@torrent-tv/proxy 2.80.0 → 2.80.1

package/services/piece-store/shared-piece-store.js

@@ -1,1512 +1,1531 @@
-/**
- * @file Torrent pieces in shared memory, with disk as the second tier.
- *
- * Replaces the chunk store WebTorrent would otherwise build for itself. Two
- * reasons, one forced and one chosen.
- *
- * **Forced.** WebTorrent's own piece cache hands out the buffer it keeps using
- * and slices again on the next read. The torrent client now runs on its own
- * thread, and moving a piece to the main thread by transferring ownership
- * detached the cache's memory: proxy 2.9.71-2.9.73 answered every read with an
- * empty body (`Stream ends prematurely at 0`) and the failure was invisible,
- * because the error never reached the reader. Owning the memory ourselves
- * removes the question of whose it was.
- *
- * **Chosen.** Memory this side of the thread boundary can be *shared* memory,
- * which the main thread reads without receiving bytes as a copy — see
- * {@link SharedPieceStore#locate}. Measured on the field host: a piece copy
- * costs 3.64 ms, a read from the page cache into a buffer we own 7.63 ms, and
- * re-downloading a piece from the swarm ~1430 ms. So memory first, disk under
- * it, and never the swarm twice.
- *
- * Per piece allocation: each resident piece owns its own `SharedArrayBuffer`.
- * Evicting a piece deletes its entry and the memory is reclaimable by GC.
- * `committed` therefore equals `resident`, not a high-water mark.
- *
- * **Room for a piece is an owned reservation, not a shared number.**
- * `#claimSlot` hands back a release the caller runs in a `finally`; nothing
- * else touches `#outstandingPieces`. It was a counter incremented in one
- * function and decremented in another, and every consequence of that was a
- * defect: a failure between the two lost a slot for the life of the process,
- * `put`'s error path guessed at the correction and could take back a
- * reservation belonging to a different claim, and one lost reservation made
- * the five-second "everything is pinned" error permanently unreachable, so a
- * read retried every 50 ms for ever without completing or failing. Read out of
- * the field failure of 2026-08-31 (`research/worker-heap-oom-2026-08-31.md`).
- *
- * What this deliberately does NOT do is manage the disk as a cache of its own.
- * Pieces evicted from memory are written once and read back on demand; the file
- * is discarded whole when the torrent goes away. libtorrent 2.0 and webtor's
- * seeder both concluded that a hand-rolled disk cache earns less than it costs,
- * and nothing here disagrees.
- */
-
-import os from "node:os";
-import { readFileSync } from "node:fs";
-import { PieceLru } from "./piece-lru.js";
-import { DiskTier } from "./disk-tier.js";
-
-/**
- * Live stores, so the worker can report on them.
- *
- * @type {Set<SharedPieceStore>}
- */
-const liveStores = new Set();
-
-export function collectStoreStats() {
-  return [...liveStores].map((store) => store.stats());
-}
-
-export function findSharedStore(torrent) {
-  let candidate = torrent?.store;
-  for (let depth = 0; candidate && depth < 8; depth += 1) {
-    if (candidate instanceof SharedPieceStore) {
-      return candidate;
-    }
-    candidate = candidate.store;
-  }
-  return null;
-}
-
-/**
- * The largest fall in the machine's available memory that this process has
- * seen and did not cause itself.
- *
- * What has to be left for everything else on the machine. It cannot be a share
- * of what is free — a share is a number chosen out of nothing, which is what
- * `AVAILABLE_MEMORY_SHARE = 0.25`, `MIN_BUDGET_BYTES` and
- * `MEMORY_BUDGET_CEILING_BYTES` were until 2026-09-02, all three traceable to
- * one observation of one host on 2026-08-03. It is measured instead: between
- * two readings, how much available memory went away beyond what the stores
- * themselves took. On the field host that quantity is large — while the proxy
- * held 76-133 MB overnight the machine's available memory fell from 2378 MB to
- * 306 MB and came back — and on a quiet host it stays near zero, which is the
- * right answer there.
- *
- * Starts at zero: nothing is reserved until somebody else has been seen to
- * need it.
- */
-const otherDemand = { falls: [], lastAvailableBytes: 0, lastStoreBytes: 0 };
-
-/**
- * How many observations of other processes' demand are kept.
- *
- * A window rather than a high-water, and for the reason the block re-use gap is
- * one too: a single spike would otherwise stand for the life of the process and
- * squeeze the stores against something that happened once, hours ago.
- */
-const OTHER_DEMAND_SAMPLES = 60;
-
-/**
- * Note what the machine had, and how much of the change was not ours.
- *
- * @param {number} availableBytes
- * @param {number} storeBytes - What the live stores hold right now.
- * @returns {number} The reserve, in bytes.
- */
-export function noteMachineMemory(availableBytes, storeBytes) {
-  if (otherDemand.lastAvailableBytes > 0) {
-    const fell = otherDemand.lastAvailableBytes - availableBytes;
-    const ours = storeBytes - otherDemand.lastStoreBytes;
-    otherDemand.falls.push(Math.max(0, fell - ours));
-    if (otherDemand.falls.length > OTHER_DEMAND_SAMPLES) {
-      otherDemand.falls.shift();
-    }
-  }
-  otherDemand.lastAvailableBytes = availableBytes;
-  otherDemand.lastStoreBytes = storeBytes;
-  return machineReserveBytes();
-}
-
-/** What has recently been observed to be needed by everything that is not us. */
-export function machineReserveBytes() {
-  return otherDemand.falls.length === 0 ? 0 : Math.max(...otherDemand.falls);
-}
-
-/** Forget what other processes have needed. For tests, which share a module. */
-export function forgetMachineMemory() {
-  otherDemand.falls = [];
-  otherDemand.lastAvailableBytes = 0;
-  otherDemand.lastStoreBytes = 0;
-}
-
-/**
- * How much memory the stores may hold between them.
- *
- * `MemAvailable` is what could be allocated on top of what is already held, so
- * the stores' own bytes are added back: the pair is the ceiling the stores
- * could reach. The reserve is what has been seen to be needed elsewhere.
- *
- * @param {number} availableBytes
- * @param {number} storeBytes
- * @param {number} reserveBytes
- * @returns {number}
- */
-export function machineAllowanceBytes(availableBytes, storeBytes, reserveBytes) {
-  return Math.max(0, Math.max(availableBytes, 0) + Math.max(storeBytes, 0) - Math.max(reserveBytes, 0));
-}
-
-/**
- * Divide what the machine allows between the stores, by what each is asking
- * for.
- *
- * A store asks for the pieces its readers have declared. When everyone's ask
- * fits, everyone gets it and the machine's limit never binds — which is the
- * usual case, since two readers of one film declare 32-192 MB against gigabytes
- * of free memory. When the asks do not fit, each store is cut in proportion to
- * what it asked, so a store wanting little is not cut to make room for one
- * wanting much.
- *
- * @param {number[]} wantedBytes - What each store is asking for, in order.
- * @param {number} allowanceBytes
- * @returns {number[]} What each store may hold, in the same order.
- */
-export function divideAllowance(wantedBytes, allowanceBytes) {
-  const total = wantedBytes.reduce((sum, want) => sum + Math.max(0, want), 0);
-  if (total <= allowanceBytes || total === 0) {
-    return wantedBytes.map((want) => Math.max(0, want));
-  }
-  return wantedBytes.map((want) => Math.floor(allowanceBytes * (Math.max(0, want) / total)));
-}
-
-export function reviseStoreBudgets() {
-  const stores = [...liveStores];
-  const held = stores.reduce((sum, store) => sum + store.residentBytes, 0);
-  const available = availableMemorySync();
-  const reserve = noteMachineMemory(available, held);
-  const allowance = machineAllowanceBytes(available, held, reserve);
-  const shares = divideAllowance(stores.map((store) => store.wantedBytes), allowance);
-  const revised = [];
-  for (const [position, store] of stores.entries()) {
-    revised.push(store.reviseGrowthCeiling(shares[position]));
-  }
-  return revised;
-}
-
-function defaultMemoryBytes() {
-  const stores = [...liveStores];
-  const held = stores.reduce((sum, store) => sum + store.residentBytes, 0);
-  const available = availableMemorySync();
-  const allowance = machineAllowanceBytes(available, held, machineReserveBytes());
-  // A store being created has no readers, so it has no demand to state and no
-  // basis for asking for more or less than the others. It opens on an equal
-  // share and the first revision — within a minute, and within seconds of a
-  // read starting — replaces that with what its readers actually declare.
-  // Deliberately not the whole allowance: on a machine with gigabytes free that
-  // would let a torrent nobody is reading yet fill memory before the first
-  // revision arrives.
-  return Math.floor(allowance / (stores.length + 1));
-}
-
-function availableMemorySync() {
-  try {
-    const text = readFileSync("/proc/meminfo", "utf8");
-    const match = /^MemAvailable:\s+(\d+)\s+kB$/m.exec(text);
-    if (match) {
-      return Number(match[1]) * 1024;
-    }
-  } catch {
-  }
-  return os.freemem();
-}
-
-/**
- * How many blocks of piece memory this thread has let go of, and how many the
- * collector has actually taken back.
- *
- * Since the store keeps a pool, one block serves many pieces, so this counts
- * blocks and not pieces — and a healthy store allocates only as many as its
- * allowance permits, so both numbers are now small and nearly equal.
- *
- * The one reading that separates the two explanations of the 700 MB nobody can
- * account for (roadmap item 2, field 2026-08-31): if the two numbers track each
- * other, this code is holding nothing and whatever grows is below us, in the
- * allocator — on musl there is no `malloc_trim` and no way to ask. If the gap
- * widens, a reference of ours outlives the piece, and then a heap snapshot can
- * name the holder.
- *
- * What it does NOT prove: a `SharedArrayBuffer`'s memory is shared between the
- * isolates, so this thread's handle going means only that THIS thread let go.
- * The main thread counts its own (`torrent-worker/client.js`), and the pair is
- * what answers the question.
- */
-const released = { count: 0, collected: 0 };
-const collector = typeof FinalizationRegistry === "function"
-  ? new FinalizationRegistry(() => {
-    released.collected += 1;
-  })
-  : null;
-
-/**
- * What the collector has taken back against what was let go.
- *
- * @returns {{ released: number, collected: number }}
- */
-export function pieceBufferCollection() {
-  return { released: released.count, collected: released.collected };
-}
-
-const MIN_RESIDENT_PIECES = 2;
-/**
- * How long a claim may go without ANYTHING moving before it gives up.
- *
- * Measured against progress, not against activity. The earlier rule skipped
- * this timer entirely while a spill was in flight or a reservation was held —
- * so one reservation that was never returned made the timer unreachable and a
- * read retried every 50 ms for the life of the process, never completing and
- * never failing (field 2026-08-31).
- */
-const PINNED_WAIT_MS = 5_000;
-
-/**
- * How many revival ages are kept for the median. A window, not a history: two
- * hundred covers several minutes of the busiest session measured (7575
- * revivals in 44 minutes) and costs two hundred numbers.
- */
-const REVIVAL_AGE_SAMPLES = 200;
-
-/**
- * How many block re-use gaps are kept. The same window as the revival ages, and
- * for the same reason: what is wanted is the rhythm of recent work, not a
- * history of it.
- */
-const REUSE_GAP_SAMPLES = 200;
-
-/** How many write durations are kept for the median. */
-const WRITE_DURATION_SAMPLES = 50;
-/** How many admission times are kept, and how far back they are counted. */
-const ARRIVAL_SAMPLES = 100;
-const ARRIVAL_WINDOW_MS = 10_000;
-
-/**
- * The middle value of a sample, or null when there is nothing to take a middle
- * of. Null rather than zero: no revivals and instant revivals are different
- * facts and must not print the same.
- *
- * @param {number[]} values
- * @returns {number | null}
- */
-function median(values) {
-  if (values.length === 0) {
-    return null;
-  }
-  const sorted = [...values].sort((left, right) => left - right);
-  const middle = Math.floor(sorted.length / 2);
-  return sorted.length % 2 === 0
-    ? Math.round((sorted[middle - 1] + sorted[middle]) / 2)
-    : sorted[middle];
-}
-const CLAIM_RETRY_MS = 50;
-
-export class SharedPieceStore {
-  #chunkLength;
-  #lastChunkLength;
-  #lastChunkIndex;
-  #growthCeiling;
-  /** Piece index → SharedArrayBuffer of that piece */
-  #buffers = new Map();
-  /** @type {Map<number, Promise<void>>} */
-  #evicting = new Map();
-  /**
-   * Slots claimed but not yet filled.
-   *
-   * Handed out by {@link SharedPieceStore##claimSlot} as a release function the
-   * caller must call in a `finally`, never as a number one function increments
-   * and another decrements. The counter used to be paired across
-   * `#claimSlot`/`#registerPiece`, so any failure between the two lost a slot
-   * for the life of the process, and `put`'s error path tried to correct that
-   * by guessing — which could take back a reservation belonging to a different
-   * claim and let the store admit past its allowance.
-   */
-  #outstandingPieces = 0;
-  /** When something last actually moved: a piece admitted, spilled or unpinned. */
-  #lastProgressAt = 0;
-  #waiters = [];
-  #lru;
-  #disk;
-  #closed = false;
-  #name;
-  #counters = {
-    fromMemory: 0,
-    fromDisk: 0,
-    spills: 0,
-    revivals: 0,
-    blockedByPins: 0,
-    waitedForPins: 0,
-    evictedOnRevise: 0,
-    spillFailures: 0,
-    // Whether the store is doing its job or being asked to hold more than it
-    // has room for. An eviction that had to take a piece a reader declared it
-    // wants is the second, and it comes back from disk moments later
-    // (roadmap item 9).
-    evictedProtected: 0,
-    evictedDistanceSum: 0,
-    evictedWithDistance: 0,
-    // Where an arriving piece went. A piece nobody has declared is being
-    // downloaded ahead of every reader; putting it in memory pushes out one
-    // that IS declared, which is then read back from disk moments later
-    // (roadmap item 9).
-    admittedInsideWindow: 0,
-    admittedOutsideWindow: 0,
-    admittedToDisk: 0,
-    /** Blocks given back to the operating system. */
-    blocksReleased: 0,
-    /**
-     * A block put back for re-use while a reader still held the piece. Zero by
-     * construction — a pinned piece is never evicted, and a re-put of a pinned
-     * one drops its block instead of recycling it. Counted because if it ever
-     * stops being zero, a consumer is reading another piece's bytes, and that
-     * is invisible from anywhere else.
