@serfab/cadre-core 1.9.0 → 1.10.0

package/src/cadre-node.ts

@@ -36,7 +36,9 @@ import type {
   CadrePeerVoucherFields,
   CadrePeerRow,
   PeerAddressRecord,
-  ResolveDeviceTokenOpts
+  ResolveDeviceTokenOpts,
+  PendingJoinRow,
+  PendingJoinStatus
 } from './types.js';
 import { controlClusterPolicy, CONTROL_REPLICATION_BREADTH, DEFAULT_CHECKIN_WINDOW_MS, DEFAULT_CONNECTION_MONITOR } from './types.js';
 import { sign } from '@optimystic/quereus-plugin-crypto';
@@ -103,10 +105,18 @@ import {
   CONTROL_COHORT_DIAL_ADDRESS_ATTEMPTS,
   type PeerDialBudget
 } from './peer-dial.js';
-import { ADMISSION_DECISION_TIMEOUT_MS, declaredCohortReadDeadlineMs, peerJoinPushBudget, relayAdmissionReserveDeadlineMs, relayReservationBudgetMs, relayedDialBudgetMs, resolveLinkRoundTripMs } from './link-budget.js';
+import { ADMISSION_DECISION_TIMEOUT_MS, declaredCohortReadDeadlineMs, optimysticDialLimits, peerJoinPushBudget, relayAdmissionReserveDeadlineMs, relayReservationBudgetMs, relayedDialBudgetMs, resolveLinkRoundTripMs } from './link-budget.js';
 import { EnrollmentService } from './enrollment.js';
 import { HibernationManager, type HibernationCallbacks } from './hibernation-manager.js';
-import { ControlDatabase, isStrandIdConflict, type RevokedRowRef } from './control-database.js';
+import { ControlDatabase, isPendingJoinConflict, isStrandIdConflict, pendingJoinId, type RevokedRowRef } from './control-database.js';
+import { FormationPostApprovalError } from './strand-formation-rejection.js';
+import {
+  PendingJoinRunner,
+  membershipInvitesToStage,
+  parseStoredDisclosure,
+  requestedPendingJoin,
+  type PendingJoinFields
+} from './pending-join-runner.js';
 import type { ControlRetryAbandonment } from './control-retry.js';
 import { SeedBootstrapService, type SeedEventCallbacks, type SeedBootstrapConfig } from './seed-bootstrap.js';
 import type { SeedTrustPolicy } from './seed-trust-policy.js';
@@ -114,6 +124,8 @@ import {
   StrandSolicitationService,
   type StrandSolicitationServiceOptions
 } from './strand-solicitation.js';
+import { ControlFormationUsageRecorder } from './control-formation-recorder.js';
+import { selectInvitationSiblingAddrs, type InvitationSibling } from './invitation-bootstrap.js';
 import {
   createMembershipConnectionGater,
   DEFAULT_ENROLLMENT_WINDOW_MS,
@@ -378,6 +390,15 @@ export class CadreNode implements SAppIdLookup {
   private enrollmentService: EnrollmentService;
   private seedBootstrapService: SeedBootstrapService | null = null;
   private strandSolicitationService: StrandSolicitationService | null = null;
+  /**
+   * The service whose formation handler is registered on the control node. Trails
+   * {@link strandSolicitationService} while {@link initializeStrandSolicitation} swaps a new
+   * one in: the field is set at once so concurrent callers reuse it, the handler only when
+   * its queued swap runs.
+   */
+  private registeredSolicitation: StrandSolicitationService | null = null;
+  /** Tail of the queued formation-handler swaps ({@link swapFormationResponder}). */
+  private solicitationSwaps: Promise<void> = Promise.resolve();
   private strandWakeService: StrandWakeService | null = null;
   /**
    * Control-network strand-address responder. Answers a co-cadre sibling's
@@ -696,6 +717,22 @@ export class CadreNode implements SAppIdLookup {
    */
   private readonly pendingMembershipInvites = new Map<string, StrandMembershipInvite>();
 
+  /**
+   * Invite keys of the membership invitations this process has staged, by
+   * {@link adoptFormationMembershipInvite} or from a `joined` `PendingJoin` row
+   * ({@link stageMembershipInvitesFromPendingJoins}), so one the reconciler has settled is not
+   * staged again by the next pending-join pass. Same lifetime as
+   * {@link pendingMembershipInvites}: a restarted process stages each live one once more, and the
+   * reconciler settles it again (spent, or the party already seated).
+   */
+  private readonly stagedMembershipInviteKeys = new Set<string>();
+
+  /**
+   * The pending-join retry loop (`pending-join-runner.ts`): built by {@link start}, stopped by
+   * {@link cleanup}. It acts only while this machine is an enrolled owner.
+   */
+  private pendingJoinRunner: PendingJoinRunner | null = null;
+
   /**
    * Founder launches refused because the strand was founded before the per-party
    * identity split (`PreSplitStrandIdentityError`), keyed by strandId. Read by
@@ -1266,6 +1303,12 @@ export class CadreNode implements SAppIdLookup {
       });
       await this.strandAddrService.initialize(this.controlNode);
 
