { return parts.segmentStore; },
+    get sourceFiles() { return parts.sourceFiles; },
+    get startupWaitMs() { return parts.startupWaitMs; },
+    get viewers() { return parts.viewers; },
+    invalidateWaits: (output) => parts.serving.invalidateWaits(output),
+    // Whether a request of this generation is still taken, what it belongs to,
+    // what answered it, and the hold a response keeps while it is being sent.
+    acceptsGeneration: (consumerId, stated) => acceptsGeneration(parts.viewers, consumerId, stated),
+    generationOfRequest: (consumerId, stated) => generationOfRequest(parts.viewers, consumerId, stated),
+    noteGivenOutput: (consumerId, generation, askedHeight, segmentIndex, outputKey) =>
+      noteGivenOutput(parts.viewers, consumerId, generation, askedHeight, segmentIndex, outputKey),
+    holdForResponse: (consumerId, outputKey) => holdForResponse(parts.viewers, consumerId, outputKey),
+    noteServingVerdict: (consumerId, verdict) => noteServingVerdict(parts.viewers, consumerId, verdict),
+    servingVerdictOf: (consumerId) => servingVerdictOf(parts.viewers, consumerId),
+  });
+  // What this host takes to create an output and to produce its first segment, kept across restarts until the synthetic figure can replace it (CLAUDE.md, host timings).
+  parts.hostTimings = new HostTimings({
+    get segmentDurationSec() { return parts.segmentDurationSec; },
+    get softwarePresetBenchmark() { return parts.softwarePresetBenchmark; },
+    get stateDir() { return parts.stateDir; },
+  });
+  parts.enabled = Boolean(enabled);
+  parts.ffmpegBin = ffmpegBin;
+  parts.keyframeTables = keyframeTables;
+  // Where measurements about this host are kept between runs. Empty means
+  // beside the installed proxy; a deployment with somewhere persistent to
+  // write names it (--state-dir).
+  parts.stateDir = typeof stateDir === "string" ? stateDir : "";
+  // Output container (fMP4/CMAF or MPEG-TS). Everything container-specific —
+  // muxer args, file naming, playlist header, per-segment correction — lives
+  // in this module; nothing here branches on the format.
+  parts.segmentFormat = resolveSegmentFormat(segmentFormatId);
+  // Optional accessor for media info the playback planner already probed for
+  // (sourceKey, fileIndex), so session create can skip its own ffmpeg scan.
+  parts.getCachedMediaInfo = typeof getCachedMediaInfo === "function" ? getCachedMediaInfo : null;
+  // The file's audio tracks, for the master playlist's rendition group. Same
+  // inventory the browser's audio menu is built from.
+  parts.getCachedAudioTracks = typeof getCachedAudioTracks === "function" ? getCachedAudioTracks : null;
+  // What a file declares about itself, read by the container layer from the
+  // same header its track table comes from: format, duration, and where its
+  // own timeline begins. The last of those is why this exists — a soundtrack
+  // shipped as its own file has a timeline of its own, and asking ffmpeg for
+  // it meant reading a header this proxy had already read.
+  parts.getContainerMediaInfo =
+    typeof getContainerMediaInfo === "function" ? getContainerMediaInfo : null;
+  // Fetch one whole file of a source, as a bounded read rather than a
+  // selection. Used to pull a soundtrack that ships beside the picture onto
+  // the disk while the swarm has capacity to spare — see
+  // `#fetchSpareSoundtracks`. Optional: a proxy wired without it simply reads
+  // such a soundtrack when it is played.
+  parts.fetchWholeFile = typeof fetchWholeFile === "function" ? fetchWholeFile : null;
+  // Optional async accessor for a source's live download stats, used by the
+  // realtime budget to tell a CPU limit from a download-starved input:
+  // (sourceKey, fileIndex) => Promise<{ downloadSpeed, fileLength, fileProgress } | null>.
+  parts.getSourceStats = typeof getSourceStats === "function" ? getSourceStats : null;
+  parts.setPriorityMap = typeof setPriorityMap === "function" ? setPriorityMap : null;
+  parts.contentionPenalties = contentionPenalties instanceof Map ? contentionPenalties : null;
+  // Seconds of film per second when the picture is COPIED, measured at
+  // startup. Nothing else prices that branch: the other benchmarks measure
+  // encoding and decoding, and a copy does neither.
