@torrent-tv/proxy 2.88.2 → 2.88.4

package/server.js

@@ -1,725 +1,725 @@
-/**
- * @file Proxy HTTP server bootstrap.
- *
- * Creates and configures the Fastify application, registers all routes and
- * plugins, then starts listening on the first available port at or above the
- * requested one.
- */
-
-import Fastify from "fastify";
-import fastifyCors from "@fastify/cors";
-import fastifyHelmet from "@fastify/helmet";
-import fastifyStatic from "@fastify/static";
-import getPort from "get-port";
-import path from "node:path";
-import { createRequire } from "node:module";
-import { fileURLToPath } from "node:url";
-import { handleHealthGet } from "./routes/health/get.js";
-import { handleHealthzGet } from "./routes/healthz/get.js";
-import { handleApiDeliverySinkGet } from "./routes/api/delivery-sink/get.js";
-import { handleApiLinkProbeGet } from "./routes/api/link-probe/get.js";
-import { handleApiSourcesPost } from "./routes/api/sources/post.js";
-import { handleApiSourceStatsGet } from "./routes/api/sources/stats/get.js";
-import { handleApiSourceFilesGet } from "./routes/api/sources/files/get.js";
-import { handleApiSourceWarmPost } from "./routes/api/sources/warm/post.js";
-import { handleApiPlaybackPlanPost } from "./routes/api/playback-plan/post.js";
-import { handleApiPlaybackPlanAudioTracksPost } from "./routes/api/playback-plan/audio-tracks/post.js";
-import { handleApiClientLogsPost } from "./routes/api/client-logs/post.js";
-import { createClientLogFiles } from "./utils/client-log-file.js";
-import { handleApiSubtitlesGet } from "./routes/api/subtitles/get.js";
-import { handleApiTranscodeSessionsPost } from "./routes/api/transcode-sessions/post.js";
-import { handleApiTranscodeSessionsProgressGet } from "./routes/api/transcode-sessions/progress/get.js";
-import { handleApiTranscodeSessionReleasePost } from "./routes/api/transcode-sessions/release/post.js";
-import { handleApiTranscodeSessionNetReportPost } from "./routes/api/transcode-sessions/net-report/post.js";
-import { handleApiTranscodeSessionFragmentFarPost } from "./routes/api/transcode-sessions/fragment-far/post.js";
-import { handleApiTranscodeSessionSeekPost } from "./routes/api/transcode-sessions/seek/post.js";
-import { handleStreamGet } from "./routes/stream/get.js";
-import { handleTranscodeSessionFileGet } from "./routes/transcode/session-file/get.js";
-import { handleTranscodeVariantFileGet } from "./routes/transcode/variant-file/get.js";
-import { handleTranscodeAudioFileGet } from "./routes/transcode/audio-file/get.js";
-import { handleTranscodeVariantWarmGet } from "./routes/transcode/variant-warm/get.js";
-import { handleTranscodeAudioWarmGet } from "./routes/transcode/audio-warm/get.js";
-import { createSourceRegistry } from "./store/source-registry.js";
-import { WorkerTorrentPool } from "./services/torrent/worker/pool-adapter.js";
-import { wireOutputs } from "./services/server/wire-outputs.js";
-import { createPlaybackPlanner } from "./services/media/playback-planner.js";
-import { KeyframeTables } from "./services/media/KeyframeTables.js";
-import { contentsOf } from "./services/torrent/Contents.js";
-import { SubtitleOrchestrator } from "./services/media/SubtitleOrchestrator.js";
-import { containerOrchestrator, CONTAINER_HEAD_BYTES } from "./services/media/ContainerOrchestrator.js";
-import {
-  warmSubtitleCues,
-  cuesHeldFor,
-  subtitleTracksOf,
-  declaredSubtitleTracksOf,
-  forgetSubtitles
-} from "./services/media/SubtitleCues.js";
-import { detectVideoEncoder, benchmarkDecodeCost, benchmarkContention, benchmarkCopySpeed, detectTonemapSupport, softwareDescriptor } from "./services/encode/hwaccel.js";
-import { calibrateEncoder, HostCalibration } from "./services/encode/calibration.js";
-import { readHostFingerprint } from "./services/encode/fingerprint.js";
-import { measureStartAndStop } from "./services/encode/start-stop-cost.js";
-import { logger } from "./utils/logger.js";
-import { completedFilesRoot } from "./services/storage/files/CompletedFiles.js";
-
-const __filename = fileURLToPath(import.meta.url);
-const __dirname = path.dirname(__filename);
-const require = createRequire(import.meta.url);
-const { version } = require("./package.json");
-const publicRoot = path.resolve(__dirname, "./public");
-
-/**
- * Build a list of candidate port numbers starting at `startPort`.
- *
- * @param {number} startPort
- * @param {number} [maxAttempts=51]
- * @returns {number[]}
- */
-function buildPortCandidates(startPort, maxAttempts = 51) {
-  const ports = [];
-  for (let index = 0; index < maxAttempts; index += 1) {
-    ports.push(startPort + index);
-  }
-  return ports;
-}
-
-/**
- * @typedef {Object} ProxyServerOptions
- * @property {string}  host           - Bind host (e.g. "127.0.0.1" or "0.0.0.0").
- * @property {number}  port           - Preferred listen port.
- * @property {boolean} transcodeAudio - Whether HLS audio transcoding is enabled.
- * @property {string}  ffmpegBin      - Path to the ffmpeg executable.
- * @property {number}  [memoryBytes]  - Per-torrent budget for pieces held in memory (undefined = store default).
- * @property {string}  [segmentFormat] - HLS output container: "fmp4" (default) or "mpegts".
- * @property {string}  [stateDir] - Where to keep what this host has measured about itself.
- */
-
-/**
- * Create, configure, and start the proxy HTTP server.
- *
- * @param {ProxyServerOptions} options
- * @returns {Promise<{ app: import("fastify").FastifyInstance, port: number }>}
- */
-export async function startProxyServer({
-  host, port, transcodeAudio, ffmpegBin, memoryBytes, segmentFormat, stateDir, onSubtitleCues, diagnostics, budgetPolicy,
-  deliverySink = false,
-  // Where the proxy writes its own log. Its DIRECTORY is what matters here:
-  // the browser's half of every session is written beside it, so the two are
-  // on one durable disk and join by name.
-  logFile = ""
-}) {
-  const app = Fastify({
-    // No practical body-size limit — the proxy server is localhost-only and
-    // receives torrent source payloads that may be arbitrarily large.
-    bodyLimit: 256 * 1024 * 1024 // 256 MB
-  });
-
-  await app.register(fastifyHelmet, {
-    // Proxy serves media to a different origin (registry UI), so CORP must allow cross-origin usage.
-    crossOriginResourcePolicy: {
-      policy: "cross-origin"
-    }
-  });
-  await app.register(fastifyCors, {
-    origin: true,
-    methods: ["GET", "POST", "OPTIONS"],
-    allowedHeaders: ["Content-Type", "Range"]
-  });
-
-  // Allow browser requests from an HTTPS page to this private-network proxy
-  // without triggering Chromium's Private Network Access permission prompt.
-  app.addHook("onRequest", async (_req, reply) => {
-    reply.header("Access-Control-Allow-Private-Network", "true");
-  });
-
-  const sourceRegistry = createSourceRegistry(200);
-  // The torrent runs on its own thread. Profiling a live seek (2026-08-02)
-  // found the main thread ~85% occupied by WebTorrent — buffer concatenation
-  // ~15%, wire updates ~9%, garbage collection ~5% — while three of four cores
-  // idled. Serving a segment shared that thread, so reading an already-finished
-  // 10 MB file took 12-23 s against 125 ms to hand it to the channel. The
-  // adapter keeps TorrentPool's interface, so nothing downstream changed.
-  const torrentPool = new WorkerTorrentPool({
-    memoryBytes,
-    stateDir,
-    // Pieces arriving is ANNOUNCED by the thread that owns the swarm; what is
-    // done about it — walking a file's new subtitle clusters and pushing what
-    // came out — happens here, because that is a reading of what the file says
-    // about itself. The walk used to run in that thread, which is what put a
-    // container parse there.
-    // The ONE forward reference here, and it is safe by construction: this fires
-    // only once the worker thread is up and a piece has verified, which is long
-    // after the declarations below have run.
-    onPiecesArrived: ({ sourceKey, fileIndexes }) => {
-      for (const fileIndex of fileIndexes) {
-        void pushFreshCues(sourceKey, fileIndex);
-      }
-    }
-  });
-  const selectedPort = await getPort({
-    port: buildPortCandidates(port)
-  });
-  // Auto-detect the best available H.264 encoder (hardware-accelerated or
-  // software) once at startup, with a real test-encode and graceful fallback.
-  // Only needed when transcoding can occur.
-  let videoEncoder = transcodeAudio
-    ? await detectVideoEncoder({ ffmpegBin, logger })
-    : null;
-  // WHAT THIS HOST DOES, measured before any viewer exists. Every one of these
-  // was gated on the chosen encoder being SOFTWARE, and two of the three are
-  // not about the encoder at all: a host with a GPU decodes in software just
-  // the same — no hardware decoder is asked for anywhere — and what a second
-  // job costs is a property of the machine. So a GPU host had no decode cost,
-  // no contention penalty and no encoder throughput, and the quality offer,
-  // which is arithmetic over those three, had nothing to compute from.
