@torrent-tv/proxy 2.87.4 → 2.87.6

package/services/encode/hwaccel.js

@@ -1,1894 +1,1894 @@
-/**
- * @file Hardware-accelerated H.264 encoder auto-detection.
- *
- * Probes the ffmpeg build and the host for a usable hardware H.264 encoder
- * (NVENC / QSV / VAAPI / V4L2 M2M), verifying each candidate with a real
- * test-encode before selecting it. Falls back to software libx264 when no
- * hardware encoder is present or working.
- *
- * Deployment-agnostic: relies only on ffmpeg, the filesystem and
- * `process.platform`; makes no assumptions about Home Assistant or any
- * specific host. A garbled or unsupported hardware path simply fails its
- * test-encode and is skipped, so the worst case is software encoding.
- *
- * A descriptor exposes:
- *   - `name`        human-readable encoder id (e.g. "h264_vaapi")
- *   - `kind`        "software" | "vaapi" | "qsv" | "nvenc" | "v4l2m2m"
- *   - `device`      device node path or null
- *   - `inputArgs`   ffmpeg args inserted before `-i` (decode/hwaccel setup)
- *   - `buildVideoArgs({ targetWidth, targetHeight, segmentDurationSec })`
- *                   ffmpeg video filter + encoder args inserted after `-map`s
- */
-
-import { spawn } from "node:child_process";
-import { mkdtempSync, readdirSync, rmSync, statSync } from "node:fs";
-import { mkdtemp, readFile, rm } from "node:fs/promises";
-import os from "node:os";
-import path from "node:path";
-import { fitDecodeCost } from "./decode-cost-fit.js";
-import { penaltiesFrom } from "./contention.js";
-import { fileURLToPath } from "node:url";
-import {
-  parseFfmpegBitrateKbps,
-  parseFfmpegDurationSeconds,
-  parseFfmpegVideoDimensions,
-  parseFfmpegVideoFps
-} from "../media/ffmpeg-banner.js";
-
-import { keyFrameArgs, TRANSCODE_FPS } from "./args.js";
-import { throughputAt } from "./throughput.js";
-import { LADDER_HEIGHTS } from "./output/ladder.js";
-// The five kinds, one class each. Detection and benchmarking stay in this file;
-// how a kind is driven belongs to the kind.
-import {
-  NvencEncoder,
-  QsvEncoder,
-  SoftwareEncoder,
-  V4l2m2mEncoder,
-  VaapiEncoder
-} from "./index.js";
-// Re-exported so every caller goes on importing these figures from here:
-// the same calculation, moved to sit beside the encoder kinds built from it.
-export {
-  chooseOutputFps,
-  maxrateKbpsFor,
-  limitRowFor,
-  nominalKbpsFor,
-  TRANSCODE_FPS
-} from "./args.js";
-
-const BENCHMARK_REF_W = 640;
-const BENCHMARK_REF_H = 360;
-const BENCHMARK_DURATION_SEC = 3;
-/**
- * How many raw frames a throughput reading at this size is taken over.
- *
- * The reference size keeps the three seconds it has always had. A larger frame
- * gets as many frames as the same number of bytes buys, but never fewer than
- * one second of film: the frames are looped, and a loop shorter than the
- * encoder's reach — its reference frames and its B-frames — would let it find
- * an identical frame to predict from and read far faster than any film.
- *
- * @param {{ width: number, height: number }} frame
- * @returns {number}
- */
-export function rawFrameCountFor(frame) {
-  const referenceBytes = BENCHMARK_REF_W * BENCHMARK_REF_H * 1.5 * TRANSCODE_FPS * BENCHMARK_DURATION_SEC;
-  const frameBytes = frame.width * frame.height * 1.5;
-  return Math.max(TRANSCODE_FPS, Math.floor(referenceBytes / frameBytes));
-}
-/**
- * The narrowest window a slope may be taken over. Measured 2026-08-15: at a
- * fifth of a second the readings were noisy enough to put `faster` and
- * `veryfast` BELOW `fast`, which libx264 cannot do — and `pickSoftwarePreset`
- * walks the list assuming it ascends. Half a second was still noisy enough for that
- * (measured again: veryfast below faster, twice), so a full second it is —
- * about six seconds of startup for a ladder the whole budget then rests on.
- */
-const ENCODE_BENCHMARK_WINDOW_SEC = 1;
-/** The narrowest window that may be used when a run ends early. */
-const ENCODE_BENCHMARK_MIN_WINDOW_SEC = 0.2;
-/** Above this a reading is a fault, not a fast machine. */
-const ENCODE_BENCHMARK_MAX_PLAUSIBLE_SPEED = 1000;
-/**
- * A preset that has not reported twice in this long is hung, not slow: reports
- * arrive twice a second whatever the encoding speed.
- */
-const ENCODE_BENCHMARK_TIMEOUT_MS = 10_000;
-/**
- * How many times the calibration clip is joined to itself to measure copying.
- *
- * Not a figure about the machine: it is how much film the reading needs to have
- * in front of it. A copy runs at hundreds of times realtime, and the slope is
- * taken over a window of one second, so the input has to hold more film than the
- * fastest plausible host gets through in that second. Forty laps of a five-second
- * clip is 200 s of film, which covers the ceiling `slopeOf` will accept.
- */
-/** Progress reports arrive line by line. */
-const NEWLINE = String.fromCharCode(10);
-// Producing one second of video per second of clock. Not a margin and not a
-// choice — the definition of keeping up, and the bar when nothing better is
-// known about the supply this step will meet.
-const REALTIME = 1;
-// The bar where decoding CANNOT be priced — no calibration fit, or a source the
-// probe said too little about. This one is not measured and cannot be: the
-// prediction it guards counts encoding only, which on the field host was
-// several times optimistic, and there is no reading on such a host to correct
-// it with. It is left at the figure it has had since before decoding was
-// priced, because lowering it to realtime would make the least-measured hosts
-// the most permissive. Where decoding IS priced, nothing chosen remains.
-const UNPRICED_DECODE_BAR = 1.8;
-
-
-// The five kinds live in `encode/`, one class each, and these keep the names
-// every caller already uses. A kind states its own arguments and its own
-// ladder of speed settings; detection and benchmarking stay here.
-/** @returns {import("./hwaccel.js").VideoEncoderDescriptor} */
-export function softwareDescriptor() {
-  return new SoftwareEncoder();
-}
-
-/**
- * @param {string} device
- * @returns {import("./hwaccel.js").VideoEncoderDescriptor}
- */
-function vaapiDescriptor(device) {
-  return new VaapiEncoder(device);
-}
-
-/**
- * @param {string} device
- * @returns {import("./hwaccel.js").VideoEncoderDescriptor}
- */
-function qsvDescriptor(device) {
-  return new QsvEncoder(device);
-}
-
-/** @returns {import("./hwaccel.js").VideoEncoderDescriptor} */
-function nvencDescriptor() {
-  return new NvencEncoder();
-}
-
-/** @returns {import("./hwaccel.js").VideoEncoderDescriptor} */
-function v4l2m2mDescriptor() {
-  return new V4l2m2mEncoder();
-}
-
-
-
-/**
- * @typedef {Object} VideoEncoderDescriptor
- * @property {string} name
- * @property {"software"|"vaapi"|"qsv"|"nvenc"|"v4l2m2m"} kind
- * @property {string|null} device
- * @property {string[]} inputArgs
- * @property {(opts: { targetWidth: number, targetHeight: number, segmentDurationSec: number, preset?: string, fps?: number, tonemap?: boolean, forcedKeyframeTimes?: number[] | null, rateControl?: { maxrateKbps: number, bufsizeKbps: number, level: string | null } | null }) => string[]} buildVideoArgs
- */
-
-/**
- * Run ffmpeg and resolve with its exit code and captured output.
- *
- * @param {string} ffmpegBin
- * @param {string[]} args
- * @param {number} [timeoutMs=12000]
- * @returns {Promise<{ code: number, stdout: string, stderr: string }>}
- */
-export function runFfmpeg(ffmpegBin, args, timeoutMs = 12000) {
-  return new Promise((resolve) => {
-    let stdout = "";
-    let stderr = "";
-    let settled = false;
-    let child;
-    const finish = (code) => {
-      if (settled) {
-        return;
-      }
-      settled = true;
-      resolve({ code, stdout, stderr });
-    };
-    try {
-      child = spawn(ffmpegBin, args, { stdio: ["ignore", "pipe", "pipe"], windowsHide: true });
-    } catch {
-      finish(-1);
-      return;
-    }
-    const timer = setTimeout(() => {
-      try {
-        child.kill("SIGKILL");
-      } catch {
-        // already gone
-      }
-      finish(-1);
-    }, timeoutMs);
-    child.stdout.on("data", (chunk) => {
-      stdout += String(chunk);
-    });
-    child.stderr.on("data", (d) => {
-      stderr += String(d);
-    });
-    child.on("error", () => {
-      clearTimeout(timer);
-      finish(-1);
-    });
-    child.on("exit", (code) => {
-      clearTimeout(timer);
-      finish(code ?? -1);
-    });
-  });
-}
-
-/** @returns {string[]} /dev/dri/renderD* nodes (VAAPI/QSV). */
-function listRenderNodes() {
-  try {
-    return readdirSync("/dev/dri")
-      .filter((n) => n.startsWith("renderD"))
-      .map((n) => `/dev/dri/${n}`)
-      .sort();
-  } catch {
-    return [];
-  }
-}
-
-/** @returns {boolean} Whether any /dev/nvidia* node exists (NVENC). */
-function hasNvidiaDevice() {
-  try {
-    return readdirSync("/dev").some((n) => /^nvidia(\d+)?$/.test(n));
-  } catch {
-    return false;
-  }
-}
-
-/** @returns {boolean} Whether any /dev/video* node exists (V4L2 M2M). */
-function hasV4l2Device() {
-  try {
-    return readdirSync("/dev").some((n) => /^video\d+$/.test(n));
-  } catch {
-    return false;
-  }
-}
-
-/**
- * Build a full ffmpeg command that encodes a short, *moving* synthetic clip
- * (testsrc2 — far more representative than a static black frame) through the
- * candidate encoder into real HLS segments in `outDir`, with keyframes forced
- * on segment boundaries. Verifying the resulting segments (see
- * {@link verifySegmentsDecodeCleanly}) catches encoders that silently produce
- * a corrupted or non-IDR-aligned stream (e.g. some V4L2 M2M builds).
- *
- * @param {VideoEncoderDescriptor} descriptor
- * @param {number} segmentDurationSec
- * @param {string} outDir
- * @returns {string[]}
- */
-function buildEncoderTestArgs(descriptor, segmentDurationSec, outDir) {
-  const durationSec = Math.max(8, segmentDurationSec * 3);
-  const source = ["-f", "lavfi", "-i", `testsrc2=s=640x360:r=${TRANSCODE_FPS}:d=${durationSec}`];
-  const kf = keyFrameArgs(segmentDurationSec);
-
-  /** @type {string[]} */
-  let pre = ["-hide_banner", "-loglevel", "error"];
-  /** @type {string[]} */
-  let encode;
-  switch (descriptor.kind) {
-    case "vaapi":
-      pre = [...pre, "-vaapi_device", String(descriptor.device)];
-      encode = ["-vf", "format=nv12,hwupload", "-c:v", "h264_vaapi", "-qp", "24", ...kf];
-      break;
-    case "qsv":
-      pre = [...pre, "-qsv_device", String(descriptor.device)];
-      encode = ["-vf", "hwupload=extra_hw_frames=16,format=qsv", "-c:v", "h264_qsv", "-global_quality", "24", ...kf];
-      break;
-    case "nvenc":
-      encode = ["-c:v", "h264_nvenc", "-preset", "p4", "-cq", "24", "-pix_fmt", "yuv420p", ...kf];
-      break;
-    case "v4l2m2m":
-      encode = ["-pix_fmt", "yuv420p", "-c:v", "h264_v4l2m2m", "-num_capture_buffers", "32", "-b:v", "3M", "-g", String(TRANSCODE_FPS * segmentDurationSec), ...kf];
-      break;
-    default:
-      encode = ["-c:v", "libx264", "-preset", "ultrafast", "-pix_fmt", "yuv420p", ...kf];
-      break;
-  }
-
-  const hlsOut = [
-    "-f", "hls",
-    "-hls_time", String(segmentDurationSec),
-    "-hls_list_size", "0",
-    "-hls_flags", "independent_segments",
-    // fMP4 (CMAF) — matches the runtime pipeline (`encode/run-command.js`).
-    "-hls_segment_type", "fmp4",
-    "-hls_fmp4_init_filename", "init.mp4",
-    "-hls_segment_filename", path.join(outDir, "seg-%03d.m4s"),
-    path.join(outDir, "index.m3u8")
-  ];
-  return [...pre, ...source, ...encode, ...hlsOut];
-}
-
-/**
- * Verify the HLS segments produced by the test encode are valid: at least two
- * segments exist, and each decodes standalone without errors. A segment that
- * does not begin with a keyframe (broken/corrupted output) emits decode errors
- * when read on its own, which fails this check.
- *
- * @param {string} ffmpegBin
- * @param {string} outDir
- * @returns {Promise<boolean>}
- */
-export async function verifySegmentsDecodeCleanly(ffmpegBin, outDir) {
-  let files;
-  try {
-    files = readdirSync(outDir).filter((n) => /^seg-\d+\.m4s$/.test(n));
-  } catch {
-    return false;
-  }
-  if (files.length < 2) {
-    return false;
-  }
-  // fMP4: parameter sets (SPS/PPS) live in init.mp4, not in each segment.
-  // Decode the whole playlist (ffmpeg's own, which references init.mp4 via
-  // #EXT-X-MAP), so every segment is exercised together with the init. Any
-  // corrupt / non-conformant segment (e.g. some V4L2 M2M builds emit a stray
-  // no-picture access unit) surfaces as a decode error here.
-  const result = await runFfmpeg(
-    ffmpegBin,
-    ["-hide_banner", "-loglevel", "error", "-i", path.join(outDir, "index.m3u8"), "-f", "null", "-"],
-    12000
-  );
-  return result.code === 0 && result.stderr.trim().length === 0;
-}
-
-/**
- * Detect the best usable H.264 encoder. Always resolves (falls back to
- * software libx264). Each hardware candidate is verified with a real
- * test-encode before being selected.
- *
- * @param {{ ffmpegBin: string, logger?: { info: (m: string) => void, warn: (m: string) => void }, segmentDurationSec?: number }} options
- * @returns {Promise<VideoEncoderDescriptor>}
- */
-export async function detectVideoEncoder({ ffmpegBin, logger, segmentDurationSec = 4 }) {
-  const log = logger ?? { info: () => {}, warn: () => {} };
-  const software = softwareDescriptor();
-
-  const { code, stdout } = await runFfmpeg(ffmpegBin, ["-hide_banner", "-encoders"], 10000);
-  if (code !== 0) {
-    log.warn("hwaccel: could not list ffmpeg encoders; using software libx264");
-    return software;
-  }
-  const has = (name) => stdout.includes(name);
-
-  /** @type {VideoEncoderDescriptor[]} */
-  const candidates = [];
-  const renderNodes = listRenderNodes();
-  if (has("h264_nvenc") && hasNvidiaDevice()) {
-    candidates.push(nvencDescriptor());
-  }
-  if (has("h264_qsv") && renderNodes.length > 0) {
-    candidates.push(qsvDescriptor(renderNodes[0]));
-  }
-  if (has("h264_vaapi") && renderNodes.length > 0) {
-    candidates.push(vaapiDescriptor(renderNodes[0]));
-  }
-  // h264_v4l2m2m (ARM SoC / Raspberry Pi / HA Yellow). It is gated behind the
-  // strict keyframe-alignment test below, because some V4L2 M2M builds silently
-  // emit a corrupted / non-IDR-aligned stream; the test rejects those and the
-  // host falls back to software libx264.
-  if (has("h264_v4l2m2m") && hasV4l2Device()) {
-    candidates.push(v4l2m2mDescriptor());
-  }
-
-  for (const candidate of candidates) {
-    const dir = mkdtempSync(path.join(os.tmpdir(), "tt-hwtest-"));
-    let ok = false;
-    try {
-      const encoded = await runFfmpeg(
-        ffmpegBin,
-        buildEncoderTestArgs(candidate, segmentDurationSec, dir),
-        25000
-      );
-      if (encoded.code === 0) {
-        ok = await verifySegmentsDecodeCleanly(ffmpegBin, dir);
-      }
-    } finally {
-      try {
-        rmSync(dir, { recursive: true, force: true });
-      } catch {
-        // best effort
-      }
-    }
-    if (ok) {
-      log.info(
-        `hwaccel: using hardware encoder ${candidate.name}` +
-          `${candidate.device ? ` (${candidate.device})` : ""}`
-      );
-      return candidate;
-    }
-    log.warn(`hwaccel: ${candidate.name} failed the HLS keyframe-alignment test; skipping`);
-  }
-
-  log.info("hwaccel: no working hardware encoder; using software libx264");
-  return software;
-}
-
-/**
- * Detect whether this ffmpeg build has the filters needed for the HDR→SDR
- * tone-map chain (`zscale`, from libzimg, and `tonemap`). Both are required;
- * when either is missing, HDR sources are re-encoded without tone mapping
- * (washed-out but playable). Always resolves.
- *
- * @param {{ ffmpegBin: string, logger?: { info: (m: string) => void, warn: (m: string) => void } }} options
- * @returns {Promise<boolean>}
- */
-export async function detectTonemapSupport({ ffmpegBin, logger }) {
-  const log = logger ?? { info: () => {}, warn: () => {} };
-  const { code, stdout } = await runFfmpeg(ffmpegBin, ["-hide_banner", "-filters"], 10000);
-  if (code !== 0) {
-    log.warn("hwaccel: could not list ffmpeg filters; HDR tone mapping disabled");
-    return false;
-  }
-  // `-filters` prints one filter per line: "... zscale  ...", "... tonemap ...".
-  const hasZscale = /\bzscale\b/.test(stdout);
-  const hasTonemap = /\btonemap\b/.test(stdout);
-  const supported = hasZscale && hasTonemap;
-  log.info(
-    `hwaccel: HDR tone mapping ${supported ? "available" : "unavailable"} ` +
-      `(zscale=${hasZscale} tonemap=${hasTonemap})`
-  );
-  return supported;
-}
-
-
-// The clips the decode cost is fitted from. They ship with the package
-// (`assets/calibration/`), cut from Netflix Open Content "Meridian" (CC-BY 4.0)
-// — real, grainy live action, because a generated `testsrc2` clip decodes 158 %
-// away from a real film where these are 11 % away (measured 2026-08-14).
-//
-// Three sizes at two bitrates each, with the axes varied INDEPENDENTLY. The set
-// this replaced was three clips for three unknowns, two of them at the same
-// size: an exact system, which cannot fail visibly. On 2026-08-17 it returned
-// `0.007542 × Mpx/s + 0.000000 × Mbit/s + 0.0000 s/s` — the bitrate term and
-// the constant exactly zero — and the prediction on top of it was 1.8-2.2x
-// optimistic. Six points leave three spare, so the fit has a residual, and a
-// term the data does not determine can be refused instead of published as a
-// zero that looks measured. See `assets/calibration/NOTICE.md`.
-//
-// One set PER CODEC FAMILY, because a family is what the model describes. The
-// fit used to be H.264 only, while a video that has to be RE-ENCODED is by
-// definition one the browser could not play — which is to say HEVC, 10-bit or
-// AV1 — and those decode dearer per pixel on the same box. Pricing them with
-// H.264 constants is the one case the model is always asked about and was never
-// measured on.
-//
-// A family that has no set of its own is priced with H.264's, which is what
-// happened to every family before this; the line says so rather than implying
-// it. AV1 has no set yet: the survey of 2026-07-10 found it rare where HEVC was
-// 18 % of releases, so it waits for the same treatment.
-const CALIBRATION_SETS = {
-  h264: [
-    "cal-h264-1080-hi.mp4",
-    "cal-h264-1080-lo.mp4",
-    "cal-h264-720-hi.mp4",
-    "cal-h264-720-lo.mp4",
-    "cal-h264-480-hi.mp4",
-    "cal-h264-480-lo.mp4"
-  ],
-  hevc: [
-    "cal-hevc-1080-hi.mp4",
-    "cal-hevc-1080-lo.mp4",
-    "cal-hevc-480-hi.mp4",
-    "cal-hevc-480-lo.mp4"
-  ],
-  hevc10: [
-    "cal-hevc10-1080-hi.mp4",
-    "cal-hevc10-1080-lo.mp4",
-    "cal-hevc10-480-hi.mp4",
-    "cal-hevc10-480-lo.mp4"
-  ]
-};
-export const CALIBRATION_REFERENCE_CLIP = "cal-h264-1080-hi.mp4";
-export const CALIBRATION_DIR = path.join(path.dirname(fileURLToPath(import.meta.url)), "..", "..", "assets", "calibration");
-// How wide the measured window must be before the slope is trusted, and how
-// long to wait for it at most.
