@serfab/cadre-core 1.17.0 → 1.19.0

package/dist/cadre-node.js

@@ -21,7 +21,7 @@ import { MemoryBootstrapPeerStore } from './bootstrap-peer-store.js';
 import { MemoryStrandNetworkStateStore } from './strand-network-state.js';
 import { MemoryEnrolledMachineStore } from './enrolled-machine-store.js';
 import { JoinedStrandSession, KeyStoreJoinedStrandStore, MemoryJoinedStrandStore } from './joined-strand-store.js';
-import { mergePeerAddrs, groupAddrsByPeerId } from './peer-addr-book.js';
+import { addrKey, mergePeerAddrs, groupAddrsByPeerId } from './peer-addr-book.js';
 import { strandFretPeerAddrs } from './strand-fret-addrs.js';
 import { verifyCadrePeerVoucher, verifyInvitationAdmission, verifyRevocationSigner } from './peer-authorization.js';
 import { deriveOwnerAnchor } from './owner-anchor-sync.js';
@@ -30,32 +30,35 @@ import { describeOwnerKeys, namedOwnerKey, planOwnerKeyRemoval, revouchWithinPla
 import { ed25519PublicKeyB64FromPeerId } from './ed25519-key.js';
 import { signPeerRecord, verifyPeerRecordSignature, isPeerRecordFresh, orderSignalingFirst, isSignalingAddr, currentMemberTrustPolicy, DEFAULT_PEER_RECORD_MAX_AGE_MS, DEFAULT_PEER_RECORD_HEARTBEAT_MS } from './peer-record.js';
 import { signDeviceTokenRecord, verifyDeviceTokenSignature, isPushPlatform } from './device-token.js';
-import { StrandWatcher } from './strand-watcher.js';
+import { StrandWatcher, DEFAULT_STRAND_WATCH_BACKSTOP_MS } from './strand-watcher.js';
 import { StrandInstanceManager, liveStrandStatus } from './strand-instance-manager.js';
 import { PeerJoinBackfill } from './peer-join-backfill.js';
 import { deriveCohortMembers } from './strand-cohort.js';
-import { selectControlCohortDials, DEFAULT_CONTROL_COHORT_RECONCILE_MS, DEFAULT_CONTROL_COHORT_TARGET_DEGREE } from './control-cohort.js';
+import { selectControlCohortDials, nextReconcileWakeMs, CLAIM_RESTART_REDIAL_STEP_MS, CONTROL_COHORT_REDIAL_MS, DEFAULT_CONTROL_COHORT_RECONCILE_MS, DEFAULT_CONTROL_COHORT_TARGET_DEGREE } from './control-cohort.js';
+import { scheduleRuns } from './table-change-watch.js';
+import { MEMBERSHIP_VIEW_RECONNECT_RELOAD_MS } from './membership-view.js';
 import { dialPeerAddrs, SelfRelayOnlyError, CONTROL_COHORT_DIAL_ADDRESS_ATTEMPTS, trailingPeerId, withTrailingPeerId } from './peer-dial.js';
 import { ADMISSION_DECISION_TIMEOUT_MS, declaredCohortReadDeadlineMs, optimysticDialLimits, peerJoinPushBudget, relayReservationBudgetMs, relayedDialBudgetMs, relayedRequestBudgetMs, resolveLinkRoundTripMs } from './link-budget.js';
 import { EnrollmentService } from './enrollment.js';
 import { HibernationManager } from './hibernation-manager.js';
 import { ControlDatabase, generateStampId, isPendingJoinConflict, isStrandIdConflict, IssuedRowNotHereYetError, pendingJoinId } from './control-database.js';
-import { CadreInviteHandler, encodeCadreInvitation, redeemInPasses, signRedeemRequest } from './cadre-invite-protocol.js';
+import { CadreInviteHandler, encodeCadreInvitation, redeemInPasses, sanitizeAddrs, signRedeemRequest } from './cadre-invite-protocol.js';
 import { OwnerGrantHandler, OwnerGrantRejectedError, encodeOwnerGrant, redeemOwnerGrantInPasses, signOwnerGrantRequest } from './owner-grant-protocol.js';
 import { FormationPostApprovalError } from './strand-formation-rejection.js';
 import { PendingJoinRunner, membershipInvitesToStage, parseStoredDisclosure, requestedPendingJoin } from './pending-join-runner.js';
-import { SeedBootstrapService, mergeSeedPeers, seedDigest } from './seed-bootstrap.js';
+import { SeedBootstrapService, mergeSeedPeers, seedDialTargets, seedDigest } from './seed-bootstrap.js';
 import { claimSecretTrustPolicy } from './seed-trust-policy.js';
 import { ClaimRefusedError, claimProof, parseClaimSecret } from './claim-proof.js';
 import { StrandSolicitationService } from './strand-solicitation.js';
 import { ControlFormationUsageRecorder } from './control-formation-recorder.js';
 import { cadreInvitationReach, selectCadreInvitationMembers, selectInvitationSiblingAddrs } from './invitation-bootstrap.js';
 import { createMembershipConnectionGater, UnauthorizedReservationBudget } from './membership-connection-gater.js';
-import { controlNetworkName, installControlProtocolGuard } from './control-protocol-guard.js';
+import { controlNetworkName, installControlProtocolGuard, servesControlDatabase } from './control-protocol-guard.js';
 import { StrandWakeService, dialWake } from './strand-wake-protocol.js';
 import { StrandAddrService, collectStrandAddrs } from './strand-addr-protocol.js';
 import { DelegateAdmissionStore, extractCircuitRelayTargets, dueRelayAnnounces, prunePeerStrandKeys, peerStrandKey, recordRelayRefusals, relaysNotRefusing } from './delegate-admission.js';
 import { relayCircuitAddrs, resolveListenAddrs, resolveTransportOptions, RelayReservationFailedError } from './relay-addrs.js';
+import { selfAddressesChanged } from './self-address-update.js';
 import { nodeTransportsOption } from './node-transports.js';
 import { resolveRelayServer } from './relay-server.js';
 import { replacesAdvertisedAddrs, resolveAnnounceAddrs } from './announce-addrs.js';
@@ -73,8 +76,8 @@ const timing = debug('sereus:cadre:timing');
  *
  * Ten minutes sits comfortably inside the peerStore's one-hour address expiry
  * (`MAX_ADDRESS_AGE`, see `peer-addr-book.ts`) with headroom for several missed
- * passes, and far above the 15 s reconcile cadence the refresh rides on, so the
- * strand-addr RPC fan-out stays cheap.
+ * passes, and far above the reconcile pass's 2-minute backstop the refresh rides
+ * on, so the strand-addr RPC fan-out stays cheap.
  */
 export const STRAND_PEER_ADDR_REFRESH_MS = 10 * 60 * 1000;
 /**
@@ -83,8 +86,9 @@ export const STRAND_PEER_ADDR_REFRESH_MS = 10 * 60 * 1000;
  * phone asking before its `CadrePeer` row has replicated to the sibling is refused
  * and its delegate grant goes unrecorded, and waiting the full
  * {@link STRAND_PEER_ADDR_REFRESH_MS} for the next try is the delay this bounds.
- * Four reconcile ticks, so a sibling that keeps failing costs one timed-out RPC a
- * minute rather than one a tick.
+ * A pass arms its own wake for the earliest due time it leaves behind
+ * (`nextReconcileWakeMs`), so a sibling that keeps failing costs one timed-out
+ * RPC a minute, with no pass in between.
  */
 export const STRAND_PEER_ADDR_RETRY_MS = 60 * 1000;
 /**
@@ -451,8 +455,8 @@ export class CadreNode {
          *
          * Cleared by {@link cleanup} only, deliberately NOT by {@link detachStrand}: the watcher's
          * filter rejection stops a replica through that very method, and forgetting the id there
-         * would have the next poll find it unknown again, re-admit the strand provisionally and
-         * relaunch the replica — once every other poll, forever. A strand id's sApp never changes,
+         * would have the next watcher run find it unknown again, re-admit the strand provisionally
+         * and relaunch the replica — on every other run, forever. A strand id's sApp never changes,
          * and the map is bounded by the strands this party has published, like
          * {@link discoveredStrands}.
          */
@@ -480,19 +484,24 @@ export class CadreNode {
          * wait that ends in mutual denial (the upstream gate is fail-closed on
          * timeout, unlike the fail-open connection gater). Optimystic's predicate is
          * synchronous by contract. So the set is refreshed OUT OF BAND instead — after
-         * start, on every {@link reconcileControlCohort} pass (15s cadence), and
-         * immediately after every LOCAL membership mutation so a just-vouched peer
-         * is admitted without waiting for the timer. That last refresh is AUTOMATIC:
-         * the control database notifies {@link refreshMembershipGate} after every
+         * start, at the head of every {@link reconcileControlCohort} pass, whenever a
+         * membership table changes (the gate watch {@link start} takes through
+         * `ControlDatabase.watchMembershipTables`, which hears a local commit and a change
+         * replicated into this node's storage alike, within a 50 ms window), and immediately
+         * after every LOCAL membership mutation so a just-vouched peer is admitted
+         * before the writer's call returns. That last refresh is AUTOMATIC: the
+         * control database notifies {@link refreshMembershipGate} after every
          * committed `CadrePeer` write (see `ControlDatabase.mutateCadrePeer`), so no
-         * writer has to remember. A change that arrives by REPLICATION is picked up
-         * on the next timed refresh — bounded staleness, acceptable because this gate
-         * is defense in depth: rows an unadmitted peer manages to write are still
-         * disbelieved at read time by the voucher-anchored predicate. For cohort
-         * choice the same window means a member whose row arrived by replication is
-         * left out of this node's control-block cohorts until that refresh.
-         * Optimystic reads the predicate live on every lookup, so replacing this set
-         * is the whole re-evaluation; there is no per-peer reclassify call to make.
+         * writer has to remember. A change this node's storage never hears of — a
+         * machine outside the block's cohort on a party larger than the replication
+         * breadth — is picked up at the pass's 2-minute backstop: bounded staleness,
+         * acceptable because this gate is defense in depth: rows an unadmitted peer
+         * manages to write are still disbelieved at read time by the voucher-anchored
+         * predicate. For cohort choice the same window means a member whose row
+         * arrived by replication is left out of this node's control-block cohorts
+         * until that refresh. Optimystic reads the predicate live on every lookup, so
+         * replacing this set is the whole re-evaluation; there is no per-peer
+         * reclassify call to make.