-  //
-  // The decode model first and the throughput second — they are independent
-  // (the rungs are timed on raw frames), but the order keeps the two figures
-  // side by side in the log.
-  const decodeCostModel = transcodeAudio
-    ? await benchmarkDecodeCost({ ffmpegBin, logger })
-    : null;
-  // What a second job costs on this host. Measured because the budget adds
-  // independent prices and this host says two jobs that each fit alone do not
-  // fit together — 2.6× on the addon box (2026-08-18).
-  const contentionPenalties = transcodeAudio
-    ? await benchmarkContention({ ffmpegBin, logger })
-    : null;
-  // WHICH MODES THIS HOST MAY ENCODE WITH, AND WHAT EACH COSTS AT EVERY SIZE
-  // (roadmap item 97, step 14). Every setting an output can actually be
-  // encoded at is put through the product's own arguments and its segments
-  // decoded, then timed at the sizes of the ladder; a mode that fails either is
-  // not used, and a size nothing was timed at is not offered. Software is
-  // calibrated too when detection chose a device: it is what a failing device
-  // falls back to for the rest of the process, and it must not be priced by
-  // the device's figures.
-  //
-  // And WHICH CONFIGURATION all of it describes — the ffmpeg and x264 builds,
-  // the processor and the threads, the device and its driver — which is what
-  // decides whether anything kept from an earlier run still applies.
-  const fingerprint = videoEncoder
-    ? await readHostFingerprint({ ffmpegBin, encoder: videoEncoder })
-    : null;
-  /** @type {Record<string, object[]>} */
-  const calibratedModes = {};
-  if (videoEncoder) {
-    calibratedModes[videoEncoder.kind] =
-      (await calibrateEncoder({ ffmpegBin, encoder: videoEncoder, logger })).modes;
-    if (videoEncoder.kind !== "software") {
-      calibratedModes.software =
-        (await calibrateEncoder({ ffmpegBin, encoder: softwareDescriptor(), logger })).modes;
-      // A device that passed detection's own test and then none of whose modes
-      // passed the product's is not used: detection encodes a synthetic pattern
-      // with arguments of its own, and it is the product's arguments a viewer
-      // is served with.
-      if (calibratedModes[videoEncoder.kind].length === 0) {
-        logger.warn(
-          `calibration: ${videoEncoder.name} passed detection but none of its modes produced correct segments ` +
-          "through the product's own arguments; encoding in software instead"
-        );
-        videoEncoder = softwareDescriptor();
-      }
-    }
-  }
-  const calibration = videoEncoder ? new HostCalibration({ byKind: calibratedModes, fingerprint }) : null;
-  // What this host does with a picture it does NOT re-encode. Every other
-  // startup measurement prices encoding or decoding, and a copied picture does
-  // neither — it reads packets and writes them out again — so that whole branch
-  // had no speed until its own run had been running long enough to report one.
-  // The encoding layer decides where encoders go from arrivals, and an arrival
-  // cannot be computed without a speed, so the moment it mattered most was the
-  // moment nothing was known. Measured whatever the encoder is: copying does not
-  // touch it.
-  const copySpeedX = transcodeAudio
-    ? await benchmarkCopySpeed({ ffmpegBin, logger })
-    : null;
-  // WHAT A START AND A STOP COST HERE, before any viewer exists. Both decide one
-  // thing — leave an encoder where it stands, or kill it and start another —
-  // and both used to be learned only from runs that had ENDED, so at a cold
-  // open they were zero. Zero does not read as "not measured": it reads as
-  // "free", and a free move is always taken. Field 2026-09-08: an encoder moved
-  // between two adjacent numbers every half second and produced nothing.
-  const startStopCost = transcodeAudio
-    ? await measureStartAndStop({ ffmpegBin, encoder: videoEncoder, logger })
-    : null;
-  // Whether this ffmpeg build can tone-map HDR→SDR (zscale + tonemap filters).
-  // Detected once; the session manager applies the tonemap chain only for HDR
-  // sources on the software path when available.
-  const tonemapSupported = transcodeAudio
-    ? await detectTonemapSupport({ ffmpegBin, logger })
-    : false;
-  // Where every file's keyframe table lives, and the only thing that reads one.
-  // Built here rather than inside the session manager because it is not a fact
-  // about a session: the playback planner warms it while a file is being opened
-  // and no session exists yet, and a session created later is handed the very
-  // object that warm read filled in.
-  // Subtitles are two layers put together, and this is the one place that knows
-  // about both: what a file STATES about its subtitle tracks and how a cue is
-  // written out is the media layer's, while which clusters may be read, where
-  // the cursor stands and what one walk at a time means are the TORRENT's own
-  // rules. The media layer used to import the second directly.
-  /**
-   * One file of one torrent, reduced to what the subtitle walk may know: a
-   * name, a length, which ranges are downloaded whole, and how to read one of
-   * those without asking the swarm. Piece length, file offsets and the bitfield
-   * stay in the torrent's thread.
-   *
-   * @param {string} sourceKey
-   * @param {number} fileIndex
-   * @param {object} [known] - The torrent handle when the caller already has it.
-   * @returns {Promise<import("./services/media/SubtitleCues.js").HeldFile | null>}
-   */
-  const heldFileFor = async (sourceKey, fileIndex, known = null) => {
-    let torrent = known;
-    if (!torrent) {
-      const record = sourceRegistry.get(sourceKey);
-      if (!record) {
-        return null;
-      }
-      torrent = await torrentPool.getTorrent(record.sourceType, record.source);
-    }
-    const file = torrent?.files?.[fileIndex];
-    if (!file || !(file.length > 0)) {
-      return null;
-    }
-    return {
-      sourceKey,
-      fileIndex,
-      name: String(file.name ?? ""),
-      length: file.length,
-      heldRanges: () => torrentPool.heldRangesOf(torrent, fileIndex),
-      readHeld: (start, end) => torrentPool.readHeldOf(torrent, fileIndex, start, end)
-    };
-  };
-  /** @type {Set<string>} */
-  const walksInFlight = new Set();
-  /**
-   * Walk whatever new cues a file now holds and push them to the viewers.
-   *
-   * @param {string} sourceKey
-   * @param {number} fileIndex
-   * @returns {Promise<void>}
-   */
-  const pushFreshCues = async (sourceKey, fileIndex) => {
-    // A pass that arrives while the previous one is still walking is dropped
-    // rather than queued: `verified` fires per piece, so on a fast download
-    // these arrive many times a second, and a queue of identical passes would
-    // only postpone the one that has something new to find. The walk is
-    // serialized per file inside `SubtitleCues` anyway.
-    const key = `${sourceKey}:${fileIndex}`;
-    if (walksInFlight.has(key)) {
-      return;
-    }
-    walksInFlight.add(key);
-    try {
-      const file = await heldFileFor(sourceKey, fileIndex);
-      if (!file) {
-        return;
-      }
-      for (const entry of await warmSubtitleCues(file)) {
-        const span = entry.spanStartSeconds === null
-          ? "empty"
-          : `${entry.spanStartSeconds.toFixed(1)}-${entry.spanEndSeconds.toFixed(1)}s`;
-        logger.info(
-          `subtitle push ${sourceKey.slice(0, 8)}:${fileIndex} track ${entry.trackIndex}: ` +
-          `${entry.cues.length} new cue(s) covering ${span}, ` +
-          `clusters walked ${entry.walkedClusters}/${entry.indexedClusters}, cursor ${entry.cursor}`
-        );
-        onSubtitleCues?.({ sourceKey, fileIndex, ...entry });
-      }
-    } catch (error) {
-      logger.warn(`subtitle push ${sourceKey.slice(0, 8)}:${fileIndex} failed: ${error?.message ?? error}`);
-    } finally {
-      walksInFlight.delete(key);
-    }
-  };
-  const subtitles = new SubtitleOrchestrator(containerOrchestrator, {
-    warm: async (torrent, fileIndex, sourceKey) => {
-      const file = await heldFileFor(sourceKey, fileIndex, torrent);
-      return file ? warmSubtitleCues(file) : [];
-    },
-    held: async (torrent, fileIndex, sourceKey, trackNumber) => {
-      const file = await heldFileFor(sourceKey, fileIndex, torrent);
-      return file ? cuesHeldFor(file, trackNumber) : null;
-    },
-    tracksOf: async (torrent, fileIndex, sourceKey) => {
-      const file = await heldFileFor(sourceKey, fileIndex, torrent);
-      return file ? subtitleTracksOf(file) : [];
-    },
-    declaredTracksOf: async (torrent, fileIndex, sourceKey) => {
-      const file = await heldFileFor(sourceKey, fileIndex, torrent);
-      return file ? declaredSubtitleTracksOf(file) : [];
-    },
-    forget: forgetSubtitles
-  });
-  // WHAT A FILE STATES ABOUT ITSELF IS READ HERE, ON THIS THREAD.