-//
-// Half a second, and it is the TIMING noise that sets it rather than the amount
-// of video: the slope is output time against wall time, both read from the same
-// two progress lines, and the jitter in stamping one is milliseconds — so half
-// a second of window is a fraction of a percent of error on any host. What used
-// to make a longer window necessary was the clip restarting inside it, and that
-// is gone: the stream is continuous now. Measured 2026-08-22 against the
-// continuous-pass truth on a desktop: -3.0 % and +0.6 % at half a second, with
-// the readings spread 2-6 %, against -25 % and -33 % for the loop it replaces.
-// Half a second also costs the startup about 0.6 s per clip less, which matters
-// because every clip of every codec family is paid for before any viewer
-// exists.
-const DECODE_WINDOW_MIN_SEC = 0.5;
-const DECODE_WINDOW_MAX_MS = 8000;
-
-/**
- * Read what a calibration clip IS from the decode run's own output: the
- * dimensions, the frame rate and the bitrate ffmpeg reports for it. Read rather
- * than declared, so replacing a clip cannot silently invalidate the fit.
- *
- * @param {string} stderr
- * @returns {{ megapixelsPerSecond: number, megabitsPerSecond: number, durationSeconds: number } | null}
- */
-function parseClipCharacteristics(stderr) {
-  // The same readers the session manager uses on the same banner — one parser
-  // per fact, so a second copy cannot drift from the first.
-  const { width, height } = parseFfmpegVideoDimensions(stderr);
-  const rate = parseFfmpegVideoFps(stderr);
-  const seconds = parseFfmpegDurationSeconds(stderr);
-  const kbps = parseFfmpegBitrateKbps(stderr);
-  if (!(width > 0) || !(height > 0) || !(rate > 0) || !(seconds > 0) || !(kbps > 0)) {
-    return null;
-  }
-  return {
-    megapixelsPerSecond: (width * height * rate) / 1e6,
-    megabitsPerSecond: kbps / 1000,
-    durationSeconds: seconds
-  };
-}
-
-/**
- * Measure what DECODING costs on this host, as seconds of work per second of
- * video, and solve it into three host constants:
- *
- *   decodeCost = a × Mpixel/s + b × Mbit/s + c
- *
- * Why it exists: the preset benchmark below measures ENCODING only, and a
- * re-encode pays for both halves. Measured 2026-08-14 on the addon host, that
- * omission made the budget offer a 240p rung it then ran at 0.39-0.95× — the
- * benchmark said the host cleared the bar 2.5× over. With the decode term the
- * same file predicts within 4.8 %; without it the error on that rung was 209 %.
- *
- * The constants are properties of the HOST, so this runs once at startup (about
- * 5 s on a CM4) and any source is then priced from figures the probe already
- * has — nothing is added to a session's cold start.
- *
- * They are also properties of the CODEC, and the clips are H.264: HEVC, AV1 and
- * 10-bit decode dearer per pixel on the same machine, and a source that has to
- * be re-encoded is by definition one this browser could not play, which is
- * usually not H.264. So the fit is optimistic exactly there. Closing that needs
- * clips in those codecs, and is its own roadmap item.
- *
- * @param {{ ffmpegBin: string, logger?: { info: (m: string) => void, warn: (m: string) => void }, clipsDir?: string }} options
- * @returns {Promise<{ pixelTerm: number, bitrateTerm: number, constantTerm: number } | null>}
- */
-/**
- * What a second job costs on this host, measured rather than assumed.
- *
- * The budget adds seconds of work per second of content — this encode, plus
- * that decode, plus what is already committed — and the addon host contradicted
- * that directly on 2026-08-18: decoding ran at 2.10-2.25x alone, 0.79-0.90x
- * with one encoder beside it and 0.56-0.64x with two. The same work costs 2.6×
- * more for having company. Heat is not the cause (the hot idle machine was the
- * fastest reading of all); four cores sharing one path to memory is.
- *
- * So it is measured the way everything else here is: the same clip decoded
- * alone, then decoded again while an encoder of the same clip runs beside it.
- * The ratio is the penalty. The encoder is stopped as soon as the reading is
- * taken, and the whole thing costs one decode plus one short encode.
- *
- * @param {{ ffmpegBin: string, logger?: { info: (m: string) => void, warn: (m: string) => void }, clipsDir?: string, upTo?: number }} options
- * @returns {Promise<Map<number, number> | null>} Penalties by how many other
- *   jobs were running, or null when the readings could not be taken.
- */
-export async function benchmarkContention({ ffmpegBin, logger, clipsDir = CALIBRATION_DIR, upTo = 2 }) {
-  const log = logger ?? { info: () => {}, warn: () => {} };
-  // The cheapest clip in the set: this measures the MACHINE's behaviour under
-  // company, not the clip's own cost, so the smallest one says it soonest.
-  const clip = path.join(clipsDir, "cal-h264-480-lo.mp4");
-  const startedAt = Date.now();
-  // Lifted once and decoded three times from the same bytes. Going through
-  // `measureDecodeSlope` lifted it again for every reading — three process
-  // starts on a path that is awaited before the proxy's tunnel opens, for a
-  // remux whose result had not changed.
-  const streams = await extractFamilyStreams(ffmpegBin, [clip], "h264");
-  const stream = streams?.[0];
-  if (!stream) {
-    log.warn("hwaccel: contention could not be measured; costs will be added as though jobs were independent");
-    return null;
-  }
-  const alone = await decodePipedStream(ffmpegBin, stream, log);
-  if (!alone?.speed) {
-    log.warn("hwaccel: contention could not be measured; costs will be added as though jobs were independent");
-    return null;
-  }
-  /** @type {Array<{ others: number, speed: number }>} */
-  const beside = [];
-  /** @type {import("node:child_process").ChildProcess[]} */
-  const load = [];
-  try {
-    for (let others = 1; others <= Math.max(1, upTo); others += 1) {
-      load.push(
-        spawn(
-          ffmpegBin,
-          [
-            "-hide_banner", "-loglevel", "error", "-nostats",
-            "-stream_loop", "-1", "-i", clip,
-            "-an", "-c:v", "libx264", "-preset", "fast", "-f", "null", "-"
-          ],
-          { stdio: ["ignore", "ignore", "ignore"], windowsHide: true }
-        )
-      );
-      // Let the encoder reach its own speed before reading anything: an encode
-      // measured in its first moments is measuring the process starting.
-      await new Promise((resolve) => {
-        setTimeout(resolve, 2_000);
-      });
-      const withCompany = await decodePipedStream(ffmpegBin, stream, log);
-      if (withCompany?.speed) {
-        beside.push({ others, speed: withCompany.speed });
-      }
-    }
-  } finally {
-    for (const child of load) {
-      try {
-        child.kill("SIGKILL");
-      } catch {
-        // Already gone: the reading is what mattered, and nothing else uses it.
-      }
-    }
-  }
-  const penalties = penaltiesFrom(alone.speed, beside);
-  if (!penalties) {
-    log.warn("hwaccel: contention readings said nothing; costs will be added as though jobs were independent");
-    return null;
-  }
-  log.info(
-    `hwaccel: a second job costs ${[...penalties.entries()]
-      .map(([others, penalty]) => `${penalty.toFixed(2)}x beside ${others}`)
-      .join(", ")} ` +
-    `(decode alone ${alone.speed.toFixed(2)}x, ` +
-    `${beside.map((reading) => `${reading.speed.toFixed(2)}x beside ${reading.others}`).join(", ")}, ` +
-    `measured in ${((Date.now() - startedAt) / 1000).toFixed(1)}s)`
-  );
-  return penalties;
-}
-
-export async function benchmarkDecodeCost({ ffmpegBin, logger, clipsDir = CALIBRATION_DIR }) {
-  const log = logger ?? { info: () => {}, warn: () => {} };
-  const startedAllAt = Date.now();
-  /** @type {Record<string, { pixelTerm: number, bitrateTerm: number, constantTerm: number }>} */
-  const families = {};
-  for (const [family, clips] of Object.entries(CALIBRATION_SETS)) {
-    const fitted = await fitOneFamily({ ffmpegBin, log, clipsDir, family, clips });
-    if (fitted) {
-      families[family] = fitted;
-    }
-  }
-  if (!families.h264) {
-    // H.264 is the family every other one falls back to, so without it there is
-    // no model at all rather than a partial one. Which families DID fit is said
-    // anyway: on a fast host the H.264 clips decode at 20-80x and the readings
-    // stop being ordered — measured 2026-08-20 on a desktop, 1080p at 9.35
-    // Mbit/s costing 0.0307 s/s against 720p at 9.94 costing 0.0472, which is
-    // not a thing a decoder does — so a failure here is a measurement problem
-    // and not a missing file, and the line has to let those be told apart.
-    log.warn(
-      "hwaccel: decode cost unknown — the H.264 clips did not fit" +
-        (Object.keys(families).length > 0
-          ? `, though ${Object.keys(families).join(" and ")} did`
-          : "")
-    );
-    return null;
-  }
-  const missing = Object.keys(CALIBRATION_SETS).filter((family) => !families[family]);
-  log.info(
-    `hwaccel: decode cost measured for ${Object.keys(families).join(", ")}` +
-      (missing.length > 0 ? `; ${missing.join(" and ")} priced as H.264` : "") +
-      ` (in ${((Date.now() - startedAllAt) / 1000).toFixed(1)}s)`
-  );
-  return { families, ...families.h264 };
-}
-
-/**
- * Fit one codec family's decode cost from its own clips.
- *
- * @param {{ ffmpegBin: string, log: { info: Function, warn: Function }, clipsDir: string, family: string, clips: string[] }} params
- * @returns {Promise<{ pixelTerm: number, bitrateTerm: number, constantTerm: number } | null>}
- */
-async function fitOneFamily({ ffmpegBin, log, clipsDir, family, clips }) {
-  const startedAllAt = Date.now();
-  /** @type {Array<{ megapixelsPerSecond: number, megabitsPerSecond: number, costSecondsPerSecond: number }>} */
-  const samples = [];
-  // Every clip of the family is lifted out of its container FIRST, in one
-  // ffmpeg run. See `extractFamilyStreams` for why one run rather than one per
-  // clip, and why before the measurements rather than beside them.
-  const streams = await extractFamilyStreams(
-    ffmpegBin,
-    clips.map((clip) => path.join(clipsDir, clip)),
-    family
-  );
-  if (!streams) {
-    log.warn(
-      `hwaccel: ${family} cannot be lifted out of its container — no Annex-B filter is mapped for it, ` +
-        `so its clips were never measured`
-    );
-    return null;
-  }
-  for (const [index, clip] of clips.entries()) {
-    const stream = streams[index];
-    const measured = stream ? await decodePipedStream(ffmpegBin, stream, log) : null;
-    if (!measured?.speed) {
-      log.warn(
-        `hwaccel: decode benchmark "${clip}" said nothing; ${family} not measured` +
-          (measured?.error ? ` — ${measured.error}` : " — the clip could not be lifted out of its container")
-      );
-      return null;
-    }
-    const cost = 1 / measured.speed;
-    samples.push({
-      megapixelsPerSecond: measured.megapixelsPerSecond,
-      megabitsPerSecond: measured.megabitsPerSecond,
-      costSecondsPerSecond: cost
-    });
-    log.info(
-      `hwaccel: decode "${clip}" ${measured.megapixelsPerSecond.toFixed(1)} Mpx/s ` +
-        `${measured.megabitsPerSecond.toFixed(2)} Mbit/s -> ${measured.speed.toFixed(1)}x ` +
-        `(cost ${cost.toFixed(4)} s/s, over ${measured.windowSec.toFixed(1)}s of decoding)`
-    );
-  }
-  const fitted = fitDecodeCost(samples);
-  if (!fitted) {
-    log.warn(`hwaccel: ${family} decode cost could not be fitted to these measurements`);
-    return null;
-  }
-  log.info(
-    `hwaccel: ${family} decode cost = ${fitted.pixelTerm.toFixed(6)} × Mpx/s + ${fitted.bitrateTerm.toFixed(6)} × Mbit/s ` +
-      `+ ${fitted.constantTerm.toFixed(4)} s/s (${fitted.shape} from ${fitted.samples} clips, ` +
-      `typical disagreement ${fitted.residualRms.toFixed(4)} s/s` +
-      // Named rather than implied: a zero in the line above means "not
-      // measured" for a dropped term and "measured to be nothing" otherwise,
-      // and those are different claims.
-      (fitted.dropped.length > 0 ? `, ${fitted.dropped.join(" and ")} not determined by these clips` : "") +
-      `, measured in ${((Date.now() - startedAllAt) / 1000).toFixed(1)}s)`
-  );
-  return { pixelTerm: fitted.pixelTerm, bitrateTerm: fitted.bitrateTerm, constantTerm: fitted.constantTerm };
-}
-
-/**
- * The bitstream filter and demuxer that turn a clip's video track into a
- * continuous elementary stream, by codec family.
- *
- * H.264 and HEVC in MP4 keep their parameter sets in the container's `avcC` /
- * `hvcC` and their access units length-prefixed; Annex-B carries them inline,
- * with start codes, which is what makes plain byte concatenation a valid
- * stream. That is the property this whole measurement rests on.
- */
-const ANNEX_B_BY_FAMILY = {
-  h264: { filter: "h264_mp4toannexb", demuxer: "h264" },
-  hevc: { filter: "hevc_mp4toannexb", demuxer: "hevc" },
-  hevc10: { filter: "hevc_mp4toannexb", demuxer: "hevc" }
-};
-
-/**
- * The last complaint in an ffmpeg stderr, for a line that has to say why.
- *
- * @param {string} stderr
- * @returns {string}
- */
-function lastErrorLine(stderr) {
-  const lines = String(stderr ?? "")
-    .split(/\r?\n/)
-    .map((line) => line.trim())
-    .filter((line) => line.length > 0);
-  return lines[lines.length - 1] ?? "";
-}
-
-/**
- * How long the lift may take before it is abandoned. It is a remux of a few
- * megabytes, so this is not a budget — it is the difference between a startup
- * that reports a failure and one that never finishes. Every other ffmpeg run in
- * this file has such a bound; this one did not, and it is awaited before the
- * proxy's tunnel opens.
- */
-const EXTRACT_TIMEOUT_MS = 20_000;
-
-/**
- * Lift a whole family's clips out of their containers, as Annex-B elementary
- * streams, in ONE ffmpeg run.
- *
- * No re-encoding — the frames are copied — so the work itself is trivial and
- * the cost is almost entirely the process. Doing one process per clip added
- * 11 s to the startup here (fourteen clips at about 0.83 s each), and running
- * them concurrently did not help: six at once took 4.75 s against 0.89 s for
- * one, so the machine serialises them. One run with many inputs and many
- * outputs costs one process.
- *
- * The outputs go to temporary files because several outputs cannot share one
- * pipe; they are read into memory and deleted immediately, and nothing about
- * this measurement is kept between runs.
- *
- * Before the measurements, never beside them: a remux running next to a decode
- * is a second job on the machine, and this benchmark exists to find out what
- * ONE job costs here.
- *
- * @param {string} ffmpegBin
- * @param {string[]} clipPaths
- * @param {string} family
- * @returns {Promise<Array<{ bytes: Buffer, demuxer: string, megapixelsPerSecond: number, megabitsPerSecond: number, fps: number } | null> | null>}
- *   One entry per clip, in order; null when the family cannot be lifted at all.
- */
-async function extractFamilyStreams(ffmpegBin, clipPaths, family) {
-  const shape = ANNEX_B_BY_FAMILY[family];
-  // A family with no mapping is a hard failure, not a silent fallback to
-  // H.264's filter. AV1 has no Annex-B form at all (its packaging is OBU), and
-  // MPEG-2 and VC-1 have no `*_mp4toannexb` filter — so the three families the
-  // roadmap plans next cannot come through here, and finding that out as
-  // "the clip failed" would send the reader after the clip.
-  if (!shape) {
-    return null;
-  }
-  const workDir = await mkdtemp(path.join(os.tmpdir(), "ttv-calibration-"));
-  const outputs = clipPaths.map((_, index) => path.join(workDir, `stream-${index}.${shape.demuxer}`));
-  /** @type {string[]} */
-  const args = ["-hide_banner", "-loglevel", "info", "-nostats", "-y"];
-  for (const clipPath of clipPaths) {
-    args.push("-i", clipPath);
-  }
-  for (const [index, output] of outputs.entries()) {
-    args.push("-map", `${index}:v:0`, "-c:v", "copy", "-bsf:v", shape.filter, "-f", shape.demuxer, output);
-  }
-  const stderr = await runCapturingStderr(ffmpegBin, args, EXTRACT_TIMEOUT_MS);
-  try {
-    if (stderr === null) {
-      return null;
-    }
-    // One banner block per input, in the order they were given. Read rather
-    // than declared, so replacing a clip cannot silently invalidate the fit
-    // that rests on it.
-    const blocks = splitInputBlocks(stderr, clipPaths.length);
-    return await Promise.all(clipPaths.map(async (_, index) => {
-      const block = blocks[index];
-      if (!block) {
-        return null;
-      }
-      const clipInfo = parseClipCharacteristics(block);
-      const fps = parseFfmpegVideoFps(block);
-      if (!clipInfo || !(fps > 0)) {
-        return null;
-      }
-      let bytes;
-      try {
-        bytes = await readFile(outputs[index]);
-      } catch {
-        return null;
-      }
-      if (bytes.length === 0) {
-        return null;
-      }
-      return {
-        bytes,
-        demuxer: shape.demuxer,
-        megapixelsPerSecond: clipInfo.megapixelsPerSecond,
-        megabitsPerSecond: clipInfo.megabitsPerSecond,
-        fps
-      };
-    }));
-  } finally {
-    await rm(workDir, { recursive: true, force: true }).catch(() => {});
-  }
-}
-
-/**
- * The part of an ffmpeg banner describing each input, in order.
- *
- * ffmpeg prints one `Input #N, …` block per input and then the stream mapping;
- * the parsers here read a single input's facts, so they are given a single
- * input's text rather than the whole banner.
- *
- * @param {string} stderr
- * @param {number} count
- * @returns {string[]}
- */
-function splitInputBlocks(stderr, count) {
-  /** @type {string[]} */
-  const blocks = [];
-  for (let index = 0; index < count; index += 1) {
-    const from = stderr.indexOf(`Input #${index},`);
-    if (from < 0) {
-      blocks.push("");
-      continue;
-    }
-    const nextInput = stderr.indexOf(`Input #${index + 1},`, from);
-    const mapping = stderr.indexOf("Stream mapping:", from);
-    const ends = [nextInput, mapping].filter((at) => at > from);
-    blocks.push(stderr.slice(from, ends.length > 0 ? Math.min(...ends) : stderr.length));
-  }
-  return blocks;
-}
-
-/**
- * Run ffmpeg to completion and return its stderr, or null when it failed or
- * outlasted its bound.
- *
- * @param {string} ffmpegBin
- * @param {string[]} args
- * @param {number} timeoutMs
- * @returns {Promise<string | null>}
- */
-function runCapturingStderr(ffmpegBin, args, timeoutMs) {
-  return new Promise((resolve) => {
-    let stderr = "";
-    let settled = false;
-    let child;
-    const settle = (value) => {
-      if (settled) {
-        return;
-      }
-      settled = true;
-      clearTimeout(timer);
-      try {
-        child?.kill("SIGKILL");
-      } catch {
-        // already gone
-      }
-      resolve(value);
-    };
-    const timer = setTimeout(() => settle(null), timeoutMs);
-    try {
-      child = spawn(ffmpegBin, args, { stdio: ["ignore", "ignore", "pipe"], windowsHide: true });
-    } catch {
-      settle(null);
-      return;
-    }
-    child.stderr.on("data", (chunk) => {
-      stderr += String(chunk);
-    });
-    child.on("error", () => settle(null));
-    child.on("close", (code) => settle(code === 0 ? stderr : null));
-  });
-}
-
-/**
- * Measure how fast this host DECODES a clip, from ffmpeg's own report of how
- * much video it has processed.
- *
- * Two things are deliberately outside the measurement.
- *
- * **The process starting.** Wall-clock around the process cannot answer this:
- * starting ffmpeg costs about a second, and on a quick machine a five-second
- * clip decodes in a tenth of that, so the measurement would be of the program
- * starting. Progress lines arrive AFTER it has started, and the slope between
- * two of them — video processed against time taken — contains no part of the
- * startup by construction.
- *
- * **The clip restarting.** This used to loop the clip with `-stream_loop -1`,
- * and a loop is not free: measured 2026-08-22 on a desktop, a restart costs
- * 0.03 s on the 480p clip and 0.12 s on the 1080p one — the decoder tearing
- * down and re-allocating its frame buffers, which is why the price rises with
- * the picture. A five-second clip decoded at 55x restarts eleven times a
- * second, so that cost DOMINATED the reading: the same clips measured 53.7x
- * looped against 80.3x in one continuous pass, and 11.8x against 15.8x. Worse,
- * the bias is not shared — it depends on the clip's own resolution and on how
- * fast the host is — so it does not cancel out of the fit, it tilts it. That is
- * the fast-host failure recorded on 2026-08-20, where 1080p read cheaper than
- * 720p, which is not a thing a decoder does.
