@serfab/cadre-core 1.3.0 → 1.5.0

package/src/index.ts

@@ -119,8 +119,16 @@ export {
 
 // Storage scope keys — what `CadreNodeConfig.storage.provider` is called with. A
 // strand's key is its strand id; the control database's key carries the party id,
-// so two parties on one device never share a control store.
-export { controlStorageScope, isControlStorageScope } from './storage-scope.js';
+// so two parties on one device never share a control store. Every key stays within
+// `[A-Za-z0-9._-]`: the control key by base64url encoding, a strand's by
+// `assertStrandScopeKey`, which every strand launch runs.
+export {
+  controlStorageScope,
+  isControlStorageScope,
+  isValidStrandScopeKey,
+  assertStrandScopeKey,
+  InvalidStrandIdError
+} from './storage-scope.js';
 
 // Strand database
 export { StrandDatabase, type StrandDatabaseConfig } from './strand-database.js';
@@ -162,6 +170,7 @@ export {
   PeerJoinBackfill,
   DEFAULT_PEER_JOIN_BACKFILL,
   MAX_BLOCK_MESSAGE_BYTES,
+  PEER_JOIN_BACKFILL_WARN_AFTER_FAILURES,
   type PeerJoinBackfillConfig,
   type PeerJoinBackfillDeps,
   type PeerJoinBackfillResult,
@@ -432,11 +441,28 @@ export {
   directBeforeRelayed,
   DEFAULT_CONTROL_COHORT_DIAL_TIMEOUT_MS,
   DEFAULT_CONTROL_COHORT_PER_ADDRESS_DIAL_TIMEOUT_MS,
+  CONTROL_COHORT_DIAL_ADDRESS_ATTEMPTS,
   DEFAULT_PEER_DIAL_BUDGET,
   type PeerDialBudget,
   type AddrDialer
 } from './peer-dial.js';
 
+// Dial and reservation budgets counted in link round trips rather than fixed milliseconds, so
+// one declared assumption about the link moves them all — see `NetworkConfig.linkRoundTripMs`.
+export {
+  DECLARED_LINK_ROUND_TRIP_MS,
+  RELAYED_DIAL_ROUND_TRIPS,
+  RELAY_RESERVATION_ROUND_TRIPS,
+  CIRCUIT_REQUEST_ROUND_TRIPS,
+  PUSH_TRANSFER_ALLOWANCE_MS,
+  resolveLinkRoundTripMs,
+  relayedDialBudgetMs,
+  relayReservationBudgetMs,
+  circuitRequestBudgetMs,
+  peerJoinPushBudget,
+  type PeerJoinPushBudget
+} from './link-budget.js';
+
 // Seed trust policy (trust anchor for incoming seeds)
 export {
   anchoredTrustPolicy,
@@ -493,6 +519,7 @@ export {
   DEFAULT_ENROLLMENT_WINDOW_MS,
   ADMISSION_DECISION_TIMEOUT_MS,
   RELAY_ADMISSION_RESERVE_DEADLINE_MS,
+  RELAY_ADMISSION_CLOSE_TIMEOUT_MS,
   MAX_UNAUTHORIZED_RELAY_RESERVATIONS,
   UNAUTHORIZED_RESERVATION_TTL_MS,
   UnauthorizedReservationBudget,