          */
         this.authorizedControlPeers = new Set();
         /**
@@ -530,9 +539,9 @@ export class CadreNode {
          * When this node last ANNOUNCED a delegate, keyed `<targetPeerId>\n<strandId>`
          * (the peer announced TO, not the delegate). Throttles
          * {@link refreshDelegateGrants} to once per `DELEGATE_GRANT_TTL_MS / 2` per
-         * (relay, strand) so the 15 s reconcile tick never becomes per-tick RPC
-         * chatter; the launch/resume announce passes record here too. Keys whose
-         * strand is no longer running are pruned on each reconcile pass.
+         * (relay, strand) so a run of change-driven reconcile passes never becomes
+         * per-pass RPC chatter; the launch/resume announce passes record here too.
+         * Keys whose strand is no longer running are pruned on each reconcile pass.
          */
         this.delegateAnnounceAt = new Map();
         /**
@@ -559,12 +568,13 @@ export class CadreNode {
          * strand, keyed `peerStrandKey(siblingControlPeerId, strandId)` → epoch ms; a
          * missing key is due. Bounds {@link refreshStrandPeerAddrs}'s fan-out per
          * (sibling, strand) rather than per strand, so a sibling that connects after a
-         * pass is asked on the next 15 s tick instead of waiting out a stamp another
-         * sibling's answer set. An answer (even an empty one) makes the sibling due again
-         * in {@link STRAND_PEER_ADDR_REFRESH_MS}; no answer, in
-         * {@link STRAND_PEER_ADDR_RETRY_MS}. Keys whose strand stopped running or whose
-         * sibling is no longer connected are pruned every pass, so a resumed strand and a
-         * reconnected sibling are both asked at once.
+         * pass is asked on the pass its connection starts instead of waiting out a stamp
+         * another sibling's answer set. An answer (even an empty one) makes the sibling
+         * due again in {@link STRAND_PEER_ADDR_REFRESH_MS}; no answer, in
+         * {@link STRAND_PEER_ADDR_RETRY_MS}, and each pass arms its own wake for the
+         * earliest due time it leaves behind (`nextReconcileWakeMs`). Keys whose strand
+         * stopped running or whose sibling is no longer connected are pruned every pass,
+         * so a resumed strand and a reconnected sibling are both asked at once.
          */
         this.strandAddrAskDueAt = new Map();
         /**
@@ -646,6 +656,18 @@ export class CadreNode {
          * `unpublishStrand` / `forgetJoinedStrand` as {@link formationStrandAddrs} is.
          */
         this.pendingMembershipInvites = new Map();
+        /**
+         * The party key {@link adoptFormationMembershipInvite} just seated or read, by strand id,
+         * so the launch that follows a join does not read the `StrandPartyKey` row again. Taken
+         * (deleted) by the {@link resolveStrandPartyKey} that uses it, so any later launch reads
+         * the row; dropped by {@link detachStrand}. The joiner's counterpart of the founder's
+         * `StrandConfig.partyMemberPrivateKey`, without the app carrying the secret.
+         *
+         * NOTE: an entry is not dropped when the row is removed, because no `CadreNode` path removes
+         * it; an embedder calling `getControlDatabase().deleteStrandPartyKey` between a join and its
+         * launch would get the deleted key. If a node-level rotation is added, drop the entry there.
+         */
+        this.formedPartyKeys = new Map();
         /**
          * Invite keys of the membership invitations this process has staged, by
          * {@link adoptFormationMembershipInvite} or from a `JoinSuccess` row
@@ -739,7 +761,7 @@ export class CadreNode {
         this.joinedStrandStore = null;
         /**
          * This session's joined strands: {@link joinedStrandStore} and the party-wide
-         * `JoinedStrand` table, unioned with the control rows each watcher poll, published and
+         * `JoinedStrand` table, unioned with the control rows each watcher run, published and
          * removed by the reconcile pass. Per SESSION, unlike the store — rebuilt at every
          * {@link start}, which is what ends a revoked join's "keep offering until this session
          * ends" (see `JoinedStrandSession.forgetAfterThisSession`).
@@ -752,17 +774,41 @@ export class CadreNode {
         /** Listener that re-publishes the self record when reachable addresses change. */
         this.selfPeerUpdateHandler = null;
         /**
-         * Recurring proactive control-cohort dial cadence (see
-         * {@link reconcileControlCohort}). Wired alongside {@link recordRefreshTimer}
-         * in {@link startRecordRefresh} and torn down symmetrically in
-         * {@link stopRecordRefresh}; `.unref()`'d so it never keeps the loop alive.
+         * The reconcile pass's schedule (see {@link reconcileControlCohort}): single-flight
+         * runs on `table-change-watch.ts`'s rule, with no table subscription of its own — the
+         * gate watch ({@link membershipGateWatch}), the control connection listeners,
+         * `self:peer:update`, and the pass's own wake ({@link armReconcileWake}) `trigger` it,
+         * and its backstop (`network.controlCohort.reconcileMs`, default
+         * `DEFAULT_CONTROL_COHORT_RECONCILE_MS`) runs it when nothing has. Built alongside
+         * {@link recordRefreshTimer} in {@link startRecordRefresh}, about 1 s after start, and
+         * disposed in {@link stopRecordRefresh}; its timers never keep the loop alive.
+         */
+        this.controlReconcileRuns = null;
+        /**
+         * The subscription on the membership tables (`ControlDatabase.watchMembershipTables`) that
+         * re-materializes the membership gate and triggers {@link controlReconcileRuns} on every
+         * change ({@link onMembershipGateTablesChanged}).
+         * Taken in {@link start} right after the membership listener is wired, disposed in
+         * {@link cleanup} before the control database closes.
+         */
+        this.membershipGateWatch = null;
+        /** The second membership-view drop after a sibling connects ({@link armMembershipViewReload}); cleared in {@link stopRecordRefresh}. */
+        this.membershipViewReloadTimer = null;
+        /**
+         * Consecutive passes whose dial of a selected sibling or a cold-start bootstrap peer
+         * failed (`ControlCohortPassOutcome.unconnected`), for the redial ladder
+         * ({@link armReconcileWake}); 0 after a pass none of whose dials did. Per session:
+         * reset in {@link stopRecordRefresh}.
          */
-        this.controlCohortReconcileTimer = null;
+        this.reconcileRedialStreak = 0;
+        /** Peers a {@link redialClaimedNode} loop is running for, so a repeated claim starts no second one. */
+        this.claimRedialPeers = new Set();
         /**
-         * Single-flight guard for {@link reconcileControlCohort}. The eager start pass,
-         * the recurring interval, and the `self:peer:update` trigger can fire close
-         * together; collapsing concurrent passes into one in-flight run prevents two
-         * passes from double-dialing the same siblings (mirrors {@link registerSelfInFlight}).
+         * Single-flight guard for {@link reconcileControlCohort}. An explicit call, the
+         * `self:peer:update` trigger and a scheduled run can fire close together; collapsing
+         * concurrent passes into one in-flight run prevents two passes from double-dialing the
+         * same siblings (mirrors {@link registerSelfInFlight}). A scheduled run waits for this
+         * pass and then starts its own rather than joining it.
          */
         this.reconcileControlCohortInFlight = null;
         /**
@@ -785,19 +831,27 @@ export class CadreNode {
          * machine in the party is already discarding this node's address as stale.
          */
         this.lastSelfRecordPublishAt = null;
+        /**
+         * The newest record this process signed for its own `CadrePeer` row: written by either path
+         * of {@link publishSelfRecord}, or handed to a member to write by {@link redeemCadreInvitation}.
+         * What {@link publishSelfRecordUnlessCurrent} compares the row against. Per session, cleared
+         * with {@link lastSelfRecordPublishAt}: a row carried over from an earlier process is never
+         * taken as current.
+         */
+        this.lastSignedSelfRecord = null;
         /** Say-once latch for {@link escalateIfSelfRecordStale}; re-armed by the next publish. */
         this.selfRecordStaleWarned = false;
         // ── Write-while-alone re-replication (control-write-ensure-replicated) ──────
         /**
          * Re-replication queue for owner `authorizePeer` writes that committed while
-         * this node was alone — `controlNode.getConnections().length === 0` at write
-         * time means the Optimystic commit was local-only (the block's cluster was ≤1)
-         * and never broadcast. Maps the affected subject peerId → 'authorize', to
+         * this node was alone ({@link committedAlone}) — the Optimystic commit was
+         * local-only (the block's cohort was this node) and never broadcast. Maps the
+         * affected subject peerId → 'authorize', to
          * re-issue once the cohort grows. Removals are NOT tracked here — a delete
          * leaves a `Revocation` tombstone, tracked by {@link pendingRevocations}; a
          * removePeer also clears any queued authorize for the same subject (the row is
-         * gone, so re-issuing the insert would be wrong). Drained on the 0→≥1
-         * control-connection transition by {@link drainPendingControlReplication}.
+         * gone, so re-issuing the insert would be wrong). Drained when a sibling becomes
+         * reachable ({@link updateControlSiblingReachability}) by {@link drainPendingControlReplication}.
          */
         this.pendingPeerWrites = new Map();
         /**
@@ -821,14 +875,6 @@ export class CadreNode {
          * exactly a "removals from before this lifetime" case).
          */
         this.reissuedHeldRevocations = false;
-        /**
-         * Whether this process has seen the singleton `Revocation` ledger marker filed — its own
-         * `'opened'` or an `'already-open'` answer, in {@link openRevocationLedgerIfDue}. A
-         * record of work done, not a cached authorization answer: the marker can never be
-         * deleted (`NoDelete`), so this can only go stale if the database itself is replaced,
-         * which is why stop() clears it with the other per-process replication state.
