@serfab/cadre-core 1.16.0 → 1.18.0

package/src/strand-addr-protocol.ts

@@ -1,538 +1,547 @@
-/**
- * Control-network strand-address RPC.
- *
- * A strand runs as its own libp2p node (`strand-<id>`, random port), separate
- * from the control node (`control-<partyId>`), with its own transport peerId
- * derived from the cadre identity key (`strand-transport-key.ts`). To seed a
- * strand's mesh, a node needs a sibling's **strand-network** address — but
- * `CadrePeer.Multiaddr` only stores the **control** node's addresses, and a
- * control address now names a different peer entirely, not merely a different
- * port. This protocol resolves the strand address on demand: a node asks its
- * connected co-cadre siblings "what are your live strand-`X` multiaddrs?" and
- * uses the union of their answers as the seed.
- *
- * Modeled directly on `strand-wake-protocol.ts`: a dedicated libp2p protocol id,
- * 4-byte big-endian length-prefixed JSON frames, `node.handle` for the receiver,
- * `node.dialProtocol` for the client, and the shared `ControlStream` primitives
- * from `control-stream.ts`. The exchange is a single request → single response on
- * one stream, so each side reads to EOF (under a read timeout) and decodes one
- * frame via the shared {@link decodeLengthPrefixedFrame} guard.
- *
- * **Authorization (v1):** like wake, the request carries no signature. Membership
- * is checked before this handler runs, by the control node's protocol guard
- * (`control-protocol-guard.ts`, the `members` class: `CadreNode.isAuthorizedMember`,
- * a voucher on the requester's `CadrePeer` row verified against the node-local
- * trusted-owner anchor). A peer it refuses gets its stream reset, never a reply, so
- * an outsider that published its own rows into the replicated control DB can neither
- * harvest live strand addresses nor announce a delegate. Cross-party strand bootstrap
- * is a different mechanism (strand formation / `MemberPeer`) and is out of scope here.
- */
-
-import debug from 'debug';
-import type { Libp2p, Connection, PeerId } from '@libp2p/interface';
-import type { Multiaddr } from '@multiformats/multiaddr';
-import { peerIdFromString } from '@libp2p/peer-id';
-import type { StrandAddrRequest, StrandAddrResponse, StrandAddrStatus } from './types.js';
-import { decodeLengthPrefixedFrame } from './seed-bootstrap.js';
-import { orderSignalingFirst } from './peer-record.js';
-import { type ControlStream, withDeadline, exchangeFrame, readStreamToEnd, replyAndClose } from './control-stream.js';
-import { relayedRequestBudgetMs } from './link-budget.js';
-
-const log = debug('sereus:cadre:strand-addr');
-
-/** Protocol id for the control-network strand-address RPC (parallel to `/sereus/strand-wake/1.0.0`). */
-export const STRAND_ADDR_PROTOCOL = '/sereus/strand-addr/1.0.0';
-
-/**
- * Maximum strand-addr frame size. Responses are tiny (a strand id + a few
- * multiaddrs), so this is a defensive cap — far below the 1MB seed ceiling —
- * that bounds the bytes a peer can make the receiver buffer per stream.
- */
-const MAX_ADDR_SIZE = 64 * 1024;
-
-/** Default time the receiver waits for an inbound request frame before aborting (ms). */
-// eslint-disable-next-line no-restricted-syntax -- link-independent: a receiver cap on one small request frame on a stream the peer already opened; it bounds a peer that opens a stream and never sends, not the dial
-const DEFAULT_ADDR_READ_TIMEOUT_MS = 10_000;
-
-/** Default cap on concurrent inbound strand-addr streams a single peer can pin open. */
-const DEFAULT_MAX_CONCURRENT_ADDRS = 100;
-
-/**
- * Read a libp2p stream to EOF and decode the single length-prefixed JSON frame
- * it carries. Bounded by `timeoutMs` (a never-half-closing peer is aborted, not
- * awaited forever) and capped at {@link MAX_ADDR_SIZE}, reusing the shared
- * {@link decodeLengthPrefixedFrame} guard for the prefix/length checks.
- */
-async function readFrame<T>(stream: ControlStream, timeoutMs: number): Promise<T> {
-  const data = await readStreamToEnd(stream, { maxBytes: MAX_ADDR_SIZE, timeoutMs, label: 'Strand-addr' });
-  const body = decodeLengthPrefixedFrame(data, MAX_ADDR_SIZE);
-  return JSON.parse(new TextDecoder().decode(body)) as T;
-}
-
-/** An address-free reply, for every path but a lookup: over the cap, an unreadable request, a lookup that threw. */
-function addrlessResponse(status: StrandAddrStatus, strandId: string): StrandAddrResponse {
-  return { status, strandId, multiaddrs: [] };
-}
-
-/**
- * Dependencies the {@link StrandAddrService} receiver needs from its host
- * (`CadreNode`), injected so the service is testable without a full node.
- */
-export interface StrandAddrServiceOptions {
-  /**
-   * The local strand instance's dialable strand-network multiaddrs, ordered
-   * signaling-first. Returns `[]` when the strand is not running / has no live
-   * node (hibernating, quiescing, or never participated). Kept injectable so the
-   * decision logic is unit-testable without a full node.
-   */
-  getStrandMultiaddrs(strandId: string): string[];
-  /**
-   * Called when a member's request carries a `delegatePeerId` — the derived
-   * transport peerId its strand-`strandId` node runs as, with a validated
-   * (parseable, non-self) peerId. Every caller is an authorized member: the
-   * protocol guard admitted it before this handler ran.
-   * `CadreNode` injects its delegate-admission grant recorder so this node's
-   * connection gate — and thus its circuit-relay server, when it runs one —
-   * admits that peerId (see `delegate-admission.ts`). Optional: absent, an
-   * announce is ignored and the address lookup proceeds unchanged.
-   */
-  onDelegateAnnounce?(announcerPeerId: string, strandId: string, delegatePeerId: string): void;
-  /**
-   * Time to wait for the inbound request frame before aborting the read (ms).
-   * Defaults to {@link DEFAULT_ADDR_READ_TIMEOUT_MS}. Bounds a buggy/compromised
-   * own-cadre node that opens a stream and never half-closes its write end.
-   */
-  readTimeoutMs?: number;
-  /**
-   * Cap on concurrent inbound strand-addr streams (defaults to
-   * {@link DEFAULT_MAX_CONCURRENT_ADDRS}). Over the cap, an `unavailable` reply is
-   * returned without looking up any strand address.