+      // Answer strand formation for this party from now on. Every machine of a party
+      // holds the replicated FormationInvite/FormationUsage rows, so an always-on one can
+      // check a token for an inviter that is offline. After the control database is up,
+      // so the responder never reads a database that cannot answer yet.
+      await this.installDefaultFormationResponder();
+
       // Server-side push-wake fan-out: only when push is configured.
       if (this.config.push) {
         this.pushFanoutService = this.buildPushFanout(this.config.push);
@@ -1306,6 +1349,9 @@ export class CadreNode implements SAppIdLookup {
       // Schedule self-registration in background
       this.scheduleSelfRegistration();
 
+      // Last: its first pass may dial another party's machines.
+      this.startPendingJoinRunner();
+
     } catch (error) {
       log('Failed to start CadreNode: %o', error);
       // `driveControlRelayReservation` throws AFTER `control:connected` was emitted,
@@ -1904,11 +1950,15 @@ export class CadreNode implements SAppIdLookup {
       // replaces the default whole (see DEFAULT_CONNECTION_MONITOR).
       connectionMonitor: network?.connectionMonitor ?? DEFAULT_CONNECTION_MONITOR,
       // The declared link every cadre budget is derived from, stated to Optimystic too so it
-      // derives its own deadlines from the same number: its request dials and responses, its
-      // block pushes, and libp2p's `dialTimeout` and `inboundUpgradeTimeout`. Always stated,
-      // default included: undeclared, Optimystic keeps LAN deadlines that cannot open a
-      // relayed connection at the link sereus supports (`link-budget.ts`).
+      // derives its own deadlines from the same number: its request responses, its block
+      // pushes, and libp2p's `inboundUpgradeTimeout`. Always stated, default included:
+      // undeclared, Optimystic keeps LAN deadlines that cannot open a relayed connection at the
+      // link sereus supports (`link-budget.ts`).
       linkRoundTripMs: resolveLinkRoundTripMs(network?.linkRoundTripMs),
+      // The three limits on opening a connection (libp2p's per-address and whole-dial limits,
+      // and Optimystic's request dial): Optimystic's derivation from the same link plus the
+      // relay's and the called machine's admission decisions, which run on this dial's clock.
+      ...optimysticDialLimits(network?.linkRoundTripMs),
       // `{ wsPort }` when a listen entry names WebSocket, otherwise `{}` — and always
       // `{}` when `network.transports` is set, since the embedder owns transport policy
       // then. Spread NEXT to `transports` because the two answer the same question.
@@ -1991,11 +2041,9 @@ export class CadreNode implements SAppIdLookup {
    *     `StrandSolicitationService.hasOutstandingInvitation`). A formation
    *     initiator is another party's peer by design and its token is only
    *     checkable inside the protocol, so the gate asks the coarser question
-   *     "does this node expect a stranger at all?". Merely REGISTERING the
-   *     responder ({@link initializeStrandSolicitation}) no longer suspends
-   *     stranger denial — eager registration (as `reference-app-rn` does at
-   *     bring-up) and `formStrand`'s lazy initialization both leave the gate
-   *     armed, because neither mints an invitation.
+   *     "does this node expect a stranger at all?". REGISTERING the responder
+   *     does not suspend stranger denial: every node registers one at
+   *     {@link start}, and registering mints no invitation.
    *
    * Ordering is semantically free (the checks are OR'd) but decides who pays:
    * checks 1-3 are in-memory, 4/5 share one control-DB read, and only a peer
@@ -2027,6 +2075,11 @@ export class CadreNode implements SAppIdLookup {
    * layer is fail-open behind `ADMISSION_DECISION_TIMEOUT_MS`, so the live
    * read is safe here — unlike the per-stream gate, which must consult the
    * materialized {@link authorizedControlPeers} snapshot instead.
+   * NOTE: check 6 adds a second control read (`hasOutstandingFormationInvite`) for a
+   * stranger with no locally minted invitation in play, on every node now that every
+   * node runs the responder — a relay-enabled storage node included. If stranger
+   * connections to such a node ever arrive fast enough for that read to show, cache
+   * the answer for a few seconds.
    * NOTE: on a relay-DISABLED node, a sibling whose membership row has not yet
    * replicated here is denied until the row converges (typically via the
    * owner); either side's next outbound reconcile dial (outbound is never
@@ -2918,6 +2971,15 @@ export class CadreNode implements SAppIdLookup {
    * handling, is what keeps a mixed list from taking a whole peer offline.
    */
   async resolvePeerAddrs(peerId: string, opts: ResolveOpts = {}): Promise<Multiaddr[]> {
+    return (await this.resolvePeerRecord(peerId, opts))?.addrs ?? [];
+  }
+
+  /**
+   * {@link resolvePeerAddrs} with the record's `UpdatedAt` kept beside the addresses, for a
+   * caller that ranks peers by how recently they published. `null` wherever that returns `[]`
+   * for a failed gate.
+   */
+  private async resolvePeerRecord(peerId: string, opts: ResolveOpts): Promise<{ addrs: Multiaddr[]; updatedAt: number } | null> {
     if (!this.controlDatabase) {
       throw new Error('CadreNode must be started before resolving peer addrs');
     }
@@ -2925,14 +2987,14 @@ export class CadreNode implements SAppIdLookup {
     const record = await this.controlDatabase.queryPeerRecord(peerId);
     if (!record) {
       log('resolvePeerAddrs: no record for %s', peerId);
-      return [];
+      return null;
     }
 