-  //
-  // The container layer is built from one function — `readRange(start, end)` —
-  // and the bytes behind it live in shared memory, so the parse has no reason
-  // to happen anywhere else. It used to happen in the torrent thread for one
-  // stated reason, that "the main thread cannot open a read stream on one of
-  // its files": true of WebTorrent's own API, and not of the bytes. The price
-  // was three commands, a second `ContainerOrchestrator` in that thread, and
-  // every answer carried back across the channel — which is what made "where is
-  // this fact kept" a question at all.
-  //
-  // The edges are fetched first because the file this is asked about is usually
-  // one nobody has played: a sidecar soundtrack is asked about before anyone
-  // has chosen it, and a Matroska file's Cues sit at the END behind a SeekHead
-  // in the head, so a read that has neither waits for the swarm twice over.
-  const containerOver = async ({ sourceKey, fileIndex, tailBytes = 0 }) => {
-    const record = sourceRegistry.get(sourceKey);
-    if (!record) {
-      return null;
-    }
-    const torrent = await torrentPool.getTorrent(record.sourceType, record.source);
-    const file = torrent?.files?.[fileIndex];
-    if (!file || !(file.length > 0)) {
-      return null;
-    }
-    try {
-      await torrentPool.prefetchFileEdges(torrent, fileIndex, {
-        headBytes: CONTAINER_HEAD_BYTES,
-        tailBytes,
-        timeoutMs: 60_000
-      });
-    } catch {
-      // A prefetch that failed is not a reason to skip the read: the read
-      // fetches what it needs itself, only more slowly.
-    }
-    return {
-      sourceKey,
-      fileIndex,
-      readRange: (start, end) =>
-        torrentPool.readRangeOf(torrent, fileIndex, start, Math.min(end, file.length - 1)),
-      fileSize: file.length,
-      label: String(file.name ?? "")
-    };
-  };
-  const keyframeTables = new KeyframeTables({
-    // Read by the same container that answers the track table and the media
-    // info, from the same header, and now in the same thread as the session
-    // that is waiting for it.
-    readTable: async ({ sourceKey, fileIndex }) => {
-      // Both edges: the Cues of a Matroska file are at the end.
-      const params = await containerOver({ sourceKey, fileIndex, tailBytes: CONTAINER_HEAD_BYTES });
-      if (!params) {
-        return null;
-      }
-      const index = await containerOrchestrator.getKeyframeIndex(params);
-      return {
-        times: index?.times ?? null,
-        tolerance: index?.tolerance ?? 0,
-        // Which container answered, whether or not it produced a table: the
-        // refusal that follows names it, and a measurement of how often an
-        // index disagrees with its own file cannot be read without it.
-        format: (await containerOrchestrator.getContainer(params))?.formatName ?? "unrecognised"
-      };
-    }
-  });
-  const outputParts = wireOutputs({
-    enabled: transcodeAudio,
-    keyframeTables,
-    ffmpegBin,
-    localBindHost: host,
-    localPort: selectedPort,
-    videoEncoder,
-    calibration,
-    decodeCostModel,
-    contentionPenalties,
-    copySpeedX,
-    startStopCost,
-    tonemapSupported,
-    segmentFormatId: segmentFormat,
-    stateDir,
-    diagnostics,
-    // Live download stats accessor for the realtime budget: lets it tell a
-    // CPU-bound transcode from a download-starved input before downscaling.
-    // What every torrent here has moved, so the proxy can price its own
-    // downloading, hashing and delivery against the machine (roadmap item 7).
-    // What the spilled pieces weigh on the torrent thread, and how to tell them
-    // their share of the disk. The owner of the disk is on this side, where the
-    // segments are; the pieces are on the other.
-    spillDisk: typeof torrentPool.allowSpillBytes === "function"
-      ? {
-          held: () => torrentPool.spilledBytes ?? 0,
-          allow: (bytes) => torrentPool.allowSpillBytes(bytes)
-        }
-      : null,
-    // Files downloaded whole, kept as files. They live on the torrent thread
-    // too, and until 2026-09-14 they had no bound of any kind — a 2.8 GB film
-    // on a host whose disk is often a 32 GB card.
-    wholeFiles: typeof torrentPool.allowWholeFileBytes === "function"
-      ? {
-          held: () => torrentPool.wholeFileBytes ?? 0,
-          allow: (bytes) => torrentPool.allowWholeFileBytes(bytes),
-          root: completedFilesRoot()
-        }
-      : null,
-    // The pieces held IN memory, on the torrent thread. Told their share by the
-    // same owner that divides the disk: what does not fit here is spilled
-    // there, so the two cannot be divided apart.
-    memoryClaimant: typeof torrentPool.allowMemoryBytes === "function"
-      ? {
-          held: () => torrentPool.memoryClaim?.held ?? 0,
-          wanted: () => torrentPool.memoryClaim?.wanted ?? 0,
-          allow: (bytes) => torrentPool.allowMemoryBytes(bytes)
-        }
-      : null,
-    diagnosticsRoot: stateDir || "",
-    budgetPolicy,
-    getTorrentTotals: async () => {
-      if (typeof torrentPool.getTorrentTotals !== "function") {
-        return null;
-      }
-      return torrentPool.getTorrentTotals();
-    },
-    // The priority map, on its way to the downloading. It lives in another
-    // thread, so the map crosses the worker channel; what it does with it —
-    // seconds into bytes, what to ask the swarm for, what to keep in memory —
-    // is its own business.
-    setPriorityMap: async ({ sourceKey, fileIndex, durationSeconds, zones }) => {
-      const record = sourceRegistry.get(sourceKey);
-      if (!record) {
-        return;
-      }
-      try {
-        await torrentPool.setPriorityMap({ sourceKey, fileIndex, durationSeconds, zones });
-      } catch {
-        // Best effort: the map is republished on the next change, and the
-        // downloading goes on serving reads meanwhile.
-      }
-    },
-    getSourceStats: async (sourceKey, fileIndex) => {
-      const record = sourceRegistry.get(sourceKey);
-      if (!record) {
-        return null;
-      }
-      try {
-        const torrent = await torrentPool.getTorrent(record.sourceType, record.source);
-        // Awaited for the same reason as the stats route: this now crosses a
-        // thread boundary and returns a promise.
-        return await torrentPool.getFileStats(torrent, Number.isInteger(fileIndex) ? fileIndex : null);
-      } catch {
-        return null;
-      }
-    },
-    // Reuse the media info the planner already probed for this file (same
-    // ffmpeg scan) so createSession skips its own probe. Late-bound: invoked
-    // only at session-create time, after playbackPlanner is initialised.
-    getCachedMediaInfo: (params) => playbackPlanner.getCachedMediaInfo(params),
-    // The file's audio tracks, for the master playlist's rendition group. Already
-    // probed for the browser's audio menu; read from there rather than probed again.
-    getCachedAudioTracks: (params) => playbackPlanner.getCachedAudioTracks(params),
-    // What a file declares about itself, read by the container layer from the
-    // same header its track table comes from. This is how the session learns
-    // where a soundtrack shipped as its own file begins — it used to spawn an
-    // ffmpeg over this proxy's own HTTP to ask the same question of the same
-    // bytes, and that read cost 8.1 s of every cold start (field 2026-09-03).
-    getContainerMediaInfo: async ({ sourceKey, fileIndex }) => {
-      try {
-        const params = await containerOver({ sourceKey, fileIndex });
-        return params ? await containerOrchestrator.getMediaInfo(params) : null;
-      } catch {
-        return null;
-      }
-    },
-    // Pull one whole file onto the disk. Used for a soundtrack that ships beside
-    // the picture, once the encoder is as far ahead of the viewer as it is
-    // allowed to get — the one moment the swarm's capacity is demonstrably
-    // spare. A bounded read of the file's whole length, NOT `file.select()`:
-    // selecting a file alongside the readers' own windows is what made a seek
-    // wait 93 s while the swarm fetched 2.47 GB in file order (see
-    // `#syncSelections` in `torrent/torrent-pool.js`).
-    fetchWholeFile: async ({ sourceKey, fileIndex }) => {
-      const record = sourceRegistry.get(sourceKey);
-      if (!record) {
-        return;
-      }
-      const torrent = await torrentPool.getTorrent(record.sourceType, record.source);
-      // The same background fill the warm-up starts when a file is chosen, not a
-      // second way of doing it. It is guarded against running twice on one file,
-      // so the two triggers converge instead of putting two readers on the same
-      // soundtrack — and only one of them would have stood aside for the
-      // picture. This trigger remains for the session that never had a warm-up
-      // before it.
-      await torrentPool.fillFileInBackground?.(torrent, fileIndex);
-    }
-  });
-  // What the last life of this process left on the disk. The kernel kills this
-  // one often enough for that to be an ordinary state rather than an odd one —
-  // twice in a single viewing on 2026-09-02 — and when it does, no exit handler
-  // runs and nothing is cleared up. So this is both the cleanup and the only
-  // record that those encoders ended at all: it says what it found before it
-  // decides anything, keeps the segments whose closure is proven, and removes
-  // the one piece per output that was being written when the process died.