+  parts.copySpeedX = Number.isFinite(copySpeedX) && copySpeedX > 0 ? copySpeedX : null;
+  // Totals across every torrent this proxy holds, used to price what the
+  // torrent itself costs the machine (item 7). Optional: a proxy wired
+  // without it simply never learns that figure.
+  parts.getTorrentTotals = typeof getTorrentTotals === "function" ? getTorrentTotals : null;
+  // The spilled pieces, as a pair of closures over the torrent thread: what
+  // they weigh and how to tell them their share. Not the pool — the pool is
+  // on the other side of the thread boundary and this side holds none of it.
+  parts.spillDisk = spillDisk;
+  parts.wholeFiles = wholeFiles;
+  parts.diagnostics = diagnostics;
+  parts.diagnosticsRoot = diagnosticsRoot;
+  parts.memoryClaimant = memoryClaimant;
+  parts.budgetPolicy = budgetPolicy;
+  // Detected H.264 encoder descriptor (hardware or software). Defaults to
+  // software libx264 when no detection result is supplied. May be downgraded
+  // to software at runtime if a hardware encode fails.
+  parts.encoders = new EncoderSelection({ detected: videoEncoder, softwareDescriptor });
+  // WHAT THIS HOST HAS SEEN ITS OWN ADMITTED ENCODES DO, kept between runs and
+  // used only on the configuration it was seen on (roadmap item 97, step 14).
+  // A refinement of figures that exist without it: nothing is selected,
+  // admitted or served differently when it is empty. Absent where the startup
+  // calibration was not run, which is only a wiring made without it.
+  parts.localObservations = null;
+  if (calibration?.fingerprint) {
+    parts.localObservations = new LocalObservations({
+      fingerprint: calibration.fingerprint,
+      filePath: path.join(typeof stateDir === "string" && stateDir.length > 0 ? stateDir : PROXY_ROOT, "local-observations.json"),
+      logger
+    });
+    parts.localObservations.load(calibration.kinds().map((kind) => configurationKeyOf(calibration.fingerprint, {
+      kind,
+      name: kind === "software" ? softwareDescriptor().name : (videoEncoder?.kind === kind ? videoEncoder.name : kind)
+    })));
+  }
+  /**
+   * What the segments an output has closed carried, from their sizes and the
+   * spans its timeline gives them: the average over all of them and the
+   * largest one.
+   *
+   * @param {object} session
+   * @returns {{ averageKbps: number, peakKbps: number } | null}
+   */
+  const segmentRatesOf = (session) => {
+    const sizes = parts.segmentStore.sizesOf(session.outputKey);
+    const timeline = session.timeline;
+    if (!timeline || sizes.size === 0) {
+      return null;
+    }
+    let bits = 0;
+    let seconds = 0;
+    let peakKbps = 0;
+    for (const [index, size] of sizes) {
+      const start = timeline.publishedStartOf(index);
+      const end = index < timeline.segmentCount ? timeline.publishedStartOf(index + 1) : null;
+      if (!(Number.isFinite(start) && Number.isFinite(end) && end > start) || !(size > 0)) {
+        continue;
+      }
+      bits += size * 8;
+      seconds += end - start;
+      peakKbps = Math.max(peakKbps, (size * 8) / (end - start) / 1000);
+    }
+    return seconds > 0 ? { averageKbps: bits / seconds / 1000, peakKbps } : null;
+  };
+  // The modes of the encoder IN USE NOW, as measured at startup: slowest
+  // first, each with its throughput by size. Read through the selection rather
+  // than copied, because the encoder in use can change — a failing device
+  // moves this host to software for the rest of the process, and from then on
+  // it is software's own modes that price and choose, not the device's.
+  // Null where nothing was measured, and then nothing is re-encoded here.
+  parts.calibration = calibration ?? null;
+  const givenModes = Array.isArray(softwarePresetBenchmark) ? softwarePresetBenchmark : null;
+  Object.defineProperty(parts, "softwarePresetBenchmark", {
+    get: () => (parts.calibration
+      ? parts.calibration.modesFor(parts.encoders.current?.kind)
+      : givenModes),
+    enumerable: true
+  });
+  // Host decode cost solved at startup from the calibration clips:
+  // `a × Mpixel/s + b × Mbit/s + c` seconds of decoding per second of video.
+  // A re-encode pays for this as well as for the encoder, and leaving it out
+  // is what made the budget offer rungs this host ran at a third of realtime.
+  // Null when the clips are missing or the fit was rejected — the budget then
+  // prices the encoder alone, as it did before.
