@torrent-tv/proxy 2.89.6 → 2.89.7

package/services/encode/EncodePlan.js

@@ -1,1368 +1,1368 @@
-/**
- * @file How many encoders there should be on one output, and where each of them
- * belongs — decided from numbers alone.
- *
- * The decision is separated from carrying it out on purpose. Every rule below
- * was previously a condition somewhere inside an eleven-thousand-line file,
- * reachable only by starting a real ffmpeg, and each of them was written for
- * one viewer:
- *
- * - a run was placed at the position of whoever asked, and never at the first
- *   thing missing, so a viewer moving into a stretch already on disk restarted
- *   an encoder to make it a second time;
- * - a run had no end at all — neither `-to` nor `-t` appeared anywhere — so it
- *   ran until something killed it, and two runs on one output could not exist
- *   without writing over each other;
- * - nothing stopped a run that had caught up with material somebody else had
- *   already made.
- *
- * The rule this file exists to express, stated by the user 2026-09-04:
- *
- * > Viewers are always independent and always reuse what can be reused. The
- * > number of encoders is however many are needed; how many are needed follows
- * > from which sets of output parameters are wanted and where the viewers stand
- * > inside each. Segments produced by ANY encoder are available to ANY viewer,
- * > and which viewer asked never enters the question.
- *
- * So no name of a viewer reaches this file. It is given what is wanted, what
- * exists, what is being made, and what the machine can afford.
- *
- * **A viewer decides the ORDER the map is walked in and, through the budget, how
- * many processes walk it. Nothing else.** Stated by the user 2026-09-05, and it
- * is the rule the rest of this file now follows: while a file is being encoded
- * it is encoded WHOLE, in the order the map dictates. Who wants which segment
- * decides which gap is closed first, never whether a run may go on living.
- *
- * **A run is therefore never stopped for standing outside a viewer's window.**
- * It used to be, and the two decisions that produced that were in direct
- * contradiction — measured in the field 2026-09-05 on a viewer watching an
- * episode:
- *
- * 1. this file commanded a start inside the window, at #46;
- * 2. `planRunInterval` in the session manager moved the start to #78, because
- *    it counted a suspended run's claim as reaching `head + look-ahead`;
- * 3. this file then saw a run at #78 against a window of [27, 57], found no
- *    overlap, and killed it as "nothing it was given is wanted";
- * 4. neither coverage nor demand had changed, so the same start was commanded
- *    again — 350-700ms per cycle, dozens of times, no segment ever produced,
- *    the viewer's picture stopped for 125 seconds.
- *
- * Both of those other authorities are gone (roadmap item 76, step 5). What is
- * left is this file, and the only reasons it stops a run are: nobody is
- * watching the output at all; the machine affords fewer processes; or there is
- * nothing left unmade anywhere in the track.
- *
- * **A run's end comes from the coverage**, never from a window: it runs until
- * it meets material somebody else has made or is making, or until the end of
- * the film.
- */
-
-/**
- * One encoder that is running now.
- *
- * @typedef {object} LiveRun
- * @property {string} id
- * @property {number} from - The first number it was given.
- * @property {number} to - The last number it was given, inclusive.
- * @property {number} head - The next number it will produce. Its position.
- * @property {number} speedX - The run's processing speed against realtime:
- *   film made between closed pieces over its own working time
- *   (`EncodeRun.speedX`), with input waits and stops taken out. Zero or less
- *   means nothing has measured it yet, and then no comparison involving its
- *   speed can be made.
- */
-
-/**
- * What a viewer is waiting for. Which viewer is deliberately absent.
- *
- * @typedef {object} WantedSpan
- * @property {number} from
- * @property {number} to
- */
-
-/**
- * @typedef {{ type: "start", from: number, to: number, because: string }
- *   | { type: "move", run: object, from: number, to: number, because: string }
- *   | { type: "stop", run: object, because: string }
- *   | { type: "keep", run: object, from: number, to: number }} PlanAction
- */
-
-/**
- * Decide what to do with the encoders on one output.
- *
- * @param {object} params
- * @param {import("./CoverageMap.js").CoverageMap} params.coverage - What has
- *   been made and what is being made.
- * @param {WantedSpan[]} params.windows - What viewers are waiting for, one
- *   window each. Empty means nobody is watching this output.
- * @param {LiveRun[]} params.runs - The encoders running on it now.
- * @param {number} params.maxRuns - How many encoders this machine can afford on
- *   this output. Comes from the same arithmetic that decides the quality offer;
- *   it is measured per host and never chosen here.
- * @param {number} params.segmentSeconds - How much film one segment holds.
- * @param {number} [params.killCostSec] - How long stopping an encoder takes,
- *   measured on this host from its own runs. Zero until something has measured
- *   it, which makes moving one look cheaper than it is and is said here so the
- *   bias is known.
- * @param {number} [params.now] - The clock, injected. This layer is arithmetic
- *   and reads no clock of its own; how old a run is is one of its inputs.
- * @param {number} [params.firstByteWaitSec] - How long a fresh encoder takes to
- *   produce anything: process start, opening the input, and the first piece.
- *   Measured the same way. It replaced a constant of 0.12 s taken from one
- *   host and charged to every other.
- * @param {(others: number) => number} [params.contentionPenaltyFor] - How much
- *   slower ONE encoder runs with that many others beside it, measured on this
- *   host. Without it every extra process looks free, and the score then wants an
- *   encoder per piece: at exactly realtime each next piece is marginally late
- *   however many are running, so another one always seemed to help a little.
- *   Unmeasured is 1, and then the budget is the only thing bounding the count.
- * @returns {PlanAction[]} Stops first, then moves, then starts, so that a plan
- *   carried out in order never holds two encoders where it means to hold one.
- */
-export function planEncoders({
-  coverage,
-  windows,
-  runs,
-  maxRuns,
-  segmentSeconds,
-  killCostSec = 0,
-  firstByteWaitSec = 0,
-  now = Date.now(),
-  refetchSecPerFilmSecond = 0,
-  contentionPenaltyFor = () => 1,
-  speedX = 0
-}) {
-  /** @type {PlanAction[]} */
-  const stops = [];
-  /** @type {PlanAction[]} */
-  const moves = [];
-  /** @type {PlanAction[]} */
-  const starts = [];
-  /** @type {PlanAction[]} */
-  const keeps = [];
-
-  const wanted = Array.isArray(windows) ? windows : [];
-  const live = Array.isArray(runs) ? runs : [];
-
-  // Nobody is watching this output: every encoder on it is making segments for
-  // no one. This is the case a look-ahead cannot answer, because look-ahead
-  // asks how far AHEAD of a viewer a run is and there is no viewer.
-  if (wanted.length === 0) {
-    for (const run of live) {
-      stops.push({ type: "stop", run, because: "nobody is watching this output" });
-    }
-    return stops;
-  }
-
-  const untilNeeded = deadlineReaderFor(wanted, segmentSeconds);
-  // Segments produced per second, from the fastest measured encoder here.
-  // Seconds of film per second, divided by the film one piece holds.
-  //
-  // A RUN WORKING ON THIS OUTPUT OUTRANKS THE BENCHMARK. The startup figure is
-  // what this host does on reference clips; a run here is what it does on THIS
-  // material, and that is the more specific statement. Taken as a floor instead
-  // — the larger of the two — an encoder reporting half realtime was scored as
-  // though it ran at twice, and nothing was ever late.
-  const working = live.reduce((best, run) => Math.max(best, run.speedX || 0), 0);
-  const rate = segmentSeconds > 0 ? (working > 0 ? working : speedX) / segmentSeconds : 0;
-  // How long one piece takes AT THE RATE ACTUALLY IN FORCE. Concurrent encoders
-  // slow each other — measured on this host — so an arrangement's own encoder count
-  // decides it, and the delays below are computed per arrangement for that
-  // reason. Taken from the unpenalised rate instead, a piece looked cheaper the
-  // more encoders there were, which made extra encoders look free: the plan bought a
-  // second encoder where one served, and the arrivals it was compared on were
-  // computed at the slower rate all along.
-  //
-  // `Infinity` where nothing has been measured, which is what "no speed" means
-  // and what makes every arrangement equally hopeless rather than equally free.
-  const pieceAt = (howManyEncoders) => {
-    const inForce = rate / contentionPenaltyFor(Math.max(0, howManyEncoders - 1));
-    return inForce > 0 ? 1 / inForce : Number.POSITIVE_INFINITY;
-  };
-  // What a encoder costs to take away from where it stands and put somewhere else:
-  // its death, the start of another, and the wait for the first bytes there.
-  // Taking an encoder somewhere else is stopping this one and waiting for the
-  // next to produce. Both halves are measured on this host.
-  // WHAT A START AND A MOVE OWE BEFORE THE PIECE THEY STAND ON EXISTS, and
-  // neither is infinite.
-  //
-  // The measured `firstByteWaitSec` is spawn to first piece, so it already
-  // contains one piece's encoding. Separated, because the two scale differently:
-  // the piece costs more when encoders share the machine, the spawn does not.
-  //
-  //   spawn overhead = measured first output - what one piece costs alone
-  //   a fresh encoder owes    spawn overhead + the piece at the rate in force
-  //   a moved one owes        the kill, and then the same
-  //
-  // With nothing measured the overhead is zero and a fresh encoder owes exactly
-  // one piece — which is the honest floor rather than a guess: a piece cannot
-  // appear before it is encoded, and how fast this host encodes is measured
-  // before any viewer exists. There was an `Infinity` here for the cost of a
-  // move, on the reasoning that an unmeasured price must not license an
-  // irreversible act; it was an exception in a model that needs none, and this
-  // is the same statement made by arithmetic.
-  const spawnOverheadSec = Math.max(0, firstByteWaitSec - (rate > 0 ? 1 / rate : 0));
-  const moveSec = killCostSec + spawnOverheadSec;
-
-  // WHAT A RUN STILL HAS TO GO BEFORE IT PRODUCES ANYTHING — the measured time
-  // to a first piece, less the time it has already been alive.
-  //
-  // This is the memory the score was missing. It is computed afresh whenever
-  // anything changes, and every arrangement used to be priced as though it were
-  // the last decision anybody would take: a run that started 10 ms ago was
-  // assumed to produce instantly, so killing it and starting another looked like
-  // a straight gain. A move is justified by a benefit that arrives when the
-  // moved run produces something; taken again before it has, the benefit is
-  // never collected and the cost is paid twice, three times, forty times.
-  //
-  // A run 0.8 s old has 0.14 s left to go against 0.94 s to move it, so it is
-  // left alone; one working for half a minute has nothing left, and a move
-  // happens exactly when the film it would reach sooner is worth the restart. No
-  // state to keep and nothing to choose: one measured figure minus elapsed time.
-  const finishesItsPieceIn = (run, perPiece) => {
-    // WHEN THIS RUN FINISHES THE PIECE IT IS STANDING ON.
-    //
-    // Two cases and neither is a choice: a run that has produced something is
-    // working at a known rate, so the piece under it lands within one piece's
-    // worth of time; a run that has produced nothing is still in its warm-up,
-    // and what is left of that is the measured time to a first piece less the
-    // time it has already been alive.
-    //
-    // This is the memory the score was missing. It is computed afresh whenever
-    // anything changes, and each arrangement used to be priced as though it were
-    // the last decision anybody would take: a run 0.8 s into a 0.9 s piece was
-    // charged a whole piece for it, the same as one about to start, so killing
-    // it and beginning again elsewhere looked cheaper by ten milliseconds. A
-    // move is justified by a benefit that arrives when the moved run produces
-    // something; taken again before it has, the benefit is never collected and
-    // the cost is paid twice, three times, forty times.
-    //
-    // One measured figure minus elapsed time. Nothing to keep and nothing to
-    // choose.
-    if (Number(run.head) !== Number(run.from)) {
-      return perPiece;
-    }
-    const startedAt = Number(run.startedAt);
-    if (!Number.isFinite(startedAt) || startedAt <= 0) {
-      return firstByteWaitSec;
-    }
-    return Math.max(0, firstByteWaitSec - (now - startedAt) / 1000);
-  };
-
-  // WHAT EACH BODY OWES BEFORE THE PIECE IT STANDS ON EXISTS, priced at the rate
-  // the arrangement itself puts in force.
-  //
-  // Each encoder states its own debt where it is created, as a function of what one
-  // piece costs — because only there is it known what the encoder IS, and here only
-  // how many of them there are. So there is nothing to dispatch on: no kind, no
-  // tag, no case analysis. The count is known before any debt is needed, which
-  // is why this is one pass over the encoders rather than a figure computed once
-  // outside.
-  const priced = (encoders) => {
-    const perPiece = pieceAt(encoders.length);
-    return encoders.map((encoder) => ({ at: encoder.at, delaySec: encoder.owes(perPiece) }));
-  };
-
-
-  // WHERE A SECOND ENCODER MEETS THE ONE ALREADY ON A STRETCH, and whether it is
-  // worth having.
-  //
-  // A stretch of unmade film runs from `from` to `to`. Whoever is already on it
-  // stands at `from` and owes `w` before the piece under it exists; a fresh one
-  // placed at `x` owes `d` — its start, and then a whole piece — and both then
-  // work at the rate two encoders leave each other, which is measured on this
-  // host and is not half by assumption.
-  //
-  //   the one already there closes [from, x-1]:  w + (x - 1 - from) / r
-  //   the fresh one closes         [x, to]:      d + (to - x) / r
-  //
-  // The first rises with `x` and the second falls, so the stretch is closed
-  // soonest where they cross:
-  //
-  //   x* = (from + to + 1) / 2  +  (d - w) * r / 2
-  //
-  // The midpoint, shifted forward by half the difference of what the two owe,
-  // expressed in pieces. Halving is the special case `d = w`, which holds when
-  // both are fresh — and it was applied to every case, including the common one
-  // where `d` is a start plus a piece and `w` is the tail of a piece already
-  // being made.
-  //
-  // AND WHETHER TO SPLIT AT ALL, which halving never asked. Two encoders under
-  // contention against one at full speed:
-  //
-  //   one:  w + (to - from - 1) / rate
-  //   two:  w + (x* - 1 - from) / r
-  //
-  // Nothing back where the second does not win. On the addon host at 1920x1080
-  // the measured penalty for a second encoder is 1.98 — it takes very nearly
-  // all of the first's speed — so there a second almost never pays, and it was
-  // being placed regardless.
-  const splitAt = (from, to) => {
-    const pieces = to - from + 1;
-    if (!(pieces > 1) || !(rate > 0)) {
-      return null;
-    }
-    const together = rate / contentionPenaltyFor(live.length);
-    if (!(together > 0)) {
-      return null;
-    }
-    const perPieceTogether = 1 / together;
-    // What the one already on this stretch owes, or a fresh one's debt where
-    // nobody is on it — and then both sides owe the same and the meeting point
-    // is the middle, as it should be.
-    const onIt = live.find((run) => Number(run.head) === from) ?? null;
-    const owedThere = onIt
-      ? finishesItsPieceIn(onIt, perPieceTogether)
-      : firstByteWaitSec + perPieceTogether;
-    const owedFresh = firstByteWaitSec + perPieceTogether;
-    const meeting = (from + to + 1) / 2 + ((owedFresh - owedThere) * together) / 2;
-    const x = Math.min(to, Math.max(from + 1, Math.round(meeting)));
-    const withTwo = owedThere + (x - 1 - from) * perPieceTogether;
-    const withOne = owedThere + (pieces - 1) / rate;
-    return withTwo < withOne ? x : null;
-  };
-
-  // ------------------------------------------------------------------ WHERE
-  //
-  // A question about the FILM, and about nothing else: which numbers are
-  // missing, when each is needed, how fast this machine encodes, how many
-  // processes it can hold. No encoder that happens to be running enters it,
-  // which is why it can be answered by arithmetic.