-   */
-  maxConcurrent?: number;
-}
-
-/**
- * Receiver side of the strand-address RPC. Registers a `STRAND_ADDR_PROTOCOL`
- * handler on the control node and, for each inbound {@link StrandAddrRequest}
- * (from a peer the protocol guard has already admitted as an authorized member),
- * replies with the local strand instance's live multiaddrs — `ok` with an empty
- * list when the strand is not running, `unavailable` when it could not answer at all.
- */
-export class StrandAddrService {
-  private readonly options: StrandAddrServiceOptions;
-  private readonly readTimeoutMs: number;
-  private readonly maxConcurrent: number;
-  private node: Libp2p | null = null;
-  /** In-flight inbound strand-addr streams, used to enforce {@link maxConcurrent}. */
-  private activeStreams = 0;
-
-  constructor(options: StrandAddrServiceOptions) {
-    this.options = options;
-    this.readTimeoutMs = options.readTimeoutMs ?? DEFAULT_ADDR_READ_TIMEOUT_MS;
-    this.maxConcurrent = options.maxConcurrent ?? DEFAULT_MAX_CONCURRENT_ADDRS;
-  }
-
-  /** Number of in-flight inbound strand-addr streams. */
-  get activeCount(): number {
-    return this.activeStreams;
-  }
-
-  /**
-   * Register the strand-addr protocol handler on the control node. Rejects when
-   * libp2p refuses the registration (a duplicate handler, a peer-store write
-   * failure); the node reference is kept only once the handler is in place.
-   */
-  async initialize(node: Libp2p): Promise<void> {
-    // `runOnLimitedConnection: true` is REQUIRED for the relay path: a NAT'd
-    // sibling is reached over a circuit-relay connection, which libp2p marks
-    // "limited" (the relay caps its data/duration). Without this the receiver
-    // would refuse the inbound stream on exactly the connection the protocol is
-    // designed to use (same reasoning as wake; see the relay note on
-    // `collectStrandAddrs`).
-    await node.handle(STRAND_ADDR_PROTOCOL, async (rawStream: unknown, rawConnection: unknown) => {
-      const remotePeerId = (rawConnection as Connection).remotePeer.toString();
-      await this.handleStream(rawStream as ControlStream, remotePeerId);
-    }, { runOnLimitedConnection: true });
-    this.node = node;
-    log('StrandAddrService registered handler: %s', STRAND_ADDR_PROTOCOL);
-  }
-
-  /** Unregister the handler and release the node reference. */
-  async shutdown(): Promise<void> {
-    if (this.node) {
-      await this.node.unhandle(STRAND_ADDR_PROTOCOL);
-      this.node = null;
-      log('StrandAddrService shutdown');
-    }
-  }
-
-  /**
-   * Read the inbound request, decide the response, and write it back.
-   *
-   * Three hardening layers, all reported as an `unavailable`
-   * {@link StrandAddrResponse} rather than a dropped/hung stream: a concurrency cap
-   * (over {@link maxConcurrent}, reply without looking up any address), a read
-   * timeout (a peer that never half-closes is aborted inside {@link readFrame}/
-   * `readStreamToEnd`), and the existing malformed/oversized-frame guard. A
-   * lookup that throws is `unavailable` too, never an empty `ok` the asker would
-   * wait ten minutes on.
-   */
-  private async handleStream(stream: ControlStream, remotePeerId: string): Promise<void> {
-    log('Incoming strand-addr request from: %s', remotePeerId);
-
-    if (this.activeStreams >= this.maxConcurrent) {
-      log('Rejecting strand-addr from %s: %d concurrent streams at cap %d', remotePeerId, this.activeStreams, this.maxConcurrent);
-      // The request frame is unread here, so the strand id is unknown — reply
-      // under an empty strand id; the asker retries an `unavailable` sibling soon.
-      await replyAndClose(stream, addrlessResponse('unavailable', ''), 'Strand-addr');
-      return;
-    }
-
-    this.activeStreams++;
-    try {
-      await replyAndClose(stream, await this.answerStream(stream, remotePeerId), 'Strand-addr');
-    } finally {
-      this.activeStreams--;
-    }
-  }
-
-  /** Read and decide one inbound request; any failure becomes an `unavailable` response. */
-  private async answerStream(stream: ControlStream, remotePeerId: string): Promise<StrandAddrResponse> {
-    try {
-      const request = await readFrame<StrandAddrRequest>(stream, this.readTimeoutMs);
-      return await this.processAddrRequest(request, remotePeerId);
-    } catch (err) {
-      log('Error handling strand-addr request from %s: %o', remotePeerId, err);
-      // Malformed/oversized/timed-out request, or a lookup that threw: reply
-      // under an empty strand id rather than hanging.
-      return addrlessResponse('unavailable', '');
-    }
-  }
-
-  /**
-   * Decide the response for a decoded request. Exposed (not private) so the
-   * decision matrix can be unit-tested directly (mirrors wake's
-   * `processWakeRequest`).
-   *
-   * - Request carrying a `delegatePeerId` → recorded via
-   *   {@link StrandAddrServiceOptions.onDelegateAnnounce} before the lookup.
-   * - Strand not running locally → `ok` with empty `multiaddrs` (`getStrandMultiaddrs` → `[]`).
-   * - Running strand → `ok` with the strand's live, signaling-first multiaddrs.
-   *
-   * A throwing lookup propagates; {@link handleStream} answers it `unavailable`.
-   */
-  async processAddrRequest(request: StrandAddrRequest, remotePeerId: string): Promise<StrandAddrResponse> {
-    this.recordDelegateAnnounce(request, remotePeerId);
-
-    const multiaddrs = this.options.getStrandMultiaddrs(request.strandId);
-    log('Strand-addr for %s → %d addr(s)', request.strandId, multiaddrs.length);
-    return { status: 'ok', strandId: request.strandId, multiaddrs };
-  }
-
-  /**
-   * Validate and forward a member's delegate announcement. A malformed field
-   * (unparsable peerId) and a self-announcement (`delegatePeerId` equal to the
-   * announcer's own peerId — a member is admitted directly, never by grant)
-   * are logged and dropped; the address lookup proceeds either way.