package/src/link-budget.ts

@@ -0,0 +1,217 @@
+/**
+ * Dial and reservation budgets derived from ONE declared link round trip, instead of fixed
+ * millisecond numbers chosen on a fast local network.
+ *
+ * Reaching another machine through a relay costs a FIXED NUMBER OF EXCHANGES, not a fixed
+ * number of milliseconds. So every budget written as milliseconds has a link speed above which
+ * it can never open a connection at all — and the slower the link, the more of them fall over,
+ * silently, because an abandoned dial looks the same as a peer that is not there. Counting round
+ * trips instead is what stops that class of defect coming back: the next person to widen a
+ * budget changes ONE declaration, and the next person to add a dial writes down its round-trip
+ * count beside the arithmetic.
+ *
+ * ── The instrument ──
+ *
+ * Every count and every number below comes from
+ * `packages/integration-tests/src/scenarios/relayed-dial-cost-by-latency.integration.ts`, which
+ * is committed and opt-in (`RELAY_DIAL_COST=1`). **Its doc comment is the single home of the
+ * measurement**; this module holds only what the derivation needs. Re-run it before changing
+ * anything here.
+ *
+ * ── The unit ──
+ *
+ * ONE LINK ROUND TRIP is a message from one machine reaching the other and a reply coming
+ * back, over the path those two machines actually use — through the relay, when that is the
+ * only way they can reach each other. In the instrument, which injects a constant ONE-WAY
+ * frame delay `d`, one link round trip is `2d`.
+ *
+ * A leg between a node and the RELAY is half a link round trip, because it crosses one of the
+ * two hops rather than both. That is why the reservation drive's cost reads as "four
+ * node-to-relay round trips" in one telling and "two link round trips" in another: same
+ * quantity, different unit. This module uses the LINK round trip throughout, so there is one
+ * unit to reason in.
+ *
+ * ── The counts, and what they cost ──
+ *
+ * | operation                                            | link round trips | at the declared default |
+ * | ---------------------------------------------------- | ---------------- | ----------------------- |
+ * | open a relayed connection to another machine          | 4                | 8 000 ms                |
+ * | dial the relay itself                                 | 1                | 2 000 ms                |
+ * | request a reservation on an open relay connection     | 1                | 2 000 ms                |
+ * | negotiate a protocol over an established circuit      | 1                | 2 000 ms                |
+ *
+ * **Measured** 2026-09-26, one Windows machine, loopback dedicated relay, re-run for this
+ * change: a relayed dial took 20-25 ms at no delay, **7 255-7 279 ms at 900 ms one-way**, and
+ * **12 066-12 094 ms at 1 500 ms one-way**. Four link round trips of pure delay would be 7 200
+ * and 12 000, so the real cost sits about 60-95 ms ABOVE the arithmetic — the handshakes and a
+ * phone's pure-JS crypto, which no delay figure contains. That residue is small but it is why
+ * {@link DECLARED_LINK_ROUND_TRIP_MS} carries headroom over the band sereus supports rather
+ * than matching it exactly: a declaration equal to the measured round trip would derive a
+ * budget marginally BELOW the dial it has to contain.
+ *
+ * ── The ceiling this does NOT lift ──
+ *
+ * Above roughly 1 250 ms one-way — a 2.5-second link round trip — no relayed connection can be
+ * established by this stack whatever cadre declares here. Two budgets of 10 000 ms each inside
+ * libp2p bound the same dial and cannot be reached from sereus at all:
+ * `connectionManager.dialTimeout` (libp2p's own default, which `@optimystic/db-p2p`'s
+ * `libp2p-node-base.ts` neither sets nor exposes) and `connectionManager.inboundUpgradeTimeout`
+ * (which it sets to 10 000). Both appear in the same re-run: at 1 500 ms one-way the dial under
+ * libp2p's default failed with `The operation was aborted due to timeout`, and with the shipped
+ * inbound-upgrade budget the DIALER's own 12 094 ms dial resolved while the LISTENER had
+ * already thrown the half-built connection away — so the first stream over it died with
+ * `Unexpected EOF - stream closed while reading 0/1 bytes` and the listener reported no peer at
+ * all. That is why a too-slow link produces silence rather than an error.
+ *
+ * Two consequences for a reader choosing a declaration. Below that ceiling, the DEFAULT
+ * declaration of 2 000 ms does not reach it: four round trips of 2 000 ms is 8 000 ms, which
+ * covers a one-way delay up to 1 000 ms, so a deployment somewhere between 1 000 and 1 250 ms
+ * one-way has to declare its own (about 2 200-2 500) to use the band libp2p still allows. And
+ * declaring much ABOVE 2 500 buys nothing: the budgets here grow, and
+ * the connection still fails inside libp2p. Lifting that needs an upstream change and a
+ * decision about how slow a link sereus intends to carry —
+ * `tickets/blocked/how-slow-a-relayed-link-does-sereus-carry`. A reader who raised these
+ * numbers and still cannot connect at 1 500 ms one-way has met that ceiling, not this module.
+ */
+
+/**
+ * The link round trip cadre assumes when a host declares none, in milliseconds. A deployment
+ * that knows its own link states it with `NetworkConfig.linkRoundTripMs` instead.
+ *
+ * **Why 2 000.** It states the slowest link sereus already claims to support, and two other
+ * budgets fix that claim independently: `COHORT_READ_DEADLINE_MS`
+ * (`quereus-plugin-sereus/src/cluster-size.ts`, 5 000 ms) and
+ * `DEFAULT_STRAND_FIRST_SYNC_TIMEOUT_MS` (`strand-first-sync-gate.ts`, 120 000 ms) were both
+ * sized against two parties reaching each other only through a public relay at a round trip
+ * near 1.8 s. Declaring 2 000 rounds that band up and leaves about 720 ms over the worst
+ * relayed dial measured at it (7 279 ms against a derived 8 000 ms).
+ *
+ * Raising it is how a deployment that KNOWS its link is slower moves every derived budget at
+ * once. The cost is the ordinary cost of a longer deadline: a peer that is genuinely gone holds