-  outputParts.lifecycle.adoptSegmentsLeftBehind();
-  const playbackPlanner = createPlaybackPlanner({
-    ffmpegBin,
-    transcodeAudioEnabled: transcodeAudio,
-    localBaseUrl: outputParts.localBaseUrl,
-    sourceRegistry,
-    torrentPool,
-    sidecarsFromTorrent: (torrent, fileIndex) => contentsOf(torrent).sidecarsOf(fileIndex),
-    // What a file declares about its own tracks, parsed on this thread from the
-    // header the swarm delivered. The planner used to ask the torrent pool for
-    // this, which meant the media layer asking the torrent layer to parse a
-    // container on its behalf, in the other thread, with the answer carried
-    // back over the channel.
-    declaredTracksOf: async ({ sourceKey, fileIndex }) => {
-      const params = await containerOver({ sourceKey, fileIndex });
-      return params ? await containerOrchestrator.getTracks(params) : [];
-    },
-    warmKeyframeIndex: (params) => keyframeTables.warm(params),
-    expectedFirstSegmentMs: () => outputParts.hostTimings.expectedFirstSegmentMs(),
-    expectedSessionCreateMs: () => outputParts.hostTimings.expectedSessionCreateMs(),
-    // The quality menu is on screen from the moment a file is opened, so the
-    // heights this host can actually serve have to be answerable before any
-    // encoder exists — from the probe and the startup benchmarks alone.
-    predictOfferedHeights: (mediaInfo) => outputParts.quality.predictOfferedHeights(mediaInfo)
-  });
-
-  app.get("/health", async (req, reply) => handleHealthGet(req, reply, { version }));
-  app.get("/healthz", async (req, reply) => handleHealthzGet(req, reply, { version }));
-  // Off unless --delivery-sink was given; see the route for why it exists.
-  app.get("/api/delivery-sink", async (req, reply) =>
-    handleApiDeliverySinkGet(req, reply, { enabled: deliverySink === true })
-  );
-  // Bytes a browser can time its link with before any film has been chosen.
-  app.get("/api/link-probe", async (req, reply) => handleApiLinkProbeGet(req, reply));
-  app.post("/api/sources", async (req, reply) =>
-    handleApiSourcesPost(req, reply, { sourceRegistry })
-  );
-  app.get("/api/sources/:sourceKey/stats", async (req, reply) =>
-    handleApiSourceStatsGet(req, reply, { sourceRegistry, torrentPool })
-  );
-  app.get("/api/sources/:sourceKey/files", async (req, reply) =>
-    handleApiSourceFilesGet(req, reply, { sourceRegistry, torrentPool })
-  );
-  app.post("/api/sources/:sourceKey/warm", async (req, reply) =>
-    handleApiSourceWarmPost(req, reply, {
-      sourceRegistry,
-      torrentPool,
-      playbackPlanner,
-      // How long the file runs, read on this thread from the header. The warm
-      // turns a position in seconds into a byte offset and needs it; it used to
-      // read the container itself, in the torrent thread, to find out.
-      durationOf: async (params) => {
-        const over = await containerOver(params);
-        return over ? (await containerOrchestrator.getMediaInfo(over))?.durationSeconds ?? null : null;
-      }
-    })
-  );
-  // The browser's own log, kept beside the proxy's. See the route's own file
-  // for why it is here and not only on the registry server.
-  const clientLogs = logFile ? createClientLogFiles(logFile) : null;
-  app.post("/api/client-logs", async (req, reply) =>
-    handleApiClientLogsPost(req, reply, { clientLogs })
-  );
-  app.post("/api/playback-plan", async (req, reply) =>
-    handleApiPlaybackPlanPost(req, reply, { playbackPlanner })
-  );
-  app.post("/api/playback-plan/audio-tracks", async (req, reply) =>
-    handleApiPlaybackPlanAudioTracksPost(req, reply, { playbackPlanner })
-  );
-  app.get("/api/subtitles", async (req, reply) =>
-    handleApiSubtitlesGet(req, reply, {
-      sourceRegistry,
-      torrentPool,
-      ffmpegBin,
-      localBaseUrl: outputParts.localBaseUrl,
-      viewers: outputParts.viewers,
-      subtitles
-    })
-  );
-  app.get("/stream", async (req, reply) =>
-    handleStreamGet(req, reply, {
-      sourceRegistry,
-      torrentPool,
-      // So a session that has produced nothing yet can still show it is being
-      // fed. The route knows only files; the session id rides on the URL the
-      // session itself built.
-      noteInputBytes: (sessionId, bytes) => outputParts.encodeRuns.noteInputBytes(sessionId, bytes)
-    })
-  );
-  app.post("/api/transcode-sessions", async (req, reply) =>
-    handleApiTranscodeSessionsPost(req, reply, { viewerRequests: outputParts.viewerRequests, renditions: outputParts.renditions, quality: outputParts.quality, outputs: outputParts.outputs, lookaheadSeconds: outputParts.lookaheadSeconds, sourceRegistry, torrentPool })
-  );
-  app.post("/api/transcode-sessions/:sessionId/release", async (req, reply) =>
-    handleApiTranscodeSessionReleasePost(req, reply, { lifecycle: outputParts.lifecycle })
-  );
-  app.get("/api/transcode-sessions/:sessionId/progress", async (req, reply) =>
-    handleApiTranscodeSessionsProgressGet(req, reply, { viewerRequests: outputParts.viewerRequests })
-  );
-  app.post("/api/transcode-sessions/:sessionId/net-report", async (req, reply) =>
-    // A viewer's statement about itself goes to the VIEWER layer, not through
-    // the session manager: what it needs is the live sessions and the registry
-    // of viewers, and nothing about encoding.
-    handleApiTranscodeSessionNetReportPost(req, reply, {
-      outputs: outputParts.outputs,
-      viewers: outputParts.viewers,
-      renditions: outputParts.renditions,
-      quality: outputParts.quality
-    })
-  );
-  app.post("/api/transcode-sessions/:sessionId/fragment-far", async (req, reply) =>
-    handleApiTranscodeSessionFragmentFarPost(req, reply, { serving: outputParts.serving })
-  );
-  app.post("/api/transcode-sessions/:sessionId/seek", async (req, reply) =>
-    handleApiTranscodeSessionSeekPost(req, reply, { viewerRequests: outputParts.viewerRequests })
-  );
-  app.get("/transcode/:sessionId/:fileName", async (req, reply) =>
-    handleTranscodeSessionFileGet(req, reply, { serving: outputParts.serving, viewerRequests: outputParts.viewerRequests })
-  );
-  // A quality variant's files. Registered before the static handler for the
-  // same reason as the line above, and kept a separate route rather than a
-  // wildcard so the height stays a parsed parameter.
-  // Registered BEFORE the variant file route: `warm` is not a file name, and
-  // Fastify matches a static segment ahead of a parameter either way — stated
-  // here so the order is not "tidied" into a bug.
-  app.get("/transcode/:sessionId/v/:height/warm", async (req, reply) =>
-    handleTranscodeVariantWarmGet(req, reply, { renditions: outputParts.renditions, serving: outputParts.serving, viewerRequests: outputParts.viewerRequests })
-  );
-  app.get("/transcode/:sessionId/a/:track/warm", async (req, reply) =>
-    handleTranscodeAudioWarmGet(req, reply, { renditions: outputParts.renditions, serving: outputParts.serving, viewerRequests: outputParts.viewerRequests })
-  );
-  app.get("/transcode/:sessionId/a/:trackIndex/:fileName", async (req, reply) =>
-    handleTranscodeAudioFileGet(req, reply, { renditions: outputParts.renditions, serving: outputParts.serving, viewerRequests: outputParts.viewerRequests })
-  );
-  app.get("/transcode/:sessionId/v/:height/:fileName", async (req, reply) =>
-    handleTranscodeVariantFileGet(req, reply, { renditions: outputParts.renditions, serving: outputParts.serving, viewerRequests: outputParts.viewerRequests })
-  );
-  await app.register(fastifyStatic, {
-    root: publicRoot,
-    prefix: "/",
-    serveDotFiles: true
-  });
-
-  app.addHook("onClose", async () => {
-    // Order matters: stop the ffmpeg readers (HLS sessions) before destroying
-    // the torrents whose files they read from, then remove the torrent data.
-    await outputParts.lifecycle.disposeAll();
-    await torrentPool.destroyAll();
-  });
-
-  await app.listen({ host, port: selectedPort });
-  return {
-    app,
-    port: selectedPort,
-    // Asked over the tunnel when the proxy a viewer landed on has refused their
-    // file: could THIS host sustain it? Answered from the startup benchmarks
-    // and a description, so it needs no torrent and costs milliseconds.
-    outputParts,
-    // The browser only ever knows a source by its REGISTRY key (a hash of the
-    // raw request bytes, scoped to one API session) — never the torrent
-    // pool's own key (the content's infohash, shared across a magnet and a
-    // `.torrent` naming the same film). The subtitle push subscription is
-    // recorded from a browser request and published from the pool's side, so
-    // resolving one into the other is what lets the two ends agree on what
-    // they are both calling "sourceKey".
-    sourceRegistry,
-    // Which films this host holds, for the health poll the browser makes before
-    // it picks a proxy. The pool is on the worker thread and this is the way to
-    // it from the process that answers that poll.
