@torrent-tv/proxy 2.89.1 → 2.89.3

package/services/transport/delivery-probe.js

@@ -1,842 +1,842 @@
-/**
- * @file Numbered delivery probes, and the verdict they make possible.
- *
- * The delivery freeze (roadmap item 11) looks identical from the proxy in two
- * cases that need opposite fixes: usrsctp stopped transmitting, or the browser
- * closed its receive window because the page stopped draining the channel. The
- * proxy's own counters cannot separate them — libdatachannel's `bytesSent`
- * counts bytes ACCEPTED into usrsctp, not bytes put on the wire — so the
- * reading has to come from the far end.
- *
- * Two facts make that cheap. The reverse direction keeps working throughout the
- * freeze (browser to proxy requests arrive and are answered for the whole
- * episode, 88 min in the 2026-08-24 case), so the browser can always report.
- * And SCTP orders per STREAM, so a probe on a channel opened UNORDERED and
- * WITHOUT retransmission passes head-of-line blocking in another stream but
- * neither a closed receive window nor a transmitter that stopped.
- *
- * So: number a probe every {@link PROBE_INTERVAL_MS} on every channel of the
- * connection, have the browser echo back the highest number it has seen on
- * each, and read the gaps:
- *
- *   every channel current                to flowing
- *   ordered behind, unreliable current   to a retransmission stuck in a stream
- *   both behind, echoes still arriving   to the association stopped transmitting
- *   no echo at all                       to the reverse direction went too
- *
- * The verdict is computed from the gaps, not chosen, and every line prints the
- * numbers that produced it.
- */
-
-/**
- * How often a probe is numbered and sent on every channel.
- *
- * Half a second, because the transport heartbeat is five and that was the whole
- * resolution the 2026-08-24 episode had: the onset could be placed no closer
- * than the five seconds between two lines. Sending is cheap - a probe is a few
- * dozen bytes - and the interval is NOT the verdict: a healthy burst can hold a
- * probe up behind queued data, so how many probes may be outstanding is worked
- * out per channel by {@link allowedGap} from the bytes queued ahead of the
- * probe and the rate they are leaving at.
- */
-export const PROBE_INTERVAL_MS = 500;
-
-/**
- * How many probes may be outstanding on a channel before it counts as behind.
- *
- * DERIVED per channel, not chosen. A probe is handed to the same association as
- * the data, and SCTP orders per stream but SCHEDULES per association: one
- * congestion window, one send buffer. So a probe waits for whatever is queued
- * ahead of it whichever channel it rides on - the unordered one included - and
- * a fixed threshold cannot tell that wait from a stopped association. Measured
- * 2026-08-26: `association-stopped` printed with all three channels at gap 4-7
- * while 110-150 Mbps crossed that same association and 7.34 GB went through it
- * without a single failure.
- *
- * What the wait costs is arithmetic on two measured quantities: the bytes
- * queued ahead of the probe, and the rate at which this connection is getting
- * bytes away. Add one round trip for the echo to come back. A probe is behind
- * only when it is later than that.
- *
- * There is a third term, and leaving it out made this worse than the constant
- * it replaced. The browser answers on ITS own schedule, not ours: it batches
- * what it has seen and echoes on a timer, and that timer is throttled to about
- * once a second whenever the tab is hidden. So with an empty queue the
- * allowance collapsed to one probe and the bound to half a second, while echoes
- * legitimately arrived every second - measured 2026-08-27, 67 `association-
- * stopped` and 66 `reverse-direction-gone` against 84 `flowing` on a connection
- * carrying 3.4 MB/s with every queue at zero. The peer's own cadence is
- * measurable on the same connection, so it is measured and added rather than
- * assumed.
- *
- * There is a fourth term, and without it the third one arrives too late to
- * help. `echoIntervalMs` is the widest gap between two echoes this connection
- * has SHOWN, so it can only grow after a late echo has landed — and a browser
- * whose tab has just been hidden goes quiet before it has taught us anything.
- * The browser measures its own event-loop delay and puts it in every echo, so
- * the delay is known BEFORE the silence rather than after it: an echo timer
- * set for half a second cannot fire until the loop runs, and a loop reported
- * as 4297 ms behind cannot answer sooner than that. Field 2026-09-03: a viewer
- * paused at 15:24:21, the tab went hidden at 15:24:35, `loopLag` climbed
- * 681 → 1881 → 4297 → 5957 ms, and at 15:26:05 the probes read `gap 12 of 11`
- * and printed `association-stopped` on a connection that was `flowing` again
- * seven seconds later. Adding the peer's own reported lag makes that allowance
- * 15 instead of 11, and the gap of 12 is then what it was — a browser whose
- * timers are frozen, not an association that stopped.
- *
- * @param {{ queuedBytes: number, bytesPerSecond: number, rttMs: number, echoIntervalMs?: number, peerLoopLagMs?: number, intervalMs?: number }} state
- * @returns {number | null} Probes that may legitimately be outstanding, or null
- *   when no rate has been measured yet and nothing can be said.
- */
-export function allowedGap({
-  queuedBytes,
-  bytesPerSecond,
-  rttMs,
-  echoIntervalMs = 0,
-  peerLoopLagMs = 0,
-  intervalMs = PROBE_INTERVAL_MS
-}) {
-  if (!(bytesPerSecond > 0) || !(intervalMs > 0)) {
-    return null;
-  }
-  // At least one: a probe sent and not yet echoed is the ordinary state.
-  return Math.max(1, Math.ceil(allowedWaitMs({ queuedBytes, bytesPerSecond, rttMs, echoIntervalMs, peerLoopLagMs }) / intervalMs));
-}
-
-/**
- * How long a probe may legitimately take to be reported back, in milliseconds.
- *
- * THE QUANTITY {@link allowedGap} COMPUTES AND THEN THROWS AWAY by dividing it
- * into probes. Probes are the wrong unit and always were: the same probe goes
- * down every channel INCLUDING the one carrying the film, and SCTP schedules
- * per association, so a probe waits behind queued video exactly as a segment
- * does. What is behind is then not the peer's answer but the probe itself —
- * and the honest measure of that is the age of the newest probe the peer has
- * seen, which this is compared against.
- *
- * Four terms, every one measured: the queue's own drain time at the rate this
- * connection is achieving, the crossing, the peer's reporting cadence, and how
- * late the peer's event loop is running.
- *
- * @param {{ queuedBytes: number, bytesPerSecond: number, rttMs: number,
- *   echoIntervalMs?: number, peerLoopLagMs?: number }} state
- * @returns {number}
- */
-export function allowedWaitMs({ queuedBytes, bytesPerSecond, rttMs, echoIntervalMs = 0, peerLoopLagMs = 0 }) {
-  if (!(bytesPerSecond > 0)) {
-    return 0;
-  }
-  const drainMs = (Math.max(queuedBytes, 0) / bytesPerSecond) * 1000;
-  return drainMs + Math.max(rttMs, 0) + Math.max(echoIntervalMs, 0) + Math.max(peerLoopLagMs, 0);
-}
-
-/**
- * How far back the send times of probes are kept.
- *
- * A probe older than the point at which the reverse direction is called gone
- * can say nothing further, and that bound is itself derived per connection —
- * this is the outer edge of it, kept so the map cannot grow with the session.
- */
-const PROBE_HISTORY_MS = 10 * 60 * 1000;
-
-/** The label the browser gives the unordered, non-retransmitting channel. */
-export const UNRELIABLE_LABEL = "proxy-fast";
-
-/**
- * How old an echo may be before the reverse direction counts as gone, when
- * nothing better can be derived.
- *
- * Used only while no rate has been measured. Otherwise the caller passes
- * `echoStaleMs`, worked out the same way as {@link allowedGap}: the browser
- * only echoes a probe it has RECEIVED, so an echo waits behind our own queue
- * exactly as the probe did.
- */
-const ECHO_STALE_FALLBACK_MS = 5_000;
-
-/** How often the probe state is written to the log while nothing changes. */
-const REPORT_INTERVAL_MS = 5_000;
-
-/**
- * Whether the `seen` counter has stopped advancing for longer than this
- * connection's own history says a healthy gap between two advances ever
- * takes.
- *
- * The `association-stopped` verdict alone is not enough to act on: a
- * connection can sit BEHIND by a bounded, roughly constant amount for
- * minutes (measured 2026-08-28, session on a backgrounded tab — gap held at
- * 6-7 probes for 95+ seconds while `seen` kept climbing right along with
- * `sent`) without anything being wrong. What a true wedge shows instead,
- * measured the same day against a session already known to be one
- * (`d85ae4f5`): `seen` FROZEN at one value for over a minute while `sent`
- * climbs unbounded. So the question is not "is there a gap" but "has the
- * highest-seen number stopped moving at all, for longer than it has ever
- * legitimately taken this connection to report an advance" — the same shape
- * as {@link wedgeIsCertain} in `transport/data-channel-handler.js`, applied to the
- * probe's own counter instead of the transport's byte counter.
- *
- * THE FLOOR IS NOT ONE PROBE INTERVAL. "The longest gap this connection has
- * ever shown" is nothing at all in its first minutes, and the floor under it
- * was a single interval — 500 ms — so any two-second quiet stretch while the
- * browser filled its cushion read as longer than anything healthy ever seen.
- * Field 2026-09-11: two captures of 180 s each, triggered at `wedged 1s` and
- * `wedged 2s`, on a connection with `rtt=5ms`, every queue at 0 B and the
- * viewer watching; 65 MB of them were left on the addon's disk.
- *
- * What replaces it is the connection's own answer to how long a legitimate
- * report may take: the peer reports twice a second, the report has to cross,
- * and the peer's own event loop may be late — all three measured, all three
- * already summed for `echoStaleMs`. A stretch shorter than that proves nothing
- * whatever about the association.