+ * the operation for that much longer before it is abandoned and retried. Above about 2 500 it
+ * buys nothing — see the module doc's ceiling paragraph.
+ */
+export const DECLARED_LINK_ROUND_TRIP_MS = 2000;
+
+/**
+ * Link round trips one relayed connection setup costs: transport, encryption and multiplexer
+ * handshakes to the relay, the circuit open, then both handshakes again end to end through it.
+ * Measured at 8 one-way delays, which is 4 link round trips.
+ */
+export const RELAYED_DIAL_ROUND_TRIPS = 4;
+
+/**
+ * Link round trips one whole reservation drive is budgeted for
+ * (`relay-reservation.ts`'s `driveRelayReservation`, whose ONE deadline covers the dial, the
+ * reservation request and the wait together).
+ *
+ * The protocol work measured is 2 — dial the relay (1) plus request the reservation (1). It is
+ * budgeted at 4, the relayed-dial count, because the wait afterwards may instead be satisfied
+ * by libp2p's own relay discovery, which repeats both legs after identify has run; 4 is
+ * therefore the largest thing this one deadline can be asked to contain, and the poll interval
+ * (`DEFAULT_RELAY_RESERVE_POLL_MS`) rounds up on top of it.
+ */
+export const RELAY_RESERVATION_ROUND_TRIPS = 4;
+
+/**
+ * Link round trips one request-and-answer over an ALREADY-OPEN circuit costs: the protocol
+ * negotiation (1, measured at 2 one-way delays) plus the request and its response (1). It does
+ * not include a dial — a caller that may have to open the connection budgets
+ * {@link RELAYED_DIAL_ROUND_TRIPS} separately, which is exactly what the two-field shape of
+ * Optimystic's `dialTimeoutMs` / `responseTimeoutMs` pair is for.
+ */
+export const CIRCUIT_REQUEST_ROUND_TRIPS = 2;
+
+/**
+ * What a block-transfer push message's BYTES get to cross, on top of the latency
+ * {@link CIRCUIT_REQUEST_ROUND_TRIPS} covers, in milliseconds.
+ *
+ * This part is bandwidth, not latency, so it does not scale with the declared round trip: a
+ * chunk is capped at `PeerJoinBackfillConfig.maxChunkBytes` (1 MiB by default) and crossing it
+ * costs whatever the link's throughput costs.
+ *
+ * NOTE: 6 000 ms is not a measurement of any throughput — it is the residue of the 10 000 ms
+ * this deadline shipped as before it was split into a latency part and a transfer part, chosen
+ * so that at the default declaration the derived value is the same 10 000 and no fast link got
+ * slower. Nobody has measured how long 1 MiB takes to cross a relayed mobile link. If pushes
+ * ever time out with the transfer only part done, measure that and raise THIS, not the declared
+ * round trip.
+ */
+export const PUSH_TRANSFER_ALLOWANCE_MS = 6000;
+
+/**
+ * Resolve a host's declared link round trip, falling back to
+ * {@link DECLARED_LINK_ROUND_TRIP_MS}.
+ *
+ * Validated here rather than downstream because nothing downstream would: every consumer
+ * multiplies this value into a `setTimeout` deadline, where a zero silently turns a budget into
+ * "give up immediately" and a `NaN` turns it into "never time out".
+ *
+ * Both node bring-up paths call this EAGERLY — `CadreNode.start()`, and
+ * `StrandInstanceManager.buildStrandRuntime` behind `addStrand`/`resumeStrand` — so a bad
+ * declaration fails the same start that Optimystic's own check on `cohortQueryTimeoutMs` fails.
+ * The eager call is what makes that true: every budget derived here is behind a condition (a
+ * node with no control storage builds no catch-up, a node with no relay addrs drives no
+ * reservation), so waiting for a first consumer would let a broken declaration boot and then
+ * throw inside a best-effort path that logs and carries on.
+ */
+export function resolveLinkRoundTripMs(linkRoundTripMs?: number): number {
+	if (linkRoundTripMs === undefined) {
+		return DECLARED_LINK_ROUND_TRIP_MS;
+	}
+	if (!Number.isFinite(linkRoundTripMs) || linkRoundTripMs <= 0) {
+		throw new Error(
+			`network.linkRoundTripMs must be a finite number of milliseconds above zero, not ${String(linkRoundTripMs)}`
+		);
+	}
+	return linkRoundTripMs;
+}
+
+/**
+ * Deadline for OPENING a connection to another machine that may only be reachable through a
+ * relay: {@link RELAYED_DIAL_ROUND_TRIPS} at the declared link round trip.
+ */
+export function relayedDialBudgetMs(linkRoundTripMs?: number): number {
+	return RELAYED_DIAL_ROUND_TRIPS * resolveLinkRoundTripMs(linkRoundTripMs);
+}
+
+/**
+ * Deadline for one whole relay reservation drive: {@link RELAY_RESERVATION_ROUND_TRIPS} at the
+ * declared link round trip.
+ */
+export function relayReservationBudgetMs(linkRoundTripMs?: number): number {
+	return RELAY_RESERVATION_ROUND_TRIPS * resolveLinkRoundTripMs(linkRoundTripMs);
+}
+
+/**
+ * Deadline for one request and its answer over a circuit that is already open —
+ * {@link CIRCUIT_REQUEST_ROUND_TRIPS} at the declared link round trip, plus `transferAllowanceMs`
+ * for the payload's own bytes.
+ */
+export function circuitRequestBudgetMs(
+	linkRoundTripMs?: number,
+	transferAllowanceMs = PUSH_TRANSFER_ALLOWANCE_MS
+): number {
+	return CIRCUIT_REQUEST_ROUND_TRIPS * resolveLinkRoundTripMs(linkRoundTripMs) + transferAllowanceMs;
+}
+
+/** The two deadlines one peer-join catch-up push needs, both derived from the declared link. */
+export interface PeerJoinPushBudget {
+	/** Opening the connection to the peer being caught up — a relayed dial, in the worst case. */
+	dialTimeoutMs: number;
+	/** The push request and its answer over that connection, including the chunk's own bytes. */
+	responseTimeoutMs: number;
+}
+
+/**
+ * The dial and response deadlines a peer-join catch-up push needs at a declared link. One
+ * helper rather than two call-site expressions, because the two construction sites
+ * (`CadreNode`'s control catch-up and `StrandInstanceManager`'s per-strand one) and
+ * `DEFAULT_PEER_JOIN_BACKFILL` must not drift apart.
+ */
+export function peerJoinPushBudget(linkRoundTripMs?: number): PeerJoinPushBudget {
+	return {
+		dialTimeoutMs: relayedDialBudgetMs(linkRoundTripMs),
+		responseTimeoutMs: circuitRequestBudgetMs(linkRoundTripMs)
+	};
+}