-    torrentPool
-  };
-}
+/**
+ * @file Proxy HTTP server bootstrap.
+ *
+ * Creates and configures the Fastify application, registers all routes and
+ * plugins, then starts listening on the first available port at or above the
+ * requested one.
+ */
+
+import Fastify from "fastify";
+import fastifyCors from "@fastify/cors";
+import fastifyHelmet from "@fastify/helmet";
+import fastifyStatic from "@fastify/static";
+import getPort from "get-port";
+import path from "node:path";
+import { createRequire } from "node:module";
+import { fileURLToPath } from "node:url";
+import { handleHealthGet } from "./routes/health/get.js";
+import { handleHealthzGet } from "./routes/healthz/get.js";
+import { handleApiDeliverySinkGet } from "./routes/api/delivery-sink/get.js";
+import { handleApiLinkProbeGet } from "./routes/api/link-probe/get.js";
+import { handleApiSourcesPost } from "./routes/api/sources/post.js";
+import { handleApiSourceStatsGet } from "./routes/api/sources/stats/get.js";
+import { handleApiSourceFilesGet } from "./routes/api/sources/files/get.js";
+import { handleApiSourceWarmPost } from "./routes/api/sources/warm/post.js";
+import { handleApiPlaybackPlanPost } from "./routes/api/playback-plan/post.js";
+import { handleApiPlaybackPlanAudioTracksPost } from "./routes/api/playback-plan/audio-tracks/post.js";
+import { handleApiClientLogsPost } from "./routes/api/client-logs/post.js";
+import { createClientLogFiles } from "./utils/client-log-file.js";
+import { handleApiSubtitlesGet } from "./routes/api/subtitles/get.js";
+import { handleApiTranscodeSessionsPost } from "./routes/api/transcode-sessions/post.js";
+import { handleApiTranscodeSessionsProgressGet } from "./routes/api/transcode-sessions/progress/get.js";
+import { handleApiTranscodeSessionReleasePost } from "./routes/api/transcode-sessions/release/post.js";
+import { handleApiTranscodeSessionNetReportPost } from "./routes/api/transcode-sessions/net-report/post.js";
+import { handleApiTranscodeSessionFragmentFarPost } from "./routes/api/transcode-sessions/fragment-far/post.js";
+import { handleApiTranscodeSessionSeekPost } from "./routes/api/transcode-sessions/seek/post.js";
+import { handleStreamGet } from "./routes/stream/get.js";
+import { handleTranscodeSessionFileGet } from "./routes/transcode/session-file/get.js";
+import { handleTranscodeVariantFileGet } from "./routes/transcode/variant-file/get.js";
+import { handleTranscodeAudioFileGet } from "./routes/transcode/audio-file/get.js";
+import { handleTranscodeVariantWarmGet } from "./routes/transcode/variant-warm/get.js";
+import { handleTranscodeAudioWarmGet } from "./routes/transcode/audio-warm/get.js";
+import { createSourceRegistry } from "./store/source-registry.js";
+import { WorkerTorrentPool } from "./services/torrent/worker/pool-adapter.js";
+import { wireOutputs } from "./services/server/wire-outputs.js";
+import { createPlaybackPlanner } from "./services/media/playback-planner.js";
+import { KeyframeTables } from "./services/media/KeyframeTables.js";
+import { contentsOf } from "./services/torrent/Contents.js";
+import { SubtitleOrchestrator } from "./services/media/SubtitleOrchestrator.js";
+import { containerOrchestrator, CONTAINER_HEAD_BYTES } from "./services/media/ContainerOrchestrator.js";
+import {
+  warmSubtitleCues,
+  cuesHeldFor,
+  subtitleTracksOf,
+  declaredSubtitleTracksOf,
+  forgetSubtitles
+} from "./services/media/SubtitleCues.js";
+import { detectVideoEncoder, benchmarkDecodeCost, benchmarkContention, benchmarkCopySpeed, detectTonemapSupport, softwareDescriptor } from "./services/encode/hwaccel.js";
+import { calibrateEncoder, HostCalibration } from "./services/encode/calibration.js";
+import { readHostFingerprint } from "./services/encode/fingerprint.js";
+import { measureStartAndStop } from "./services/encode/start-stop-cost.js";
+import { logger } from "./utils/logger.js";
+import { completedFilesRoot } from "./services/storage/files/CompletedFiles.js";
+
+const __filename = fileURLToPath(import.meta.url);
+const __dirname = path.dirname(__filename);
+const require = createRequire(import.meta.url);
+const { version } = require("./package.json");
+const publicRoot = path.resolve(__dirname, "./public");
+
+/**
+ * Build a list of candidate port numbers starting at `startPort`.
+ *
+ * @param {number} startPort
+ * @param {number} [maxAttempts=51]
+ * @returns {number[]}
+ */
+function buildPortCandidates(startPort, maxAttempts = 51) {
+  const ports = [];
+  for (let index = 0; index < maxAttempts; index += 1) {
+    ports.push(startPort + index);
+  }
+  return ports;
+}
+
+/**
+ * @typedef {Object} ProxyServerOptions
+ * @property {string}  host           - Bind host (e.g. "127.0.0.1" or "0.0.0.0").
+ * @property {number}  port           - Preferred listen port.
+ * @property {boolean} transcodeAudio - Whether HLS audio transcoding is enabled.
+ * @property {string}  ffmpegBin      - Path to the ffmpeg executable.
+ * @property {number}  [memoryBytes]  - Per-torrent budget for pieces held in memory (undefined = store default).
+ * @property {string}  [segmentFormat] - HLS output container: "fmp4" (default) or "mpegts".
+ * @property {string}  [stateDir] - Where to keep what this host has measured about itself.
+ */
+
+/**
+ * Create, configure, and start the proxy HTTP server.
+ *
+ * @param {ProxyServerOptions} options
+ * @returns {Promise<{ app: import("fastify").FastifyInstance, port: number }>}
+ */
+export async function startProxyServer({
+  host, port, transcodeAudio, ffmpegBin, memoryBytes, segmentFormat, stateDir, onSubtitleCues, diagnostics, budgetPolicy,
+  deliverySink = false,
+  // Where the proxy writes its own log. Its DIRECTORY is what matters here:
+  // the browser's half of every session is written beside it, so the two are
+  // on one durable disk and join by name.
+  logFile = ""
+}) {
+  const app = Fastify({
+    // No practical body-size limit — the proxy server is localhost-only and
+    // receives torrent source payloads that may be arbitrarily large.
+    bodyLimit: 256 * 1024 * 1024 // 256 MB
+  });
+
+  await app.register(fastifyHelmet, {
+    // Proxy serves media to a different origin (registry UI), so CORP must allow cross-origin usage.
+    crossOriginResourcePolicy: {
+      policy: "cross-origin"
+    }
+  });
+  await app.register(fastifyCors, {
+    origin: true,
+    methods: ["GET", "POST", "OPTIONS"],
+    allowedHeaders: ["Content-Type", "Range"]
+  });
+
+  // Allow browser requests from an HTTPS page to this private-network proxy
+  // without triggering Chromium's Private Network Access permission prompt.
+  app.addHook("onRequest", async (_req, reply) => {
+    reply.header("Access-Control-Allow-Private-Network", "true");
+  });
+
+  const sourceRegistry = createSourceRegistry(200);
+  // The torrent runs on its own thread. Profiling a live seek (2026-08-02)
+  // found the main thread ~85% occupied by WebTorrent — buffer concatenation
+  // ~15%, wire updates ~9%, garbage collection ~5% — while three of four cores
+  // idled. Serving a segment shared that thread, so reading an already-finished
+  // 10 MB file took 12-23 s against 125 ms to hand it to the channel. The
+  // adapter keeps TorrentPool's interface, so nothing downstream changed.
+  const torrentPool = new WorkerTorrentPool({
+    memoryBytes,
+    stateDir,
+    // Pieces arriving is ANNOUNCED by the thread that owns the swarm; what is
+    // done about it — walking a file's new subtitle clusters and pushing what
+    // came out — happens here, because that is a reading of what the file says
+    // about itself. The walk used to run in that thread, which is what put a
+    // container parse there.
+    // The ONE forward reference here, and it is safe by construction: this fires
+    // only once the worker thread is up and a piece has verified, which is long
+    // after the declarations below have run.
+    onPiecesArrived: ({ sourceKey, fileIndexes }) => {
+      for (const fileIndex of fileIndexes) {
+        void pushFreshCues(sourceKey, fileIndex);
+      }
+    }
+  });
+  const selectedPort = await getPort({
+    port: buildPortCandidates(port)
+  });
+  // Auto-detect the best available H.264 encoder (hardware-accelerated or
+  // software) once at startup, with a real test-encode and graceful fallback.
+  // Only needed when transcoding can occur.
+  let videoEncoder = transcodeAudio
+    ? await detectVideoEncoder({ ffmpegBin, logger })
+    : null;
+  // WHAT THIS HOST DOES, measured before any viewer exists. Every one of these
+  // was gated on the chosen encoder being SOFTWARE, and two of the three are
+  // not about the encoder at all: a host with a GPU decodes in software just
+  // the same — no hardware decoder is asked for anywhere — and what a second
+  // job costs is a property of the machine. So a GPU host had no decode cost,
+  // no contention penalty and no encoder throughput, and the quality offer,
+  // which is arithmetic over those three, had nothing to compute from.