-  const positions = placeEncoders({
-    coverage,
-    windows: wanted,
-    howMany: maxRuns,
-    // EVERY LIVE ENCODER IS PRE-PLACED, because that is what "somebody already
-    // gets here in time" means. A number one of them reaches before it is
-    // needed is not a position at all; a number none of them reaches is, and
-    // needs a encoder brought to it. There is no third case, and in particular no
-    // separate question of whether an encoder should drive on or be moved:
-    // driving is simply its arrival, and its arrival is priced in one place.
-    firstGap: gapFinderFor(coverage, new Set(live), rate, segmentSeconds * refetchSecPerFilmSecond),
-    deadlineAt: untilNeeded,
-    splitAt
-  });
-
-  // -------------------------------------------------------------------- WHO
-  //
-  // ARGMIN OF THE OBJECTIVE, EVALUATED. Not a rule that approximates it.
-  //
-  // Every way of filling the positions is scored by `latenessOf` and the best is
-  // taken. There are at most a handful of positions and a handful of encoders, so
-  // the enumeration is exact: no local rule stands in for the objective, and
-  // none can therefore disagree with another.
-  //
-  // Four such rules were written before this and all four had to go — "place
-  // where a number is late", "take a encoder that serves nothing", "take one whose
-  // work is needed later than this", "drive on or move, by cost". Each looked
-  // like a consequence of the model and each approximated it from a different
-  // side, so together they contradicted one another and the answer depended on
-  // which ran first.
-  /** One decision per live encoder, so none can be decided twice. @type {Map<object, PlanAction>} */
-  const decided = new Map();
-  const room = Math.max(0, maxRuns - live.length);
-  const refetchPerSegment = segmentSeconds * refetchSecPerFilmSecond;
-
-  let arrangements = [{ fill: [], used: new Set(), fresh: 0 }];
-  for (let index = 0; index < positions.length; index += 1) {
-    const next = [];
-    for (const arrangement of arrangements) {
-      next.push({ fill: [...arrangement.fill, null], used: arrangement.used, fresh: arrangement.fresh });
-      // A FRESH PROCESS IS OFFERED BEFORE ANY WORKING BODY, so that when the two
-      // score the same the working one is left alone. Taking it is free in the
-      // arithmetic — its output stays on disk — but it is not free in fact: the
-      // run it belongs to has a position, a warm input and a measured speed, and
-      // all three are thrown away for nothing.
-      if (arrangement.fresh < room) {
-        next.push({
-          fill: [...arrangement.fill, "new"],
-          used: arrangement.used,
-          fresh: arrangement.fresh + 1
-        });
-      }
-      for (const run of live) {
-        if (arrangement.used.has(run)) {
-          continue;
-        }
-        next.push({
-          fill: [...arrangement.fill, run],
-          used: new Set([...arrangement.used, run]),
-          fresh: arrangement.fresh
-        });
-      }
-    }
-    arrangements = next;
-  }
-
-  let best = null;
-  let bestScore = null;
-  for (const arrangement of arrangements) {
-    const encoders = [];
-    for (let index = 0; index < positions.length; index += 1) {
-      const filler = arrangement.fill[index];
-      if (filler === null) {
-        continue;
-      }
-      if (filler === "new") {
-        // A encoder that does not exist yet owes its own start and then the piece.
-        encoders.push({ at: positions[index], owes: (piece) => spawnOverheadSec + piece });
-        continue;
-      }
-      const head = Number(filler.head);
-      // Left where it stands it owes what is left of the piece under it; taken
-      // somewhere else it owes the killing, the start and a whole piece.
-      encoders.push({
-        at: positions[index],
-        owes: head === positions[index]
-          ? (piece) => finishesItsPieceIn(filler, piece)
-          : (piece) => moveSec + piece
-      });
-    }
-    // Bodies nobody was given a position for go on working where they stand,
-    // and their coverage counts: the file is encoded whole.
-    //
-    // A encoder given no end pays a restart the moment anybody is placed inside the
-    // road it would drive: where a run stops is fixed when its process starts,
-    // so it has to be cut and begun again at its own head. That price was
-    // invisible here, and an arrangement was scored as free when it was not.
-    for (const run of live) {
-      if (arrangement.used.has(run)) {
-        continue;
-      }
-      const head = Number(run.head);
-      const endless = Number(run.to) < Number(run.from);
-      const cutInFront = arrangement.fill.some((filler, index) =>
-        filler !== null && positions[index] > head
-        && (endless || positions[index] <= Number(run.to)));
-      encoders.push({
-        at: head,
-        owes: cutInFront
-          ? (piece) => moveSec + piece
-          : (piece) => finishesItsPieceIn(run, piece)
-      });
-    }
-    const scored = latenessOf(priced(encoders), coverage, wanted, untilNeeded, rate / contentionPenaltyFor(Math.max(0, encoders.length - 1)), refetchPerSegment, segmentSeconds);
-    if (bestScore === null || cheaperThan(scored, bestScore)) {
-      bestScore = scored;
-      best = arrangement;
-    }
-  }
-
-  // A BODY STANDING ON FILM THAT EXISTS is the one arrangement the enumeration
-  // above cannot reach: the gap in front of it is nobody's deadline, so it is
-  // never a position, and the encoder is left to make three hundred pieces a second
-  // time. Each such encoder is offered its own first gap and the SAME score decides
-  // — moving costs a restart on everything downstream, staying costs the repeat.
-  //
-  // Offered one at a time rather than folded into the enumeration because the
-  // enumeration is exponential in the number of positions, and this is called
-  // again on every piece produced. One extra evaluation per encoder against
-  // several thousand arrangements is the difference between arithmetic and a
-  // stalled proxy.
-  const placement = new Map();
-  for (let index = 0; index < positions.length; index += 1) {
-    const filler = best ? best.fill[index] : null;
-    if (filler && filler !== "new") {
-      placement.set(filler, positions[index]);
-    }
-  }
-  const encodersOf = (override) => {
-    const encoders = [];
-    for (const run of live) {
-      if (override.has(run) && override.get(run) === null) {
-        // Asked what the film looks like WITHOUT this one.
-        continue;
-      }
-      const at = override.has(run) ? override.get(run) : (placement.get(run) ?? Number(run.head));
-      const head = Number(run.head);
-      encoders.push({
-        at,
-        owes: at === head
-          ? (piece) => finishesItsPieceIn(run, piece)
-          : (piece) => moveSec + piece
-      });
-    }
-    for (let index = 0; index < positions.length; index += 1) {
-      if ((best ? best.fill[index] : null) === "new") {
-        encoders.push({ at: positions[index], owes: (piece) => spawnOverheadSec + piece });
-      }
-    }
-    return priced(encoders);
-  };
-  const scoreOf = (override) => {
-    const encoders = encodersOf(override);
-    return latenessOf(encoders, coverage, wanted, untilNeeded,
-      rate / contentionPenaltyFor(Math.max(0, encoders.length - 1)), refetchPerSegment, segmentSeconds);
-  };
-  for (const run of live) {
-    if (placement.has(run)) {
-      continue;
-    }
-    const gap = coverage.firstGapFrom(run.head, undefined, run);
-    if (gap === null || gap === Number(run.head)) {
-      continue;
-    }
-    const asIs = scoreOf(new Map());
-    const moved = scoreOf(new Map([[run, gap]]));
-    if (cheaperThan(moved, asIs)) {
-      placement.set(run, gap);
-    }
-  }
-
-  /**
-   * Would the film be worse off without this encoder? Asked of the same score.
-   *
-   * @param {object} run
-   * @returns {boolean}
-   */
-  const worseWithout = (run) => {
-    const kept = encodersOf(new Map());
-    const without = encodersOf(new Map([[run, null]]));
-    const scoreOf_ = (encoders) => latenessOf(encoders, coverage, wanted, untilNeeded,
-      rate / contentionPenaltyFor(Math.max(0, encoders.length - 1)), refetchPerSegment, segmentSeconds);
-    return cheaperThan(scoreOf_(kept), scoreOf_(without));
-  };
-
-  const stretchAt = (from) => endOfStretch(from, Math.min(
-    coverage.unmadeRunFrom(from),
-    coverage.freeRunFrom(from, new Set(live))
-  ));
-
-  for (let index = 0; index < positions.length; index += 1) {
-    if ((best ? best.fill[index] : null) !== "new") {
-      continue;
-    }
-    const at = positions[index];
-    starts.push({
-      type: "start",
-      from: at,
-      to: stretchAt(at),
-      because: `#${at} is wanted and nobody reaches it in time`
-    });
-  }
-
-  for (const run of live) {
-    const head = Number(run.head);
-    const at = placement.has(run) ? placement.get(run) : head;
-    if (at !== head) {
-      decided.set(run, {
-        type: "move",
-        run,
-        from: at,
-        to: stretchAt(at),
-        because: `standing at #${head} scores worse than standing at #${at}, counting ` +
-          "both how late the film would be and the work that would be done twice"
-      });
-      continue;
-    }
-    // It stays where it is — unless holding it changes nothing.
-    //
-    // A encoder left over from where a viewer used to be goes on costing the
-    // machine a process while another encoder already reaches everything it
-    // would. The score says so directly: take it away and see. Removing it is
-    // refused the moment it makes anything later or leaves film abandoned, so
-    // this cannot quietly drop the encoder somebody is waiting on.
-    if (!placement.has(run) && !worseWithout(run)) {
-      stops.push({
-        type: "stop",
-        run,
-        because: "the film is no worse off without it"
-      });
-      continue;
-    }
-    decided.set(run, { type: "keep", run, from: head, to: run.to });
-  }
-
-  // THE MACHINE'S LIMIT BINDS, whatever the map wants. It is measured — the
-  // processor, the swarm and the piece store each give a figure and the smallest
-  // wins — and an encoder over it is one the host cannot feed. Which of them
-  // goes is the same question as any other here: the one the film misses least,
-  // by the same score.
-  while (decided.size + starts.length > maxRuns) {
-    let cheapest = null;
-    let cheapestScore = null;
-    for (const [run, action] of decided) {
-      if (action.type !== "keep") {
-        continue;
-      }
-      const without = latenessOf(encodersOf(new Map([[run, null]])), coverage, wanted, untilNeeded,
-        rate / contentionPenaltyFor(Math.max(0, live.length - 2)), refetchPerSegment, segmentSeconds);
-      if (cheapestScore === null || cheaperThan(without, cheapestScore)) {
-        cheapestScore = without;
-        cheapest = run;
-      }
-    }
-    if (cheapest === null) {
-      break;
-    }
-    decided.delete(cheapest);
-    placement.delete(cheapest);
-    stops.push({
-      type: "stop",
-      run: cheapest,
-      because: `the machine holds ${maxRuns} encoder(s) on this output and this is the one ` +
-        "the film misses least"
-    });
-  }
-
-  for (const action of decided.values()) {
-    if (action.type === "keep") {
-      keeps.push(action);
-    } else {
-      moves.push(action);
-    }
-  }
-
-  // ONE ENCODER'S WORK ENDS WHERE THE NEXT ONE'S BEGINS.
-  //
-  // A free stretch may run to the end of the track, and an encoder given all of
-  // it stands in the road of every encoder placed behind it: they write the
-  // same names, and each one's output is the other's "material somebody else
-  // made", so they stop one another. Field 2026-09-05: three encoders started
-  // on one track within 200 ms, each into the road another was already writing.
-  // Fifteen readers on a piece store that holds sixteen pieces followed, half
-  // of all evictions took a piece a reader had declared, and `/stream` began
-  // handing out bytes that were not the file's.
-  //
-  // The bound is taken from the NEXT ENCODER'S START, not from a band edge: a
-  // band edge travels with the viewer, so every step forward would leave a
-  // sliver just past the previous encoder and buy an encoder for it.
-  // A RUN THAT IS STAYING IS IN THE SORT TOO, because its road can be taken.
-  //
-  // It used to be left out, on the reading that a run staying put keeps what it
-  // was given. That is true of the stretch it was GIVEN and false of the road it
-  // will actually drive: a run with no end carries no `-to` and walks to the end
-  // of the film, so an encoder placed in front of it writes the same names.
-  const placed = [...moves, ...starts, ...keeps].sort(
-    (left, right) => /** @type {any} */ (left).from - /** @type {any} */ (right).from
-  );
-  for (let index = 0; index < placed.length - 1; index += 1) {
-    const here = /** @type {{ type: string, run?: object, from: number, to: number }} */ (placed[index]);
-    const next = /** @type {{ from: number }} */ (placed[index + 1]);
-    if (here.to >= 0 && here.to < next.from) {
-      continue;
-    }
-    here.to = next.from - 1;
-    if (here.type !== "keep") {
-      continue;
-    }
-    // SHORTENING A LIVE RUN'S ROAD MEANS STOPPING IT, not merely writing a
-    // smaller number down. Where a run's end goes is fixed when its process
-    // starts, so one that was given none keeps producing past any bound decided
-    // later and would write one piece into the new encoder's road — two
-    // processes on one name, which is the collision this whole pass exists to
-    // prevent. So it ends here and begins again at its own head with a real end;
-    // the viewer in front pays a restart, which is a cost this file already
-    // prices rather than a interruption nobody counted.
-    keeps.splice(keeps.indexOf(here), 1);
-    moves.push({
-      type: "move",
-      run: here.run,
-      from: here.from,
-      to: here.to,
-      because:
-        `an encoder is needed at #${next.from}, which this run would reach only by ` +
-        "encoding through; it takes the road up to there and ends by itself"
-    });
-  }
-
-  return [...stops, ...moves, ...starts, ...keeps];
-}
-
-
-
-/**
- * THE OBJECTIVE, as a value that can be compared.
- *
- * THREE COUNTS OF SECONDS, COMPARED IN ORDER. The order is the user's, stated
- * 2026-09-06, and a later count decides only where the earlier ones tie:
- *
- * 1. SECONDS ANYBODY SPENDS LOOKING AT A SPINNER. Nothing outranks it, at any
- *    size. Walked forward in film order rather than summed piece by piece: a
- *    viewer who is stopped is not watching, so a wait moves every deadline
- *    behind it by its own length;
- *
- * 2. WHEN THE FILM IN FRONT OF THE VIEWERS IS FINISHED — the last piece of it to
- *    be made, whichever encoder makes it. A stretch no encoder will ever reach
- *    counts as never, which is what stops the front being abandoned;
- *
- * 3. WHEN THE WHOLE FILE IS FINISHED — the film behind the viewers included,
- *    plus what the swarm pays to fetch anything a second time. Film nobody is
- *    waiting for still has value: a viewer seeking back into a part that exists
- *    starts playing at once, and seeking back is what people do in the first
- *    minutes while they find their place. So spare capacity goes to finishing
- *    the file. This is where "the file is encoded WHOLE" lives; it used to be a
- *    penalty for film below the lowest encoder, which said the same thing as a
- *    patch and said it about one edge of the track only.
- *
- * WHY AN ENCODER MAY STAND BEHIND A VIEWER while film in front is still unmade:
- * encoders work at the same time, so one in front and one behind can finish the
- * file sooner than two in front. Where nobody is stalled and the front is closed
- * just as fast, the file being done sooner is the answer — count 3 deciding a
- * tie in 1 and 2, which is exactly what the order is for.