package/src/membership-connection-gater.ts

@@ -88,9 +88,10 @@
  *    annex the party's relay capacity.
  *  - An `'admit-for-relay'` connection that is NOT reserving is dropped: it has
  *    {@link RELAY_ADMISSION_RESERVE_DEADLINE_MS} to get a reservation ADMITTED
- *    at that hook, after which the underlying connection is aborted. The
- *    guarantee above thus weakens on relay-enabled nodes from "never in the
- *    conversation" to "in it briefly, and can speak nothing".
+ *    at that hook, after which the gate CLOSES the underlying connection so
+ *    both ends let go of the socket. The guarantee above thus weakens on
+ *    relay-enabled nodes from "never in the conversation" to "in it briefly,
+ *    and can speak nothing".
  *
  * The deadline clears on reservation ADMISSION, not on the reservation's own
  * success: this gate cannot observe the server-side `reserve()` outcome, so a
@@ -158,6 +159,7 @@ import debug from 'debug';
 import type { ConnectionGater, PeerId, MultiaddrConnection } from '@libp2p/interface';
 import { SEED_PROTOCOL } from './seed-bootstrap.js';
 import { FORMATION_PROTOCOL } from './strand-formation-protocol.js';
+import { withDeadline } from './control-stream.js';
 
 const log = debug('sereus:cadre:connection-gater');
 