     // publicKey <-> peerId binding: the stored key must be the one embedded in
     // the requested Ed25519 peer id (also rejects a non-Ed25519 / missing key).
     if (!record.publicKey || ed25519PublicKeyB64FromPeerId(peerId) !== record.publicKey) {
       log('resolvePeerAddrs: publicKey does not match peerId for %s', peerId);
-      return [];
+      return null;
     }
 
     // Self-signature over (peerId, addrs, updatedAt).
@@ -2943,14 +3005,14 @@ export class CadreNode implements SAppIdLookup {
       log('resolvePeerAddrs: signature verification failed for %s (updatedAt=%d, addrs=%o, sig=%s)',
         peerId, record.updatedAt, record.addrs,
         record.sig ? `${record.sig.slice(0, 16)}…` : '(empty)');
-      return [];
+      return null;
     }
 
     // Freshness: never hand back a dead relay reservation.
     const maxAgeMs = opts.maxAgeMs ?? DEFAULT_PEER_RECORD_MAX_AGE_MS;
     if (!isPeerRecordFresh(record.updatedAt, maxAgeMs, Date.now())) {
       log('resolvePeerAddrs: record for %s is stale (updatedAt=%d, maxAgeMs=%d)', peerId, record.updatedAt, maxAgeMs);
-      return [];
+      return null;
     }
 
     // Pluggable trust gate (defaults to current-member).
@@ -2963,7 +3025,7 @@ export class CadreNode implements SAppIdLookup {
     });
     if (!trusted) {
       log('resolvePeerAddrs: trust policy rejected %s', peerId);
-      return [];
+      return null;
     }
 
     // Order signaling-first (the on-record order was what we verified above),
@@ -2974,7 +3036,7 @@ export class CadreNode implements SAppIdLookup {
     if (opts.signalingOnly) {
       addrs = addrs.filter(isSignalingAddr);
     }
-    return this.normalizeDialAddrs(this.parseMultiaddrs(addrs), peerId);
+    return { addrs: this.normalizeDialAddrs(this.parseMultiaddrs(addrs), peerId), updatedAt: record.updatedAt };
   }
 
   /**
@@ -3535,15 +3597,18 @@ export class CadreNode implements SAppIdLookup {
    *
    * The list may name transports this node cannot dial (a lent node reports TCP
    * and `/ws` addresses to a phone that dials WebSockets only). That needs no
-   * filtering here: libp2p's dial queue (`calculateMultiaddrs`, libp2p 3.1.3)
+   * filtering here: libp2p's dial queue (`calculateMultiaddrs`, libp2p 3.3.11)
    * rejects an address no transport can dial before touching the network, so
    * such an address costs a log line.
    *
    * The list is NOT handed to one `dial()`. libp2p tries a multi-address dial's
-   * addresses one after another under a single deadline, with no limit per
-   * address, and sorts loopback addresses last — so one or two addresses that
-   * never answer used up the whole deadline before the address that worked was
-   * tried. {@link dialPeerAddrs} dials each address on its own time limit instead.
+   * addresses one after another under a single deadline, and sorts loopback
+   * addresses last — so one or two addresses that never answer use up most of
+   * the deadline before the address that works is tried. libp2p 3.3's
+   * per-address `addressDialTimeout` does not prevent that: cadre sizes it for a
+   * cold relayed open through two admission decisions (`optimysticDialLimits`,
+   * 39 s at the default declared link).
+   * {@link dialPeerAddrs} dials each address on its own time limit instead.
    */
   private async resolveControlDialAddrs(peerId: string, resolved: Multiaddr[]): Promise<Multiaddr[]> {
     if (resolved.length > 0) {
@@ -3932,6 +3997,9 @@ export class CadreNode implements SAppIdLookup {
 