- *
- * So the clip is fed to the decoder as ONE stream instead. An Annex-B
- * elementary stream carries its parameter sets inline, so writing the same
- * bytes again is simply more stream — the decoder never re-initialises, and
- * there is no restart inside the window to measure. Verified against the
- * continuous-pass truth on the same host: -0.2 % and -5.5 %, against -25 % and
- * -33 % for the loop. Nothing is written to disk and the process is killed as
- * soon as the window is wide enough.
- *
- * Exported because the property that broke here is checkable and was not being
- * checked: a bigger picture must cost more than a smaller one of the same
- * bitrate, and under the loop it did not.
- *
- * @param {string} ffmpegBin
- * @param {string} clipPath
- * @param {string} [family="h264"]
- * @returns {Promise<{ speed: number, windowSec: number, megapixelsPerSecond: number, megabitsPerSecond: number } | null>}
- */
-export async function measureDecodeSlope(ffmpegBin, clipPath, family = "h264") {
-  const streams = await extractFamilyStreams(ffmpegBin, [clipPath], family);
-  const stream = streams?.[0];
-  if (!stream) {
-    return null;
-  }
-  const measured = await decodePipedStream(ffmpegBin, stream);
-  return measured?.speed ? measured : null;
-}
-
-/**
- * Decode an elementary stream fed from memory, and report the slope.
- *
- * @param {string} ffmpegBin
- * @param {{ bytes: Buffer, demuxer: string, megapixelsPerSecond: number, megabitsPerSecond: number, fps: number }} stream
- * @returns {Promise<{ speed: number, windowSec: number, megapixelsPerSecond: number, megabitsPerSecond: number } | null>}
- */
-function decodePipedStream(ffmpegBin, stream, log = { info: () => {}, warn: () => {} }) {
-  return new Promise((resolve) => {
-    const args = [
-      "-hide_banner", "-loglevel", "error", "-nostats",
-      // A raw stream states no frame rate, so the one the container declared is
-      // given back to it. It decides how output time advances, and therefore
-      // what "seconds of video per second of clock" means.
-      "-f", stream.demuxer, "-framerate", String(stream.fps), "-i", "pipe:0",
-      "-an", "-f", "null", "-",
-      "-progress", "pipe:1"
-    ];
-    /** @type {Array<{ wallSec: number, outSec: number }>} */
-    const samples = [];
-    let stdout = "";
-    // Kept because this path depends on three things the old one did not: the
-    // raw demuxer accepting the frame rate, the bitstream filter having
-    // produced something parsable, and the fed concatenation being decodable.
-    // Without it the only trace of any of those failing is "said nothing".
-    let stderr = "";
-    let settled = false;
-    let child;
-    const startedAt = Date.now();
-    // Whether the FEED, not the decoder, could be what this reading measures
-    // (item 4(d2)). `child.stdin.write()` returning false was tried as the
-    // signal — never once true would mean this process was never ahead of the
-    // pipe — and measured false on every reading taken while writing this,
-    // including clips this same host decodes at 15-80x with room to spare, so
-    // it does not discriminate: `write()`'s return value tracks Node's own
-    // internal watermark against the size of what was just handed to it, not
-    // real drain state, and answered "no slack" identically whether the pipe
-    // or the decoder was the true limit. Rather than publish a verdict that
-    // reads the same in both cases, only the byte count is kept, for the
-    // MB/s figure below — a number to read, not a boolean to trust.
-    let bytesWritten = 0;
-    const finish = () => {
-      if (settled) {
-        return;
-      }
-      settled = true;
-      clearTimeout(timer);
-      try {
-        child?.stdin?.destroy();
-      } catch {
-        // already gone
-      }
-      try {
-        child?.kill("SIGKILL");
-      } catch {
-        // already gone
-      }
-      // The first sample still carries the startup — it reports whatever was
-      // processed while the process was coming up. Everything is measured from
-      // the second onwards.
-      const first = samples[1];
-      const last = samples[samples.length - 1];
-      if (!first || !last) {
-        resolve({ error: lastErrorLine(stderr) || "the decoder reported no progress" });
-        return;
-      }
-      const windowSec = last.wallSec - first.wallSec;
-      const producedSec = last.outSec - first.outSec;
-      if (!(windowSec >= DECODE_WINDOW_MIN_SEC) || !(producedSec > 0)) {
-        resolve({ error: lastErrorLine(stderr) || `the window was ${windowSec.toFixed(2)}s of ${producedSec.toFixed(2)}s produced` });
-        return;
-      }
-      const speed = producedSec / windowSec;
-      // Diagnostic only — logged, not acted on. Both figures are taken over the
-      // SAME window the speed itself is: bytesWritten is snapshotted alongside
-      // every progress sample, so this compares like against like rather than
-      // the achieved rate over the whole run (which starts before the first
-      // kept sample and reads systematically low against the window's own
-      // rate for no reason but that mismatch — measured while building this).
-      const windowBytes = last.bytesWritten - first.bytesWritten;
-      const achievedMBps = windowSec > 0 ? windowBytes / windowSec / 1e6 : 0;
-      const requiredMBps = ((stream.megabitsPerSecond * 1e6) / 8) * speed / 1e6;
-      // "Far apart" is stated, not left to the reader to eyeball: outside a
-      // factor of 1.5 either way is bigger than the write-timing slop this
-      // comparison carries on a healthy reading.
-      const farApart = requiredMBps > 0 && (achievedMBps / requiredMBps < 1 / 1.5 || achievedMBps / requiredMBps > 1.5);
-      log.info(
-        `hwaccel: decode pipe fed ${achievedMBps.toFixed(1)} MB/s, ${requiredMBps.toFixed(1)} MB/s ` +
-        `needed for ${speed.toFixed(2)}x` +
-        (farApart ? " — far enough apart to be worth a second look" : "")
-      );
-      resolve({
-        speed,
-        windowSec,
-        megapixelsPerSecond: stream.megapixelsPerSecond,
-        megabitsPerSecond: stream.megabitsPerSecond,
-        pipeThroughputMBps: achievedMBps
-      });
-    };
-    const timer = setTimeout(finish, DECODE_WINDOW_MAX_MS);
-    try {
-      child = spawn(ffmpegBin, args, { stdio: ["pipe", "pipe", "pipe"], windowsHide: true });
-    } catch (error) {
-      clearTimeout(timer);
-      settled = true;
-      resolve({ error: error instanceof Error ? error.message : String(error) });
-      return;
-    }
-    child.stderr.on("data", (chunk) => {
-      stderr += String(chunk);
-    });
-    // Keep the decoder fed. `write` returning false means the pipe is full, and
-    // the next copy goes on the `drain` — so the decoder is never starved and
-    // this process never buffers more than the pipe holds.
-    const writeOnce = () => {
-      const accepted = child.stdin.write(stream.bytes);
-      bytesWritten += stream.bytes.length;
-      return accepted;
-    };
-    const feed = () => {
-      while (!settled && child.stdin.writable && writeOnce()) {
-        // Written straight through; go round again.
-      }
-    };
-    child.stdin.on("drain", feed);
-    // The kill closes the pipe under the writer; that is the intended end.
-    child.stdin.on("error", () => {});
-    child.stdout.on("data", (chunk) => {
-      stdout += String(chunk);
-      let newline = stdout.indexOf("\n");
-      while (newline >= 0) {
-        const line = stdout.slice(0, newline).trim();
-        stdout = stdout.slice(newline + 1);
-        if (line.startsWith("out_time_ms=")) {
-          const microseconds = Number(line.slice("out_time_ms=".length));
-          if (Number.isFinite(microseconds)) {
-            samples.push({ wallSec: (Date.now() - startedAt) / 1000, outSec: microseconds / 1e6, bytesWritten });
-          }
-        }
-        newline = stdout.indexOf("\n");
-      }
-      if (samples.length >= 2 && samples[samples.length - 1].wallSec - samples[1].wallSec >= DECODE_WINDOW_MIN_SEC) {
-        finish();
-      }
-    });
-    child.on("error", (error) => {
-      if (settled) {
-        return;
-      }
-      clearTimeout(timer);
-      settled = true;
-      try {
-        child?.stdin?.destroy();
-      } catch {
-        // already gone
-      }
-      try {
-        child?.kill("SIGKILL");
-      } catch {
-        // already gone
-      }
-      resolve({ error: error instanceof Error ? error.message : String(error) });
-    });
-    child.on("close", finish);
-    feed();
-  });
-}
-
-
-/**
- * How many times realtime this host can DECODE a source of these
- * characteristics, from the startup fit. `null` when the fit is unavailable or
- * the source figures are not known.
- *
- * @param {{ pixelTerm: number, bitrateTerm: number, constantTerm: number } | null} model
- * @param {{ megapixelsPerSecond: number, megabitsPerSecond: number }} source
- * @returns {number | null}
- */
-export function decodeSpeedFor(model, source) {
-  if (!model) {
-    return null;
-  }
-  const pixels = Number(source?.megapixelsPerSecond);
-  const bits = Number(source?.megabitsPerSecond);
-  if (!Number.isFinite(pixels) || pixels <= 0 || !Number.isFinite(bits) || bits < 0) {
-    return null;
-  }
-  const cost = model.pixelTerm * pixels + model.bitrateTerm * bits + model.constantTerm;
-  if (!(cost > 0)) {
-    return null;
-  }
-  return 1 / cost;
-}
-
-/**
- * How many times realtime a re-encode of this source at this output pixel rate
- * would run: decoding and encoding share the machine, so their costs add and
- * their speeds combine as
- *
- *   1 / (1/decodeSpeed + 1/encodeSpeed)
- *
- * Checked 2026-08-14 on the rung that broke playback: 1/(1/2.31 + 1/5.99) =
- * 1.67× against 1.48× measured. With no decode fit this falls back to the
- * encode speed alone — which is what the budget did before, and which
- * overestimated that rung five to eleven times.
- *
- * @param {{ decodeModel: { pixelTerm: number, bitrateTerm: number, constantTerm: number } | null, encodePixelsPerSec: number, outputPixelsPerSec: number, source: { megapixelsPerSecond: number, megabitsPerSecond: number } | null }} params
- * @returns {number | null}
- */
-export function predictedRealtimeSpeed({
-  decodeModel,
-  encodePixelsPerSec,
-  outputPixelsPerSec,
-  source,
-  observedDecodeCostSec = null
-}) {
-  if (!Number.isFinite(encodePixelsPerSec) || encodePixelsPerSec <= 0) {
-    return null;
-  }
-  if (!Number.isFinite(outputPixelsPerSec) || outputPixelsPerSec <= 0) {
-    return null;
-  }
-  const encodeSpeed = encodePixelsPerSec / outputPixelsPerSec;
-  // What this very file has been seen to cost, when it has been: the clips are
-  // H.264 and a source that has to be re-encoded usually is not, so a figure
-  // taken from the encoder actually running on THIS source beats any model of
-  // a stand-in. It arrives seconds into playback and replaces the estimate.
-  const decodeSpeed = Number.isFinite(observedDecodeCostSec) && observedDecodeCostSec > 0
-    ? 1 / observedDecodeCostSec
-    : (source ? decodeSpeedFor(decodeModel, source) : null);
-  if (decodeSpeed === null) {
-    return encodeSpeed;
-  }
-  return 1 / (1 / decodeSpeed + 1 / encodeSpeed);
-}
-
-/**
- * Whether this host can hold realtime, with the margin, while re-encoding this
- * source to this output pixel rate — and the predicted speed either way, so a
- * refusal can say what it refused on.
- *
- * The encoder figure is the FASTEST benchmarked preset: it is the best this
- * host can do, so a rung it cannot hold cannot be held at any quality setting.
- *
- * @param {{ benchmark: Array<{ preset: string, pixelsPerSec: number }>, decodeModel?: object | null, source?: { megapixelsPerSecond: number, megabitsPerSecond: number } | null, outputPixelsPerSec: number, requiredSpeed?: number | null }} params
- * @returns {{ speed: number | null, sustainable: boolean }}
- */
-export function canSustainOutput({
-  benchmark,
-  decodeModel = null,
-  source = null,
-  outputPixelsPerSec,
-  observedDecodeCostSec = null,
-  concurrentCostSec = 0,
-  requiredSpeed = null,
-  // The frame this output is encoded at. Given, the encoder's figure is the
-  // one measured AT that size (`throughput.js`), and a size nothing was
-  // measured at is not sustainable: this host has not shown it can do it.
-  frame = null
-}) {
-  if (!Array.isArray(benchmark)) {
-    // Not calibrated at all: a wiring made without the startup calibration,
-    // which the proxy itself never is when it re-encodes (`server.js`). There is
-    // nothing to judge by, and nothing is judged.
-    return { speed: null, sustainable: true };
-  }
-  if (benchmark.length === 0) {
-    // CALIBRATED, and no mode of this encoder qualified (roadmap item 97, step
-    // 14). Unknown capacity is not capacity: nothing has shown this machine can
-    // hold a re-encode at all. The copied picture needs no encoder and is not
-    // asked here.
-    return { speed: null, sustainable: false };
-  }
-  const encodePixelsPerSec = cheapestPresetPixelsPerSec(benchmark, frame);
-  if (encodePixelsPerSec === null) {
-    return { speed: null, sustainable: false };
-  }
-  const observed = Number.isFinite(observedDecodeCostSec) && observedDecodeCostSec > 0
-    ? observedDecodeCostSec
-    : null;
-  if (observed === null && !isDecodePriced({ decodeModel, source })) {
-    // An encoder-only figure was several times too optimistic on the rung this
-    // check exists for, so it is not fit to refuse anything. Without the decode
-    // term the ladder is offered whole, exactly as it was before.
-    return { speed: null, sustainable: true };
-  }
-  const alone = predictedRealtimeSpeed({
-    decodeModel,
-    encodePixelsPerSec,
-    outputPixelsPerSec,
-    source,
-    observedDecodeCostSec: observed
-  });
-  // What ELSE will be running while this rung is. A rung is never the only
-  // thing on the machine: the picture it accompanies is being copied or
-  // encoded, an audio track may have its own encoder, and a warm-up is two
-  // encoders by design. Measured on the addon host, a copy alone takes about an
-  // eighth of the machine per second of video, and the field case of
-  // 2026-08-15 adds up exactly: 0.125 for the copy plus ~1.05 for the rung is
-  // more than the one second per second the machine has, which is what was
-  // observed.
-  //
-  // Zero when nothing else is known to be running, or when nothing has been
-  // measured yet — then this is a LOWER bound on the cost and the check is as
-  // permissive as it was before.
-  const speed = alone === null || !(concurrentCostSec > 0)
-    ? alone
-    : 1 / (1 / alone + concurrentCostSec);
-  if (speed === null) {
-    return { speed: null, sustainable: true };
-  }
-  return { speed, sustainable: speed >= speedBar(requiredSpeed) };
-}
-
-/**
- * The speed a step has to reach to be worth offering.
- *
- * Realtime is not enough on its own: a step that produces exactly one second
- * per second never recovers the seconds lost while its reader waits for the
- * swarm, so it survives its own supply only if what it gains between
- * interruptions covers what one interruption costs. That is measured per file
- * and per swarm by the reader — `1 + worst wait / median interval`, in
- * `torrent/supply-margin.js` — and on the field torrent of 2026-08-17 it came to 1.67
- * against the 1.5 that used to stand here, and to 4.04-8.14 on a torrent whose
- * swarm no encoder could have kept up with.
- *
- * Where that figure does not exist yet — fewer than two interruptions measured
- * — the bar is realtime. It is the one thing that can be said without
- * measuring the swarm, and the offer is restated as soon as the reader has
- * something to say.
- *
- * @param {number | null | undefined} requiredSpeed - What this file's own
- *   interruptions demand, when they have been measured.
- * @returns {number}
- */
-export function speedBar(requiredSpeed) {
-  return Number.isFinite(requiredSpeed) && requiredSpeed > REALTIME ? requiredSpeed : REALTIME;
-}
-
-/**
- * The bar for a cost description — the supply's demand where decoding is
- * priced, and never below the unpriced-decode bar where it is not.
- *
- * @param {{ decodeModel?: object | null, source?: object | null, observedDecodeCostSec?: number | null, requiredSpeed?: number | null }} cost
- * @returns {number}
- */
-function barFor(cost) {
-  const measured = speedBar(cost?.requiredSpeed);
-  return isDecodePriced(cost) ? measured : Math.max(UNPRICED_DECODE_BAR, measured);
-}
-
-/**
- * Benchmark software libx264 presets on this host. Encodes a short synthetic
- * clip at a fixed reference resolution with each preset and measures encoder
- * throughput in pixels/second. The session manager uses this to pick, per
- * stream, the highest-quality preset that still encodes the actual
- * (source-capped) resolution faster than realtime.
- *
- * Runs once at startup; bounded by a per-encode timeout. Presets that fail are
- * omitted from the result.
- *
- * @param {{ ffmpegBin: string, logger?: { info: (m: string) => void, warn: (m: string) => void } }} options
- * @returns {Promise<Array<{ preset: string, pixelsPerSec: number }>>} Ordered slowest→fastest.
- */
-export async function benchmarkSoftwarePresets({ ffmpegBin, logger, encoder = null }) {
-  const log = logger ?? { info: () => {}, warn: () => {} };
-
-  // REAL footage, decoded ONCE into raw frames, and the presets are then timed
-  // on those frames.
-  //
-  // Two reasons, both measured. The pattern this replaced (`testsrc2`) has flat
-  // areas and no grain and encodes 1.23x cheaper than film on the same machine
-  // and preset — an error that always points at offering a rung the host cannot
-  // hold. And feeding a compressed clip to each preset instead would put
-  // decoding and scaling inside the measurement: subtracting them afterwards
-  // compares a wall clock that includes process startup against a decode figure
-  // measured to exclude it, while inside one ffmpeg the two halves overlap. On
-  // the fastest preset — the one every ladder decision reads as the ceiling —
-  // that subtraction is most of the number being measured, so a small error in
-  // it becomes a large error in the answer.
-  //
-  // Raw frames remove all of it: no decoder, no scaler, nothing to subtract,
-  // and no dependence on the decode model. The cost is 25 MB of memory in a
-  // pipe for a few seconds.
-  const rawFramesPath = await decodeToRawFrames(ffmpegBin, log);
-  if (rawFramesPath === null) {
-    // Said once more, in the words that matter to whoever reads the log next:
-    // with no benchmark, `#sustainableHeights` filters nothing and every rung
-    // is offered, which is the failure of 2026-08-14 in full.
-    log.warn("hwaccel: the quality ladder is UNFILTERED on this host — nothing measured the encoder");
-    return [];
-  }
-  /** @type {Array<{ preset: string, pixelsPerSec: number }>} */
-  const results = [];
-  try {
-    // THE CHOSEN ENCODER'S OWN LADDER, whatever kind it is. NVENC walks p1…p7,
-    // QSV veryfast…veryslow, VAAPI its quality levels. A kind with no ladder is
-    // measured once, which is still a reading where there was none at all.
-    const ladder = encoder?.speedLadder;
-    const rungs = Array.isArray(ladder?.values) && ladder.values.length > 0 ? ladder.values : [null];
-    for (const rung of rungs) {
-      const speed = await measureEncodeSlope(ffmpegBin, encoder, rung, rawFramesPath);
-      if (speed === null) {
-        log.warn(`hwaccel: the benchmark of "${rung ?? encoder?.name}" produced no usable reading; skipping`);
-        continue;
-      }
-      const pixelsPerSec = BENCHMARK_REF_W * BENCHMARK_REF_H * TRANSCODE_FPS * speed;
-      results.push({ preset: rung ?? encoder?.name, pixelsPerSec });
-      log.info(
-        `hwaccel: ${encoder?.name} "${rung ?? "as it comes"}" ~= ${(pixelsPerSec / 1e6).toFixed(1)} Mpx/s ` +
-          `(${speed.toFixed(2)}x @ ${BENCHMARK_REF_W}x${BENCHMARK_REF_H}, real footage)`
-      );
-    }
-  } finally {
-    // The encoder was killed a moment ago and on Windows the handle outlives
-    // the signal, so removal is retried and its failure is not worth a session:
-    // this is a temp directory the operating system will clear anyway.
-    try {
-      rmSync(path.dirname(rawFramesPath), { recursive: true, force: true, maxRetries: 5, retryDelay: 200 });
-    } catch (error) {
-      log.warn(`hwaccel: could not remove the benchmark's raw frames: ${error instanceof Error ? error.message : String(error)}`);
-    }
-  }
-  return results;
-}
-
-/**
- * How fast one preset encodes, from ffmpeg's own reports of how much video it
- * has written — not from the clock around the process.
- *
- * Timing whole runs measures the run STARTING. Measured 2026-08-15 on a desktop
- * that spawns ffmpeg in ~0.4 s: three seconds of raw frames encoded that way
- * put `fast` and `ultrafast` within 1.24x of each other, when libx264's own
- * presets differ by several times — the constant had swallowed the difference.
- * The slope between two progress reports contains no part of the startup.