+  parts.decodeCostModel = decodeCostModel ?? null;
+  // Whether this ffmpeg build can tone-map HDR→SDR (zscale + tonemap filters).
+  // Gates the tonemap chain for HDR sources on the software path.
+  parts.tonemapSupported = Boolean(tonemapSupported);
+  parts.segmentDurationSec = segmentDurationSec;
+  // How far ahead of the viewer this proxy lets an encoder run, in seconds of
+  // playback. Stated rather than kept private, because the browser's forward
+  // buffer is bounded by the same quantity and used to carry a copy of its
+  // own: a hand-written 30 s, justified in a comment by a DIFFERENT constant
+  // (the eight segments that bound a request ahead of the ENCODE HEAD), so
+  // three quarters of what the encoder had already produced was thrown away.
+  // One figure, said by the side that owns it (roadmap item 4).
+  parts.lookaheadSeconds = LOOKAHEAD_PAUSE_SECONDS;
+  parts.sessionTtlMs = sessionTtlMs;
+  parts.startupWaitMs = startupWaitMs;
+  parts.localBaseUrl = buildHttpBaseUrl(localBindHost, localPort);
+  // Everyone watching anything, one object per person rather than one per
+  // person per output. What a viewer chose, where they are and which outputs
+  // they are watching are facts about the person; kept per output they were
+  // three copies of which two were always stale.
+  // Every change to who is watching what re-decides which encoders should
+  // exist, because that decision reads nothing else about viewers. It used to
+  // be re-taken on a five-second timer instead, which made a just-created
+  // output wait up to five seconds before anything noticed it had a viewer.
+  parts.viewers = new Viewers({ onChange: () => parts.encodeRuns.planEncodersSoon() });
+  // The outputs that exist, and every question about them: the picture a step
+  // belongs to, the steps, the soundtracks, the height a step is named by.
+  parts.outputs = new OutputCatalog({
+    fileLengthOf: (session) => parts.hostLoad.fileLengthByKey.get(session.file.key) ?? 0,
+    largestPieceOf: (address) => parts.segmentStore.largestPiece(address)
+  });
+  // What this host learned last time it ran. Without it every restart shows
+  // the first viewer a figure with no measurement behind it.
+  parts.hostTimings.loadHostTimings();
+  // Where produced segments live, addressed by WHAT they are rather than by
+  // which session's encoder wrote them. Two sessions of one output — two
+  // viewers who opened the same film at different places — write into one
+  // directory and each serves what the other has already made. Injectable so
+  // a test can give it a root of its own.
+  parts.segmentStore = segmentStore instanceof SegmentStore
+    ? segmentStore
+    : new SegmentStore({ logger });
+  // The store has grown: whether it still fits its share is asked now rather
+  // than on the next cleanup pass.
+  parts.segmentStore.onPublished(() => parts.lifecycle.keepWithinRoomSoon());
+  // A segment closed may be the one a viewer's move between two limits of a
+  // height is waiting for.
+  parts.segmentStore.onPublished((key, index) => parts.renditions.noteSegmentPublished(key, index));
+  // What encoders there should be on each output, and where. Given the two
+  // things only this class can answer: how many this machine can afford, and
+  // how to make one.
+  // What encoding costs this machine, and which heights follow from it. Given
+  // what it cannot work out for itself: which sessions belong to one file, the
+  // host's own readings AT THE MOMENT OF THE QUESTION rather than copied, how
+  // a soundtrack is keyed, how many encoders are running, and what a file
+  // costs merely by being fetched.
+  parts.encodeCost = new EncodeCost({
+    outputs: parts.outputs,
+    host: () => ({
+      benchmark: parts.softwarePresetBenchmark,
+      decodeModel: parts.decodeCostModel,
+      contentionPenalties: parts.contentionPenalties,
+      copySpeedX: parts.copySpeedX,
+      availability: parts.hostLoad.hostAvailability,
+      // WHICH encoder this host settled on. Only the software ladder is
+      // benchmarked, so a reading taken off a hardware encoder cannot be
+      // split into its decode and encode halves and is not filed as one.
+      encoderKind: parts.encoders.current?.kind ?? null,
+      // The slowest this output's mode has been seen running alone on
+      // comparable material on this configuration, or null.