+  //
+  // The decode model first and the throughput second — they are independent
+  // (the rungs are timed on raw frames), but the order keeps the two figures
+  // side by side in the log.
+  const decodeCostModel = transcodeAudio
+    ? await benchmarkDecodeCost({ ffmpegBin, logger })
+    : null;
+  // What a second job costs on this host. Measured because the budget adds
+  // independent prices and this host says two jobs that each fit alone do not
+  // fit together — 2.6× on the addon box (2026-08-18).
+  const contentionPenalties = transcodeAudio
+    ? await benchmarkContention({ ffmpegBin, logger })
+    : null;
+  // WHICH MODES THIS HOST MAY ENCODE WITH, AND WHAT EACH COSTS AT EVERY SIZE
+  // (roadmap item 97, step 14). Every setting an output can actually be
+  // encoded at is put through the product's own arguments and its segments
+  // decoded, then timed at the sizes of the ladder; a mode that fails either is
+  // not used, and a size nothing was timed at is not offered. Software is
+  // calibrated too when detection chose a device: it is what a failing device
+  // falls back to for the rest of the process, and it must not be priced by
+  // the device's figures.
+  //
+  // And WHICH CONFIGURATION all of it describes — the ffmpeg and x264 builds,
+  // the processor and the threads, the device and its driver — which is what
+  // decides whether anything kept from an earlier run still applies.
+  const fingerprint = videoEncoder
+    ? await readHostFingerprint({ ffmpegBin, encoder: videoEncoder })
+    : null;
+  /** @type {Record<string, object[]>} */
+  const calibratedModes = {};
+  if (videoEncoder) {
+    calibratedModes[videoEncoder.kind] =
+      (await calibrateEncoder({ ffmpegBin, encoder: videoEncoder, logger })).modes;
+    if (videoEncoder.kind !== "software") {
+      calibratedModes.software =
+        (await calibrateEncoder({ ffmpegBin, encoder: softwareDescriptor(), logger })).modes;
+      // A device that passed detection's own test and then none of whose modes
+      // passed the product's is not used: detection encodes a synthetic pattern
+      // with arguments of its own, and it is the product's arguments a viewer
+      // is served with.
+      if (calibratedModes[videoEncoder.kind].length === 0) {
+        logger.warn(
+          `calibration: ${videoEncoder.name} passed detection but none of its modes produced correct segments ` +
+          "through the product's own arguments; encoding in software instead"
+        );
+        videoEncoder = softwareDescriptor();
+      }
+    }
+  }
+  const calibration = videoEncoder ? new HostCalibration({ byKind: calibratedModes, fingerprint }) : null;
+  // What this host does with a picture it does NOT re-encode. Every other
+  // startup measurement prices encoding or decoding, and a copied picture does
+  // neither — it reads packets and writes them out again — so that whole branch
+  // had no speed until its own run had been running long enough to report one.
+  // The encoding layer decides where encoders go from arrivals, and an arrival
+  // cannot be computed without a speed, so the moment it mattered most was the
+  // moment nothing was known. Measured whatever the encoder is: copying does not
+  // touch it.
+  const copySpeedX = transcodeAudio
+    ? await benchmarkCopySpeed({ ffmpegBin, logger })
+    : null;
+  // WHAT A START AND A STOP COST HERE, before any viewer exists. Both decide one
+  // thing — leave an encoder where it stands, or kill it and start another —
+  // and both used to be learned only from runs that had ENDED, so at a cold
+  // open they were zero. Zero does not read as "not measured": it reads as
+  // "free", and a free move is always taken. Field 2026-09-08: an encoder moved
+  // between two adjacent numbers every half second and produced nothing.
+  const startStopCost = transcodeAudio
+    ? await measureStartAndStop({ ffmpegBin, encoder: videoEncoder, logger })
+    : null;
+  // Whether this ffmpeg build can tone-map HDR→SDR (zscale + tonemap filters).
+  // Detected once; the session manager applies the tonemap chain only for HDR
+  // sources on the software path when available.
+  const tonemapSupported = transcodeAudio
+    ? await detectTonemapSupport({ ffmpegBin, logger })
+    : false;
+  // Where every file's keyframe table lives, and the only thing that reads one.
+  // Built here rather than inside the session manager because it is not a fact
+  // about a session: the playback planner warms it while a file is being opened
+  // and no session exists yet, and a session created later is handed the very
+  // object that warm read filled in.
+  // Subtitles are two layers put together, and this is the one place that knows
+  // about both: what a file STATES about its subtitle tracks and how a cue is
+  // written out is the media layer's, while which clusters may be read, where
+  // the cursor stands and what one walk at a time means are the TORRENT's own
+  // rules. The media layer used to import the second directly.
+  /**
+   * One file of one torrent, reduced to what the subtitle walk may know: a
+   * name, a length, which ranges are downloaded whole, and how to read one of
+   * those without asking the swarm. Piece length, file offsets and the bitfield
+   * stay in the torrent's thread.
+   *
+   * @param {string} sourceKey
+   * @param {number} fileIndex
+   * @param {object} [known] - The torrent handle when the caller already has it.
+   * @returns {Promise<import("./services/media/SubtitleCues.js").HeldFile | null>}
+   */
+  const heldFileFor = async (sourceKey, fileIndex, known = null) => {
+    let torrent = known;
+    if (!torrent) {
+      const record = sourceRegistry.get(sourceKey);
+      if (!record) {
+        return null;
+      }
+      torrent = await torrentPool.getTorrent(record.sourceType, record.source);
+    }
+    const file = torrent?.files?.[fileIndex];
+    if (!file || !(file.length > 0)) {
+      return null;
+    }
+    return {
+      sourceKey,
+      fileIndex,
+      name: String(file.name ?? ""),
+      length: file.length,
+      heldRanges: () => torrentPool.heldRangesOf(torrent, fileIndex),
+      readHeld: (start, end) => torrentPool.readHeldOf(torrent, fileIndex, start, end)
+    };
+  };
+  /** @type {Set<string>} */
+  const walksInFlight = new Set();
+  /**
+   * Walk whatever new cues a file now holds and push them to the viewers.
+   *
+   * @param {string} sourceKey
+   * @param {number} fileIndex
+   * @returns {Promise<void>}
+   */
+  const pushFreshCues = async (sourceKey, fileIndex) => {
+    // A pass that arrives while the previous one is still walking is dropped
+    // rather than queued: `verified` fires per piece, so on a fast download
+    // these arrive many times a second, and a queue of identical passes would
+    // only postpone the one that has something new to find. The walk is
+    // serialized per file inside `SubtitleCues` anyway.
+    const key = `${sourceKey}:${fileIndex}`;
+    if (walksInFlight.has(key)) {
+      return;
+    }
+    walksInFlight.add(key);
+    try {
+      const file = await heldFileFor(sourceKey, fileIndex);
+      if (!file) {
+        return;
+      }
+      for (const entry of await warmSubtitleCues(file)) {
+        const span = entry.spanStartSeconds === null
+          ? "empty"
+          : `${entry.spanStartSeconds.toFixed(1)}-${entry.spanEndSeconds.toFixed(1)}s`;
+        logger.info(
+          `subtitle push ${sourceKey.slice(0, 8)}:${fileIndex} track ${entry.trackIndex}: ` +
+          `${entry.cues.length} new cue(s) covering ${span}, ` +
+          `clusters walked ${entry.walkedClusters}/${entry.indexedClusters}, cursor ${entry.cursor}`
+        );
+        onSubtitleCues?.({ sourceKey, fileIndex, ...entry });
+      }
+    } catch (error) {
+      logger.warn(`subtitle push ${sourceKey.slice(0, 8)}:${fileIndex} failed: ${error?.message ?? error}`);
+    } finally {
+      walksInFlight.delete(key);
+    }
+  };
+  const subtitles = new SubtitleOrchestrator(containerOrchestrator, {
+    warm: async (torrent, fileIndex, sourceKey) => {
+      const file = await heldFileFor(sourceKey, fileIndex, torrent);
+      return file ? warmSubtitleCues(file) : [];
+    },
+    held: async (torrent, fileIndex, sourceKey, trackNumber) => {
+      const file = await heldFileFor(sourceKey, fileIndex, torrent);
+      return file ? cuesHeldFor(file, trackNumber) : null;
+    },
+    tracksOf: async (torrent, fileIndex, sourceKey) => {
+      const file = await heldFileFor(sourceKey, fileIndex, torrent);
+      return file ? subtitleTracksOf(file) : [];
+    },
+    declaredTracksOf: async (torrent, fileIndex, sourceKey) => {
+      const file = await heldFileFor(sourceKey, fileIndex, torrent);
+      return file ? declaredSubtitleTracksOf(file) : [];
+    },
+    forget: forgetSubtitles
+  });
+  // WHAT A FILE STATES ABOUT ITSELF IS READ HERE, ON THIS THREAD.