@@ -197,7 +199,7 @@ export const ADMISSION_DECISION_TIMEOUT_MS = 2_000;
 /**
  * How long an `'admit-for-relay'` connection may exist without a relay
  * reservation being ADMITTED at the `denyInboundRelayReservation` hook, after
- * which the gate aborts the underlying connection. A reserving client asks for
+ * which the gate closes the underlying connection. A reserving client asks for
  * its slot immediately after the connection upgrades (`relay-reservation.ts`
  * dials and requests in one drive; a strand node's configured circuit listener
  * does the same from inside its own `listen()`), so a connection idle past this
@@ -205,6 +207,19 @@ export const ADMISSION_DECISION_TIMEOUT_MS = 2_000;
  */
 export const RELAY_ADMISSION_RESERVE_DEADLINE_MS = 5_000;
 
+/**
+ * Bound on the graceful close of an expired relay-only connection. It exists
+ * because `AbstractMultiaddrConnection.close()` awaits an `idle`/`drain` event
+ * when the connection still has unsent bytes, and that wait has no timeout of
+ * its own — an unsignalled one never ends. This gate writes nothing to a
+ * stranger, so the wait is not reachable today; the bound is here so a timer
+ * callback can never hold an unending await.
+ *
+ * Applied through {@link withDeadline}, not `AbortSignal.timeout`: the latter is
+ * not reliably present on React Native/Hermes, which loads this same module.
+ */
+export const RELAY_ADMISSION_CLOSE_TIMEOUT_MS = 2_000;
+
 /**
  * Default cap on concurrent relay reservations held by peers the membership
  * check could not place (see {@link UnauthorizedReservationBudget}). Small on
@@ -351,7 +366,7 @@ export class UnauthorizedReservationBudget {
  *
  * An `'admit-for-relay'` verdict admits the connection and arms a
  * `reserveDeadlineMs` timer against it; the timer is disarmed when a
- * reservation for that peer is ADMITTED at the reservation hook, and aborts the
+ * reservation for that peer is ADMITTED at the reservation hook, and closes the
  * underlying `MultiaddrConnection` when it fires first.
  *
  * NOTE: `base` is spread, so a gater passed as a CLASS INSTANCE would lose its
@@ -470,7 +485,7 @@ export async function decideWithinDeadline<T>(
   }
 }
 