- *
- * @param {{ stuckForMs: number, longestHealthySeenGapMs: number, intervalMs?: number,
- *   legitimateReportMs?: number }} state
- * @returns {{ isCertain: boolean, needMs: number }}
- */
-export function probeWedgeIsCertain({
-  stuckForMs,
-  longestHealthySeenGapMs,
-  intervalMs = PROBE_INTERVAL_MS,
-  legitimateReportMs = 0
-}) {
-  const floorMs = Math.max(intervalMs, Number.isFinite(legitimateReportMs) ? legitimateReportMs : 0);
-  const needMs = Math.max(longestHealthySeenGapMs, floorMs);
-  return { isCertain: stuckForMs >= needMs, needMs };
-}
-
-/**
- * One connection's probe state.
- *
- * @typedef {Object} ProbeConnection
- * @property {string} tag
- * @property {Map<import('node-datachannel').DataChannel, string>} channels
- * @property {number} seq        - Highest probe number sent.
- * @property {number} sentAt     - When that probe was sent.
- * @property {Map<string, number>} seen - Label to the highest number the browser reported.
- * @property {number} echoAt     - When the last echo arrived (0 = never).
- * @property {number} echoes     - How many echoes have arrived.
- * @property {string} verdict    - Last verdict reported, so a change is logged at once.
- * @property {number} reportedAt - When the state was last written to the log.
- * @property {number} lastSeenAdvanceAt - When any label's `seen` value last increased (0 = never yet).
- * @property {number} longestHealthySeenGapMs - The longest gap between two advances this connection has shown while not flagged as a wedge.
- * @property {number | null} peerBytes - The far end's transport-level received total, as last reported.
- * @property {number | null} peerBytesAtTick - The same, as it stood at the previous tick.
- * @property {number | null} peerLoopLagMs - The far end's own event-loop delay, as last reported.
- * @property {string | null} peerVisibility - Whether the far end's tab is visible or hidden, as last reported.
- * @property {boolean} probeCaptureStarted - One evidence-gathering attempt per wedge; reset once `seen` advances again.
- * @property {ReturnType<typeof setInterval> | null} timer
- */
-
-/**
- * Read the gaps and say what they mean.
- *
- * Exported so the rule is testable without a connection: the same numbers
- * always produce the same word.
- *
- * `allowed` carries, per channel label, how many probes may legitimately be
- * outstanding right now — {@link allowedGap} computes it from that channel's
- * own queue and the connection's measured rate. A label with no entry, or an
- * entry of null, cannot be judged: with no rate measured there is nothing to
- * divide the queue by, and the verdict says that rather than inventing one.
- *
- * `peerBytesAdvancing` is the fact that outranks every gap here. The gaps are
- * counted against an allowance whose only load-dependent term is OUR OWN
- * queue — and that queue is empty by construction, because it drains the moment
- * libdatachannel accepts the bytes, whether or not usrsctp then puts them on
- * the wire. So a browser filling its cushion as fast as the link allows shows
- * an empty queue, a small allowance and a large gap, which reads exactly like a
- * stopped association. Measured 2026-08-28: four `association-stopped` in the
- * first two minutes of a healthy session, on a connection whose every queue was
- * at 0 B and whose viewer never saw the picture stop. Deepening the browser's
- * cushion from 30 s to 120 s (roadmap item 4) made the burst four times longer
- * and the false verdict correspondingly likelier.
- *
- * What separates the two is not the size of the backlog — it is large in both —
- * but whether bytes are still arriving. The browser reports its own
- * transport-level received total with every echo, so the question is answered
- * by a counter rather than by a threshold: while that total is advancing, bytes
- * ARE crossing and no verdict of a stopped association can stand, however far
- * behind the probes are. `null` where the far end does not report it, and then
- * the rule falls back to what it says without the term.
- *
- * `peerLoopLagMs` and `peerVisibility` are printed but never compared against
- * anything here: they are what made the allowance as wide as it is, and a
- * reader of the log cannot check that arithmetic unless the terms are on the
- * same line as the result.
- *
- * @param {{ seq: number, seen: Map<string, number> | Record<string, number>, labels: string[], echoes: number, echoAgeMs: number | null, allowed?: Map<string, number | null> | Record<string, number | null>, echoStaleMs?: number, peerBytesAdvancing?: boolean | null, peerLoopLagMs?: number | null, peerVisibility?: string | null, peerPending?: number | null }} state
- * @returns {{ verdict: string, detail: string }}
- */
-export function readProbeState(state) {
-  const seenOf = (label) =>
-    state.seen instanceof Map ? state.seen.get(label) : state.seen?.[label];
-  /**
-   * @param {unknown} source
-   * @param {string} label
-   * @returns {number | null}
-   */
-  const timeOf = (source, label) => {
-    const value = source instanceof Map ? source.get(label) : /** @type {any} */ (source)?.[label];
-    return Number.isFinite(value) ? Number(value) : null;
-  };
-  const allowedOf = (label) => {
-    const source = state.allowed;
-    const value = source instanceof Map ? source.get(label) : source?.[label];
-    return Number.isInteger(value) ? Number(value) : null;
-  };
-  const parts = [];
-  let isOrderedBehind = false;
-  let isUnreliableBehind = false;
-  let isUnreliableKnown = false;
-  let isJudgeable = false;
-  for (const label of state.labels) {
-    const seen = seenOf(label);
-    const gap = Number.isInteger(seen) ? state.seq - Number(seen) : null;
-    const allowance = allowedOf(label);
-    // BY TIME WHERE IT IS KNOWN. The count is what the line prints, because it
-    // is what a reader recognises; what decides is a time.
-    //
-    // The measured one-way delay first: it is the forward direction itself,
-    // with the two clocks reconciled from the exchange. The age of the newest
-    // reported probe is the fallback, and it is larger than the thing itself by
-    // the peer's reporting cadence and the way back — which is why its own
-    // allowance carries both of those and the one-way allowance carries
-    // neither.
-    const oneWay = timeOf(state.oneWayMs, label);
-    const oneWayAllowed = timeOf(state.allowedOneWayMs, label);
-    const lagMs = oneWay !== null && oneWayAllowed !== null ? oneWay : timeOf(state.behindMs, label);
-    const mayWaitMs = oneWay !== null && oneWayAllowed !== null
-      ? oneWayAllowed
-      : timeOf(state.allowedWaitMs, label);
-    const lagText = lagMs === null || mayWaitMs === null
-      ? ""
-      : ` ${Math.round(lagMs)}ms of ${Math.round(mayWaitMs)}ms${oneWay !== null ? " one way" : ""}`;
-    parts.push(`${label}=${seen ?? "?"}(gap ${gap ?? "?"} of ${allowance ?? "?"}${lagText})`);
-    if (allowance === null) {
-      continue;
-    }
-    isJudgeable = true;
-    const isBehind = lagMs !== null && mayWaitMs !== null
-      ? lagMs > mayWaitMs
-      : gap === null || gap > allowance;
-    if (label === UNRELIABLE_LABEL) {
-      isUnreliableKnown = true;
-      isUnreliableBehind = isBehind;
-    } else if (isBehind) {
-      isOrderedBehind = true;
-    }
-  }
-  const isPeerAdvancing = state.peerBytesAdvancing === true;
-  const detail =
-    `sent=${state.seq} ${parts.join(" ")} ` +
-    `echoAge=${state.echoAgeMs === null ? "never" : `${state.echoAgeMs}ms`}` +
-    (state.peerBytesAdvancing === null || state.peerBytesAdvancing === undefined
-      ? ""
-      : ` peerBytes=${isPeerAdvancing ? "advancing" : "still"}`) +
-    (Number.isFinite(state.peerLoopLagMs) ? ` peerLoopLag=${Math.round(Number(state.peerLoopLagMs))}ms` : "") +
-    (state.peerVisibility ? ` peerTab=${state.peerVisibility}` : "") +
-    // WHAT THE PEER SAYS IT IS STILL WAITING FOR. The far end has sent this
-    // twice a second since the probe echo was built and nothing read it, while
-    // every verdict here was reasoned out from this side alone. It is evidence
-    // and not a verdict: the receiver declares its own transport lost, because
-    // it knows first — on 2026-09-06 the browser said so sixteen minutes before
-    // this end did.
-    (Number.isFinite(state.peerPending) ? ` peerWaiting=${Number(state.peerPending)}` : "");
-
-  if (state.echoes === 0) {
-    return { verdict: "no-echo-yet", detail };
-  }
-  if (!isJudgeable) {
-    return { verdict: "no-rate-yet", detail };
-  }
-  const staleAfterMs = Number.isFinite(state.echoStaleMs) && state.echoStaleMs > 0
-    ? state.echoStaleMs
-    : ECHO_STALE_FALLBACK_MS;
-  if (state.echoAgeMs !== null && state.echoAgeMs > staleAfterMs) {
-    return { verdict: "reverse-direction-gone", detail };
-  }
-  if (!isOrderedBehind && !(isUnreliableKnown && isUnreliableBehind)) {
-    return { verdict: "flowing", detail };
-  }
-  // Bytes are still arriving at the far end. Whatever the probe gaps say, this
-  // association has not stopped — the probes are behind a backlog, which is
-  // what filling a cushion looks like from here.
-  if (isPeerAdvancing) {
-    return { verdict: "flowing", detail };
-  }
-  if (isOrderedBehind && isUnreliableKnown && !isUnreliableBehind) {
-    return { verdict: "stream-stuck", detail };
-  }
-  if (isOrderedBehind) {
-    return {
-      verdict: isUnreliableKnown ? "association-stopped" : "ordered-behind-no-comparison",
-      detail
-    };
-  }
-  return { verdict: "unreliable-behind-only", detail };
-}
-
-/**
- * Create the probe service. One instance serves every session.
- *
- * @param {Object} options
- * @param {(message: string) => void} options.log
- * @param {number} [options.intervalMs]
- * @param {(sessionId: string) => object | null} [options.getTransportSnapshot]
- *   Needed only to hand the witness a remote endpoint when this probe is the
- *   one declaring a wedge.