-   */
-  private recordDelegateAnnounce(request: StrandAddrRequest, remotePeerId: string): void {
-    const { strandId, delegatePeerId } = request;
-    if (delegatePeerId === undefined || !this.options.onDelegateAnnounce) {
-      return;
-    }
-    if (delegatePeerId === remotePeerId) {
-      log('Ignoring self-announcement from %s (strand %s)', remotePeerId, strandId);
-      return;
-    }
-    try {
-      peerIdFromString(delegatePeerId);
-    } catch (err) {
-      log('Ignoring malformed delegatePeerId %o from %s: %o', delegatePeerId, remotePeerId, err);
-      return;
-    }
-    this.options.onDelegateAnnounce(remotePeerId, strandId, delegatePeerId);
-  }
-}
-
-/** A sibling to ask for its strand address. */
-export interface StrandAddrPeer {
-  /** The sibling's control-network peer id (also its strand-node peer id). */
-  peerId: string;
-  /**
-   * Pre-resolved control-network multiaddrs, used as a fallback when no control
-   * connection is already open. Dialing by {@link peerId} is preferred (it
-   * reuses an existing connection); these are tried only if that fails.
-   */
-  addrs?: Multiaddr[];
-}
-
-/** Options for {@link collectStrandAddrs}. */
-export interface CollectStrandAddrsOptions {
-  /**
-   * The asker's declared link round trip (`NetworkConfig.linkRoundTripMs`), from which the
-   * per-dial timeout is derived when {@link timeoutMs} is not given. Unset means the declared
-   * default.
-   */
-  linkRoundTripMs?: number;
-  /** Per-dial timeout in ms (default: {@link attemptTimeoutMs}). */
-  timeoutMs?: number;
-  /** Override the protocol id (defaults to {@link STRAND_ADDR_PROTOCOL}). */
-  protocolId?: string;
-  /**
-   * The derived transport peerId the caller's own strand-`strandId` node runs
-   * (or is about to run) as. Set on every request this collection sends, so
-   * each receiver records a delegate admission grant for it — see
-   * {@link StrandAddrRequest.delegatePeerId}. Omit for a pure address lookup.
-   */
-  delegatePeerId?: string;
-}
-
-/**
- * What asking one sibling produced, as {@link collectStrandAddrs} reports it.
- *
- * - `answered` — status `ok` with at least one address.
- * - `empty` — status `ok` with no address: the sibling does not run the strand
- *   right now, a normal steady state.
- * - `unavailable` — the sibling replied that it could not answer.
- * - `unreachable` — no reply at all: every dial target failed or timed out, the
- *   reply was malformed, or the sibling's protocol guard reset the stream because
- *   it does not (yet) count the asker as an authorized member.
- */
-export type StrandAddrOutcome = 'answered' | 'empty' | 'unavailable' | 'unreachable';
-
-/** Result of {@link collectStrandAddrs}. */
-export interface StrandAddrCollection {
-  /** Deduplicated union of every answer, signaling-first. */
-  addrs: string[];
-  /** One entry per candidate (self excluded), keyed by the sibling's control peerId. */
-  outcomes: Map<string, StrandAddrOutcome>;
-}
-
-/** One sibling's share of a {@link StrandAddrCollection}. */
-interface SiblingAnswer {
-  outcome: StrandAddrOutcome;
-  multiaddrs: string[];
-}
-
-/**
- * Client side: ask each candidate sibling for its live strand-`strandId`
- * multiaddrs and return the **deduplicated union** of every answer, ordered
- * signaling-first, alongside each sibling's {@link StrandAddrOutcome}.
- *
- * Best-effort per peer: a failed/timed-out/empty sibling contributes no address
- * and never fails the collection — an empty union (no sibling online or running
- * the strand) is an acceptable seed that self-heals on the next resume/reconcile
- * pass. `outcomes` is what lets a caller retry the siblings that could not answer
- * sooner than the ones that answered with nothing. The local node (`node.peerId`)
- * is excluded so we never RPC ourselves or seed with our own strand address.
- *
- * Dials run concurrently but the union preserves candidate order, so the result
- * is deterministic regardless of which sibling answers first.
- */
-export async function collectStrandAddrs(
-  node: Libp2p,
-  peers: StrandAddrPeer[],
-  strandId: string,
-  options: CollectStrandAddrsOptions = {}
-): Promise<StrandAddrCollection> {
-  const selfId = node.peerId.toString();
-  const request: StrandAddrRequest = options.delegatePeerId !== undefined
-    ? { strandId, delegatePeerId: options.delegatePeerId }
-    : { strandId };
-  const candidates = peers.filter(p => p.peerId !== selfId);
-
-  // Concurrent dials; `dialOneSibling` folds per-peer failure into an outcome, so
-  // `Promise.all` never rejects and the result array stays in candidate order.
-  const answers = await Promise.all(
-    candidates.map(async (peer) => ({ peerId: peer.peerId, ...await dialOneSibling(node, peer, request, options) }))
-  );
-
-  const outcomes = new Map<string, StrandAddrOutcome>();
-  // Deduplicated union in candidate order, then signaling-first for a usable
-  // dial sequence (relay/`p2p-circuit` ahead of direct addrs).
-  const seen = new Set<string>();
-  const union: string[] = [];
-  for (const { peerId, outcome, multiaddrs } of answers) {
-    outcomes.set(peerId, outcome);
-    for (const addr of multiaddrs) {
-      if (!seen.has(addr)) {
-        seen.add(addr);
-        union.push(addr);
-      }
-    }
-  }
-  return { addrs: orderSignalingFirst(union), outcomes };
-}
-
-/**
- * Ask one sibling for its strand address. Tries each dial target in order (peerId
- * first to reuse an open control connection, then explicit addrs) until one
- * produces a well-formed reply, and reports that reply's outcome; a total failure
- * is `unreachable`, so a single dead sibling never aborts the collection. Never
- * throws.
- *
- * A reply of any status ends the loop: every target reaches the same responder,
- * so asking it again by another address would get the same answer.
- *
- * A total failure logs ONE line naming every target and its cause, because
- * per-target lines scattered through a concurrent fan-out do not reassemble into
- * "this sibling was unreachable".
- *
- * NOTE: cost is (targets × `timeoutMs`, 28.5 s each at the default declared link) with
- * no whole-sibling budget, the same shape `dialWake` bounds with
- * `DEFAULT_WAKE_DIAL_BUDGET_MS`. Siblings are asked concurrently, so a collection costs
- * its slowest sibling, not the sum. Fine today — `dialTargets` yields the peerId plus
- * whatever `resolvePeerAddrs` returned, which for a cadre device is one or two
- * addresses. If a sibling's record ever carries a long address list, give this the same
- * whole-sibling budget.
- */
-async function dialOneSibling(
-  node: Libp2p,
-  peer: StrandAddrPeer,
-  request: StrandAddrRequest,
-  options: CollectStrandAddrsOptions
-): Promise<SiblingAnswer> {
-  const protocolId = options.protocolId ?? STRAND_ADDR_PROTOCOL;
-  const timeoutMs = attemptTimeoutMs(options);
-  const targets = dialTargets(peer);
-
-  if (targets.length === 0) {
-    log('No dial target for sibling %s', peer.peerId);
-    return { outcome: 'unreachable', multiaddrs: [] };
-  }
-
-  const failures: string[] = [];
-  for (const target of targets) {
-    try {
-      // One deadline per attempt: its signal aborts the in-flight dialProtocol and
-      // resets the live stream, so neither the connect nor the response-read leaks.