+      observedAloneSpeed: (session) => parts.localObservations?.slowestAloneSpeed(session.spec, contentOf(session.file)) ?? null
+    }),
+    // HOW MANY ENCODER PROCESSES ARE RUNNING, asked of the one thing that
+    // makes and unmakes them. This used to be two counts of one fact, taken
+    // by walking the session registry — one per session, one per run — and
+    // they agreed only while a session held at most one run.
+    runningEncoders: () => parts.encodeOrchestrator.runningCount(),
+    encodersRunningNow: () => parts.encodeOrchestrator.runningCount(),
+    torrentCostSecFor: (session) => parts.hostLoad.torrentCostSecFor(session),
+    boundBy: (session) => parts.quality.classifyTranscodeBound(session),
+    runsFor: (session) => parts.encodeRuns.runsOf(session),
+    stateFor: (session) => parts.encodeRuns.runStateOf(session),
+    // The run's OWN progress: a speed is a pair of readings of one process.
+    progressFor: (_session, run) => run?.progress?.snapshot?.() ?? null
+  });
+  // WHICH HEIGHTS ARE ON THE MENU, which is the arithmetic above plus three
+  // things that are nothing to do with it: whose answer it is, what may never
+  // be withdrawn, and when the answer may be reused.
+  parts.qualityOffer = new QualityOffer({
+    encodeCost: parts.encodeCost,
+    occupiedCostSec: (fileKey) => parts.admission?.occupiedCost({ exceptFileKey: fileKey }).costSec ?? 0,
+    occupancyKnownFor: (fileKey) => (parts.admission?.occupiedCost({ exceptFileKey: fileKey }).unpriced ?? 0) === 0,
+    outputs: parts.outputs,
+    stateFor: (session) => parts.encodeRuns.runStateOf(session),
+    // WHICH HEIGHTS A LIVE VIEWER HAS ON SCREEN, as numbers. Who is watching
+    // what is the viewer layer's, and which session is which height is the
+    // film's shape; neither travels — what crosses is the list of heights.
+    heightsOnScreen: (owner) => parts.quality.heightsOnScreen(owner),
+    // WHAT THE SWARM IS DOING WITH THIS FILE. Three readings, taken by
+    // whoever reads it and handed over as three numbers: the speed this
+    // file's own interruptions demand, the megabytes a second a viewer draws
+    // through it, and what a megabyte costs this process.
+    supplyFor: (file) => ({
+      requiredSpeed: parts.hostLoad.requiredSpeedFor(file.sourceKey, file.fileIndex),
+      megabytesPerSecond: parts.hostLoad.torrentMegabytesPerSecond(
+        file.sourceKey,
+        file.fileIndex,
+        file.lengthBytes ?? parts.hostLoad.fileLengthByKey.get(file.key) ?? null,
+        file.durationSeconds
+      ),
+      costPerMegabyte: parts.hostLoad.observedTorrentCostPerMegabyte
+    })
+  });
+  // The priority map, built from where the viewers are and handed to both
+  // sides that act on it. The downloading lives in another thread, so its
+  // copy travels over the worker channel.
+  parts.priority = new PriorityOrchestrator({
+    publish: ({ sourceKey, fileIndex, durationSeconds, zones }) => {
+      void Promise.resolve(
+        parts.setPriorityMap?.({ sourceKey, fileIndex, durationSeconds, zones })
+      ).catch(() => {});
+    },
+    get viewers() { return parts.viewers; },
+    // WHERE THE TWO FACTS MEET, and this is the only place that holds both.
+    // Which step is on somebody's screen belongs to the person; which output
+    // a step supersedes belongs to the film's shape. Neither layer is handed
+    // the other — one gets a plain id, the other is read for one field.
+    watchedBy: (session, viewer) => !parts.outputs.supersededBy(session, viewer.activeVariantId ?? null),
+    allowanceFor: (session) => minimumBufferFrom({
+      segmentSeconds: parts.segmentDurationSec,
+      worstSupplyWaitSec: parts.hostLoad.supplyFor(session.file)?.worstWaitSec
+    })?.seconds ?? parts.segmentDurationSec
+  });
+  // WHETHER THE WHOLE MACHINE HAS A PLACE for one more encoder, across outputs.
+  // Given what it cannot hold itself: what runs (the orchestrator's), what is
+  // being prepared (the viewers' records), what an output costs (the encoding
+  // cost) and how much of the machine nobody has priced (the host's reading).