- * @param {{ maybeCapture: (trigger: object) => boolean }} [options.witness]
- * @param {{ maybeRead: (reasonText: string) => boolean }} [options.usrsctpState]
- * @returns {{
- *   attach: (sessionId: string, tag: string, label: string, channel: import('node-datachannel').DataChannel) => void,
- *   detach: (sessionId: string, channel: import('node-datachannel').DataChannel) => void,
- *   noteEcho: (sessionId: string, echo: object) => void,
- *   dispose: () => void
- * }}
- */
-export function createDeliveryProbe({
-  log,
-  intervalMs = PROBE_INTERVAL_MS,
-  readDelivery,
-  getTransportSnapshot,
-  witness,
-  usrsctpState
-}) {
-  /** @type {Map<string, ProbeConnection>} */
-  const connections = new Map();
-
-  /**
-   * Send this tick's probe on every channel of one connection, then report.
-   *
-   * @param {ProbeConnection} connection
-   * @returns {void}
-   */
-  function tick(connection) {
-    const now = Date.now();
-    connection.seq += 1;
-    connection.sentAt = now;
-    // WHEN each probe went out, so that what the peer reports can be read as a
-    // time rather than as a count of probes. Pruned to the oldest probe any
-    // judgement could still be about: once a probe is older than the point at
-    // which the reverse direction is called gone, its age says nothing further.
-    connection.sentAtBySeq.set(connection.seq, now);
-    for (const [probeNumber, sentAt] of connection.sentAtBySeq) {
-      if (now - sentAt > PROBE_HISTORY_MS) {
-        connection.sentAtBySeq.delete(probeNumber);
-      } else {
-        break;
-      }
-    }
-    const message = JSON.stringify({ type: "probe", seq: connection.seq, sentAt: now });
-    for (const channel of connection.channels.keys()) {
-      try {
-        channel.sendMessage(message);
-      } catch {
-        // A channel closing between the check and the send is ordinary.
-      }
-    }
-
-    // What this connection is getting away, and how far behind it therefore
-    // sits. Both come from the send-queue watcher, which measures them anyway.
-    const delivery = typeof readDelivery === "function" ? readDelivery(connection.id) : null;
-    const bytesPerSecond = Number(delivery?.bytesPerSecond) || 0;
-    const rttMs = Number(delivery?.rttMs) || 0;
-    /** @type {Map<string, number | null>} */
-    const allowed = new Map();
-    /** How long each channel's newest unreported probe may legitimately be. */
-    const allowedWait = new Map();
-    /** The same, for the measured one-way time where the clocks are reconciled. */
-    const allowedOneWay = new Map();
-    for (const [channel, label] of connection.channels) {
-      let queuedBytes = 0;
-      try {
-        queuedBytes = typeof channel.bufferedAmount === "function" ? channel.bufferedAmount() : 0;
-      } catch {
-        queuedBytes = 0;
-      }
-      const allowance = allowedGap({
-        queuedBytes,
-        bytesPerSecond,
-        rttMs,
-        echoIntervalMs: connection.echoIntervalMs,
-        peerLoopLagMs: connection.peerLoopLagMs ?? 0,
-        intervalMs
-      });
-      const waitMs = allowedWaitMs({
-        queuedBytes,
-        bytesPerSecond,
-        rttMs,
-        echoIntervalMs: connection.echoIntervalMs,
-        peerLoopLagMs: connection.peerLoopLagMs ?? 0
-      });
-      const heldWait = allowedWait.get(label);
-      if (!Number.isFinite(heldWait) || waitMs > heldWait) {
-        allowedWait.set(label, waitMs);
-      }
-      // What a probe may take ONE WAY: the queue's own drain time and half the
-      // crossing. Neither the peer's reporting cadence nor its loop delay
-      // belongs here — with the clocks reconciled they are no longer in the
-      // measurement they would have to be allowed for.
-      const oneWayAllowed = allowedWaitMs({ queuedBytes, bytesPerSecond, rttMs: rttMs / 2 });
-      const heldOneWay = allowedOneWay.get(label);
-      if (!Number.isFinite(heldOneWay) || oneWayAllowed > heldOneWay) {
-        allowedOneWay.set(label, oneWayAllowed);
-      }
-      // Several channels can carry one label only in malformed cases; the
-      // larger allowance is the safer of the two.
-      const held = allowed.get(label);
-      if (allowance !== null && (!Number.isInteger(held) || allowance > held)) {
-        allowed.set(label, allowance);
-      } else if (!allowed.has(label)) {
-        allowed.set(label, allowance);
-      }
-    }
-    const widest = [...allowed.values()].reduce(
-      (most, value) => (Number.isInteger(value) && value > most ? value : most),
-      0
-    );
-    // Advancing SINCE THE PREVIOUS TICK, not since the connection began: the
-    // question is whether bytes are crossing now. Prefer channel bytes —
-    // transport's bytesReceived advances on SACKs (140 B/s on a wedge) and
-    // gives a false "advancing" every other tick, while channel bytes are flat.
-    let peerBytesAdvancing = null;
-    if (connection.peerChannelBytes !== null) {
-      peerBytesAdvancing =
-        connection.peerChannelBytesAtTick === null ||
-        connection.peerChannelBytes > connection.peerChannelBytesAtTick;
-      connection.peerChannelBytesAtTick = connection.peerChannelBytes;
-      // keep transport's tick in sync for fallback logging
-      connection.peerBytesAtTick = connection.peerBytes;
-    } else if (connection.peerBytes !== null) {
-      peerBytesAdvancing =
-        connection.peerBytesAtTick === null || connection.peerBytes > connection.peerBytesAtTick;
-      connection.peerBytesAtTick = connection.peerBytes;
-    }
-    const { verdict, detail } = readProbeState({
-      seq: connection.seq,
-      seen: connection.seen,
-      peerBytesAdvancing,
-      peerPending: connection.peerPending,
-      labels: [...new Set(connection.channels.values())],
-      echoes: connection.echoes,
-      echoAgeMs: connection.echoAt === 0 ? null : now - connection.echoAt,
-      peerLoopLagMs: connection.peerLoopLagMs,
-      peerVisibility: connection.peerVisibility,
-      allowed,
-      // WHAT IS BEHIND, IN TIME. For each channel: how old the newest probe the
-      // peer has reported seeing is. The same probe goes down every channel
-      // including the one carrying the film, and SCTP schedules per
-      // association, so a probe waits behind queued video exactly as a segment
-      // does — which means a count of outstanding probes measures the queue,
-      // not the association. This is the queue's own time, against the time the
-      // queue is allowed to take.
-      behindMs: Object.fromEntries(
-        [...connection.seen.entries()].map(([label, newestSeenNumber]) => {
-          const sentAt = connection.sentAtBySeq.get(newestSeenNumber);
-          return [label, Number.isFinite(sentAt) ? now - sentAt : null];
-        })
-      ),
-      allowedWaitMs: Object.fromEntries(allowedWait.entries()),
-      // The measured forward delay, where the two clocks have been reconciled.
-      // Preferred over the age above because it is the thing itself: the age
-      // also carries the peer's reporting cadence and the way back.
-      oneWayMs: Object.fromEntries(connection.oneWayMs.entries()),
-      allowedOneWayMs: Object.fromEntries(allowedOneWay.entries()),
-      // Same arithmetic for the echo's own age: the peer cannot answer sooner
-      // than its own cadence allows, nor sooner than its own event loop runs.
-      echoStaleMs:
-        widest > 0
-          ? widest * intervalMs +
-            rttMs +
-            connection.echoIntervalMs +
-            Math.max(connection.peerLoopLagMs ?? 0, 0)
-          : 0
-    });
-    if (verdict !== connection.verdict || now - connection.reportedAt >= REPORT_INTERVAL_MS) {
-      connection.verdict = verdict;
-      connection.reportedAt = now;
-      log(`[dc-probe] ${connection.tag} ${verdict} — ${detail} at=${new Date(now).toISOString()}`);
-    }
-
-    // `association-stopped` alone is not certainty — see probeWedgeIsCertain.
-    // A connection that is merely lagging by a bounded amount reaches this
-    // verdict too (a hidden tab's own echo cadence, measured 2026-08-28), and
-    // `seen` keeps advancing right along with it. Only a `seen` value that has
-    // stopped moving ENTIRELY, for longer than this connection has ever shown
-    // as a legitimate gap, is the wedge this exists to catch.
-    const stuckForMs = connection.lastSeenAdvanceAt === 0 ? 0 : now - connection.lastSeenAdvanceAt;
-    if (verdict === "association-stopped") {
-      const { isCertain, needMs } = probeWedgeIsCertain({
-        stuckForMs,
-        longestHealthySeenGapMs: connection.longestHealthySeenGapMs,
-        // The same sum the echo's own staleness is judged by: the peer reports
-        // twice a second, the report has to cross, and the peer's loop may be
-        // late. Nothing shorter than that says anything.
-        // BOTH cadences, because they beat against each other: a probe sent
-        // just after the peer composed a report shows up only in the NEXT one.
-        // So the longest legitimate gap between two advances is one probe
-        // interval plus one report interval plus the crossing plus whatever
-        // the peer's own loop is behind by.
-        legitimateReportMs:
-          intervalMs +
-          connection.echoIntervalMs +
-          rttMs +
-          Math.max(connection.peerLoopLagMs ?? 0, 0)
-      });
-      if (isCertain && !connection.probeCaptureStarted) {
-        connection.probeCaptureStarted = true;
-        const reasonText =
-          `probe seen-counter unmoved ${Math.round(stuckForMs / 1000)}s against the ` +
-          `${(needMs / 1000).toFixed(1)}s this connection's own history says is legitimate`;
-        if (witness) {
-          const snapshot = getTransportSnapshot?.(connection.id) ?? null;
-          const started = witness.maybeCapture({
-            sessionId: connection.id,
-            tag: connection.tag,
-            label: "probe",
-            remote: snapshot?.remote ?? null,
-            queuedBytes: 0,
-            stuckForMs
-          });
-          if (!started) {
-            connection.probeCaptureStarted = false;
-          }
-        }
-        if (usrsctpState) {
-          usrsctpState.maybeRead(reasonText);
-        }
-      }
-    } else {
-      // Not association-stopped any more: whatever was flagged has cleared,
-      // and a later wedge on the same connection deserves its own attempt.