-         */
-        this.revocationLedgerOpened = false;
         /** This node's own `CadrePeer` self-write committed local-only (re-touched on growth). */
         this.pendingSelfPeerWrite = false;
         /** This node's own `DeviceToken` self-write committed local-only (re-touched on growth). */
@@ -843,14 +889,24 @@ export class CadreNode {
         /** Single-flight guard for the re-replication drain (mirrors {@link registerSelfInFlight}). */
         this.drainControlReplicationInFlight = null;
         /**
-         * Tracks the control-connection presence edge so the drain fires only on the
-         * 0→≥1 transition (the earliest point a re-issue can broadcast), not on every
-         * subsequent `connection:open`. Re-armed to false once connections return to 0.
+         * Connected control peers whose identify says they serve the control database
+         * ({@link onControlPeerIdentified}): the peers Optimystic can count into a control block's
+         * cohort, before its membership check. A peer joins on its first `peer:identify` of a
+         * connected period, not on `connection:open`: until it identifies, every control block's
+         * cohort is this node alone. It leaves when its last connection closes. Read through
+         * {@link hasReachableControlSibling}, which applies the membership check.
          */
-        this.hasControlConnection = false;
-        /** `connection:open` listener driving the growth-triggered drain (teardown in {@link stopRecordRefresh}). */
-        this.controlConnectionOpenHandler = null;
-        /** `connection:close` listener re-arming the growth edge (teardown in {@link stopRecordRefresh}). */
+        this.identifiedControlPeers = new Set();
+        /**
+         * {@link hasReachableControlSibling} at its last evaluation, so the drain fires on the
+         * false→true edge only (the earliest point a re-issue reaches the cohort), not on every later
+         * identify. Re-armed to false when no reachable sibling remains
+         * ({@link updateControlSiblingReachability}).
+         */
+        this.controlSiblingReachable = false;
+        /** `peer:identify` listener: {@link onControlPeerIdentified} (teardown in {@link stopRecordRefresh}). */
+        this.controlPeerIdentifyHandler = null;
+        /** `connection:close` listener: {@link onControlConnectionClosed} (teardown in {@link stopRecordRefresh}). */
         this.controlConnectionCloseHandler = null;
         this.config = config;
         this.hostUnclaimedStrands = config.hostUnclaimedStrands ?? config.profile === 'storage';
@@ -1013,7 +1069,7 @@ export class CadreNode {
             // closed before any network bring-up, and the retained dial targets are
             // loaded before the first reconcile pass could consult them.
             this.initializeBootstrapPeerStore();
-            // The remembered cross-party joins, which the strand watcher built below polls
+            // The remembered cross-party joins, which the strand watcher built below reads
             // beside the control rows. Before network bring-up, so a mis-scoped injected store
             // fails closed like the two above.
             this.initializeJoinedStrandStore();
@@ -1092,6 +1148,20 @@ export class CadreNode {
             // rest of start() and the refresh early-returns while it is — a write in that
             // window is covered by start's own pass below.
             this.controlDatabase.setMembershipChangeListener((reason) => this.refreshMembershipGate(reason));
+            // The gate watch: a change to any membership table, replicated in or committed here,
+            // re-materializes the gate and starts a reconcile pass. Separate from the pass so a
+            // replicated revocation does not wait behind a pass spending its dial budget on an
+            // unreachable sibling. The watch handler only arms a timer, so nothing runs under the
+            // write lock the membership listener is called with; `onChange` runs from that timer.
+            //
+            // NOTE: a table this `Database` has not yet touched hears local commits only
+            // (`table-change-watch.ts`). The start refresh reads `CadrePeer`, `OwnerKey` and
+            // `Revocation`; the four invitation and grant tables are first read when some row is
+            // admitted through one, and a load the committed-read path served is not kept and
+            // arms no bridge, so a chain row that replicates in before then waits for the next
+            // change on a touched table or the backstop. Nothing depends on a chain row until a
+            // `CadrePeer` or `OwnerKey` row refers to it, and that row is itself such a change.
+            this.membershipGateWatch = this.controlDatabase.watchMembershipTables(() => this.onMembershipGateTablesChanged(), { label: 'membership-gate' });
             // Committed-delete seam → write-while-alone revocation queue. Same wiring
             // window argument as the membership listener above.
             this.controlDatabase.setGuardedDeleteListener((revocation) => this.noteGuardedDelete(revocation));
@@ -1104,7 +1174,7 @@ export class CadreNode {
             this.strandWatcher = new StrandWatcher(queryable, {
                 onStrandAdded: async (strand) => this.handleStrandAdded(strand),
                 onStrandRemoved: async (strandId) => this.handleStrandRemoved(strandId)
-            }, this.config.strandFilter ?? { mode: 'all' }, this.config.strandWatchInterval ?? 5000, this // CadreNode implements SAppIdLookup
+            }, this.config.strandFilter ?? { mode: 'all' }, this.config.strandWatchInterval ?? DEFAULT_STRAND_WATCH_BACKSTOP_MS, this // CadreNode implements SAppIdLookup
             );
             t0 = performance.now();
             await this.strandWatcher.start();
@@ -1174,8 +1244,8 @@ export class CadreNode {
             timing('[start] total: %dms', Math.round(performance.now() - tTotal));
             log('CadreNode started successfully');
             // Wire the control-connection growth listeners immediately (not via the
-            // delayed refresh path) so no early `connection:open` is missed — the
-            // write-while-alone re-replication drain must fire on the first 0→≥1 edge.
+            // delayed refresh path) so no early `peer:identify` is missed — the
+            // write-while-alone re-replication drain must fire when the first sibling is reachable.
             this.wireControlConnectionListeners();
             // Arm the control network's peer-join block catch-up now that the control
             // DB (which the push-time membership gate reads) is up. Armed even when
@@ -1524,7 +1594,7 @@ export class CadreNode {
                 `but this node serves party ${partyId} — refusing to offer another party's joined strands`);
         }
         this.joinedStrandStore = store;
-        this.joinedStrands = new JoinedStrandSession(store, this.createPartyJoinedStrandLedger());
+        this.joinedStrands = new JoinedStrandSession(store, this.createPartyJoinedStrandLedger(), () => this.strandWatcher?.refresh());
     }
     /**
      * Construct (or adopt) the node-local strand network state (see
@@ -1665,8 +1735,8 @@ export class CadreNode {
     }
     /**
      * Arm the CONTROL network's peer-join block catch-up: push every block in the
-     * control database's own raw store to each newly connected AUTHORIZED member
-     * this runtime has not yet caught up. This is what physically replicates
+     * control database's own raw store to each AUTHORIZED member on each
+     * (re)connection, skipping what it is known to hold. This is what physically replicates
      * control blocks committed while the writer was alone — the named
      * collection-header blocks written once at genesis above all, whose revision
      * never moves again, so no later commit ever carries them to a member that
@@ -1675,7 +1745,9 @@ export class CadreNode {
      * for peers it knew about before the restart).
      *
      * Unlike the per-strand instances (see `strand-instance-manager.ts`), pushes
-     * are gated on {@link isAuthorizedMember}, judged at push time: the control
+     * are gated on the authorized-member snapshot ({@link authorizedControlPeers},
+     * which {@link refreshAuthorizedControlPeers} materializes from the same rows and
+     * rule {@link isAuthorizedMember} queries), judged at push time: the control
      * network's inbound connection gate deliberately admits non-members in
      * several states (seed delivery to an un-enrolled node, an open enrollment
      * window, an outstanding invitation, configured bootstrap/relay peers), and
@@ -1691,22 +1763,18 @@ export class CadreNode {
      * The one non-member the gate admits is a device whose invitation
      * redemption this member is admitting ({@link admitRedeemingDevice}).
      *
-     * No-ops (logged) when the embedder configured no control storage or the
-     * node exposes no key network — the backfill would have nothing to read or
-     * no way to dial.
+     * No-ops (logged) when {@link controlBackfillStorage} answers nothing (disabled,
+     * or no control storage configured) or the node exposes no key network — the
+     * backfill would have nothing to read or no way to dial.
      */
     startControlBackfill() {
-        if (this.config.controlBackfill?.enabled === false || !this.controlNode) {
+        if (!this.controlNode) {
             return;
         }
-        // Through the resolver, not the {@link controlStorage} memo field: the resolver is
-        // idempotent (it returns the same wrapped store `buildControlNodeOptions` already
-        // resolved), so the catch-up cannot be silently disarmed by a future reordering that
-        // moves this call ahead of the one that populated the field.
-        const storage = this.resolveControlStorage();
+        const storage = this.controlBackfillStorage();
         const keyNetwork = this.controlNode.keyNetwork;
         if (!storage || !keyNetwork) {
-            log('Control peer-join block catch-up is inert: %s', !storage ? 'no control storage configured' : 'control node exposes no keyNetwork');
+            log('Control peer-join block catch-up is inert: %s', !storage ? 'disabled, or no control storage configured' : 'control node exposes no keyNetwork');
             return;
         }
         const networkName = this.controlNetworkName();
@@ -1719,7 +1787,11 @@ export class CadreNode {
             // under — derived from the same networkName binding the node options
             // used, never re-spelled here.
             protocolPrefix: `/optimystic/${networkName}`,
-            authorizePeer: async (peerId) => this.admittingDevices.has(peerId) || await this.isAuthorizedMember(peerId)
+            reportBlockHeld: this.controlNode.reportBlockHeld,
+            // The membership snapshot, not {@link isControlNetworkMember}: that predicate admits
+            // everyone while the snapshot is empty and admits bootstrap/relay peers outright, and
+            // neither may receive the control store.
+            authorizePeer: async (peerId) => this.admittingDevices.has(peerId) || this.authorizedControlPeers.has(peerId)
         }, {
             // A shorter settle window than the strand default (1000 ms): the control
             // store is small (a party's membership — dozens of blocks), a re-push is
@@ -1774,7 +1846,7 @@ export class CadreNode {
         const release = this.countAdmittingDevice(peerId);
         try {
             const result = await backfill.forceCatchUpPeer(peerIdFromString(peerId));
-            log('Device %s caught up before its admission: offered=%d accepted=%d rejected=%d', peerId, result.offered, result.accepted, result.rejected.length);
+            log('Device %s caught up before its admission: offered=%d alreadyHeld=%d accepted=%d rejected=%d', peerId, result.offered, result.alreadyHeld, result.accepted, result.rejected.length);
             return release;
         }
         catch (error) {
@@ -1838,6 +1910,19 @@ export class CadreNode {
         this.controlStorage = wrapStorageWithCache(resolved, scope);
         return this.controlStorage;
     }
+    /**
+     * The store the control network's peer-join block catch-up reads, or `undefined` when this
+     * runtime does not arm it: `controlBackfill.enabled: false`, or no control storage. Both of
+     * its readers ask here so they cannot disagree: {@link startControlBackfill} arms on it, and
+     * {@link buildControlNodeOptions} turns Optimystic's rebalance reaction off on it.