     // 4. In-session pending owner authorize writes.
     await this.drainPendingPeerWrites();
+
+    // 5. `PendingJoin` rows this process wrote while alone. Their tombstones went out in step 1.
+    await this.pendingJoinRunner?.reissueWritesMadeAlone();
   }
 
   /**
@@ -4611,6 +4679,10 @@ export class CadreNode implements SAppIdLookup {
     // it is stopped here.
     this.relayReserveSupervisor?.stop();
     this.relayReserveSupervisor = null;
+    // Before anything it reads is torn down. An attempt in flight is not awaited (NOTE at
+    // `PendingJoinRunner.stop`).
+    this.pendingJoinRunner?.stop();
+    this.pendingJoinRunner = null;
     // Drop the rest of the relay-reservation posture: a torn-down node holds no
     // reservation, so neither a restarted instance nor a caller inspecting a failed
     // start reports the previous attempt's `reserved`.
@@ -4658,11 +4730,14 @@ export class CadreNode implements SAppIdLookup {
       this.pushFanoutService = null;
     }
 
-    // Unregister strand solicitation service
-    if (this.strandSolicitationService && this.controlNode) {
-      await this.strandSolicitationService.unregisterResponder(this.controlNode);
-      this.strandSolicitationService = null;
+    // Unregister the formation responder once any queued swap has run, so no swap
+    // registers a handler after this. A restarted node installs a fresh one.
+    await this.solicitationSwaps;
+    if (this.registeredSolicitation && this.controlNode) {
+      await this.registeredSolicitation.unregisterResponder(this.controlNode);
     }
+    this.registeredSolicitation = null;
+    this.strandSolicitationService = null;
 
     // Stop strand watcher
     if (this.strandWatcher) {
@@ -6258,9 +6333,9 @@ export class CadreNode implements SAppIdLookup {
    * caller runs the drive regardless).
    *
    * NOTE: against a relay that is DOWN this hook costs up to two strand-addr
-   * timeouts (dial by peer id, then by addr; 23 s each at the default declared link round trip)
+   * timeouts (dial by peer id, then by addr; 28.5 s each at the default declared link round trip)
    * before the reservation drive even starts, so one failed re-drive holds the supervisor
-   * `driving` for those 46 s plus the drive's own deadline — 64 s at the default, and longer on
+   * `driving` for those 57 s plus the drive's own deadline — 75 s at the default, and longer on
    * a host that declared a slower one (`link-budget.ts`).
    * Bounded and harmless while the relay is unreachable anyway; if recovery
    * latency after a relay comes back ever matters, skip the announce when the
@@ -7253,6 +7328,9 @@ export class CadreNode implements SAppIdLookup {
       ...(this.trustedOwnerStore ? { trustedOwners: this.trustedOwnerStore } : {}),
     }), this.controlNode, this.controlDatabase);
     log('Seed bootstrap service initialized');
+    // Embedders wire the owner key after start(), so the runner's first pass usually found
+    // this machine not yet an owner; without this it would wait a full poll interval.
+    this.pendingJoinRunner?.kick();
   }
 
   /**
@@ -7973,20 +8051,33 @@ export class CadreNode implements SAppIdLookup {
   // ============================================================================
 
   /**
-   * Initialize the strand solicitation service.
-   * This enables forming strands with other parties via open invitations.
+   * Replace this node's strand solicitation service — the formation responder every node
+   * installs at {@link start}, and the initiator side of {@link formStrand}. Optional: call it
+   * only to customize the responder (an approver, a provisioner, formation deadlines). The
+   * new service's handler replaces the previous one's on the control node, and invitations
+   * the previous service minted stay outstanding.
+   *
+   * Without `options.formationUsageRecorder`, tokens are checked against this party's
+   * replicated `FormationInvite`/`FormationUsage` rows ({@link ControlFormationUsageRecorder}).
+   * A responder that accepts every token is reachable only by constructing
+   * {@link StrandSolicitationService} directly.
    *
    * @param options - Configuration for the solicitation service
    */
   async initializeStrandSolicitation(options?: StrandSolicitationServiceOptions): Promise<void> {
-    if (!this.controlNode) {
+    const controlNode = this.controlNode;
+    const controlDatabase = this.controlDatabase;
+    if (!controlNode || !controlDatabase) {
       throw new Error('CadreNode must be started before initializing strand solicitation');
     }
 