-     */
-    returnedWhilePinned: 0,
-    /** Spills that found the disk already holding identical bytes. */
-    spillsSkipped: 0,
-    /**
-     * Blocks a second registration of the same piece displaced. Expected to be
-     * small and non-zero: two callers racing for one piece is ordinary. What is
-     * NOT ordinary is the block going missing when it happens, which is what
-     * killed the process on 2026-09-02.
-     */
-    blocksDisplaced: 0,
-    /** Admissions that waited for the disk instead of evicting another piece. */
-    waitedForDisk: 0,
-    /**
-     * Admissions that grew memory because the disk had stopped answering.
-     * Non-zero means the store exceeded its allowance on purpose, and by how
-     * many pieces.
-     */
-    grewWaitingForDisk: 0,
-    /**
-     * Times a block was wanted, none was free, and the pool was already at its
-     * ceiling. Zero by construction — a block is only taken after a slot has
-     * been claimed, and slots are what the ceiling counts — so a number here
-     * means the two have come apart and the pool is growing past its allowance.
-     */
-    blocksBeyondCeiling: 0
-  };
-  /**
-   * Blocks that hold no piece, most recently freed last.
-   *
-   * A block is one piece's worth of memory. Allocating a new one for every
-   * piece meant 7575 allocations of 4 MiB in 44 minutes on 2026-09-02, each
-   * released only when the collector got to it — which is why the process held
-   * 1.86 GB while its own accounting said 352 MB. Blocks are taken from here
-   * and put back here instead (roadmap item 2).
-   *
-   * @type {{ buffer: SharedArrayBuffer, freedAt: number }[]}
-   */
-  #freeBlocks = [];
-  /** Blocks that exist at all: free plus holding a piece. */
-  #blocksAllocated = 0;
-  /** Whether a reader has ever declared a window here. See `wantedBytes`. */
-  #everHadReader = false;
-  /** Whether the last revision had to exceed the machine's share. */
-  #beyondTheMachine = false;
-  /**
-   * How long a block sat free before it was taken again, in milliseconds.
-   * Bounded, because what is wanted is the longest gap of RECENT work: an
-   * all-time maximum would be raised by one long pause and then never let a
-   * block go again.
-   *
-   * @type {number[]}
-   */
-  #reuseGaps = [];
-  /**
-   * How long recent writes to disk took, in milliseconds, and when pieces were
-   * admitted. Together they say how much room the store needs beyond what the
-   * readers ask for: while one write is finishing, more pieces arrive, and each
-   * needs somewhere to go. Without that room every arrival evicts something,
-   * which is what produced 233 evictions against 119 completed writes in a
-   * minute on 2026-09-02.
-   *
-   * @type {number[]}
-   */
-  #writeDurations = [];
-  /** When the last few pieces were admitted, for the arrival rate. */
-  #admittedAt = [];
-  /** Piece index → when it was written out, for the age it comes back at. */
-  #spilledAt = new Map();
-  /**
-   * Ages, in milliseconds, of the last revivals — bounded, because the figure
-   * wanted is a median and not a history. A piece that comes back seconds
-   * after it left should not have left.
-   *
-   * @type {number[]}
-   */
-  #revivalAges = [];
-
-  constructor(chunkLength, options = {}) {
-    if (!Number.isInteger(chunkLength) || chunkLength < 1) {
-      throw new Error(`Chunk length must be a positive integer, got ${chunkLength}.`);
-    }
-    this.#chunkLength = chunkLength;
-
-    const totalLength = Number.isFinite(options.length) ? options.length : 0;
-    this.#lastChunkIndex = totalLength > 0 ? Math.ceil(totalLength / chunkLength) - 1 : -1;
-    const remainder = totalLength % chunkLength;
-    this.#lastChunkLength = remainder === 0 ? chunkLength : remainder;
-
-    const memoryBytes = Number.isFinite(options.memoryBytes) && options.memoryBytes > 0
-      ? options.memoryBytes
-      : defaultMemoryBytes();
-    // Only where this store STARTS. It is not kept, so it cannot come back as a
-    // cap on the revision the way `#capacity` did.
-    const openingCeiling = Math.max(MIN_RESIDENT_PIECES, Math.floor(memoryBytes / chunkLength));
-    this.#growthCeiling = openingCeiling;
-    this.#lru = new PieceLru(openingCeiling);
-    this.#name = options.name ?? "pieces";
-    // `options.disk` exists so a test can hold a write open or make one fail on
-    // purpose. Four of the defects fixed here live in what happens when the
-    // disk tier does not answer immediately or at all, and none of them is
-    // reachable from outside without saying so.
-    this.#disk = options.disk ?? new DiskTier({
-      directory: options.path ?? ".",
-      name: `${this.#name}.pieces`,
-      chunkLength
-    });
-    liveStores.add(this);
-  }
-
-  stats() {
-    const resident = this.#buffers.size;
-    const residentBytes = resident * this.#chunkLength;
-    // Last piece may be short but stats historically use chunkLength.
-    return {
-      name: this.#name,
-      resident,
-      capacity: this.#growthCeiling,
-      residentBytes,
-      allocatedSlots: this.#blocksAllocated,
-      // What the process HOLDS, which with a pool is the blocks that exist —
-      // not the pieces in them. Holding and using are different quantities and
-      // the difference is the point of the reading.
-      committedBytes: this.#blocksAllocated * this.#chunkLength,
-      budgetBytes: this.#growthCeiling * this.#chunkLength,
-      pinned: this.#lru.pinnedCount,
-      // Slots claimed and not yet filled. Reported because a reservation that
-      // is never returned is invisible until the store cannot admit anything,
-      // and by then the reason is long gone. At rest this is zero.
-      outstanding: this.#outstandingPieces,
-      spilled: this.#disk.size,
-      spilledBytes: this.#disk.size * this.#chunkLength,
-      // What the readers between them are asking this store to keep, against
-      // what it may hold. A union wider than the capacity cannot be held
-      // however the eviction is ordered, and that is the difference between a
-      // policy to fix and arithmetic to accept (roadmap item 9).
-      demand: this.#lru.demand(),
-      // What the process actually holds for this store, which is not the same
-      // as what it is using: blocks are kept for re-use. The difference is the
-      // spare, and it is the whole of the question whether consumption only
-      // grows (roadmap item 2).
-      blocksAllocated: this.#blocksAllocated,
-      blocksFree: this.#freeBlocks.length,
-      blocksDisplaced: this.#counters.blocksDisplaced,
-      blocksInUse: this.#blocksInUse(),
-      blocksInFlight: this.#blocksInFlight(),
-      blocksBeyondCeiling: this.#counters.blocksBeyondCeiling,
-      blockBytes: this.#blocksAllocated * this.#chunkLength,
-      reuseGapMs: this.#reuseGapCeilingMs(),
-      revivalAgeMedianMs: median(this.#revivalAges),
-      revivalAgeSamples: this.#revivalAges.length,
-      revivedWithinFiveSeconds: this.#revivalAges.filter((age) => age <= 5_000).length,
-      ...this.#counters
-    };
-  }
-
-  /**
-   * Whether the machine's share of memory is smaller than one reader's window.
-   *
-   * The store holds the window anyway — refusing would leave the read it is
-   * serving unable to finish, which is worse — but it is the honest measure of
-   * "this machine cannot take any more", and it is measured rather than
-   * guessed: it is the last revision's own comparison of what the machine
-   * allowed against what the widest reader declared.
-   */
-  get isBeyondTheMachine() {
-    return this.#beyondTheMachine;
-  }
-
-  /** What this store holds right now, in bytes. */
-  get residentBytes() {
-    return this.#buffers.size * this.#chunkLength;
-  }
-
-  /**
-   * What this store is asking to be allowed to hold, in bytes.
-   *
-   * The union of its readers' declared windows — what they have said they will
-   * need — never below the widest single window, because a store that cannot
-   * hold one reader's window cannot complete that reader's read at all: every
-   * resident piece ends up pinned, the read returns zero bytes and ffmpeg takes
-   * that for the end of the file (field 2026-08-15, roadmap item 9).
-   *
-   * With no reader declaring anything there is no demand to speak of, so the
-   * store asks for what the machine allows and the first revision after a read
-   * begins brings it down to what that read needs.
-   */
-  get wantedBytes() {
-    const demand = this.#lru.demand();
-    if (demand.readers > 0) {
-      this.#everHadReader = true;
-      const pieces = Math.max(MIN_RESIDENT_PIECES, demand.unionPieces, demand.widestPieces);
-      // Plus room to absorb what arrives while a write is finishing. Asking for
-      // exactly what the readers want leaves no free place ever, so every
-      // arrival evicts one of them — measured 2026-09-02: `6 reader(s) want 23
-      // piece(s) of 23 the store may hold`, and 233 evictions in the minute
-      // that followed.
-      return (pieces + this.slackPieces()) * this.#chunkLength;
-    }
-    // Readers that have GONE are not the same as readers that have not arrived.
-    // A store whose readers ended has nothing to hold pieces for — its torrent
-    // sits until the pool's idle timer removes it, which needs a refcount of
-    // zero and can be a quarter of an hour away — so it asks for nothing and
-    // its memory goes back to the machine now. A store that has never had a
-    // reader is being filled for one that is on its way, and asks for what it
-    // was opened with until the first read says what it needs.
-    return this.#everHadReader
-      ? (MIN_RESIDENT_PIECES + this.slackPieces()) * this.#chunkLength
-      : this.#growthCeiling * this.#chunkLength;
-  }
-
-  reviseGrowthCeiling(allowedBytes) {
-    // No cap at what the machine could spare when this store was CREATED.
-    // `#capacity` was computed once in the constructor and used as an upper
-    // bound here, so the allowance could only ever fall: a torrent opened while
-    // the machine was full kept a small allowance for its whole life, however
-    // much memory was freed afterwards (roadmap item 2, 2026-09-02).
-    const wanted = Math.floor(Number(allowedBytes) / this.#chunkLength);
-    // Never below one reader's whole window while a reader exists, even when
-    // the machine's share says less. A store that cannot hold the window of the
-    // read it is serving cannot complete that read at all: every resident piece
-    // ends up pinned, the read returns zero bytes and ffmpeg takes that for the
-    // end of the file, which killed every encoder on that file in the field on
-    // 2026-08-15. Exceeding the share is the lesser failure, and the line below
-    // says when it happens.
-    const demand = this.#lru.demand();
-    this.#growthCeiling = Math.max(
-      MIN_RESIDENT_PIECES,
-      demand.readers > 0 ? demand.widestPieces : MIN_RESIDENT_PIECES,
-      Number.isFinite(wanted) ? wanted : MIN_RESIDENT_PIECES
-    );
-    const belowAWindow = demand.readers > 0
-      && Number.isFinite(wanted)
-      && wanted < demand.widestPieces;
-    this.#beyondTheMachine = belowAWindow;
-    // The LRU is told too. It was constructed with the store's original
-    // capacity and never revised, so `isFull()` answered against a number that
-    // had not been the limit for some time — dormant only because nothing calls
-    // it, which is a trap for whoever calls it next.
-    this.#lru.setCapacity(this.#growthCeiling);
-    // With per-piece buffers memory CAN be given back immediately, unlike the
-    // old growable pool. Eagerly evict excess to honour the new ceiling.
-    let evicted = 0;
-    while (this.#buffers.size > this.#growthCeiling) {
-      const { index: victim, protectionYielded, distance } = this.#lru.evictionChoice();
-      if (victim === null) {
-        break;
-      }
-      const victimBuffer = this.#buffers.get(victim);
-      if (victimBuffer === undefined) {
-        this.#lru.remove(victim);
-        continue;
-      }
-      this.#buffers.delete(victim);
-      this.#lru.remove(victim);
-      evicted += 1;
-      this.#counters.evictedOnRevise += 1;
-      this.#noteEviction(protectionYielded, distance);
-      // Nobody awaits this spill, so its failure has to end here. Rethrowing
-      // made it an unhandled rejection, and an unhandled rejection in the
-      // torrent worker ends the thread — a second way to lose the torrent
-      // client, on top of the one that already loses it.
-      void this.#spill(victim, victimBuffer).catch(() => {
-        this.#counters.spillFailures += 1;
-      });
-    }
-    // Blocks the store is no longer using and has waited long enough to give
-    // up. The allowance falling is the moment to ask, because that is when the
-    // machine has been shown to need the memory.
-    const releasedBlocks = this.sweepFreeBlocks();
-    return {
-      name: this.#name,
-      ceilingBytes: this.#growthCeiling * this.#chunkLength,
-      committedBytes: this.#blocksAllocated * this.#chunkLength,
-      evicted,
-      releasedBlocks,
-      belowAWindow
-    };
-  }
-
-  get chunkLength() {
-    return this.#chunkLength;
-  }
-
-  get capacity() {
-    return this.#growthCeiling;
-  }
-
-  get residentCount() {
-    return this.#buffers.size;
-  }
-
-  get spilledCount() {
-    return this.#disk.size;
-  }
-
-  #lengthOf(index) {
-    return index === this.#lastChunkIndex ? this.#lastChunkLength : this.#chunkLength;
-  }
-
-  /**
-   * Where a resident piece sits, or `null` if not resident.
-   * Returns the piece's own SharedArrayBuffer and the intra-piece range.
-   * @param {number} index
-   * @returns {{ buffer: SharedArrayBuffer, offset: number, length: number } | null}
-   */
-  locate(index) {
-    const buffer = this.#buffers.get(index);
-    if (buffer === undefined) {
-      return null;
-    }
-    return { buffer, offset: 0, length: this.#lengthOf(index) };
-  }
-
-  /**
-   * Direct access to a piece's buffer for zero-copy consumers.
-   * @param {number} index
-   * @returns {SharedArrayBuffer | undefined}
-   */
-  getPieceBuffer(index) {
-    return this.#buffers.get(index);
-  }
-
-  pin(index) {
-    this.#lru.pin(index);
-  }
-
-  unpin(index) {
-    this.#lru.unpin(index);
-    this.#noteProgress();
-  }
-
-  /**
-   * Reserve room for one piece.
-   *
-   * @returns {Promise<() => void>} The release, which the caller MUST call in a
-   *   `finally`. Calling it twice is harmless.
-   */
-  async #claimSlot() {
-    // How long THIS claim has been trying, kept here and not on the store.