+     *
+     * Through the resolver, not the {@link controlStorage} memo field: the resolver is
+     * idempotent, so the catch-up cannot be silently disarmed by a future reordering that moves
+     * a read ahead of the one that populated the field.
+     */
+    controlBackfillStorage() {
+        return this.config.controlBackfill?.enabled === false ? undefined : this.resolveControlStorage();
+    }
     /**
      * Warn the operator when `network.announceAddrs` will silently cost this node the
      * relay reachability it also configured.
@@ -1940,6 +2025,19 @@ export class CadreNode {
                 cohortQueryTimeoutMs: declaredCohortReadDeadlineMs(network)
             }),
             arachnode: { enableRingZulu: profile === 'storage' },
+            // One join catch-up per node, for the strand node's reason (`strand-instance-manager.ts`,
+            // at its `rebalance` option): Optimystic's rebalance growth arm would push a joining
+            // member each control block again, one request per block. Every peer the growth arm
+            // pushes to, the control catch-up pushes to as well, except the ones it deliberately
+            // withholds the control store from: the cohort predicate (`isNetworkMember` below) admits
+            // everyone while the membership snapshot is empty, and bootstrap/relay peers outright.
+            // With the catch-up disabled the reaction stays on.
+            //
+            // NOTE: no ring shifts and no reclaimable shed blocks on this node either, unused today.
+            // If storage nodes ever need them, turn the reaction back on: the catch-up already reports
+            // its accepted pushes to the rebalance monitor (`reportBlockHeld`); measure whether the
+            // growth arm still re-sends.
+            ...(this.controlBackfillStorage() && { rebalance: { enabled: false } }),
             ...(identityKey && { privateKey: identityKey }),
             // With `network.relayAddrs` set, built with libp2p's relay discovery off (given a
             // transport list to wrap — `node-transports.ts`), so the supervisor
@@ -2002,10 +2100,8 @@ export class CadreNode {
             // Who this node SENDS control writes to (cohort and coordinator choice), where
             // `authorizeInboundStream` is who may send to it. Without it Optimystic picks cohort
             // members from the protocols a peer's peer-store record lists, which any connected
-            // peer can fill. The stream gate's rule, unlogged (Optimystic asks several times per
-            // peer per cohort assembly and counts refusals on its own debug line), plus a device
-            // whose verified redemption is being written, which must hold the rows that admit it.
-            isNetworkMember: (peerId) => this.isControlNetworkMember(peerId) || this.admittingDevices.has(peerId)
+            // peer can fill.
+            isNetworkMember: (peerId) => this.isControlCohortMember(peerId)
         };
         return nodeOptions;
     }
@@ -2085,8 +2181,8 @@ export class CadreNode {
      * (it redeems here by seating the row from its bundle, after which check 7
      * admits) or from a sibling whose membership row is still replicating (check 5
      * admits once it lands; if it has not landed by the deadline, the connection
-     * is closed and either side's next reconcile dial re-forms it, since outbound
-     * dials are never gated). The protocol guard refuses all of them every
+     * is closed and the reconcile pass the close starts on either side re-forms it,
+     * since outbound dials are never gated). The protocol guard refuses all of them every
      * members-only protocol meanwhile. A circuit-relay reservation admitted by
      * {@link admitControlRelayReservation} disarms the deadline (see
      * `membership-connection-gater.ts` → "The relay-reservation seam").
@@ -2310,6 +2406,17 @@ export class CadreNode {
         }
         return this.authorizedControlPeers.size === 0 || this.authorizedControlPeers.has(remotePeerId);
     }
+    /**
+     * Whether this node sends control writes to `remotePeerId`: Optimystic's `isNetworkMember`
+     * for the control node (cohort and coordinator choice), and the membership half of
+     * {@link hasReachableControlSibling}. The stream gate's rule ({@link isControlNetworkMember}),
+     * unlogged, because Optimystic asks several times per peer per cohort assembly and counts
+     * refusals on its own debug line; plus a device whose verified redemption is being written
+     * ({@link admittingDevices}), which must hold the rows that admit it.
+     */
+    isControlCohortMember(remotePeerId) {
+        return this.isControlNetworkMember(remotePeerId) || this.admittingDevices.has(remotePeerId);
+    }
     /**
      * Per-stream authorization for the protocol guard's `members-snapshot` class
      * (`control-protocol-guard.ts`): the four Optimystic control-DB protocols
@@ -2349,7 +2456,8 @@ export class CadreNode {
      * this listener with its write lock HELD (`notifyMembershipChanged`), so a retrying
      * read here would sleep its backoff holding the lock and stall every other local
      * writer. Nothing is lost — this refresh already keeps the previous snapshot on
-     * failure and is re-driven by the next membership write and by the timed reconcile.
+     * failure and is re-driven by the next membership change (local or replicated, through
+     * the gate watch) and by the reconcile pass's backstop.
      *
      * The trusted-owner anchor sync ({@link syncOwnerAnchor}) runs FIRST, so the snapshot
      * built here judges `CadrePeer` rows against the anchor as the `OwnerKey` table now
@@ -2424,15 +2532,20 @@ export class CadreNode {
             // commit is what would authorize it), and the gate re-judges at push time.
             //
             // NOTE: every refresh re-arms every connected peer that is not yet caught
-            // up, and each re-armed pass costs one `CadrePeer` query at the gate. A
-            // connection that is admitted but never authorized — a configured
+            // up. A connection that is admitted but never authorized — a configured
             // bootstrap or relay peer, the steady state for a NAT'd node — is
-            // therefore re-judged on every membership write and every timed reconcile
-            // tick, forever. Negligible at a cadre's handful of connections; if a node
-            // ever holds many such connections, or refreshes get frequent, skip
-            // re-arming peers whose last pass was denied under the SAME membership
-            // snapshot instead of re-arming unconditionally.
+            // therefore re-judged on every refresh, forever: every local membership
+            // write, every replicated membership change (the gate watch), and every
+            // reconcile pass, down to the 2-minute backstop on an idle party. Each
+            // such pass is one timer and a lookup in the snapshot just set (the gate
+            // reads no control database). If a node ever holds many such connections,
+            // or refreshes get frequent, skip re-arming peers whose last pass was
+            // denied under the SAME membership snapshot instead of re-arming
+            // unconditionally.
             void this.controlBackfill?.scheduleConnectedPeers();
+            // Same join order for the growth edge: a sibling that identified before this commit
+            // authorized it becomes reachable here.
+            this.updateControlSiblingReachability();
             await this.hangUpRevokedControlPeers(rows, loadTombstones);
             await this.noteOwnRemoval(rows, loadTombstones);
         }
@@ -2619,9 +2732,10 @@ export class CadreNode {
         return this.bootstrapPeerIds;
     }
     /**
-     * What the strand watcher polls: this party's control rows plus a row per cross-party
+     * What the strand watcher reads: this party's control rows plus a row per cross-party
      * join (see {@link joinedStrands}), so a joined strand is offered and relaunched on
-     * the same path as the party's own.
+     * the same path as the party's own. It runs on a change to either table; the
+     * machine-local joins wake it through {@link JoinedStrandSession}'s change callback.
      */
     createStrandQueryable() {
         return {
@@ -2633,7 +2747,14 @@ export class CadreNode {
                 log('Querying strands from control database');
                 const control = await this.controlDatabase.queryStrands();
                 return this.joinedStrands ? await this.joinedStrands.withControlRows(control) : control;
-            }
+            },
+            watchChanges: (onChange, options) => {
+                if (!this.controlDatabase) {
+                    throw new Error('CadreNode must initialize its control database before watching its strand tables');
+                }
+                return this.controlDatabase.watchTables(['CadreControl.Strand', 'CadreControl.JoinedStrand'], onChange, options);
+            },
+            lastReadPartial: () => this.joinedStrands?.lastReadPartial() ?? false
         };
     }
     /**
@@ -2648,10 +2769,6 @@ export class CadreNode {
             return this.controlDatabase;
         };
         return {
-            // NOTE: one more control read per watcher poll (5 s), and for a party that never
-            // joined anything it reads a never-written block, which consults the cohort. If poll
-            // cost shows up in a device profile, cache the list and re-read it on the reconcile
-            // cadence.
             list: () => database().queryJoinedStrands(),
             names: async (strandId) => (await database().queryStrand(strandId)) !== null
                 || (await database().queryJoinedStrand(strandId)) !== null,
@@ -2729,26 +2846,30 @@ export class CadreNode {
      * after start (non-blocking). {@link registerSelf} is idempotent and safely
      * no-ops when it cannot yet sign/insert (e.g. owner key not installed),
      * so the timer is harmless even when registration only becomes possible later.
+     * A record this process already put in place that is still current (a founding owner's
+     * explicit publish, a redeemed invitation's admission, a restart's first-growth re-touch) is
+     * not written again ({@link publishSelfRecordUnlessCurrent}). Whether the node is connected
+     * yet does not matter: a device that redeems within a second of starting is.
      */
     scheduleSelfRegistration() {
         this.selfRegistrationTimer = setTimeout(() => {
             void (async () => {
                 try {
-                    await this.registerSelf();
+                    await this.publishSelfRecordUnlessCurrent();
                 }
                 catch (error) {
-                    // Background task — a failed publish must not crash the node. A restarted machine
-                    // whose own CadrePeer row replicates in while this update runs fails here with
-                    // ConcurrentModificationError (seen on the owner's restart in every traced run of
-                    // cadre-invite-any-member); the next connected reconcile pass retries
+                    // Background task — a failed publish must not crash the node. A publish that loses
+                    // its row to another writer — the owner's re-touch, or a restarted machine's own row
+                    // replicating in while the update runs — is re-run by the write funnel over the row
+                    // it finds (`ControlDatabase.updateSelfPeerRecord`); one that lost past the retry
+                    // budget fails here, and the next connected reconcile pass is the fallback
                     // (publishFirstSelfRecord).
                     log('Self-registration failed: %o', error);
                 }
-                // Wire the ongoing refresh + control-cohort reconcile cadence, then run
-                // an eager reconcile pass once the node has settled (the recurring
-                // interval was just armed inside startRecordRefresh).