-/** One armed not-reserving deadline: the connection it will abort, and its timer. */
+/** One armed not-reserving deadline: the connection it will close, and its timer. */
 interface PendingReserveDeadline {
   maConn: MultiaddrConnection;
   timer: ReturnType<typeof setTimeout>;
@@ -479,11 +494,11 @@ interface PendingReserveDeadline {
 /**
  * The not-reserving deadlines for `'admit-for-relay'` connections, keyed by
  * remote peerId (a Set per peer — one peer can hold several in-flight
- * connections). `arm` starts a timer that aborts the connection; `disarm`
+ * connections). `arm` starts a timer that closes the connection; `disarm`
  * (called when a reservation for that peer is admitted) cancels every pending
  * timer for the peer. Timers are unref'd so an armed deadline never holds a
  * process open, and a timer that fires against an already-closed connection is
- * a swallowed no-op.
+ * a no-op (`close()` returns at once on any status but `open`).
  *
  * Disarming is final for the connections it cancelled: a peer that reserved
  * once and then lets its reservation lapse keeps a mute connection (the
@@ -528,11 +543,41 @@ class PendingReserveDeadlines {
     if (entries?.size === 0) {
       this.byPeer.delete(remotePeerId);
     }
-    log('Relay-only admission expired for %s — no reservation admitted within %dms, aborting the connection', remotePeerId, this.deadlineMs);
+    log('Relay-only admission expired for %s — no reservation admitted within %dms, closing the connection', remotePeerId, this.deadlineMs);
+    void this.drop(remotePeerId, entry.maConn);
+  }
+
+  /**
+   * End an expired relay-only connection: a bounded graceful close, with
+   * `abort()` only as the escalation. The order cannot be reversed — `abort()`
+   * marks the connection `aborted`, and `close()` returns immediately on any
+   * status that is not `open`, so aborting first would make the close a silent
+   * no-op. Closing a connection the peer already dropped returns at once
+   * without throwing, so the fallback stays unreached in the ordinary case.
+   */
+  private async drop(remotePeerId: string, maConn: MultiaddrConnection): Promise<void> {
+    try {
+      await withDeadline(
+        RELAY_ADMISSION_CLOSE_TIMEOUT_MS,
+        `relay-only close for ${remotePeerId}`,
+        (signal) => maConn.close({ signal })
+      );
+      return;
+    } catch (error) {
+      log('Closing the expired relay-only connection from %s failed — aborting: %o', remotePeerId, error);
+    }
+    // NOTE: this fallback cannot actually free a WebSocket. `@libp2p/websockets`'
+    // `sendReset()` calls `websocket.close(1006)`, and 1006 is a reserved code
+    // RFC 6455 forbids an endpoint from sending, so `ws` throws and
+    // `AbstractMessageStream.abort()` swallows it — the local end flips to
+    // `aborted` while the socket stays up. Measured against
+    // @libp2p/websockets@10.1.3 (see tickets/blocked/report-libp2p-websockets-abort-close-code.md).
+    // `MultiaddrConnection` exposes no way down to the raw socket, so this is as
+    // far as this layer can go. Revisit — drop the fallback, or go back to a
+    // plain `abort()` — once that transport sends a legal reset code.
     try {
-      entry.maConn.abort(new Error(`relay-only admission expired: no relay reservation admitted within ${this.deadlineMs}ms`));
+      maConn.abort(new Error(`relay-only admission expired: no relay reservation admitted within ${this.deadlineMs}ms`));
     } catch (error) {
-      // Aborting a connection that already closed on its own is not a failure.
       log('Aborting the expired relay-only connection from %s threw: %o', remotePeerId, error);
     }
   }

package/src/peer-dial.ts

@@ -27,32 +27,47 @@ import debug from 'debug';
 import type { Connection } from '@libp2p/interface';
 import type { Multiaddr } from '@multiformats/multiaddr';
 import { withDeadline } from './control-stream.js';
+import { relayedDialBudgetMs } from './link-budget.js';
 
 const log = debug('sereus:cadre:peer-dial');
 