-    // It was a field, `#pinnedWaitStartedAt`, shared by the two waits inside
-    // `#claimSlotOnce` — the one for the disk and the one for a piece that may
-    // be evicted — and each of them zeroed it on giving up, which restarted the
-    // other's clock. Measured 2026-09-03: the claim cycled for ever, five
-    // seconds per side, `grewWaitingForDisk` and `waitedForPins` both climbing
-    // once every five seconds while `blockedByPins` stayed at 0 — the refusal
-    // was unreachable. A claim's patience belongs to the claim; with several
-    // claimants a shared field is wrong anyway, since one caller giving up
-    // would reset the wait of every other.
-    const waitingSince = Date.now();
-    for (;;) {
-      if (this.#closed) {
-        throw new Error("Piece store is closed.");
-      }
-      const ok = await this.#claimSlotOnce(waitingSince);
-      if (ok) {
-        let released = false;
-        return () => {
-          if (released) {
-            return;
-          }
-          released = true;
-          this.#outstandingPieces -= 1;
-          this.#noteProgress();
-        };
-      }
-      await this.#waitForSlot();
-    }
-  }
-
-  /**
-   * Sleep until something moves, or until the retry interval, whichever first.
-   *
-   * One handler, idempotent, and its timer is cleared when it is woken. The
-   * earlier version attached a fresh pair of handlers to EVERY pending spill on
-   * every attempt and left a timer running each time, so a claim that could not
-   * be satisfied allocated in proportion to attempts times pending spills. A
-   * settling spill now wakes the store itself, which is where that belongs.
-   *
-   * @returns {Promise<void>}
-   */
-  #waitForSlot() {
-    return new Promise((resolve) => {
-      let settled = false;
-      /** @type {ReturnType<typeof setTimeout> | null} */
-      let retry = null;
-      const finish = () => {
-        if (settled) {
-          return;
-        }
-        settled = true;
-        if (retry !== null) {
-          clearTimeout(retry);
-        }
-        resolve();
-      };
-      this.#waiters.push(finish);
-      retry = setTimeout(finish, CLAIM_RETRY_MS);
-      retry.unref?.();
-    });
-  }
-
-  #registerPiece(index, buffer) {
-    // A piece may already be here. Both paths that register one look first and
-    // then await — `put` waits for a slot, `#revive` waits for the disk — and
-    // in that gap another caller can register the same index. Overwriting the
-    // entry used to drop the previous block on the floor: its memory was not
-    // returned to the pool and the pool's own count was never decremented, so
-    // the count climbed past the ceiling for ever and the pool degenerated into
-    // allocating a fresh block per piece. Field 2026-09-02: a store holding
-    // THREE pieces reported 812 MB committed against 68 MB allowed, 739 blocks
-    // allocated and 676 still alive, and the process was killed at 4.37 GB.
-    const displaced = this.#buffers.get(index);
-    this.#buffers.set(index, buffer);
-    if (displaced !== undefined && displaced !== buffer) {
-      this.#counters.blocksDisplaced += 1;
-      this.#returnBlock(displaced);
-    }
-    this.#lru.touch(index);
-    this.#noteProgress();
-  }
-
-  /**
-   * Write a piece out and account for it, from the one place that does it.
-   *
-   * @param {number} index
-   * @param {SharedArrayBuffer} buffer
-   * @returns {Promise<void>}
-   */
-  #spill(index, buffer) {
-    // The disk may already hold these very bytes. `#revive` reads a piece back
-    // into memory and leaves the copy on disk, and only `put` removes it — so a
-    // piece that was revived and not re-put is identical to what is already
-    // written, and writing it again is work for nothing. There were 7575
-    // revivals in one session on 2026-09-02, and every later eviction of one of
-    // them wrote a second time (roadmap item 66: 14.4 GB written in a single
-    // viewing).
-    if (this.#disk.has(index)) {
-      this.#counters.spills += 1;
-      this.#counters.spillsSkipped += 1;
-      this.#spilledAt.set(index, Date.now());
-      this.#returnBlock(buffer);
-      this.#noteProgress();
-      return Promise.resolve();
-    }
-
-    const bytes = Buffer.from(buffer, 0, this.#lengthOf(index));
-    const startedAt = Date.now();
-    const spill = this.#disk.write(index, bytes).then(
-      () => {
-        this.#noteWriteDuration(Date.now() - startedAt);
-        this.#counters.spills += 1;
-        this.#spilledAt.set(index, Date.now());
-        this.#evicting.delete(index);
-        // Only now. The write reads out of this block, so a block handed to
-        // another piece before the write finished would put that piece's bytes
-        // into this piece's place in the file.
-        this.#returnBlock(buffer);
-        this.#noteProgress();
-      },
-      (error) => {
-        this.#evicting.delete(index);
-        // The block is no longer holding anything either way; keeping it out of
-        // the pool because the write failed would lose it for good.
-        this.#returnBlock(buffer);
-        this.#noteProgress();
-        throw error;
-      }
-    );
-    this.#evicting.set(index, spill);
-    return spill;
-  }
-
-  /** Something actually moved: wake whoever is waiting and restart the clock. */
-  #noteProgress() {
-    this.#lastProgressAt = Date.now();
-    this.#wake();
-  }
-
-  /**
-   * Record how long one write to disk took.
-   *
-   * @param {number} durationMs
-   * @returns {void}
-   */
-  #noteWriteDuration(durationMs) {
-    this.#writeDurations.push(Math.max(0, durationMs));
-    if (this.#writeDurations.length > WRITE_DURATION_SAMPLES) {
-      this.#writeDurations.shift();
-    }
-  }
-
-  /** Record that a piece arrived, for the arrival rate. */
-  #noteArrival() {
-    const now = Date.now();
-    this.#admittedAt.push(now);
-    while (this.#admittedAt.length > ARRIVAL_SAMPLES
-      || (this.#admittedAt.length > 0 && now - this.#admittedAt[0] > ARRIVAL_WINDOW_MS)) {
-      this.#admittedAt.shift();
-    }
-  }
-
-  /**
-   * How many pieces the store needs room for beyond what the readers ask for.
-   *
-   * Measured, not chosen: the pieces that arrive while one write to disk is
-   * finishing. Without this room every arrival must evict something, and each
-   * eviction holds its block until its write completes — so a disk slower than
-   * the swarm turns every admission into one more block held. On 2026-09-02
-   * that was 233 evictions against 119 completed writes in a minute, and 203
-   * blocks held with three pieces resident.
-   *
-   * Zero until both quantities have been seen, because a slack invented before
-   * anything is measured is a chosen number, and this store has been bitten by
-   * those.
-   *
-   * @returns {number} Pieces.
-   */
-  slackPieces() {
-    if (this.#writeDurations.length === 0 || this.#admittedAt.length < 2) {
-      return 0;
-    }
-    const sorted = [...this.#writeDurations].sort((left, right) => left - right);
-    const writeMs = sorted[Math.floor(sorted.length / 2)];
-    const spanMs = this.#admittedAt[this.#admittedAt.length - 1] - this.#admittedAt[0];
-    if (!(spanMs > 0)) {
-      return 0;
-    }
-    const perMs = (this.#admittedAt.length - 1) / spanMs;
-    return Math.ceil(perMs * writeMs);
-  }
-
-  /**
-   * Blocks that are not free: resident pieces, blocks being written out, and
-   * blocks taken but not yet registered.
-   *
-   * This is what the ceiling has to bound, and until 2026-09-02 it bounded
-   * resident pieces instead. The difference is exactly the memory that goes
-   * missing from the count: a piece being spilled leaves `#buffers` the moment
-   * the eviction begins, while its block stays held until the write it feeds
-   * has finished.
-   *
-   * @returns {number}
-   */
-  #blocksInUse() {
-    return this.#blocksAllocated - this.#freeBlocks.length;
-  }
-
-  /**
-   * Blocks held by writes that have not finished.
-   *
-   * @returns {number}
-   */
-  #blocksInFlight() {
-    return Math.max(0, this.#blocksInUse() - this.#buffers.size);
-  }
-
-  /**
-   * Whether admitting one more piece would need something evicted first.
-   *
-   * Reservations count: a slot claimed and not yet filled is as taken as a
-   * resident piece.
-   *
-   * @returns {boolean}
-   */
-  #isFullNow() {
-    return this.#blocksInUse() + this.#outstandingPieces >= this.#growthCeiling;
-  }
-
-  /**
-   * Write an arriving piece straight to disk, without it ever occupying memory.
-   *
-   * Registered in `#evicting` like a spill, so a `get` for this index waits for
-   * the write instead of finding the piece on neither tier. Deliberately NOT
-   * recorded in `#spilledAt`: that clock measures how long an EVICTED piece
-   * stayed away, and a piece that was never resident has no such age.
-   *
-   * @param {number} index
-   * @param {Uint8Array} bytes
-   * @returns {Promise<void>}
-   */
-  #writeThrough(index, bytes) {
-    // Copied, not viewed. `DiskTier.write` opens the file before it reads the
-    // bytes, so a view onto the caller's buffer could be written to in between
-    // and the file would get the wrong data. The spill path may pass a view
-    // because that memory is ours; this buffer belongs to the torrent client.
-    // It costs nothing extra: the piece was being copied into a shared buffer
-    // on this path before, and now it is copied here instead.
-    const copy = Buffer.from(bytes.subarray(0, Math.min(bytes.length, this.#lengthOf(index))));
-    const write = this.#disk.write(index, copy).then(
-      () => {
-        this.#evicting.delete(index);
-        this.#noteProgress();
-      },
-      (error) => {
-        this.#evicting.delete(index);
-        this.#noteProgress();
-        throw error;
-      }
-    );
-    this.#evicting.set(index, write);
-    return write;
-  }
-
-  /**
-   * Record how long a piece stayed on disk before it was wanted again.
-   *
-   * A piece that comes back seconds after it left was evicted from a working
-   * set that does not fit, and the write and the read were both waste. Kept as
-   * a bounded window of ages because the figure wanted is a median, not a
-   * history.
-   *
-   * @param {number} index
-   * @returns {void}
-   */
-  #noteRevival(index) {
-    const spilledAt = this.#spilledAt.get(index);
-    if (spilledAt === undefined) {
-      return;
-    }
-    this.#spilledAt.delete(index);
-    this.#revivalAges.push(Date.now() - spilledAt);
-    if (this.#revivalAges.length > REVIVAL_AGE_SAMPLES) {
-      this.#revivalAges.shift();
-    }
-  }
-
-  /**
-   * Record what an eviction had to take.
-   *
-   * @param {boolean} protectionYielded - The victim was inside a window a
-   *   reader had declared, and was taken anyway because nothing else was free.
-   * @param {number} distance - Pieces from the nearest declared window, -1 when
-   *   no reader declared one.
-   * @returns {void}
-   */
-  #noteEviction(protectionYielded, distance) {
-    if (protectionYielded) {
-      this.#counters.evictedProtected += 1;
-    }
-    if (distance >= 0) {
-      this.#counters.evictedDistanceSum += distance;
-      this.#counters.evictedWithDistance += 1;
-    }
-  }
-
-  #wake() {
-    const waiting = this.#waiters;
-    this.#waiters = [];
-    for (const resolve of waiting) {
-      resolve();
-    }
-  }
-
-  /**
-   * One attempt at a reservation.
-   *
-   * @param {number} waitingSince - When the claim this attempt belongs to began
-   *   trying. Both waits below are measured from the later of it and the last
-   *   time the store moved, so neither can restart the other's clock — see
-   *   `#claimSlot`.
-   * @returns {Promise<boolean>} Whether a slot was reserved.
-   */
-  async #claimSlotOnce(waitingSince) {
-    // Reserve before suspension so concurrent callers see the reservation.
-    if (this.#blocksInUse() + this.#outstandingPieces < this.#growthCeiling) {
-      this.#outstandingPieces += 1;
-      return true;
-    }
-    const stillFor = () => Date.now() - Math.max(waitingSince, this.#lastProgressAt);
-
-    // Full, and evicting would make it worse rather than better. A piece being
-    // written out has already left `#buffers` while its block is still held —
-    // the write reads from that block — so evicting another one converts a
-    // resident block into an in-flight block and takes a fresh block for the
-    // arrival: the memory in use goes UP by one per admission for as long as
-    // the disk is behind. Field 2026-09-02: 233 evictions against 119 completed
-    // writes in a minute, 203 blocks held with three pieces resident, and the
-    // process killed at 4.37 GB.
-    //
-    // So when the disk is what the store is waiting for, it waits. A completing
-    // write calls `#noteProgress`, which wakes whoever is here.
-    if (this.#blocksInFlight() > 0 && this.#blocksInUse() >= this.#growthCeiling) {
-      if (stillFor() < PINNED_WAIT_MS) {
-        this.#counters.waitedForDisk += 1;
-        return false;
-      }
-      // The disk has stopped answering. Falling through to eviction is the
-      // lesser failure: it grows memory, and the line above says by how much,
-      // where refusing would fail the read outright. The clock is NOT restarted
-      // here: the wait that follows asks the same question of the same claim,
-      // and zeroing it was half of the loop described in `#claimSlot`.
-      this.#counters.grewWaitingForDisk += 1;
-    }
-
-    const { index: victim, protectionYielded, distance } = this.#lru.evictionChoice();
-    if (victim === null) {
-      // Nothing may leave. Wait while the store is still MOVING — a spill
-      // completing, a piece admitted, a pin released — and give up when it has
-      // not moved for PINNED_WAIT_MS, whatever is nominally in flight. The old
-      // rule asked whether anything was in flight rather than whether anything
-      // had happened, which is why one lost reservation could hold a read here
-      // for ever.
-      if (stillFor() < PINNED_WAIT_MS) {
-        this.#counters.waitedForPins += 1;
-        return false;
-      }
-      this.#counters.blockedByPins += 1;
-      throw new Error(
-        `Every resident piece is pinned and nothing moved for ${PINNED_WAIT_MS}ms; no slot can be freed.`
-      );
-    }
-
-    const victimBuffer = this.#buffers.get(victim);
-    if (victimBuffer === undefined) {
-      // Should not happen: LRU says resident but buffer missing.
-      this.#lru.remove(victim);
-      return false;
-    }
-
-    // Claim atomically before await.
-    this.#buffers.delete(victim);
-    this.#lru.remove(victim);
-    this.#outstandingPieces += 1;
-    this.#noteEviction(protectionYielded, distance);
-
-    try {
-      await this.#spill(victim, victimBuffer);
-    } catch (error) {
-      // The caller never received a release for this reservation, so it is
-      // given back here rather than left outstanding for ever.
-      this.#outstandingPieces -= 1;
-      this.#noteProgress();
-      throw error;
-    }
-    // Outstanding stays +1 for the caller; the slot for the new piece is now free.