-      const response = await withDeadline(
-        timeoutMs,
-        `Strand-addr dial ${peer.peerId}`,
-        (signal) => sendStrandAddr(node, target, protocolId, request, timeoutMs, signal),
-      );
-      return siblingAnswer(response);
-    } catch (err) {
-      const error = err instanceof Error ? err : new Error(String(err));
-      failures.push(`${describeTarget(target)} — ${error.message}`);
-      log('Strand-addr dial to %s via %s failed: %o', peer.peerId, describeTarget(target), err);
-    }
-  }
-  log('Strand-addr dial to %s failed on all %d target(s), contributing no addrs: %s',
-    peer.peerId, failures.length, failures.join('; '));
-  return { outcome: 'unreachable', multiaddrs: [] };
-}
-
-/**
- * Deadline for ONE attempt to ask a sibling — dial, request, response — unless the caller set
- * its own: `relayedRequestBudgetMs` at the declared link (28.5 s at the default).
- *
- * An attempt is either a dial by peer id, which normally reuses an open control connection
- * (`CIRCUIT_REQUEST_ROUND_TRIPS`, 2), or the fallback to the sibling's control addresses, a
- * fresh and possibly relayed dial. One deadline covers both, so it is sized for the fallback.
- *
- * NOTE: the receiver's membership check (two live control reads) runs inside this deadline and
- * is not counted. In steady state those reads touch only held blocks and do not consult the
- * cohort. If a wake or address request is seen timing out while the receiver's membership read
- * is consulting, count one membership decision in this deadline or answer the check from the
- * materialized authorized-peer snapshot. See docs/cadre-consistency.md → "Deadlines Over
- * Optimystic's Reads and Commits".
- */
-function attemptTimeoutMs(options: CollectStrandAddrsOptions): number {
-  return options.timeoutMs ?? relayedRequestBudgetMs(options.linkRoundTripMs);
-}
-
-/** Map a validated reply to its outcome; only an `ok` reply contributes addresses. */
-function siblingAnswer(response: StrandAddrResponse): SiblingAnswer {
-  if (response.status !== 'ok') {
-    return { outcome: response.status, multiaddrs: [] };
-  }
-  return {
-    outcome: response.multiaddrs.length > 0 ? 'answered' : 'empty',
-    multiaddrs: response.multiaddrs
-  };
-}
-
-/**
- * Dial targets for a sibling, peerId first. Dialing by peerId lets libp2p reuse
- * an already-open control connection (the common case — the wiring ticket passes
- * connected siblings); the explicit control addrs are a fallback for when no
- * connection is open. An unparsable peerId is skipped (rely on addrs).
- */
-function dialTargets(peer: StrandAddrPeer): Array<PeerId | Multiaddr> {
-  const targets: Array<PeerId | Multiaddr> = [];
-  try {
-    targets.push(peerIdFromString(peer.peerId));
-  } catch (err) {
-    log('Unparsable peer id %s, falling back to addrs: %o', peer.peerId, err);
-  }
-  if (peer.addrs) {
-    targets.push(...peer.addrs);
-  }
-  return targets;
-}
-
-/** Short label for a dial target (peerId or multiaddr) for logging. */
-function describeTarget(target: PeerId | Multiaddr): string {
-  return target.toString();
-}
-
-/**
- * Open one stream to a target, send the request, half-close, and read the
- * response.
- *
- * `signal` is the per-attempt deadline from {@link dialOneSibling}: it goes to
- * `dialProtocol` so a timeout during connect aborts the dial, and into
- * {@link exchangeFrame} so a timeout after the stream is open resets it —
- * releasing the otherwise unbounded response-read.
- */
-async function sendStrandAddr(
-  node: Libp2p,
-  target: PeerId | Multiaddr,
-  protocolId: string,
-  request: StrandAddrRequest,
-  timeoutMs: number,
-  signal: AbortSignal,
-): Promise<StrandAddrResponse> {
-  // `runOnLimitedConnection: true`: the sibling may be reachable only over a
-  // circuit-relay connection (the signaling-first addr), which libp2p marks
-  // "limited". The exchange is a single tiny request→response well within the
-  // relay's data/duration cap, so opening it on the limited connection is safe
-  // and is the whole point of dialing over the relay.
-  const rawStream = await node.dialProtocol(target, protocolId, { runOnLimitedConnection: true, signal });
-
-  const response = await exchangeFrame(
-    rawStream as unknown as ControlStream,
-    signal,
-    request,
-    (stream) => readFrame<unknown>(stream, timeoutMs),
-    'Strand-addr dial aborted by timeout',
-  );
-  if (!isStrandAddrResponse(response)) {
-    throw new Error('Malformed strand-addr response');
-  }
-
-  log('Strand-addr response: %s, %d addr(s) for %s', response.status, response.multiaddrs.length, response.strandId);
-  return response;
-}
-
-const STRAND_ADDR_STATUSES: ReadonlySet<string> = new Set<StrandAddrStatus>(['ok', 'unavailable']);
-
-/**
- * Shape check on a decoded reply: the responder is another machine, so a reply
- * missing its status (an older responder, or a buggy one) or carrying anything but
- * strings in `multiaddrs` is a failed exchange, not an answer.
- *
- * Deliberately not `sanitizeStrandAddrs` (strand-formation-protocol.ts): that drops
- * bad entries and caps the list at 16, which suits a cross-party formation result
- * but would silently truncate a sibling's full address list here.
- */
-function isStrandAddrResponse(value: unknown): value is StrandAddrResponse {
-  if (typeof value !== 'object' || value === null) {
-    return false;
-  }
-  const { status, strandId, multiaddrs } = value as Record<string, unknown>;
-  return typeof status === 'string' && STRAND_ADDR_STATUSES.has(status)
-    && typeof strandId === 'string'
-    && Array.isArray(multiaddrs) && multiaddrs.every((addr) => typeof addr === 'string');
-}
+/**
+ * Control-network strand-address RPC.
+ *
+ * A strand runs as its own libp2p node (`strand-<id>`, random port), separate
+ * from the control node (`control-<partyId>`), with its own transport peerId
+ * derived from the cadre identity key (`strand-transport-key.ts`). To seed a
+ * strand's mesh, a node needs a sibling's **strand-network** address — but
+ * `CadrePeer.Multiaddr` only stores the **control** node's addresses, and a
+ * control address now names a different peer entirely, not merely a different
+ * port. This protocol resolves the strand address on demand: a node asks its
+ * connected co-cadre siblings "what are your live strand-`X` multiaddrs?" and
+ * uses the union of their answers as the seed.