-      connection.probeCaptureStarted = false;
-    }
-  }
-
-  return {
-    attach(sessionId, tag, label, channel) {
-      let connection = connections.get(sessionId);
-      if (!connection) {
-        connection = {
-          id: sessionId,
-          tag,
-          // The longest this peer has ever taken between two echoes. Its own
-          // schedule, measured rather than assumed - a hidden tab answers
-          // about once a second because the browser throttles the timer.
-          echoIntervalMs: 0,
-          channels: new Map(),
-          seq: 0,
-          sentAt: 0,
-          seen: new Map(),
-          echoAt: 0,
-          echoes: 0,
-          // The far end's own transport-level received total, and its value at
-          // the previous tick. Null until a browser that reports it has echoed.
-          /** @type {number | null} */
-          peerBytes: null,
-          // Requests the far end says it is still waiting for, as last reported.
-          peerPending: null,
-          /** @type {number | null} */
-          peerBytesAtTick: null,
-          // Sum of per-channel bytes (report.channels[].bytes). Transport's
-          // bytesReceived advances on SACKs (140 B/s on a wedge) and gives a
-          // false "advancing" every other tick — channel bytes are flat on a wedge.
-          /** @type {number | null} */
-          peerChannelBytes: null,
-          /** @type {number | null} */
-          peerChannelBytesAtTick: null,
-          // The far end's own event-loop delay and tab state, as last reported.
-          // The lag is a term in every allowance below; the tab state is only
-          // printed, so that a wide allowance can be read back to its cause.
-          /** @type {number | null} */
-          peerLoopLagMs: null,
-          /** @type {string | null} */
-          peerVisibility: null,
-          verdict: "",
-          reportedAt: 0,
-          lastSeenAdvanceAt: 0,
-          longestHealthySeenGapMs: 0,
-          sentAtBySeq: new Map(),
-          // THE TWO CLOCKS, separated by the exchange rather than assumed to
-          // agree. Every report carries when each channel's newest probe
-          // arrived at the peer and when the report itself left; with the time
-          // this proxy stamped into that probe, and the time the report
-          // arrives, that is the four timestamps an offset is computed from.
-          //
-          // The estimate is kept from the report with the SMALLEST round trip
-          // seen, because that is the one that queued the least — and the two
-          // directions here are as unequal as they get, film one way and almost
-          // nothing the other, which is exactly where an offset taken from a
-          // busy moment is wrong by half the difference.
-          /** @type {number | null} */
-          clockOffsetMs: null,
-          /** @type {number} */
-          offsetFromRoundTripMs: Number.POSITIVE_INFINITY,
-          /** How long the newest probe took to reach the peer, by channel. @type {Map<string, number>} */
-          oneWayMs: new Map(),
-          probeCaptureStarted: false,
-          timer: null
-        };
-        connections.set(sessionId, connection);
-      }
-      connection.channels.set(channel, label);
-      if (connection.timer === null) {
-        const held = connection;
-        connection.timer = setInterval(() => tick(held), intervalMs);
-        // The probe must never be the reason a process stays alive.
-        if (typeof connection.timer.unref === "function") {
-          connection.timer.unref();
-        }
-      }
-    },
-
-    detach(sessionId, channel) {
-      const connection = connections.get(sessionId);
-      if (!connection) {
-        return;
-      }
-      connection.channels.delete(channel);
-      if (connection.channels.size === 0) {
-        if (connection.timer !== null) {
-          clearInterval(connection.timer);
-          connection.timer = null;
-        }
-        if (connections.get(sessionId) === connection) {
-          connections.delete(sessionId);
-        }
-      }
-    },
-
-    noteEcho(sessionId, echo) {
-      const connection = connections.get(sessionId);
-      if (!connection || !echo || typeof echo !== "object") {
-        return;
-      }
-      const now = Date.now();
-      const seen = echo.seen;
-      if (seen && typeof seen === "object") {
-        let advanced = false;
-        for (const [label, value] of Object.entries(seen)) {
-          if (!Number.isInteger(value)) {
-            continue;
-          }
-          const previous = connection.seen.get(label);
-          if (!Number.isInteger(previous) || value > previous) {
-            advanced = true;
-          }
-          connection.seen.set(label, value);
-        }
-        // What a wedge shows is this counter frozen, not merely behind — see
-        // probeWedgeIsCertain. The gap since the last time ANY label moved is
-        // this connection's own answer to "how long may a healthy report take
-        // to arrive", recorded only while nothing is currently flagged (the
-        // same guard `longestHealthyFlatMs` uses): a stretch already under
-        // suspicion must not teach the detector to tolerate it.
-        if (advanced) {
-          if (connection.lastSeenAdvanceAt !== 0 && !connection.probeCaptureStarted) {
-            const gap = now - connection.lastSeenAdvanceAt;
-            if (gap > connection.longestHealthySeenGapMs) {
-              connection.longestHealthySeenGapMs = gap;
-            }
-          }
-          connection.lastSeenAdvanceAt = now;
-        }
-      }
-      // THE FOUR TIMESTAMPS. `t1` is when this proxy sent the probe the peer
-      // is reporting, `t2` when the peer received it, `t3` when the peer sent
-      // this report, `t4` is now. From them the round trip is
-      // `(t4 - t1) - (t3 - t2)` — the peer's own thinking time removed — and
-      // the difference between the two clocks is `((t2 - t1) + (t3 - t4)) / 2`.
-      //
-      // The offset is kept from the report whose round trip was SMALLEST: that
-      // arithmetic assumes the two directions are equally quick, which here
-      // they are not — film one way, almost nothing the other — and the least
-      // queued sample is the one where the assumption costs least. A fresh
-      // minimum replaces the estimate, so a clock that jumps is followed within
-      // a few reports rather than believed for ever.
-      const peerSentAt = Number(echo?.sentAt);
-      const seenAt = echo?.seenAt;
-      if (Number.isFinite(peerSentAt) && seenAt && typeof seenAt === "object") {
-        for (const [label, peerSawAtRaw] of Object.entries(seenAt)) {
-          const peerSawAt = Number(peerSawAtRaw);
-          const newestSeenNumber = connection.seen.get(label);
-          const weSentAt = Number.isInteger(newestSeenNumber)
-            ? connection.sentAtBySeq.get(newestSeenNumber)
-            : undefined;
-          if (!Number.isFinite(peerSawAt) || !Number.isFinite(weSentAt)) {
-            continue;
-          }
-          const roundTrip = (now - weSentAt) - (peerSentAt - peerSawAt);
-          if (roundTrip >= 0 && roundTrip < connection.offsetFromRoundTripMs) {
-            connection.offsetFromRoundTripMs = roundTrip;
-            connection.clockOffsetMs = ((peerSawAt - weSentAt) + (peerSentAt - now)) / 2;
-          }
-          if (connection.clockOffsetMs !== null) {
-            // What the probe itself took, one way, with the clocks reconciled.
-            // Negative only if the offset is stale, and then it says so rather
-            // than being hidden.
-            connection.oneWayMs.set(label, peerSawAt - weSentAt - connection.clockOffsetMs);
-          }
-        }
-      }
-      // What the far end says it has received at the transport level. It is the
-      // one figure that separates a backlog from a stopped association, and it
-      // arrives on the direction that goes on working through a freeze.
-      // How many requests the far end says it is still waiting for. Theirs to
-      // count: this end knows what it has answered, not what reached them.
-      const peerPending = Number(echo?.report?.pending);
-      connection.peerPending = Number.isFinite(peerPending) && peerPending >= 0 ? peerPending : null;
-      const peerBytes = Number(echo?.report?.transportBytesReceived);
-      if (Number.isFinite(peerBytes) && peerBytes >= 0) {
-        connection.peerBytes = peerBytes;
-      }
-      // Sum of per-channel bytes. Transport's bytesReceived advances on SACKs
-      // (140 B/s on a wedge) and gives a false "advancing" every other tick —
-      // channel bytes are flat on a wedge, so they are the correct signal.
-      const peerChannels = echo?.report?.channels;
-      if (peerChannels && typeof peerChannels === "object") {
-        let bytesAcrossChannels = 0;
-        let hasChannelReport = false;
-        for (const channelReport of Object.values(peerChannels)) {
-          if (channelReport && Number.isFinite(channelReport.bytes)) {
-            bytesAcrossChannels += channelReport.bytes;
-            hasChannelReport = true;
-          }
-        }
-        if (hasChannelReport) {
-          connection.peerChannelBytes = bytesAcrossChannels;
-        }
-      }
-      // How far behind the far end's own event loop is running. A browser that
-      // cannot run its timers cannot answer a probe, and every allowance here
-      // is a wait the answer has to fit inside — so this is a term in the
-      // arithmetic, not a note. It is the peer's own measurement, taken on the
-      // peer, and it arrives on the direction that survives a freeze.
-      const peerLoopLag = Number(echo?.report?.loopLagMs);
-      if (Number.isFinite(peerLoopLag) && peerLoopLag >= 0) {
-        connection.peerLoopLagMs = peerLoopLag;
-      }
-      const peerVisibility = echo?.report?.visibility;
-      if (typeof peerVisibility === "string" && peerVisibility.length > 0) {
-        connection.peerVisibility = peerVisibility;
-      }
-      if (connection.echoAt !== 0) {
-        const sinceLast = now - connection.echoAt;
-        if (sinceLast > connection.echoIntervalMs) {
-          connection.echoIntervalMs = sinceLast;
-        }
-      }
-      connection.echoAt = now;
-      connection.echoes += 1;
-    },
-
-    dispose() {
-      for (const connection of connections.values()) {
-        if (connection.timer !== null) {
-          clearInterval(connection.timer);
-          connection.timer = null;
-        }
-      }
-      connections.clear();
-    }
-  };
-}
+/**
+ * @file Numbered delivery probes, and the verdict they make possible.