-    return true;
-  }
-
-  /**
-   * Bring a spilled piece back into memory, once, however many callers ask.
-   *
-   * @param {number} index
-   * @returns {Promise<SharedArrayBuffer | null>} `null` when the piece is on
-   *   neither tier.
-   */
-  async #revive(index) {
-    const spill = this.#evicting.get(index);
-    if (spill) {
-      await spill.catch(() => undefined);
-    }
-
-    if (!this.#disk.has(index)) {
-      return null;
-    }
-
-    const release = await this.#claimSlot();
-    try {
-      // Another caller may have brought it back while this one waited for a
-      // slot. Registering a second buffer for the same piece would leave
-      // whoever holds the first reading memory nothing evicts.
-      const already = this.#buffers.get(index);
-      if (already !== undefined) {
-        this.#lru.touch(index);
-        this.#counters.fromMemory += 1;
-        return already;
-      }
-      const target = this.#takeBlock();
-      try {
-        // Only this piece's own length: the block is a full piece long and the
-        // last piece of a file is shorter, so reading the whole block would ask
-        // the file for bytes past its end.
-        await this.#disk.read(index, Buffer.from(target, 0, this.#lengthOf(index)));
-      } catch (error) {
-        this.#returnBlock(target);
-        throw error;
-      }
-      this.#registerPiece(index, target);
-      this.#counters.fromDisk += 1;
-      this.#counters.revivals += 1;
-      this.#noteRevival(index);
-      return target;
-    } finally {
-      release();
-    }
-  }
-
-  /**
-   * A block to hold one piece: the most recently freed one, or a new one.
-   *
-   * Every block is a full piece long, whatever piece will live in it. The last
-   * piece of a file is shorter, and it occupies a full block with only its own
-   * bytes meaningful — `#lengthOf` is what decides how much is ever read out.
-   * Uniform blocks are what makes them interchangeable at all.
-   *
-   * Most recently freed first, deliberately: a few blocks then carry the whole
-   * of a busy store's traffic and the rest age out of use, which is what makes
-   * {@link SharedPieceStore#sweepFreeBlocks} able to tell a spare block from a
-   * working one.
-   *
-   * @returns {SharedArrayBuffer}
-   */
-  #takeBlock() {
-    const spare = this.#freeBlocks.pop();
-    if (spare !== undefined) {
-      this.#noteReuseGap(Date.now() - spare.freedAt);
-      return spare.buffer;
-    }
-    // Nothing spare and the pool is already as large as it is allowed to be.
-    // This cannot happen while the accounting is sound: a block is taken only
-    // after a slot has been claimed, and slots are exactly what the ceiling
-    // counts. It is recorded rather than hidden because when it does happen the
-    // pool grows without bound, and every block it then allocates is used once
-    // and thrown to a collector that has no reason to run — the heap stays at
-    // 50 MB of 2240 while the process reaches four gigabytes.
-    // Strictly greater: reaching the ceiling exactly is what a full store looks
-    // like, and a slot has just been claimed for this block. Being ALREADY past
-    // it and allocating anyway is the state that runs away.
-    if (this.#blocksAllocated > this.#growthCeiling) {
-      this.#counters.blocksBeyondCeiling += 1;
-    }
-    this.#blocksAllocated += 1;
-    return this.#watchForCollection(new SharedArrayBuffer(this.#chunkLength));
-  }
-
-  /**
-   * Put a block back for re-use, or give it up.
-   *
-   * Given up when the allowance has fallen below the number of blocks that
-   * exist: keeping it would hold memory the machine has just been shown to
-   * need. Otherwise it waits in the free list for the next piece.
-   *
-   * @param {SharedArrayBuffer | undefined} buffer
-   * @returns {void}
-   */
-  #returnBlock(buffer) {
-    if (buffer === undefined) {
-      return;
-    }
-    if (this.#closed || this.#blocksAllocated > this.#growthCeiling) {
-      // Never below zero: `close` gives up every block at once, and a spill
-      // that was already in flight resolves afterwards and arrives here.
-      if (this.#blocksAllocated > 0) {
-        this.#blocksAllocated -= 1;
-        this.#counters.blocksReleased += 1;
-      }
-      return;
-    }
-    this.#freeBlocks.push({ buffer, freedAt: Date.now() });
-  }
-
-  /**
-   * Record how long a block waited to be used again.
-   *
-   * @param {number} gapMs
-   * @returns {void}
-   */
-  #noteReuseGap(gapMs) {
-    this.#reuseGaps.push(Math.max(0, gapMs));
-    if (this.#reuseGaps.length > REUSE_GAP_SAMPLES) {
-      this.#reuseGaps.shift();
-    }
-  }
-
-  /**
-   * The longest a block has recently waited before being wanted again, or null
-   * when no block has yet been re-used.
-   *
-   * This is the store's own working rhythm, measured rather than chosen: while
-   * a film is being watched a block is taken again within milliseconds, because
-   * one is taken for every piece that arrives. A block that has been sitting
-   * longer than the longest of those waits is not part of the work.
-   *
-   * @returns {number | null}
-   */
-  #reuseGapCeilingMs() {
-    if (this.#reuseGaps.length === 0) {
-      return null;
-    }
-    return Math.max(...this.#reuseGaps);
-  }
-
-  /**
-   * Give up blocks that have sat unused longer than this store's own working
-   * rhythm.
-   *
-   * The case it is for: a torrent whose peers have gone. Its readers are still
-   * attached, so nothing removes the torrent — the pool's idle timer needs a
-   * refcount of zero and never starts. Its allowance falls to what those
-   * readers declared, the pieces beyond it are written out, and their blocks
-   * would otherwise wait in the free list for peers that may not return.
-   *
-   * @param {number} [now]
-   * @returns {number} Blocks given up.
-   */
-  sweepFreeBlocks(now = Date.now()) {
-    const ceiling = this.#reuseGapCeilingMs();
-    if (ceiling === null) {
-      return 0;
-    }
-    const keeping = [];
-    let released = 0;
-    for (const spare of this.#freeBlocks) {
-      if (now - spare.freedAt <= ceiling) {
-        keeping.push(spare);
-        continue;
-      }
-      this.#blocksAllocated -= 1;
-      this.#counters.blocksReleased += 1;
-      released += 1;
-    }
-    this.#freeBlocks = keeping;
-    return released;
-  }
-
-  /**
-   * Count this buffer as one this thread will have to let go of, and notice
-   * when the collector takes it. See {@link pieceBufferCollection}.
-   *
-   * @param {SharedArrayBuffer} buffer
-   * @returns {SharedArrayBuffer} The same buffer.
-   */
-  #watchForCollection(buffer) {
-    released.count += 1;
-    collector?.register(buffer, null);
-    return buffer;
-  }
-
-  /**
-   * A fresh buffer holding this piece's bytes.
-   *
-   * @param {number} index
-   * @param {Uint8Array} bytes
-   * @returns {SharedArrayBuffer}
-   */
-  #copyIntoNewBuffer(index, bytes) {
-    const length = this.#lengthOf(index);
-    const block = this.#takeBlock();
-    const view = Buffer.from(block, 0, length);
-    if (bytes.copy) {
-      bytes.copy(view, 0, 0, length);
-    } else {
-      view.set(bytes.subarray(0, length), 0);
-    }
-    return block;
-  }
-
-  /**
-   * Drop the disk copy of a piece that memory now holds — after any spill of
-   * that same piece has finished.
-   *
-   * `DiskTier.write` records the index when it COMPLETES, so forgetting while a
-   * spill of that index is still running let the completing write put it back,
-   * and a later read then returned the stale bytes.
-   *
-   * @param {number} index
-   * @returns {Promise<void>}
-   */
-  async #forgetOnDisk(index) {
-    const spill = this.#evicting.get(index);
-    if (spill) {
-      await spill.catch(() => undefined);
-    }
-    this.#disk.forget(index);
-    this.#spilledAt.delete(index);
-  }
-
-  put(index, bytes, callback = () => undefined) {
-    if (this.#closed) {
-      queueMicrotask(() => callback(new Error("Piece store is closed.")));
-      return;
-    }
-
-    const write = async () => {
-      // Already resident: the buffer is replaced, not added, so no slot is
-      // needed and none is claimed. A fresh buffer rather than a write into the
-      // old one, so a reader holding the old reference cannot see a torn write.
-      if (this.#buffers.has(index)) {
-        const previous = this.#buffers.get(index);
-        // A fresh block rather than a write into the old one, so a reader
-        // holding the old reference cannot see a torn write.
-        this.#buffers.set(index, this.#copyIntoNewBuffer(index, bytes));
-        this.#lru.touch(index);
-        // And the old block goes back for re-use only if nobody is reading it.
-        // A pinned piece has a view onto its memory somewhere; that block is
-        // given up instead, and the pool allocates another when it needs one.
-        if (this.#lru.isPinned(index)) {
-          if (this.#blocksAllocated > 0) {
-            this.#blocksAllocated -= 1;
-            this.#counters.blocksReleased += 1;
-          }
-        } else {
-          this.#returnBlock(previous);
-        }
-        await this.#forgetOnDisk(index);
-        this.#noteProgress();
-        return;
-      }
-
-      this.#noteArrival();
-      const declared = this.#lru.wants(index);
-      if (declared) {
-        this.#counters.admittedInsideWindow += 1;
-      } else {
-        this.#counters.admittedOutsideWindow += 1;
-      }
-
-      // A piece that will be wanted LATER than the one it would displace goes
-      // straight to disk instead of being admitted.
-      //
-      // Two cases, and the second was missing until 2026-09-02. The first: the
-      // arrival is in nobody's window at all. The second: it IS in somebody's
-      // window, but further from every read head than the piece the store would
-      // have to evict to make room for it — so admitting it would write out the
-      // nearer piece and read it back sooner. The field session had six readers
-      // whose windows covered the whole file, so the first case never applied
-      // and `0 of those went straight to disk` while the store spilled 233
-      // pieces in a minute.
-      //
-      // Only when SOMETHING is declared: before the first read there is no
-      // basis for calling one piece more wanted than another.
-      const worseThanTheVictim = () => {
-        const arriving = this.#lru.waitFor(index);
-        if (arriving < 0) {
-          return false;
-        }
-        const victim = this.#lru.nextVictim();
-        return victim.index !== null && victim.wait >= 0 && arriving > victim.wait;
-      };
-      if (this.#lru.protectedCount > 0 && this.#isFullNow()
-        && (!declared || worseThanTheVictim())) {
-        this.#counters.admittedToDisk += 1;
-        await this.#writeThrough(index, bytes);
-        this.#noteProgress();
-        return;
-      }
-
-      const release = await this.#claimSlot();
-      try {
-        this.#registerPiece(index, this.#copyIntoNewBuffer(index, bytes));
-        await this.#forgetOnDisk(index);
-      } finally {
-        release();
-      }
-    };
-
-    write().then(() => callback(null), (error) => callback(error));
-  }
-
-  get(index, options, callback) {
-    if (typeof options === "function") {
-      return this.get(index, undefined, options);
-    }
-    const done = callback ?? (() => undefined);
-    if (this.#closed) {
-      queueMicrotask(() => done(new Error("Piece store is closed.")));
-      return;
-    }
-
-    const pieceLength = this.#lengthOf(index);
-    const offset = options?.offset ?? 0;
-    const length = options?.length ?? pieceLength - offset;
-
-    const fetch = async () => {
-      const buffer = this.#buffers.get(index);
-      if (buffer !== undefined) {
-        this.#lru.touch(index);
-        this.#counters.fromMemory += 1;
-        return Buffer.from(Buffer.from(buffer, offset, length));
-      }
-
-      const revived = await this.#revive(index);
-      if (revived === null) {
-        throw new Error(`Piece ${index} is not in the store.`);
-      }
-      return Buffer.from(Buffer.from(revived, offset, length));
-    };
-
-    fetch().then((bytes) => done(null, bytes), (error) => done(error));
-  }
-
-  protectRange(readerId, from, to) {
-    this.#lru.protect(readerId, from, to);
-  }
-
-  protectedRanges() {
-    return this.#lru.protectedRanges();
-  }
-
-  releaseProtection(readerId) {
-    this.#lru.unprotect(readerId);
-  }
-
-  warmRange(from, to, limit = Math.max(1, Math.floor(this.#growthCeiling / 4))) {
-    if (this.#closed || !Number.isInteger(from) || !Number.isInteger(to)) {
-      return 0;
-    }
-    let started = 0;
-    for (let index = from; index <= to && started < limit; index += 1) {
-      if (this.#buffers.has(index) || !this.#disk.has(index)) {
-        continue;
-      }
-      started += 1;
-      void this.reside(index).catch(() => undefined);
-    }
-    return started;
-  }
-
-  async reside(index) {
-    if (this.#closed) {
-      throw new Error("Piece store is closed.");
-    }
-
-    const buffer = this.#buffers.get(index);
-    if (buffer !== undefined) {
-      this.#lru.touch(index);
-      this.#counters.fromMemory += 1;
-      return this.locate(index);
-    }
-
-    const revived = await this.#revive(index);
-    if (revived === null) {
-      return null;
-    }
-    return { buffer: revived, offset: 0, length: this.#lengthOf(index) };
-  }
-
-  close(callback = () => undefined) {
-    this.#closed = true;
-    liveStores.delete(this);
-    this.#buffers.clear();
-    this.#spilledAt.clear();
-    this.#counters.blocksReleased += this.#blocksAllocated;
-    this.#blocksAllocated = 0;
-    this.#freeBlocks = [];
-    // Whoever is waiting for a slot is woken and finds the store closed, which
-    // is an error they can report. Left asleep they simply never returned.
-    this.#wake();
-    this.#disk.close().then(() => callback(null), (error) => callback(error));
-  }
-
-  destroy(callback = () => undefined) {
-    this.#closed = true;
-    liveStores.delete(this);
-    this.#buffers.clear();
-    this.#spilledAt.clear();
-    this.#counters.blocksReleased += this.#blocksAllocated;
-    this.#blocksAllocated = 0;
-    this.#freeBlocks = [];
-    this.#wake();
-    this.#disk.destroy().then(() => callback(null), (error) => callback(error));
-  }
-}
+/**
+ * @file Torrent pieces in shared memory, with disk as the second tier.
+ *
+ * Replaces the chunk store WebTorrent would otherwise build for itself. Two
+ * reasons, one forced and one chosen.
+ *
+ * **Forced.** WebTorrent's own piece cache hands out the buffer it keeps using
+ * and slices again on the next read. The torrent client now runs on its own
+ * thread, and moving a piece to the main thread by transferring ownership
+ * detached the cache's memory: proxy 2.9.71-2.9.73 answered every read with an
+ * empty body (`Stream ends prematurely at 0`) and the failure was invisible,
+ * because the error never reached the reader. Owning the memory ourselves
+ * removes the question of whose it was.