+                // Wire the ongoing refresh and the control-cohort reconcile schedule, then
+                // run an eager reconcile pass now that the node has settled.
                 this.startRecordRefresh();
-                void this.reconcileControlCohort().catch((error) => log('Control-cohort reconcile (start) failed: %o', error));
+                this.controlReconcileRuns?.trigger(0);
             })();
         }, 1000); // Small delay to ensure node is fully started
         // Node-only: don't keep the event loop alive solely for this timer.
@@ -2820,16 +2941,15 @@ export class CadreNode {
             return 'skipped';
         }
         // Collected only past that skip, which a node waiting for its row replicated in takes on
-        // every connected reconcile pass ({@link publishFirstSelfRecord}).
+        // reconcile pass made while a sibling is reachable ({@link publishFirstSelfRecord}).
         const addrs = await this.collectSelfAddrs();
         if (!existing && ownerService) {
             // First-time row: requires an owner signature (the node is its own
             // owner). insertSelfPeerRecord throws if no owner key is present.
             const record = this.signSelfRecord(peerId, signingKey, addrs, null);
+            const aloneBefore = this.committedAlone();
             if (await ownerService.insertSelfPeerRecord(record)) {
-                if (this.committedAlone())
-                    this.pendingSelfPeerWrite = true;
-                this.noteSelfRecordPublished();
+                this.noteSelfPeerWritten(record, aloneBefore);
                 log('registerSelf: inserted own CadrePeer record (owner-signed, updatedAt=%d, %d addrs, sig=%s…)', record.updatedAt, addrs.length, record.sig.slice(0, 16));
                 return 'inserted';
             }
@@ -2843,43 +2963,97 @@ export class CadreNode {
             // first publish", add a fourth SelfRegistrationOutcome rather than re-labelling.
             log('registerSelf: own CadrePeer row appeared mid-publish (concurrent authorize); self-updating to carry the signature');
         }
-        // Reached on the fall-through above, where the pre-race read missed the row a
-        // concurrent authorize then seated — so re-read to get that row's `UpdatedAt`, which
-        // the strictly-greater self-update rule is measured against. The re-read can still
-        // come back null (see the branch below).
-        //
-        // NOTE: when `existing` is set it predates this publish, so a removePeer(self) racing
-        // in would leave the UPDATE below matching no rows while still reporting 'refreshed'.
-        // Closing that needs updateSelfPeerRecord to report rows-affected — only worth doing
-        // if removing self ever becomes something that happens concurrently in practice.
-        const current = existing ?? await this.controlDatabase.queryPeerRecord(peerId);
-        if (!current) {
+        // The update body reads the row under the write lock and signs over what it finds, on
+        // every attempt — so neither `existing` (a read that predates this publish) nor a first
+        // attempt that lost the row to another writer decides the `UpdatedAt` the strictly-greater
+        // self-update rule is measured against, and a removal landing mid-publish re-reads as
+        // absent instead of matching no rows under a 'refreshed'. The record it returns is the
+        // one in the row.
+        const aloneBefore = this.committedAlone();
+        const record = await this.controlDatabase.updateSelfPeerRecord(peerId, (current) => this.signSelfRecord(peerId, signingKey, addrs, current));
+        if (record === null) {
             // Two causes, and the second is a steady state rather than a race: the row really
             // vanished mid-publish (next refresh re-inserts), or this node was REVOKED — its row
             // is physically present, so insertCadrePeer's in-lock existence check no-op'd above,
-            // but queryPeerRecord reads it as absent because its stamp is retired. The revoked
-            // case repeats every heartbeat and never re-inserts; that is intended (a revoked node
-            // has nothing to publish), so this stays a skip, not an error.
+            // but the body reads it as absent because its stamp is retired. The revoked case
+            // repeats every heartbeat and never re-inserts; that is intended (a revoked node has
+            // nothing to publish), so this stays a skip, not an error.
             log('registerSelf: own CadrePeer row not readable (vanished mid-publish, or this node is revoked and its stamp retired); skipping');
             return 'skipped';
         }
-        const record = this.signSelfRecord(peerId, signingKey, addrs, current);
-        await this.controlDatabase.updateSelfPeerRecord(record);
-        if (this.committedAlone())
-            this.pendingSelfPeerWrite = true;
-        this.noteSelfRecordPublished();
+        this.noteSelfPeerWritten(record, aloneBefore);
         log('registerSelf: refreshed own CadrePeer record (updatedAt=%d, %d addrs, sig=%s…)', record.updatedAt, addrs.length, record.sig.slice(0, 16));
         return 'refreshed';
     }
     /**
-     * Record that this node's own address record just landed, and re-arm the stale-record
-     * warning. Called from BOTH publishing paths of {@link publishSelfRecord} and nowhere
-     * else — a `skipped` publish wrote nothing, so it must not refresh the stamp.
+     * Bookkeeping for a committed write of this node's own `CadrePeer` row by {@link publishSelfRecord}:
+     * `record` is the one the row holds (for the update path, the one its body signed on the
+     * attempt that committed, not the first it signed). Queued for the drain's re-touch
+     * ({@link pendingSelfPeerWrite}) when no sibling was reachable when the write started or when
+     * it ended: a sibling that identifies while the write is out may do so after the write's cohort
+     * was chosen, and the first-growth drain skips a row holding {@link lastSignedSelfRecord} unless
+     * this flag is set. Both fields are set here together so no write can leave the record without
+     * the flag.
+     */
+    noteSelfPeerWritten(record, aloneBefore) {
+        if (aloneBefore || this.committedAlone())
+            this.pendingSelfPeerWrite = true;
+        this.lastSignedSelfRecord = record;
+        this.noteSelfRecordPublished();
+    }
+    /**
+     * Record that this node's own address record is in place as of `at`, and re-arm the
+     * stale-record warning. Called for BOTH publishing paths of {@link publishSelfRecord}
+     * ({@link noteSelfPeerWritten}), and by {@link publishSelfRecordUnlessCurrent} with the stamp
+     * of a record already in place — a `skipped` publish wrote nothing and found nothing, so it
+     * must not refresh the stamp.
      */
-    noteSelfRecordPublished() {
-        this.lastSelfRecordPublishAt = Date.now();
+    noteSelfRecordPublished(at = Date.now()) {
+        this.lastSelfRecordPublishAt = at;
         this.selfRecordStaleWarned = false;
     }
+    /**
+     * {@link registerSelf}, unless the own row already holds the record this process last signed
+     * ({@link lastSignedSelfRecord}) and that record is current ({@link isCurrentSelfRecord});
+     * then nothing is written and the record counts as published at its own stamp. For a publish
+     * that puts a record in place (the boot timer, the reconcile pass's first publish, the
+     * first-growth drain), never the heartbeat, which exists to re-stamp freshness. `'current'`
+     * is this method's own answer; {@link registerSelf} always writes.
+     */
+    async publishSelfRecordUnlessCurrent() {
+        if (this.registerSelfInFlight) {
+            return this.registerSelfInFlight;
+        }
+        const record = this.lastSignedSelfRecord;
+        if (record !== null && await this.isCurrentSelfRecord(record)) {
+            this.noteSelfRecordPublished(record.updatedAt);
+            log('registerSelf: own CadrePeer record already current (updatedAt=%d); not written again', record.updatedAt);
+            return 'current';
+        }
+        return this.registerSelf();
+    }
+    /**
+     * Is `record` still the one to publish: signed less than half a heartbeat ago, with this
+     * node's profile, over the addresses it would publish now, and the signature the own row
+     * holds? The age bound keeps a record that is not re-written fresh for resolvers: at most half
+     * a heartbeat old here and re-stamped by the heartbeat within one more, so never older than
+     * 1.5 heartbeats (11.25 min) against {@link DEFAULT_PEER_RECORD_MAX_AGE_MS}. Cheap checks
+     * first: the row comes from the membership view, the addresses from the app's resolver.
+     */
+    async isCurrentSelfRecord(record) {
+        if (Date.now() - record.updatedAt >= DEFAULT_PEER_RECORD_HEARTBEAT_MS / 2 || record.profile !== this.config.profile) {
+            return false;
+        }
+        const row = await this.controlDatabase?.queryPeerRecord(record.peerId);
+        if (row?.sig !== record.sig) {
+            return false;
+        }
+        // Sanitized, as the record a redemption carries is (`signRedeemRequest`); a published
+        // record is not, so one whose list sanitizing would change (over 16 addresses) is never
+        // current and is written as before.
+        const addrs = sanitizeAddrs(await this.collectSelfAddrs());
+        return addrs.length === record.addrs.length && addrs.every((addr, i) => addr === record.addrs[i]);
+    }
     /**
      * Report an abandoned control write to the embedding app.
      *
@@ -3007,19 +3181,24 @@ export class CadreNode {
      * process does not learn it was removed ({@link noteOwnRemoval}), and pays that read on every
      * pass.
      *
-     * Connected-only: a row written elsewhere arrives over a connection, and an owner's publish
-     * committed alone is local-only and forks its own history — a restarted owner whose embedder
-     * runs a pass before it reconnects would write its row on the stale side of the fork, and
-     * then never see rows written while it was down (`cadre-invite-any-member` runs that pass and
-     * fails every time without this condition). The boot-time publish and the first-growth drain already try at the node's lone and
-     * first-connection moments; this retries on the passes after.
+     * Only while a sibling is reachable ({@link hasReachableControlSibling}): a row written
+     * elsewhere arrives from a sibling, and an owner's publish committed alone is local-only and
+     * forks its own history — a restarted owner whose embedder runs a pass before it reconnects
+     * would write its row on the stale side of the fork, and then never see rows written while it
+     * was down (`cadre-invite-any-member` runs that pass and fails every time without this
+     * condition). A connection is not enough: until the sibling identifies, the write's cohort is
+     * this node alone. The boot-time publish and the first-growth drain already try at the
+     * node's lone and first-sibling moments; this retries on the passes after.