- *
- * WHY THE COUNTS ARE COMPARED AND NOT ADDED: seconds of somebody waiting and
- * seconds until a distant stretch exists are not the same thing, and no measured
- * quantity says how many of one are worth one of the other. Adding them would
- * mean choosing that exchange rate, which is inventing a number.
- *
- * @param {{ at: number, delaySec: number }[]} encoders - Where each encoder would
- *   stand, and how long before it produces anything there: nothing where it is
- *   already standing, a move or a start otherwise.
- * @param {import("./CoverageMap.js").CoverageMap} coverage
- * @param {WantedSpan[]} wanted
- * @param {(index: number) => number} untilNeeded
- * @param {number} rate - Segments per second. Always a real figure: this host
- *   measures what it encodes at on startup, before any viewer exists, and every
- *   run that works then refines it. There is no "unmeasured" case to answer.
- * @param {number} refetchSecPerSegment
- * @param {number} segmentSeconds
- * @returns {{ stall: number, ahead: number, whole: number }} Three counts of
- *   seconds, compared in that order by {@link cheaperThan}.
- */
-function latenessOf(encoders, coverage, wanted, untilNeeded, rate, refetchSecPerSegment, segmentSeconds) {
-  const first = Math.min(...wanted.map((span) => span.from));
-  const last = Math.max(...wanted.map((span) => span.to));
-  // What a number nobody reaches at all counts as. The film's own length is the
-  // honest bound — nothing can be later than never — and a finite figure is what
-  // lets two hopeless arrangements still be told apart by the rest of the sum.
-  const never = (last + 1) * segmentSeconds;
-
-  // WHICH SIDE OF THE VIEWERS a piece is on. The map states it; nothing here
-  // works it out from positions, and nothing here knows where a viewer stands.
-  //
-  // It was read off the deadline before — no time stated meant behind — and that
-  // is true only of a viewer who is playing. A paused viewer has no times
-  // anywhere, so their whole film read as behind them, "ahead before behind"
-  // had nothing to compare, and the encoder was free to wander to the start of
-  // the file. Which side a stretch is on and how soon it is wanted are two
-  // different facts, and the map states both.
-  // WHAT RANK THE MAP GIVES THIS NUMBER — the highest, where zones overlap,
-  // because a number two viewers want is wanted as much as the more urgent of
-  // them wants it.
-  //
-  // This replaced a boolean, "is it behind everybody", and the boolean was the
-  // whole of what the objective knew about the map's own order. The map states
-  // ten ranks on a film — p100 at the number a viewer is stopped on, doubling
-  // zones down to p91 for the far tail, p1 for what lies behind them — and all
-  // of that was collapsed into two buckets and then converted to seconds, where
-  // "never" for the film behind is the film's own length. On a 48-minute file
-  // that is 2024 s, which outvotes everything: field 2026-09-08, one viewer got
-  // three encoders, two of them on film behind them, and the run serving them
-  // was killed to make room for one.
-  const rankAt = (at) => {
-    let rank = 0;
-    for (const span of wanted) {
-      if (at >= span.from && at <= span.to) {
-        rank = Math.max(rank, Number(span.priority) || 0);
-      }
-    }
-    return rank;
-  };
-  // The ranks the map actually states, most urgent first. The comparison is over
-  // these and nothing else, so a rank can never be outvoted by a lower one
-  // however many seconds are at stake there.
-  const ranks = [...new Set(wanted.map((span) => Number(span.priority) || 0))]
-    .sort((left, right) => right - left);
-
-  // EVERY COUNT IS OVER THE FILM, NOT OVER THE ENCODERS. When a piece is made
-  // depends on which encoder reaches it soonest, and the encoder that reaches
-  // film in front of the viewers may well be standing behind them.
-  //
-  // Counted over the encoders instead — each charged to the side it stands on —
-  // the score had a hole that swallowed everything: an arrangement with every
-  // encoder BEHIND the viewers had nothing charged to the film in front, so its
-  // second term was zero, which is the best value there is. The plan then
-  // abandoned the film in front of a viewer and put both encoders at the start
-  // of the file, which is the opposite of the rule it is supposed to obey.
-  //
-  // AND THE WAITING IS WALKED FORWARD, not summed piece by piece.
-  //
-  // Not a sum of each piece's own lateness. A viewer who is stopped is not
-  // watching, so everything after the piece they are stopped on is needed that
-  // much later too: one wait moves every deadline behind it by its own length.
-  //
-  // Summed independently instead, the far tail of a long file outvoted the film
-  // under the viewer's feet — measured, and it placed the only encoder at #114
-  // while the viewer stood at #100, because thirteen pieces of certain waiting
-  // "cost" less than 886 distant pieces arriving a little later. Walking the
-  // clock forward makes that trade impossible: abandoning the near film delays
-  // the far film by at least as much.
-  let stalled = 0;
-  let wastedSwarm = 0;
-  /** Seconds anybody waits past a deadline, per rank. @type {Map<number, number>} */
-  const lateAt = new Map(ranks.map((rank) => [rank, 0]));
-  /** When the last number of a rank is made, per rank. @type {Map<number, number>} */
-  const doneAt = new Map(ranks.map((rank) => [rank, 0]));
-  for (let index = first; index <= last; index += 1) {
-    // Which encoder gets to this piece first, and when. One standing on it is
-    // already there; one behind it must work its way up, re-making anything
-    // already made on the way, which costs its own time and the swarm's.
-    let soonest = Number.POSITIVE_INFINITY;
-    let byWhom = null;
-    for (const encoder of encoders) {
-      if (encoder.at > index) {
-        continue;
-      }
-      // WHEN THIS BODY REACHES THIS NUMBER. `delaySec` is when it finishes the
-      // piece it is STANDING ON, so the pieces after it are the only ones still
-      // to be encoded at `rate`.
-      //
-      // It used to be `(index - at + 1) / rate` on top of the delay, which
-      // charges every encoder a whole piece for the one it is already working on. A
-      // fresh encoder does start from nothing, so for it that is right — and it is
-      // now inside its own delay. A run 0.8 s into a 0.9 s piece does not, and
-      // charging it 0.94 s for that piece is what made moving it look cheaper
-      // than leaving it: from #58 it was priced at 1.89 s to reach #59 against
-      // 1.88 s for a kill and a cold start, a difference of ten milliseconds
-      // that is nothing but the double charge. Field 2026-09-08: 39 moves in one
-      // session, 24 of them between three adjacent numbers, and the picture
-      // stood still for 116.7 s.
-      const arrival = encoder.delaySec
-        + (index - encoder.at) / rate
-        + coverage.madeBetween(encoder.at, index) * refetchSecPerSegment;
-      if (arrival < soonest) {
-        soonest = arrival;
-        byWhom = encoder;
-      }
-    }
-    if (coverage.isReady(index)) {
-      // It exists. Nobody waits for it and nothing is owed — but whoever passes
-      // over it makes it a second time, and the swarm fetches its bytes again.
-      if (byWhom !== null) {
-        wastedSwarm += refetchSecPerSegment;
-      }
-      continue;
-    }
-    // A piece nobody is working towards arrives never. There is no third case:
-    // the host measures what it encodes at, and what it copies at, before any
-    // viewer exists, so a speed is always a real number and an arrival can
-    // always be computed.
-    // Nothing arrives later than never, which is the bound the film's own length
-    // gives. It is a definition rather than a guard: it also makes the score
-    // total on a host whose startup measured nothing at all, where every arrival
-    // is beyond reckoning and every arrangement is therefore equally hopeless.
-    const when = byWhom === null ? never : Math.min(soonest, never);
-    const rank = rankAt(index);
-    doneAt.set(rank, Math.max(doneAt.get(rank) ?? 0, when));
-    const deadline = untilNeeded(index);
-    if (Number.isFinite(deadline)) {
-      const due = deadline + stalled;
-      const waited = Math.max(0, when - due);
-      lateAt.set(rank, (lateAt.get(rank) ?? 0) + waited);
-      stalled += waited;
-    }
-  }
-
-  return {
-    // THE MAP'S OWN ORDER, AS A VECTOR. One pair per rank the map states, most
-    // urgent rank first: how long anybody waits at that rank, then when the last
-    // number of it is made.
-    //
-    // Compared position by position, so a rank is never outvoted by a lower one
-    // — which is the whole of what was asked for: nobody stares at a spinner;
-    // then the film in front of the viewers is encoded as fast as it can be, band
-    // by band as the map ranks them; then, with whatever is left over and only
-    // then, the film behind them, in case somebody seeks back.
-    //
-    // No weights, and none possible: a weight would let seconds at one rank buy
-    // seconds at another, and it would be a number nobody measured. The map is
-    // the source of truth about what matters, and it already says so.
-    byRank: ranks.flatMap((rank) => [lateAt.get(rank) ?? 0, doneAt.get(rank) ?? 0]),
-    // HOW MANY ENCODERS IT TAKES. Ranked below every rank of the map and above
-    // the swarm's bill, so it cannot buy one where the map is indifferent — and
-    // the map IS indifferent about spare capacity, which is what bought an
-    // encoder for film nobody waits for.
-    encoders: encoders.length,
-    // WHAT THE SWARM PAYS for anything fetched twice, which delays everything.
-    wasted: wastedSwarm
-  };
-}
-
-/**
- * Is the first arrangement cheaper than the second?
- *
- * Three things in order, stated by the user 2026-09-06: nobody stares at a
- * spinner; then the film in front of the viewers is finished soonest; then the
- * whole file is. A later one decides only where the earlier ones tie.
- *
- * That order is why an encoder may stand BEHIND a viewer while film in front is
- * still unmade: encoders work at the same time, so one in front and one behind
- * can finish the file sooner than two in front — and where nobody is stalled and
- * the front is closed just as fast, the file being done sooner is the answer.
- *
- * @param {{ stall: number, ahead: number, whole: number }} left
- * @param {{ stall: number, ahead: number, whole: number }} right
- * @returns {boolean}
- */
-function cheaperThan(left, right) {
-  // POSITION BY POSITION, in the map's own order of ranks. A difference at a
-  // higher rank settles it, and nothing at a lower one can reopen it.
-  //
-  // There is no margin here and there must not be one. A threshold — "a move
-  // must beat staying by at least what moving costs" — was written while the
-  // arrival arithmetic charged a run for a piece it had already half made, and
-  // it was a prop under a comparison that was wrong rather than indifferent.
-  // With the arithmetic right the two are 1.08 s against 1.88 s, and nothing
-  // needs propping.
-  const size = Math.max(left.byRank.length, right.byRank.length);
-  for (let index = 0; index < size; index += 1) {
-    const here = left.byRank[index] ?? 0;
-    const there = right.byRank[index] ?? 0;
-    if (here !== there) {
-      return here < there;
-    }
-  }
-  // Where every rank is served identically, fewer encoders. The map is
-  // indifferent, so the machine decides: a process, a reader of the piece store
-  // and the swarm's bandwidth are all paid by the viewers the ranks above are
-  // about.
-  if (left.encoders !== right.encoders) {
-    return left.encoders < right.encoders;
-  }
-  return left.wasted < right.wasted;
-}
-
-/**
- * How long until a number is needed, read off the map.
- *
- * The map states it per stretch, for the stretch's NEAR EDGE, because a stretch
- * is met at its beginning. Every number inside is needed no sooner than that, so
- * taking the stretch's figure for all of them is the safe reading: it can only
- * make the filling earlier than it has to be, never later.
- *
- * Where the map says nothing, nobody is coming and nothing can be late.
- *
- * Inside a stretch the time GROWS with the distance, because a viewer covers a
- * second of film in a second: the number `n` places past the near edge is
- * reached `n` segments of film later. Taking the near edge's figure for every
- * number inside instead makes a whole stretch due at once — measured while
- * building this: the first stretch is as wide as the measured allowance, so its
- * far end was demanded instantly and an encoder was placed on a number another
- * one was already writing.
- *
- * @param {WantedSpan[]} windows
- * @param {number} segmentSeconds - How much film one number holds.
- * @returns {(index: number) => number}
- */
-/** @param {WantedSpan[]} windows */
-function firstOf(windows) {
-  return Math.min(...windows.map((span) => span.from));
-}
-
-/** @param {WantedSpan[]} windows */
-function lastOf(windows) {
-  return Math.max(...windows.map((span) => span.to));
-}
-
-function deadlineReaderFor(windows, segmentSeconds) {
-  const perSegment = segmentSeconds > 0 ? segmentSeconds : 0;
-  return (index) => {
-    let soonest = Number.POSITIVE_INFINITY;
-    for (const span of windows) {
-      if (index < span.from || index > span.to) {
-        continue;
-      }
-      const stated = /** @type {{ withinSeconds?: number }} */ (span).withinSeconds;
-      // A stretch stated with no time is somebody waiting at its near edge: that
-      // is what stating one means. The rest of it grows with the distance, the
-      // same as a stated one — read as due all at once instead, a window as wide
-      // as a viewer's cushion demanded its far end instantly and bought an
-      // encoder to stand beside one already working.
-      // `null` IS A STATEMENT AND IT SAYS NOBODY IS COMING. `undefined` is the
-      // absence of one, and a caller that knows only a position is somebody
-      // waiting at it.
-      //
-      // Read through `Number()`, `null` becomes 0 — due NOW — so the film BEHIND
-      // the viewers, which the map marks with exactly that, was the most urgent
-      // material in the file. Everything followed from it: it bought encoders,
-      // it took the run standing in front of the viewer because that run was the
-      // nearest encoder to it, and it did so again on every pass. Field 2026-09-08:
-      // 39 moves in one session, 24 between three adjacent numbers, one viewer
-      // on three encoders, and the picture stood still for 116.7 s in three
-      // interruptions, the worst of them 91.8 s.
-      //
-      // The map has always said it plainly — `{"from":0,"to":57,"priority":1,
-      // "withinSeconds":null,"behind":true}` is in the log of every session — and
-      // this line turned it into its opposite. Fourth time in this repository
-      // that the input to a calculation was not what the calculation assumed.
-      const within = stated === undefined ? 0 : (stated === null ? Number.NaN : Number(stated));
-      if (!Number.isFinite(within)) {
-        // Stated as no time at all: nobody is coming here.
-        continue;
-      }
-      const here = within + (index - span.from) * perSegment;
-      if (here < soonest) {
-        soonest = here;
-      }
-    }
-    return soonest;
-  };
-}
-
-/**
- * WHERE ENCODERS BELONG, from the model rather than from a list of cases.
- *
- * The problem this solves, stated exactly:
- *
- * - the track is a line of segment numbers; `M` are the ones not made;
- * - each `x` carries a DEADLINE `D(x)`, the seconds until somebody needs it.
- *   That is what the priority map is a reading of — a viewer moving forward
- *   covers a second of film in a second, so the time until they are at `x` is
- *   the distance to it. `Infinity` where nobody is coming;
- * - an encoder is a SEQUENTIAL producer: placed at `a`, it delivers `a + j` at
- *   time `(j + 1) / r`, where `r` is segments per second, measured. It cannot
- *   skip, so its whole schedule follows from where it starts;
- * - the machine affords `k` of them, measured.
- *
- * Two consequences fall out and need no rule of their own. Placements
- * `a_1 < ... < a_k` PARTITION the line: encoder `i` is useful only on
- * `[a_i, a_{i+1})`, because past that its neighbour got there first. And a
- * segment served by encoder `i` arrives at `(x - a_i + 1) / r`, which is
- * therefore also the answer to "when would the encoder already placed before it
- * get here" — the second half of the comparison, and the half that was missing.