     const service = new StrandSolicitationService({
       ...options,
+      formationUsageRecorder: options?.formationUsageRecorder ?? new ControlFormationUsageRecorder(controlDatabase),
       partyId: this.config.controlNetwork.partyId,
-      cadrePeerAddrs: this.getMultiaddrs(),
+      // Read per formation: the responder is installed during start, before a relay
+      // reservation gives a relay-only node any address, and addresses change after.
+      cadrePeerAddrs: () => this.getMultiaddrs(),
       // Every formation deadline derives from this node's declared link, like every other
       // dial budget here; a caller's own declaration on the formation config still wins.
       formationConfig: {
@@ -8004,24 +8095,61 @@ export class CadreNode implements SAppIdLookup {
         this.issueStrandMembershipInvite(strandId, signal)
     });
 
-    // Set before the registration is awaited, so a concurrent createOpenInvitation /
-    // formStrand uses this service instead of building a second one whose handler
-    // would collide; a failed registration puts the previous service back.
-    // NOTE: same stray-handler caveat as installSeedBootstrapService — a failed
-    // peer-store merge leaves the formation handler registered, so a retry is a duplicate.
     const previous = this.strandSolicitationService;
+    if (previous) {
+      service.adoptMintedInvitations(previous);
+    }
+    // Set before the swap is awaited, so a concurrent createOpenInvitation / formStrand
+    // uses this service instead of building another.
     this.strandSolicitationService = service;
+    const swap = this.solicitationSwaps.then(() => this.swapFormationResponder(service, controlNode));
+    // The queue only orders swaps; a failed one rejects this call's own await below.
+    this.solicitationSwaps = swap.catch(() => undefined);
+    await swap;
+    log('Strand solicitation service initialized');
+  }
+
+  /**
+   * Answer formation from start. Log-and-continue: a node that cannot is still useful, and
+   * {@link createOpenInvitation} / {@link formStrand} install the responder when it is missing.
+   */
+  private async installDefaultFormationResponder(): Promise<void> {
+    try {
+      await this.initializeStrandSolicitation();
+    } catch (error) {
+      log('Formation responder not installed at start; createOpenInvitation/formStrand will retry: %o', error);
+    }
+  }
+
+  /**
+   * Move the control node's formation handler from the registered service to `service`.
+   * Queued by {@link initializeStrandSolicitation} so swaps run one at a time: libp2p allows
+   * one handler per protocol id, and two overlapping swaps would each find the other's
+   * handler in the way. A failed swap registers the previous service again, points
+   * {@link strandSolicitationService} back at it unless a later call has replaced it, and
+   * rethrows.
+   *
+   * NOTE: same stray-handler caveat as {@link installSeedBootstrapService} — a failed
+   * peer-store merge leaves the formation handler registered, so the restore below (and a
+   * retry) rejects as a duplicate.
+   */
+  private async swapFormationResponder(service: StrandSolicitationService, controlNode: Libp2p): Promise<void> {
+    const previous = this.registeredSolicitation;
     try {
-      await service.registerResponder(this.controlNode);
+      await previous?.unregisterResponder(controlNode);
+      await service.registerResponder(controlNode);
+      this.registeredSolicitation = service;
     } catch (error) {
       if (this.strandSolicitationService === service) this.strandSolicitationService = previous;
+      await previous?.registerResponder(controlNode).catch((restoreError: unknown) =>
+        log('Re-registering the previous formation responder failed; none answers until the next initializeStrandSolicitation: %o', restoreError));
       throw error;
     }
-    log('Strand solicitation service initialized');
   }
 
   /**
-   * Get the strand solicitation service (for advanced use)
+   * Get the strand solicitation service (for advanced use). Non-null on a started node unless
+   * the install at {@link start} failed.
    */
   getStrandSolicitationService(): StrandSolicitationService | null {
     return this.strandSolicitationService;
@@ -8030,20 +8158,25 @@ export class CadreNode implements SAppIdLookup {
   /**
    * Create an open invitation for others to form strands with this party.
    *
+   * The invitation's bootstrap list names this machine first (it is the one most likely to
+   * run the host strand), then a few of the party's other machines
+   * ({@link siblingInvitationAddrs}), so a joiner can still form while this one is offline.
+   *
    * @param sAppId - The sApp to use for formed strands
    * @param expirationMs - How long the invitation is valid (ms from now)
    * @returns The open invitation to share out-of-band
+   * @throws when neither this machine nor any other machine of the party has an address
    */
   async createOpenInvitation(
     sAppId: string,
     expirationMs: number = 24 * 60 * 60 * 1000 // 24 hours default
   ): Promise<OpenInvitation> {
     if (!this.strandSolicitationService) {
-      // Create a temporary service for creating invitations
+      // Only when the install at start failed.
       await this.initializeStrandSolicitation();
     }
 