+     *
+     * A machine admitted by invitation finds its row already holding the record its redemption
+     * carried, and writes nothing ({@link publishSelfRecordUnlessCurrent}).
      */
     async publishFirstSelfRecord() {
         if (this.lastSelfRecordPublishAt !== null || this.isAwaitingClaim() || this.removal !== null
-            || this.getControlConnectionCount() === 0) {
+            || !this.hasReachableControlSibling()) {
             return;
         }
-        const outcome = await this.registerSelf().catch((error) => {
+        const outcome = await this.publishSelfRecordUnlessCurrent().catch((error) => {
             log('reconcileControlCohort: first self-record publish failed (continuing): %o', error);
             return 'skipped';
         });
@@ -3055,15 +3234,23 @@ export class CadreNode {
                 this.escalateIfSelfRecordStale(reason, error);
             });
         };
-        const reconcile = (reason) => {
-            void this.reconcileControlCohort().catch((error) => log('Control-cohort reconcile (%s) failed: %o', reason, error));
-        };
+        const backstopMs = this.controlReconcileBackstopMs();
+        // The pass's schedule. A run never joins a pass already in flight: that pass may have
+        // read its rows before the change that woke this run, so joining it would drop the
+        // change until the backstop. It waits for the pass, then starts its own.
+        this.controlReconcileRuns = scheduleRuns(async () => {
+            await this.reconcileControlCohortInFlight?.catch(() => undefined);
+            await this.reconcileControlCohort().catch((error) => log('Control-cohort reconcile (scheduled) failed: %o', error));
+        }, { label: 'control-reconcile', backstopMs });
         // Address churn re-publishes the self record AND re-checks the control cohort:
         // a relay-reservation rotation / NAT change can drop a sibling connection, so
-        // the next pass should re-observe and re-dial it.
-        this.selfPeerUpdateHandler = () => {
+        // the next pass should re-observe and re-dial it. Only address churn: libp2p also
+        // fires this for a protocol handler registered or removed (see selfAddressesChanged).
+        this.selfPeerUpdateHandler = (event) => {
+            if (!selfAddressesChanged(event.detail))
+                return;
             republish('self:peer:update');
-            reconcile('self:peer:update');
+            this.controlReconcileRuns?.trigger(0);
         };
         this.controlNode.addEventListener('self:peer:update', this.selfPeerUpdateHandler);
         // A node AUTHORIZED AFTER IT BOOTED (seeded, claimed, invited) gets its first
@@ -3083,12 +3270,28 @@ export class CadreNode {
         // dependable there.
         this.recordRefreshTimer = setInterval(() => republish('heartbeat'), DEFAULT_PEER_RECORD_HEARTBEAT_MS);
         this.recordRefreshTimer?.unref?.();
-        const reconcileMs = this.config.network?.controlCohort?.reconcileMs ?? DEFAULT_CONTROL_COHORT_RECONCILE_MS;
-        this.controlCohortReconcileTimer = setInterval(() => reconcile('interval'), reconcileMs);
-        this.controlCohortReconcileTimer?.unref?.();
-        log('Record refresh wired (heartbeat=%dms, cohortReconcile=%dms)', DEFAULT_PEER_RECORD_HEARTBEAT_MS, reconcileMs);
+        log('Record refresh wired (heartbeat=%dms, cohortReconcileBackstop=%dms)', DEFAULT_PEER_RECORD_HEARTBEAT_MS, backstopMs);
     }
-    /** Tear down the self-record refresh timers + listener (see {@link cleanup}). */
+    /** The reconcile schedule's backstop: `network.controlCohort.reconcileMs`, default `DEFAULT_CONTROL_COHORT_RECONCILE_MS`. */
+    controlReconcileBackstopMs() {
+        return this.config.network?.controlCohort?.reconcileMs ?? DEFAULT_CONTROL_COHORT_RECONCILE_MS;
+    }
+    /**
+     * The gate watch's run ({@link membershipGateWatch}): re-materialize the gate, so a
+     * replicated revocation hangs up the removed machine now (`hangUpRevokedControlPeers`
+     * ends every refresh), then start a reconcile pass for whatever else the change means
+     * (a new sibling to dial, a tombstoned row to reap). A local membership write refreshes
+     * the gate twice — once from the awaited membership listener, so the writer sees its
+     * change before returning, and once here, 50 ms later. The listener's refresh runs under
+     * the write lock, so its reads take the committed-read path and the view keeps nothing
+     * from them; this refresh is the read that refills the view after the write, which some
+     * later caller would otherwise pay. Accepted: one reload per change, not two.
+     */
+    async onMembershipGateTablesChanged() {
+        await this.refreshMembershipGate('table-change');
+        this.controlReconcileRuns?.trigger(0);
+    }
+    /** Tear down the self-record refresh timers, the reconcile schedule and the listeners (see {@link cleanup}). */
     stopRecordRefresh() {
         if (this.selfRegistrationTimer) {
             clearTimeout(this.selfRegistrationTimer);
@@ -3098,24 +3301,30 @@ export class CadreNode {
             clearInterval(this.recordRefreshTimer);
             this.recordRefreshTimer = null;
         }
-        if (this.controlCohortReconcileTimer) {
-            clearInterval(this.controlCohortReconcileTimer);
-            this.controlCohortReconcileTimer = null;
+        if (this.controlReconcileRuns) {
+            this.controlReconcileRuns.dispose();
+            this.controlReconcileRuns = null;
+        }
+        if (this.membershipViewReloadTimer) {
+            clearTimeout(this.membershipViewReloadTimer);
+            this.membershipViewReloadTimer = null;
         }
+        this.reconcileRedialStreak = 0;
         if (this.controlNode) {
             if (this.selfPeerUpdateHandler) {
                 this.controlNode.removeEventListener('self:peer:update', this.selfPeerUpdateHandler);
             }
-            if (this.controlConnectionOpenHandler) {
-                this.controlNode.removeEventListener('connection:open', this.controlConnectionOpenHandler);
+            if (this.controlPeerIdentifyHandler) {
+                this.controlNode.removeEventListener('peer:identify', this.controlPeerIdentifyHandler);
             }
             if (this.controlConnectionCloseHandler) {
                 this.controlNode.removeEventListener('connection:close', this.controlConnectionCloseHandler);
             }
         }
         this.selfPeerUpdateHandler = null;
-        this.controlConnectionOpenHandler = null;
+        this.controlPeerIdentifyHandler = null;
         this.controlConnectionCloseHandler = null;
+        this.identifiedControlPeers.clear();
         // Reset the write-while-alone re-replication state so a stop()→start() cycle
         // re-arms the growth edge and re-runs BOTH one-shot passes. Any queued
         // local-only writes are moot once the node is torn down: owner inserts are
@@ -3123,10 +3332,9 @@ export class CadreNode {
         // first-growth revocation sweep — which is why the sweep flag must reset too,
         // or the second lifetime would skip the sweep and strand exactly the
         // committed-alone tombstones it exists to carry.
-        this.hasControlConnection = false;
+        this.controlSiblingReachable = false;
         this.reconstructedLocalOnlyWrites = false;
         this.reissuedHeldRevocations = false;
-        this.revocationLedgerOpened = false;
         this.pendingPeerWrites.clear();
         this.pendingRevocations.clear();
         this.pendingSelfPeerWrite = false;
@@ -3135,6 +3343,7 @@ export class CadreNode {
         // from scratch, and carrying the previous lifetime's stamp forward would either
         // suppress the stale-record warning or fire it on a node that just came up.
         this.lastSelfRecordPublishAt = null;
+        this.lastSignedSelfRecord = null;
         this.selfRecordStaleWarned = false;
     }
     /**
@@ -3311,26 +3520,37 @@ export class CadreNode {
      * Run one proactive control-cohort dial pass to keep this node connected to its
      * cadre siblings (so the `CadreControl` collections form a replicating cohort).
      * Public so the cohort-growth-driven re-replication path
-     * (`control-write-ensure-replicated`) and tests can trigger a pass on demand;
-     * normally driven by the eager start pass, the recurring interval, and
-     * `self:peer:update` (all wired in {@link startRecordRefresh}).
+     * (`control-write-ensure-replicated`), `addDrone`'s callers, `claimNode`,
+     * `redeemCadreInvitation` and tests can run a pass on demand. On its own a pass
+     * runs when something it acts on changes, and otherwise at a backstop
+     * ({@link controlReconcileRuns}): once after start; on `self:peer:update`; when a
+     * membership table changes, local commit or replicated ({@link membershipGateWatch});
+     * when a control connection to an authorized sibling opens or closes
+     * ({@link onControlSiblingConnectionChange}); on the redial and strand-address wakes
+     * a pass arms for itself ({@link armReconcileWake}); and at the backstop,
+     * `network.controlCohort.reconcileMs` (default `DEFAULT_CONTROL_COHORT_RECONCILE_MS`,
+     * 2 minutes) after the last pass settled. A connection to a relay, a bootstrap node or
+     * another party's machine starts no pass: the pass acts on none of them, beyond
+     * counting them as "connected" for its write steps, which therefore run at the next
+     * membership change or backstop on a one-machine party whose only connection is a
+     * relay.
      *
      * Concurrent triggers collapse into a single in-flight pass
      * (see {@link reconcileControlCohortInFlight}) so two passes never double-dial;
-     * a call that joins an in-flight pass resolves to THAT pass's result.
+     * a call that joins an in-flight pass resolves to THAT pass's result (a scheduled
+     * run waits for it instead, so the change that woke the run is seen).
      * Best-effort throughout: a failure to resolve/dial any one sibling is logged
      * and the pass continues; the whole pass is a no-op when the node is alone.
      *
      * Resolves to the peers the pass dialled ({@link ControlCohortReconcileResult}),
      * so a caller can tell a link this pass opened from one something else opened —
-     * the pass skips a peer that is already connected. The timer and event triggers
-     * discard it.
+     * the pass skips a peer that is already connected. The scheduled runs discard it.