- *
- * The frames are written repeatedly so there is runway to measure over,
- * whatever the preset's speed.
- *
- * @param {string} ffmpegBin
- * @param {string} preset
- * @param {string} rawFramesPath
- * @returns {Promise<number | null>} Video seconds encoded per second of clock.
- */
-/**
- * Video seconds produced per second of clock, from ffmpeg's own reports.
- *
- * Startup is excluded by taking a DIFFERENCE: it lands in the wall clock of
- * every report equally, so it cancels between two of them. (The decode
- * benchmark drops its first report instead, because there the first one is
- * emitted at out_time zero; here reports with no time yet are discarded before
- * they arrive, so the first kept one is already running.)
- *
- * @param {Array<{ wallSec: number, outSec: number }>} samples
- * @param {number} [minimumWindowSec=ENCODE_BENCHMARK_WINDOW_SEC]
- * @returns {number | null}
- */
-export function slopeOf(samples, minimumWindowSec = ENCODE_BENCHMARK_WINDOW_SEC) {
-  const first = samples[0];
-  const last = samples[samples.length - 1];
-  if (!first || !last || first === last) {
-    return null;
-  }
-  const took = last.wallSec - first.wallSec;
-  const produced = last.outSec - first.outSec;
-  if (!(took >= minimumWindowSec) || !(produced > 0)) {
-    return null;
-  }
-  const slope = produced / took;
-  // Nothing encodes a thousand times realtime. A figure above that is a
-  // measurement fault, and letting it through opens the whole ladder.
-  return slope <= ENCODE_BENCHMARK_MAX_PLAUSIBLE_SPEED ? slope : null;
-}
-
-export function measureEncodeSlope(ffmpegBin, encoder, rung, rawFramesPath, frame = { width: BENCHMARK_REF_W, height: BENCHMARK_REF_H }) {
-  return new Promise((resolve) => {
-    const args = [
-      "-hide_banner", "-loglevel", "error", "-nostats",
-      // What a device-backed encoder needs before the input — the device. Not
-      // its decoding setup: this is fed raw frames and there is nothing to
-      // decode, and `-hwaccel vaapi` over rawvideo fails to open.
-      ...(typeof encoder?.benchmarkInputArgs === "function" ? encoder.benchmarkInputArgs() : []),
-      "-stream_loop", "-1",
-      "-f", "rawvideo", "-pix_fmt", "yuv420p",
-      "-s", `${frame.width}x${frame.height}`, "-r", String(TRANSCODE_FPS),
-      "-i", rawFramesPath,
-      // THE ENCODER SAYS HOW TO MEASURE ITSELF. It was libx264 written here, so
-      // a host with NVENC, QSV, VAAPI or V4L2M2M measured its encoder not at all
-      // and the quality offer, which is arithmetic over pixels per second, had
-      // no pixels per second to work with.
-      ...encoder.benchmarkArgs(rung),
-      "-f", "null", "-",
-      "-progress", "pipe:1"
-    ];
-    /** @type {Array<{ wallSec: number, outSec: number }>} */
-    const samples = [];
-    let settled = false;
-    let buffered = "";
-    let child;
-    const startedAt = Date.now();
-    const finish = (value) => {
-      if (settled) {
-        return;
-      }
-      settled = true;
-      clearTimeout(timer);
-      try {
-        child?.kill("SIGKILL");
-      } catch {
-        // already gone
-      }
-      resolve(value);
-    };
-    const timer = setTimeout(() => finish(null), ENCODE_BENCHMARK_TIMEOUT_MS);
-    try {
-      child = spawn(ffmpegBin, args, { stdio: ["ignore", "pipe", "ignore"], windowsHide: true });
-    } catch {
-      finish(null);
-      return;
-    }
-    // The frames come from a FILE, read on repeat by ffmpeg itself. Fed through
-    // a pipe instead, the fastest presets measured the pipe: `ultrafast` on a
-    // desktop wants raw frames at hundreds of megabytes a second, which no
-    // writer here can supply, and the reading then describes the feeding rather
-    // than the encoder.
-    child.stdout.on("data", (chunk) => {
-      buffered += String(chunk);
-      let newline = buffered.indexOf(NEWLINE);
-      while (newline >= 0) {
-        const line = buffered.slice(0, newline).trim();
-        buffered = buffered.slice(newline + 1);
-        if (line.startsWith("out_time_ms=")) {
-          const outSec = Number(line.slice("out_time_ms=".length)) / 1e6;
-          // `N/A` is not the only way ffmpeg says "no position yet": some builds
-          // print the smallest signed 64-bit integer, which IS finite and would
-          // be taken for a position nine trillion seconds before the start.
-          if (Number.isFinite(outSec) && outSec >= 0) {
-            samples.push({ wallSec: (Date.now() - startedAt) / 1000, outSec });
-          }
-        }
-        newline = buffered.indexOf(NEWLINE);
-      }
-      const slope = slopeOf(samples);
-      if (slope !== null) {
-        finish(slope);
-      }
-    });
-    child.on("error", () => finish(null));
-    // A preset that finished before the window was wide enough is measured from
-    // whatever it did report, provided two reports exist at all.
-    // A preset that finished before the wide window was covered is still
-    // measured — but never over a window of nothing. Two reports a millisecond
-    // apart would divide a frame of video by that millisecond and call the host
-    // twenty times faster than it is, and one such reading becomes the figure
-    // every ladder decision is taken from.
-    child.on("exit", () => finish(slopeOf(samples, ENCODE_BENCHMARK_MIN_WINDOW_SEC)));
-  });
-}
-
-/**
- * The benchmark's footage as raw frames: the calibration clip, looped to the
- * benchmark's length and scaled to its size, decoded once.
- *
- * @param {string} ffmpegBin
- * @param {{ info: (m: string) => void, warn: (m: string) => void }} log
- * @returns {Promise<string | null>} Path to the raw frames, or null.
- */
-export async function decodeToRawFrames(ffmpegBin, log, frame = { width: BENCHMARK_REF_W, height: BENCHMARK_REF_H }) {
-  // A benchmark may leave a host unmeasured; it may never stop it from
-  // starting. Before this the temp directory was made outside any guard, so a
-  // read-only or missing TMPDIR rejected the promise that starts the proxy.
-  let directory;
-  try {
-    directory = mkdtempSync(path.join(os.tmpdir(), "torrent-tv-bench-"));
-  } catch (error) {
-    log.warn(
-      `hwaccel: no writable temp directory for the preset benchmark (${error instanceof Error ? error.message : String(error)}); ` +
-      "presets unmeasured, so no quality rung will be refused on this host"
-    );
-    return null;
-  }
-  const rawPath = path.join(directory, "frames.yuv");
-  const args = [
-    "-hide_banner", "-loglevel", "error",
-    "-stream_loop", "-1",
-    "-i", path.join(CALIBRATION_DIR, CALIBRATION_REFERENCE_CLIP),
-    "-frames:v", String(rawFrameCountFor(frame)),
-    "-vf", `scale=${frame.width}:${frame.height},fps=${TRANSCODE_FPS}`,
-    "-an", "-f", "rawvideo", "-pix_fmt", "yuv420p", "-y", rawPath
-  ];
-  const { code } = await runFfmpeg(ffmpegBin, args, 30000);
-  const expectedBytes = frame.width * frame.height * 1.5 * rawFrameCountFor(frame);
-  let written = 0;
-  try {
-    written = statSync(rawPath).size;
-  } catch {
-    written = 0;
-  }
-  if (code !== 0 || written < expectedBytes * 0.9) {
-    log.warn(
-      "hwaccel: could not decode the calibration clip for the preset benchmark " +
-      `(${written} of ~${Math.round(expectedBytes)} bytes); presets unmeasured, ` +
-      "so no quality rung will be refused on this host"
-    );
-    try {
-      rmSync(directory, { recursive: true, force: true, maxRetries: 5, retryDelay: 200 });
-    } catch {
-      // A temp directory the operating system will clear; not worth a start-up.
-    }
-    return null;
-  }
-  return rawPath;
-}
-
-/**
- * Pick the highest-quality (slowest) benchmarked preset that can encode
- * `pixelsPerSecNeeded` with the speed margin. Falls back to the fastest
- * benchmarked preset, or `"ultrafast"` when no benchmark is available.
- *
- * @param {Array<{ preset: string, pixelsPerSec: number }>} benchmark - slowest→fastest
- * @param {number} pixelsPerSecNeeded
- * @param {{ decodeModel?: object | null, source?: { megapixelsPerSecond: number, megabitsPerSecond: number } | null, requiredSpeed?: number | null }} [cost]
- * @returns {string}
- */
-/**
- * What this host can do at its CHEAPEST preset — the ceiling of the ladder.
- *
- * Deliberately not the largest reading in the array. The list is in quality
- * order, so its last measured entry is the cheapest preset; taking the maximum
- * instead would let one noisy reading of an expensive preset raise the bar that
- * decides which rungs are offered, and a rung offered on noise is a rung the
- * host cannot hold. For choosing a preset the direction of that error is
- * harmless; for deciding what to offer it is not, so the two use different
- * statistics on purpose.
- *
- * @param {Array<{ preset: string, pixelsPerSec: number }>} benchmark
- * @returns {number}
- */
-function cheapestPresetPixelsPerSec(benchmark, frame = null) {
-  const cheapest = benchmark[benchmark.length - 1];
-  return cheapest ? throughputAt(cheapest, frame) : null;
-}
-
-export function pickSoftwarePreset(benchmark, pixelsPerSecNeeded, cost = {}, frame = null) {
-  if (!Array.isArray(benchmark) || benchmark.length === 0) {
-    return "ultrafast";
-  }
-  const observed = Number.isFinite(cost?.observedDecodeCostSec) && cost.observedDecodeCostSec > 0
-    ? cost.observedDecodeCostSec
-    : null;
-  const bar = barFor(cost);
-  // The FIRST entry that clears the bar wins — the list is in quality order, so
-  // that is the best picture this host can hold. Every entry is examined rather
-  // than the walk stopping at the first miss, because the measurements do not
-  // always ascend with the list: on a busy machine on 2026-08-15 `faster` read
-  // below `fast` twice.
-  for (const entry of benchmark) {
-    // At the size this output is encoded at: a preset not measured there is
-    // not a mode this host has shown it can hold there.
-    const encodePixelsPerSec = throughputAt(entry, frame);
-    if (encodePixelsPerSec === null) {
-      continue;
-    }
-    const speed = predictedRealtimeSpeed({
-      decodeModel: cost.decodeModel ?? null,
-      encodePixelsPerSec,
-      outputPixelsPerSec: pixelsPerSecNeeded,
-      source: cost.source ?? null,
-      observedDecodeCostSec: observed
-    });
-    if (speed !== null && speed >= bar) {
-      return entry.preset;
-    }
-  }
-  // Nothing clears the bar: the cheapest preset, which is the last in quality
-  // order. Returning whichever preset measured fastest would hand an expensive
-  // one to a host that has just been shown to hold no rung at all.
-  return benchmark[benchmark.length - 1].preset;
-}
-
-/**
- * Whether a cost description can actually price decoding — a fit AND a source
- * to apply it to. Without both, every prediction is encoder-only.
- *
- * @param {{ decodeModel?: object | null, source?: object | null }} cost
- * @returns {boolean}
- */
-function isDecodePriced(cost) {
-  if (Number.isFinite(cost?.observedDecodeCostSec) && cost.observedDecodeCostSec > 0) {
-    return true; // measured on the source itself, which needs no fit to stand on
-  }
-  return Boolean(cost?.decodeModel) && Boolean(cost?.source);
-}
-
-/**
- * Build the resolution ladder for a ceiling box. Returns candidate output
- * dimensions from the ceiling downward, preserving the ceiling's aspect ratio,
- * each even-sized. The ceiling itself is always the top rung; the ladder's
- * standard heights (`LADDER_HEIGHTS`, the same list the quality menu offers) at
- * or above it are skipped (never upscale). Deduped by height.
- *
- * @param {number} ceilingWidth
- * @param {number} ceilingHeight
- * @returns {Array<{ width: number, height: number }>} high→low
- */
-export function buildResolutionLadder(ceilingWidth, ceilingHeight) {
-  const cw = Number.isInteger(ceilingWidth) && ceilingWidth > 0 ? ceilingWidth : 0;
-  const ch = Number.isInteger(ceilingHeight) && ceilingHeight > 0 ? ceilingHeight : 0;
-  if (!cw || !ch) {
-    return [];
-  }
-  const even = (v) => {
-    const r = Math.round(v);
-    return Math.max(2, r - (r % 2));
-  };
-  /** @type {Array<{ width: number, height: number }>} */
-  const rungs = [{ width: cw, height: ch }];
-  for (const h of LADDER_HEIGHTS) {
-    if (h >= ch) {
-      continue; // at/above the ceiling — the ceiling rung already covers it
-    }
-    rungs.push({ width: even(cw * (h / ch)), height: h });
-  }
-  const seen = new Set();
-  return rungs.filter((rung) => {
-    if (seen.has(rung.height)) {
-      return false;
-    }
-    seen.add(rung.height);
-    return true;
-  });
-}
-
-/**
- * Choose the software encode settings (resolution + preset) that fit the
- * realtime budget on this host. From the resolution ladder (ceiling downward),
- * pick the HIGHEST rung whose encode throughput — predicted from the startup
- * benchmark's fastest preset — clears the speed this file's own supply
- * demands (`speedBar`). Then, at that resolution, pick the highest-quality
- * preset that still clears it. When even the lowest rung cannot clear it, use the lowest rung with
- * the fastest preset (best effort — a smaller picture beats sub-realtime
- * playback at full size). Returns null when no benchmark or ceiling is
- * available (the caller keeps the ceiling resolution and the default preset).
- *
- * @param {Array<{ preset: string, pixelsPerSec: number }>} benchmark - slowest→fastest
- * @param {{ width: number, height: number }} ceiling
- * @param {number} outputFps
- * @param {{ decodeModel?: object | null, source?: { megapixelsPerSecond: number, megabitsPerSecond: number } | null, requiredSpeed?: number | null }} [cost]
- * @returns {{ width: number, height: number, preset: string, ladder: Array<{ width: number, height: number }>, rungIndex: number } | null}
- */
-export function chooseSoftwareEncodeSettings(benchmark, ceiling, outputFps, cost = {}) {
-  if (!Array.isArray(benchmark) || benchmark.length === 0) {
-    return null;
-  }
-  const fps = Number.isFinite(outputFps) && outputFps > 0 ? outputFps : TRANSCODE_FPS;
-  const ladder = buildResolutionLadder(ceiling?.width, ceiling?.height);
-  if (ladder.length === 0) {
-    return null;
-  }
-  const bar = barFor(cost);
-  let chosenIndex = ladder.length - 1; // default: lowest rung (best effort)
-  for (let i = 0; i < ladder.length; i += 1) {
-    // The cheapest preset's throughput AT THIS RUNG's size.
-    const fastest = cheapestPresetPixelsPerSec(benchmark, ladder[i]);
-    if (fastest === null) {
-      continue;
-    }
-    const speed = predictedRealtimeSpeed({
-      decodeModel: cost.decodeModel ?? null,
-      encodePixelsPerSec: fastest,
-      outputPixelsPerSec: ladder[i].width * ladder[i].height * fps,
-      source: cost.source ?? null
-    });
-    if (speed !== null && speed >= bar) {
-      chosenIndex = i;
-      break;
-    }
-  }
-  const chosen = ladder[chosenIndex];
-  const preset = pickSoftwarePreset(benchmark, chosen.width * chosen.height * fps, cost, chosen);
-  return { width: chosen.width, height: chosen.height, preset, ladder, rungIndex: chosenIndex };
-}
-
-/**
- * How fast this machine COPIES a picture, in seconds of film per second.
- *
- * The startup measurements price encoding and decoding, and a copied picture
- * does neither: it reads packets and writes them out again. That left one whole
- * branch of what this proxy does with no figure at all, and a figure is what
- * every decision in the encoding layer is made from — where to put an encoder,
- * how many to run, whether anybody will be left waiting. Without it a copied
- * output was planned with no speed until its own run had been running long
- * enough to report one, which is exactly the moment the plan matters most.
- *
- * Measured the same way as the others: ffmpeg's own progress, read as a slope
- * over a window, so the process starting is outside the figure.
- *
- * The clip is joined to itself first rather than looped with `-stream_loop`.
- * Looping charges a re-initialisation per lap — measured on the addon host at
- * 0.03 s for 480p and 0.12 s for 1080p — and a copy of a five-second clip laps
- * many times a second, so the reading would have been mostly re-initialisation.
- *
- * @param {{ ffmpegBin: string, logger?: { info: (m: string) => void, warn: (m: string) => void }, clipsDir?: string }} params
- * @returns {Promise<number | null>} Seconds of film per second, or null where
- *   the reading could not be taken. Null means unmeasured and is never a
- *   substitute for a number.
- */
-export async function benchmarkCopySpeed({ ffmpegBin, logger, clipsDir = CALIBRATION_DIR }) {
-  const log = logger ?? { info: () => {}, warn: () => {} };
-  const startedAt = Date.now();
-  // The largest clip in the set. A copy moves BYTES, so what it can do is a
-  // statement about the biggest pictures this host will be asked to pass
-  // through, and the small ones are covered by the same figure.
-  const clip = path.join(clipsDir, "cal-h264-1080-hi.mp4");
-  const speed = await measureCopySlope(ffmpegBin, clip);
-  if (!(speed > 0)) {
-    log.warn("hwaccel: copying could not be measured; a copied picture will be planned from its own run instead");
-    return null;
-  }
-  log.info(
-    `hwaccel: this host copies a picture at ${speed.toFixed(0)}x realtime ` +
-    `(measured in ${((Date.now() - startedAt) / 1000).toFixed(1)}s)`
-  );
-  return speed;
-}
-
-/**
- * Seconds of film per second, copying one file, read from ffmpeg's progress.
- *
- * @param {string} ffmpegBin
- * @param {string} filePath
- * @returns {Promise<number | null>}
- */
-function measureCopySlope(ffmpegBin, filePath) {
-  return new Promise((resolve) => {
-    const args = [
-      "-hide_banner", "-loglevel", "error", "-nostats",
-      // Played over and over, because a copy gets through a five-second clip in
-      // milliseconds and a slope needs a window to be taken over. Looping
-      // charges the demuxer being re-opened once a lap, so what comes out is a
-      // FLOOR on what this host can copy — the safe direction, since a plan made
-      // from it expects copying to be slower than it is.
-      "-stream_loop", "-1", "-i", filePath,
-      // What a copied output does: packets in, packets out, nothing decoded and
-      // nothing encoded. Written nowhere, so the figure is this machine's own
-      // handling and not the disk under a temp directory.
-      "-c", "copy", "-f", "null", "-",
-      // Progress is reported every half second by default, which over a window
-      // of one second is two readings. This asks for twenty.
-      "-stats_period", "0.05",
-      "-progress", "pipe:1"
-    ];
-    /** @type {Array<{ wallSec: number, outSec: number }>} */
-    const samples = [];
-    let settled = false;
-    let buffered = "";
-    let child;
-    const startedAt = Date.now();
-    const finish = (value) => {
-      if (settled) {
-        return;
-      }
-      settled = true;
-      clearTimeout(timer);
-      try {
-        child?.kill("SIGKILL");
-      } catch {
-        // already gone
-      }
-      resolve(value);
-    };
-    const timer = setTimeout(() => finish(null), ENCODE_BENCHMARK_TIMEOUT_MS);
-    try {
-      child = spawn(ffmpegBin, args, { stdio: ["ignore", "pipe", "ignore"], windowsHide: true });
-    } catch {
-      finish(null);
-      return;
-    }
-    child.stdout.on("data", (chunk) => {
-      buffered += String(chunk);
-      let newline = buffered.indexOf(NEWLINE);
-      while (newline >= 0) {
-        const line = buffered.slice(0, newline).trim();
-        buffered = buffered.slice(newline + 1);
-        if (line.startsWith("out_time_ms=")) {
-          const outSec = Number(line.slice("out_time_ms=".length)) / 1e6;
-          if (Number.isFinite(outSec) && outSec >= 0) {
-            samples.push({ wallSec: (Date.now() - startedAt) / 1000, outSec });
-          }
-        }
-        newline = buffered.indexOf(NEWLINE);
-      }
-      const slope = slopeOf(samples);
-      if (slope !== null) {
-        finish(slope);
-      }
-    });
-    child.on("error", () => finish(null));
-    child.on("exit", () => finish(slopeOf(samples, ENCODE_BENCHMARK_MIN_WINDOW_SEC)));
-  });
-}
+/**
+ * @file Hardware-accelerated H.264 encoder auto-detection.
+ *
+ * Probes the ffmpeg build and the host for a usable hardware H.264 encoder
+ * (NVENC / QSV / VAAPI / V4L2 M2M), verifying each candidate with a real
+ * test-encode before selecting it. Falls back to software libx264 when no
+ * hardware encoder is present or working.