+ *
+ * **Chosen.** Memory this side of the thread boundary can be *shared* memory,
+ * which the main thread reads without receiving bytes as a copy — see
+ * {@link SharedPieceStore#locate}. Measured on the field host: a piece copy
+ * costs 3.64 ms, a read from the page cache into a buffer we own 7.63 ms, and
+ * re-downloading a piece from the swarm ~1430 ms. So memory first, disk under
+ * it, and never the swarm twice.
+ *
+ * Per piece allocation: each resident piece owns its own `SharedArrayBuffer`.
+ * Evicting a piece deletes its entry and the memory is reclaimable by GC.
+ * `committed` therefore equals `resident`, not a high-water mark.
+ *
+ * **Room for a piece is an owned reservation, not a shared number.**
+ * `#claimSlot` hands back a release the caller runs in a `finally`; nothing
+ * else touches `#outstandingPieces`. It was a counter incremented in one
+ * function and decremented in another, and every consequence of that was a
+ * defect: a failure between the two lost a slot for the life of the process,
+ * `put`'s error path guessed at the correction and could take back a
+ * reservation belonging to a different claim, and one lost reservation made
+ * the five-second "everything is pinned" error permanently unreachable, so a
+ * read retried every 50 ms for ever without completing or failing. Read out of
+ * the field failure of 2026-08-31 (`research/worker-heap-oom-2026-08-31.md`).
+ *
+ * What this deliberately does NOT do is manage the disk as a cache of its own.
+ * Pieces evicted from memory are written once and read back on demand; the file
+ * is discarded whole when the torrent goes away. libtorrent 2.0 and webtor's
+ * seeder both concluded that a hand-rolled disk cache earns less than it costs,
+ * and nothing here disagrees.
+ */
+
+import os from "node:os";
+import { readFileSync } from "node:fs";
+import { PieceLru } from "./piece-lru.js";
+import { DiskTier } from "./disk-tier.js";
+
+/**
+ * Live stores, so the worker can report on them.
+ *
+ * @type {Set<SharedPieceStore>}
+ */
+const liveStores = new Set();
+
+export function collectStoreStats() {
+  return [...liveStores].map((store) => store.stats());
+}
+
+export function findSharedStore(torrent) {
+  let candidate = torrent?.store;
+  for (let depth = 0; candidate && depth < 8; depth += 1) {
+    if (candidate instanceof SharedPieceStore) {
+      return candidate;
+    }
+    candidate = candidate.store;
+  }
+  return null;
+}
+
+/**
+ * The largest fall in the machine's available memory that this process has
+ * seen and did not cause itself.
+ *
+ * What has to be left for everything else on the machine. It cannot be a share
+ * of what is free — a share is a number chosen out of nothing, which is what
+ * `AVAILABLE_MEMORY_SHARE = 0.25`, `MIN_BUDGET_BYTES` and
+ * `MEMORY_BUDGET_CEILING_BYTES` were until 2026-09-02, all three traceable to
+ * one observation of one host on 2026-08-03. It is measured instead: between
+ * two readings, how much available memory went away beyond what the stores
+ * themselves took. On the field host that quantity is large — while the proxy
+ * held 76-133 MB overnight the machine's available memory fell from 2378 MB to
+ * 306 MB and came back — and on a quiet host it stays near zero, which is the
+ * right answer there.
+ *
+ * Starts at zero: nothing is reserved until somebody else has been seen to
+ * need it.
+ */
+const otherDemand = { falls: [], lastAvailableBytes: 0, lastStoreBytes: 0 };
+
+/**
+ * How many observations of other processes' demand are kept.
+ *
+ * A window rather than a high-water, and for the reason the block re-use gap is
+ * one too: a single spike would otherwise stand for the life of the process and
+ * squeeze the stores against something that happened once, hours ago.
+ */
+const OTHER_DEMAND_SAMPLES = 60;
+
+/**
+ * Note what the machine had, and how much of the change was not ours.
+ *
+ * @param {number} availableBytes
+ * @param {number} storeBytes - What the live stores hold right now.
+ * @returns {number} The reserve, in bytes.
+ */
+export function noteMachineMemory(availableBytes, storeBytes) {
+  if (otherDemand.lastAvailableBytes > 0) {
+    const fell = otherDemand.lastAvailableBytes - availableBytes;
+    const ours = storeBytes - otherDemand.lastStoreBytes;
+    otherDemand.falls.push(Math.max(0, fell - ours));
+    if (otherDemand.falls.length > OTHER_DEMAND_SAMPLES) {
+      otherDemand.falls.shift();
+    }
+  }
+  otherDemand.lastAvailableBytes = availableBytes;
+  otherDemand.lastStoreBytes = storeBytes;
+  return machineReserveBytes();
+}
+
+/** What has recently been observed to be needed by everything that is not us. */
+export function machineReserveBytes() {
+  return otherDemand.falls.length === 0 ? 0 : Math.max(...otherDemand.falls);
+}
+
+/** Forget what other processes have needed. For tests, which share a module. */
+export function forgetMachineMemory() {
+  otherDemand.falls = [];
+  otherDemand.lastAvailableBytes = 0;
+  otherDemand.lastStoreBytes = 0;
+}
+
+/**
+ * How much memory the stores may hold between them.
+ *
+ * `MemAvailable` is what could be allocated on top of what is already held, so
+ * the stores' own bytes are added back: the pair is the ceiling the stores
+ * could reach. The reserve is what has been seen to be needed elsewhere.
+ *
+ * @param {number} availableBytes
+ * @param {number} storeBytes
+ * @param {number} reserveBytes
+ * @returns {number}
+ */
+export function machineAllowanceBytes(availableBytes, storeBytes, reserveBytes) {
+  return Math.max(0, Math.max(availableBytes, 0) + Math.max(storeBytes, 0) - Math.max(reserveBytes, 0));
+}
+
+/**
+ * Divide what the machine allows between the stores, by what each is asking
+ * for.
+ *
+ * A store asks for the pieces its readers have declared. When everyone's ask
+ * fits, everyone gets it and the machine's limit never binds — which is the
+ * usual case, since two readers of one film declare 32-192 MB against gigabytes
+ * of free memory. When the asks do not fit, each store is cut in proportion to
+ * what it asked, so a store wanting little is not cut to make room for one
+ * wanting much.
+ *
+ * @param {number[]} wantedBytes - What each store is asking for, in order.
+ * @param {number} allowanceBytes
+ * @returns {number[]} What each store may hold, in the same order.
+ */
+export function divideAllowance(wantedBytes, allowanceBytes) {
+  const total = wantedBytes.reduce((sum, want) => sum + Math.max(0, want), 0);
+  if (total <= allowanceBytes || total === 0) {
+    return wantedBytes.map((want) => Math.max(0, want));
+  }
+  return wantedBytes.map((want) => Math.floor(allowanceBytes * (Math.max(0, want) / total)));
+}
+
+export function reviseStoreBudgets() {
+  const stores = [...liveStores];
+  const held = stores.reduce((sum, store) => sum + store.residentBytes, 0);
+  const available = availableMemorySync();
+  const reserve = noteMachineMemory(available, held);
+  const allowance = machineAllowanceBytes(available, held, reserve);
+  const shares = divideAllowance(stores.map((store) => store.wantedBytes), allowance);
+  const revised = [];
+  for (const [position, store] of stores.entries()) {
+    revised.push(store.reviseGrowthCeiling(shares[position]));
+  }
+  return revised;
+}
+
+function defaultMemoryBytes() {
+  const stores = [...liveStores];
+  const held = stores.reduce((sum, store) => sum + store.residentBytes, 0);
+  const available = availableMemorySync();
+  const allowance = machineAllowanceBytes(available, held, machineReserveBytes());
+  // A store being created has no readers, so it has no demand to state and no
+  // basis for asking for more or less than the others. It opens on an equal
+  // share and the first revision — within a minute, and within seconds of a
+  // read starting — replaces that with what its readers actually declare.
+  // Deliberately not the whole allowance: on a machine with gigabytes free that
+  // would let a torrent nobody is reading yet fill memory before the first
+  // revision arrives.
+  return Math.floor(allowance / (stores.length + 1));
+}
+
+function availableMemorySync() {
+  try {
+    const text = readFileSync("/proc/meminfo", "utf8");
+    const match = /^MemAvailable:\s+(\d+)\s+kB$/m.exec(text);
+    if (match) {
+      return Number(match[1]) * 1024;
+    }
+  } catch {
+  }
+  return os.freemem();
+}
+
+/**
+ * How many blocks of piece memory this thread has let go of, and how many the
+ * collector has actually taken back.
+ *
+ * Since the store keeps a pool, one block serves many pieces, so this counts
+ * blocks and not pieces — and a healthy store allocates only as many as its
+ * allowance permits, so both numbers are now small and nearly equal.
+ *
+ * The one reading that separates the two explanations of the 700 MB nobody can
+ * account for (roadmap item 2, field 2026-08-31): if the two numbers track each
+ * other, this code is holding nothing and whatever grows is below us, in the
+ * allocator — on musl there is no `malloc_trim` and no way to ask. If the gap
+ * widens, a reference of ours outlives the piece, and then a heap snapshot can
+ * name the holder.
+ *
+ * What it does NOT prove: a `SharedArrayBuffer`'s memory is shared between the
+ * isolates, so this thread's handle going means only that THIS thread let go.
+ * The main thread counts its own (`torrent-worker/client.js`), and the pair is
+ * what answers the question.
+ */
+const released = { count: 0, collected: 0 };
+const collector = typeof FinalizationRegistry === "function"
+  ? new FinalizationRegistry(() => {
+    released.collected += 1;
+  })
+  : null;
+
+/**
+ * What the collector has taken back against what was let go.
+ *
+ * @returns {{ released: number, collected: number }}
+ */
+export function pieceBufferCollection() {
+  return { released: released.count, collected: released.collected };
+}
+
+const MIN_RESIDENT_PIECES = 2;
+/**
+ * How long a claim may go without ANYTHING moving before it gives up.
+ *
+ * Measured against progress, not against activity. The earlier rule skipped
+ * this timer entirely while a spill was in flight or a reservation was held —
+ * so one reservation that was never returned made the timer unreachable and a
+ * read retried every 50 ms for the life of the process, never completing and
+ * never failing (field 2026-08-31).
+ */
+const PINNED_WAIT_MS = 5_000;
+
+/**
+ * How many revival ages are kept for the median. A window, not a history: two
+ * hundred covers several minutes of the busiest session measured (7575
+ * revivals in 44 minutes) and costs two hundred numbers.
+ */
+const REVIVAL_AGE_SAMPLES = 200;
+
+/**
+ * How many block re-use gaps are kept. The same window as the revival ages, and
+ * for the same reason: what is wanted is the rhythm of recent work, not a
+ * history of it.
+ */
+const REUSE_GAP_SAMPLES = 200;
+
+/** How many write durations are kept for the median. */
+const WRITE_DURATION_SAMPLES = 50;
+/** How many admission times are kept, and how far back they are counted. */
+const ARRIVAL_SAMPLES = 100;
+const ARRIVAL_WINDOW_MS = 10_000;
+
+/**
+ * The middle value of a sample, or null when there is nothing to take a middle
+ * of. Null rather than zero: no revivals and instant revivals are different
+ * facts and must not print the same.
+ *
+ * @param {number[]} values
+ * @returns {number | null}
+ */
+function median(values) {
+  if (values.length === 0) {
+    return null;
+  }
+  const sorted = [...values].sort((left, right) => left - right);
+  const middle = Math.floor(sorted.length / 2);
+  return sorted.length % 2 === 0
+    ? Math.round((sorted[middle - 1] + sorted[middle]) / 2)
+    : sorted[middle];
+}
+const CLAIM_RETRY_MS = 50;
+
+export class SharedPieceStore {
+  #chunkLength;
+  #lastChunkLength;
+  #lastChunkIndex;
+  #growthCeiling;
+  /** Piece index → SharedArrayBuffer of that piece */
+  #buffers = new Map();
+  /** @type {Map<number, Promise<void>>} */
+  #evicting = new Map();
+  /**
+   * Slots claimed but not yet filled.
+   *
+   * Handed out by {@link SharedPieceStore##claimSlot} as a release function the
+   * caller must call in a `finally`, never as a number one function increments
+   * and another decrements. The counter used to be paired across
+   * `#claimSlot`/`#registerPiece`, so any failure between the two lost a slot
+   * for the life of the process, and `put`'s error path tried to correct that
+   * by guessing — which could take back a reservation belonging to a different
+   * claim and let the store admit past its allowance.
+   */
+  #outstandingPieces = 0;
+  /** When something last actually moved: a piece admitted, spilled or unpinned. */
+  #lastProgressAt = 0;
+  #waiters = [];
+  #lru;
+  #disk;
+  #closed = false;
+  #name;
+  #counters = {
+    fromMemory: 0,
+    fromDisk: 0,
+    spills: 0,
+    revivals: 0,
+    blockedByPins: 0,
+    waitedForPins: 0,
+    evictedOnRevise: 0,
+    spillFailures: 0,
+    // Whether the store is doing its job or being asked to hold more than it
+    // has room for. An eviction that had to take a piece a reader declared it
+    // wants is the second, and it comes back from disk moments later
+    // (roadmap item 9).
+    evictedProtected: 0,
+    evictedDistanceSum: 0,
+    evictedWithDistance: 0,
+    // Where an arriving piece went. A piece nobody has declared is being
+    // downloaded ahead of every reader; putting it in memory pushes out one
+    // that IS declared, which is then read back from disk moments later
+    // (roadmap item 9).
+    admittedInsideWindow: 0,
+    admittedOutsideWindow: 0,
+    admittedToDisk: 0,
+    /** Blocks given back to the operating system. */
+    blocksReleased: 0,
+    /**
+     * A block put back for re-use while a reader still held the piece. Zero by
+     * construction — a pinned piece is never evicted, and a re-put of a pinned
+     * one drops its block instead of recycling it. Counted because if it ever
+     * stops being zero, a consumer is reading another piece's bytes, and that
+     * is invisible from anywhere else.
+     */
+    returnedWhilePinned: 0,
+    /** Spills that found the disk already holding identical bytes. */
+    spillsSkipped: 0,
+    /**
+     * Blocks a second registration of the same piece displaced. Expected to be
+     * small and non-zero: two callers racing for one piece is ordinary. What is
+     * NOT ordinary is the block going missing when it happens, which is what
+     * killed the process on 2026-09-02.