- *
- * `x` is LATE when it arrives after `D(x)`. The objective is no late segments;
- * where `k` does not stretch to that, lateness beginning as far to the right as
- * possible.
- *
- * THE ALGORITHM is first-fit, left to right:
- *
- *     for each missing x with a finite deadline, ascending:
- *         if some encoder already placed at a satisfies (x - a + 1)/r <= D(x):
- *             it covers x
- *         else:
- *             place an encoder at x
- *
- * A LIVE run enters as an encoder already placed at its own head. There is no
- * special case for it.
- *
- * WHY IT IS OPTIMAL. The leftmost missing number with a finite deadline must be
- * covered by somebody. An encoder placed exactly on it delivers it at the
- * earliest time any placement can, `1/r`, and covers the longest suffix any
- * placement can — starting further left only re-makes material and arrives
- * later, starting further right does not cover it at all. So the greedy choice
- * is never worse than any other, and the usual exchange argument carries it to
- * the whole line. This is the known result for FIXED-ORDER scheduling with
- * deadlines, where first-fit is optimal at unit processing times, and a segment
- * is one unit. General machine minimisation with release times and deadlines is
- * NP-hard; this case is polynomial because the order is forced and each machine
- * covers a contiguous stretch.
- *
- * WHAT WAS TRIED FIRST AND WAS WRONG, kept because each looked reasonable:
- *
- * - `(h - p) * s / (1 - s)`, how long a run stays in front of a viewer moving
- *   forward. It answers a different question: a viewer stopped with an empty
- *   buffer needs the segment now, and at exactly realtime that formula says
- *   "for ever" while the viewer waits thirteen minutes;
- * - whether a run's head lies inside a wanted band — which ties an encoder to
- *   whoever is standing there, and this layer must never know that;
- * - the run's head as a barrier, everything above it placeable. It has no time
- *   in it at all, so it cannot tell two segments ahead from two hundred.
- *
- * Each was a case, not a model. The deadline is the model.
- *
- * @param {import("./CoverageMap.js").CoverageMap} coverage
- * @param {Set<object>} surviving - Runs that will still be alive, as encoders
- *   already placed at their own heads.
- * @param {number} rate - Segments produced per second by one encoder, measured.
- *   Zero when nothing has measured it, and then no arrival time can be computed
- *   and every claimed number is left alone.
- * @returns {(at: number, bound: number, deadlineAt: (index: number) => number, alsoPlaced?: number[]) => number | null}
- */
-function gapFinderFor(coverage, surviving, rate, refetchSecPerSegment = 0) {
-  /** Encoders already placed: where each stands, and how far its road runs. */
-  const placed = [];
-  for (const run of surviving) {
-    const head = Number(/** @type {{ head?: number }} */ (run).head);
-    placed.push({
-      at: Number.isFinite(head) ? head : Number(/** @type {{ from: number }} */ (run).from),
-      // A live run's road, so that placing inside it can be priced. A run given
-      // no end drives to the end of the film, which is what makes the price real.
-      // A run given no end drives to the end of the film, which is what makes
-      // the price of cutting in front of it real. Written out rather than
-      // imported: this file depends on nothing, and that is what lets it be
-      // exercised with plain values alone.
-      to: Number(/** @type {{ to: number }} */ (run).to) < Number(/** @type {{ from: number }} */ (run).from)
-        ? Number.POSITIVE_INFINITY
-        : Number(/** @type {{ to: number }} */ (run).to)
-    });
-  }
-  return (at, bound, deadlineAt, alsoPlaced) => {
-    const start = Number.isInteger(at) && at > 0 ? at : 0;
-    const last = Number.isInteger(bound) ? bound : -1;
-    // THE LATE NUMBER THAT IS DUE SOONEST, not the leftmost one.
-    //
-    // With room for every placement the two are the same answer. With a budget
-    // that binds they are not, and the objective decides: lateness pushed as far
-    // to the right as possible means the soonest deadline is served first. A
-    // walk by number gave the one machine to a viewer due in ten minutes while
-    // another stood waiting with an empty buffer.
-    //
-    // Ties go to the smaller number, so the answer does not depend on the order
-    // the map happens to be in.
-    let best = null;
-    let bestDue = Number.POSITIVE_INFINITY;
-    for (let index = start; index <= last; index += 1) {
-      if (coverage.isReady(index)) {
-        continue;
-      }
-      const deadline = deadlineAt(index);
-      if (!Number.isFinite(deadline)) {
-        // NOBODY IS COMING HERE, so nothing can be late — but the film is still
-        // wanted, and this is where spare capacity goes. The number is proposed;
-        // whether an encoder is actually spent on it is the score's answer, and
-        // the score puts anything anybody is waiting for first.
-        return index;
-      }
-      // When would the SOONEST of those already placed get here? Encoders placed
-      // EARLIER IN THIS PASS count: the first one placed for a viewer covers the
-      // stretch in front of them, and without counting it the walk placed a
-      // second and a third on the very next numbers — three processes a segment
-      // apart for one person, which is the waste this model exists to refuse.
-      let soonest = Number.POSITIVE_INFINITY;
-      for (const a of [...placed.map((live) => live.at), ...(alsoPlaced ?? [])]) {
-        if (a > index) {
-          // Standing past it. Encoders only move forward, so it never will.
-          continue;
-        }
-        if (a === index) {
-          // Standing ON it. No placement is faster than the one already made.
-          soonest = 0;
-          break;
-        }
-        // WHEN THIS BODY GETS HERE, and both terms of it.
-        //
-        // Its own encoding of everything between, and the swarm's price for the
-        // film it would fetch a SECOND time — every number between that is
-        // already made, it makes again. That second term is why "should this
-        // encoder drive on or be moved" is not a question of its own: an
-        // encoder with three hundred made pieces in front of it is simply slow
-        // to arrive, and the model compares arrivals. Asked separately it was a
-        // second authority over the same encoder, and the two disagreed.
-        const arrival = (index - a + 1) / rate
-          + coverage.madeBetween(a, index) * refetchSecPerSegment;
-        if (arrival < soonest) {
-          soonest = arrival;
-        }
-      }
-      if (soonest <= deadline) {
-        // Somebody gets here in time. Nothing to decide.
-        continue;
-      }
-      // IT IS LATE, AND THAT IS ALL THIS DECIDES. Whether filling it is worth
-      // the price is not asked here: this only proposes candidates, and the
-      // score decides how many of them are taken and by whom. Asked here as
-      // well, it was a second cost model beside the objective — with its own
-      // idea of what a process costs — and the two disagreed at exactly
-      // realtime, where every next piece is marginally late and each looked
-      // worth its own encoder.
-      if (deadline < bestDue) {
-        best = index;
-        bestDue = deadline;
-      }
-    }
-    return best;
-  };
-}
-
-/**
- * The last number of a stretch that begins at `from` and is `length` long.
- *
- * `-1` when the length is not finite, which is this layer's word for a run with
- * no end: the film's length is not known, so there is nothing to stop it at, and
- * a number invented here would be an end nobody measured.
- *
- * @param {number} from
- * @param {number} length
- * @returns {number}
- */
-function endOfStretch(from, length) {
-  return Number.isFinite(length) ? from + Math.max(1, length) - 1 : -1;
-}
-
-/**
- * The lowest number a viewer is waiting for that is not ready — what the plan
- * is judged by.
- *
- * Not used to decide anything: it is the figure a log line carries, so that a
- * plan that keeps producing while a viewer waits is visible rather than
- * inferred.
- *
- * @param {import("./CoverageMap.js").CoverageMap} coverage
- * @param {WantedSpan[]} windows
- * @returns {number | null}
- */
-export function firstUnmetWant(coverage, windows) {
-  let lowest = null;
-  for (const span of windows ?? []) {
-    for (let at = span.from; at <= span.to; at += 1) {
-      if (!coverage.isReady(at)) {
-        if (lowest === null || at < lowest) {
-          lowest = at;
-        }
-        break;
-      }
-    }
-  }
-  return lowest;
-}
-
-/**
- * Where to put the encoders this machine can afford.
- *
- * Two things are wanted of a division of the film, and they are wanted in this
- * order:
- *
- * 1. **the viewer must not stop.** An encoder starting at `q` stays ahead of a
- *    viewer at `p` while `y / s <= q + y - p`, so it holds `(q - p) * s / (1-s)`
- *    of film and no more. Beyond that the viewer catches it, and the next
- *    encoder has to be standing there. That is where the first ones go, and it
- *    is why the stretches grow: the further off one starts, the later the
- *    viewer arrives and the longer it may work;
- * 2. **the film should be finished as soon as possible.** Once the viewer is
- *    safe, whatever is left is divided EQUALLY between the encoders that
- *    remain: equal shares finish together, and any other division finishes when
- *    its longest share does. That is what makes seeking cheap — the film exists.
- *
- * At or above realtime the first requirement is met by one encoder for the
- * whole film, and every other encoder goes to the second — which is the common
- * case on a copied picture, and is why "one viewer, one encoder" was never the
- * rule.
- *
- * @param {object} params
- * @param {import("./CoverageMap.js").CoverageMap} params.coverage
- * @param {WantedSpan[]} params.windows - The merged map, in this output's own
- *   numbering. Its highest-numbered band starts where the viewer is.
- * @param {number} params.howMany - What the machine affords.
- * @param {number} params.speedX - Measured. Zero when nothing has measured it,
- *   and then only the first requirement can be served.
- * @param {(at: number, bound: number, deadlineAt: (index: number) => number, placed: number[]) => number | null} [params.firstGap] -
- *   Where a gap may be opened. Defaults to the map's own answer; the plan hands
- *   in one that also counts a number a live run has claimed but will not reach
- *   before it is needed, which is the only way anybody beyond a working encoder
- *   is served.
- * @param {(index: number) => number} [params.deadlineAt] - Seconds until that
- *   number is needed. `Infinity` where nobody is coming. Absent means every
- *   stated want is due now.
- * @returns {number[]} Where to start each encoder, ascending.
- */
-export function placeEncoders({
-  coverage,
-  windows,
-  howMany,
-  firstGap = null,
-  deadlineAt = null,
-  // WHERE TWO ENCODERS SHARE A STRETCH, and whether a second is worth having at
-  // all. Derived by whoever holds the measurements — the rates, the contention
-  // penalty, what a start costs — because this function is positional and holds
-  // none of them. The default is the midpoint, which IS the answer when both are
-  // fresh and owe the same, and is what a caller with nothing measured falls
-  // back on.
-  splitAt = (from, to) => from + Math.floor((to - from + 1) / 2)
-}) {
-  if (!(howMany > 0) || windows.length === 0) {
-    return [];
-  }
-  // NOW when the caller says nothing. A stated want with no time is somebody
-  // waiting on it — that is what stating one means — so the honest reading is
-  // that it is due. `Infinity` is a statement in its own right and has to be
-  // made: it says nobody is coming.
-  const untilNeeded = deadlineAt ?? (() => 0);
-  /** Where this pass has placed so far — each one covers what it can reach. */
-  const placedHere = [];
-  const gapAt = firstGap
-    ? (at, bound) => firstGap(at, bound, untilNeeded, placedHere)
-    : (at, bound) => coverage.firstGapFrom(at, bound);
-
-
-  // CANDIDATES COME FROM THE PRIORITY MAP, IN THE ORDER THE MAP STATES.
-  //
-  // The map already answers every question that was being re-derived here. Its
-  // ranks say what matters most — the number a viewer is stopped on, then what
-  // is in front of them band by band, then the rest of the track, and last of
-  // all what lies behind them. A pause flattens those ranks; a seek moves them;
-  // a second viewer merges into them. So walking the map in its own order is
-  // what "ahead before behind" means, and nothing here has to work out where the
-  // viewers are.
-  //
-  // It was not read that way. This walked the film by number and proposed
-  // whatever was late, then a second pass divided the leftovers — an order of
-  // its own invention, which put the beginning of the file before the film in
-  // front of a viewer and, at one point, proposed #0, #1 and #2 as three
-  // separate places.
-  //
-  // One candidate per zone: the first number in it nobody has and nobody
-  // reaches in time. Zones with no deadline can have nothing late in them, so
-  // there it is simply the first number nobody has — which is how spare capacity
-  // comes to finish the file.
-  /** @type {number[]} */
-  const places = [];
-  const byRank = [...windows].sort(
-    (left, right) => (right.priority ?? 0) - (left.priority ?? 0) || left.from - right.from
-  );
-  for (const zone of byRank) {
-    if (places.length >= howMany) {
-      break;
-    }
-    const at = gapAt(zone.from, zone.to);
-    if (at === null || places.includes(at)) {
-      continue;
-    }
-    places.push(at);
-    placedHere.push(at);
-  }
-
-  // AND WHERE TO SPLIT WHAT IS LEFT, for the capacity the map has not spent.
-  //
-  // The search above proposes only what is LATE, so once one encoder covers a
-  // zone in time that zone proposes nothing more — and a machine that holds four
-  // ran one. Finishing a contiguous stretch soonest with several machines of the
-  // same speed means dividing it between them, which is where these come from:
-  // the widest run of film between two encoders, split.
-  //
-  // Proposing is not spending. The score decides whether another process is
-  // worth it, and its first term — how late the film is — always outranks its
-  // second, so this can never take capacity from somebody waiting.
-  while (places.length < howMany) {
-    const edges = [...places].sort((left, right) => left - right);
-    let widestFrom = null;
-    let widest = 0;
-    for (let index = 0; index <= edges.length; index += 1) {
-      const from = index === 0 ? firstOf(windows) : edges[index - 1] + 1;
-      const to = index === edges.length ? lastOf(windows) : edges[index] - 1;
-      const room_ = coverage.unmadeRunFrom(from);
-      if (to >= from && room_ > widest) {
-        // WHERE THE TWO OF THEM MEET, calculated rather than halved. Halving is
-        // the answer only when both encoders are fresh and owe the same, and it
-        // was applied to every case — one already working is partway through a
-        // piece while a new one owes its whole start, so the point where they
-        // finish together lies further on. Nothing back means a second encoder
-        // does not pay for itself here at all: two under this host's measured
-        // contention against one at full speed, a comparison halving never made.
-        const meet = splitAt(from, Math.min(from + room_ - 1, to));
-        if (meet !== null) {
-          widest = room_;
-          widestFrom = meet;
-        }
-      }
-    }
-    if (widestFrom === null) {
-      break;
-    }
-    const at = coverage.firstGapFrom(widestFrom, lastOf(windows));
-    if (at === null || places.includes(at)) {
-      break;
-    }
-    places.push(at);
-    placedHere.push(at);
-  }
-
-  // These are CANDIDATES, not decisions. What is late proposes first, because
-  // that is what a viewer feels; what is merely unmade proposes after it. The
-  // score decides which of them are worth a process, and a paused viewer, who
-  // states no deadline at all, therefore still leaves the file being finished
-  // rather than the machine falling idle.
-  return places.sort((left, right) => left - right);
-}
-
+/**
+ * @file How many encoders there should be on one output, and where each of them
+ * belongs — decided from numbers alone.