-    const bootstrap = this.getMultiaddrs();
+    const bootstrap = [...this.getMultiaddrs(), ...await this.siblingInvitationAddrs()];
     if (bootstrap.length === 0) {
       throw new Error('No multiaddrs available for invitation');
     }
@@ -8056,11 +8189,44 @@ export class CadreNode implements SAppIdLookup {
   }
 
   /**
-   * Form a strand with a responder via an open invitation.
+   * Addresses of the party's other machines for an invitation's bootstrap list
+   * ({@link selectInvitationSiblingAddrs} picks which). The source is each authorized
+   * member's signed `CadrePeer` record ({@link resolvePeerRecord}), so a machine whose record
+   * is missing, stale or untrusted is left out; live connections only order them.
+   *
+   * Best-effort: a failed read leaves the invitation naming this machine alone, which is what
+   * it named before siblings were added.
+   */
+  private async siblingInvitationAddrs(): Promise<string[]> {
+    const controlNode = this.controlNode;
+    if (!controlNode || !this.controlDatabase) {
+      return [];
+    }
+    try {
+      const siblings: InvitationSibling[] = [];
+      for (const { peerId } of await this.listAuthorizedMembers()) {
+        const resolved = await this.resolvePeerRecord(peerId, {});
+        if (resolved && resolved.addrs.length > 0) {
+          siblings.push({ peerId, updatedAt: resolved.updatedAt, addrs: resolved.addrs.map(String) });
+        }
+      }
+      const connected = new Set(controlNode.getConnections().map((c) => c.remotePeer.toString()));
+      return selectInvitationSiblingAddrs(siblings, connected);
+    } catch (error) {
+      log('createOpenInvitation: resolving the other machines failed; the invitation names only this one: %o', error);
+      return [];
+    }
+  }
+
+  /**
+   * Form a strand with a responder via an open invitation: one attempt, from this machine, with
+   * nothing recorded party-wide. {@link requestJoin} is the path that keeps trying.
    *
    * @param invitation - The open invitation received out-of-band
    * @param disclosure - Identity/context information to share with the responder
    * @returns The member key and strand info if successful
+   * @throws `FormationRejectedError` or `FormationUnreachableError` when the formation fails, and
+   *   `FormationPostApprovalError` when it was approved and a step on this machine then failed
    */
   async formStrand(
     invitation: OpenInvitation,
@@ -8091,6 +8257,190 @@ export class CadreNode implements SAppIdLookup {
     return result;
   }
 