+ *
+ * Modeled directly on `strand-wake-protocol.ts`: a dedicated libp2p protocol id,
+ * 4-byte big-endian length-prefixed JSON frames, `node.handle` for the receiver,
+ * `node.dialProtocol` for the client, and the shared `ControlStream` primitives
+ * from `control-stream.ts`. The exchange is a single request → single response on
+ * one stream, so each side reads to EOF (under a read timeout) and decodes one
+ * frame via the shared {@link decodeLengthPrefixedFrame} guard.
+ *
+ * **Authorization (v1):** like wake, the request carries no signature. Membership
+ * is checked before this handler runs, by the control node's protocol guard
+ * (`control-protocol-guard.ts`, the `members` class: `CadreNode.isAuthorizedMember`,
+ * a voucher on the requester's `CadrePeer` row verified against the node-local
+ * trusted-owner anchor). A peer it refuses gets its stream reset, never a reply, so
+ * an outsider that published its own rows into the replicated control DB can neither
+ * harvest live strand addresses nor announce a delegate. Cross-party strand bootstrap
+ * is a different mechanism (strand formation / `MemberPeer`) and is out of scope here.
+ */
+
+import debug from 'debug';
+import type { Libp2p, Connection, PeerId } from '@libp2p/interface';
+import type { Multiaddr } from '@multiformats/multiaddr';
+import { peerIdFromString } from '@libp2p/peer-id';
+import type { StrandAddrRequest, StrandAddrResponse, StrandAddrStatus } from './types.js';
+import { decodeLengthPrefixedFrame } from './seed-bootstrap.js';
+import { orderSignalingFirst } from './peer-record.js';
+import { type ControlStream, withDeadline, exchangeFrame, readStreamToEnd, replyAndClose } from './control-stream.js';
+import { relayedRequestBudgetMs } from './link-budget.js';
+
+const log = debug('sereus:cadre:strand-addr');
+
+/** Protocol id for the control-network strand-address RPC (parallel to `/sereus/strand-wake/1.0.0`). */
+export const STRAND_ADDR_PROTOCOL = '/sereus/strand-addr/1.0.0';
+
+/**
+ * Maximum strand-addr frame size. Responses are tiny (a strand id + a few
+ * multiaddrs), so this is a defensive cap — far below the 1MB seed ceiling —
+ * that bounds the bytes a peer can make the receiver buffer per stream.
+ */
+const MAX_ADDR_SIZE = 64 * 1024;
+
+/** Default time the receiver waits for an inbound request frame before aborting (ms). */
+// eslint-disable-next-line no-restricted-syntax -- link-independent: a receiver cap on one small request frame on a stream the peer already opened; it bounds a peer that opens a stream and never sends, not the dial
+const DEFAULT_ADDR_READ_TIMEOUT_MS = 10_000;
+
+/** Default cap on concurrent inbound strand-addr streams a single peer can pin open. */
+const DEFAULT_MAX_CONCURRENT_ADDRS = 100;
+
+/**
+ * Read a libp2p stream to EOF and decode the single length-prefixed JSON frame
+ * it carries. Bounded by `timeoutMs` (a never-half-closing peer is aborted, not
+ * awaited forever) and capped at {@link MAX_ADDR_SIZE}, reusing the shared
+ * {@link decodeLengthPrefixedFrame} guard for the prefix/length checks.
+ */
+async function readFrame<T>(stream: ControlStream, timeoutMs: number): Promise<T> {
+  const data = await readStreamToEnd(stream, { maxBytes: MAX_ADDR_SIZE, timeoutMs, label: 'Strand-addr' });
+  const body = decodeLengthPrefixedFrame(data, MAX_ADDR_SIZE);
+  return JSON.parse(new TextDecoder().decode(body)) as T;
+}
+
+/** An address-free reply, for every path but a lookup: over the cap, an unreadable request, a lookup that threw. */
+function addrlessResponse(status: StrandAddrStatus, strandId: string): StrandAddrResponse {
+  return { status, strandId, multiaddrs: [] };
+}
+
+/**
+ * Dependencies the {@link StrandAddrService} receiver needs from its host
+ * (`CadreNode`), injected so the service is testable without a full node.
+ */
+export interface StrandAddrServiceOptions {
+  /**
+   * The local strand instance's dialable strand-network multiaddrs, ordered
+   * signaling-first. Returns `[]` when the strand is not running / has no live
+   * node (hibernating, quiescing, or never participated). Kept injectable so the
+   * decision logic is unit-testable without a full node.
+   */
+  getStrandMultiaddrs(strandId: string): string[];
+  /**
+   * Called when a member's request carries a `delegatePeerId` — the derived
+   * transport peerId its strand-`strandId` node runs as, with a validated
+   * (parseable, non-self) peerId. Every caller is an authorized member: the
+   * protocol guard admitted it before this handler ran.
+   * `CadreNode` injects its delegate-admission grant recorder so this node's
+   * connection gate — and thus its circuit-relay server, when it runs one —
+   * admits that peerId (see `delegate-admission.ts`). Optional: absent, an
+   * announce is ignored and the address lookup proceeds unchanged.
+   */
+  onDelegateAnnounce?(announcerPeerId: string, strandId: string, delegatePeerId: string): void;
+  /**
+   * Time to wait for the inbound request frame before aborting the read (ms).
+   * Defaults to {@link DEFAULT_ADDR_READ_TIMEOUT_MS}. Bounds a buggy/compromised
+   * own-cadre node that opens a stream and never half-closes its write end.
+   */
+  readTimeoutMs?: number;
+  /**
+   * Cap on concurrent inbound strand-addr streams (defaults to
+   * {@link DEFAULT_MAX_CONCURRENT_ADDRS}). Over the cap, an `unavailable` reply is
+   * returned without looking up any strand address.
+   */
+  maxConcurrent?: number;
+}
+
+/**
+ * Receiver side of the strand-address RPC. Registers a `STRAND_ADDR_PROTOCOL`
+ * handler on the control node and, for each inbound {@link StrandAddrRequest}
+ * (from a peer the protocol guard has already admitted as an authorized member),
+ * replies with the local strand instance's live multiaddrs — `ok` with an empty
+ * list when the strand is not running, `unavailable` when it could not answer at all.