+ *
+ * The delivery freeze (roadmap item 11) looks identical from the proxy in two
+ * cases that need opposite fixes: usrsctp stopped transmitting, or the browser
+ * closed its receive window because the page stopped draining the channel. The
+ * proxy's own counters cannot separate them — libdatachannel's `bytesSent`
+ * counts bytes ACCEPTED into usrsctp, not bytes put on the wire — so the
+ * reading has to come from the far end.
+ *
+ * Two facts make that cheap. The reverse direction keeps working throughout the
+ * freeze (browser to proxy requests arrive and are answered for the whole
+ * episode, 88 min in the 2026-08-24 case), so the browser can always report.
+ * And SCTP orders per STREAM, so a probe on a channel opened UNORDERED and
+ * WITHOUT retransmission passes head-of-line blocking in another stream but
+ * neither a closed receive window nor a transmitter that stopped.
+ *
+ * So: number a probe every {@link PROBE_INTERVAL_MS} on every channel of the
+ * connection, have the browser echo back the highest number it has seen on
+ * each, and read the gaps:
+ *
+ *   every channel current                to flowing
+ *   ordered behind, unreliable current   to a retransmission stuck in a stream
+ *   both behind, echoes still arriving   to the association stopped transmitting
+ *   no echo at all                       to the reverse direction went too
+ *
+ * The verdict is computed from the gaps, not chosen, and every line prints the
+ * numbers that produced it.
+ */
+
+/**
+ * How often a probe is numbered and sent on every channel.
+ *
+ * Half a second, because the transport heartbeat is five and that was the whole
+ * resolution the 2026-08-24 episode had: the onset could be placed no closer
+ * than the five seconds between two lines. Sending is cheap - a probe is a few
+ * dozen bytes - and the interval is NOT the verdict: a healthy burst can hold a
+ * probe up behind queued data, so how many probes may be outstanding is worked
+ * out per channel by {@link allowedGap} from the bytes queued ahead of the
+ * probe and the rate they are leaving at.
+ */
+export const PROBE_INTERVAL_MS = 500;
+
+/**
+ * How many probes may be outstanding on a channel before it counts as behind.
+ *
+ * DERIVED per channel, not chosen. A probe is handed to the same association as
+ * the data, and SCTP orders per stream but SCHEDULES per association: one
+ * congestion window, one send buffer. So a probe waits for whatever is queued
+ * ahead of it whichever channel it rides on - the unordered one included - and
+ * a fixed threshold cannot tell that wait from a stopped association. Measured
+ * 2026-08-26: `association-stopped` printed with all three channels at gap 4-7
+ * while 110-150 Mbps crossed that same association and 7.34 GB went through it
+ * without a single failure.
+ *
+ * What the wait costs is arithmetic on two measured quantities: the bytes
+ * queued ahead of the probe, and the rate at which this connection is getting
+ * bytes away. Add one round trip for the echo to come back. A probe is behind
+ * only when it is later than that.
+ *
+ * There is a third term, and leaving it out made this worse than the constant
+ * it replaced. The browser answers on ITS own schedule, not ours: it batches
+ * what it has seen and echoes on a timer, and that timer is throttled to about
+ * once a second whenever the tab is hidden. So with an empty queue the
+ * allowance collapsed to one probe and the bound to half a second, while echoes
+ * legitimately arrived every second - measured 2026-08-27, 67 `association-
+ * stopped` and 66 `reverse-direction-gone` against 84 `flowing` on a connection
+ * carrying 3.4 MB/s with every queue at zero. The peer's own cadence is
+ * measurable on the same connection, so it is measured and added rather than
+ * assumed.
+ *
+ * There is a fourth term, and without it the third one arrives too late to
+ * help. `echoIntervalMs` is the widest gap between two echoes this connection
+ * has SHOWN, so it can only grow after a late echo has landed — and a browser
+ * whose tab has just been hidden goes quiet before it has taught us anything.
+ * The browser measures its own event-loop delay and puts it in every echo, so
+ * the delay is known BEFORE the silence rather than after it: an echo timer
+ * set for half a second cannot fire until the loop runs, and a loop reported
+ * as 4297 ms behind cannot answer sooner than that. Field 2026-09-03: a viewer
+ * paused at 15:24:21, the tab went hidden at 15:24:35, `loopLag` climbed
+ * 681 → 1881 → 4297 → 5957 ms, and at 15:26:05 the probes read `gap 12 of 11`
+ * and printed `association-stopped` on a connection that was `flowing` again
+ * seven seconds later. Adding the peer's own reported lag makes that allowance
+ * 15 instead of 11, and the gap of 12 is then what it was — a browser whose
+ * timers are frozen, not an association that stopped.
+ *
+ * @param {{ queuedBytes: number, bytesPerSecond: number, rttMs: number, echoIntervalMs?: number, peerLoopLagMs?: number, intervalMs?: number }} state
+ * @returns {number | null} Probes that may legitimately be outstanding, or null
+ *   when no rate has been measured yet and nothing can be said.
+ */
+export function allowedGap({
+  queuedBytes,
+  bytesPerSecond,
+  rttMs,
+  echoIntervalMs = 0,
+  peerLoopLagMs = 0,
+  intervalMs = PROBE_INTERVAL_MS
+}) {
+  if (!(bytesPerSecond > 0) || !(intervalMs > 0)) {
+    return null;
+  }
+  // At least one: a probe sent and not yet echoed is the ordinary state.
+  return Math.max(1, Math.ceil(allowedWaitMs({ queuedBytes, bytesPerSecond, rttMs, echoIntervalMs, peerLoopLagMs }) / intervalMs));
+}
+
+/**
+ * How long a probe may legitimately take to be reported back, in milliseconds.
+ *
+ * THE QUANTITY {@link allowedGap} COMPUTES AND THEN THROWS AWAY by dividing it
+ * into probes. Probes are the wrong unit and always were: the same probe goes
+ * down every channel INCLUDING the one carrying the film, and SCTP schedules
+ * per association, so a probe waits behind queued video exactly as a segment
+ * does. What is behind is then not the peer's answer but the probe itself —
+ * and the honest measure of that is the age of the newest probe the peer has
+ * seen, which this is compared against.
+ *
+ * Four terms, every one measured: the queue's own drain time at the rate this
+ * connection is achieving, the crossing, the peer's reporting cadence, and how
+ * late the peer's event loop is running.
+ *
+ * @param {{ queuedBytes: number, bytesPerSecond: number, rttMs: number,
+ *   echoIntervalMs?: number, peerLoopLagMs?: number }} state
+ * @returns {number}
+ */
+export function allowedWaitMs({ queuedBytes, bytesPerSecond, rttMs, echoIntervalMs = 0, peerLoopLagMs = 0 }) {
+  if (!(bytesPerSecond > 0)) {
+    return 0;
+  }
+  const drainMs = (Math.max(queuedBytes, 0) / bytesPerSecond) * 1000;
+  return drainMs + Math.max(rttMs, 0) + Math.max(echoIntervalMs, 0) + Math.max(peerLoopLagMs, 0);
+}
+
+/**
+ * How far back the send times of probes are kept.
+ *
+ * A probe older than the point at which the reverse direction is called gone
+ * can say nothing further, and that bound is itself derived per connection —
+ * this is the outer edge of it, kept so the map cannot grow with the session.
+ */
+const PROBE_HISTORY_MS = 10 * 60 * 1000;
+
+/** The label the browser gives the unordered, non-retransmitting channel. */
+export const UNRELIABLE_LABEL = "proxy-fast";
+
+/**
+ * How old an echo may be before the reverse direction counts as gone, when
+ * nothing better can be derived.
+ *
+ * Used only while no rate has been measured. Otherwise the caller passes
+ * `echoStaleMs`, worked out the same way as {@link allowedGap}: the browser
+ * only echoes a probe it has RECEIVED, so an echo waits behind our own queue
+ * exactly as the probe did.
+ */
+const ECHO_STALE_FALLBACK_MS = 5_000;
+
+/** How often the probe state is written to the log while nothing changes. */
+const REPORT_INTERVAL_MS = 5_000;
+
+/**
+ * Whether the `seen` counter has stopped advancing for longer than this
+ * connection's own history says a healthy gap between two advances ever
+ * takes.
+ *
+ * The `association-stopped` verdict alone is not enough to act on: a
+ * connection can sit BEHIND by a bounded, roughly constant amount for
+ * minutes (measured 2026-08-28, session on a backgrounded tab — gap held at
+ * 6-7 probes for 95+ seconds while `seen` kept climbing right along with
+ * `sent`) without anything being wrong. What a true wedge shows instead,
+ * measured the same day against a session already known to be one
+ * (`d85ae4f5`): `seen` FROZEN at one value for over a minute while `sent`
+ * climbs unbounded. So the question is not "is there a gap" but "has the
+ * highest-seen number stopped moving at all, for longer than it has ever
+ * legitimately taken this connection to report an advance" — the same shape
+ * as {@link wedgeIsCertain} in `transport/data-channel-handler.js`, applied to the
+ * probe's own counter instead of the transport's byte counter.
+ *
+ * THE FLOOR IS NOT ONE PROBE INTERVAL. "The longest gap this connection has
+ * ever shown" is nothing at all in its first minutes, and the floor under it
+ * was a single interval — 500 ms — so any two-second quiet stretch while the
+ * browser filled its cushion read as longer than anything healthy ever seen.
+ * Field 2026-09-11: two captures of 180 s each, triggered at `wedged 1s` and
+ * `wedged 2s`, on a connection with `rtt=5ms`, every queue at 0 B and the
+ * viewer watching; 65 MB of them were left on the addon's disk.
+ *
+ * What replaces it is the connection's own answer to how long a legitimate
+ * report may take: the peer reports twice a second, the report has to cross,
+ * and the peer's own event loop may be late — all three measured, all three
+ * already summed for `echoStaleMs`. A stretch shorter than that proves nothing
+ * whatever about the association.