+ *
+ * Deployment-agnostic: relies only on ffmpeg, the filesystem and
+ * `process.platform`; makes no assumptions about Home Assistant or any
+ * specific host. A garbled or unsupported hardware path simply fails its
+ * test-encode and is skipped, so the worst case is software encoding.
+ *
+ * A descriptor exposes:
+ *   - `name`        human-readable encoder id (e.g. "h264_vaapi")
+ *   - `kind`        "software" | "vaapi" | "qsv" | "nvenc" | "v4l2m2m"
+ *   - `device`      device node path or null
+ *   - `inputArgs`   ffmpeg args inserted before `-i` (decode/hwaccel setup)
+ *   - `buildVideoArgs({ targetWidth, targetHeight, segmentDurationSec })`
+ *                   ffmpeg video filter + encoder args inserted after `-map`s
+ */
+
+import { spawn } from "node:child_process";
+import { mkdtempSync, readdirSync, rmSync, statSync } from "node:fs";
+import { mkdtemp, readFile, rm } from "node:fs/promises";
+import os from "node:os";
+import path from "node:path";
+import { fitDecodeCost } from "./decode-cost-fit.js";
+import { penaltiesFrom } from "./contention.js";
+import { fileURLToPath } from "node:url";
+import {
+  parseFfmpegBitrateKbps,
+  parseFfmpegDurationSeconds,
+  parseFfmpegVideoDimensions,
+  parseFfmpegVideoFps
+} from "../media/ffmpeg-banner.js";
+
+import { keyFrameArgs, TRANSCODE_FPS } from "./args.js";
+import { throughputAt } from "./throughput.js";
+import { LADDER_HEIGHTS } from "./output/ladder.js";
+// The five kinds, one class each. Detection and benchmarking stay in this file;
+// how a kind is driven belongs to the kind.
+import {
+  NvencEncoder,
+  QsvEncoder,
+  SoftwareEncoder,
+  V4l2m2mEncoder,
+  VaapiEncoder
+} from "./index.js";
+// Re-exported so every caller goes on importing these figures from here:
+// the same calculation, moved to sit beside the encoder kinds built from it.
+export {
+  chooseOutputFps,
+  maxrateKbpsFor,
+  limitRowFor,
+  nominalKbpsFor,
+  TRANSCODE_FPS
+} from "./args.js";
+
+const BENCHMARK_REF_W = 640;
+const BENCHMARK_REF_H = 360;
+const BENCHMARK_DURATION_SEC = 3;
+/**
+ * How many raw frames a throughput reading at this size is taken over.
+ *
+ * The reference size keeps the three seconds it has always had. A larger frame
+ * gets as many frames as the same number of bytes buys, but never fewer than
+ * one second of film: the frames are looped, and a loop shorter than the
+ * encoder's reach — its reference frames and its B-frames — would let it find
+ * an identical frame to predict from and read far faster than any film.
+ *
+ * @param {{ width: number, height: number }} frame
+ * @returns {number}
+ */
+export function rawFrameCountFor(frame) {
+  const referenceBytes = BENCHMARK_REF_W * BENCHMARK_REF_H * 1.5 * TRANSCODE_FPS * BENCHMARK_DURATION_SEC;
+  const frameBytes = frame.width * frame.height * 1.5;
+  return Math.max(TRANSCODE_FPS, Math.floor(referenceBytes / frameBytes));
+}
+/**
+ * The narrowest window a slope may be taken over. Measured 2026-08-15: at a
+ * fifth of a second the readings were noisy enough to put `faster` and
+ * `veryfast` BELOW `fast`, which libx264 cannot do — and `pickSoftwarePreset`
+ * walks the list assuming it ascends. Half a second was still noisy enough for that
+ * (measured again: veryfast below faster, twice), so a full second it is —
+ * about six seconds of startup for a ladder the whole budget then rests on.
+ */
+const ENCODE_BENCHMARK_WINDOW_SEC = 1;
+/** The narrowest window that may be used when a run ends early. */
+const ENCODE_BENCHMARK_MIN_WINDOW_SEC = 0.2;
+/** Above this a reading is a fault, not a fast machine. */
+const ENCODE_BENCHMARK_MAX_PLAUSIBLE_SPEED = 1000;
+/**
+ * A preset that has not reported twice in this long is hung, not slow: reports
+ * arrive twice a second whatever the encoding speed.
+ */
+const ENCODE_BENCHMARK_TIMEOUT_MS = 10_000;
+/**
+ * How many times the calibration clip is joined to itself to measure copying.
+ *
+ * Not a figure about the machine: it is how much film the reading needs to have
+ * in front of it. A copy runs at hundreds of times realtime, and the slope is
+ * taken over a window of one second, so the input has to hold more film than the
+ * fastest plausible host gets through in that second. Forty laps of a five-second
+ * clip is 200 s of film, which covers the ceiling `slopeOf` will accept.
+ */
+/** Progress reports arrive line by line. */
+const NEWLINE = String.fromCharCode(10);
+// Producing one second of video per second of clock. Not a margin and not a
+// choice — the definition of keeping up, and the bar when nothing better is
+// known about the supply this step will meet.
+const REALTIME = 1;
+// The bar where decoding CANNOT be priced — no calibration fit, or a source the
+// probe said too little about. This one is not measured and cannot be: the
+// prediction it guards counts encoding only, which on the field host was
+// several times optimistic, and there is no reading on such a host to correct
+// it with. It is left at the figure it has had since before decoding was
+// priced, because lowering it to realtime would make the least-measured hosts
+// the most permissive. Where decoding IS priced, nothing chosen remains.
+const UNPRICED_DECODE_BAR = 1.8;
+
+
+// The five kinds live in `encode/`, one class each, and these keep the names
+// every caller already uses. A kind states its own arguments and its own
+// ladder of speed settings; detection and benchmarking stay here.
+/** @returns {import("./hwaccel.js").VideoEncoderDescriptor} */
+export function softwareDescriptor() {
+  return new SoftwareEncoder();
+}
+
+/**
+ * @param {string} device
+ * @returns {import("./hwaccel.js").VideoEncoderDescriptor}
+ */
+function vaapiDescriptor(device) {
+  return new VaapiEncoder(device);
+}
+
+/**
+ * @param {string} device
+ * @returns {import("./hwaccel.js").VideoEncoderDescriptor}
+ */
+function qsvDescriptor(device) {
+  return new QsvEncoder(device);
+}
+
+/** @returns {import("./hwaccel.js").VideoEncoderDescriptor} */
+function nvencDescriptor() {
+  return new NvencEncoder();
+}
+
+/** @returns {import("./hwaccel.js").VideoEncoderDescriptor} */
+function v4l2m2mDescriptor() {
+  return new V4l2m2mEncoder();
+}
+
+
+
+/**
+ * @typedef {Object} VideoEncoderDescriptor
+ * @property {string} name
+ * @property {"software"|"vaapi"|"qsv"|"nvenc"|"v4l2m2m"} kind
+ * @property {string|null} device
+ * @property {string[]} inputArgs
+ * @property {(opts: { targetWidth: number, targetHeight: number, segmentDurationSec: number, preset?: string, fps?: number, tonemap?: boolean, forcedKeyframeTimes?: number[] | null, rateControl?: { maxrateKbps: number, bufsizeKbps: number, level: string | null } | null }) => string[]} buildVideoArgs
+ */
+
+/**
+ * Run ffmpeg and resolve with its exit code and captured output.
+ *
+ * @param {string} ffmpegBin
+ * @param {string[]} args
+ * @param {number} [timeoutMs=12000]
+ * @returns {Promise<{ code: number, stdout: string, stderr: string }>}
+ */
+export function runFfmpeg(ffmpegBin, args, timeoutMs = 12000) {
+  return new Promise((resolve) => {
+    let stdout = "";
+    let stderr = "";
+    let settled = false;
+    let child;
+    const finish = (code) => {
+      if (settled) {
+        return;
+      }
+      settled = true;
+      resolve({ code, stdout, stderr });
+    };
+    try {
+      child = spawn(ffmpegBin, args, { stdio: ["ignore", "pipe", "pipe"], windowsHide: true });
+    } catch {
+      finish(-1);
+      return;
+    }
+    const timer = setTimeout(() => {
+      try {
+        child.kill("SIGKILL");
+      } catch {
+        // already gone
+      }
+      finish(-1);
+    }, timeoutMs);
+    child.stdout.on("data", (chunk) => {
+      stdout += String(chunk);
+    });
+    child.stderr.on("data", (d) => {
+      stderr += String(d);
+    });
+    child.on("error", () => {
+      clearTimeout(timer);
+      finish(-1);
+    });
+    child.on("exit", (code) => {
+      clearTimeout(timer);
+      finish(code ?? -1);
+    });
+  });
+}
+
+/** @returns {string[]} /dev/dri/renderD* nodes (VAAPI/QSV). */
+function listRenderNodes() {
+  try {
+    return readdirSync("/dev/dri")
+      .filter((n) => n.startsWith("renderD"))
+      .map((n) => `/dev/dri/${n}`)
+      .sort();
+  } catch {
+    return [];
+  }
+}
+
+/** @returns {boolean} Whether any /dev/nvidia* node exists (NVENC). */
+function hasNvidiaDevice() {
+  try {
+    return readdirSync("/dev").some((n) => /^nvidia(\d+)?$/.test(n));
+  } catch {
+    return false;
+  }
+}
+
+/** @returns {boolean} Whether any /dev/video* node exists (V4L2 M2M). */
+function hasV4l2Device() {
+  try {
+    return readdirSync("/dev").some((n) => /^video\d+$/.test(n));
+  } catch {
+    return false;
+  }
+}
+
+/**
+ * Build a full ffmpeg command that encodes a short, *moving* synthetic clip
+ * (testsrc2 — far more representative than a static black frame) through the
+ * candidate encoder into real HLS segments in `outDir`, with keyframes forced
+ * on segment boundaries. Verifying the resulting segments (see
+ * {@link verifySegmentsDecodeCleanly}) catches encoders that silently produce
+ * a corrupted or non-IDR-aligned stream (e.g. some V4L2 M2M builds).
+ *
+ * @param {VideoEncoderDescriptor} descriptor
+ * @param {number} segmentDurationSec
+ * @param {string} outDir
+ * @returns {string[]}
+ */
+function buildEncoderTestArgs(descriptor, segmentDurationSec, outDir) {
+  const durationSec = Math.max(8, segmentDurationSec * 3);
+  const source = ["-f", "lavfi", "-i", `testsrc2=s=640x360:r=${TRANSCODE_FPS}:d=${durationSec}`];
+  const kf = keyFrameArgs(segmentDurationSec);
+
+  /** @type {string[]} */
+  let pre = ["-hide_banner", "-loglevel", "error"];
+  /** @type {string[]} */
+  let encode;
+  switch (descriptor.kind) {
+    case "vaapi":
+      pre = [...pre, "-vaapi_device", String(descriptor.device)];
+      encode = ["-vf", "format=nv12,hwupload", "-c:v", "h264_vaapi", "-qp", "24", ...kf];
+      break;
+    case "qsv":
+      pre = [...pre, "-qsv_device", String(descriptor.device)];
+      encode = ["-vf", "hwupload=extra_hw_frames=16,format=qsv", "-c:v", "h264_qsv", "-global_quality", "24", ...kf];
+      break;
+    case "nvenc":
+      encode = ["-c:v", "h264_nvenc", "-preset", "p4", "-cq", "24", "-pix_fmt", "yuv420p", ...kf];
+      break;
+    case "v4l2m2m":
+      encode = ["-pix_fmt", "yuv420p", "-c:v", "h264_v4l2m2m", "-num_capture_buffers", "32", "-b:v", "3M", "-g", String(TRANSCODE_FPS * segmentDurationSec), ...kf];
+      break;
+    default:
+      encode = ["-c:v", "libx264", "-preset", "ultrafast", "-pix_fmt", "yuv420p", ...kf];
+      break;
+  }
+
+  const hlsOut = [
+    "-f", "hls",
+    "-hls_time", String(segmentDurationSec),
+    "-hls_list_size", "0",
+    "-hls_flags", "independent_segments",
+    // fMP4 (CMAF) — matches the runtime pipeline (`encode/run-command.js`).
+    "-hls_segment_type", "fmp4",
+    "-hls_fmp4_init_filename", "init.mp4",
+    "-hls_segment_filename", path.join(outDir, "seg-%03d.m4s"),
+    path.join(outDir, "index.m3u8")
+  ];
+  return [...pre, ...source, ...encode, ...hlsOut];
+}
+
+/**
+ * Verify the HLS segments produced by the test encode are valid: at least two
+ * segments exist, and each decodes standalone without errors. A segment that
+ * does not begin with a keyframe (broken/corrupted output) emits decode errors
+ * when read on its own, which fails this check.
+ *
+ * @param {string} ffmpegBin
+ * @param {string} outDir
+ * @returns {Promise<boolean>}
+ */
+export async function verifySegmentsDecodeCleanly(ffmpegBin, outDir) {
+  let files;
+  try {
+    files = readdirSync(outDir).filter((n) => /^seg-\d+\.m4s$/.test(n));
+  } catch {
+    return false;
+  }
+  if (files.length < 2) {
+    return false;
+  }
+  // fMP4: parameter sets (SPS/PPS) live in init.mp4, not in each segment.
+  // Decode the whole playlist (ffmpeg's own, which references init.mp4 via
+  // #EXT-X-MAP), so every segment is exercised together with the init. Any
+  // corrupt / non-conformant segment (e.g. some V4L2 M2M builds emit a stray
+  // no-picture access unit) surfaces as a decode error here.
+  const result = await runFfmpeg(
+    ffmpegBin,
+    ["-hide_banner", "-loglevel", "error", "-i", path.join(outDir, "index.m3u8"), "-f", "null", "-"],
+    12000
+  );
+  return result.code === 0 && result.stderr.trim().length === 0;
+}
+
+/**
+ * Detect the best usable H.264 encoder. Always resolves (falls back to
+ * software libx264). Each hardware candidate is verified with a real
+ * test-encode before being selected.
+ *
+ * @param {{ ffmpegBin: string, logger?: { info: (m: string) => void, warn: (m: string) => void }, segmentDurationSec?: number }} options
+ * @returns {Promise<VideoEncoderDescriptor>}
+ */
+export async function detectVideoEncoder({ ffmpegBin, logger, segmentDurationSec = 4 }) {
+  const log = logger ?? { info: () => {}, warn: () => {} };
+  const software = softwareDescriptor();
+
+  const { code, stdout } = await runFfmpeg(ffmpegBin, ["-hide_banner", "-encoders"], 10000);
+  if (code !== 0) {
+    log.warn("hwaccel: could not list ffmpeg encoders; using software libx264");
+    return software;
+  }
+  const has = (name) => stdout.includes(name);
+
+  /** @type {VideoEncoderDescriptor[]} */
+  const candidates = [];
+  const renderNodes = listRenderNodes();
+  if (has("h264_nvenc") && hasNvidiaDevice()) {
+    candidates.push(nvencDescriptor());
+  }
+  if (has("h264_qsv") && renderNodes.length > 0) {
+    candidates.push(qsvDescriptor(renderNodes[0]));
+  }
+  if (has("h264_vaapi") && renderNodes.length > 0) {
+    candidates.push(vaapiDescriptor(renderNodes[0]));
+  }
+  // h264_v4l2m2m (ARM SoC / Raspberry Pi / HA Yellow). It is gated behind the
+  // strict keyframe-alignment test below, because some V4L2 M2M builds silently
+  // emit a corrupted / non-IDR-aligned stream; the test rejects those and the
+  // host falls back to software libx264.
+  if (has("h264_v4l2m2m") && hasV4l2Device()) {
+    candidates.push(v4l2m2mDescriptor());
+  }
+
+  for (const candidate of candidates) {
+    const dir = mkdtempSync(path.join(os.tmpdir(), "tt-hwtest-"));
+    let ok = false;
+    try {
+      const encoded = await runFfmpeg(
+        ffmpegBin,
+        buildEncoderTestArgs(candidate, segmentDurationSec, dir),
+        25000
+      );
+      if (encoded.code === 0) {
+        ok = await verifySegmentsDecodeCleanly(ffmpegBin, dir);
+      }
+    } finally {
+      try {
+        rmSync(dir, { recursive: true, force: true });
+      } catch {
+        // best effort
+      }
+    }
+    if (ok) {
+      log.info(
+        `hwaccel: using hardware encoder ${candidate.name}` +
+          `${candidate.device ? ` (${candidate.device})` : ""}`
+      );
+      return candidate;
+    }
+    log.warn(`hwaccel: ${candidate.name} failed the HLS keyframe-alignment test; skipping`);
+  }
+
+  log.info("hwaccel: no working hardware encoder; using software libx264");
+  return software;
+}
+
+/**
+ * Detect whether this ffmpeg build has the filters needed for the HDR→SDR
+ * tone-map chain (`zscale`, from libzimg, and `tonemap`). Both are required;
+ * when either is missing, HDR sources are re-encoded without tone mapping
+ * (washed-out but playable). Always resolves.
+ *
+ * @param {{ ffmpegBin: string, logger?: { info: (m: string) => void, warn: (m: string) => void } }} options
+ * @returns {Promise<boolean>}
+ */
+export async function detectTonemapSupport({ ffmpegBin, logger }) {
+  const log = logger ?? { info: () => {}, warn: () => {} };
+  const { code, stdout } = await runFfmpeg(ffmpegBin, ["-hide_banner", "-filters"], 10000);
+  if (code !== 0) {
+    log.warn("hwaccel: could not list ffmpeg filters; HDR tone mapping disabled");
+    return false;
+  }
+  // `-filters` prints one filter per line: "... zscale  ...", "... tonemap ...".
+  const hasZscale = /\bzscale\b/.test(stdout);
+  const hasTonemap = /\btonemap\b/.test(stdout);
+  const supported = hasZscale && hasTonemap;
+  log.info(
+    `hwaccel: HDR tone mapping ${supported ? "available" : "unavailable"} ` +
+      `(zscale=${hasZscale} tonemap=${hasTonemap})`
+  );
+  return supported;
+}
+
+
+// The clips the decode cost is fitted from. They ship with the package
+// (`assets/calibration/`), cut from Netflix Open Content "Meridian" (CC-BY 4.0)
+// — real, grainy live action, because a generated `testsrc2` clip decodes 158 %
+// away from a real film where these are 11 % away (measured 2026-08-14).
+//
+// Three sizes at two bitrates each, with the axes varied INDEPENDENTLY. The set
+// this replaced was three clips for three unknowns, two of them at the same
+// size: an exact system, which cannot fail visibly. On 2026-08-17 it returned
+// `0.007542 × Mpx/s + 0.000000 × Mbit/s + 0.0000 s/s` — the bitrate term and
+// the constant exactly zero — and the prediction on top of it was 1.8-2.2x
+// optimistic. Six points leave three spare, so the fit has a residual, and a
+// term the data does not determine can be refused instead of published as a
+// zero that looks measured. See `assets/calibration/NOTICE.md`.
+//
+// One set PER CODEC FAMILY, because a family is what the model describes. The
+// fit used to be H.264 only, while a video that has to be RE-ENCODED is by
+// definition one the browser could not play — which is to say HEVC, 10-bit or
+// AV1 — and those decode dearer per pixel on the same box. Pricing them with
+// H.264 constants is the one case the model is always asked about and was never
+// measured on.
+//
+// A family that has no set of its own is priced with H.264's, which is what
+// happened to every family before this; the line says so rather than implying
+// it. AV1 has no set yet: the survey of 2026-07-10 found it rare where HEVC was
+// 18 % of releases, so it waits for the same treatment.
+const CALIBRATION_SETS = {
+  h264: [
+    "cal-h264-1080-hi.mp4",
+    "cal-h264-1080-lo.mp4",
+    "cal-h264-720-hi.mp4",
+    "cal-h264-720-lo.mp4",
+    "cal-h264-480-hi.mp4",
+    "cal-h264-480-lo.mp4"
+  ],
+  hevc: [
+    "cal-hevc-1080-hi.mp4",
+    "cal-hevc-1080-lo.mp4",
+    "cal-hevc-480-hi.mp4",
+    "cal-hevc-480-lo.mp4"
+  ],
+  hevc10: [
+    "cal-hevc10-1080-hi.mp4",
+    "cal-hevc10-1080-lo.mp4",
+    "cal-hevc10-480-hi.mp4",
+    "cal-hevc10-480-lo.mp4"
+  ]
+};
+export const CALIBRATION_REFERENCE_CLIP = "cal-h264-1080-hi.mp4";
+export const CALIBRATION_DIR = path.join(path.dirname(fileURLToPath(import.meta.url)), "..", "..", "assets", "calibration");
+// How wide the measured window must be before the slope is trusted, and how
+// long to wait for it at most.
+//
+// Half a second, and it is the TIMING noise that sets it rather than the amount
+// of video: the slope is output time against wall time, both read from the same
+// two progress lines, and the jitter in stamping one is milliseconds — so half
+// a second of window is a fraction of a percent of error on any host. What used
+// to make a longer window necessary was the clip restarting inside it, and that
+// is gone: the stream is continuous now. Measured 2026-08-22 against the
+// continuous-pass truth on a desktop: -3.0 % and +0.6 % at half a second, with
+// the readings spread 2-6 %, against -25 % and -33 % for the loop it replaces.