+ */
+export class StrandAddrService {
+  private readonly options: StrandAddrServiceOptions;
+  private readonly readTimeoutMs: number;
+  private readonly maxConcurrent: number;
+  private node: Libp2p | null = null;
+  /** In-flight inbound strand-addr streams, used to enforce {@link maxConcurrent}. */
+  private activeStreams = 0;
+
+  constructor(options: StrandAddrServiceOptions) {
+    this.options = options;
+    this.readTimeoutMs = options.readTimeoutMs ?? DEFAULT_ADDR_READ_TIMEOUT_MS;
+    this.maxConcurrent = options.maxConcurrent ?? DEFAULT_MAX_CONCURRENT_ADDRS;
+  }
+
+  /** Number of in-flight inbound strand-addr streams. */
+  get activeCount(): number {
+    return this.activeStreams;
+  }
+
+  /**
+   * Register the strand-addr protocol handler on the control node. Rejects when
+   * libp2p refuses the registration (a duplicate handler, a peer-store write
+   * failure); the node reference is kept only once the handler is in place.
+   */
+  async initialize(node: Libp2p): Promise<void> {
+    // `runOnLimitedConnection: true` is REQUIRED for the relay path: a NAT'd
+    // sibling is reached over a circuit-relay connection, which libp2p marks
+    // "limited" (the relay caps its data/duration). Without this the receiver
+    // would refuse the inbound stream on exactly the connection the protocol is
+    // designed to use (same reasoning as wake; see the relay note on
+    // `collectStrandAddrs`).
+    await node.handle(STRAND_ADDR_PROTOCOL, async (rawStream: unknown, rawConnection: unknown) => {
+      const remotePeerId = (rawConnection as Connection).remotePeer.toString();
+      await this.handleStream(rawStream as ControlStream, remotePeerId);
+    }, { runOnLimitedConnection: true });
+    this.node = node;
+    log('StrandAddrService registered handler: %s', STRAND_ADDR_PROTOCOL);
+  }
+
+  /** Unregister the handler and release the node reference. */
+  async shutdown(): Promise<void> {
+    if (this.node) {
+      await this.node.unhandle(STRAND_ADDR_PROTOCOL);
+      this.node = null;
+      log('StrandAddrService shutdown');
+    }
+  }
+
+  /**
+   * Read the inbound request, decide the response, and write it back.
+   *
+   * Three hardening layers, all reported as an `unavailable`
+   * {@link StrandAddrResponse} rather than a dropped/hung stream: a concurrency cap
+   * (over {@link maxConcurrent}, reply without looking up any address), a read
+   * timeout (a peer that never half-closes is aborted inside {@link readFrame}/
+   * `readStreamToEnd`), and the existing malformed/oversized-frame guard. A
+   * lookup that throws is `unavailable` too, never an empty `ok` the asker would
+   * wait ten minutes on.
+   */
+  private async handleStream(stream: ControlStream, remotePeerId: string): Promise<void> {
+    log('Incoming strand-addr request from: %s', remotePeerId);
+
+    if (this.activeStreams >= this.maxConcurrent) {
+      log('Rejecting strand-addr from %s: %d concurrent streams at cap %d', remotePeerId, this.activeStreams, this.maxConcurrent);
+      // The request frame is unread here, so the strand id is unknown — reply
+      // under an empty strand id; the asker retries an `unavailable` sibling soon.
+      await replyAndClose(stream, addrlessResponse('unavailable', ''), 'Strand-addr');
+      return;
+    }
+
+    this.activeStreams++;
+    try {
+      await replyAndClose(stream, await this.answerStream(stream, remotePeerId), 'Strand-addr');
+    } finally {
+      this.activeStreams--;
+    }
+  }
+
+  /** Read and decide one inbound request; any failure becomes an `unavailable` response. */
+  private async answerStream(stream: ControlStream, remotePeerId: string): Promise<StrandAddrResponse> {
+    try {
+      const request = await readFrame<StrandAddrRequest>(stream, this.readTimeoutMs);
+      return await this.processAddrRequest(request, remotePeerId);
+    } catch (err) {
+      log('Error handling strand-addr request from %s: %o', remotePeerId, err);
+      // Malformed/oversized/timed-out request, or a lookup that threw: reply
+      // under an empty strand id rather than hanging.
+      return addrlessResponse('unavailable', '');
+    }
+  }
+
+  /**
+   * Decide the response for a decoded request. Exposed (not private) so the
+   * decision matrix can be unit-tested directly (mirrors wake's
+   * `processWakeRequest`).
+   *
+   * - Request carrying a `delegatePeerId` → recorded via
+   *   {@link StrandAddrServiceOptions.onDelegateAnnounce} before the lookup.
+   * - Strand not running locally → `ok` with empty `multiaddrs` (`getStrandMultiaddrs` → `[]`).
+   * - Running strand → `ok` with the strand's live, signaling-first multiaddrs.
+   *
+   * A throwing lookup propagates; {@link handleStream} answers it `unavailable`.
+   */
+  async processAddrRequest(request: StrandAddrRequest, remotePeerId: string): Promise<StrandAddrResponse> {
+    this.recordDelegateAnnounce(request, remotePeerId);
+
+    const multiaddrs = this.options.getStrandMultiaddrs(request.strandId);
+    log('Strand-addr for %s → %d addr(s)', request.strandId, multiaddrs.length);
+    return { status: 'ok', strandId: request.strandId, multiaddrs };
+  }
+
+  /**
+   * Validate and forward a member's delegate announcement. A malformed field
+   * (unparsable peerId) and a self-announcement (`delegatePeerId` equal to the
+   * announcer's own peerId — a member is admitted directly, never by grant)
+   * are logged and dropped; the address lookup proceeds either way.
+   */
+  private recordDelegateAnnounce(request: StrandAddrRequest, remotePeerId: string): void {
+    const { strandId, delegatePeerId } = request;
+    if (delegatePeerId === undefined || !this.options.onDelegateAnnounce) {
+      return;
+    }
+    if (delegatePeerId === remotePeerId) {
+      log('Ignoring self-announcement from %s (strand %s)', remotePeerId, strandId);
+      return;
+    }
+    try {
+      peerIdFromString(delegatePeerId);
+    } catch (err) {
+      log('Ignoring malformed delegatePeerId %o from %s: %o', delegatePeerId, remotePeerId, err);
+      return;
+    }
+    this.options.onDelegateAnnounce(remotePeerId, strandId, delegatePeerId);
+  }
+}
+
+/** A sibling to ask for its strand address. */
+export interface StrandAddrPeer {
+  /** The sibling's control-network peer id (also its strand-node peer id). */
+  peerId: string;
+  /**
+   * Pre-resolved control-network multiaddrs, used as a fallback when no control
+   * connection is already open. Dialing by {@link peerId} is preferred (it
+   * reuses an existing connection); these are tried only if that fails.