+ *
+ * @param {{ stuckForMs: number, longestHealthySeenGapMs: number, intervalMs?: number,
+ *   legitimateReportMs?: number }} state
+ * @returns {{ isCertain: boolean, needMs: number }}
+ */
+export function probeWedgeIsCertain({
+  stuckForMs,
+  longestHealthySeenGapMs,
+  intervalMs = PROBE_INTERVAL_MS,
+  legitimateReportMs = 0
+}) {
+  const floorMs = Math.max(intervalMs, Number.isFinite(legitimateReportMs) ? legitimateReportMs : 0);
+  const needMs = Math.max(longestHealthySeenGapMs, floorMs);
+  return { isCertain: stuckForMs >= needMs, needMs };
+}
+
+/**
+ * One connection's probe state.
+ *
+ * @typedef {Object} ProbeConnection
+ * @property {string} tag
+ * @property {Map<import('node-datachannel').DataChannel, string>} channels
+ * @property {number} seq        - Highest probe number sent.
+ * @property {number} sentAt     - When that probe was sent.
+ * @property {Map<string, number>} seen - Label to the highest number the browser reported.
+ * @property {number} echoAt     - When the last echo arrived (0 = never).
+ * @property {number} echoes     - How many echoes have arrived.
+ * @property {string} verdict    - Last verdict reported, so a change is logged at once.
+ * @property {number} reportedAt - When the state was last written to the log.
+ * @property {number} lastSeenAdvanceAt - When any label's `seen` value last increased (0 = never yet).
+ * @property {number} longestHealthySeenGapMs - The longest gap between two advances this connection has shown while not flagged as a wedge.
+ * @property {number | null} peerBytes - The far end's transport-level received total, as last reported.
+ * @property {number | null} peerBytesAtTick - The same, as it stood at the previous tick.
+ * @property {number | null} peerLoopLagMs - The far end's own event-loop delay, as last reported.
+ * @property {string | null} peerVisibility - Whether the far end's tab is visible or hidden, as last reported.
+ * @property {boolean} probeCaptureStarted - One evidence-gathering attempt per wedge; reset once `seen` advances again.
+ * @property {ReturnType<typeof setInterval> | null} timer
+ */
+
+/**
+ * Read the gaps and say what they mean.
+ *
+ * Exported so the rule is testable without a connection: the same numbers
+ * always produce the same word.
+ *
+ * `allowed` carries, per channel label, how many probes may legitimately be
+ * outstanding right now — {@link allowedGap} computes it from that channel's
+ * own queue and the connection's measured rate. A label with no entry, or an
+ * entry of null, cannot be judged: with no rate measured there is nothing to
+ * divide the queue by, and the verdict says that rather than inventing one.
+ *
+ * `peerBytesAdvancing` is the fact that outranks every gap here. The gaps are
+ * counted against an allowance whose only load-dependent term is OUR OWN
+ * queue — and that queue is empty by construction, because it drains the moment
+ * libdatachannel accepts the bytes, whether or not usrsctp then puts them on
+ * the wire. So a browser filling its cushion as fast as the link allows shows
+ * an empty queue, a small allowance and a large gap, which reads exactly like a
+ * stopped association. Measured 2026-08-28: four `association-stopped` in the
+ * first two minutes of a healthy session, on a connection whose every queue was
+ * at 0 B and whose viewer never saw the picture stop. Deepening the browser's
+ * cushion from 30 s to 120 s (roadmap item 4) made the burst four times longer
+ * and the false verdict correspondingly likelier.
+ *
+ * What separates the two is not the size of the backlog — it is large in both —
+ * but whether bytes are still arriving. The browser reports its own
+ * transport-level received total with every echo, so the question is answered
+ * by a counter rather than by a threshold: while that total is advancing, bytes
+ * ARE crossing and no verdict of a stopped association can stand, however far
+ * behind the probes are. `null` where the far end does not report it, and then
+ * the rule falls back to what it says without the term.
+ *
+ * `peerLoopLagMs` and `peerVisibility` are printed but never compared against
+ * anything here: they are what made the allowance as wide as it is, and a
+ * reader of the log cannot check that arithmetic unless the terms are on the
+ * same line as the result.
+ *
+ * @param {{ seq: number, seen: Map<string, number> | Record<string, number>, labels: string[], echoes: number, echoAgeMs: number | null, allowed?: Map<string, number | null> | Record<string, number | null>, echoStaleMs?: number, peerBytesAdvancing?: boolean | null, peerLoopLagMs?: number | null, peerVisibility?: string | null, peerPending?: number | null }} state
+ * @returns {{ verdict: string, detail: string }}
+ */
+export function readProbeState(state) {
+  const seenOf = (label) =>
+    state.seen instanceof Map ? state.seen.get(label) : state.seen?.[label];
+  /**
+   * @param {unknown} source
+   * @param {string} label
+   * @returns {number | null}
+   */
+  const timeOf = (source, label) => {
+    const value = source instanceof Map ? source.get(label) : /** @type {any} */ (source)?.[label];
+    return Number.isFinite(value) ? Number(value) : null;
+  };
+  const allowedOf = (label) => {
+    const source = state.allowed;
+    const value = source instanceof Map ? source.get(label) : source?.[label];
+    return Number.isInteger(value) ? Number(value) : null;
+  };
+  const parts = [];
+  let isOrderedBehind = false;
+  let isUnreliableBehind = false;
+  let isUnreliableKnown = false;
+  let isJudgeable = false;
+  for (const label of state.labels) {
+    const seen = seenOf(label);
+    const gap = Number.isInteger(seen) ? state.seq - Number(seen) : null;
+    const allowance = allowedOf(label);
+    // BY TIME WHERE IT IS KNOWN. The count is what the line prints, because it
+    // is what a reader recognises; what decides is a time.
+    //
+    // The measured one-way delay first: it is the forward direction itself,
+    // with the two clocks reconciled from the exchange. The age of the newest
+    // reported probe is the fallback, and it is larger than the thing itself by
+    // the peer's reporting cadence and the way back — which is why its own
+    // allowance carries both of those and the one-way allowance carries
+    // neither.
+    const oneWay = timeOf(state.oneWayMs, label);
+    const oneWayAllowed = timeOf(state.allowedOneWayMs, label);
+    const lagMs = oneWay !== null && oneWayAllowed !== null ? oneWay : timeOf(state.behindMs, label);
+    const mayWaitMs = oneWay !== null && oneWayAllowed !== null
+      ? oneWayAllowed
+      : timeOf(state.allowedWaitMs, label);
+    const lagText = lagMs === null || mayWaitMs === null
+      ? ""
+      : ` ${Math.round(lagMs)}ms of ${Math.round(mayWaitMs)}ms${oneWay !== null ? " one way" : ""}`;
+    parts.push(`${label}=${seen ?? "?"}(gap ${gap ?? "?"} of ${allowance ?? "?"}${lagText})`);
+    if (allowance === null) {
+      continue;
+    }
+    isJudgeable = true;
+    const isBehind = lagMs !== null && mayWaitMs !== null
+      ? lagMs > mayWaitMs
+      : gap === null || gap > allowance;
+    if (label === UNRELIABLE_LABEL) {
+      isUnreliableKnown = true;
+      isUnreliableBehind = isBehind;
+    } else if (isBehind) {
+      isOrderedBehind = true;
+    }
+  }
+  const isPeerAdvancing = state.peerBytesAdvancing === true;
+  const detail =
+    `sent=${state.seq} ${parts.join(" ")} ` +
+    `echoAge=${state.echoAgeMs === null ? "never" : `${state.echoAgeMs}ms`}` +
+    (state.peerBytesAdvancing === null || state.peerBytesAdvancing === undefined
+      ? ""
+      : ` peerBytes=${isPeerAdvancing ? "advancing" : "still"}`) +
+    (Number.isFinite(state.peerLoopLagMs) ? ` peerLoopLag=${Math.round(Number(state.peerLoopLagMs))}ms` : "") +
+    (state.peerVisibility ? ` peerTab=${state.peerVisibility}` : "") +
+    // WHAT THE PEER SAYS IT IS STILL WAITING FOR. The far end has sent this
+    // twice a second since the probe echo was built and nothing read it, while
+    // every verdict here was reasoned out from this side alone. It is evidence
+    // and not a verdict: the receiver declares its own transport lost, because
+    // it knows first — on 2026-09-06 the browser said so sixteen minutes before
+    // this end did.
+    (Number.isFinite(state.peerPending) ? ` peerWaiting=${Number(state.peerPending)}` : "");
+
+  if (state.echoes === 0) {
+    return { verdict: "no-echo-yet", detail };
+  }
+  if (!isJudgeable) {
+    return { verdict: "no-rate-yet", detail };
+  }
+  const staleAfterMs = Number.isFinite(state.echoStaleMs) && state.echoStaleMs > 0
+    ? state.echoStaleMs
+    : ECHO_STALE_FALLBACK_MS;
+  if (state.echoAgeMs !== null && state.echoAgeMs > staleAfterMs) {
+    return { verdict: "reverse-direction-gone", detail };
+  }
+  if (!isOrderedBehind && !(isUnreliableKnown && isUnreliableBehind)) {
+    return { verdict: "flowing", detail };
+  }
+  // Bytes are still arriving at the far end. Whatever the probe gaps say, this
+  // association has not stopped — the probes are behind a backlog, which is
+  // what filling a cushion looks like from here.
+  if (isPeerAdvancing) {
+    return { verdict: "flowing", detail };
+  }
+  if (isOrderedBehind && isUnreliableKnown && !isUnreliableBehind) {
+    return { verdict: "stream-stuck", detail };
+  }
+  if (isOrderedBehind) {
+    return {
+      verdict: isUnreliableKnown ? "association-stopped" : "ordered-behind-no-comparison",
+      detail
+    };
+  }
+  return { verdict: "unreliable-behind-only", detail };
+}
+
+/**
+ * Create the probe service. One instance serves every session.
+ *
+ * @param {Object} options
+ * @param {(message: string) => void} options.log
+ * @param {number} [options.intervalMs]
+ * @param {(sessionId: string) => object | null} [options.getTransportSnapshot]
+ *   Needed only to hand the witness a remote endpoint when this probe is the
+ *   one declaring a wedge.