+     */
+    blocksDisplaced: 0,
+    /** Admissions that waited for the disk instead of evicting another piece. */
+    waitedForDisk: 0,
+    /**
+     * Admissions that grew memory because the disk had stopped answering.
+     * Non-zero means the store exceeded its allowance on purpose, and by how
+     * many pieces.
+     */
+    grewWaitingForDisk: 0,
+    /**
+     * Times a block was wanted, none was free, and the pool was already at its
+     * ceiling. Zero by construction — a block is only taken after a slot has
+     * been claimed, and slots are what the ceiling counts — so a number here
+     * means the two have come apart and the pool is growing past its allowance.
+     */
+    blocksBeyondCeiling: 0
+  };
+  /**
+   * Blocks that hold no piece, most recently freed last.
+   *
+   * A block is one piece's worth of memory. Allocating a new one for every
+   * piece meant 7575 allocations of 4 MiB in 44 minutes on 2026-09-02, each
+   * released only when the collector got to it — which is why the process held
+   * 1.86 GB while its own accounting said 352 MB. Blocks are taken from here
+   * and put back here instead (roadmap item 2).
+   *
+   * @type {{ buffer: SharedArrayBuffer, freedAt: number }[]}
+   */
+  #freeBlocks = [];
+  /** Blocks that exist at all: free plus holding a piece. */
+  #blocksAllocated = 0;
+  /** Whether a reader has ever declared a window here. See `wantedBytes`. */
+  #everHadReader = false;
+  /** Whether the last revision had to exceed the machine's share. */
+  #beyondTheMachine = false;
+  /**
+   * How long a block sat free before it was taken again, in milliseconds.
+   * Bounded, because what is wanted is the longest gap of RECENT work: an
+   * all-time maximum would be raised by one long pause and then never let a
+   * block go again.
+   *
+   * @type {number[]}
+   */
+  #reuseGaps = [];
+  /**
+   * How long recent writes to disk took, in milliseconds, and when pieces were
+   * admitted. Together they say how much room the store needs beyond what the
+   * readers ask for: while one write is finishing, more pieces arrive, and each
+   * needs somewhere to go. Without that room every arrival evicts something,
+   * which is what produced 233 evictions against 119 completed writes in a
+   * minute on 2026-09-02.
+   *
+   * @type {number[]}
+   */
+  #writeDurations = [];
+  /** When the last few pieces were admitted, for the arrival rate. */
+  #admittedAt = [];
+  /** Piece index → when it was written out, for the age it comes back at. */
+  #spilledAt = new Map();
+  /**
+   * Ages, in milliseconds, of the last revivals — bounded, because the figure
+   * wanted is a median and not a history. A piece that comes back seconds
+   * after it left should not have left.
+   *
+   * @type {number[]}
+   */
+  #revivalAges = [];
+
+  constructor(chunkLength, options = {}) {
+    if (!Number.isInteger(chunkLength) || chunkLength < 1) {
+      throw new Error(`Chunk length must be a positive integer, got ${chunkLength}.`);
+    }
+    this.#chunkLength = chunkLength;
+
+    const totalLength = Number.isFinite(options.length) ? options.length : 0;
+    this.#lastChunkIndex = totalLength > 0 ? Math.ceil(totalLength / chunkLength) - 1 : -1;
+    const remainder = totalLength % chunkLength;
+    this.#lastChunkLength = remainder === 0 ? chunkLength : remainder;
+
+    const memoryBytes = Number.isFinite(options.memoryBytes) && options.memoryBytes > 0
+      ? options.memoryBytes
+      : defaultMemoryBytes();
+    // Only where this store STARTS. It is not kept, so it cannot come back as a
+    // cap on the revision the way `#capacity` did.
+    const openingCeiling = Math.max(MIN_RESIDENT_PIECES, Math.floor(memoryBytes / chunkLength));
+    this.#growthCeiling = openingCeiling;
+    this.#lru = new PieceLru(openingCeiling);
+    this.#name = options.name ?? "pieces";
+    // `options.disk` exists so a test can hold a write open or make one fail on
+    // purpose. Four of the defects fixed here live in what happens when the
+    // disk tier does not answer immediately or at all, and none of them is
+    // reachable from outside without saying so.
+    this.#disk = options.disk ?? new DiskTier({
+      directory: options.path ?? ".",
+      name: `${this.#name}.pieces`,
+      chunkLength
+    });
+    liveStores.add(this);
+  }
+
+  stats() {
+    const resident = this.#buffers.size;
+    const residentBytes = resident * this.#chunkLength;
+    // Last piece may be short but stats historically use chunkLength.
+    return {
+      name: this.#name,
+      resident,
+      capacity: this.#growthCeiling,
+      residentBytes,
+      allocatedSlots: this.#blocksAllocated,
+      // What the process HOLDS, which with a pool is the blocks that exist —
+      // not the pieces in them. Holding and using are different quantities and
+      // the difference is the point of the reading.
+      committedBytes: this.#blocksAllocated * this.#chunkLength,
+      budgetBytes: this.#growthCeiling * this.#chunkLength,
+      pinned: this.#lru.pinnedCount,
+      // Slots claimed and not yet filled. Reported because a reservation that
+      // is never returned is invisible until the store cannot admit anything,
+      // and by then the reason is long gone. At rest this is zero.
+      outstanding: this.#outstandingPieces,
+      spilled: this.#disk.size,
+      spilledBytes: this.#disk.size * this.#chunkLength,
+      // What the readers between them are asking this store to keep, against
+      // what it may hold. A union wider than the capacity cannot be held
+      // however the eviction is ordered, and that is the difference between a
+      // policy to fix and arithmetic to accept (roadmap item 9).
+      demand: this.#lru.demand(),
+      // What the process actually holds for this store, which is not the same
+      // as what it is using: blocks are kept for re-use. The difference is the
+      // spare, and it is the whole of the question whether consumption only
+      // grows (roadmap item 2).
+      blocksAllocated: this.#blocksAllocated,
+      blocksFree: this.#freeBlocks.length,
+      blocksDisplaced: this.#counters.blocksDisplaced,
+      blocksInUse: this.#blocksInUse(),
+      blocksInFlight: this.#blocksInFlight(),
+      blocksBeyondCeiling: this.#counters.blocksBeyondCeiling,
+      blockBytes: this.#blocksAllocated * this.#chunkLength,
+      reuseGapMs: this.#reuseGapCeilingMs(),
+      revivalAgeMedianMs: median(this.#revivalAges),
+      revivalAgeSamples: this.#revivalAges.length,
+      revivedWithinFiveSeconds: this.#revivalAges.filter((age) => age <= 5_000).length,
+      ...this.#counters
+    };
+  }
+
+  /**
+   * Whether the machine's share of memory is smaller than one reader's window.
+   *
+   * The store holds the window anyway — refusing would leave the read it is
+   * serving unable to finish, which is worse — but it is the honest measure of
+   * "this machine cannot take any more", and it is measured rather than
+   * guessed: it is the last revision's own comparison of what the machine
+   * allowed against what the widest reader declared.
+   */
+  get isBeyondTheMachine() {
+    return this.#beyondTheMachine;
+  }
+
+  /** What this store holds right now, in bytes. */
+  get residentBytes() {
+    return this.#buffers.size * this.#chunkLength;
+  }
+
+  /**
+   * What this store is asking to be allowed to hold, in bytes.
+   *
+   * The union of its readers' declared windows — what they have said they will
+   * need — never below the widest single window, because a store that cannot
+   * hold one reader's window cannot complete that reader's read at all: every
+   * resident piece ends up pinned, the read returns zero bytes and ffmpeg takes
+   * that for the end of the file (field 2026-08-15, roadmap item 9).
+   *
+   * With no reader declaring anything there is no demand to speak of, so the
+   * store asks for what the machine allows and the first revision after a read
+   * begins brings it down to what that read needs.
+   */
+  get wantedBytes() {
+    const demand = this.#lru.demand();
+    if (demand.readers > 0) {
+      this.#everHadReader = true;
+      const pieces = Math.max(MIN_RESIDENT_PIECES, demand.unionPieces, demand.widestPieces);
+      // Plus room to absorb what arrives while a write is finishing. Asking for
+      // exactly what the readers want leaves no free place ever, so every
+      // arrival evicts one of them — measured 2026-09-02: `6 reader(s) want 23
+      // piece(s) of 23 the store may hold`, and 233 evictions in the minute
+      // that followed.
+      return (pieces + this.slackPieces()) * this.#chunkLength;
+    }
+    // Readers that have GONE are not the same as readers that have not arrived.
+    // A store whose readers ended has nothing to hold pieces for — its torrent
+    // sits until the pool's idle timer removes it, which needs a refcount of
+    // zero and can be a quarter of an hour away — so it asks for nothing and
+    // its memory goes back to the machine now. A store that has never had a
+    // reader is being filled for one that is on its way, and asks for what it
+    // was opened with until the first read says what it needs.
+    return this.#everHadReader
+      ? (MIN_RESIDENT_PIECES + this.slackPieces()) * this.#chunkLength
+      : this.#growthCeiling * this.#chunkLength;
+  }
+
+  reviseGrowthCeiling(allowedBytes) {
+    // No cap at what the machine could spare when this store was CREATED.
+    // `#capacity` was computed once in the constructor and used as an upper
+    // bound here, so the allowance could only ever fall: a torrent opened while
+    // the machine was full kept a small allowance for its whole life, however
+    // much memory was freed afterwards (roadmap item 2, 2026-09-02).
+    const wanted = Math.floor(Number(allowedBytes) / this.#chunkLength);
+    // Never below one reader's whole window while a reader exists, even when
+    // the machine's share says less. A store that cannot hold the window of the
+    // read it is serving cannot complete that read at all: every resident piece
+    // ends up pinned, the read returns zero bytes and ffmpeg takes that for the
+    // end of the file, which killed every encoder on that file in the field on
+    // 2026-08-15. Exceeding the share is the lesser failure, and the line below
+    // says when it happens.
+    const demand = this.#lru.demand();
+    this.#growthCeiling = Math.max(
+      MIN_RESIDENT_PIECES,
+      demand.readers > 0 ? demand.widestPieces : MIN_RESIDENT_PIECES,
+      Number.isFinite(wanted) ? wanted : MIN_RESIDENT_PIECES
+    );
+    const belowAWindow = demand.readers > 0
+      && Number.isFinite(wanted)
+      && wanted < demand.widestPieces;
+    this.#beyondTheMachine = belowAWindow;
+    // The LRU is told too. It was constructed with the store's original
+    // capacity and never revised, so `isFull()` answered against a number that
+    // had not been the limit for some time — dormant only because nothing calls
+    // it, which is a trap for whoever calls it next.
+    this.#lru.setCapacity(this.#growthCeiling);
+    // With per-piece buffers memory CAN be given back immediately, unlike the
+    // old growable pool. Eagerly evict excess to honour the new ceiling.
+    let evicted = 0;
+    while (this.#buffers.size > this.#growthCeiling) {
+      const { index: victim, protectionYielded, distance } = this.#lru.evictionChoice();
+      if (victim === null) {
+        break;
+      }
+      const victimBuffer = this.#buffers.get(victim);
+      if (victimBuffer === undefined) {
+        this.#lru.remove(victim);
+        continue;
+      }
+      this.#buffers.delete(victim);
+      this.#lru.remove(victim);
+      evicted += 1;
+      this.#counters.evictedOnRevise += 1;
+      this.#noteEviction(protectionYielded, distance);
+      // Nobody awaits this spill, so its failure has to end here. Rethrowing
+      // made it an unhandled rejection, and an unhandled rejection in the
+      // torrent worker ends the thread — a second way to lose the torrent
+      // client, on top of the one that already loses it.
+      void this.#spill(victim, victimBuffer).catch(() => {
+        this.#counters.spillFailures += 1;
+      });
+    }
+    // Blocks the store is no longer using and has waited long enough to give
+    // up. The allowance falling is the moment to ask, because that is when the
+    // machine has been shown to need the memory.
+    const releasedBlocks = this.sweepFreeBlocks();
+    return {
+      name: this.#name,
+      ceilingBytes: this.#growthCeiling * this.#chunkLength,
+      committedBytes: this.#blocksAllocated * this.#chunkLength,
+      evicted,
+      releasedBlocks,
+      belowAWindow
+    };
+  }
+
+  get chunkLength() {
+    return this.#chunkLength;
+  }
+
+  get capacity() {
+    return this.#growthCeiling;
+  }
+
+  get residentCount() {
+    return this.#buffers.size;
+  }
+
+  get spilledCount() {
+    return this.#disk.size;
+  }
+
+  #lengthOf(index) {
+    return index === this.#lastChunkIndex ? this.#lastChunkLength : this.#chunkLength;
+  }
+
+  /**
+   * Where a resident piece sits, or `null` if not resident.
+   * Returns the piece's own SharedArrayBuffer and the intra-piece range.
+   * @param {number} index
+   * @returns {{ buffer: SharedArrayBuffer, offset: number, length: number } | null}
+   */
+  locate(index) {
+    const buffer = this.#buffers.get(index);
+    if (buffer === undefined) {
+      return null;
+    }
+    return { buffer, offset: 0, length: this.#lengthOf(index) };
+  }
+
+  /**
+   * Direct access to a piece's buffer for zero-copy consumers.
+   * @param {number} index
+   * @returns {SharedArrayBuffer | undefined}
+   */
+  getPieceBuffer(index) {
+    return this.#buffers.get(index);
+  }
+
+  pin(index) {
+    this.#lru.pin(index);
+  }
+
+  unpin(index) {
+    this.#lru.unpin(index);
+    this.#noteProgress();
+  }
+
+  /**
+   * Reserve room for one piece.
+   *
+   * @returns {Promise<() => void>} The release, which the caller MUST call in a
+   *   `finally`. Calling it twice is harmless.
+   */
+  async #claimSlot() {
+    // How long THIS claim has been trying, kept here and not on the store.
+    // It was a field, `#pinnedWaitStartedAt`, shared by the two waits inside
+    // `#claimSlotOnce` — the one for the disk and the one for a piece that may
+    // be evicted — and each of them zeroed it on giving up, which restarted the
+    // other's clock. Measured 2026-09-03: the claim cycled for ever, five
+    // seconds per side, `grewWaitingForDisk` and `waitedForPins` both climbing
+    // once every five seconds while `blockedByPins` stayed at 0 — the refusal
+    // was unreachable. A claim's patience belongs to the claim; with several
+    // claimants a shared field is wrong anyway, since one caller giving up
+    // would reset the wait of every other.