+  parts.admission = new EncodeAdmission({
+    liveRunsByAddress: () => parts.encodeOrchestrator.liveRunsByAddress(),
+    preparedAddresses: () => {
+      const addresses = new Set();
+      for (const id of outputsBeingPrepared(parts.viewers)) {
+        const key = parts.outputs.get(id)?.outputKey;
+        if (key) {
+          addresses.add(key);
+        }
+      }
+      return addresses;
+    },
+    // Every output a present viewer RECEIVES — by the same rule the per-output
+    // priority map is built by — that still has something left to make. The
+    // place an output is opened on is held from the moment its viewer is put
+    // on it, not from when the plan gets round to its first encoder.
+    watchedAddresses: () => {
+      const addresses = new Set();
+      for (const session of parts.outputs.values()) {
+        const address = session.outputKey;
+        if (!address || addresses.has(address) || finishedOutput(session)) {
+          continue;
+        }
+        for (const viewer of parts.viewers.forOutput(session).values()) {
+          if (viewer.isPresent() && !parts.outputs.supersededBy(session, viewer.activeVariantId ?? null)) {
+            addresses.add(address);
+            break;
+          }
+        }
+      }
+      return addresses;
+    },
+    finished: (address) => parts.outputs.outputsOn(address).some((session) => finishedOutput(session)),
+    loadOf: (address) => parts.encodeCost.loadOfOutput(address),
+    loadForCandidate: (spec, file) => parts.encodeCost.loadForCandidate(spec, file),
+    availability: () => parts.hostLoad.hostAvailability ?? null
+  });
+  parts.encodeOrchestrator = new EncodeOrchestrator({
+    admission: parts.admission,
+    // What the viewers actually waited for, by band. The ledger is the
+    // priority layer's; the encoding is handed a way to ask it.
+    describeWaits: (address) => waits.describe(address),
+    maxRunsFor: (address) => parts.encodeRuns.maxRunsForOutput(address),
+    makeRun: ({ address, from, to, because }) => parts.encodeRuns.makeRunAt(address, from, to, because),
+    segmentSeconds: parts.segmentDurationSec,
+    contentionPenalties: parts.contentionPenalties,
+    startingSpeedFor: (address) => parts.encodeCost.speedForOutput(address),
+    segmentCoverage,
+    // HOW IT ASKS TO DECIDE AGAIN. A plan that refuses to place anything
+    // because an output's input is away needs something to bring it back:
+    // nothing about the state changes while the data is missing, so no event
+    // arrives on its own.
+    planSoon: () => parts.encodeRuns.planEncodersSoon(),
+    logger
+  });
+  // What a start and a stop were measured to cost here, before any viewer
+  // existed. Without it both read zero at a cold open, and zero is not
+  // "unmeasured" — it is "free", which is what moved an encoder between two
+  // adjacent numbers every half second in the field.
+  parts.encodeOrchestrator.noteStartupCosts(startStopCost);
+  // Where each file is cut, held once per file and grid rather than once per
+  // session. Two sessions of one film MUST agree about this to the
+  // millisecond — a segment made by either has to be appendable where the
+  // other's would have gone — and until now they agreed by copying, which is
+  // a thing somebody has to remember to do and which drifted twice in the
+  // field. They share the table now.
+  parts.timelines = new Timelines();
+  // The files this proxy is serving, one object per file however many
+  // sessions are of it. It holds the file's key — which every cache about a
+  // file is keyed by — its name, and the facts a probe returned.
+  parts.sourceFiles = new SourceFiles();
+  parts.cleanupTimer = setInterval(() => {
+    void parts.lifecycle.cleanupExpired();
+  }, CLEANUP_INTERVAL_MS);
+  parts.cleanupTimer.unref();
+  // How long after material stops being read somebody asks for it again — the
+  // one term of the keeping period that is guessed rather than measured, and
+  // the only place it can be measured from.
+  parts.returns = new Returns();
+  // One owner of the disk, and the list of what takes it lives with the owner.
+  parts.machineBudget = wireMachineBudget({
+    segmentStore: parts.segmentStore,
+    spill: parts.spillDisk,
+    wholeFiles: parts.wholeFiles,
+    diagnostics: parts.diagnostics,
+    diagnosticsRoot: parts.diagnosticsRoot,
+    memory: parts.memoryClaimant,
+    policy: parts.budgetPolicy ?? undefined,
+    readFree: freeBytesFor,
+    logger
+  });
+  // Realtime-budget monitor: only meaningful for the software encoder with a
+  // benchmark (the only path that can pick/step resolution). Cheap no-op scan
+  // otherwise.