+ * @param {{ maybeCapture: (trigger: object) => boolean }} [options.witness]
+ * @param {{ maybeRead: (reasonText: string) => boolean }} [options.usrsctpState]
+ * @returns {{
+ *   attach: (sessionId: string, tag: string, label: string, channel: import('node-datachannel').DataChannel) => void,
+ *   detach: (sessionId: string, channel: import('node-datachannel').DataChannel) => void,
+ *   noteEcho: (sessionId: string, echo: object) => void,
+ *   dispose: () => void
+ * }}
+ */
+export function createDeliveryProbe({
+  log,
+  intervalMs = PROBE_INTERVAL_MS,
+  readDelivery,
+  getTransportSnapshot,
+  witness,
+  usrsctpState
+}) {
+  /** @type {Map<string, ProbeConnection>} */
+  const connections = new Map();
+
+  /**
+   * Send this tick's probe on every channel of one connection, then report.
+   *
+   * @param {ProbeConnection} connection
+   * @returns {void}
+   */
+  function tick(connection) {
+    const now = Date.now();
+    connection.seq += 1;
+    connection.sentAt = now;
+    // WHEN each probe went out, so that what the peer reports can be read as a
+    // time rather than as a count of probes. Pruned to the oldest probe any
+    // judgement could still be about: once a probe is older than the point at
+    // which the reverse direction is called gone, its age says nothing further.
+    connection.sentAtBySeq.set(connection.seq, now);
+    for (const [probeNumber, sentAt] of connection.sentAtBySeq) {
+      if (now - sentAt > PROBE_HISTORY_MS) {
+        connection.sentAtBySeq.delete(probeNumber);
+      } else {
+        break;
+      }
+    }
+    const message = JSON.stringify({ type: "probe", seq: connection.seq, sentAt: now });
+    for (const channel of connection.channels.keys()) {
+      try {
+        channel.sendMessage(message);
+      } catch {
+        // A channel closing between the check and the send is ordinary.
+      }
+    }
+
+    // What this connection is getting away, and how far behind it therefore
+    // sits. Both come from the send-queue watcher, which measures them anyway.
+    const delivery = typeof readDelivery === "function" ? readDelivery(connection.id) : null;
+    const bytesPerSecond = Number(delivery?.bytesPerSecond) || 0;
+    const rttMs = Number(delivery?.rttMs) || 0;
+    /** @type {Map<string, number | null>} */
+    const allowed = new Map();
+    /** How long each channel's newest unreported probe may legitimately be. */
+    const allowedWait = new Map();
+    /** The same, for the measured one-way time where the clocks are reconciled. */
+    const allowedOneWay = new Map();
+    for (const [channel, label] of connection.channels) {
+      let queuedBytes = 0;
+      try {
+        queuedBytes = typeof channel.bufferedAmount === "function" ? channel.bufferedAmount() : 0;
+      } catch {
+        queuedBytes = 0;
+      }
+      const allowance = allowedGap({
+        queuedBytes,
+        bytesPerSecond,
+        rttMs,
+        echoIntervalMs: connection.echoIntervalMs,
+        peerLoopLagMs: connection.peerLoopLagMs ?? 0,
+        intervalMs
+      });
+      const waitMs = allowedWaitMs({
+        queuedBytes,
+        bytesPerSecond,
+        rttMs,
+        echoIntervalMs: connection.echoIntervalMs,
+        peerLoopLagMs: connection.peerLoopLagMs ?? 0
+      });
+      const heldWait = allowedWait.get(label);
+      if (!Number.isFinite(heldWait) || waitMs > heldWait) {
+        allowedWait.set(label, waitMs);
+      }
+      // What a probe may take ONE WAY: the queue's own drain time and half the
+      // crossing. Neither the peer's reporting cadence nor its loop delay
+      // belongs here — with the clocks reconciled they are no longer in the
+      // measurement they would have to be allowed for.
+      const oneWayAllowed = allowedWaitMs({ queuedBytes, bytesPerSecond, rttMs: rttMs / 2 });
+      const heldOneWay = allowedOneWay.get(label);
+      if (!Number.isFinite(heldOneWay) || oneWayAllowed > heldOneWay) {
+        allowedOneWay.set(label, oneWayAllowed);
+      }
+      // Several channels can carry one label only in malformed cases; the
+      // larger allowance is the safer of the two.
+      const held = allowed.get(label);
+      if (allowance !== null && (!Number.isInteger(held) || allowance > held)) {
+        allowed.set(label, allowance);
+      } else if (!allowed.has(label)) {
+        allowed.set(label, allowance);
+      }
+    }
+    const widest = [...allowed.values()].reduce(
+      (most, value) => (Number.isInteger(value) && value > most ? value : most),
+      0
+    );
+    // Advancing SINCE THE PREVIOUS TICK, not since the connection began: the
+    // question is whether bytes are crossing now. Prefer channel bytes —
+    // transport's bytesReceived advances on SACKs (140 B/s on a wedge) and
+    // gives a false "advancing" every other tick, while channel bytes are flat.
+    let peerBytesAdvancing = null;
+    if (connection.peerChannelBytes !== null) {
+      peerBytesAdvancing =
+        connection.peerChannelBytesAtTick === null ||
+        connection.peerChannelBytes > connection.peerChannelBytesAtTick;
+      connection.peerChannelBytesAtTick = connection.peerChannelBytes;
+      // keep transport's tick in sync for fallback logging
+      connection.peerBytesAtTick = connection.peerBytes;
+    } else if (connection.peerBytes !== null) {
+      peerBytesAdvancing =
+        connection.peerBytesAtTick === null || connection.peerBytes > connection.peerBytesAtTick;
+      connection.peerBytesAtTick = connection.peerBytes;
+    }
+    const { verdict, detail } = readProbeState({
+      seq: connection.seq,
+      seen: connection.seen,
+      peerBytesAdvancing,
+      peerPending: connection.peerPending,
+      labels: [...new Set(connection.channels.values())],
+      echoes: connection.echoes,
+      echoAgeMs: connection.echoAt === 0 ? null : now - connection.echoAt,
+      peerLoopLagMs: connection.peerLoopLagMs,
+      peerVisibility: connection.peerVisibility,
+      allowed,
+      // WHAT IS BEHIND, IN TIME. For each channel: how old the newest probe the
+      // peer has reported seeing is. The same probe goes down every channel
+      // including the one carrying the film, and SCTP schedules per
+      // association, so a probe waits behind queued video exactly as a segment
+      // does — which means a count of outstanding probes measures the queue,
+      // not the association. This is the queue's own time, against the time the
+      // queue is allowed to take.
+      behindMs: Object.fromEntries(
+        [...connection.seen.entries()].map(([label, newestSeenNumber]) => {
+          const sentAt = connection.sentAtBySeq.get(newestSeenNumber);
+          return [label, Number.isFinite(sentAt) ? now - sentAt : null];
+        })
+      ),
+      allowedWaitMs: Object.fromEntries(allowedWait.entries()),
+      // The measured forward delay, where the two clocks have been reconciled.
+      // Preferred over the age above because it is the thing itself: the age
+      // also carries the peer's reporting cadence and the way back.
+      oneWayMs: Object.fromEntries(connection.oneWayMs.entries()),
+      allowedOneWayMs: Object.fromEntries(allowedOneWay.entries()),
+      // Same arithmetic for the echo's own age: the peer cannot answer sooner
+      // than its own cadence allows, nor sooner than its own event loop runs.
+      echoStaleMs:
+        widest > 0
+          ? widest * intervalMs +
+            rttMs +
+            connection.echoIntervalMs +
+            Math.max(connection.peerLoopLagMs ?? 0, 0)
+          : 0
+    });
+    if (verdict !== connection.verdict || now - connection.reportedAt >= REPORT_INTERVAL_MS) {
+      connection.verdict = verdict;
+      connection.reportedAt = now;
+      log(`[dc-probe] ${connection.tag} ${verdict} — ${detail} at=${new Date(now).toISOString()}`);
+    }
+
+    // `association-stopped` alone is not certainty — see probeWedgeIsCertain.
+    // A connection that is merely lagging by a bounded amount reaches this
+    // verdict too (a hidden tab's own echo cadence, measured 2026-08-28), and
+    // `seen` keeps advancing right along with it. Only a `seen` value that has
+    // stopped moving ENTIRELY, for longer than this connection has ever shown
+    // as a legitimate gap, is the wedge this exists to catch.
+    const stuckForMs = connection.lastSeenAdvanceAt === 0 ? 0 : now - connection.lastSeenAdvanceAt;
+    if (verdict === "association-stopped") {
+      const { isCertain, needMs } = probeWedgeIsCertain({
+        stuckForMs,
+        longestHealthySeenGapMs: connection.longestHealthySeenGapMs,
+        // The same sum the echo's own staleness is judged by: the peer reports
+        // twice a second, the report has to cross, and the peer's loop may be
+        // late. Nothing shorter than that says anything.
+        // BOTH cadences, because they beat against each other: a probe sent
+        // just after the peer composed a report shows up only in the NEXT one.
+        // So the longest legitimate gap between two advances is one probe
+        // interval plus one report interval plus the crossing plus whatever
+        // the peer's own loop is behind by.
+        legitimateReportMs:
+          intervalMs +
+          connection.echoIntervalMs +
+          rttMs +
+          Math.max(connection.peerLoopLagMs ?? 0, 0)
+      });
+      if (isCertain && !connection.probeCaptureStarted) {
+        connection.probeCaptureStarted = true;
+        const reasonText =
+          `probe seen-counter unmoved ${Math.round(stuckForMs / 1000)}s against the ` +
+          `${(needMs / 1000).toFixed(1)}s this connection's own history says is legitimate`;
+        if (witness) {
+          const snapshot = getTransportSnapshot?.(connection.id) ?? null;
+          const started = witness.maybeCapture({
+            sessionId: connection.id,
+            tag: connection.tag,
+            label: "probe",
+            remote: snapshot?.remote ?? null,
+            queuedBytes: 0,
+            stuckForMs
+          });
+          if (!started) {
+            connection.probeCaptureStarted = false;
+          }
+        }
+        if (usrsctpState) {
+          usrsctpState.maybeRead(reasonText);
+        }
+      }
+    } else {
+      // Not association-stopped any more: whatever was flagged has cleared,
+      // and a later wedge on the same connection deserves its own attempt.