+  //
+  // The container layer is built from one function — `readRange(start, end)` —
+  // and the bytes behind it live in shared memory, so the parse has no reason
+  // to happen anywhere else. It used to happen in the torrent thread for one
+  // stated reason, that "the main thread cannot open a read stream on one of
+  // its files": true of WebTorrent's own API, and not of the bytes. The price
+  // was three commands, a second `ContainerOrchestrator` in that thread, and
+  // every answer carried back across the channel — which is what made "where is
+  // this fact kept" a question at all.
+  //
+  // The edges are fetched first because the file this is asked about is usually
+  // one nobody has played: a sidecar soundtrack is asked about before anyone
+  // has chosen it, and a Matroska file's Cues sit at the END behind a SeekHead
+  // in the head, so a read that has neither waits for the swarm twice over.
+  const containerOver = async ({ sourceKey, fileIndex, tailBytes = 0 }) => {
+    const record = sourceRegistry.get(sourceKey);
+    if (!record) {
+      return null;
+    }
+    const torrent = await torrentPool.getTorrent(record.sourceType, record.source);
+    const file = torrent?.files?.[fileIndex];
+    if (!file || !(file.length > 0)) {
+      return null;
+    }
+    try {
+      await torrentPool.prefetchFileEdges(torrent, fileIndex, {
+        headBytes: CONTAINER_HEAD_BYTES,
+        tailBytes,
+        timeoutMs: 60_000
+      });
+    } catch {
+      // A prefetch that failed is not a reason to skip the read: the read
+      // fetches what it needs itself, only more slowly.
+    }
+    return {
+      sourceKey,
+      fileIndex,
+      readRange: (start, end) =>
+        torrentPool.readRangeOf(torrent, fileIndex, start, Math.min(end, file.length - 1)),
+      fileSize: file.length,
+      label: String(file.name ?? "")
+    };
+  };
+  const keyframeTables = new KeyframeTables({
+    // Read by the same container that answers the track table and the media
+    // info, from the same header, and now in the same thread as the session
+    // that is waiting for it.
+    readTable: async ({ sourceKey, fileIndex }) => {
+      // Both edges: the Cues of a Matroska file are at the end.
+      const params = await containerOver({ sourceKey, fileIndex, tailBytes: CONTAINER_HEAD_BYTES });
+      if (!params) {
+        return null;
+      }
+      const index = await containerOrchestrator.getKeyframeIndex(params);
+      return {
+        times: index?.times ?? null,
+        tolerance: index?.tolerance ?? 0,
+        // Which container answered, whether or not it produced a table: the
+        // refusal that follows names it, and a measurement of how often an
+        // index disagrees with its own file cannot be read without it.
+        format: (await containerOrchestrator.getContainer(params))?.formatName ?? "unrecognised"
+      };
+    }
+  });
+  const outputParts = wireOutputs({
+    enabled: transcodeAudio,
+    keyframeTables,
+    ffmpegBin,
+    localBindHost: host,
+    localPort: selectedPort,
+    videoEncoder,
+    calibration,
+    decodeCostModel,
+    contentionPenalties,
+    copySpeedX,
+    startStopCost,
+    tonemapSupported,
+    segmentFormatId: segmentFormat,
+    stateDir,
+    diagnostics,
+    // Live download stats accessor for the realtime budget: lets it tell a
+    // CPU-bound transcode from a download-starved input before downscaling.
+    // What every torrent here has moved, so the proxy can price its own
+    // downloading, hashing and delivery against the machine (roadmap item 7).
+    // What the spilled pieces weigh on the torrent thread, and how to tell them
+    // their share of the disk. The owner of the disk is on this side, where the
+    // segments are; the pieces are on the other.
+    spillDisk: typeof torrentPool.allowSpillBytes === "function"
+      ? {
+          held: () => torrentPool.spilledBytes ?? 0,
+          allow: (bytes) => torrentPool.allowSpillBytes(bytes)
+        }
+      : null,
+    // Files downloaded whole, kept as files. They live on the torrent thread
+    // too, and until 2026-09-14 they had no bound of any kind — a 2.8 GB film
+    // on a host whose disk is often a 32 GB card.
+    wholeFiles: typeof torrentPool.allowWholeFileBytes === "function"
+      ? {
+          held: () => torrentPool.wholeFileBytes ?? 0,
+          allow: (bytes) => torrentPool.allowWholeFileBytes(bytes),
+          root: completedFilesRoot()
+        }
+      : null,
+    // The pieces held IN memory, on the torrent thread. Told their share by the
+    // same owner that divides the disk: what does not fit here is spilled
+    // there, so the two cannot be divided apart.
+    memoryClaimant: typeof torrentPool.allowMemoryBytes === "function"
+      ? {
+          held: () => torrentPool.memoryClaim?.held ?? 0,
+          wanted: () => torrentPool.memoryClaim?.wanted ?? 0,
+          allow: (bytes) => torrentPool.allowMemoryBytes(bytes)
+        }
+      : null,
+    diagnosticsRoot: stateDir || "",
+    budgetPolicy,
+    getTorrentTotals: async () => {
+      if (typeof torrentPool.getTorrentTotals !== "function") {
+        return null;
+      }
+      return torrentPool.getTorrentTotals();
+    },
+    // The priority map, on its way to the downloading. It lives in another
+    // thread, so the map crosses the worker channel; what it does with it —
+    // seconds into bytes, what to ask the swarm for, what to keep in memory —
+    // is its own business.
+    setPriorityMap: async ({ sourceKey, fileIndex, durationSeconds, zones }) => {
+      const record = sourceRegistry.get(sourceKey);
+      if (!record) {
+        return;
+      }
+      try {
+        await torrentPool.setPriorityMap({ sourceKey, fileIndex, durationSeconds, zones });
+      } catch {
+        // Best effort: the map is republished on the next change, and the
+        // downloading goes on serving reads meanwhile.
+      }
+    },
+    getSourceStats: async (sourceKey, fileIndex) => {
+      const record = sourceRegistry.get(sourceKey);
+      if (!record) {
+        return null;
+      }
+      try {
+        const torrent = await torrentPool.getTorrent(record.sourceType, record.source);
+        // Awaited for the same reason as the stats route: this now crosses a
+        // thread boundary and returns a promise.
+        return await torrentPool.getFileStats(torrent, Number.isInteger(fileIndex) ? fileIndex : null);
+      } catch {
+        return null;
+      }
+    },
+    // Reuse the media info the planner already probed for this file (same
+    // ffmpeg scan) so createSession skips its own probe. Late-bound: invoked
+    // only at session-create time, after playbackPlanner is initialised.
+    getCachedMediaInfo: (params) => playbackPlanner.getCachedMediaInfo(params),
+    // The file's audio tracks, for the master playlist's rendition group. Already
+    // probed for the browser's audio menu; read from there rather than probed again.
+    getCachedAudioTracks: (params) => playbackPlanner.getCachedAudioTracks(params),
+    // What a file declares about itself, read by the container layer from the
+    // same header its track table comes from. This is how the session learns
+    // where a soundtrack shipped as its own file begins — it used to spawn an
+    // ffmpeg over this proxy's own HTTP to ask the same question of the same
+    // bytes, and that read cost 8.1 s of every cold start (field 2026-09-03).
+    getContainerMediaInfo: async ({ sourceKey, fileIndex }) => {
+      try {
+        const params = await containerOver({ sourceKey, fileIndex });
+        return params ? await containerOrchestrator.getMediaInfo(params) : null;
+      } catch {
+        return null;
+      }
+    },
+    // Pull one whole file onto the disk. Used for a soundtrack that ships beside
+    // the picture, once the encoder is as far ahead of the viewer as it is
+    // allowed to get — the one moment the swarm's capacity is demonstrably
+    // spare. A bounded read of the file's whole length, NOT `file.select()`:
+    // selecting a file alongside the readers' own windows is what made a seek
+    // wait 93 s while the swarm fetched 2.47 GB in file order (see
+    // `#syncSelections` in `torrent/torrent-pool.js`).
+    fetchWholeFile: async ({ sourceKey, fileIndex }) => {
+      const record = sourceRegistry.get(sourceKey);
+      if (!record) {
+        return;
+      }
+      const torrent = await torrentPool.getTorrent(record.sourceType, record.source);
+      // The same background fill the warm-up starts when a file is chosen, not a
+      // second way of doing it. It is guarded against running twice on one file,
+      // so the two triggers converge instead of putting two readers on the same
+      // soundtrack — and only one of them would have stood aside for the
+      // picture. This trigger remains for the session that never had a warm-up
+      // before it.
+      await torrentPool.fillFileInBackground?.(torrent, fileIndex);
+    }
+  });
+  // What the last life of this process left on the disk. The kernel kills this
+  // one often enough for that to be an ordinary state rather than an odd one —
+  // twice in a single viewing on 2026-09-02 — and when it does, no exit handler
+  // runs and nothing is cleared up. So this is both the cleanup and the only
+  // record that those encoders ended at all: it says what it found before it
+  // decides anything, keeps the segments whose closure is proven, and removes
+  // the one piece per output that was being written when the process died.