+// Half a second also costs the startup about 0.6 s per clip less, which matters
+// because every clip of every codec family is paid for before any viewer
+// exists.
+const DECODE_WINDOW_MIN_SEC = 0.5;
+const DECODE_WINDOW_MAX_MS = 8000;
+
+/**
+ * Read what a calibration clip IS from the decode run's own output: the
+ * dimensions, the frame rate and the bitrate ffmpeg reports for it. Read rather
+ * than declared, so replacing a clip cannot silently invalidate the fit.
+ *
+ * @param {string} stderr
+ * @returns {{ megapixelsPerSecond: number, megabitsPerSecond: number, durationSeconds: number } | null}
+ */
+function parseClipCharacteristics(stderr) {
+  // The same readers the session manager uses on the same banner — one parser
+  // per fact, so a second copy cannot drift from the first.
+  const { width, height } = parseFfmpegVideoDimensions(stderr);
+  const rate = parseFfmpegVideoFps(stderr);
+  const seconds = parseFfmpegDurationSeconds(stderr);
+  const kbps = parseFfmpegBitrateKbps(stderr);
+  if (!(width > 0) || !(height > 0) || !(rate > 0) || !(seconds > 0) || !(kbps > 0)) {
+    return null;
+  }
+  return {
+    megapixelsPerSecond: (width * height * rate) / 1e6,
+    megabitsPerSecond: kbps / 1000,
+    durationSeconds: seconds
+  };
+}
+
+/**
+ * Measure what DECODING costs on this host, as seconds of work per second of
+ * video, and solve it into three host constants:
+ *
+ *   decodeCost = a × Mpixel/s + b × Mbit/s + c
+ *
+ * Why it exists: the preset benchmark below measures ENCODING only, and a
+ * re-encode pays for both halves. Measured 2026-08-14 on the addon host, that
+ * omission made the budget offer a 240p rung it then ran at 0.39-0.95× — the
+ * benchmark said the host cleared the bar 2.5× over. With the decode term the
+ * same file predicts within 4.8 %; without it the error on that rung was 209 %.
+ *
+ * The constants are properties of the HOST, so this runs once at startup (about
+ * 5 s on a CM4) and any source is then priced from figures the probe already
+ * has — nothing is added to a session's cold start.
+ *
+ * They are also properties of the CODEC, and the clips are H.264: HEVC, AV1 and
+ * 10-bit decode dearer per pixel on the same machine, and a source that has to
+ * be re-encoded is by definition one this browser could not play, which is
+ * usually not H.264. So the fit is optimistic exactly there. Closing that needs
+ * clips in those codecs, and is its own roadmap item.
+ *
+ * @param {{ ffmpegBin: string, logger?: { info: (m: string) => void, warn: (m: string) => void }, clipsDir?: string }} options
+ * @returns {Promise<{ pixelTerm: number, bitrateTerm: number, constantTerm: number } | null>}
+ */
+/**
+ * What a second job costs on this host, measured rather than assumed.
+ *
+ * The budget adds seconds of work per second of content — this encode, plus
+ * that decode, plus what is already committed — and the addon host contradicted
+ * that directly on 2026-08-18: decoding ran at 2.10-2.25x alone, 0.79-0.90x
+ * with one encoder beside it and 0.56-0.64x with two. The same work costs 2.6×
+ * more for having company. Heat is not the cause (the hot idle machine was the
+ * fastest reading of all); four cores sharing one path to memory is.
+ *
+ * So it is measured the way everything else here is: the same clip decoded
+ * alone, then decoded again while an encoder of the same clip runs beside it.
+ * The ratio is the penalty. The encoder is stopped as soon as the reading is
+ * taken, and the whole thing costs one decode plus one short encode.
+ *
+ * @param {{ ffmpegBin: string, logger?: { info: (m: string) => void, warn: (m: string) => void }, clipsDir?: string, upTo?: number }} options
+ * @returns {Promise<Map<number, number> | null>} Penalties by how many other
+ *   jobs were running, or null when the readings could not be taken.
+ */
+export async function benchmarkContention({ ffmpegBin, logger, clipsDir = CALIBRATION_DIR, upTo = 2 }) {
+  const log = logger ?? { info: () => {}, warn: () => {} };
+  // The cheapest clip in the set: this measures the MACHINE's behaviour under
+  // company, not the clip's own cost, so the smallest one says it soonest.
+  const clip = path.join(clipsDir, "cal-h264-480-lo.mp4");
+  const startedAt = Date.now();
+  // Lifted once and decoded three times from the same bytes. Going through
+  // `measureDecodeSlope` lifted it again for every reading — three process
+  // starts on a path that is awaited before the proxy's tunnel opens, for a
+  // remux whose result had not changed.
+  const streams = await extractFamilyStreams(ffmpegBin, [clip], "h264");
+  const stream = streams?.[0];
+  if (!stream) {
+    log.warn("hwaccel: contention could not be measured; costs will be added as though jobs were independent");
+    return null;
+  }
+  const alone = await decodePipedStream(ffmpegBin, stream, log);
+  if (!alone?.speed) {
+    log.warn("hwaccel: contention could not be measured; costs will be added as though jobs were independent");
+    return null;
+  }
+  /** @type {Array<{ others: number, speed: number }>} */
+  const beside = [];
+  /** @type {import("node:child_process").ChildProcess[]} */
+  const load = [];
+  try {
+    for (let others = 1; others <= Math.max(1, upTo); others += 1) {
+      load.push(
+        spawn(
+          ffmpegBin,
+          [
+            "-hide_banner", "-loglevel", "error", "-nostats",
+            "-stream_loop", "-1", "-i", clip,
+            "-an", "-c:v", "libx264", "-preset", "fast", "-f", "null", "-"
+          ],
+          { stdio: ["ignore", "ignore", "ignore"], windowsHide: true }
+        )
+      );
+      // Let the encoder reach its own speed before reading anything: an encode
+      // measured in its first moments is measuring the process starting.
+      await new Promise((resolve) => {
+        setTimeout(resolve, 2_000);
+      });
+      const withCompany = await decodePipedStream(ffmpegBin, stream, log);
+      if (withCompany?.speed) {
+        beside.push({ others, speed: withCompany.speed });
+      }
+    }
+  } finally {
+    for (const child of load) {
+      try {
+        child.kill("SIGKILL");
+      } catch {
+        // Already gone: the reading is what mattered, and nothing else uses it.
+      }
+    }
+  }
+  const penalties = penaltiesFrom(alone.speed, beside);
+  if (!penalties) {
+    log.warn("hwaccel: contention readings said nothing; costs will be added as though jobs were independent");
+    return null;
+  }
+  log.info(
+    `hwaccel: a second job costs ${[...penalties.entries()]
+      .map(([others, penalty]) => `${penalty.toFixed(2)}x beside ${others}`)
+      .join(", ")} ` +
+    `(decode alone ${alone.speed.toFixed(2)}x, ` +
+    `${beside.map((reading) => `${reading.speed.toFixed(2)}x beside ${reading.others}`).join(", ")}, ` +
+    `measured in ${((Date.now() - startedAt) / 1000).toFixed(1)}s)`
+  );
+  return penalties;
+}
+
+export async function benchmarkDecodeCost({ ffmpegBin, logger, clipsDir = CALIBRATION_DIR }) {
+  const log = logger ?? { info: () => {}, warn: () => {} };
+  const startedAllAt = Date.now();
+  /** @type {Record<string, { pixelTerm: number, bitrateTerm: number, constantTerm: number }>} */
+  const families = {};
+  for (const [family, clips] of Object.entries(CALIBRATION_SETS)) {
+    const fitted = await fitOneFamily({ ffmpegBin, log, clipsDir, family, clips });
+    if (fitted) {
+      families[family] = fitted;
+    }
+  }
+  if (!families.h264) {
+    // H.264 is the family every other one falls back to, so without it there is
+    // no model at all rather than a partial one. Which families DID fit is said
+    // anyway: on a fast host the H.264 clips decode at 20-80x and the readings
+    // stop being ordered — measured 2026-08-20 on a desktop, 1080p at 9.35
+    // Mbit/s costing 0.0307 s/s against 720p at 9.94 costing 0.0472, which is
+    // not a thing a decoder does — so a failure here is a measurement problem
+    // and not a missing file, and the line has to let those be told apart.
+    log.warn(
+      "hwaccel: decode cost unknown — the H.264 clips did not fit" +
+        (Object.keys(families).length > 0
+          ? `, though ${Object.keys(families).join(" and ")} did`
+          : "")
+    );
+    return null;
+  }
+  const missing = Object.keys(CALIBRATION_SETS).filter((family) => !families[family]);
+  log.info(
+    `hwaccel: decode cost measured for ${Object.keys(families).join(", ")}` +
+      (missing.length > 0 ? `; ${missing.join(" and ")} priced as H.264` : "") +
+      ` (in ${((Date.now() - startedAllAt) / 1000).toFixed(1)}s)`
+  );
+  return { families, ...families.h264 };
+}
+
+/**
+ * Fit one codec family's decode cost from its own clips.
+ *
+ * @param {{ ffmpegBin: string, log: { info: Function, warn: Function }, clipsDir: string, family: string, clips: string[] }} params
+ * @returns {Promise<{ pixelTerm: number, bitrateTerm: number, constantTerm: number } | null>}
+ */
+async function fitOneFamily({ ffmpegBin, log, clipsDir, family, clips }) {
+  const startedAllAt = Date.now();
+  /** @type {Array<{ megapixelsPerSecond: number, megabitsPerSecond: number, costSecondsPerSecond: number }>} */
+  const samples = [];
+  // Every clip of the family is lifted out of its container FIRST, in one
+  // ffmpeg run. See `extractFamilyStreams` for why one run rather than one per
+  // clip, and why before the measurements rather than beside them.
+  const streams = await extractFamilyStreams(
+    ffmpegBin,
+    clips.map((clip) => path.join(clipsDir, clip)),
+    family
+  );
+  if (!streams) {
+    log.warn(
+      `hwaccel: ${family} cannot be lifted out of its container — no Annex-B filter is mapped for it, ` +
+        `so its clips were never measured`
+    );
+    return null;
+  }
+  for (const [index, clip] of clips.entries()) {
+    const stream = streams[index];
+    const measured = stream ? await decodePipedStream(ffmpegBin, stream, log) : null;
+    if (!measured?.speed) {
+      log.warn(
+        `hwaccel: decode benchmark "${clip}" said nothing; ${family} not measured` +
+          (measured?.error ? ` — ${measured.error}` : " — the clip could not be lifted out of its container")
+      );
+      return null;
+    }
+    const cost = 1 / measured.speed;
+    samples.push({
+      megapixelsPerSecond: measured.megapixelsPerSecond,
+      megabitsPerSecond: measured.megabitsPerSecond,
+      costSecondsPerSecond: cost
+    });
+    log.info(
+      `hwaccel: decode "${clip}" ${measured.megapixelsPerSecond.toFixed(1)} Mpx/s ` +
+        `${measured.megabitsPerSecond.toFixed(2)} Mbit/s -> ${measured.speed.toFixed(1)}x ` +
+        `(cost ${cost.toFixed(4)} s/s, over ${measured.windowSec.toFixed(1)}s of decoding)`
+    );
+  }
+  const fitted = fitDecodeCost(samples);
+  if (!fitted) {
+    log.warn(`hwaccel: ${family} decode cost could not be fitted to these measurements`);
+    return null;
+  }
+  log.info(
+    `hwaccel: ${family} decode cost = ${fitted.pixelTerm.toFixed(6)} × Mpx/s + ${fitted.bitrateTerm.toFixed(6)} × Mbit/s ` +
+      `+ ${fitted.constantTerm.toFixed(4)} s/s (${fitted.shape} from ${fitted.samples} clips, ` +
+      `typical disagreement ${fitted.residualRms.toFixed(4)} s/s` +
+      // Named rather than implied: a zero in the line above means "not
+      // measured" for a dropped term and "measured to be nothing" otherwise,
+      // and those are different claims.
+      (fitted.dropped.length > 0 ? `, ${fitted.dropped.join(" and ")} not determined by these clips` : "") +
+      `, measured in ${((Date.now() - startedAllAt) / 1000).toFixed(1)}s)`
+  );
+  return { pixelTerm: fitted.pixelTerm, bitrateTerm: fitted.bitrateTerm, constantTerm: fitted.constantTerm };
+}
+
+/**
+ * The bitstream filter and demuxer that turn a clip's video track into a
+ * continuous elementary stream, by codec family.
+ *
+ * H.264 and HEVC in MP4 keep their parameter sets in the container's `avcC` /
+ * `hvcC` and their access units length-prefixed; Annex-B carries them inline,
+ * with start codes, which is what makes plain byte concatenation a valid
+ * stream. That is the property this whole measurement rests on.
+ */
+const ANNEX_B_BY_FAMILY = {
+  h264: { filter: "h264_mp4toannexb", demuxer: "h264" },
+  hevc: { filter: "hevc_mp4toannexb", demuxer: "hevc" },
+  hevc10: { filter: "hevc_mp4toannexb", demuxer: "hevc" }
+};
+
+/**
+ * The last complaint in an ffmpeg stderr, for a line that has to say why.
+ *
+ * @param {string} stderr
+ * @returns {string}
+ */
+function lastErrorLine(stderr) {
+  const lines = String(stderr ?? "")
+    .split(/\r?\n/)
+    .map((line) => line.trim())
+    .filter((line) => line.length > 0);
+  return lines[lines.length - 1] ?? "";
+}
+
+/**
+ * How long the lift may take before it is abandoned. It is a remux of a few
+ * megabytes, so this is not a budget — it is the difference between a startup
+ * that reports a failure and one that never finishes. Every other ffmpeg run in
+ * this file has such a bound; this one did not, and it is awaited before the
+ * proxy's tunnel opens.
+ */
+const EXTRACT_TIMEOUT_MS = 20_000;
+
+/**
+ * Lift a whole family's clips out of their containers, as Annex-B elementary
+ * streams, in ONE ffmpeg run.
+ *
+ * No re-encoding — the frames are copied — so the work itself is trivial and
+ * the cost is almost entirely the process. Doing one process per clip added
+ * 11 s to the startup here (fourteen clips at about 0.83 s each), and running
+ * them concurrently did not help: six at once took 4.75 s against 0.89 s for
+ * one, so the machine serialises them. One run with many inputs and many
+ * outputs costs one process.
+ *
+ * The outputs go to temporary files because several outputs cannot share one
+ * pipe; they are read into memory and deleted immediately, and nothing about
+ * this measurement is kept between runs.
+ *
+ * Before the measurements, never beside them: a remux running next to a decode
+ * is a second job on the machine, and this benchmark exists to find out what
+ * ONE job costs here.
+ *
+ * @param {string} ffmpegBin
+ * @param {string[]} clipPaths
+ * @param {string} family
+ * @returns {Promise<Array<{ bytes: Buffer, demuxer: string, megapixelsPerSecond: number, megabitsPerSecond: number, fps: number } | null> | null>}
+ *   One entry per clip, in order; null when the family cannot be lifted at all.
+ */
+async function extractFamilyStreams(ffmpegBin, clipPaths, family) {
+  const shape = ANNEX_B_BY_FAMILY[family];
+  // A family with no mapping is a hard failure, not a silent fallback to
+  // H.264's filter. AV1 has no Annex-B form at all (its packaging is OBU), and
+  // MPEG-2 and VC-1 have no `*_mp4toannexb` filter — so the three families the
+  // roadmap plans next cannot come through here, and finding that out as
+  // "the clip failed" would send the reader after the clip.
+  if (!shape) {
+    return null;
+  }
+  const workDir = await mkdtemp(path.join(os.tmpdir(), "ttv-calibration-"));
+  const outputs = clipPaths.map((_, index) => path.join(workDir, `stream-${index}.${shape.demuxer}`));
+  /** @type {string[]} */
+  const args = ["-hide_banner", "-loglevel", "info", "-nostats", "-y"];
+  for (const clipPath of clipPaths) {
+    args.push("-i", clipPath);
+  }
+  for (const [index, output] of outputs.entries()) {
+    args.push("-map", `${index}:v:0`, "-c:v", "copy", "-bsf:v", shape.filter, "-f", shape.demuxer, output);
+  }
+  const stderr = await runCapturingStderr(ffmpegBin, args, EXTRACT_TIMEOUT_MS);
+  try {
+    if (stderr === null) {
+      return null;
+    }
+    // One banner block per input, in the order they were given. Read rather
+    // than declared, so replacing a clip cannot silently invalidate the fit
+    // that rests on it.
+    const blocks = splitInputBlocks(stderr, clipPaths.length);
+    return await Promise.all(clipPaths.map(async (_, index) => {
+      const block = blocks[index];
+      if (!block) {
+        return null;
+      }
+      const clipInfo = parseClipCharacteristics(block);
+      const fps = parseFfmpegVideoFps(block);
+      if (!clipInfo || !(fps > 0)) {
+        return null;
+      }
+      let bytes;
+      try {
+        bytes = await readFile(outputs[index]);
+      } catch {
+        return null;
+      }
+      if (bytes.length === 0) {
+        return null;
+      }
+      return {
+        bytes,
+        demuxer: shape.demuxer,
+        megapixelsPerSecond: clipInfo.megapixelsPerSecond,
+        megabitsPerSecond: clipInfo.megabitsPerSecond,
+        fps
+      };
+    }));
+  } finally {
+    await rm(workDir, { recursive: true, force: true }).catch(() => {});
+  }
+}
+
+/**
+ * The part of an ffmpeg banner describing each input, in order.
+ *
+ * ffmpeg prints one `Input #N, …` block per input and then the stream mapping;
+ * the parsers here read a single input's facts, so they are given a single
+ * input's text rather than the whole banner.
+ *
+ * @param {string} stderr
+ * @param {number} count
+ * @returns {string[]}
+ */
+function splitInputBlocks(stderr, count) {
+  /** @type {string[]} */
+  const blocks = [];
+  for (let index = 0; index < count; index += 1) {
+    const from = stderr.indexOf(`Input #${index},`);
+    if (from < 0) {
+      blocks.push("");
+      continue;
+    }
+    const nextInput = stderr.indexOf(`Input #${index + 1},`, from);
+    const mapping = stderr.indexOf("Stream mapping:", from);
+    const ends = [nextInput, mapping].filter((at) => at > from);
+    blocks.push(stderr.slice(from, ends.length > 0 ? Math.min(...ends) : stderr.length));
+  }
+  return blocks;
+}
+
+/**
+ * Run ffmpeg to completion and return its stderr, or null when it failed or
+ * outlasted its bound.
+ *
+ * @param {string} ffmpegBin
+ * @param {string[]} args
+ * @param {number} timeoutMs
+ * @returns {Promise<string | null>}
+ */
+function runCapturingStderr(ffmpegBin, args, timeoutMs) {
+  return new Promise((resolve) => {
+    let stderr = "";
+    let settled = false;
+    let child;
+    const settle = (value) => {
+      if (settled) {
+        return;
+      }
+      settled = true;
+      clearTimeout(timer);
+      try {
+        child?.kill("SIGKILL");
+      } catch {
+        // already gone
+      }
+      resolve(value);
+    };
+    const timer = setTimeout(() => settle(null), timeoutMs);
+    try {
+      child = spawn(ffmpegBin, args, { stdio: ["ignore", "ignore", "pipe"], windowsHide: true });
+    } catch {
+      settle(null);
+      return;
+    }
+    child.stderr.on("data", (chunk) => {
+      stderr += String(chunk);
+    });
+    child.on("error", () => settle(null));
+    child.on("close", (code) => settle(code === 0 ? stderr : null));
+  });
+}
+
+/**
+ * Measure how fast this host DECODES a clip, from ffmpeg's own report of how
+ * much video it has processed.
+ *
+ * Two things are deliberately outside the measurement.
+ *
+ * **The process starting.** Wall-clock around the process cannot answer this:
+ * starting ffmpeg costs about a second, and on a quick machine a five-second
+ * clip decodes in a tenth of that, so the measurement would be of the program
+ * starting. Progress lines arrive AFTER it has started, and the slope between
+ * two of them — video processed against time taken — contains no part of the
+ * startup by construction.
+ *
+ * **The clip restarting.** This used to loop the clip with `-stream_loop -1`,
+ * and a loop is not free: measured 2026-08-22 on a desktop, a restart costs
+ * 0.03 s on the 480p clip and 0.12 s on the 1080p one — the decoder tearing
+ * down and re-allocating its frame buffers, which is why the price rises with
+ * the picture. A five-second clip decoded at 55x restarts eleven times a
+ * second, so that cost DOMINATED the reading: the same clips measured 53.7x
+ * looped against 80.3x in one continuous pass, and 11.8x against 15.8x. Worse,
+ * the bias is not shared — it depends on the clip's own resolution and on how
+ * fast the host is — so it does not cancel out of the fit, it tilts it. That is
+ * the fast-host failure recorded on 2026-08-20, where 1080p read cheaper than
+ * 720p, which is not a thing a decoder does.