+    const waitingSince = Date.now();
+    for (;;) {
+      if (this.#closed) {
+        throw new Error("Piece store is closed.");
+      }
+      const ok = await this.#claimSlotOnce(waitingSince);
+      if (ok) {
+        let released = false;
+        return () => {
+          if (released) {
+            return;
+          }
+          released = true;
+          this.#outstandingPieces -= 1;
+          this.#noteProgress();
+        };
+      }
+      await this.#waitForSlot();
+    }
+  }
+
+  /**
+   * Sleep until something moves, or until the retry interval, whichever first.
+   *
+   * One handler, idempotent, and its timer is cleared when it is woken. The
+   * earlier version attached a fresh pair of handlers to EVERY pending spill on
+   * every attempt and left a timer running each time, so a claim that could not
+   * be satisfied allocated in proportion to attempts times pending spills. A
+   * settling spill now wakes the store itself, which is where that belongs.
+   *
+   * @returns {Promise<void>}
+   */
+  #waitForSlot() {
+    return new Promise((resolve) => {
+      let settled = false;
+      /** @type {ReturnType<typeof setTimeout> | null} */
+      let retry = null;
+      const finish = () => {
+        if (settled) {
+          return;
+        }
+        settled = true;
+        if (retry !== null) {
+          clearTimeout(retry);
+        }
+        resolve();
+      };
+      this.#waiters.push(finish);
+      retry = setTimeout(finish, CLAIM_RETRY_MS);
+      retry.unref?.();
+    });
+  }
+
+  #registerPiece(index, buffer) {
+    // A piece may already be here. Both paths that register one look first and
+    // then await — `put` waits for a slot, `#revive` waits for the disk — and
+    // in that gap another caller can register the same index. Overwriting the
+    // entry used to drop the previous block on the floor: its memory was not
+    // returned to the pool and the pool's own count was never decremented, so
+    // the count climbed past the ceiling for ever and the pool degenerated into
+    // allocating a fresh block per piece. Field 2026-09-02: a store holding
+    // THREE pieces reported 812 MB committed against 68 MB allowed, 739 blocks
+    // allocated and 676 still alive, and the process was killed at 4.37 GB.
+    const displaced = this.#buffers.get(index);
+    this.#buffers.set(index, buffer);
+    if (displaced !== undefined && displaced !== buffer) {
+      this.#counters.blocksDisplaced += 1;
+      this.#returnBlock(displaced);
+    }
+    this.#lru.touch(index);
+    this.#noteProgress();
+  }
+
+  /**
+   * Write a piece out and account for it, from the one place that does it.
+   *
+   * @param {number} index
+   * @param {SharedArrayBuffer} buffer
+   * @returns {Promise<void>}
+   */
+  #spill(index, buffer) {
+    // The disk may already hold these very bytes. `#revive` reads a piece back
+    // into memory and leaves the copy on disk, and only `put` removes it — so a
+    // piece that was revived and not re-put is identical to what is already
+    // written, and writing it again is work for nothing. There were 7575
+    // revivals in one session on 2026-09-02, and every later eviction of one of
+    // them wrote a second time (roadmap item 66: 14.4 GB written in a single
+    // viewing).
+    if (this.#disk.has(index)) {
+      this.#counters.spills += 1;
+      this.#counters.spillsSkipped += 1;
+      this.#spilledAt.set(index, Date.now());
+      this.#returnBlock(buffer);
+      this.#noteProgress();
+      return Promise.resolve();
+    }
+
+    const bytes = Buffer.from(buffer, 0, this.#lengthOf(index));
+    const startedAt = Date.now();
+    const spill = this.#disk.write(index, bytes).then(
+      () => {
+        this.#noteWriteDuration(Date.now() - startedAt);
+        this.#counters.spills += 1;
+        this.#spilledAt.set(index, Date.now());
+        this.#evicting.delete(index);
+        // Only now. The write reads out of this block, so a block handed to
+        // another piece before the write finished would put that piece's bytes
+        // into this piece's place in the file.
+        this.#returnBlock(buffer);
+        this.#noteProgress();
+      },
+      (error) => {
+        this.#evicting.delete(index);
+        // The block is no longer holding anything either way; keeping it out of
+        // the pool because the write failed would lose it for good.
+        this.#returnBlock(buffer);
+        this.#noteProgress();
+        throw error;
+      }
+    );
+    this.#evicting.set(index, spill);
+    return spill;
+  }
+
+  /** Something actually moved: wake whoever is waiting and restart the clock. */
+  #noteProgress() {
+    this.#lastProgressAt = Date.now();
+    this.#wake();
+  }
+
+  /**
+   * Record how long one write to disk took.
+   *
+   * @param {number} durationMs
+   * @returns {void}
+   */
+  #noteWriteDuration(durationMs) {
+    this.#writeDurations.push(Math.max(0, durationMs));
+    if (this.#writeDurations.length > WRITE_DURATION_SAMPLES) {
+      this.#writeDurations.shift();
+    }
+  }
+
+  /** Record that a piece arrived, for the arrival rate. */
+  #noteArrival() {
+    const now = Date.now();
+    this.#admittedAt.push(now);
+    while (this.#admittedAt.length > ARRIVAL_SAMPLES
+      || (this.#admittedAt.length > 0 && now - this.#admittedAt[0] > ARRIVAL_WINDOW_MS)) {
+      this.#admittedAt.shift();
+    }
+  }
+
+  /**
+   * How many pieces the store needs room for beyond what the readers ask for.
+   *
+   * Measured, not chosen: the pieces that arrive while one write to disk is
+   * finishing. Without this room every arrival must evict something, and each
+   * eviction holds its block until its write completes — so a disk slower than
+   * the swarm turns every admission into one more block held. On 2026-09-02
+   * that was 233 evictions against 119 completed writes in a minute, and 203
+   * blocks held with three pieces resident.
+   *
+   * Zero until both quantities have been seen, because a slack invented before
+   * anything is measured is a chosen number, and this store has been bitten by
+   * those.
+   *
+   * @returns {number} Pieces.
+   */
+  slackPieces() {
+    if (this.#writeDurations.length === 0 || this.#admittedAt.length < 2) {
+      return 0;
+    }
+    const sorted = [...this.#writeDurations].sort((left, right) => left - right);
+    const writeMs = sorted[Math.floor(sorted.length / 2)];
+    const spanMs = this.#admittedAt[this.#admittedAt.length - 1] - this.#admittedAt[0];
+    if (!(spanMs > 0)) {
+      return 0;
+    }
+    const perMs = (this.#admittedAt.length - 1) / spanMs;
+    return Math.ceil(perMs * writeMs);
+  }
+
+  /**
+   * Blocks that are not free: resident pieces, blocks being written out, and
+   * blocks taken but not yet registered.
+   *
+   * This is what the ceiling has to bound, and until 2026-09-02 it bounded
+   * resident pieces instead. The difference is exactly the memory that goes
+   * missing from the count: a piece being spilled leaves `#buffers` the moment
+   * the eviction begins, while its block stays held until the write it feeds
+   * has finished.
+   *
+   * @returns {number}
+   */
+  #blocksInUse() {
+    return this.#blocksAllocated - this.#freeBlocks.length;
+  }
+
+  /**
+   * Blocks held by writes that have not finished.
+   *
+   * @returns {number}
+   */
+  #blocksInFlight() {
+    return Math.max(0, this.#blocksInUse() - this.#buffers.size);
+  }
+
+  /**
+   * Whether admitting one more piece would need something evicted first.
+   *
+   * Reservations count: a slot claimed and not yet filled is as taken as a
+   * resident piece.
+   *
+   * @returns {boolean}
+   */
+  #isFullNow() {
+    return this.#blocksInUse() + this.#outstandingPieces >= this.#growthCeiling;
+  }
+
+  /**
+   * Write an arriving piece straight to disk, without it ever occupying memory.
+   *
+   * Registered in `#evicting` like a spill, so a `get` for this index waits for
+   * the write instead of finding the piece on neither tier. Deliberately NOT
+   * recorded in `#spilledAt`: that clock measures how long an EVICTED piece
+   * stayed away, and a piece that was never resident has no such age.
+   *
+   * @param {number} index
+   * @param {Uint8Array} bytes
+   * @returns {Promise<void>}
+   */
+  #writeThrough(index, bytes) {
+    // Copied, not viewed. `DiskTier.write` opens the file before it reads the
+    // bytes, so a view onto the caller's buffer could be written to in between
+    // and the file would get the wrong data. The spill path may pass a view
+    // because that memory is ours; this buffer belongs to the torrent client.
+    // It costs nothing extra: the piece was being copied into a shared buffer
+    // on this path before, and now it is copied here instead.
+    const copy = Buffer.from(bytes.subarray(0, Math.min(bytes.length, this.#lengthOf(index))));
+    const write = this.#disk.write(index, copy).then(
+      () => {
+        this.#evicting.delete(index);
+        this.#noteProgress();
+      },
+      (error) => {
+        this.#evicting.delete(index);
+        this.#noteProgress();
+        throw error;
+      }
+    );
+    this.#evicting.set(index, write);
+    return write;
+  }
+
+  /**
+   * Record how long a piece stayed on disk before it was wanted again.
+   *
+   * A piece that comes back seconds after it left was evicted from a working
+   * set that does not fit, and the write and the read were both waste. Kept as
+   * a bounded window of ages because the figure wanted is a median, not a
+   * history.
+   *
+   * @param {number} index
+   * @returns {void}
+   */
+  #noteRevival(index) {
+    const spilledAt = this.#spilledAt.get(index);
+    if (spilledAt === undefined) {
+      return;
+    }
+    this.#spilledAt.delete(index);
+    this.#revivalAges.push(Date.now() - spilledAt);
+    if (this.#revivalAges.length > REVIVAL_AGE_SAMPLES) {
+      this.#revivalAges.shift();
+    }
+  }
+
+  /**
+   * Record what an eviction had to take.
+   *
+   * @param {boolean} protectionYielded - The victim was inside a window a
+   *   reader had declared, and was taken anyway because nothing else was free.
+   * @param {number} distance - Pieces from the nearest declared window, -1 when
+   *   no reader declared one.
+   * @returns {void}
+   */
+  #noteEviction(protectionYielded, distance) {
+    if (protectionYielded) {
+      this.#counters.evictedProtected += 1;
+    }
+    if (distance >= 0) {
+      this.#counters.evictedDistanceSum += distance;
+      this.#counters.evictedWithDistance += 1;
+    }
+  }
+
+  #wake() {
+    const waiting = this.#waiters;
+    this.#waiters = [];
+    for (const resolve of waiting) {
+      resolve();
+    }
+  }
+
+  /**
+   * One attempt at a reservation.
+   *
+   * @param {number} waitingSince - When the claim this attempt belongs to began
+   *   trying. Both waits below are measured from the later of it and the last
+   *   time the store moved, so neither can restart the other's clock — see
+   *   `#claimSlot`.
+   * @returns {Promise<boolean>} Whether a slot was reserved.
+   */
+  async #claimSlotOnce(waitingSince) {
+    // Reserve before suspension so concurrent callers see the reservation.
+    if (this.#blocksInUse() + this.#outstandingPieces < this.#growthCeiling) {
+      this.#outstandingPieces += 1;
+      return true;
+    }
+    const stillFor = () => Date.now() - Math.max(waitingSince, this.#lastProgressAt);
+
+    // Full, and evicting would make it worse rather than better. A piece being
+    // written out has already left `#buffers` while its block is still held —
+    // the write reads from that block — so evicting another one converts a
+    // resident block into an in-flight block and takes a fresh block for the
+    // arrival: the memory in use goes UP by one per admission for as long as
+    // the disk is behind. Field 2026-09-02: 233 evictions against 119 completed
+    // writes in a minute, 203 blocks held with three pieces resident, and the
+    // process killed at 4.37 GB.
+    //
+    // So when the disk is what the store is waiting for, it waits. A completing
+    // write calls `#noteProgress`, which wakes whoever is here.
+    if (this.#blocksInFlight() > 0 && this.#blocksInUse() >= this.#growthCeiling) {
+      if (stillFor() < PINNED_WAIT_MS) {
+        this.#counters.waitedForDisk += 1;
+        return false;
+      }
+      // The disk has stopped answering. Falling through to eviction is the
+      // lesser failure: it grows memory, and the line above says by how much,
+      // where refusing would fail the read outright. The clock is NOT restarted
+      // here: the wait that follows asks the same question of the same claim,
+      // and zeroing it was half of the loop described in `#claimSlot`.
+      this.#counters.grewWaitingForDisk += 1;
+    }
+
+    const { index: victim, protectionYielded, distance } = this.#lru.evictionChoice();
+    if (victim === null) {
+      // Nothing may leave. Wait while the store is still MOVING — a spill
+      // completing, a piece admitted, a pin released — and give up when it has
+      // not moved for PINNED_WAIT_MS, whatever is nominally in flight. The old
+      // rule asked whether anything was in flight rather than whether anything
+      // had happened, which is why one lost reservation could hold a read here
+      // for ever.
+      if (stillFor() < PINNED_WAIT_MS) {
+        this.#counters.waitedForPins += 1;
+        return false;
+      }
+      this.#counters.blockedByPins += 1;
+      throw new Error(
+        `Every resident piece is pinned and nothing moved for ${PINNED_WAIT_MS}ms; no slot can be freed.`
+      );
+    }
+
+    const victimBuffer = this.#buffers.get(victim);
+    if (victimBuffer === undefined) {
+      // Should not happen: LRU says resident but buffer missing.
+      this.#lru.remove(victim);
+      return false;
+    }
+
+    // Claim atomically before await.
+    this.#buffers.delete(victim);
+    this.#lru.remove(victim);
+    this.#outstandingPieces += 1;
+    this.#noteEviction(protectionYielded, distance);
+
+    try {
+      await this.#spill(victim, victimBuffer);
+    } catch (error) {
+      // The caller never received a release for this reservation, so it is
+      // given back here rather than left outstanding for ever.
+      this.#outstandingPieces -= 1;
+      this.#noteProgress();
+      throw error;
+    }
+    // Outstanding stays +1 for the caller; the slot for the new piece is now free.
+    return true;
+  }
+
+  /**
+   * Bring a spilled piece back into memory, once, however many callers ask.
+   *
+   * @param {number} index
+   * @returns {Promise<SharedArrayBuffer | null>} `null` when the piece is on
+   *   neither tier.