+ *
+ * The decision is separated from carrying it out on purpose. Every rule below
+ * was previously a condition somewhere inside an eleven-thousand-line file,
+ * reachable only by starting a real ffmpeg, and each of them was written for
+ * one viewer:
+ *
+ * - a run was placed at the position of whoever asked, and never at the first
+ *   thing missing, so a viewer moving into a stretch already on disk restarted
+ *   an encoder to make it a second time;
+ * - a run had no end at all — neither `-to` nor `-t` appeared anywhere — so it
+ *   ran until something killed it, and two runs on one output could not exist
+ *   without writing over each other;
+ * - nothing stopped a run that had caught up with material somebody else had
+ *   already made.
+ *
+ * The rule this file exists to express, stated by the user 2026-09-04:
+ *
+ * > Viewers are always independent and always reuse what can be reused. The
+ * > number of encoders is however many are needed; how many are needed follows
+ * > from which sets of output parameters are wanted and where the viewers stand
+ * > inside each. Segments produced by ANY encoder are available to ANY viewer,
+ * > and which viewer asked never enters the question.
+ *
+ * So no name of a viewer reaches this file. It is given what is wanted, what
+ * exists, what is being made, and what the machine can afford.
+ *
+ * **A viewer decides the ORDER the map is walked in and, through the budget, how
+ * many processes walk it. Nothing else.** Stated by the user 2026-09-05, and it
+ * is the rule the rest of this file now follows: while a file is being encoded
+ * it is encoded WHOLE, in the order the map dictates. Who wants which segment
+ * decides which gap is closed first, never whether a run may go on living.
+ *
+ * **A run is therefore never stopped for standing outside a viewer's window.**
+ * It used to be, and the two decisions that produced that were in direct
+ * contradiction — measured in the field 2026-09-05 on a viewer watching an
+ * episode:
+ *
+ * 1. this file commanded a start inside the window, at #46;
+ * 2. `planRunInterval` in the session manager moved the start to #78, because
+ *    it counted a suspended run's claim as reaching `head + look-ahead`;
+ * 3. this file then saw a run at #78 against a window of [27, 57], found no
+ *    overlap, and killed it as "nothing it was given is wanted";
+ * 4. neither coverage nor demand had changed, so the same start was commanded
+ *    again — 350-700ms per cycle, dozens of times, no segment ever produced,
+ *    the viewer's picture stopped for 125 seconds.
+ *
+ * Both of those other authorities are gone (roadmap item 76, step 5). What is
+ * left is this file, and the only reasons it stops a run are: nobody is
+ * watching the output at all; the machine affords fewer processes; or there is
+ * nothing left unmade anywhere in the track.
+ *
+ * **A run's end comes from the coverage**, never from a window: it runs until
+ * it meets material somebody else has made or is making, or until the end of
+ * the film.
+ */
+
+/**
+ * One encoder that is running now.
+ *
+ * @typedef {object} LiveRun
+ * @property {string} id
+ * @property {number} from - The first number it was given.
+ * @property {number} to - The last number it was given, inclusive.
+ * @property {number} head - The next number it will produce. Its position.
+ * @property {number} speedX - The run's processing speed against realtime:
+ *   film made between closed pieces over its own working time
+ *   (`EncodeRun.speedX`), with input waits and stops taken out. Zero or less
+ *   means nothing has measured it yet, and then no comparison involving its
+ *   speed can be made.
+ */
+
+/**
+ * What a viewer is waiting for. Which viewer is deliberately absent.
+ *
+ * @typedef {object} WantedSpan
+ * @property {number} from
+ * @property {number} to
+ */
+
+/**
+ * @typedef {{ type: "start", from: number, to: number, because: string }
+ *   | { type: "move", run: object, from: number, to: number, because: string }
+ *   | { type: "stop", run: object, because: string }
+ *   | { type: "keep", run: object, from: number, to: number }} PlanAction
+ */
+
+/**
+ * Decide what to do with the encoders on one output.
+ *
+ * @param {object} params
+ * @param {import("./CoverageMap.js").CoverageMap} params.coverage - What has
+ *   been made and what is being made.
+ * @param {WantedSpan[]} params.windows - What viewers are waiting for, one
+ *   window each. Empty means nobody is watching this output.
+ * @param {LiveRun[]} params.runs - The encoders running on it now.
+ * @param {number} params.maxRuns - How many encoders this machine can afford on
+ *   this output. Comes from the same arithmetic that decides the quality offer;
+ *   it is measured per host and never chosen here.
+ * @param {number} params.segmentSeconds - How much film one segment holds.
+ * @param {number} [params.killCostSec] - How long stopping an encoder takes,
+ *   measured on this host from its own runs. Zero until something has measured
+ *   it, which makes moving one look cheaper than it is and is said here so the
+ *   bias is known.
+ * @param {number} [params.now] - The clock, injected. This layer is arithmetic
+ *   and reads no clock of its own; how old a run is is one of its inputs.
+ * @param {number} [params.firstByteWaitSec] - How long a fresh encoder takes to
+ *   produce anything: process start, opening the input, and the first piece.
+ *   Measured the same way. It replaced a constant of 0.12 s taken from one
+ *   host and charged to every other.
+ * @param {(others: number) => number} [params.contentionPenaltyFor] - How much
+ *   slower ONE encoder runs with that many others beside it, measured on this
+ *   host. Without it every extra process looks free, and the score then wants an
+ *   encoder per piece: at exactly realtime each next piece is marginally late
+ *   however many are running, so another one always seemed to help a little.
+ *   Unmeasured is 1, and then the budget is the only thing bounding the count.
+ * @returns {PlanAction[]} Stops first, then moves, then starts, so that a plan
+ *   carried out in order never holds two encoders where it means to hold one.
+ */
+export function planEncoders({
+  coverage,
+  windows,
+  runs,
+  maxRuns,
+  segmentSeconds,
+  killCostSec = 0,
+  firstByteWaitSec = 0,
+  now = Date.now(),
+  refetchSecPerFilmSecond = 0,
+  contentionPenaltyFor = () => 1,
+  speedX = 0
+}) {
+  /** @type {PlanAction[]} */
+  const stops = [];
+  /** @type {PlanAction[]} */
+  const moves = [];
+  /** @type {PlanAction[]} */
+  const starts = [];
+  /** @type {PlanAction[]} */
+  const keeps = [];
+
+  const wanted = Array.isArray(windows) ? windows : [];
+  const live = Array.isArray(runs) ? runs : [];
+
+  // Nobody is watching this output: every encoder on it is making segments for
+  // no one. This is the case a look-ahead cannot answer, because look-ahead
+  // asks how far AHEAD of a viewer a run is and there is no viewer.
+  if (wanted.length === 0) {
+    for (const run of live) {
+      stops.push({ type: "stop", run, because: "nobody is watching this output" });
+    }
+    return stops;
+  }
+
+  const untilNeeded = deadlineReaderFor(wanted, segmentSeconds);
+  // Segments produced per second, from the fastest measured encoder here.
+  // Seconds of film per second, divided by the film one piece holds.
+  //
+  // A RUN WORKING ON THIS OUTPUT OUTRANKS THE BENCHMARK. The startup figure is
+  // what this host does on reference clips; a run here is what it does on THIS
+  // material, and that is the more specific statement. Taken as a floor instead
+  // — the larger of the two — an encoder reporting half realtime was scored as
+  // though it ran at twice, and nothing was ever late.
+  const working = live.reduce((best, run) => Math.max(best, run.speedX || 0), 0);
+  const rate = segmentSeconds > 0 ? (working > 0 ? working : speedX) / segmentSeconds : 0;
+  // How long one piece takes AT THE RATE ACTUALLY IN FORCE. Concurrent encoders
+  // slow each other — measured on this host — so an arrangement's own encoder count
+  // decides it, and the delays below are computed per arrangement for that
+  // reason. Taken from the unpenalised rate instead, a piece looked cheaper the
+  // more encoders there were, which made extra encoders look free: the plan bought a
+  // second encoder where one served, and the arrivals it was compared on were
+  // computed at the slower rate all along.
+  //
+  // `Infinity` where nothing has been measured, which is what "no speed" means
+  // and what makes every arrangement equally hopeless rather than equally free.
+  const pieceAt = (howManyEncoders) => {
+    const inForce = rate / contentionPenaltyFor(Math.max(0, howManyEncoders - 1));
+    return inForce > 0 ? 1 / inForce : Number.POSITIVE_INFINITY;
+  };
+  // What a encoder costs to take away from where it stands and put somewhere else:
+  // its death, the start of another, and the wait for the first bytes there.
+  // Taking an encoder somewhere else is stopping this one and waiting for the
+  // next to produce. Both halves are measured on this host.
+  // WHAT A START AND A MOVE OWE BEFORE THE PIECE THEY STAND ON EXISTS, and
+  // neither is infinite.
+  //
+  // The measured `firstByteWaitSec` is spawn to first piece, so it already
+  // contains one piece's encoding. Separated, because the two scale differently:
+  // the piece costs more when encoders share the machine, the spawn does not.
+  //
+  //   spawn overhead = measured first output - what one piece costs alone
+  //   a fresh encoder owes    spawn overhead + the piece at the rate in force
+  //   a moved one owes        the kill, and then the same
+  //
+  // With nothing measured the overhead is zero and a fresh encoder owes exactly
+  // one piece — which is the honest floor rather than a guess: a piece cannot
+  // appear before it is encoded, and how fast this host encodes is measured
+  // before any viewer exists. There was an `Infinity` here for the cost of a
+  // move, on the reasoning that an unmeasured price must not license an
+  // irreversible act; it was an exception in a model that needs none, and this
+  // is the same statement made by arithmetic.
+  const spawnOverheadSec = Math.max(0, firstByteWaitSec - (rate > 0 ? 1 / rate : 0));
+  const moveSec = killCostSec + spawnOverheadSec;
+
+  // WHAT A RUN STILL HAS TO GO BEFORE IT PRODUCES ANYTHING — the measured time
+  // to a first piece, less the time it has already been alive.
+  //
+  // This is the memory the score was missing. It is computed afresh whenever
+  // anything changes, and every arrangement used to be priced as though it were
+  // the last decision anybody would take: a run that started 10 ms ago was
+  // assumed to produce instantly, so killing it and starting another looked like
+  // a straight gain. A move is justified by a benefit that arrives when the
+  // moved run produces something; taken again before it has, the benefit is
+  // never collected and the cost is paid twice, three times, forty times.
+  //
+  // A run 0.8 s old has 0.14 s left to go against 0.94 s to move it, so it is
+  // left alone; one working for half a minute has nothing left, and a move
+  // happens exactly when the film it would reach sooner is worth the restart. No
+  // state to keep and nothing to choose: one measured figure minus elapsed time.
+  const finishesItsPieceIn = (run, perPiece) => {
+    // WHEN THIS RUN FINISHES THE PIECE IT IS STANDING ON.
+    //
+    // Two cases and neither is a choice: a run that has produced something is
+    // working at a known rate, so the piece under it lands within one piece's
+    // worth of time; a run that has produced nothing is still in its warm-up,
+    // and what is left of that is the measured time to a first piece less the
+    // time it has already been alive.
+    //
+    // This is the memory the score was missing. It is computed afresh whenever
+    // anything changes, and each arrangement used to be priced as though it were
+    // the last decision anybody would take: a run 0.8 s into a 0.9 s piece was
+    // charged a whole piece for it, the same as one about to start, so killing
+    // it and beginning again elsewhere looked cheaper by ten milliseconds. A
+    // move is justified by a benefit that arrives when the moved run produces
+    // something; taken again before it has, the benefit is never collected and
+    // the cost is paid twice, three times, forty times.
+    //
+    // One measured figure minus elapsed time. Nothing to keep and nothing to
+    // choose.
+    if (Number(run.head) !== Number(run.from)) {
+      return perPiece;
+    }
+    const startedAt = Number(run.startedAt);
+    if (!Number.isFinite(startedAt) || startedAt <= 0) {
+      return firstByteWaitSec;
+    }
+    return Math.max(0, firstByteWaitSec - (now - startedAt) / 1000);
+  };
+
+  // WHAT EACH BODY OWES BEFORE THE PIECE IT STANDS ON EXISTS, priced at the rate
+  // the arrangement itself puts in force.
+  //
+  // Each encoder states its own debt where it is created, as a function of what one
+  // piece costs — because only there is it known what the encoder IS, and here only
+  // how many of them there are. So there is nothing to dispatch on: no kind, no
+  // tag, no case analysis. The count is known before any debt is needed, which
+  // is why this is one pass over the encoders rather than a figure computed once
+  // outside.
+  const priced = (encoders) => {
+    const perPiece = pieceAt(encoders.length);
+    return encoders.map((encoder) => ({ at: encoder.at, delaySec: encoder.owes(perPiece) }));
+  };
+
+
+  // WHERE A SECOND ENCODER MEETS THE ONE ALREADY ON A STRETCH, and whether it is
+  // worth having.
+  //
+  // A stretch of unmade film runs from `from` to `to`. Whoever is already on it
+  // stands at `from` and owes `w` before the piece under it exists; a fresh one
+  // placed at `x` owes `d` — its start, and then a whole piece — and both then
+  // work at the rate two encoders leave each other, which is measured on this
+  // host and is not half by assumption.
+  //
+  //   the one already there closes [from, x-1]:  w + (x - 1 - from) / r
+  //   the fresh one closes         [x, to]:      d + (to - x) / r
+  //
+  // The first rises with `x` and the second falls, so the stretch is closed
+  // soonest where they cross:
+  //
+  //   x* = (from + to + 1) / 2  +  (d - w) * r / 2
+  //
+  // The midpoint, shifted forward by half the difference of what the two owe,
+  // expressed in pieces. Halving is the special case `d = w`, which holds when
+  // both are fresh — and it was applied to every case, including the common one
+  // where `d` is a start plus a piece and `w` is the tail of a piece already
+  // being made.
+  //
+  // AND WHETHER TO SPLIT AT ALL, which halving never asked. Two encoders under
+  // contention against one at full speed:
+  //
+  //   one:  w + (to - from - 1) / rate
+  //   two:  w + (x* - 1 - from) / r
+  //
+  // Nothing back where the second does not win. On the addon host at 1920x1080
+  // the measured penalty for a second encoder is 1.98 — it takes very nearly
+  // all of the first's speed — so there a second almost never pays, and it was
+  // being placed regardless.
+  const splitAt = (from, to) => {
+    const pieces = to - from + 1;
+    if (!(pieces > 1) || !(rate > 0)) {
+      return null;
+    }
+    const together = rate / contentionPenaltyFor(live.length);
+    if (!(together > 0)) {
+      return null;
+    }
+    const perPieceTogether = 1 / together;
+    // What the one already on this stretch owes, or a fresh one's debt where
+    // nobody is on it — and then both sides owe the same and the meeting point
+    // is the middle, as it should be.
+    const onIt = live.find((run) => Number(run.head) === from) ?? null;
+    const owedThere = onIt
+      ? finishesItsPieceIn(onIt, perPieceTogether)
+      : firstByteWaitSec + perPieceTogether;
+    const owedFresh = firstByteWaitSec + perPieceTogether;
+    const meeting = (from + to + 1) / 2 + ((owedFresh - owedThere) * together) / 2;
+    const x = Math.min(to, Math.max(from + 1, Math.round(meeting)));
+    const withTwo = owedThere + (x - 1 - from) * perPieceTogether;
+    const withOne = owedThere + (pieces - 1) / rate;
+    return withTwo < withOne ? x : null;
+  };
+
+  // ------------------------------------------------------------------ WHERE
+  //
+  // A question about the FILM, and about nothing else: which numbers are
+  // missing, when each is needed, how fast this machine encodes, how many
+  // processes it can hold. No encoder that happens to be running enters it,
+  // which is why it can be answered by arithmetic.