+ *
+ * So the clip is fed to the decoder as ONE stream instead. An Annex-B
+ * elementary stream carries its parameter sets inline, so writing the same
+ * bytes again is simply more stream — the decoder never re-initialises, and
+ * there is no restart inside the window to measure. Verified against the
+ * continuous-pass truth on the same host: -0.2 % and -5.5 %, against -25 % and
+ * -33 % for the loop. Nothing is written to disk and the process is killed as
+ * soon as the window is wide enough.
+ *
+ * Exported because the property that broke here is checkable and was not being
+ * checked: a bigger picture must cost more than a smaller one of the same
+ * bitrate, and under the loop it did not.
+ *
+ * @param {string} ffmpegBin
+ * @param {string} clipPath
+ * @param {string} [family="h264"]
+ * @returns {Promise<{ speed: number, windowSec: number, megapixelsPerSecond: number, megabitsPerSecond: number } | null>}
+ */
+export async function measureDecodeSlope(ffmpegBin, clipPath, family = "h264") {
+  const streams = await extractFamilyStreams(ffmpegBin, [clipPath], family);
+  const stream = streams?.[0];
+  if (!stream) {
+    return null;
+  }
+  const measured = await decodePipedStream(ffmpegBin, stream);
+  return measured?.speed ? measured : null;
+}
+
+/**
+ * Decode an elementary stream fed from memory, and report the slope.
+ *
+ * @param {string} ffmpegBin
+ * @param {{ bytes: Buffer, demuxer: string, megapixelsPerSecond: number, megabitsPerSecond: number, fps: number }} stream
+ * @returns {Promise<{ speed: number, windowSec: number, megapixelsPerSecond: number, megabitsPerSecond: number } | null>}
+ */
+function decodePipedStream(ffmpegBin, stream, log = { info: () => {}, warn: () => {} }) {
+  return new Promise((resolve) => {
+    const args = [
+      "-hide_banner", "-loglevel", "error", "-nostats",
+      // A raw stream states no frame rate, so the one the container declared is
+      // given back to it. It decides how output time advances, and therefore
+      // what "seconds of video per second of clock" means.
+      "-f", stream.demuxer, "-framerate", String(stream.fps), "-i", "pipe:0",
+      "-an", "-f", "null", "-",
+      "-progress", "pipe:1"
+    ];
+    /** @type {Array<{ wallSec: number, outSec: number }>} */
+    const samples = [];
+    let stdout = "";
+    // Kept because this path depends on three things the old one did not: the
+    // raw demuxer accepting the frame rate, the bitstream filter having
+    // produced something parsable, and the fed concatenation being decodable.
+    // Without it the only trace of any of those failing is "said nothing".
+    let stderr = "";
+    let settled = false;
+    let child;
+    const startedAt = Date.now();
+    // Whether the FEED, not the decoder, could be what this reading measures
+    // (item 4(d2)). `child.stdin.write()` returning false was tried as the
+    // signal — never once true would mean this process was never ahead of the
+    // pipe — and measured false on every reading taken while writing this,
+    // including clips this same host decodes at 15-80x with room to spare, so
+    // it does not discriminate: `write()`'s return value tracks Node's own
+    // internal watermark against the size of what was just handed to it, not
+    // real drain state, and answered "no slack" identically whether the pipe
+    // or the decoder was the true limit. Rather than publish a verdict that
+    // reads the same in both cases, only the byte count is kept, for the
+    // MB/s figure below — a number to read, not a boolean to trust.
+    let bytesWritten = 0;
+    const finish = () => {
+      if (settled) {
+        return;
+      }
+      settled = true;
+      clearTimeout(timer);
+      try {
+        child?.stdin?.destroy();
+      } catch {
+        // already gone
+      }
+      try {
+        child?.kill("SIGKILL");
+      } catch {
+        // already gone
+      }
+      // The first sample still carries the startup — it reports whatever was
+      // processed while the process was coming up. Everything is measured from
+      // the second onwards.
+      const first = samples[1];
+      const last = samples[samples.length - 1];
+      if (!first || !last) {
+        resolve({ error: lastErrorLine(stderr) || "the decoder reported no progress" });
+        return;
+      }
+      const windowSec = last.wallSec - first.wallSec;
+      const producedSec = last.outSec - first.outSec;
+      if (!(windowSec >= DECODE_WINDOW_MIN_SEC) || !(producedSec > 0)) {
+        resolve({ error: lastErrorLine(stderr) || `the window was ${windowSec.toFixed(2)}s of ${producedSec.toFixed(2)}s produced` });
+        return;
+      }
+      const speed = producedSec / windowSec;
+      // Diagnostic only — logged, not acted on. Both figures are taken over the
+      // SAME window the speed itself is: bytesWritten is snapshotted alongside
+      // every progress sample, so this compares like against like rather than
+      // the achieved rate over the whole run (which starts before the first
+      // kept sample and reads systematically low against the window's own
+      // rate for no reason but that mismatch — measured while building this).
+      const windowBytes = last.bytesWritten - first.bytesWritten;
+      const achievedMBps = windowSec > 0 ? windowBytes / windowSec / 1e6 : 0;
+      const requiredMBps = ((stream.megabitsPerSecond * 1e6) / 8) * speed / 1e6;
+      // "Far apart" is stated, not left to the reader to eyeball: outside a
+      // factor of 1.5 either way is bigger than the write-timing slop this
+      // comparison carries on a healthy reading.
+      const farApart = requiredMBps > 0 && (achievedMBps / requiredMBps < 1 / 1.5 || achievedMBps / requiredMBps > 1.5);
+      log.info(
+        `hwaccel: decode pipe fed ${achievedMBps.toFixed(1)} MB/s, ${requiredMBps.toFixed(1)} MB/s ` +
+        `needed for ${speed.toFixed(2)}x` +
+        (farApart ? " — far enough apart to be worth a second look" : "")
+      );
+      resolve({
+        speed,
+        windowSec,
+        megapixelsPerSecond: stream.megapixelsPerSecond,
+        megabitsPerSecond: stream.megabitsPerSecond,
+        pipeThroughputMBps: achievedMBps
+      });
+    };
+    const timer = setTimeout(finish, DECODE_WINDOW_MAX_MS);
+    try {
+      child = spawn(ffmpegBin, args, { stdio: ["pipe", "pipe", "pipe"], windowsHide: true });
+    } catch (error) {
+      clearTimeout(timer);
+      settled = true;
+      resolve({ error: error instanceof Error ? error.message : String(error) });
+      return;
+    }
+    child.stderr.on("data", (chunk) => {
+      stderr += String(chunk);
+    });
+    // Keep the decoder fed. `write` returning false means the pipe is full, and
+    // the next copy goes on the `drain` — so the decoder is never starved and
+    // this process never buffers more than the pipe holds.
+    const writeOnce = () => {
+      const accepted = child.stdin.write(stream.bytes);
+      bytesWritten += stream.bytes.length;
+      return accepted;
+    };
+    const feed = () => {
+      while (!settled && child.stdin.writable && writeOnce()) {
+        // Written straight through; go round again.
+      }
+    };
+    child.stdin.on("drain", feed);
+    // The kill closes the pipe under the writer; that is the intended end.
+    child.stdin.on("error", () => {});
+    child.stdout.on("data", (chunk) => {
+      stdout += String(chunk);
+      let newline = stdout.indexOf("\n");
+      while (newline >= 0) {
+        const line = stdout.slice(0, newline).trim();
+        stdout = stdout.slice(newline + 1);
+        if (line.startsWith("out_time_ms=")) {
+          const microseconds = Number(line.slice("out_time_ms=".length));
+          if (Number.isFinite(microseconds)) {
+            samples.push({ wallSec: (Date.now() - startedAt) / 1000, outSec: microseconds / 1e6, bytesWritten });
+          }
+        }
+        newline = stdout.indexOf("\n");
+      }
+      if (samples.length >= 2 && samples[samples.length - 1].wallSec - samples[1].wallSec >= DECODE_WINDOW_MIN_SEC) {
+        finish();
+      }
+    });
+    child.on("error", (error) => {
+      if (settled) {
+        return;
+      }
+      clearTimeout(timer);
+      settled = true;
+      try {
+        child?.stdin?.destroy();
+      } catch {
+        // already gone
+      }
+      try {
+        child?.kill("SIGKILL");
+      } catch {
+        // already gone
+      }
+      resolve({ error: error instanceof Error ? error.message : String(error) });
+    });
+    child.on("close", finish);
+    feed();
+  });
+}
+
+
+/**
+ * How many times realtime this host can DECODE a source of these
+ * characteristics, from the startup fit. `null` when the fit is unavailable or
+ * the source figures are not known.
+ *
+ * @param {{ pixelTerm: number, bitrateTerm: number, constantTerm: number } | null} model
+ * @param {{ megapixelsPerSecond: number, megabitsPerSecond: number }} source
+ * @returns {number | null}
+ */
+export function decodeSpeedFor(model, source) {
+  if (!model) {
+    return null;
+  }
+  const pixels = Number(source?.megapixelsPerSecond);
+  const bits = Number(source?.megabitsPerSecond);
+  if (!Number.isFinite(pixels) || pixels <= 0 || !Number.isFinite(bits) || bits < 0) {
+    return null;
+  }
+  const cost = model.pixelTerm * pixels + model.bitrateTerm * bits + model.constantTerm;
+  if (!(cost > 0)) {
+    return null;
+  }
+  return 1 / cost;
+}
+
+/**
+ * How many times realtime a re-encode of this source at this output pixel rate
+ * would run: decoding and encoding share the machine, so their costs add and
+ * their speeds combine as
+ *
+ *   1 / (1/decodeSpeed + 1/encodeSpeed)
+ *
+ * Checked 2026-08-14 on the rung that broke playback: 1/(1/2.31 + 1/5.99) =
+ * 1.67× against 1.48× measured. With no decode fit this falls back to the
+ * encode speed alone — which is what the budget did before, and which
+ * overestimated that rung five to eleven times.
+ *
+ * @param {{ decodeModel: { pixelTerm: number, bitrateTerm: number, constantTerm: number } | null, encodePixelsPerSec: number, outputPixelsPerSec: number, source: { megapixelsPerSecond: number, megabitsPerSecond: number } | null }} params
+ * @returns {number | null}
+ */
+export function predictedRealtimeSpeed({
+  decodeModel,
+  encodePixelsPerSec,
+  outputPixelsPerSec,
+  source,
+  observedDecodeCostSec = null
+}) {
+  if (!Number.isFinite(encodePixelsPerSec) || encodePixelsPerSec <= 0) {
+    return null;
+  }
+  if (!Number.isFinite(outputPixelsPerSec) || outputPixelsPerSec <= 0) {
+    return null;
+  }
+  const encodeSpeed = encodePixelsPerSec / outputPixelsPerSec;
+  // What this very file has been seen to cost, when it has been: the clips are
+  // H.264 and a source that has to be re-encoded usually is not, so a figure
+  // taken from the encoder actually running on THIS source beats any model of
+  // a stand-in. It arrives seconds into playback and replaces the estimate.
+  const decodeSpeed = Number.isFinite(observedDecodeCostSec) && observedDecodeCostSec > 0
+    ? 1 / observedDecodeCostSec
+    : (source ? decodeSpeedFor(decodeModel, source) : null);
+  if (decodeSpeed === null) {
+    return encodeSpeed;
+  }
+  return 1 / (1 / decodeSpeed + 1 / encodeSpeed);
+}
+
+/**
+ * Whether this host can hold realtime, with the margin, while re-encoding this
+ * source to this output pixel rate — and the predicted speed either way, so a
+ * refusal can say what it refused on.
+ *
+ * The encoder figure is the FASTEST benchmarked preset: it is the best this
+ * host can do, so a rung it cannot hold cannot be held at any quality setting.
+ *
+ * @param {{ benchmark: Array<{ preset: string, pixelsPerSec: number }>, decodeModel?: object | null, source?: { megapixelsPerSecond: number, megabitsPerSecond: number } | null, outputPixelsPerSec: number, requiredSpeed?: number | null }} params
+ * @returns {{ speed: number | null, sustainable: boolean }}
+ */
+export function canSustainOutput({
+  benchmark,
+  decodeModel = null,
+  source = null,
+  outputPixelsPerSec,
+  observedDecodeCostSec = null,
+  concurrentCostSec = 0,
+  requiredSpeed = null,
+  // The frame this output is encoded at. Given, the encoder's figure is the
+  // one measured AT that size (`throughput.js`), and a size nothing was
+  // measured at is not sustainable: this host has not shown it can do it.
+  frame = null
+}) {
+  if (!Array.isArray(benchmark)) {
+    // Not calibrated at all: a wiring made without the startup calibration,
+    // which the proxy itself never is when it re-encodes (`server.js`). There is
+    // nothing to judge by, and nothing is judged.
+    return { speed: null, sustainable: true };
+  }
+  if (benchmark.length === 0) {
+    // CALIBRATED, and no mode of this encoder qualified (roadmap item 97, step
+    // 14). Unknown capacity is not capacity: nothing has shown this machine can
+    // hold a re-encode at all. The copied picture needs no encoder and is not
+    // asked here.
+    return { speed: null, sustainable: false };
+  }
+  const encodePixelsPerSec = cheapestPresetPixelsPerSec(benchmark, frame);
+  if (encodePixelsPerSec === null) {
+    return { speed: null, sustainable: false };
+  }
+  const observed = Number.isFinite(observedDecodeCostSec) && observedDecodeCostSec > 0
+    ? observedDecodeCostSec
+    : null;
+  if (observed === null && !isDecodePriced({ decodeModel, source })) {
+    // An encoder-only figure was several times too optimistic on the rung this
+    // check exists for, so it is not fit to refuse anything. Without the decode
+    // term the ladder is offered whole, exactly as it was before.
+    return { speed: null, sustainable: true };
+  }
+  const alone = predictedRealtimeSpeed({
+    decodeModel,
+    encodePixelsPerSec,
+    outputPixelsPerSec,
+    source,
+    observedDecodeCostSec: observed
+  });
+  // What ELSE will be running while this rung is. A rung is never the only
+  // thing on the machine: the picture it accompanies is being copied or
+  // encoded, an audio track may have its own encoder, and a warm-up is two
+  // encoders by design. Measured on the addon host, a copy alone takes about an
+  // eighth of the machine per second of video, and the field case of
+  // 2026-08-15 adds up exactly: 0.125 for the copy plus ~1.05 for the rung is
+  // more than the one second per second the machine has, which is what was
+  // observed.
+  //
+  // Zero when nothing else is known to be running, or when nothing has been
+  // measured yet — then this is a LOWER bound on the cost and the check is as
+  // permissive as it was before.
+  const speed = alone === null || !(concurrentCostSec > 0)
+    ? alone
+    : 1 / (1 / alone + concurrentCostSec);
+  if (speed === null) {
+    return { speed: null, sustainable: true };
+  }
+  return { speed, sustainable: speed >= speedBar(requiredSpeed) };
+}
+
+/**
+ * The speed a step has to reach to be worth offering.
+ *
+ * Realtime is not enough on its own: a step that produces exactly one second
+ * per second never recovers the seconds lost while its reader waits for the
+ * swarm, so it survives its own supply only if what it gains between
+ * interruptions covers what one interruption costs. That is measured per file
+ * and per swarm by the reader — `1 + worst wait / median interval`, in
+ * `torrent/supply-margin.js` — and on the field torrent of 2026-08-17 it came to 1.67
+ * against the 1.5 that used to stand here, and to 4.04-8.14 on a torrent whose
+ * swarm no encoder could have kept up with.
+ *
+ * Where that figure does not exist yet — fewer than two interruptions measured
+ * — the bar is realtime. It is the one thing that can be said without
+ * measuring the swarm, and the offer is restated as soon as the reader has
+ * something to say.
+ *
+ * @param {number | null | undefined} requiredSpeed - What this file's own
+ *   interruptions demand, when they have been measured.
+ * @returns {number}
+ */
+export function speedBar(requiredSpeed) {
+  return Number.isFinite(requiredSpeed) && requiredSpeed > REALTIME ? requiredSpeed : REALTIME;
+}
+
+/**
+ * The bar for a cost description — the supply's demand where decoding is
+ * priced, and never below the unpriced-decode bar where it is not.
+ *
+ * @param {{ decodeModel?: object | null, source?: object | null, observedDecodeCostSec?: number | null, requiredSpeed?: number | null }} cost
+ * @returns {number}
+ */
+function barFor(cost) {
+  const measured = speedBar(cost?.requiredSpeed);
+  return isDecodePriced(cost) ? measured : Math.max(UNPRICED_DECODE_BAR, measured);
+}
+
+/**
+ * Benchmark software libx264 presets on this host. Encodes a short synthetic
+ * clip at a fixed reference resolution with each preset and measures encoder
+ * throughput in pixels/second. The session manager uses this to pick, per
+ * stream, the highest-quality preset that still encodes the actual
+ * (source-capped) resolution faster than realtime.
+ *
+ * Runs once at startup; bounded by a per-encode timeout. Presets that fail are
+ * omitted from the result.
+ *
+ * @param {{ ffmpegBin: string, logger?: { info: (m: string) => void, warn: (m: string) => void } }} options
+ * @returns {Promise<Array<{ preset: string, pixelsPerSec: number }>>} Ordered slowest→fastest.
+ */
+export async function benchmarkSoftwarePresets({ ffmpegBin, logger, encoder = null }) {
+  const log = logger ?? { info: () => {}, warn: () => {} };
+
+  // REAL footage, decoded ONCE into raw frames, and the presets are then timed
+  // on those frames.
+  //
+  // Two reasons, both measured. The pattern this replaced (`testsrc2`) has flat
+  // areas and no grain and encodes 1.23x cheaper than film on the same machine
+  // and preset — an error that always points at offering a rung the host cannot
+  // hold. And feeding a compressed clip to each preset instead would put
+  // decoding and scaling inside the measurement: subtracting them afterwards
+  // compares a wall clock that includes process startup against a decode figure
+  // measured to exclude it, while inside one ffmpeg the two halves overlap. On
+  // the fastest preset — the one every ladder decision reads as the ceiling —
+  // that subtraction is most of the number being measured, so a small error in
+  // it becomes a large error in the answer.
+  //
+  // Raw frames remove all of it: no decoder, no scaler, nothing to subtract,
+  // and no dependence on the decode model. The cost is 25 MB of memory in a
+  // pipe for a few seconds.
+  const rawFramesPath = await decodeToRawFrames(ffmpegBin, log);
+  if (rawFramesPath === null) {
+    // Said once more, in the words that matter to whoever reads the log next:
+    // with no benchmark, `#sustainableHeights` filters nothing and every rung
+    // is offered, which is the failure of 2026-08-14 in full.
+    log.warn("hwaccel: the quality ladder is UNFILTERED on this host — nothing measured the encoder");
+    return [];
+  }
+  /** @type {Array<{ preset: string, pixelsPerSec: number }>} */
+  const results = [];
+  try {
+    // THE CHOSEN ENCODER'S OWN LADDER, whatever kind it is. NVENC walks p1…p7,
+    // QSV veryfast…veryslow, VAAPI its quality levels. A kind with no ladder is
+    // measured once, which is still a reading where there was none at all.
+    const ladder = encoder?.speedLadder;
+    const rungs = Array.isArray(ladder?.values) && ladder.values.length > 0 ? ladder.values : [null];
+    for (const rung of rungs) {
+      const speed = await measureEncodeSlope(ffmpegBin, encoder, rung, rawFramesPath);
+      if (speed === null) {
+        log.warn(`hwaccel: the benchmark of "${rung ?? encoder?.name}" produced no usable reading; skipping`);
+        continue;
+      }
+      const pixelsPerSec = BENCHMARK_REF_W * BENCHMARK_REF_H * TRANSCODE_FPS * speed;
+      results.push({ preset: rung ?? encoder?.name, pixelsPerSec });
+      log.info(
+        `hwaccel: ${encoder?.name} "${rung ?? "as it comes"}" ~= ${(pixelsPerSec / 1e6).toFixed(1)} Mpx/s ` +
+          `(${speed.toFixed(2)}x @ ${BENCHMARK_REF_W}x${BENCHMARK_REF_H}, real footage)`
+      );
+    }
+  } finally {
+    // The encoder was killed a moment ago and on Windows the handle outlives
+    // the signal, so removal is retried and its failure is not worth a session:
+    // this is a temp directory the operating system will clear anyway.
+    try {
+      rmSync(path.dirname(rawFramesPath), { recursive: true, force: true, maxRetries: 5, retryDelay: 200 });
+    } catch (error) {
+      log.warn(`hwaccel: could not remove the benchmark's raw frames: ${error instanceof Error ? error.message : String(error)}`);
+    }
+  }
+  return results;
+}
+
+/**
+ * How fast one preset encodes, from ffmpeg's own reports of how much video it
+ * has written — not from the clock around the process.
+ *
+ * Timing whole runs measures the run STARTING. Measured 2026-08-15 on a desktop
+ * that spawns ffmpeg in ~0.4 s: three seconds of raw frames encoded that way
+ * put `fast` and `ultrafast` within 1.24x of each other, when libx264's own
+ * presets differ by several times — the constant had swallowed the difference.