+   */
+  addrs?: Multiaddr[];
+}
+
+/** Options for {@link collectStrandAddrs}. */
+export interface CollectStrandAddrsOptions {
+  /**
+   * The asker's declared link round trip (`NetworkConfig.linkRoundTripMs`), from which the
+   * per-dial timeout is derived when {@link timeoutMs} is not given. Unset means the declared
+   * default.
+   */
+  linkRoundTripMs?: number;
+  /** Per-dial timeout in ms (default: {@link attemptTimeoutMs}). */
+  timeoutMs?: number;
+  /** Override the protocol id (defaults to {@link STRAND_ADDR_PROTOCOL}). */
+  protocolId?: string;
+  /**
+   * The derived transport peerId the caller's own strand-`strandId` node runs
+   * (or is about to run) as. Set on every request this collection sends, so
+   * each receiver records a delegate admission grant for it — see
+   * {@link StrandAddrRequest.delegatePeerId}. Omit for a pure address lookup.
+   */
+  delegatePeerId?: string;
+}
+
+/**
+ * What asking one sibling produced, as {@link collectStrandAddrs} reports it.
+ *
+ * - `answered` — status `ok` with at least one address.
+ * - `empty` — status `ok` with no address: the sibling does not run the strand
+ *   right now, a normal steady state.
+ * - `unavailable` — the sibling replied that it could not answer.
+ * - `unsupported` — the peer was reached and its protocol negotiation refused
+ *   strand-addr: it registers no handler for it (a dedicated relay, an older
+ *   build) or has not registered one yet (a node still starting).
+ * - `unreachable` — no reply at all: every dial target failed or timed out, the
+ *   reply was malformed, or the sibling's protocol guard reset the stream because
+ *   it does not (yet) count the asker as an authorized member.
+ */
+export type StrandAddrOutcome = 'answered' | 'empty' | 'unavailable' | 'unsupported' | 'unreachable';
+
+/** Result of {@link collectStrandAddrs}. */
+export interface StrandAddrCollection {
+  /** Deduplicated union of every answer, signaling-first. */
+  addrs: string[];
+  /** One entry per candidate (self excluded), keyed by the sibling's control peerId. */
+  outcomes: Map<string, StrandAddrOutcome>;
+}
+
+/** One sibling's share of a {@link StrandAddrCollection}. */
+interface SiblingAnswer {
+  outcome: StrandAddrOutcome;
+  multiaddrs: string[];
+}
+
+/**
+ * Client side: ask each candidate sibling for its live strand-`strandId`
+ * multiaddrs and return the **deduplicated union** of every answer, ordered
+ * signaling-first, alongside each sibling's {@link StrandAddrOutcome}.
+ *
+ * Best-effort per peer: a failed/timed-out/empty sibling contributes no address
+ * and never fails the collection — an empty union (no sibling online or running
+ * the strand) is an acceptable seed that self-heals on the next resume/reconcile
+ * pass. `outcomes` is what lets a caller retry the siblings that could not answer
+ * sooner than the ones that answered with nothing. The local node (`node.peerId`)
+ * is excluded so we never RPC ourselves or seed with our own strand address.
+ *
+ * Dials run concurrently but the union preserves candidate order, so the result
+ * is deterministic regardless of which sibling answers first.
+ */
+export async function collectStrandAddrs(
+  node: Libp2p,
+  peers: StrandAddrPeer[],
+  strandId: string,
+  options: CollectStrandAddrsOptions = {}
+): Promise<StrandAddrCollection> {
+  const selfId = node.peerId.toString();
+  const request: StrandAddrRequest = options.delegatePeerId !== undefined
+    ? { strandId, delegatePeerId: options.delegatePeerId }
+    : { strandId };
+  const candidates = peers.filter(p => p.peerId !== selfId);
+
+  // Concurrent dials; `dialOneSibling` folds per-peer failure into an outcome, so
+  // `Promise.all` never rejects and the result array stays in candidate order.
+  const answers = await Promise.all(
+    candidates.map(async (peer) => ({ peerId: peer.peerId, ...await dialOneSibling(node, peer, request, options) }))
+  );
+
+  const outcomes = new Map<string, StrandAddrOutcome>();
+  // Deduplicated union in candidate order, then signaling-first for a usable
+  // dial sequence (relay/`p2p-circuit` ahead of direct addrs).
+  const seen = new Set<string>();
+  const union: string[] = [];
+  for (const { peerId, outcome, multiaddrs } of answers) {
+    outcomes.set(peerId, outcome);
+    for (const addr of multiaddrs) {
+      if (!seen.has(addr)) {
+        seen.add(addr);
+        union.push(addr);
+      }
+    }
+  }
+  return { addrs: orderSignalingFirst(union), outcomes };
+}
+
+/**
+ * Ask one sibling for its strand address. Tries each dial target in order (peerId
+ * first to reuse an open control connection, then explicit addrs) until one
+ * produces a well-formed reply, and reports that reply's outcome; a total failure
+ * is `unreachable`, so a single dead sibling never aborts the collection. Never
+ * throws.
+ *
+ * A reply of any status, or a refused protocol negotiation (`unsupported`), ends
+ * the loop: every target reaches the same responder, so asking it again by another
+ * address would get the same answer.
+ *
+ * A total failure logs ONE line naming every target and its cause, because
+ * per-target lines scattered through a concurrent fan-out do not reassemble into
+ * "this sibling was unreachable".
+ *
+ * NOTE: cost is (targets × `timeoutMs`, 28.5 s each at the default declared link) with
+ * no whole-sibling budget, the same shape `dialWake` bounds with
+ * `DEFAULT_WAKE_DIAL_BUDGET_MS`. Siblings are asked concurrently, so a collection costs
+ * its slowest sibling, not the sum. Fine today — `dialTargets` yields the peerId plus
+ * whatever `resolvePeerAddrs` returned, which for a cadre device is one or two
+ * addresses. If a sibling's record ever carries a long address list, give this the same
+ * whole-sibling budget.
+ */
+async function dialOneSibling(
+  node: Libp2p,
+  peer: StrandAddrPeer,
+  request: StrandAddrRequest,
+  options: CollectStrandAddrsOptions
+): Promise<SiblingAnswer> {
+  const protocolId = options.protocolId ?? STRAND_ADDR_PROTOCOL;
+  const timeoutMs = attemptTimeoutMs(options);
+  const targets = dialTargets(peer);
+
+  if (targets.length === 0) {
+    log('No dial target for sibling %s', peer.peerId);
+    return { outcome: 'unreachable', multiaddrs: [] };
+  }
+
+  const failures: string[] = [];
+  for (const target of targets) {
+    try {
+      // One deadline per attempt: its signal aborts the in-flight dialProtocol and
+      // resets the live stream, so neither the connect nor the response-read leaks.