+  const positions = placeEncoders({
+    coverage,
+    windows: wanted,
+    howMany: maxRuns,
+    // EVERY LIVE ENCODER IS PRE-PLACED, because that is what "somebody already
+    // gets here in time" means. A number one of them reaches before it is
+    // needed is not a position at all; a number none of them reaches is, and
+    // needs a encoder brought to it. There is no third case, and in particular no
+    // separate question of whether an encoder should drive on or be moved:
+    // driving is simply its arrival, and its arrival is priced in one place.
+    firstGap: gapFinderFor(coverage, new Set(live), rate, segmentSeconds * refetchSecPerFilmSecond),
+    deadlineAt: untilNeeded,
+    splitAt
+  });
+
+  // -------------------------------------------------------------------- WHO
+  //
+  // ARGMIN OF THE OBJECTIVE, EVALUATED. Not a rule that approximates it.
+  //
+  // Every way of filling the positions is scored by `latenessOf` and the best is
+  // taken. There are at most a handful of positions and a handful of encoders, so
+  // the enumeration is exact: no local rule stands in for the objective, and
+  // none can therefore disagree with another.
+  //
+  // Four such rules were written before this and all four had to go — "place
+  // where a number is late", "take a encoder that serves nothing", "take one whose
+  // work is needed later than this", "drive on or move, by cost". Each looked
+  // like a consequence of the model and each approximated it from a different
+  // side, so together they contradicted one another and the answer depended on
+  // which ran first.
+  /** One decision per live encoder, so none can be decided twice. @type {Map<object, PlanAction>} */
+  const decided = new Map();
+  const room = Math.max(0, maxRuns - live.length);
+  const refetchPerSegment = segmentSeconds * refetchSecPerFilmSecond;
+
+  let arrangements = [{ fill: [], used: new Set(), fresh: 0 }];
+  for (let index = 0; index < positions.length; index += 1) {
+    const next = [];
+    for (const arrangement of arrangements) {
+      next.push({ fill: [...arrangement.fill, null], used: arrangement.used, fresh: arrangement.fresh });
+      // A FRESH PROCESS IS OFFERED BEFORE ANY WORKING BODY, so that when the two
+      // score the same the working one is left alone. Taking it is free in the
+      // arithmetic — its output stays on disk — but it is not free in fact: the
+      // run it belongs to has a position, a warm input and a measured speed, and
+      // all three are thrown away for nothing.
+      if (arrangement.fresh < room) {
+        next.push({
+          fill: [...arrangement.fill, "new"],
+          used: arrangement.used,
+          fresh: arrangement.fresh + 1
+        });
+      }
+      for (const run of live) {
+        if (arrangement.used.has(run)) {
+          continue;
+        }
+        next.push({
+          fill: [...arrangement.fill, run],
+          used: new Set([...arrangement.used, run]),
+          fresh: arrangement.fresh
+        });
+      }
+    }
+    arrangements = next;
+  }
+
+  let best = null;
+  let bestScore = null;
+  for (const arrangement of arrangements) {
+    const encoders = [];
+    for (let index = 0; index < positions.length; index += 1) {
+      const filler = arrangement.fill[index];
+      if (filler === null) {
+        continue;
+      }
+      if (filler === "new") {
+        // A encoder that does not exist yet owes its own start and then the piece.
+        encoders.push({ at: positions[index], owes: (piece) => spawnOverheadSec + piece });
+        continue;
+      }
+      const head = Number(filler.head);
+      // Left where it stands it owes what is left of the piece under it; taken
+      // somewhere else it owes the killing, the start and a whole piece.
+      encoders.push({
+        at: positions[index],
+        owes: head === positions[index]
+          ? (piece) => finishesItsPieceIn(filler, piece)
+          : (piece) => moveSec + piece
+      });
+    }
+    // Bodies nobody was given a position for go on working where they stand,
+    // and their coverage counts: the file is encoded whole.
+    //
+    // A encoder given no end pays a restart the moment anybody is placed inside the
+    // road it would drive: where a run stops is fixed when its process starts,
+    // so it has to be cut and begun again at its own head. That price was
+    // invisible here, and an arrangement was scored as free when it was not.
+    for (const run of live) {
+      if (arrangement.used.has(run)) {
+        continue;
+      }
+      const head = Number(run.head);
+      const endless = Number(run.to) < Number(run.from);
+      const cutInFront = arrangement.fill.some((filler, index) =>
+        filler !== null && positions[index] > head
+        && (endless || positions[index] <= Number(run.to)));
+      encoders.push({
+        at: head,
+        owes: cutInFront
+          ? (piece) => moveSec + piece
+          : (piece) => finishesItsPieceIn(run, piece)
+      });
+    }
+    const scored = latenessOf(priced(encoders), coverage, wanted, untilNeeded, rate / contentionPenaltyFor(Math.max(0, encoders.length - 1)), refetchPerSegment, segmentSeconds);
+    if (bestScore === null || cheaperThan(scored, bestScore)) {
+      bestScore = scored;
+      best = arrangement;
+    }
+  }
+
+  // A BODY STANDING ON FILM THAT EXISTS is the one arrangement the enumeration
+  // above cannot reach: the gap in front of it is nobody's deadline, so it is
+  // never a position, and the encoder is left to make three hundred pieces a second
+  // time. Each such encoder is offered its own first gap and the SAME score decides
+  // — moving costs a restart on everything downstream, staying costs the repeat.
+  //
+  // Offered one at a time rather than folded into the enumeration because the
+  // enumeration is exponential in the number of positions, and this is called
+  // again on every piece produced. One extra evaluation per encoder against
+  // several thousand arrangements is the difference between arithmetic and a
+  // stalled proxy.
+  const placement = new Map();
+  for (let index = 0; index < positions.length; index += 1) {
+    const filler = best ? best.fill[index] : null;
+    if (filler && filler !== "new") {
+      placement.set(filler, positions[index]);
+    }
+  }
+  const encodersOf = (override) => {
+    const encoders = [];
+    for (const run of live) {
+      if (override.has(run) && override.get(run) === null) {
+        // Asked what the film looks like WITHOUT this one.
+        continue;
+      }
+      const at = override.has(run) ? override.get(run) : (placement.get(run) ?? Number(run.head));
+      const head = Number(run.head);
+      encoders.push({
+        at,
+        owes: at === head
+          ? (piece) => finishesItsPieceIn(run, piece)
+          : (piece) => moveSec + piece
+      });
+    }
+    for (let index = 0; index < positions.length; index += 1) {
+      if ((best ? best.fill[index] : null) === "new") {
+        encoders.push({ at: positions[index], owes: (piece) => spawnOverheadSec + piece });
+      }
+    }
+    return priced(encoders);
+  };
+  const scoreOf = (override) => {
+    const encoders = encodersOf(override);
+    return latenessOf(encoders, coverage, wanted, untilNeeded,
+      rate / contentionPenaltyFor(Math.max(0, encoders.length - 1)), refetchPerSegment, segmentSeconds);
+  };
+  for (const run of live) {
+    if (placement.has(run)) {
+      continue;
+    }
+    const gap = coverage.firstGapFrom(run.head, undefined, run);
+    if (gap === null || gap === Number(run.head)) {
+      continue;
+    }
+    const asIs = scoreOf(new Map());
+    const moved = scoreOf(new Map([[run, gap]]));
+    if (cheaperThan(moved, asIs)) {
+      placement.set(run, gap);
+    }
+  }
+
+  /**
+   * Would the film be worse off without this encoder? Asked of the same score.
+   *
+   * @param {object} run
+   * @returns {boolean}
+   */
+  const worseWithout = (run) => {
+    const kept = encodersOf(new Map());
+    const without = encodersOf(new Map([[run, null]]));
+    const scoreOf_ = (encoders) => latenessOf(encoders, coverage, wanted, untilNeeded,
+      rate / contentionPenaltyFor(Math.max(0, encoders.length - 1)), refetchPerSegment, segmentSeconds);
+    return cheaperThan(scoreOf_(kept), scoreOf_(without));
+  };
+
+  const stretchAt = (from) => endOfStretch(from, Math.min(
+    coverage.unmadeRunFrom(from),
+    coverage.freeRunFrom(from, new Set(live))
+  ));
+
+  for (let index = 0; index < positions.length; index += 1) {
+    if ((best ? best.fill[index] : null) !== "new") {
+      continue;
+    }
+    const at = positions[index];
+    starts.push({
+      type: "start",
+      from: at,
+      to: stretchAt(at),
+      because: `#${at} is wanted and nobody reaches it in time`
+    });
+  }
+
+  for (const run of live) {
+    const head = Number(run.head);
+    const at = placement.has(run) ? placement.get(run) : head;
+    if (at !== head) {
+      decided.set(run, {
+        type: "move",
+        run,
+        from: at,
+        to: stretchAt(at),
+        because: `standing at #${head} scores worse than standing at #${at}, counting ` +
+          "both how late the film would be and the work that would be done twice"
+      });
+      continue;
+    }
+    // It stays where it is — unless holding it changes nothing.
+    //
+    // A encoder left over from where a viewer used to be goes on costing the
+    // machine a process while another encoder already reaches everything it
+    // would. The score says so directly: take it away and see. Removing it is
+    // refused the moment it makes anything later or leaves film abandoned, so
+    // this cannot quietly drop the encoder somebody is waiting on.
+    if (!placement.has(run) && !worseWithout(run)) {
+      stops.push({
+        type: "stop",
+        run,
+        because: "the film is no worse off without it"
+      });
+      continue;
+    }
+    decided.set(run, { type: "keep", run, from: head, to: run.to });
+  }
+
+  // THE MACHINE'S LIMIT BINDS, whatever the map wants. It is measured — the
+  // processor, the swarm and the piece store each give a figure and the smallest
+  // wins — and an encoder over it is one the host cannot feed. Which of them
+  // goes is the same question as any other here: the one the film misses least,
+  // by the same score.
+  while (decided.size + starts.length > maxRuns) {
+    let cheapest = null;
+    let cheapestScore = null;
+    for (const [run, action] of decided) {
+      if (action.type !== "keep") {
+        continue;
+      }
+      const without = latenessOf(encodersOf(new Map([[run, null]])), coverage, wanted, untilNeeded,
+        rate / contentionPenaltyFor(Math.max(0, live.length - 2)), refetchPerSegment, segmentSeconds);
+      if (cheapestScore === null || cheaperThan(without, cheapestScore)) {
+        cheapestScore = without;
+        cheapest = run;
+      }
+    }
+    if (cheapest === null) {
+      break;
+    }
+    decided.delete(cheapest);
+    placement.delete(cheapest);
+    stops.push({
+      type: "stop",
+      run: cheapest,
+      because: `the machine holds ${maxRuns} encoder(s) on this output and this is the one ` +
+        "the film misses least"
+    });
+  }
+
+  for (const action of decided.values()) {
+    if (action.type === "keep") {
+      keeps.push(action);
+    } else {
+      moves.push(action);
+    }
+  }
+
+  // ONE ENCODER'S WORK ENDS WHERE THE NEXT ONE'S BEGINS.
+  //
+  // A free stretch may run to the end of the track, and an encoder given all of
+  // it stands in the road of every encoder placed behind it: they write the
+  // same names, and each one's output is the other's "material somebody else
+  // made", so they stop one another. Field 2026-09-05: three encoders started
+  // on one track within 200 ms, each into the road another was already writing.
+  // Fifteen readers on a piece store that holds sixteen pieces followed, half
+  // of all evictions took a piece a reader had declared, and `/stream` began
+  // handing out bytes that were not the file's.
+  //
+  // The bound is taken from the NEXT ENCODER'S START, not from a band edge: a
+  // band edge travels with the viewer, so every step forward would leave a
+  // sliver just past the previous encoder and buy an encoder for it.
+  // A RUN THAT IS STAYING IS IN THE SORT TOO, because its road can be taken.
+  //
+  // It used to be left out, on the reading that a run staying put keeps what it
+  // was given. That is true of the stretch it was GIVEN and false of the road it
+  // will actually drive: a run with no end carries no `-to` and walks to the end
+  // of the film, so an encoder placed in front of it writes the same names.
+  const placed = [...moves, ...starts, ...keeps].sort(
+    (left, right) => /** @type {any} */ (left).from - /** @type {any} */ (right).from
+  );
+  for (let index = 0; index < placed.length - 1; index += 1) {
+    const here = /** @type {{ type: string, run?: object, from: number, to: number }} */ (placed[index]);
+    const next = /** @type {{ from: number }} */ (placed[index + 1]);
+    if (here.to >= 0 && here.to < next.from) {
+      continue;
+    }
+    here.to = next.from - 1;
+    if (here.type !== "keep") {
+      continue;
+    }
+    // SHORTENING A LIVE RUN'S ROAD MEANS STOPPING IT, not merely writing a
+    // smaller number down. Where a run's end goes is fixed when its process
+    // starts, so one that was given none keeps producing past any bound decided
+    // later and would write one piece into the new encoder's road — two
+    // processes on one name, which is the collision this whole pass exists to
+    // prevent. So it ends here and begins again at its own head with a real end;
+    // the viewer in front pays a restart, which is a cost this file already
+    // prices rather than a interruption nobody counted.
+    keeps.splice(keeps.indexOf(here), 1);
+    moves.push({
+      type: "move",
+      run: here.run,
+      from: here.from,
+      to: here.to,
+      because:
+        `an encoder is needed at #${next.from}, which this run would reach only by ` +
+        "encoding through; it takes the road up to there and ends by itself"
+    });
+  }
+
+  return [...stops, ...moves, ...starts, ...keeps];
+}
+
+
+
+/**
+ * THE OBJECTIVE, as a value that can be compared.
+ *
+ * THREE COUNTS OF SECONDS, COMPARED IN ORDER. The order is the user's, stated
+ * 2026-09-06, and a later count decides only where the earlier ones tie:
+ *
+ * 1. SECONDS ANYBODY SPENDS LOOKING AT A SPINNER. Nothing outranks it, at any
+ *    size. Walked forward in film order rather than summed piece by piece: a
+ *    viewer who is stopped is not watching, so a wait moves every deadline
+ *    behind it by its own length;
+ *
+ * 2. WHEN THE FILM IN FRONT OF THE VIEWERS IS FINISHED — the last piece of it to
+ *    be made, whichever encoder makes it. A stretch no encoder will ever reach
+ *    counts as never, which is what stops the front being abandoned;
+ *
+ * 3. WHEN THE WHOLE FILE IS FINISHED — the film behind the viewers included,
+ *    plus what the swarm pays to fetch anything a second time. Film nobody is
+ *    waiting for still has value: a viewer seeking back into a part that exists
+ *    starts playing at once, and seeking back is what people do in the first
+ *    minutes while they find their place. So spare capacity goes to finishing
+ *    the file. This is where "the file is encoded WHOLE" lives; it used to be a
+ *    penalty for film below the lowest encoder, which said the same thing as a
+ *    patch and said it about one edge of the track only.
+ *
+ * WHY AN ENCODER MAY STAND BEHIND A VIEWER while film in front is still unmade:
+ * encoders work at the same time, so one in front and one behind can finish the
+ * file sooner than two in front. Where nobody is stalled and the front is closed
+ * just as fast, the file being done sooner is the answer — count 3 deciding a
+ * tie in 1 and 2, which is exactly what the order is for.