+ * The slope between two progress reports contains no part of the startup.
+ *
+ * The frames are written repeatedly so there is runway to measure over,
+ * whatever the preset's speed.
+ *
+ * @param {string} ffmpegBin
+ * @param {string} preset
+ * @param {string} rawFramesPath
+ * @returns {Promise<number | null>} Video seconds encoded per second of clock.
+ */
+/**
+ * Video seconds produced per second of clock, from ffmpeg's own reports.
+ *
+ * Startup is excluded by taking a DIFFERENCE: it lands in the wall clock of
+ * every report equally, so it cancels between two of them. (The decode
+ * benchmark drops its first report instead, because there the first one is
+ * emitted at out_time zero; here reports with no time yet are discarded before
+ * they arrive, so the first kept one is already running.)
+ *
+ * @param {Array<{ wallSec: number, outSec: number }>} samples
+ * @param {number} [minimumWindowSec=ENCODE_BENCHMARK_WINDOW_SEC]
+ * @returns {number | null}
+ */
+export function slopeOf(samples, minimumWindowSec = ENCODE_BENCHMARK_WINDOW_SEC) {
+  const first = samples[0];
+  const last = samples[samples.length - 1];
+  if (!first || !last || first === last) {
+    return null;
+  }
+  const took = last.wallSec - first.wallSec;
+  const produced = last.outSec - first.outSec;
+  if (!(took >= minimumWindowSec) || !(produced > 0)) {
+    return null;
+  }
+  const slope = produced / took;
+  // Nothing encodes a thousand times realtime. A figure above that is a
+  // measurement fault, and letting it through opens the whole ladder.
+  return slope <= ENCODE_BENCHMARK_MAX_PLAUSIBLE_SPEED ? slope : null;
+}
+
+export function measureEncodeSlope(ffmpegBin, encoder, rung, rawFramesPath, frame = { width: BENCHMARK_REF_W, height: BENCHMARK_REF_H }) {
+  return new Promise((resolve) => {
+    const args = [
+      "-hide_banner", "-loglevel", "error", "-nostats",
+      // What a device-backed encoder needs before the input — the device. Not
+      // its decoding setup: this is fed raw frames and there is nothing to
+      // decode, and `-hwaccel vaapi` over rawvideo fails to open.
+      ...(typeof encoder?.benchmarkInputArgs === "function" ? encoder.benchmarkInputArgs() : []),
+      "-stream_loop", "-1",
+      "-f", "rawvideo", "-pix_fmt", "yuv420p",
+      "-s", `${frame.width}x${frame.height}`, "-r", String(TRANSCODE_FPS),
+      "-i", rawFramesPath,
+      // THE ENCODER SAYS HOW TO MEASURE ITSELF. It was libx264 written here, so
+      // a host with NVENC, QSV, VAAPI or V4L2M2M measured its encoder not at all
+      // and the quality offer, which is arithmetic over pixels per second, had
+      // no pixels per second to work with.
+      ...encoder.benchmarkArgs(rung),
+      "-f", "null", "-",
+      "-progress", "pipe:1"
+    ];
+    /** @type {Array<{ wallSec: number, outSec: number }>} */
+    const samples = [];
+    let settled = false;
+    let buffered = "";
+    let child;
+    const startedAt = Date.now();
+    const finish = (value) => {
+      if (settled) {
+        return;
+      }
+      settled = true;
+      clearTimeout(timer);
+      try {
+        child?.kill("SIGKILL");
+      } catch {
+        // already gone
+      }
+      resolve(value);
+    };
+    const timer = setTimeout(() => finish(null), ENCODE_BENCHMARK_TIMEOUT_MS);
+    try {
+      child = spawn(ffmpegBin, args, { stdio: ["ignore", "pipe", "ignore"], windowsHide: true });
+    } catch {
+      finish(null);
+      return;
+    }
+    // The frames come from a FILE, read on repeat by ffmpeg itself. Fed through
+    // a pipe instead, the fastest presets measured the pipe: `ultrafast` on a
+    // desktop wants raw frames at hundreds of megabytes a second, which no
+    // writer here can supply, and the reading then describes the feeding rather
+    // than the encoder.
+    child.stdout.on("data", (chunk) => {
+      buffered += String(chunk);
+      let newline = buffered.indexOf(NEWLINE);
+      while (newline >= 0) {
+        const line = buffered.slice(0, newline).trim();
+        buffered = buffered.slice(newline + 1);
+        if (line.startsWith("out_time_ms=")) {
+          const outSec = Number(line.slice("out_time_ms=".length)) / 1e6;
+          // `N/A` is not the only way ffmpeg says "no position yet": some builds
+          // print the smallest signed 64-bit integer, which IS finite and would
+          // be taken for a position nine trillion seconds before the start.
+          if (Number.isFinite(outSec) && outSec >= 0) {
+            samples.push({ wallSec: (Date.now() - startedAt) / 1000, outSec });
+          }
+        }
+        newline = buffered.indexOf(NEWLINE);
+      }
+      const slope = slopeOf(samples);
+      if (slope !== null) {
+        finish(slope);
+      }
+    });
+    child.on("error", () => finish(null));
+    // A preset that finished before the window was wide enough is measured from
+    // whatever it did report, provided two reports exist at all.
+    // A preset that finished before the wide window was covered is still
+    // measured — but never over a window of nothing. Two reports a millisecond
+    // apart would divide a frame of video by that millisecond and call the host
+    // twenty times faster than it is, and one such reading becomes the figure
+    // every ladder decision is taken from.
+    child.on("exit", () => finish(slopeOf(samples, ENCODE_BENCHMARK_MIN_WINDOW_SEC)));
+  });
+}
+
+/**
+ * The benchmark's footage as raw frames: the calibration clip, looped to the
+ * benchmark's length and scaled to its size, decoded once.
+ *
+ * @param {string} ffmpegBin
+ * @param {{ info: (m: string) => void, warn: (m: string) => void }} log
+ * @returns {Promise<string | null>} Path to the raw frames, or null.
+ */
+export async function decodeToRawFrames(ffmpegBin, log, frame = { width: BENCHMARK_REF_W, height: BENCHMARK_REF_H }) {
+  // A benchmark may leave a host unmeasured; it may never stop it from
+  // starting. Before this the temp directory was made outside any guard, so a
+  // read-only or missing TMPDIR rejected the promise that starts the proxy.
+  let directory;
+  try {
+    directory = mkdtempSync(path.join(os.tmpdir(), "torrent-tv-bench-"));
+  } catch (error) {
+    log.warn(
+      `hwaccel: no writable temp directory for the preset benchmark (${error instanceof Error ? error.message : String(error)}); ` +
+      "presets unmeasured, so no quality rung will be refused on this host"
+    );
+    return null;
+  }
+  const rawPath = path.join(directory, "frames.yuv");
+  const args = [
+    "-hide_banner", "-loglevel", "error",
+    "-stream_loop", "-1",
+    "-i", path.join(CALIBRATION_DIR, CALIBRATION_REFERENCE_CLIP),
+    "-frames:v", String(rawFrameCountFor(frame)),
+    "-vf", `scale=${frame.width}:${frame.height},fps=${TRANSCODE_FPS}`,
+    "-an", "-f", "rawvideo", "-pix_fmt", "yuv420p", "-y", rawPath
+  ];
+  const { code } = await runFfmpeg(ffmpegBin, args, 30000);
+  const expectedBytes = frame.width * frame.height * 1.5 * rawFrameCountFor(frame);
+  let written = 0;
+  try {
+    written = statSync(rawPath).size;
+  } catch {
+    written = 0;
+  }
+  if (code !== 0 || written < expectedBytes * 0.9) {
+    log.warn(
+      "hwaccel: could not decode the calibration clip for the preset benchmark " +
+      `(${written} of ~${Math.round(expectedBytes)} bytes); presets unmeasured, ` +
+      "so no quality rung will be refused on this host"
+    );
+    try {
+      rmSync(directory, { recursive: true, force: true, maxRetries: 5, retryDelay: 200 });
+    } catch {
+      // A temp directory the operating system will clear; not worth a start-up.
+    }
+    return null;
+  }
+  return rawPath;
+}
+
+/**
+ * Pick the highest-quality (slowest) benchmarked preset that can encode
+ * `pixelsPerSecNeeded` with the speed margin. Falls back to the fastest
+ * benchmarked preset, or `"ultrafast"` when no benchmark is available.
+ *
+ * @param {Array<{ preset: string, pixelsPerSec: number }>} benchmark - slowest→fastest
+ * @param {number} pixelsPerSecNeeded
+ * @param {{ decodeModel?: object | null, source?: { megapixelsPerSecond: number, megabitsPerSecond: number } | null, requiredSpeed?: number | null }} [cost]
+ * @returns {string}
+ */
+/**
+ * What this host can do at its CHEAPEST preset — the ceiling of the ladder.
+ *
+ * Deliberately not the largest reading in the array. The list is in quality
+ * order, so its last measured entry is the cheapest preset; taking the maximum
+ * instead would let one noisy reading of an expensive preset raise the bar that
+ * decides which rungs are offered, and a rung offered on noise is a rung the
+ * host cannot hold. For choosing a preset the direction of that error is
+ * harmless; for deciding what to offer it is not, so the two use different
+ * statistics on purpose.
+ *
+ * @param {Array<{ preset: string, pixelsPerSec: number }>} benchmark
+ * @returns {number}
+ */
+function cheapestPresetPixelsPerSec(benchmark, frame = null) {
+  const cheapest = benchmark[benchmark.length - 1];
+  return cheapest ? throughputAt(cheapest, frame) : null;
+}
+
+export function pickSoftwarePreset(benchmark, pixelsPerSecNeeded, cost = {}, frame = null) {
+  if (!Array.isArray(benchmark) || benchmark.length === 0) {
+    return "ultrafast";
+  }
+  const observed = Number.isFinite(cost?.observedDecodeCostSec) && cost.observedDecodeCostSec > 0
+    ? cost.observedDecodeCostSec
+    : null;
+  const bar = barFor(cost);
+  // The FIRST entry that clears the bar wins — the list is in quality order, so
+  // that is the best picture this host can hold. Every entry is examined rather
+  // than the walk stopping at the first miss, because the measurements do not
+  // always ascend with the list: on a busy machine on 2026-08-15 `faster` read
+  // below `fast` twice.
+  for (const entry of benchmark) {
+    // At the size this output is encoded at: a preset not measured there is
+    // not a mode this host has shown it can hold there.
+    const encodePixelsPerSec = throughputAt(entry, frame);
+    if (encodePixelsPerSec === null) {
+      continue;
+    }
+    const speed = predictedRealtimeSpeed({
+      decodeModel: cost.decodeModel ?? null,
+      encodePixelsPerSec,
+      outputPixelsPerSec: pixelsPerSecNeeded,
+      source: cost.source ?? null,
+      observedDecodeCostSec: observed
+    });
+    if (speed !== null && speed >= bar) {
+      return entry.preset;
+    }
+  }
+  // Nothing clears the bar: the cheapest preset, which is the last in quality
+  // order. Returning whichever preset measured fastest would hand an expensive
+  // one to a host that has just been shown to hold no rung at all.
+  return benchmark[benchmark.length - 1].preset;
+}
+
+/**
+ * Whether a cost description can actually price decoding — a fit AND a source
+ * to apply it to. Without both, every prediction is encoder-only.
+ *
+ * @param {{ decodeModel?: object | null, source?: object | null }} cost
+ * @returns {boolean}
+ */
+function isDecodePriced(cost) {
+  if (Number.isFinite(cost?.observedDecodeCostSec) && cost.observedDecodeCostSec > 0) {
+    return true; // measured on the source itself, which needs no fit to stand on
+  }
+  return Boolean(cost?.decodeModel) && Boolean(cost?.source);
+}
+
+/**
+ * Build the resolution ladder for a ceiling box. Returns candidate output
+ * dimensions from the ceiling downward, preserving the ceiling's aspect ratio,
+ * each even-sized. The ceiling itself is always the top rung; the ladder's
+ * standard heights (`LADDER_HEIGHTS`, the same list the quality menu offers) at
+ * or above it are skipped (never upscale). Deduped by height.
+ *
+ * @param {number} ceilingWidth
+ * @param {number} ceilingHeight
+ * @returns {Array<{ width: number, height: number }>} high→low
+ */
+export function buildResolutionLadder(ceilingWidth, ceilingHeight) {
+  const cw = Number.isInteger(ceilingWidth) && ceilingWidth > 0 ? ceilingWidth : 0;
+  const ch = Number.isInteger(ceilingHeight) && ceilingHeight > 0 ? ceilingHeight : 0;
+  if (!cw || !ch) {
+    return [];
+  }
+  const even = (v) => {
+    const r = Math.round(v);
+    return Math.max(2, r - (r % 2));
+  };
+  /** @type {Array<{ width: number, height: number }>} */
+  const rungs = [{ width: cw, height: ch }];
+  for (const h of LADDER_HEIGHTS) {
+    if (h >= ch) {
+      continue; // at/above the ceiling — the ceiling rung already covers it
+    }
+    rungs.push({ width: even(cw * (h / ch)), height: h });
+  }
+  const seen = new Set();
+  return rungs.filter((rung) => {
+    if (seen.has(rung.height)) {
+      return false;
+    }
+    seen.add(rung.height);
+    return true;
+  });
+}
+
+/**
+ * Choose the software encode settings (resolution + preset) that fit the
+ * realtime budget on this host. From the resolution ladder (ceiling downward),
+ * pick the HIGHEST rung whose encode throughput — predicted from the startup
+ * benchmark's fastest preset — clears the speed this file's own supply
+ * demands (`speedBar`). Then, at that resolution, pick the highest-quality
+ * preset that still clears it. When even the lowest rung cannot clear it, use the lowest rung with
+ * the fastest preset (best effort — a smaller picture beats sub-realtime
+ * playback at full size). Returns null when no benchmark or ceiling is
+ * available (the caller keeps the ceiling resolution and the default preset).
+ *
+ * @param {Array<{ preset: string, pixelsPerSec: number }>} benchmark - slowest→fastest
+ * @param {{ width: number, height: number }} ceiling
+ * @param {number} outputFps
+ * @param {{ decodeModel?: object | null, source?: { megapixelsPerSecond: number, megabitsPerSecond: number } | null, requiredSpeed?: number | null }} [cost]
+ * @returns {{ width: number, height: number, preset: string, ladder: Array<{ width: number, height: number }>, rungIndex: number } | null}
+ */
+export function chooseSoftwareEncodeSettings(benchmark, ceiling, outputFps, cost = {}) {
+  if (!Array.isArray(benchmark) || benchmark.length === 0) {
+    return null;
+  }
+  const fps = Number.isFinite(outputFps) && outputFps > 0 ? outputFps : TRANSCODE_FPS;
+  const ladder = buildResolutionLadder(ceiling?.width, ceiling?.height);
+  if (ladder.length === 0) {
+    return null;
+  }
+  const bar = barFor(cost);
+  let chosenIndex = ladder.length - 1; // default: lowest rung (best effort)
+  for (let i = 0; i < ladder.length; i += 1) {
+    // The cheapest preset's throughput AT THIS RUNG's size.
+    const fastest = cheapestPresetPixelsPerSec(benchmark, ladder[i]);
+    if (fastest === null) {
+      continue;
+    }
+    const speed = predictedRealtimeSpeed({
+      decodeModel: cost.decodeModel ?? null,
+      encodePixelsPerSec: fastest,
+      outputPixelsPerSec: ladder[i].width * ladder[i].height * fps,
+      source: cost.source ?? null
+    });
+    if (speed !== null && speed >= bar) {
+      chosenIndex = i;
+      break;
+    }
+  }
+  const chosen = ladder[chosenIndex];
+  const preset = pickSoftwarePreset(benchmark, chosen.width * chosen.height * fps, cost, chosen);
+  return { width: chosen.width, height: chosen.height, preset, ladder, rungIndex: chosenIndex };
+}
+
+/**
+ * How fast this machine COPIES a picture, in seconds of film per second.
+ *
+ * The startup measurements price encoding and decoding, and a copied picture
+ * does neither: it reads packets and writes them out again. That left one whole
+ * branch of what this proxy does with no figure at all, and a figure is what
+ * every decision in the encoding layer is made from — where to put an encoder,
+ * how many to run, whether anybody will be left waiting. Without it a copied
+ * output was planned with no speed until its own run had been running long
+ * enough to report one, which is exactly the moment the plan matters most.
+ *
+ * Measured the same way as the others: ffmpeg's own progress, read as a slope
+ * over a window, so the process starting is outside the figure.
+ *
+ * The clip is joined to itself first rather than looped with `-stream_loop`.
+ * Looping charges a re-initialisation per lap — measured on the addon host at
+ * 0.03 s for 480p and 0.12 s for 1080p — and a copy of a five-second clip laps
+ * many times a second, so the reading would have been mostly re-initialisation.
+ *
+ * @param {{ ffmpegBin: string, logger?: { info: (m: string) => void, warn: (m: string) => void }, clipsDir?: string }} params
+ * @returns {Promise<number | null>} Seconds of film per second, or null where
+ *   the reading could not be taken. Null means unmeasured and is never a
+ *   substitute for a number.
+ */
+export async function benchmarkCopySpeed({ ffmpegBin, logger, clipsDir = CALIBRATION_DIR }) {
+  const log = logger ?? { info: () => {}, warn: () => {} };
+  const startedAt = Date.now();
+  // The largest clip in the set. A copy moves BYTES, so what it can do is a
+  // statement about the biggest pictures this host will be asked to pass
+  // through, and the small ones are covered by the same figure.
+  const clip = path.join(clipsDir, "cal-h264-1080-hi.mp4");
+  const speed = await measureCopySlope(ffmpegBin, clip);
+  if (!(speed > 0)) {
+    log.warn("hwaccel: copying could not be measured; a copied picture will be planned from its own run instead");
+    return null;
+  }
+  log.info(
+    `hwaccel: this host copies a picture at ${speed.toFixed(0)}x realtime ` +
+    `(measured in ${((Date.now() - startedAt) / 1000).toFixed(1)}s)`
+  );
+  return speed;
+}
+
+/**
+ * Seconds of film per second, copying one file, read from ffmpeg's progress.
+ *
+ * @param {string} ffmpegBin
+ * @param {string} filePath
+ * @returns {Promise<number | null>}
+ */
+function measureCopySlope(ffmpegBin, filePath) {
+  return new Promise((resolve) => {
+    const args = [
+      "-hide_banner", "-loglevel", "error", "-nostats",
+      // Played over and over, because a copy gets through a five-second clip in
+      // milliseconds and a slope needs a window to be taken over. Looping
+      // charges the demuxer being re-opened once a lap, so what comes out is a
+      // FLOOR on what this host can copy — the safe direction, since a plan made
+      // from it expects copying to be slower than it is.
+      "-stream_loop", "-1", "-i", filePath,
+      // What a copied output does: packets in, packets out, nothing decoded and
+      // nothing encoded. Written nowhere, so the figure is this machine's own
+      // handling and not the disk under a temp directory.
+      "-c", "copy", "-f", "null", "-",
+      // Progress is reported every half second by default, which over a window
+      // of one second is two readings. This asks for twenty.
+      "-stats_period", "0.05",
+      "-progress", "pipe:1"
+    ];
+    /** @type {Array<{ wallSec: number, outSec: number }>} */
+    const samples = [];
+    let settled = false;
+    let buffered = "";
+    let child;
+    const startedAt = Date.now();
+    const finish = (value) => {
+      if (settled) {
+        return;
+      }
+      settled = true;
+      clearTimeout(timer);
+      try {
+        child?.kill("SIGKILL");
+      } catch {
+        // already gone
+      }
+      resolve(value);
+    };
+    const timer = setTimeout(() => finish(null), ENCODE_BENCHMARK_TIMEOUT_MS);
+    try {
+      child = spawn(ffmpegBin, args, { stdio: ["ignore", "pipe", "ignore"], windowsHide: true });
+    } catch {
+      finish(null);
+      return;
+    }
+    child.stdout.on("data", (chunk) => {
+      buffered += String(chunk);
+      let newline = buffered.indexOf(NEWLINE);
+      while (newline >= 0) {
+        const line = buffered.slice(0, newline).trim();
+        buffered = buffered.slice(newline + 1);
+        if (line.startsWith("out_time_ms=")) {
+          const outSec = Number(line.slice("out_time_ms=".length)) / 1e6;
+          if (Number.isFinite(outSec) && outSec >= 0) {
+            samples.push({ wallSec: (Date.now() - startedAt) / 1000, outSec });
+          }
+        }
+        newline = buffered.indexOf(NEWLINE);
+      }
+      const slope = slopeOf(samples);
+      if (slope !== null) {
+        finish(slope);
+      }
+    });
+    child.on("error", () => finish(null));
+    child.on("exit", () => finish(slopeOf(samples, ENCODE_BENCHMARK_MIN_WINDOW_SEC)));
+  });
+}