      */
     async reconcileControlCohort() {
         if (this.reconcileControlCohortInFlight) {
             return this.reconcileControlCohortInFlight;
         }
-        const op = this.runReconcileControlCohort();
+        const op = this.runReconcileControlCohortAndArmWake();
         this.reconcileControlCohortInFlight = op;
         try {
             return await op;
@@ -3339,38 +3559,75 @@ export class CadreNode {
             this.reconcileControlCohortInFlight = null;
         }
     }
+    /** One pass, then the wake it earns ({@link armReconcileWake}); the public result is the pass's `dialed`. */
+    async runReconcileControlCohortAndArmWake() {
+        const startedAt = Date.now();
+        const outcome = await this.runReconcileControlCohort();
+        this.armReconcileWake(outcome, startedAt);
+        return { dialed: outcome.dialed };
+    }
+    /**
+     * After every pass, explicit or scheduled: advance or reset the redial streak and
+     * `trigger` the schedule for the earliest of the pass's own reasons to run again
+     * (`nextReconcileWakeMs`): a redial for a sibling or bootstrap peer it left
+     * unconnected, or a strand-address ask falling due before the backstop. Nothing else
+     * the pass does needs a wake sooner than the backstop: delegate grants re-announce
+     * every 15 minutes inside a 30-minute TTL, the FRET and formation address re-merges
+     * guard a one-hour expiry, and the self record has its own heartbeat. A pass that
+     * runs before the schedule exists (an explicit call during the first second after
+     * start) still counts for the streak.
+     */
+    armReconcileWake(outcome, startedAt) {
+        this.reconcileRedialStreak = outcome.unconnected ? this.reconcileRedialStreak + 1 : 0;
+        const runs = this.controlReconcileRuns;
+        if (!runs) {
+            return;
+        }
+        const wakeMs = nextReconcileWakeMs({
+            redialStreak: this.reconcileRedialStreak,
+            passStartedAt: startedAt,
+            strandAddrDueAt: this.strandAddrAskDueAt.values(),
+            now: Date.now(),
+            backstopMs: this.controlReconcileBackstopMs(),
+        });
+        if (wakeMs !== undefined) {
+            log('reconcileControlCohort: next pass in %dms (redial streak %d)', wakeMs, this.reconcileRedialStreak);
+            runs.trigger(wakeMs);
+        }
+    }
     /** Body of {@link reconcileControlCohort}; serialised by its single-flight guard. */
     async runReconcileControlCohort() {
         // Shutdown / not-yet-started guard (mirrors publishSelfRecord). A pass that
         // fires after stop() began must early-return rather than touch a torn-down node.
         if (!this._running || !this.controlNode || !this.controlDatabase) {
-            return { dialed: [] };
-        }
-        // Ride this pass's cadence to refresh the per-stream gate's materialized
-        // authorized set — membership changes that ARRIVED BY REPLICATION (rather
-        // than a local write) are picked up here, bounding the snapshot's staleness
-        // to the time between pass starts: the reconcile interval, or the previous
-        // pass's length when that is longer (passes are single-flight, and one
-        // unreachable sibling's dial alone can take `controlDialBudget().totalMs`).
+            return { dialed: [], unconnected: false };
+        }
+        // Refresh the per-stream gate's materialized authorized set at the head of every
+        // pass, so the sibling enumeration below judges rows the same way the gate does.
+        // Changes reach the gate on their own through the gate watch; this refresh is the
+        // backstop's, for a change this node's storage never heard of.
         // NOTE: this refresh and the sibling enumeration below each run their own
         // CadrePeer query (two reads per pass), plus a third from
         // `refreshStrandPeerAddrs` on every pass where a strand is running AND this
         // node holds a control connection (one read for the whole pass, not one per
         // strand; it runs even when no sibling is due, because pruning departed
         // siblings needs the current target set). Each also reads Revocation first.
-        // Before the Revocation ledger marker exists (filed below, once connected)
-        // that block is missing, so every one of those reads consults the cohort
-        // about it; once the marker exists every
-        // block they touch is held and none of them does (both states pinned in
-        // control-founding-consult-budget.spec.ts). If those reads ever get costly,
-        // share one row-set across all three.
+        // All of them are served from the control database's membership views
+        // (`ControlDatabase.readMembershipRows`): an in-memory scan while the view is
+        // current, storage only for the first read after the table changed or after
+        // `MEMBERSHIP_VIEW_MAX_AGE_MS`, and an idle pass makes no cohort consult
+        // (pinned in control-founding-consult-budget.spec.ts). Threading one row set
+        // through the refresh, the reap, the enumeration and the strand-address targets
+        // would save no storage read and was declined
+        // (`control-reconcile-woken-by-changes-reads-once`); rows going stale during a
+        // pass is harmless, since a change during the pass starts another one.
         await this.refreshMembershipGate('reconcile');
         if (!this._running || !this.controlNode || !this.controlDatabase) {
-            return { dialed: [] };
+            return { dialed: [], unconnected: false };
         }
         await this.publishFirstSelfRecord();
         if (!this._running || !this.controlNode || !this.controlDatabase) {
-            return { dialed: [] };
+            return { dialed: [], unconnected: false };
         }
         // Refresh relay delegate grants BEFORE the sibling enumeration below: a
         // solo cadre with a party relay has no siblings to dial but must still keep
@@ -3378,7 +3635,7 @@ export class CadreNode {
         // the reservation and faces the connection gate again).
         await this.refreshDelegateGrants();
         if (!this._running || !this.controlNode || !this.controlDatabase) {
-            return { dialed: [] };
+            return { dialed: [], unconnected: false };
         }
         // Then re-warm each running strand's own address book from its FRET address
         // records and from whichever siblings are due an ask. Same reasoning as
@@ -3386,14 +3643,15 @@ export class CadreNode {
         // network, and every layer under cadre-core dials strand peers by bare peer id.
         await this.refreshStrandPeerAddrs();
         if (!this._running || !this.controlNode || !this.controlDatabase) {
-            return { dialed: [] };
+            return { dialed: [], unconnected: false };
         }
         const selfPeerId = this.controlNode.peerId.toString();
         // Reap guarded rows this node still holds for incarnations an already-committed
         // Revocation tombstone retires — the half of a removal replication cannot carry
-        // (a delete of an already-absent row replays as nothing). Connected-only: a reap
-        // is a write, and a write committed alone is local-only and forks this node's own
-        // history, which is the condition this work exists to stop creating
+        // (a delete of an already-absent row replays as nothing). Only while a sibling is
+        // reachable (hasReachableControlSibling, not merely connected): a reap is a write, and a
+        // write committed alone is local-only and forks this node's own history, which is the
+        // condition this work exists to stop creating
         // (tickets/blocked/forked-control-collection-sync-livelocks.md). Nothing is urgent
         // here — the stale row already reads as absent everywhere — so waiting for
         // connectivity costs nothing and keeps the reap fork-CLOSING rather than
@@ -3415,7 +3673,7 @@ export class CadreNode {
         // avoid. Bounded by how long a sweep runs, which is bounded by the backlog of
         // unreaped rows (0 in steady state). If that backlog can ever be large, move the
         // check inside reapRevokedRows' loop via an injected predicate.
-        if (this.getControlConnectionCount() > 0) {
+        if (this.hasReachableControlSibling()) {
             try {
                 const reaped = await this.controlDatabase.reapRevokedRows(selfPeerId);
                 if (reaped > 0) {
@@ -3428,32 +3686,14 @@ export class CadreNode {
                 log('reconcileControlCohort: reap pass failed (continuing): %o', error);
             }
             if (!this._running || !this.controlNode || !this.controlDatabase) {
-                return { dialed: [] };
-            }
-            // File the Revocation ledger marker once, so that table stops being a never-written
-            // block, which the storage layer re-checks with the cohort on every read (and every
-            // membership lookup and guarded insert reads it). Connected-only for the reap's
-            // reason above: a marker committed alone is local-only, and one filed by a
-            // disconnected owner while another machine creates the same collection is exactly
-            // that fork. Owner-only, and at most once per process — see openRevocationLedgerIfDue.
-            //
-            // NOTE: accepted tradeoff — a solo founder does not file the marker until it holds a
-            // control connection (normally its first sibling; a relay or bootstrap connection also
-            // counts — see getControlConnectionCount — and is the reap's same residual), so until
-            // then it keeps paying one local findCluster per read of
-            // the missing Revocation block (0.009 ms each, measured upstream; no network work on
-            // a cohort of one). Fork safety weighed over that and kept; revisit if findCluster
-            // ever shows up as material in a device profile, or if a solo-founding marker can be
-            // made fork-safe (e.g. filed inside the genesis transaction, before any other machine
-            // can hold the party's collections).
-            await this.openRevocationLedgerIfDue();
-            if (!this._running || !this.controlNode || !this.controlDatabase) {
-                return { dialed: [] };
+                return { dialed: [], unconnected: false };
             }
             // Publish this machine's unpublished cross-party joins party-wide and remove the rows a
-            // self-revocation queued (JoinedStrandSession.syncWithParty). Connected-only for the
-            // reap's reason above; owner-only by the session's own signer check. Nothing is urgent:
-            // a join reaches the party within one pass of this machine being connected.
+            // self-revocation queued (JoinedStrandSession.syncWithParty). Behind the reap's gate,
+            // for its reason; owner-only by the session's own signer check. Nothing is urgent:
+            // a join recorded locally starts no pass, so it reaches the party on the next pass
+            // while a sibling is reachable — a membership change, a sibling connecting, or the
+            // backstop, 2 minutes at most.
             try {
                 await this.joinedStrands?.syncWithParty();
             }
@@ -3461,7 +3701,7 @@ export class CadreNode {
                 log('reconcileControlCohort: joined-strand sync failed (retrying next pass): %o', error);
             }
             if (!this._running || !this.controlNode || !this.controlDatabase) {
-                return { dialed: [] };
+                return { dialed: [], unconnected: false };
             }
         }
         // 1. Enumerate known siblings. This membership read is itself a pull-on-read
@@ -3475,11 +3715,11 @@ export class CadreNode {
             // because filling this table needs a connection and getting a connection
             // needs this table. Fall back to the seed's bootstrap addresses; a solo
             // cadre has none and this is a no-op. Must not throw or busy-loop.