+  /**
+   * Ask to join through `invitation` and have the party keep trying until the join works, the
+   * invitation is used up or expires, or the request is dismissed: across restarts, and from
+   * any owner machine of the party (`docs/strands.md` → "Joining while the inviter is
+   * offline"). Records a party-wide `PendingJoin` row, runs one attempt on this machine at
+   * once, and returns the status after it: `'joined'` when the inviter answered yes, `'waiting'`
+   * when it was unreachable or not ready, `'failed'` when it refused for good.
+   *
+   * Asking again for an invitation whose request is still pending adopts that request as it
+   * is; asking again after it finished starts a fresh one. Later changes arrive as
+   * `pendingJoin:changed`. A joined strand is offered through `strand:discovered` like any
+   * {@link formStrand} join. Do not also call {@link formStrand} with the same invitation:
+   * nothing coordinates the two, and one of them gets `token-spent`.
+   *
+   * @throws on a machine that is not an enrolled owner (the row is owner-signed), and on an
+   *   invitation already past its expiration
+   */
+  async requestJoin(invitation: OpenInvitation, disclosure: StrandFormationDisclosure = {}): Promise<PendingJoinStatus> {
+    const { database, runner } = this.requirePendingJoins();
+    const signer = await this.requireOwnerSigner(
+      'requestJoin needs an enrolled owner machine of this party, because the party-wide join request is ' +
+      'owner-signed. formStrand(invitation, disclosure) joins once from this machine, with no retries.'
+    );
+    const now = Date.now();
+    const expiresAt = invitation.expiration.getTime();
+    if (!Number.isFinite(expiresAt)) {
+      throw new Error('The invitation carries no valid expiration');
+    }
+    if (now >= expiresAt) {
+      throw new Error(`The invitation expired at ${invitation.expiration.toISOString()}; ask the inviter for a fresh one`);
+    }
+    const requested = requestedPendingJoin(
+      pendingJoinId(invitation.token), this.encodeInvitation(invitation), disclosure, expiresAt, now);
+    const row = await this.recordJoinRequest(database, requested, signer);
+    log('requestJoin: pending join %s recorded; attempting now', row.Id);
+    return runner.attemptNow(row);
+  }
+
+  /**
+   * Every join asked for with {@link requestJoin} and not dismissed, pending or finished, as this
+   * machine sees it (see `PendingJoinStatus` for which states are machine-local).
+   */
+  async listPendingJoins(): Promise<PendingJoinStatus[]> {
+    const { database, runner } = this.requirePendingJoins();
+    return (await database.queryPendingJoins()).map((row) => runner.statusOf(row));
+  }
+
+  /**
+   * Remove a join request party-wide. On a pending request this cancels it: every owner machine
+   * stops at its next read, and an attempt already running does not record its outcome (a join
+   * it made stays on that machine, as any {@link formStrand} join does). Owner machines only.
+   * Resolves `false` when there was no such request.
+   */
+  async dismissPendingJoin(id: string): Promise<boolean> {
+    const { database, runner } = this.requirePendingJoins();
+    const { ownerKey, signMessage } = await this.requireOwnerSigner(
+      'dismissPendingJoin needs an enrolled owner machine of this party: the removal is owner-signed');
+    const removed = await database.deletePendingJoin(id, ownerKey, signMessage);
+    runner.forgetRow(id);
+    return removed;
+  }
+
+  private requirePendingJoins(): { database: ControlDatabase; runner: PendingJoinRunner } {
+    if (!this._running || !this.controlDatabase || !this.pendingJoinRunner) {
+      throw new Error('CadreNode must be started before using pending joins');
+    }
+    return { database: this.controlDatabase, runner: this.pendingJoinRunner };
+  }
+
+  /** This machine's owner key and signer for an owner-signed control write, or a throw carrying `refusal`. */
+  private async requireOwnerSigner(refusal: string): Promise<{ ownerKey: string; signMessage: (message: Uint8Array) => string }> {
+    const signingKey = await this.enrolledOwnerSigningKey();
+    if (!signingKey) {
+      throw new Error(refusal);
+    }
+    return { ownerKey: signingKey.publicKeyB64, signMessage: signMessageWith(signingKey.privateKeyB64) };
+  }
+
+  /**
+   * Write a join request's row, or adopt the party's existing row for the same invitation: a
+   * pending one as it is, a finished one replaced by `requested` (the user asked again).
+   */
+  private async recordJoinRequest(
+    database: ControlDatabase,
+    requested: PendingJoinFields,
+    { ownerKey, signMessage }: { ownerKey: string; signMessage: (message: Uint8Array) => string }
+  ): Promise<PendingJoinRow> {
+    let written: PendingJoinRow;
+    try {
+      written = await database.insertPendingJoin(requested, ownerKey, signMessage);
+    } catch (error) {
+      if (!isPendingJoinConflict(error)) {
+        throw error;
+      }
+      const existing = await database.queryPendingJoin(requested.Id);
+      if (!existing) {
+        throw new Error(
+          `A removed join request for this invitation (${requested.Id}) is still held on this machine until its ` +
+          'clean-up runs, which needs a connection to the party. Ask again once connected.',
+          { cause: error }
+        );
+      }
+      if (existing.Outcome === null) {
+        return existing;
+      }