+ *
+ * WHY THE COUNTS ARE COMPARED AND NOT ADDED: seconds of somebody waiting and
+ * seconds until a distant stretch exists are not the same thing, and no measured
+ * quantity says how many of one are worth one of the other. Adding them would
+ * mean choosing that exchange rate, which is inventing a number.
+ *
+ * @param {{ at: number, delaySec: number }[]} encoders - Where each encoder would
+ *   stand, and how long before it produces anything there: nothing where it is
+ *   already standing, a move or a start otherwise.
+ * @param {import("./CoverageMap.js").CoverageMap} coverage
+ * @param {WantedSpan[]} wanted
+ * @param {(index: number) => number} untilNeeded
+ * @param {number} rate - Segments per second. Always a real figure: this host
+ *   measures what it encodes at on startup, before any viewer exists, and every
+ *   run that works then refines it. There is no "unmeasured" case to answer.
+ * @param {number} refetchSecPerSegment
+ * @param {number} segmentSeconds
+ * @returns {{ stall: number, ahead: number, whole: number }} Three counts of
+ *   seconds, compared in that order by {@link cheaperThan}.
+ */
+function latenessOf(encoders, coverage, wanted, untilNeeded, rate, refetchSecPerSegment, segmentSeconds) {
+  const first = Math.min(...wanted.map((span) => span.from));
+  const last = Math.max(...wanted.map((span) => span.to));
+  // What a number nobody reaches at all counts as. The film's own length is the
+  // honest bound — nothing can be later than never — and a finite figure is what
+  // lets two hopeless arrangements still be told apart by the rest of the sum.
+  const never = (last + 1) * segmentSeconds;
+
+  // WHICH SIDE OF THE VIEWERS a piece is on. The map states it; nothing here
+  // works it out from positions, and nothing here knows where a viewer stands.
+  //
+  // It was read off the deadline before — no time stated meant behind — and that
+  // is true only of a viewer who is playing. A paused viewer has no times
+  // anywhere, so their whole film read as behind them, "ahead before behind"
+  // had nothing to compare, and the encoder was free to wander to the start of
+  // the file. Which side a stretch is on and how soon it is wanted are two
+  // different facts, and the map states both.
+  // WHAT RANK THE MAP GIVES THIS NUMBER — the highest, where zones overlap,
+  // because a number two viewers want is wanted as much as the more urgent of
+  // them wants it.
+  //
+  // This replaced a boolean, "is it behind everybody", and the boolean was the
+  // whole of what the objective knew about the map's own order. The map states
+  // ten ranks on a film — p100 at the number a viewer is stopped on, doubling
+  // zones down to p91 for the far tail, p1 for what lies behind them — and all
+  // of that was collapsed into two buckets and then converted to seconds, where
+  // "never" for the film behind is the film's own length. On a 48-minute file
+  // that is 2024 s, which outvotes everything: field 2026-09-08, one viewer got
+  // three encoders, two of them on film behind them, and the run serving them
+  // was killed to make room for one.
+  const rankAt = (at) => {
+    let rank = 0;
+    for (const span of wanted) {
+      if (at >= span.from && at <= span.to) {
+        rank = Math.max(rank, Number(span.priority) || 0);
+      }
+    }
+    return rank;
+  };
+  // The ranks the map actually states, most urgent first. The comparison is over
+  // these and nothing else, so a rank can never be outvoted by a lower one
+  // however many seconds are at stake there.
+  const ranks = [...new Set(wanted.map((span) => Number(span.priority) || 0))]
+    .sort((left, right) => right - left);
+
+  // EVERY COUNT IS OVER THE FILM, NOT OVER THE ENCODERS. When a piece is made
+  // depends on which encoder reaches it soonest, and the encoder that reaches
+  // film in front of the viewers may well be standing behind them.
+  //
+  // Counted over the encoders instead — each charged to the side it stands on —
+  // the score had a hole that swallowed everything: an arrangement with every
+  // encoder BEHIND the viewers had nothing charged to the film in front, so its
+  // second term was zero, which is the best value there is. The plan then
+  // abandoned the film in front of a viewer and put both encoders at the start
+  // of the file, which is the opposite of the rule it is supposed to obey.
+  //
+  // AND THE WAITING IS WALKED FORWARD, not summed piece by piece.
+  //
+  // Not a sum of each piece's own lateness. A viewer who is stopped is not
+  // watching, so everything after the piece they are stopped on is needed that
+  // much later too: one wait moves every deadline behind it by its own length.
+  //
+  // Summed independently instead, the far tail of a long file outvoted the film
+  // under the viewer's feet — measured, and it placed the only encoder at #114
+  // while the viewer stood at #100, because thirteen pieces of certain waiting
+  // "cost" less than 886 distant pieces arriving a little later. Walking the
+  // clock forward makes that trade impossible: abandoning the near film delays
+  // the far film by at least as much.
+  let stalled = 0;
+  let wastedSwarm = 0;
+  /** Seconds anybody waits past a deadline, per rank. @type {Map<number, number>} */
+  const lateAt = new Map(ranks.map((rank) => [rank, 0]));
+  /** When the last number of a rank is made, per rank. @type {Map<number, number>} */
+  const doneAt = new Map(ranks.map((rank) => [rank, 0]));
+  for (let index = first; index <= last; index += 1) {
+    // Which encoder gets to this piece first, and when. One standing on it is
+    // already there; one behind it must work its way up, re-making anything
+    // already made on the way, which costs its own time and the swarm's.
+    let soonest = Number.POSITIVE_INFINITY;
+    let byWhom = null;
+    for (const encoder of encoders) {
+      if (encoder.at > index) {
+        continue;
+      }
+      // WHEN THIS BODY REACHES THIS NUMBER. `delaySec` is when it finishes the
+      // piece it is STANDING ON, so the pieces after it are the only ones still
+      // to be encoded at `rate`.
+      //
+      // It used to be `(index - at + 1) / rate` on top of the delay, which
+      // charges every encoder a whole piece for the one it is already working on. A
+      // fresh encoder does start from nothing, so for it that is right — and it is
+      // now inside its own delay. A run 0.8 s into a 0.9 s piece does not, and
+      // charging it 0.94 s for that piece is what made moving it look cheaper
+      // than leaving it: from #58 it was priced at 1.89 s to reach #59 against
+      // 1.88 s for a kill and a cold start, a difference of ten milliseconds
+      // that is nothing but the double charge. Field 2026-09-08: 39 moves in one
+      // session, 24 of them between three adjacent numbers, and the picture
+      // stood still for 116.7 s.
+      const arrival = encoder.delaySec
+        + (index - encoder.at) / rate
+        + coverage.madeBetween(encoder.at, index) * refetchSecPerSegment;
+      if (arrival < soonest) {
+        soonest = arrival;
+        byWhom = encoder;
+      }
+    }
+    if (coverage.isReady(index)) {
+      // It exists. Nobody waits for it and nothing is owed — but whoever passes
+      // over it makes it a second time, and the swarm fetches its bytes again.
+      if (byWhom !== null) {
+        wastedSwarm += refetchSecPerSegment;
+      }
+      continue;
+    }
+    // A piece nobody is working towards arrives never. There is no third case:
+    // the host measures what it encodes at, and what it copies at, before any
+    // viewer exists, so a speed is always a real number and an arrival can
+    // always be computed.
+    // Nothing arrives later than never, which is the bound the film's own length
+    // gives. It is a definition rather than a guard: it also makes the score
+    // total on a host whose startup measured nothing at all, where every arrival
+    // is beyond reckoning and every arrangement is therefore equally hopeless.
+    const when = byWhom === null ? never : Math.min(soonest, never);
+    const rank = rankAt(index);
+    doneAt.set(rank, Math.max(doneAt.get(rank) ?? 0, when));
+    const deadline = untilNeeded(index);
+    if (Number.isFinite(deadline)) {
+      const due = deadline + stalled;
+      const waited = Math.max(0, when - due);
+      lateAt.set(rank, (lateAt.get(rank) ?? 0) + waited);
+      stalled += waited;
+    }
+  }
+
+  return {
+    // THE MAP'S OWN ORDER, AS A VECTOR. One pair per rank the map states, most
+    // urgent rank first: how long anybody waits at that rank, then when the last
+    // number of it is made.
+    //
+    // Compared position by position, so a rank is never outvoted by a lower one
+    // — which is the whole of what was asked for: nobody stares at a spinner;
+    // then the film in front of the viewers is encoded as fast as it can be, band
+    // by band as the map ranks them; then, with whatever is left over and only
+    // then, the film behind them, in case somebody seeks back.
+    //
+    // No weights, and none possible: a weight would let seconds at one rank buy
+    // seconds at another, and it would be a number nobody measured. The map is
+    // the source of truth about what matters, and it already says so.
+    byRank: ranks.flatMap((rank) => [lateAt.get(rank) ?? 0, doneAt.get(rank) ?? 0]),
+    // HOW MANY ENCODERS IT TAKES. Ranked below every rank of the map and above
+    // the swarm's bill, so it cannot buy one where the map is indifferent — and
+    // the map IS indifferent about spare capacity, which is what bought an
+    // encoder for film nobody waits for.
+    encoders: encoders.length,
+    // WHAT THE SWARM PAYS for anything fetched twice, which delays everything.
+    wasted: wastedSwarm
+  };
+}
+
+/**
+ * Is the first arrangement cheaper than the second?
+ *
+ * Three things in order, stated by the user 2026-09-06: nobody stares at a
+ * spinner; then the film in front of the viewers is finished soonest; then the
+ * whole file is. A later one decides only where the earlier ones tie.
+ *
+ * That order is why an encoder may stand BEHIND a viewer while film in front is
+ * still unmade: encoders work at the same time, so one in front and one behind
+ * can finish the file sooner than two in front — and where nobody is stalled and
+ * the front is closed just as fast, the file being done sooner is the answer.
+ *
+ * @param {{ stall: number, ahead: number, whole: number }} left
+ * @param {{ stall: number, ahead: number, whole: number }} right
+ * @returns {boolean}
+ */
+function cheaperThan(left, right) {
+  // POSITION BY POSITION, in the map's own order of ranks. A difference at a
+  // higher rank settles it, and nothing at a lower one can reopen it.
+  //
+  // There is no margin here and there must not be one. A threshold — "a move
+  // must beat staying by at least what moving costs" — was written while the
+  // arrival arithmetic charged a run for a piece it had already half made, and
+  // it was a prop under a comparison that was wrong rather than indifferent.
+  // With the arithmetic right the two are 1.08 s against 1.88 s, and nothing
+  // needs propping.
+  const size = Math.max(left.byRank.length, right.byRank.length);
+  for (let index = 0; index < size; index += 1) {
+    const here = left.byRank[index] ?? 0;
+    const there = right.byRank[index] ?? 0;
+    if (here !== there) {
+      return here < there;
+    }
+  }
+  // Where every rank is served identically, fewer encoders. The map is
+  // indifferent, so the machine decides: a process, a reader of the piece store
+  // and the swarm's bandwidth are all paid by the viewers the ranks above are
+  // about.
+  if (left.encoders !== right.encoders) {
+    return left.encoders < right.encoders;
+  }
+  return left.wasted < right.wasted;
+}
+
+/**
+ * How long until a number is needed, read off the map.
+ *
+ * The map states it per stretch, for the stretch's NEAR EDGE, because a stretch
+ * is met at its beginning. Every number inside is needed no sooner than that, so
+ * taking the stretch's figure for all of them is the safe reading: it can only
+ * make the filling earlier than it has to be, never later.
+ *
+ * Where the map says nothing, nobody is coming and nothing can be late.
+ *
+ * Inside a stretch the time GROWS with the distance, because a viewer covers a
+ * second of film in a second: the number `n` places past the near edge is
+ * reached `n` segments of film later. Taking the near edge's figure for every
+ * number inside instead makes a whole stretch due at once — measured while
+ * building this: the first stretch is as wide as the measured allowance, so its
+ * far end was demanded instantly and an encoder was placed on a number another
+ * one was already writing.
+ *
+ * @param {WantedSpan[]} windows
+ * @param {number} segmentSeconds - How much film one number holds.
+ * @returns {(index: number) => number}
+ */
+/** @param {WantedSpan[]} windows */
+function firstOf(windows) {
+  return Math.min(...windows.map((span) => span.from));
+}
+
+/** @param {WantedSpan[]} windows */
+function lastOf(windows) {
+  return Math.max(...windows.map((span) => span.to));
+}
+
+function deadlineReaderFor(windows, segmentSeconds) {
+  const perSegment = segmentSeconds > 0 ? segmentSeconds : 0;
+  return (index) => {
+    let soonest = Number.POSITIVE_INFINITY;
+    for (const span of windows) {
+      if (index < span.from || index > span.to) {
+        continue;
+      }
+      const stated = /** @type {{ withinSeconds?: number }} */ (span).withinSeconds;
+      // A stretch stated with no time is somebody waiting at its near edge: that
+      // is what stating one means. The rest of it grows with the distance, the
+      // same as a stated one — read as due all at once instead, a window as wide
+      // as a viewer's cushion demanded its far end instantly and bought an
+      // encoder to stand beside one already working.
+      // `null` IS A STATEMENT AND IT SAYS NOBODY IS COMING. `undefined` is the
+      // absence of one, and a caller that knows only a position is somebody
+      // waiting at it.
+      //
+      // Read through `Number()`, `null` becomes 0 — due NOW — so the film BEHIND
+      // the viewers, which the map marks with exactly that, was the most urgent
+      // material in the file. Everything followed from it: it bought encoders,
+      // it took the run standing in front of the viewer because that run was the
+      // nearest encoder to it, and it did so again on every pass. Field 2026-09-08:
+      // 39 moves in one session, 24 between three adjacent numbers, one viewer
+      // on three encoders, and the picture stood still for 116.7 s in three
+      // interruptions, the worst of them 91.8 s.
+      //
+      // The map has always said it plainly — `{"from":0,"to":57,"priority":1,
+      // "withinSeconds":null,"behind":true}` is in the log of every session — and
+      // this line turned it into its opposite. Fourth time in this repository
+      // that the input to a calculation was not what the calculation assumed.
+      const within = stated === undefined ? 0 : (stated === null ? Number.NaN : Number(stated));
+      if (!Number.isFinite(within)) {
+        // Stated as no time at all: nobody is coming here.
+        continue;
+      }
+      const here = within + (index - span.from) * perSegment;
+      if (here < soonest) {
+        soonest = here;
+      }
+    }
+    return soonest;
+  };
+}
+
+/**
+ * WHERE ENCODERS BELONG, from the model rather than from a list of cases.
+ *
+ * The problem this solves, stated exactly:
+ *
+ * - the track is a line of segment numbers; `M` are the ones not made;
+ * - each `x` carries a DEADLINE `D(x)`, the seconds until somebody needs it.
+ *   That is what the priority map is a reading of — a viewer moving forward
+ *   covers a second of film in a second, so the time until they are at `x` is
+ *   the distance to it. `Infinity` where nobody is coming;
+ * - an encoder is a SEQUENTIAL producer: placed at `a`, it delivers `a + j` at
+ *   time `(j + 1) / r`, where `r` is segments per second, measured. It cannot
+ *   skip, so its whole schedule follows from where it starts;
+ * - the machine affords `k` of them, measured.