+   */
+  async #revive(index) {
+    const spill = this.#evicting.get(index);
+    if (spill) {
+      await spill.catch(() => undefined);
+    }
+
+    if (!this.#disk.has(index)) {
+      return null;
+    }
+
+    const release = await this.#claimSlot();
+    try {
+      // Another caller may have brought it back while this one waited for a
+      // slot. Registering a second buffer for the same piece would leave
+      // whoever holds the first reading memory nothing evicts.
+      const already = this.#buffers.get(index);
+      if (already !== undefined) {
+        this.#lru.touch(index);
+        this.#counters.fromMemory += 1;
+        return already;
+      }
+      const target = this.#takeBlock();
+      try {
+        // Only this piece's own length: the block is a full piece long and the
+        // last piece of a file is shorter, so reading the whole block would ask
+        // the file for bytes past its end.
+        await this.#disk.read(index, Buffer.from(target, 0, this.#lengthOf(index)));
+      } catch (error) {
+        this.#returnBlock(target);
+        throw error;
+      }
+      this.#registerPiece(index, target);
+      this.#counters.fromDisk += 1;
+      this.#counters.revivals += 1;
+      this.#noteRevival(index);
+      return target;
+    } finally {
+      release();
+    }
+  }
+
+  /**
+   * A block to hold one piece: the most recently freed one, or a new one.
+   *
+   * Every block is a full piece long, whatever piece will live in it. The last
+   * piece of a file is shorter, and it occupies a full block with only its own
+   * bytes meaningful — `#lengthOf` is what decides how much is ever read out.
+   * Uniform blocks are what makes them interchangeable at all.
+   *
+   * Most recently freed first, deliberately: a few blocks then carry the whole
+   * of a busy store's traffic and the rest age out of use, which is what makes
+   * {@link SharedPieceStore#sweepFreeBlocks} able to tell a spare block from a
+   * working one.
+   *
+   * @returns {SharedArrayBuffer}
+   */
+  #takeBlock() {
+    const spare = this.#freeBlocks.pop();
+    if (spare !== undefined) {
+      this.#noteReuseGap(Date.now() - spare.freedAt);
+      return spare.buffer;
+    }
+    // Nothing spare and the pool is already as large as it is allowed to be.
+    // This cannot happen while the accounting is sound: a block is taken only
+    // after a slot has been claimed, and slots are exactly what the ceiling
+    // counts. It is recorded rather than hidden because when it does happen the
+    // pool grows without bound, and every block it then allocates is used once
+    // and thrown to a collector that has no reason to run — the heap stays at
+    // 50 MB of 2240 while the process reaches four gigabytes.
+    // Strictly greater: reaching the ceiling exactly is what a full store looks
+    // like, and a slot has just been claimed for this block. Being ALREADY past
+    // it and allocating anyway is the state that runs away.
+    if (this.#blocksAllocated > this.#growthCeiling) {
+      this.#counters.blocksBeyondCeiling += 1;
+    }
+    this.#blocksAllocated += 1;
+    return this.#watchForCollection(new SharedArrayBuffer(this.#chunkLength));
+  }
+
+  /**
+   * Put a block back for re-use, or give it up.
+   *
+   * Given up when the allowance has fallen below the number of blocks that
+   * exist: keeping it would hold memory the machine has just been shown to
+   * need. Otherwise it waits in the free list for the next piece.
+   *
+   * @param {SharedArrayBuffer | undefined} buffer
+   * @returns {void}
+   */
+  #returnBlock(buffer) {
+    if (buffer === undefined) {
+      return;
+    }
+    if (this.#closed || this.#blocksAllocated > this.#growthCeiling) {
+      // Never below zero: `close` gives up every block at once, and a spill
+      // that was already in flight resolves afterwards and arrives here.
+      if (this.#blocksAllocated > 0) {
+        this.#blocksAllocated -= 1;
+        this.#counters.blocksReleased += 1;
+      }
+      return;
+    }
+    this.#freeBlocks.push({ buffer, freedAt: Date.now() });
+  }
+
+  /**
+   * Record how long a block waited to be used again.
+   *
+   * @param {number} gapMs
+   * @returns {void}
+   */
+  #noteReuseGap(gapMs) {
+    this.#reuseGaps.push(Math.max(0, gapMs));
+    if (this.#reuseGaps.length > REUSE_GAP_SAMPLES) {
+      this.#reuseGaps.shift();
+    }
+  }
+
+  /**
+   * The longest a block has recently waited before being wanted again, or null
+   * when no block has yet been re-used.
+   *
+   * This is the store's own working rhythm, measured rather than chosen: while
+   * a film is being watched a block is taken again within milliseconds, because
+   * one is taken for every piece that arrives. A block that has been sitting
+   * longer than the longest of those waits is not part of the work.
+   *
+   * @returns {number | null}
+   */
+  #reuseGapCeilingMs() {
+    if (this.#reuseGaps.length === 0) {
+      return null;
+    }
+    return Math.max(...this.#reuseGaps);
+  }
+
+  /**
+   * Give up blocks that have sat unused longer than this store's own working
+   * rhythm.
+   *
+   * The case it is for: a torrent whose peers have gone. Its readers are still
+   * attached, so nothing removes the torrent — the pool's idle timer needs a
+   * refcount of zero and never starts. Its allowance falls to what those
+   * readers declared, the pieces beyond it are written out, and their blocks
+   * would otherwise wait in the free list for peers that may not return.
+   *
+   * @param {number} [now]
+   * @returns {number} Blocks given up.
+   */
+  sweepFreeBlocks(now = Date.now()) {
+    const ceiling = this.#reuseGapCeilingMs();
+    if (ceiling === null) {
+      return 0;
+    }
+    const keeping = [];
+    let released = 0;
+    for (const spare of this.#freeBlocks) {
+      if (now - spare.freedAt <= ceiling) {
+        keeping.push(spare);
+        continue;
+      }
+      this.#blocksAllocated -= 1;
+      this.#counters.blocksReleased += 1;
+      released += 1;
+    }
+    this.#freeBlocks = keeping;
+    return released;
+  }
+
+  /**
+   * Count this buffer as one this thread will have to let go of, and notice
+   * when the collector takes it. See {@link pieceBufferCollection}.
+   *
+   * @param {SharedArrayBuffer} buffer
+   * @returns {SharedArrayBuffer} The same buffer.
+   */
+  #watchForCollection(buffer) {
+    released.count += 1;
+    collector?.register(buffer, null);
+    return buffer;
+  }
+
+  /**
+   * A fresh buffer holding this piece's bytes.
+   *
+   * @param {number} index
+   * @param {Uint8Array} bytes
+   * @returns {SharedArrayBuffer}
+   */
+  #copyIntoNewBuffer(index, bytes) {
+    const length = this.#lengthOf(index);
+    const block = this.#takeBlock();
+    const view = Buffer.from(block, 0, length);
+    if (bytes.copy) {
+      bytes.copy(view, 0, 0, length);
+    } else {
+      view.set(bytes.subarray(0, length), 0);
+    }
+    return block;
+  }
+
+  /**
+   * Drop the disk copy of a piece that memory now holds — after any spill of
+   * that same piece has finished.
+   *
+   * `DiskTier.write` records the index when it COMPLETES, so forgetting while a
+   * spill of that index is still running let the completing write put it back,
+   * and a later read then returned the stale bytes.
+   *
+   * @param {number} index
+   * @returns {Promise<void>}
+   */
+  async #forgetOnDisk(index) {
+    const spill = this.#evicting.get(index);
+    if (spill) {
+      await spill.catch(() => undefined);
+    }
+    this.#disk.forget(index);
+    this.#spilledAt.delete(index);
+  }
+
+  put(index, bytes, callback = () => undefined) {
+    if (this.#closed) {
+      queueMicrotask(() => callback(new Error("Piece store is closed.")));
+      return;
+    }
+
+    const write = async () => {
+      // Already resident: the buffer is replaced, not added, so no slot is
+      // needed and none is claimed. A fresh buffer rather than a write into the
+      // old one, so a reader holding the old reference cannot see a torn write.
+      if (this.#buffers.has(index)) {
+        const previous = this.#buffers.get(index);
+        // A fresh block rather than a write into the old one, so a reader
+        // holding the old reference cannot see a torn write.
+        this.#buffers.set(index, this.#copyIntoNewBuffer(index, bytes));
+        this.#lru.touch(index);
+        // And the old block goes back for re-use only if nobody is reading it.
+        // A pinned piece has a view onto its memory somewhere; that block is
+        // given up instead, and the pool allocates another when it needs one.
+        if (this.#lru.isPinned(index)) {
+          if (this.#blocksAllocated > 0) {
+            this.#blocksAllocated -= 1;
+            this.#counters.blocksReleased += 1;
+          }
+        } else {
+          this.#returnBlock(previous);
+        }
+        await this.#forgetOnDisk(index);
+        this.#noteProgress();
+        return;
+      }
+
+      this.#noteArrival();
+      const declared = this.#lru.wants(index);
+      if (declared) {
+        this.#counters.admittedInsideWindow += 1;
+      } else {
+        this.#counters.admittedOutsideWindow += 1;
+      }
+
+      // A piece wanted LESS than the one it would displace goes straight to
+      // disk instead of being admitted.
+      //
+      // The same comparison eviction makes, in the same order, from the other
+      // end: the priority map's level first, and the distance to a read head
+      // only between pieces the map wants equally. Answering the two from
+      // different quantities would let the store admit a piece by one rule and
+      // evict it again by the other in the same second.
+      //
+      // Two cases, and the second was missing until 2026-09-02. The first: the
+      // arrival is in nobody's zone at all. The second: it IS in a zone, but
+      // one the map wants less than the zone holding the piece the store would
+      // have to evict for it. The field session had six readers whose windows
+      // covered the whole file, so the first case never applied and `0 of those
+      // went straight to disk` while the store spilled 233 pieces in a minute.
+      //
+      // Only when SOMETHING is stated: before that there is no basis for
+      // calling one piece more wanted than another.
+      const worseThanTheVictim = () => {
+        const victim = this.#lru.nextVictim();
+        if (victim.index === null) {
+          return false;
+        }
+        const arrivingWant = this.#lru.wantAt(index);
+        if (arrivingWant !== victim.want) {
+          return arrivingWant > victim.want;
+        }
+        const arriving = this.#lru.waitFor(index);
+        return arriving >= 0 && victim.wait >= 0 && arriving > victim.wait;
+      };
+      if (this.#lru.protectedCount > 0 && this.#isFullNow()
+        && (!declared || worseThanTheVictim())) {
+        this.#counters.admittedToDisk += 1;
+        await this.#writeThrough(index, bytes);
+        this.#noteProgress();
+        return;
+      }
+
+      const release = await this.#claimSlot();
+      try {
+        this.#registerPiece(index, this.#copyIntoNewBuffer(index, bytes));
+        await this.#forgetOnDisk(index);
+      } finally {
+        release();
+      }
+    };
+
+    write().then(() => callback(null), (error) => callback(error));
+  }
+
+  get(index, options, callback) {
+    if (typeof options === "function") {
+      return this.get(index, undefined, options);
+    }
+    const done = callback ?? (() => undefined);
+    if (this.#closed) {
+      queueMicrotask(() => done(new Error("Piece store is closed.")));
+      return;
+    }
+
+    const pieceLength = this.#lengthOf(index);
+    const offset = options?.offset ?? 0;
+    const length = options?.length ?? pieceLength - offset;
+
+    const fetch = async () => {
+      const buffer = this.#buffers.get(index);
+      if (buffer !== undefined) {
+        this.#lru.touch(index);
+        this.#counters.fromMemory += 1;
+        return Buffer.from(Buffer.from(buffer, offset, length));
+      }
+
+      const revived = await this.#revive(index);
+      if (revived === null) {
+        throw new Error(`Piece ${index} is not in the store.`);
+      }
+      return Buffer.from(Buffer.from(revived, offset, length));
+    };
+
+    fetch().then((bytes) => done(null, bytes), (error) => done(error));
+  }
+
+  /**
+   * State a range of pieces and how much they are wanted. The number is the
+   * priority map's own, and it is what eviction compares — see
+   * {@link PieceLru#protect}.
+   *
+   * @param {string|number} readerId
+   * @param {number} from
+   * @param {number} to
+   * @param {number} [urgency]
+   * @returns {void}
+   */
+  protectRange(readerId, from, to, urgency) {
+    this.#lru.protect(readerId, from, to, urgency);
+  }
+
+  protectedRanges() {
+    return this.#lru.protectedRanges();
+  }
+
+  releaseProtection(readerId) {
+    this.#lru.unprotect(readerId);
+  }
+
+  warmRange(from, to, limit = Math.max(1, Math.floor(this.#growthCeiling / 4))) {
+    if (this.#closed || !Number.isInteger(from) || !Number.isInteger(to)) {
+      return 0;
+    }
+    let started = 0;
+    for (let index = from; index <= to && started < limit; index += 1) {
+      if (this.#buffers.has(index) || !this.#disk.has(index)) {
+        continue;
+      }
+      started += 1;
+      void this.reside(index).catch(() => undefined);
+    }
+    return started;
+  }
+
+  async reside(index) {
+    if (this.#closed) {
+      throw new Error("Piece store is closed.");
+    }
+
+    const buffer = this.#buffers.get(index);
+    if (buffer !== undefined) {
+      this.#lru.touch(index);
+      this.#counters.fromMemory += 1;
+      return this.locate(index);
+    }
+
+    const revived = await this.#revive(index);
+    if (revived === null) {
+      return null;
+    }
+    return { buffer: revived, offset: 0, length: this.#lengthOf(index) };
+  }
+
+  close(callback = () => undefined) {
+    this.#closed = true;
+    liveStores.delete(this);
+    this.#buffers.clear();
+    this.#spilledAt.clear();
+    this.#counters.blocksReleased += this.#blocksAllocated;
+    this.#blocksAllocated = 0;
+    this.#freeBlocks = [];
+    // Whoever is waiting for a slot is woken and finds the store closed, which
+    // is an error they can report. Left asleep they simply never returned.
+    this.#wake();
+    this.#disk.close().then(() => callback(null), (error) => callback(error));
+  }
+
+  destroy(callback = () => undefined) {
+    this.#closed = true;
+    liveStores.delete(this);
+    this.#buffers.clear();
+    this.#spilledAt.clear();
+    this.#counters.blocksReleased += this.#blocksAllocated;
+    this.#blocksAllocated = 0;
+    this.#freeBlocks = [];
+    this.#wake();
+    this.#disk.destroy().then(() => callback(null), (error) => callback(error));
+  }
+}