+      const response = await withDeadline(
+        timeoutMs,
+        `Strand-addr dial ${peer.peerId}`,
+        (signal) => sendStrandAddr(node, target, protocolId, request, timeoutMs, signal),
+      );
+      return siblingAnswer(response);
+    } catch (err) {
+      const error = err instanceof Error ? err : new Error(String(err));
+      // Matched by name, as libp2p's own error classes are compared across package copies.
+      if (error.name === 'UnsupportedProtocolError') {
+        log('Strand-addr refused by %s via %s: it does not serve %s', peer.peerId, describeTarget(target), protocolId);
+        return { outcome: 'unsupported', multiaddrs: [] };
+      }
+      failures.push(`${describeTarget(target)} — ${error.message}`);
+      log('Strand-addr dial to %s via %s failed: %o', peer.peerId, describeTarget(target), err);
+    }
+  }
+  log('Strand-addr dial to %s failed on all %d target(s), contributing no addrs: %s',
+    peer.peerId, failures.length, failures.join('; '));
+  return { outcome: 'unreachable', multiaddrs: [] };
+}
+
+/**
+ * Deadline for ONE attempt to ask a sibling — dial, request, response — unless the caller set
+ * its own: `relayedRequestBudgetMs` at the declared link (28.5 s at the default).
+ *
+ * An attempt is either a dial by peer id, which normally reuses an open control connection
+ * (`CIRCUIT_REQUEST_ROUND_TRIPS`, 2), or the fallback to the sibling's control addresses, a
+ * fresh and possibly relayed dial. One deadline covers both, so it is sized for the fallback.
+ *
+ * NOTE: the receiver's membership check (two live control reads) runs inside this deadline and
+ * is not counted. In steady state those reads touch only held blocks and do not consult the
+ * cohort. If a wake or address request is seen timing out while the receiver's membership read
+ * is consulting, count one membership decision in this deadline or answer the check from the
+ * materialized authorized-peer snapshot. See docs/cadre-consistency.md → "Deadlines Over
+ * Optimystic's Reads and Commits".
+ */
+function attemptTimeoutMs(options: CollectStrandAddrsOptions): number {
+  return options.timeoutMs ?? relayedRequestBudgetMs(options.linkRoundTripMs);
+}
+
+/** Map a validated reply to its outcome; only an `ok` reply contributes addresses. */
+function siblingAnswer(response: StrandAddrResponse): SiblingAnswer {
+  if (response.status !== 'ok') {
+    return { outcome: response.status, multiaddrs: [] };
+  }
+  return {
+    outcome: response.multiaddrs.length > 0 ? 'answered' : 'empty',
+    multiaddrs: response.multiaddrs
+  };
+}
+
+/**
+ * Dial targets for a sibling, peerId first. Dialing by peerId lets libp2p reuse
+ * an already-open control connection (the common case — the wiring ticket passes
+ * connected siblings); the explicit control addrs are a fallback for when no
+ * connection is open. An unparsable peerId is skipped (rely on addrs).
+ */
+function dialTargets(peer: StrandAddrPeer): Array<PeerId | Multiaddr> {
+  const targets: Array<PeerId | Multiaddr> = [];
+  try {
+    targets.push(peerIdFromString(peer.peerId));
+  } catch (err) {
+    log('Unparsable peer id %s, falling back to addrs: %o', peer.peerId, err);
+  }
+  if (peer.addrs) {
+    targets.push(...peer.addrs);
+  }
+  return targets;
+}
+
+/** Short label for a dial target (peerId or multiaddr) for logging. */
+function describeTarget(target: PeerId | Multiaddr): string {
+  return target.toString();
+}
+
+/**
+ * Open one stream to a target, send the request, half-close, and read the
+ * response.
+ *
+ * `signal` is the per-attempt deadline from {@link dialOneSibling}: it goes to
+ * `dialProtocol` so a timeout during connect aborts the dial, and into
+ * {@link exchangeFrame} so a timeout after the stream is open resets it —
+ * releasing the otherwise unbounded response-read.
+ */
+async function sendStrandAddr(
+  node: Libp2p,
+  target: PeerId | Multiaddr,
+  protocolId: string,
+  request: StrandAddrRequest,
+  timeoutMs: number,
+  signal: AbortSignal,
+): Promise<StrandAddrResponse> {
+  // `runOnLimitedConnection: true`: the sibling may be reachable only over a
+  // circuit-relay connection (the signaling-first addr), which libp2p marks
+  // "limited". The exchange is a single tiny request→response well within the
+  // relay's data/duration cap, so opening it on the limited connection is safe
+  // and is the whole point of dialing over the relay.
+  const rawStream = await node.dialProtocol(target, protocolId, { runOnLimitedConnection: true, signal });
+
+  const response = await exchangeFrame(
+    rawStream as unknown as ControlStream,
+    signal,
+    request,
+    (stream) => readFrame<unknown>(stream, timeoutMs),
+    'Strand-addr dial aborted by timeout',
+  );
+  if (!isStrandAddrResponse(response)) {
+    throw new Error('Malformed strand-addr response');
+  }
+
+  log('Strand-addr response: %s, %d addr(s) for %s', response.status, response.multiaddrs.length, response.strandId);
+  return response;
+}
+
+const STRAND_ADDR_STATUSES: ReadonlySet<string> = new Set<StrandAddrStatus>(['ok', 'unavailable']);
+
+/**
+ * Shape check on a decoded reply: the responder is another machine, so a reply
+ * missing its status (an older responder, or a buggy one) or carrying anything but
+ * strings in `multiaddrs` is a failed exchange, not an answer.
+ *
+ * Deliberately not `sanitizeStrandAddrs` (strand-formation-protocol.ts): that drops
+ * bad entries and caps the list at 16, which suits a cross-party formation result
+ * but would silently truncate a sibling's full address list here.
+ */
+function isStrandAddrResponse(value: unknown): value is StrandAddrResponse {
+  if (typeof value !== 'object' || value === null) {
+    return false;
+  }
+  const { status, strandId, multiaddrs } = value as Record<string, unknown>;
+  return typeof status === 'string' && STRAND_ADDR_STATUSES.has(status)
+    && typeof strandId === 'string'
+    && Array.isArray(multiaddrs) && multiaddrs.every((addr) => typeof addr === 'string');
+}