+      written = await database.replacePendingJoin(existing.StampId, requested, ownerKey, signMessage);
+    }
+    this.pendingJoinRunner?.noteWritten(written);
+    return written;
+  }
+
+  /** Build and start the pending-join retry loop over this node (see {@link pendingJoinRunner}). */
+  private startPendingJoinRunner(): void {
+    const database = (): ControlDatabase => {
+      if (!this.controlDatabase) {
+        throw new Error('CadreNode is stopped: no control database for pending joins');
+      }
+      return this.controlDatabase;
+    };
+    const outcomeSigner = (): Promise<{ ownerKey: string; signMessage: (message: Uint8Array) => string }> =>
+      this.requireOwnerSigner('Cannot record a pending join outcome: this machine is not an enrolled owner of the party');
+    this.pendingJoinRunner = new PendingJoinRunner({
+      selfId: this.controlNode!.peerId.toString(),
+      linkRoundTripMs: this.config.network?.linkRoundTripMs,
+      isOwner: async () => (await this.enrolledOwnerSigningKey()) !== null,
+      // NOTE: one owner-key read and one `PendingJoin` read per pass (30 s); for a party that never
+      // asked for a join the second reads a never-written block, which consults the cohort. If it
+      // shows up in a device profile, poll more slowly while the table reads empty.
+      readRows: () => database().queryPendingJoins(),
+      readRow: (id) => database().queryPendingJoin(id),
+      attempt: (row) => this.formStrand(this.pendingJoinInvitation(row), parseStoredDisclosure(row)),
+      replace: async (expectedStampId, next) => {
+        const { ownerKey, signMessage } = await outcomeSigner();
+        return database().replacePendingJoin(expectedStampId, next, ownerKey, signMessage);
+      },
+      remove: async (id) => {
+        const { ownerKey, signMessage } = await outcomeSigner();
+        return database().deletePendingJoin(id, ownerKey, signMessage);
+      },
+      isAlone: () => this.committedAlone(),
+      sAppIdOf: (row) => this.pendingJoinSAppId(row),
+      observeRows: (rows) => this.stageMembershipInvitesFromPendingJoins(rows),
+      emit: (status) => this.emit('pendingJoin:changed', status),
+    });
+    this.pendingJoinRunner.start();
+  }
+
+  /** The invitation a `PendingJoin` row stores. A failure names the row, never the invitation, which is a bearer credential. */
+  private pendingJoinInvitation(row: PendingJoinRow): OpenInvitation {
+    try {
+      return this.decodeInvitation(row.Invitation);
+    } catch {
+      throw new Error(`PendingJoin ${row.Id}: the stored invitation does not decode`);
+    }
+  }
+
+  private pendingJoinSAppId(row: PendingJoinRow): string {
+    try {
+      return this.decodeInvitation(row.Invitation).sAppId;
+    } catch {
+      log('PendingJoin %s: the stored invitation does not decode; reporting no sApp', row.Id);
+      return '';
+    }
+  }
+
+  /**
+   * Stage the membership invitation of each `joined` row this process has not staged yet, so a
+   * closed strand joined on another owner machine is seated by whichever machine launches it
+   * first; the party key it admits is the replicated `StrandPartyKey` row the finishing machine
+   * seated. A strand that already has an invitation staged keeps it, for the reconciler to
+   * settle first.
+   */
+  private stageMembershipInvitesFromPendingJoins(rows: readonly PendingJoinRow[]): void {
+    for (const { rowId, strandId, invite } of membershipInvitesToStage(rows, this.stagedMembershipInviteKeys, Date.now())) {
+      this.stagedMembershipInviteKeys.add(invite.inviteKey);
+      if (this.pendingMembershipInvites.has(strandId)) {
+        continue;
+      }
+      this.pendingMembershipInvites.set(strandId, invite);
+      log('pending join %s: staged its membership invitation for strand %s', rowId, strandId);
+      this.strandManager.notifyMembershipInviteStaged(strandId);
+    }
+  }
+
   /**
    * Remember the strand a formation just joined (see {@link joinedStrandStore}), so it
    * is re-offered after a restart even when the app is killed before its `addStrand`, and
@@ -8111,7 +8461,8 @@ export class CadreNode implements SAppIdLookup {
         joinedAt: Date.now()
       });
     } catch (error) {
-      throw new Error(
+      throw new FormationPostApprovalError(
+        result.strandId,
         `Formation for strand ${result.strandId} was approved (its one-time token is spent), but ` +
         'remembering the join in this node\'s joined-strand store failed, so the strand would not ' +
         'come back after a restart. Fix the store (usually the configured keyStore), then redeem a ' +
@@ -8147,7 +8498,8 @@ export class CadreNode implements SAppIdLookup {
     try {
       await this.ensureStrandPartyKey(strandId);
     } catch (error) {
-      throw new Error(
+      throw new FormationPostApprovalError(
+        strandId,
         `Formation for strand ${strandId} was approved (its one-time token is spent), but ` +
         'persisting this party\'s membership identity (StrandPartyKey) failed — the joiner ' +
         'cannot become a member without it. Fix the underlying cause, then redeem a fresh ' +
@@ -8156,6 +8508,7 @@ export class CadreNode implements SAppIdLookup {
       );
     }
     this.pendingMembershipInvites.set(strandId, invite);
+    this.stagedMembershipInviteKeys.add(invite.inviteKey);
     log('formStrand: staged membership invitation for strand %s (party key persisted)', strandId);
     this.strandManager.notifyMembershipInviteStaged(strandId);
   }