+  // The quality budget: what each output is asked to step to, and the bitrate ceiling a viewer's link sets. Everything it reads of the rest of the proxy is listed here.
+  parts.quality = new QualityController({
+    // The longest an output of this mode has been seen taking to be ready for
+    // a viewer moving onto it, in milliseconds, or null.
+    observedPreparationMs: (output) => {
+      const seconds = parts.localObservations?.longestPreparationSec(output?.spec) ?? null;
+      return seconds === null ? null : seconds * 1000;
+    },
+    // WHAT ENCODING IS TOLD ABOUT A PERSON: values, by name. Not the viewer,
+    // whose record belongs to the viewer layer and is written only there.
+    consumersOn: (output) => consumersOn(parts.viewers, output),
+    qualityModeOf: (output, consumerId) => qualityModeOf(parts.viewers, output, consumerId),
+    stepOnScreenOf: (output, consumerId) => stepOnScreenOf(parts.viewers, output, consumerId),
+    standingAskOf: (output, consumerId) => standingAskOf(parts.viewers, output, consumerId),
+    askQualityOf: (output, consumerId, height, reason, now, urgent) =>
+      askQualityOf(parts.viewers, output, consumerId, height, reason, now, urgent),
+    // What a viewer's page said about their buffer and the picture they see
+    // (roadmap item 98), as values.
+    bufferOf: (output, consumerId, spanSec) => bufferOf(parts.viewers, output, consumerId, spanSec),
+    visiblePictureOf: (output, consumerId) => visiblePictureOf(parts.viewers, output, consumerId),
+    // How long this host takes to close a first segment: how soon another
+    // output could have the piece a viewer needs.
+    expectedFirstSegmentMs: () => parts.hostTimings.expectedFirstSegmentMs(),
+    dropAskOf: (output, consumerId) => dropAskOf(parts.viewers, output, consumerId),
+    // EACH viewer's own link and who is present: a thin link decides for the
+    // person on it and for nobody else (roadmap item 97, step 11), so the
+    // budget asks per viewer instead of taking the worst over all of them.
+    presentOn: (output) => presentOn(parts.viewers, output),
+    linkReportOf: (output, consumerId) => linkReportOf(parts.viewers, output, consumerId),
+    // The soundtrack a viewer receives, as part of the load their link carries.
+    viewerAudioLoadOf: (base, consumerId) => parts.renditions.viewerAudioLoadOf(base, consumerId),
+    // The first lever on a viewer's link: another limit of the height on their
+    // screen, prepared and moved to without their player being told.
+    prepareSameHeightSwitch: (output, consumerId, direction, reason) =>
+      parts.renditions.prepareSameHeightSwitch(output, consumerId, direction, reason),
+    sameHeightSwitchPending: (consumerId) => parts.renditions.sameHeightSwitchPending(consumerId),
+    sameHeightSwitchDirection: (consumerId) => parts.renditions.sameHeightSwitchDirection(consumerId),
+    cancelSameHeightSwitch: (consumerId, reason) => parts.renditions.cancelSameHeightSwitch(consumerId, reason),
+    heightReadyFor: (base, consumerId, height) => parts.renditions.heightReadyFor(base, consumerId, height),
+    isLive: (...args) => parts.encodeRuns.isLive(...args),
+    liveConsumers: (...args) => parts.renditions.liveConsumers(...args),
+    liveRunsOf: (...args) => parts.encodeRuns.liveRunsOf(...args),
+    producedNumbers: (...args) => parts.serving.producedNumbers(...args),
+    reportHostLoad: (...args) => parts.hostLoad.reportHostLoad(...args),
+    runStateOf: (...args) => parts.encodeRuns.runStateOf(...args),
+    sampleDownloadRates: (...args) => parts.hostLoad.sampleDownloadRates(...args),
+    get encodeCost() { return parts.encodeCost; },
+    get getSourceStats() { return parts.getSourceStats; },
+    get outputs() { return parts.outputs; },
+    get qualityOffer() { return parts.qualityOffer; },
+    get segmentDurationSec() { return parts.segmentDurationSec; },
+    segmentPaths,
+    get videoEncoder() { return parts.encoders.current; },
+  });
+  parts.budgetTimer = setInterval(() => {
+    parts.cushion.reportCushions();
+    void parts.quality.runQualityBudgetOnce();
+  }, BUDGET_CHECK_INTERVAL_MS);
+  parts.budgetTimer.unref();
+  return parts;
+}