+  outputParts.lifecycle.adoptSegmentsLeftBehind();
+  const playbackPlanner = createPlaybackPlanner({
+    ffmpegBin,
+    transcodeAudioEnabled: transcodeAudio,
+    localBaseUrl: outputParts.localBaseUrl,
+    sourceRegistry,
+    torrentPool,
+    sidecarsFromTorrent: (torrent, fileIndex) => contentsOf(torrent).sidecarsOf(fileIndex),
+    // What a file declares about its own tracks, parsed on this thread from the
+    // header the swarm delivered. The planner used to ask the torrent pool for
+    // this, which meant the media layer asking the torrent layer to parse a
+    // container on its behalf, in the other thread, with the answer carried
+    // back over the channel.
+    declaredTracksOf: async ({ sourceKey, fileIndex }) => {
+      const params = await containerOver({ sourceKey, fileIndex });
+      return params ? await containerOrchestrator.getTracks(params) : [];
+    },
+    warmKeyframeIndex: (params) => keyframeTables.warm(params),
+    expectedFirstSegmentMs: () => outputParts.hostTimings.expectedFirstSegmentMs(),
+    expectedSessionCreateMs: () => outputParts.hostTimings.expectedSessionCreateMs(),
+    // The quality menu is on screen from the moment a file is opened, so the
+    // heights this host can actually serve have to be answerable before any
+    // encoder exists — from the probe and the startup benchmarks alone.
+    predictOfferedHeights: (mediaInfo) => outputParts.quality.predictOfferedHeights(mediaInfo)
+  });
+
+  app.get("/health", async (req, reply) => handleHealthGet(req, reply, { version }));
+  app.get("/healthz", async (req, reply) => handleHealthzGet(req, reply, { version }));
+  // Off unless --delivery-sink was given; see the route for why it exists.
+  app.get("/api/delivery-sink", async (req, reply) =>
+    handleApiDeliverySinkGet(req, reply, { enabled: deliverySink === true })
+  );
+  // Bytes a browser can time its link with before any film has been chosen.
+  app.get("/api/link-probe", async (req, reply) => handleApiLinkProbeGet(req, reply));
+  app.post("/api/sources", async (req, reply) =>
+    handleApiSourcesPost(req, reply, { sourceRegistry })
+  );
+  app.get("/api/sources/:sourceKey/stats", async (req, reply) =>
+    handleApiSourceStatsGet(req, reply, { sourceRegistry, torrentPool })
+  );
+  app.get("/api/sources/:sourceKey/files", async (req, reply) =>
+    handleApiSourceFilesGet(req, reply, { sourceRegistry, torrentPool })
+  );
+  app.post("/api/sources/:sourceKey/warm", async (req, reply) =>
+    handleApiSourceWarmPost(req, reply, {
+      sourceRegistry,
+      torrentPool,
+      playbackPlanner,
+      // How long the file runs, read on this thread from the header. The warm
+      // turns a position in seconds into a byte offset and needs it; it used to
+      // read the container itself, in the torrent thread, to find out.
+      durationOf: async (params) => {
+        const over = await containerOver(params);
+        return over ? (await containerOrchestrator.getMediaInfo(over))?.durationSeconds ?? null : null;
+      }
+    })
+  );
+  // The browser's own log, kept beside the proxy's. See the route's own file
+  // for why it is here and not only on the registry server.
+  const clientLogs = logFile ? createClientLogFiles(logFile) : null;
+  app.post("/api/client-logs", async (req, reply) =>
+    handleApiClientLogsPost(req, reply, { clientLogs })
+  );
+  app.post("/api/playback-plan", async (req, reply) =>
+    handleApiPlaybackPlanPost(req, reply, { playbackPlanner })
+  );
+  app.post("/api/playback-plan/audio-tracks", async (req, reply) =>
+    handleApiPlaybackPlanAudioTracksPost(req, reply, { playbackPlanner })
+  );
+  app.get("/api/subtitles", async (req, reply) =>
+    handleApiSubtitlesGet(req, reply, {
+      sourceRegistry,
+      torrentPool,
+      ffmpegBin,
+      localBaseUrl: outputParts.localBaseUrl,
+      viewers: outputParts.viewers,
+      subtitles
+    })
+  );
+  app.get("/stream", async (req, reply) =>
+    handleStreamGet(req, reply, {
+      sourceRegistry,
+      torrentPool,
+      // So a session that has produced nothing yet can still show it is being
+      // fed. The route knows only files; the session id rides on the URL the
+      // session itself built.
+      noteInputBytes: (sessionId, bytes) => outputParts.encodeRuns.noteInputBytes(sessionId, bytes)
+    })
+  );
+  app.post("/api/transcode-sessions", async (req, reply) =>
+    handleApiTranscodeSessionsPost(req, reply, { viewerRequests: outputParts.viewerRequests, renditions: outputParts.renditions, quality: outputParts.quality, outputs: outputParts.outputs, lookaheadSeconds: outputParts.lookaheadSeconds, sourceRegistry, torrentPool })
+  );
+  app.post("/api/transcode-sessions/:sessionId/release", async (req, reply) =>
+    handleApiTranscodeSessionReleasePost(req, reply, { lifecycle: outputParts.lifecycle })
+  );
+  app.get("/api/transcode-sessions/:sessionId/progress", async (req, reply) =>
+    handleApiTranscodeSessionsProgressGet(req, reply, { viewerRequests: outputParts.viewerRequests })
+  );
+  app.post("/api/transcode-sessions/:sessionId/net-report", async (req, reply) =>
+    // A viewer's statement about itself goes to the VIEWER layer, not through
+    // the session manager: what it needs is the live sessions and the registry
+    // of viewers, and nothing about encoding.
+    handleApiTranscodeSessionNetReportPost(req, reply, {
+      outputs: outputParts.outputs,
+      viewers: outputParts.viewers,
+      renditions: outputParts.renditions,
+      quality: outputParts.quality
+    })
+  );
+  app.post("/api/transcode-sessions/:sessionId/fragment-far", async (req, reply) =>
+    handleApiTranscodeSessionFragmentFarPost(req, reply, { serving: outputParts.serving })
+  );
+  app.post("/api/transcode-sessions/:sessionId/seek", async (req, reply) =>
+    handleApiTranscodeSessionSeekPost(req, reply, { viewerRequests: outputParts.viewerRequests })
+  );
+  app.get("/transcode/:sessionId/:fileName", async (req, reply) =>
+    handleTranscodeSessionFileGet(req, reply, { serving: outputParts.serving, viewerRequests: outputParts.viewerRequests })
+  );
+  // A quality variant's files. Registered before the static handler for the
+  // same reason as the line above, and kept a separate route rather than a
+  // wildcard so the height stays a parsed parameter.
+  // Registered BEFORE the variant file route: `warm` is not a file name, and
+  // Fastify matches a static segment ahead of a parameter either way — stated
+  // here so the order is not "tidied" into a bug.
+  app.get("/transcode/:sessionId/v/:height/warm", async (req, reply) =>
+    handleTranscodeVariantWarmGet(req, reply, { renditions: outputParts.renditions, serving: outputParts.serving, viewerRequests: outputParts.viewerRequests })
+  );
+  app.get("/transcode/:sessionId/a/:track/warm", async (req, reply) =>
+    handleTranscodeAudioWarmGet(req, reply, { renditions: outputParts.renditions, serving: outputParts.serving, viewerRequests: outputParts.viewerRequests })
+  );
+  app.get("/transcode/:sessionId/a/:trackIndex/:fileName", async (req, reply) =>
+    handleTranscodeAudioFileGet(req, reply, { renditions: outputParts.renditions, serving: outputParts.serving, viewerRequests: outputParts.viewerRequests })
+  );
+  app.get("/transcode/:sessionId/v/:height/:fileName", async (req, reply) =>
+    handleTranscodeVariantFileGet(req, reply, { renditions: outputParts.renditions, serving: outputParts.serving, viewerRequests: outputParts.viewerRequests })
+  );
+  await app.register(fastifyStatic, {
+    root: publicRoot,
+    prefix: "/",
+    serveDotFiles: true
+  });
+
+  app.addHook("onClose", async () => {
+    // Order matters: stop the ffmpeg readers (HLS sessions) before destroying
+    // the torrents whose files they read from, then remove the torrent data.
+    await outputParts.lifecycle.disposeAll();
+    await torrentPool.destroyAll();
+  });
+
+  await app.listen({ host, port: selectedPort });
+  return {
+    app,
+    port: selectedPort,
+    // Asked over the tunnel when the proxy a viewer landed on has refused their
+    // file: could THIS host sustain it? Answered from the startup benchmarks
+    // and a description, so it needs no torrent and costs milliseconds.
+    outputParts,
+    // The browser only ever knows a source by its REGISTRY key (a hash of the
+    // raw request bytes, scoped to one API session) — never the torrent
+    // pool's own key (the content's infohash, shared across a magnet and a
+    // `.torrent` naming the same film). The subtitle push subscription is
+    // recorded from a browser request and published from the pool's side, so
+    // resolving one into the other is what lets the two ends agree on what
+    // they are both calling "sourceKey".
+    sourceRegistry,
+    // Which films this host holds, for the health poll the browser makes before
+    // it picks a proxy. The pool is on the worker thread and this is the way to
+    // it from the process that answers that poll.
+    torrentPool
+  };
+}