+      connection.probeCaptureStarted = false;
+    }
+  }
+
+  return {
+    attach(sessionId, tag, label, channel) {
+      let connection = connections.get(sessionId);
+      if (!connection) {
+        connection = {
+          id: sessionId,
+          tag,
+          // The longest this peer has ever taken between two echoes. Its own
+          // schedule, measured rather than assumed - a hidden tab answers
+          // about once a second because the browser throttles the timer.
+          echoIntervalMs: 0,
+          channels: new Map(),
+          seq: 0,
+          sentAt: 0,
+          seen: new Map(),
+          echoAt: 0,
+          echoes: 0,
+          // The far end's own transport-level received total, and its value at
+          // the previous tick. Null until a browser that reports it has echoed.
+          /** @type {number | null} */
+          peerBytes: null,
+          // Requests the far end says it is still waiting for, as last reported.
+          peerPending: null,
+          /** @type {number | null} */
+          peerBytesAtTick: null,
+          // Sum of per-channel bytes (report.channels[].bytes). Transport's
+          // bytesReceived advances on SACKs (140 B/s on a wedge) and gives a
+          // false "advancing" every other tick — channel bytes are flat on a wedge.
+          /** @type {number | null} */
+          peerChannelBytes: null,
+          /** @type {number | null} */
+          peerChannelBytesAtTick: null,
+          // The far end's own event-loop delay and tab state, as last reported.
+          // The lag is a term in every allowance below; the tab state is only
+          // printed, so that a wide allowance can be read back to its cause.
+          /** @type {number | null} */
+          peerLoopLagMs: null,
+          /** @type {string | null} */
+          peerVisibility: null,
+          verdict: "",
+          reportedAt: 0,
+          lastSeenAdvanceAt: 0,
+          longestHealthySeenGapMs: 0,
+          sentAtBySeq: new Map(),
+          // THE TWO CLOCKS, separated by the exchange rather than assumed to
+          // agree. Every report carries when each channel's newest probe
+          // arrived at the peer and when the report itself left; with the time
+          // this proxy stamped into that probe, and the time the report
+          // arrives, that is the four timestamps an offset is computed from.
+          //
+          // The estimate is kept from the report with the SMALLEST round trip
+          // seen, because that is the one that queued the least — and the two
+          // directions here are as unequal as they get, film one way and almost
+          // nothing the other, which is exactly where an offset taken from a
+          // busy moment is wrong by half the difference.
+          /** @type {number | null} */
+          clockOffsetMs: null,
+          /** @type {number} */
+          offsetFromRoundTripMs: Number.POSITIVE_INFINITY,
+          /** How long the newest probe took to reach the peer, by channel. @type {Map<string, number>} */
+          oneWayMs: new Map(),
+          probeCaptureStarted: false,
+          timer: null
+        };
+        connections.set(sessionId, connection);
+      }
+      connection.channels.set(channel, label);
+      if (connection.timer === null) {
+        const held = connection;
+        connection.timer = setInterval(() => tick(held), intervalMs);
+        // The probe must never be the reason a process stays alive.
+        if (typeof connection.timer.unref === "function") {
+          connection.timer.unref();
+        }
+      }
+    },
+
+    detach(sessionId, channel) {
+      const connection = connections.get(sessionId);
+      if (!connection) {
+        return;
+      }
+      connection.channels.delete(channel);
+      if (connection.channels.size === 0) {
+        if (connection.timer !== null) {
+          clearInterval(connection.timer);
+          connection.timer = null;
+        }
+        if (connections.get(sessionId) === connection) {
+          connections.delete(sessionId);
+        }
+      }
+    },
+
+    noteEcho(sessionId, echo) {
+      const connection = connections.get(sessionId);
+      if (!connection || !echo || typeof echo !== "object") {
+        return;
+      }
+      const now = Date.now();
+      const seen = echo.seen;
+      if (seen && typeof seen === "object") {
+        let advanced = false;
+        for (const [label, value] of Object.entries(seen)) {
+          if (!Number.isInteger(value)) {
+            continue;
+          }
+          const previous = connection.seen.get(label);
+          if (!Number.isInteger(previous) || value > previous) {
+            advanced = true;
+          }
+          connection.seen.set(label, value);
+        }
+        // What a wedge shows is this counter frozen, not merely behind — see
+        // probeWedgeIsCertain. The gap since the last time ANY label moved is
+        // this connection's own answer to "how long may a healthy report take
+        // to arrive", recorded only while nothing is currently flagged (the
+        // same guard `longestHealthyFlatMs` uses): a stretch already under
+        // suspicion must not teach the detector to tolerate it.
+        if (advanced) {
+          if (connection.lastSeenAdvanceAt !== 0 && !connection.probeCaptureStarted) {
+            const gap = now - connection.lastSeenAdvanceAt;
+            if (gap > connection.longestHealthySeenGapMs) {
+              connection.longestHealthySeenGapMs = gap;
+            }
+          }
+          connection.lastSeenAdvanceAt = now;
+        }
+      }
+      // THE FOUR TIMESTAMPS. `t1` is when this proxy sent the probe the peer
+      // is reporting, `t2` when the peer received it, `t3` when the peer sent
+      // this report, `t4` is now. From them the round trip is
+      // `(t4 - t1) - (t3 - t2)` — the peer's own thinking time removed — and
+      // the difference between the two clocks is `((t2 - t1) + (t3 - t4)) / 2`.
+      //
+      // The offset is kept from the report whose round trip was SMALLEST: that
+      // arithmetic assumes the two directions are equally quick, which here
+      // they are not — film one way, almost nothing the other — and the least
+      // queued sample is the one where the assumption costs least. A fresh
+      // minimum replaces the estimate, so a clock that jumps is followed within
+      // a few reports rather than believed for ever.
+      const peerSentAt = Number(echo?.sentAt);
+      const seenAt = echo?.seenAt;
+      if (Number.isFinite(peerSentAt) && seenAt && typeof seenAt === "object") {
+        for (const [label, peerSawAtRaw] of Object.entries(seenAt)) {
+          const peerSawAt = Number(peerSawAtRaw);
+          const newestSeenNumber = connection.seen.get(label);
+          const weSentAt = Number.isInteger(newestSeenNumber)
+            ? connection.sentAtBySeq.get(newestSeenNumber)
+            : undefined;
+          if (!Number.isFinite(peerSawAt) || !Number.isFinite(weSentAt)) {
+            continue;
+          }
+          const roundTrip = (now - weSentAt) - (peerSentAt - peerSawAt);
+          if (roundTrip >= 0 && roundTrip < connection.offsetFromRoundTripMs) {
+            connection.offsetFromRoundTripMs = roundTrip;
+            connection.clockOffsetMs = ((peerSawAt - weSentAt) + (peerSentAt - now)) / 2;
+          }
+          if (connection.clockOffsetMs !== null) {
+            // What the probe itself took, one way, with the clocks reconciled.
+            // Negative only if the offset is stale, and then it says so rather
+            // than being hidden.
+            connection.oneWayMs.set(label, peerSawAt - weSentAt - connection.clockOffsetMs);
+          }
+        }
+      }
+      // What the far end says it has received at the transport level. It is the
+      // one figure that separates a backlog from a stopped association, and it
+      // arrives on the direction that goes on working through a freeze.
+      // How many requests the far end says it is still waiting for. Theirs to
+      // count: this end knows what it has answered, not what reached them.
+      const peerPending = Number(echo?.report?.pending);
+      connection.peerPending = Number.isFinite(peerPending) && peerPending >= 0 ? peerPending : null;
+      const peerBytes = Number(echo?.report?.transportBytesReceived);
+      if (Number.isFinite(peerBytes) && peerBytes >= 0) {
+        connection.peerBytes = peerBytes;
+      }
+      // Sum of per-channel bytes. Transport's bytesReceived advances on SACKs
+      // (140 B/s on a wedge) and gives a false "advancing" every other tick —
+      // channel bytes are flat on a wedge, so they are the correct signal.
+      const peerChannels = echo?.report?.channels;
+      if (peerChannels && typeof peerChannels === "object") {
+        let bytesAcrossChannels = 0;
+        let hasChannelReport = false;
+        for (const channelReport of Object.values(peerChannels)) {
+          if (channelReport && Number.isFinite(channelReport.bytes)) {
+            bytesAcrossChannels += channelReport.bytes;
+            hasChannelReport = true;
+          }
+        }
+        if (hasChannelReport) {
+          connection.peerChannelBytes = bytesAcrossChannels;
+        }
+      }
+      // How far behind the far end's own event loop is running. A browser that
+      // cannot run its timers cannot answer a probe, and every allowance here
+      // is a wait the answer has to fit inside — so this is a term in the
+      // arithmetic, not a note. It is the peer's own measurement, taken on the
+      // peer, and it arrives on the direction that survives a freeze.
+      const peerLoopLag = Number(echo?.report?.loopLagMs);
+      if (Number.isFinite(peerLoopLag) && peerLoopLag >= 0) {
+        connection.peerLoopLagMs = peerLoopLag;
+      }
+      const peerVisibility = echo?.report?.visibility;
+      if (typeof peerVisibility === "string" && peerVisibility.length > 0) {
+        connection.peerVisibility = peerVisibility;
+      }
+      if (connection.echoAt !== 0) {
+        const sinceLast = now - connection.echoAt;
+        if (sinceLast > connection.echoIntervalMs) {
+          connection.echoIntervalMs = sinceLast;
+        }
+      }
+      connection.echoAt = now;
+      connection.echoes += 1;
+    },
+
+    dispose() {
+      for (const connection of connections.values()) {
+        if (connection.timer !== null) {
+          clearInterval(connection.timer);
+          connection.timer = null;
+        }
+      }
+      connections.clear();
+    }
+  };
+}