 /**
- * Default limit on ONE address's dial attempt, in ms.
+ * Default limit on ONE address's dial attempt, in ms, at the DEFAULT declared link round trip:
+ * `RELAYED_DIAL_ROUND_TRIPS` (4) x `DECLARED_LINK_ROUND_TRIP_MS` (2000 ms) = 8000 ms, which is
+ * what this constant was already set to as a fixed number.
  *
  * It has to cover the slowest address that should succeed, because an attempt
  * that needs longer fails the same way on every retry. That is a relayed dial on
  * a mobile link: connect to the relay (transport, encryption and multiplexer
  * handshakes, when no relay connection is open yet), open the circuit, then run
- * both handshakes again end to end through the relay at roughly twice the
- * round-trip time. At a few hundred ms per round trip that is several seconds.
- * A direct dial that works is far quicker — a phone's WebSocket dial to a node
- * forwarded over `adb reverse` measured 1.6 s end to end. 8 s covers the relayed
- * case with some margin while still abandoning a silently dropped address
- * sooner than libp2p's own 10 s dial timeout would.
+ * both handshakes again end to end through the relay. `link-budget.ts` is where that
+ * reasoning is now a measured count instead of an estimate — four link round trips, measured at
+ * 7 279 ms on a link whose round trip is 1.8 s — so the value MOVES with a host's
+ * `NetworkConfig.linkRoundTripMs` rather than pinning the band this node can reach to whatever
+ * number was typed here. A direct dial that works is far quicker — a phone's WebSocket dial to
+ * a node forwarded over `adb reverse` measured 1.6 s end to end.
+ *
+ * Deriving it rather than leaving it fixed is deliberate: at 8000 ms fixed, the per-address
+ * limit — not libp2p's own 10 s dial timeout — was the first thing to cut a relayed dial off,
+ * from about 1000 ms of one-way link delay upward. Nothing said so, and a reader raising
+ * `DEFAULT_CONTROL_COHORT_DIAL_TIMEOUT_MS` below would not have found it.
  *
  * NOTE: a `/p2p-circuit/webrtc` address also negotiates a WebRTC session after
  * the relay leg. If such dials are ever seen failing at exactly this limit, raise
- * it (or `network.controlCohort.perAddressDialTimeoutMs` on the affected nodes)
- * rather than the per-peer limit, which cannot help an address that is itself
+ * the declared link round trip (or `network.controlCohort.perAddressDialTimeoutMs` on the
+ * affected nodes) rather than the per-peer limit, which cannot help an address that is itself
  * cut off.
  *
  * Override per node with `network.controlCohort.perAddressDialTimeoutMs`.
  */
-export const DEFAULT_CONTROL_COHORT_PER_ADDRESS_DIAL_TIMEOUT_MS = 8_000;
+export const DEFAULT_CONTROL_COHORT_PER_ADDRESS_DIAL_TIMEOUT_MS = relayedDialBudgetMs();
+
+/**
+ * How many whole per-address dial attempts {@link DEFAULT_CONTROL_COHORT_DIAL_TIMEOUT_MS}
+ * budgets for: the address that works plus the three dead ones a lent cadre-host node is
+ * observed to report ahead of it (loopback and LAN WebSocket addresses, plus virtual adapters).
+ */
+export const CONTROL_COHORT_DIAL_ADDRESS_ATTEMPTS = 4;
 
 /**
  * Default limit on dialing ONE peer — every candidate address together — in ms.
@@ -66,18 +81,25 @@ export const DEFAULT_CONTROL_COHORT_PER_ADDRESS_DIAL_TIMEOUT_MS = 8_000;
  *
  * Sized to fit a typical list with dead addresses ahead of the one that works:
  * a lent cadre-host node reports two to four addresses a phone can dial
- * (loopback and LAN WebSocket addresses, plus any virtual adapters). Three
- * silently dropped addresses cost 3 × 8 s = 24 s, leaving 6 s for the working
- * address after them — well over the 1.6 s the phone measured. A peer whose
- * working address sits behind four or more dropped ones does not fit, and fails
- * each pass the same way; if that shows up, the remedy is ordering (try the
- * address that last connected first), not a larger number here.
+ * (loopback and LAN WebSocket addresses, plus any virtual adapters). So it is
+ * {@link CONTROL_COHORT_DIAL_ADDRESS_ATTEMPTS} whole per-address budgets — three silently
+ * dropped addresses, then the one that works, each getting the FULL per-address limit. A peer
+ * whose working address sits behind four or more dropped ones does not fit, and fails each pass
+ * the same way; if that shows up, the remedy is ordering (try the address that last connected
+ * first), not a larger number here.
+ *
+ * Counted as whole per-address budgets rather than the flat 30_000 it was before, because the
+ * old number left only 6 s after three dropped addresses — enough for the 1.6 s direct dial the
+ * phone measured, but NOT for the relayed dial the per-address limit is sized to cover. A peer
+ * whose only working address was a relayed one behind two dead ones therefore could not be
+ * dialled at all, and the arithmetic in this comment was what said otherwise.
  *
  * Override per node with `network.controlCohort.dialTimeoutMs` — tests that
  * drive dead addresses on purpose set it low so a pass's duration is a chosen
  * number rather than a transitive libp2p default stretched by machine load.
  */
-export const DEFAULT_CONTROL_COHORT_DIAL_TIMEOUT_MS = 30_000;
+export const DEFAULT_CONTROL_COHORT_DIAL_TIMEOUT_MS =
+	CONTROL_COHORT_DIAL_ADDRESS_ATTEMPTS * DEFAULT_CONTROL_COHORT_PER_ADDRESS_DIAL_TIMEOUT_MS;
 
 /** Time limits for dialing one peer from a list of candidate addresses. */
 export interface PeerDialBudget {