+ *
+ * Two consequences fall out and need no rule of their own. Placements
+ * `a_1 < ... < a_k` PARTITION the line: encoder `i` is useful only on
+ * `[a_i, a_{i+1})`, because past that its neighbour got there first. And a
+ * segment served by encoder `i` arrives at `(x - a_i + 1) / r`, which is
+ * therefore also the answer to "when would the encoder already placed before it
+ * get here" — the second half of the comparison, and the half that was missing.
+ *
+ * `x` is LATE when it arrives after `D(x)`. The objective is no late segments;
+ * where `k` does not stretch to that, lateness beginning as far to the right as
+ * possible.
+ *
+ * THE ALGORITHM is first-fit, left to right:
+ *
+ *     for each missing x with a finite deadline, ascending:
+ *         if some encoder already placed at a satisfies (x - a + 1)/r <= D(x):
+ *             it covers x
+ *         else:
+ *             place an encoder at x
+ *
+ * A LIVE run enters as an encoder already placed at its own head. There is no
+ * special case for it.
+ *
+ * WHY IT IS OPTIMAL. The leftmost missing number with a finite deadline must be
+ * covered by somebody. An encoder placed exactly on it delivers it at the
+ * earliest time any placement can, `1/r`, and covers the longest suffix any
+ * placement can — starting further left only re-makes material and arrives
+ * later, starting further right does not cover it at all. So the greedy choice
+ * is never worse than any other, and the usual exchange argument carries it to
+ * the whole line. This is the known result for FIXED-ORDER scheduling with
+ * deadlines, where first-fit is optimal at unit processing times, and a segment
+ * is one unit. General machine minimisation with release times and deadlines is
+ * NP-hard; this case is polynomial because the order is forced and each machine
+ * covers a contiguous stretch.
+ *
+ * WHAT WAS TRIED FIRST AND WAS WRONG, kept because each looked reasonable:
+ *
+ * - `(h - p) * s / (1 - s)`, how long a run stays in front of a viewer moving
+ *   forward. It answers a different question: a viewer stopped with an empty
+ *   buffer needs the segment now, and at exactly realtime that formula says
+ *   "for ever" while the viewer waits thirteen minutes;
+ * - whether a run's head lies inside a wanted band — which ties an encoder to
+ *   whoever is standing there, and this layer must never know that;
+ * - the run's head as a barrier, everything above it placeable. It has no time
+ *   in it at all, so it cannot tell two segments ahead from two hundred.
+ *
+ * Each was a case, not a model. The deadline is the model.
+ *
+ * @param {import("./CoverageMap.js").CoverageMap} coverage
+ * @param {Set<object>} surviving - Runs that will still be alive, as encoders
+ *   already placed at their own heads.
+ * @param {number} rate - Segments produced per second by one encoder, measured.
+ *   Zero when nothing has measured it, and then no arrival time can be computed
+ *   and every claimed number is left alone.
+ * @returns {(at: number, bound: number, deadlineAt: (index: number) => number, alsoPlaced?: number[]) => number | null}
+ */
+function gapFinderFor(coverage, surviving, rate, refetchSecPerSegment = 0) {
+  /** Encoders already placed: where each stands, and how far its road runs. */
+  const placed = [];
+  for (const run of surviving) {
+    const head = Number(/** @type {{ head?: number }} */ (run).head);
+    placed.push({
+      at: Number.isFinite(head) ? head : Number(/** @type {{ from: number }} */ (run).from),
+      // A live run's road, so that placing inside it can be priced. A run given
+      // no end drives to the end of the film, which is what makes the price real.
+      // A run given no end drives to the end of the film, which is what makes
+      // the price of cutting in front of it real. Written out rather than
+      // imported: this file depends on nothing, and that is what lets it be
+      // exercised with plain values alone.
+      to: Number(/** @type {{ to: number }} */ (run).to) < Number(/** @type {{ from: number }} */ (run).from)
+        ? Number.POSITIVE_INFINITY
+        : Number(/** @type {{ to: number }} */ (run).to)
+    });
+  }
+  return (at, bound, deadlineAt, alsoPlaced) => {
+    const start = Number.isInteger(at) && at > 0 ? at : 0;
+    const last = Number.isInteger(bound) ? bound : -1;
+    // THE LATE NUMBER THAT IS DUE SOONEST, not the leftmost one.
+    //
+    // With room for every placement the two are the same answer. With a budget
+    // that binds they are not, and the objective decides: lateness pushed as far
+    // to the right as possible means the soonest deadline is served first. A
+    // walk by number gave the one machine to a viewer due in ten minutes while
+    // another stood waiting with an empty buffer.
+    //
+    // Ties go to the smaller number, so the answer does not depend on the order
+    // the map happens to be in.
+    let best = null;
+    let bestDue = Number.POSITIVE_INFINITY;
+    for (let index = start; index <= last; index += 1) {
+      if (coverage.isReady(index)) {
+        continue;
+      }
+      const deadline = deadlineAt(index);
+      if (!Number.isFinite(deadline)) {
+        // NOBODY IS COMING HERE, so nothing can be late — but the film is still
+        // wanted, and this is where spare capacity goes. The number is proposed;
+        // whether an encoder is actually spent on it is the score's answer, and
+        // the score puts anything anybody is waiting for first.
+        return index;
+      }
+      // When would the SOONEST of those already placed get here? Encoders placed
+      // EARLIER IN THIS PASS count: the first one placed for a viewer covers the
+      // stretch in front of them, and without counting it the walk placed a
+      // second and a third on the very next numbers — three processes a segment
+      // apart for one person, which is the waste this model exists to refuse.
+      let soonest = Number.POSITIVE_INFINITY;
+      for (const a of [...placed.map((live) => live.at), ...(alsoPlaced ?? [])]) {
+        if (a > index) {
+          // Standing past it. Encoders only move forward, so it never will.
+          continue;
+        }
+        if (a === index) {
+          // Standing ON it. No placement is faster than the one already made.
+          soonest = 0;
+          break;
+        }
+        // WHEN THIS BODY GETS HERE, and both terms of it.
+        //
+        // Its own encoding of everything between, and the swarm's price for the
+        // film it would fetch a SECOND time — every number between that is
+        // already made, it makes again. That second term is why "should this
+        // encoder drive on or be moved" is not a question of its own: an
+        // encoder with three hundred made pieces in front of it is simply slow
+        // to arrive, and the model compares arrivals. Asked separately it was a
+        // second authority over the same encoder, and the two disagreed.
+        const arrival = (index - a + 1) / rate
+          + coverage.madeBetween(a, index) * refetchSecPerSegment;
+        if (arrival < soonest) {
+          soonest = arrival;
+        }
+      }
+      if (soonest <= deadline) {
+        // Somebody gets here in time. Nothing to decide.
+        continue;
+      }
+      // IT IS LATE, AND THAT IS ALL THIS DECIDES. Whether filling it is worth
+      // the price is not asked here: this only proposes candidates, and the
+      // score decides how many of them are taken and by whom. Asked here as
+      // well, it was a second cost model beside the objective — with its own
+      // idea of what a process costs — and the two disagreed at exactly
+      // realtime, where every next piece is marginally late and each looked
+      // worth its own encoder.
+      if (deadline < bestDue) {
+        best = index;
+        bestDue = deadline;
+      }
+    }
+    return best;
+  };
+}
+
+/**
+ * The last number of a stretch that begins at `from` and is `length` long.
+ *
+ * `-1` when the length is not finite, which is this layer's word for a run with
+ * no end: the film's length is not known, so there is nothing to stop it at, and
+ * a number invented here would be an end nobody measured.
+ *
+ * @param {number} from
+ * @param {number} length
+ * @returns {number}
+ */
+function endOfStretch(from, length) {
+  return Number.isFinite(length) ? from + Math.max(1, length) - 1 : -1;
+}
+
+/**
+ * The lowest number a viewer is waiting for that is not ready — what the plan
+ * is judged by.
+ *
+ * Not used to decide anything: it is the figure a log line carries, so that a
+ * plan that keeps producing while a viewer waits is visible rather than
+ * inferred.
+ *
+ * @param {import("./CoverageMap.js").CoverageMap} coverage
+ * @param {WantedSpan[]} windows
+ * @returns {number | null}
+ */
+export function firstUnmetWant(coverage, windows) {
+  let lowest = null;
+  for (const span of windows ?? []) {
+    for (let at = span.from; at <= span.to; at += 1) {
+      if (!coverage.isReady(at)) {
+        if (lowest === null || at < lowest) {
+          lowest = at;
+        }
+        break;
+      }
+    }
+  }
+  return lowest;
+}
+
+/**
+ * Where to put the encoders this machine can afford.
+ *
+ * Two things are wanted of a division of the film, and they are wanted in this
+ * order:
+ *
+ * 1. **the viewer must not stop.** An encoder starting at `q` stays ahead of a
+ *    viewer at `p` while `y / s <= q + y - p`, so it holds `(q - p) * s / (1-s)`
+ *    of film and no more. Beyond that the viewer catches it, and the next
+ *    encoder has to be standing there. That is where the first ones go, and it
+ *    is why the stretches grow: the further off one starts, the later the
+ *    viewer arrives and the longer it may work;
+ * 2. **the film should be finished as soon as possible.** Once the viewer is
+ *    safe, whatever is left is divided EQUALLY between the encoders that
+ *    remain: equal shares finish together, and any other division finishes when
+ *    its longest share does. That is what makes seeking cheap — the film exists.
+ *
+ * At or above realtime the first requirement is met by one encoder for the
+ * whole film, and every other encoder goes to the second — which is the common
+ * case on a copied picture, and is why "one viewer, one encoder" was never the
+ * rule.
+ *
+ * @param {object} params
+ * @param {import("./CoverageMap.js").CoverageMap} params.coverage
+ * @param {WantedSpan[]} params.windows - The merged map, in this output's own
+ *   numbering. Its highest-numbered band starts where the viewer is.
+ * @param {number} params.howMany - What the machine affords.
+ * @param {number} params.speedX - Measured. Zero when nothing has measured it,
+ *   and then only the first requirement can be served.
+ * @param {(at: number, bound: number, deadlineAt: (index: number) => number, placed: number[]) => number | null} [params.firstGap] -
+ *   Where a gap may be opened. Defaults to the map's own answer; the plan hands
+ *   in one that also counts a number a live run has claimed but will not reach
+ *   before it is needed, which is the only way anybody beyond a working encoder
+ *   is served.
+ * @param {(index: number) => number} [params.deadlineAt] - Seconds until that
+ *   number is needed. `Infinity` where nobody is coming. Absent means every
+ *   stated want is due now.
+ * @returns {number[]} Where to start each encoder, ascending.
+ */
+export function placeEncoders({
+  coverage,
+  windows,
+  howMany,
+  firstGap = null,
+  deadlineAt = null,
+  // WHERE TWO ENCODERS SHARE A STRETCH, and whether a second is worth having at
+  // all. Derived by whoever holds the measurements — the rates, the contention
+  // penalty, what a start costs — because this function is positional and holds
+  // none of them. The default is the midpoint, which IS the answer when both are
+  // fresh and owe the same, and is what a caller with nothing measured falls
+  // back on.
+  splitAt = (from, to) => from + Math.floor((to - from + 1) / 2)
+}) {
+  if (!(howMany > 0) || windows.length === 0) {
+    return [];
+  }
+  // NOW when the caller says nothing. A stated want with no time is somebody
+  // waiting on it — that is what stating one means — so the honest reading is
+  // that it is due. `Infinity` is a statement in its own right and has to be
+  // made: it says nobody is coming.
+  const untilNeeded = deadlineAt ?? (() => 0);
+  /** Where this pass has placed so far — each one covers what it can reach. */
+  const placedHere = [];
+  const gapAt = firstGap
+    ? (at, bound) => firstGap(at, bound, untilNeeded, placedHere)
+    : (at, bound) => coverage.firstGapFrom(at, bound);
+
+
+  // CANDIDATES COME FROM THE PRIORITY MAP, IN THE ORDER THE MAP STATES.
+  //
+  // The map already answers every question that was being re-derived here. Its
+  // ranks say what matters most — the number a viewer is stopped on, then what
+  // is in front of them band by band, then the rest of the track, and last of
+  // all what lies behind them. A pause flattens those ranks; a seek moves them;
+  // a second viewer merges into them. So walking the map in its own order is
+  // what "ahead before behind" means, and nothing here has to work out where the
+  // viewers are.
+  //
+  // It was not read that way. This walked the film by number and proposed
+  // whatever was late, then a second pass divided the leftovers — an order of
+  // its own invention, which put the beginning of the file before the film in
+  // front of a viewer and, at one point, proposed #0, #1 and #2 as three
+  // separate places.
+  //
+  // One candidate per zone: the first number in it nobody has and nobody
+  // reaches in time. Zones with no deadline can have nothing late in them, so
+  // there it is simply the first number nobody has — which is how spare capacity
+  // comes to finish the file.
+  /** @type {number[]} */
+  const places = [];
+  const byRank = [...windows].sort(
+    (left, right) => (right.priority ?? 0) - (left.priority ?? 0) || left.from - right.from
+  );
+  for (const zone of byRank) {
+    if (places.length >= howMany) {
+      break;
+    }
+    const at = gapAt(zone.from, zone.to);
+    if (at === null || places.includes(at)) {
+      continue;
+    }
+    places.push(at);
+    placedHere.push(at);
+  }
+
+  // AND WHERE TO SPLIT WHAT IS LEFT, for the capacity the map has not spent.
+  //
+  // The search above proposes only what is LATE, so once one encoder covers a
+  // zone in time that zone proposes nothing more — and a machine that holds four
+  // ran one. Finishing a contiguous stretch soonest with several machines of the
+  // same speed means dividing it between them, which is where these come from:
+  // the widest run of film between two encoders, split.
+  //
+  // Proposing is not spending. The score decides whether another process is
+  // worth it, and its first term — how late the film is — always outranks its
+  // second, so this can never take capacity from somebody waiting.
+  while (places.length < howMany) {
+    const edges = [...places].sort((left, right) => left - right);
+    let widestFrom = null;
+    let widest = 0;
+    for (let index = 0; index <= edges.length; index += 1) {
+      const from = index === 0 ? firstOf(windows) : edges[index - 1] + 1;
+      const to = index === edges.length ? lastOf(windows) : edges[index] - 1;
+      const room_ = coverage.unmadeRunFrom(from);
+      if (to >= from && room_ > widest) {
+        // WHERE THE TWO OF THEM MEET, calculated rather than halved. Halving is
+        // the answer only when both encoders are fresh and owe the same, and it
+        // was applied to every case — one already working is partway through a
+        // piece while a new one owes its whole start, so the point where they
+        // finish together lies further on. Nothing back means a second encoder
+        // does not pay for itself here at all: two under this host's measured
+        // contention against one at full speed, a comparison halving never made.
+        const meet = splitAt(from, Math.min(from + room_ - 1, to));
+        if (meet !== null) {
+          widest = room_;
+          widestFrom = meet;
+        }
+      }
+    }
+    if (widestFrom === null) {
+      break;
+    }
+    const at = coverage.firstGapFrom(widestFrom, lastOf(windows));
+    if (at === null || places.includes(at)) {
+      break;
+    }
+    places.push(at);
+    placedHere.push(at);
+  }
+
+  // These are CANDIDATES, not decisions. What is late proposes first, because
+  // that is what a viewer feels; what is merely unmade proposes after it. The
+  // score decides which of them are worth a process, and a paused viewer, who
+  // states no deadline at all, therefore still leaves the file being finished
+  // rather than the machine falling idle.
+  return places.sort((left, right) => left - right);
+}
+