-            return { dialed: await this.dialColdStartBootstrap() };
+            return this.dialColdStartBootstrap();
         }
         // Re-guard after the await: a stop() may have raced the membership read.
         if (!this._running || !this.controlNode || !this.controlDatabase) {
-            return { dialed: [] };
+            return { dialed: [], unconnected: false };
         }
         // 2. Classify backbone (always-on) members and select a bounded dial set.
         const targetDegree = this.config.network?.controlCohort?.targetDegree
@@ -3493,51 +3733,32 @@ export class CadreNode {
         //    keeping what resolved for the dial loop below.
         const resolved = await this.warmSiblingAddrBook(siblings);
         if (!this._running || !this.controlNode || !this.controlDatabase) {
-            return { dialed: [] };
+            return { dialed: [], unconnected: false };
         }
         // 4. Skip already-connected peers (no re-dial / churn for live connections).
         const connected = new Set(this.controlNode.getConnections().map((c) => c.remotePeer.toString()));
         // 5. Dial each selected, not-yet-connected sibling, best-effort, from the
-        //    addresses step 3 already resolved for it.
+        //    addresses step 3 already resolved for it. A failed dial earns the pass a
+        //    redial wake (armReconcileWake); a sibling with nothing to dial does not.
         const dialed = [];
+        let unconnected = false;
         for (const sibling of dials) {
             if (!this._running || !this.controlNode) {
-                return { dialed };
+                return { dialed, unconnected: false };
             }
             if (connected.has(sibling.peerId)) {
                 continue;
             }
-            if (await this.dialControlSibling(sibling, resolved.get(sibling.peerId) ?? [])) {
+            const outcome = await this.dialControlSibling(sibling, resolved.get(sibling.peerId) ?? []);
+            if (outcome === 'connected') {
                 dialed.push(sibling.peerId);
             }
+            else if (outcome === 'failed') {
+                unconnected = true;
+            }
         }
-        log('reconcileControlCohort: pass complete (siblings=%d, selected=%d, dialed=%d)', siblings.length, dials.length, dialed.length);
-        return { dialed };
-    }
-    /**
-     * The owner-only half of the reconcile pass's ledger-marker step: file the singleton
-     * `Revocation` marker ({@link SeedBootstrapService.openRevocationLedger}) unless this
-     * process has already seen it filed. The connectivity gate is the caller's
-     * ({@link runReconcileControlCohort}); why the marker exists is on
-     * {@link ControlDatabase.openRevocationLedger}.
-     *
-     * Sets {@link revocationLedgerOpened} on `'opened'` or `'already-open'`. A node that
-     * cannot sign as an owner does nothing: it picks the owner's marker up the first time it
-     * reads the table, like any other replicated row. Best-effort like every step of the
-     * pass — a failure is logged and leaves the flag clear, so the next connected pass retries.
-     */
-    async openRevocationLedgerIfDue() {
-        if (this.revocationLedgerOpened || !this.seedBootstrapService?.canAuthorize()) {
-            return;
-        }
-        try {
-            const outcome = await this.seedBootstrapService.openRevocationLedger();
-            this.revocationLedgerOpened = true;
-            log('reconcileControlCohort: revocation ledger marker %s', outcome);
-        }
-        catch (error) {
-            log('reconcileControlCohort: filing the revocation ledger marker failed (retrying next pass): %o', error);
-        }
+        log('reconcileControlCohort: pass complete (siblings=%d, selected=%d, dialed=%d, unconnected=%s)', siblings.length, dials.length, dialed.length, unconnected);
+        return { dialed, unconnected };
     }
     /**
      * Resolve every sibling's signed address record once and merge what resolves
@@ -3562,8 +3783,9 @@ export class CadreNode {
      * current ({@link refreshDialHint}).
      *
      * NOTE: this resolves EVERY sibling serially before the dial loop below runs,
-     * so it costs one record query per sibling per reconcile pass (~15s) and each
-     * one delays the pass's first dial. A cadre is a handful of devices, so today
+     * so it costs one record query per sibling per reconcile pass (served from the
+     * membership rows view while it is current) and each one delays the pass's first
+     * dial. A cadre is a handful of devices, so today
      * that is a few extra local reads; if cadres ever grow large, batch the records
      * into one query, merge only on change, or move the warm pass after the dials.
      */
@@ -3688,40 +3910,41 @@ export class CadreNode {
         });
     }
     /**
-     * Dial one sibling from its already-resolved addresses, best-effort. Returns
-     * whether the dial resolved (false when no address resolves or the dial fails).
+     * Dial one sibling from its already-resolved addresses, best-effort.
      *
      * A per-peer failure (NAT, offline, relay down, connection-gater denial, or
      * the {@link controlDialBudget} expiring) is logged and swallowed so one
      * unreachable sibling never aborts the pass — exactly like
      * {@link SeedBootstrapService.applySeed}'s owner-dial loop. A failed dial is
-     * simply retried on the next pass. A sibling whose every address relays
-     * through this node ({@link SelfRelayOnlyError}) is not dialed at all and gets
-     * a one-line log instead: only the sibling can reconnect.
+     * retried on the pass's redial ladder ({@link armReconcileWake}: 15 s, 30 s, 60 s,
+     * then every backstop) and on any pass a change starts sooner. A sibling with no
+     * dialable address, or whose every address relays through this node
+     * ({@link SelfRelayOnlyError}), is not dialed at all and is `undialable`: it gets a
+     * one-line log and no redial, since only a change — its record replicating in, or
+     * the sibling connecting — can make it dialable, and each of those starts a pass.
      */
     async dialControlSibling(sibling, resolved) {
         const controlNode = this.controlNode;
         if (!controlNode) {
-            return false;
+            return 'undialable';
         }
         const addrs = await this.resolveControlDialAddrs(sibling.peerId, resolved);
         if (addrs.length === 0) {
             log('reconcileControlCohort: no dialable control address for sibling %s; skipping', sibling.peerId);
-            return false;
+            return 'undialable';
         }
         try {
             log('reconcileControlCohort: dialing sibling %s (%d addr(s))', sibling.peerId, addrs.length);
             await dialPeerAddrs(controlNode, sibling.peerId, addrs, this.controlDialBudget(), `reconcileControlCohort dial of sibling ${sibling.peerId}`);
-            return true;
+            return 'connected';
         }
         catch (error) {
             if (error instanceof SelfRelayOnlyError) {
                 log('reconcileControlCohort: sibling %s is reachable only by relaying through this node; waiting for it to reconnect', sibling.peerId);
+                return 'undialable';
             }
-            else {
-                log('reconcileControlCohort: dial of sibling %s failed (continuing): %o', sibling.peerId, error);
-            }
-            return false;
+            log('reconcileControlCohort: dial of sibling %s failed (continuing): %o', sibling.peerId, error);
+            return 'failed';
         }
     }
     /**
@@ -3774,7 +3997,7 @@ export class CadreNode {
      * These addresses are never merged into the libp2p address book: they are
      * unverified, and the address book takes only verified ones (see
      * `mergePeerAddrs`). Layers that dial by bare peer id use the connection this
-     * dial opens, and after it drops the next reconcile pass dials again.
+     * dial opens, and after it drops the reconcile pass the close starts dials again.
      *
      * NOTE: an entry is consulted here only for a current sibling, so a peer revoked
      * by another owner is not dialed from it, even though the entry stays (only this
@@ -3825,8 +4048,7 @@ export class CadreNode {
      * Called for every seed this node accepts, on BOTH intake paths — the
      * {@link applySeed} wrapper (which may run on a throwaway service, so the
      * service itself is the wrong place to keep this) and the inbound
-     * `/sereus/seed/1.0.0` handler via `onSeedApplied`. Peers with no address are
-     * skipped: there is nothing to dial.
+     * `/sereus/seed/1.0.0` handler via `onSeedApplied`.
      *
      * `preferred` is the seed's own claim about its peers, exactly as
      * `SeedBootstrapService.applySeed` uses it to choose its dial targets. That is
@@ -3834,14 +4056,15 @@ export class CadreNode {
      * signature-checked against a trust-anchored signer before this runs, and the
      * flag only *selects a dial target*; a dial grants no authority.
      *
-     * Addressless peers and self are skipped by {@link retainDialTarget}.
+     * The peers retained are `seedDialTargets` (addressless peers and self dropped), the same
+     * ones the service dials; they are returned for the `seed:applied` payload.
      */
     recordSeedBootstrapPeers(seed) {
-        for (const peer of seed.peers) {
-            if (peer.preferred) {
-                this.retainDialTarget(peer.peerId, peer.multiaddrs, 'recordSeedBootstrapPeers');
-            }
+        const targets = seedDialTargets(seed, this.controlNode?.peerId.toString());
+        for (const { peerId, multiaddrs } of targets) {
+            this.retainDialTarget(peerId, multiaddrs, 'recordSeedBootstrapPeers');
         }
+        return targets;
     }
     /**
      * Retain (or replace) one peer's out-of-band dial target in
@@ -3855,9 +4078,8 @@ export class CadreNode {
      * failing a seed or an add that has already been accepted.
      *
      * Two peers are never retained. One with no address: there is nothing to dial.
-     * And self: `createSeed` projects EVERY `CadrePeer` row, so an owner that applies
-     * a seed minted after it joined finds itself in the owner list, and `addDrone`
-     * could be handed this node's own id. Retaining self would make the cold-start
+     * And self: `addDrone` could be handed this node's own id (seed intake arrives
+     * already filtered by `seedDialTargets`). Retaining self would make the cold-start
      * pass dial this node forever (the steady-state pass filters self out of its
      * sibling list for the same reason).
      */
@@ -3891,15 +4113,16 @@ export class CadreNode {
      * help: it dials only siblings enumerated from that very table. Retrying here
      * is the only way back in.
      *
-     * Unbounded, at the reconcile cadence, with no backoff: the branch is already
-     * gated by "the control database has no siblings", so it stops the moment the
-     * node is actually in the party, and a node that is stranded MUST keep trying —
-     * a give-up rule would turn a transient outage into a permanent one. The steady
-     * cost is one dial per bootstrap peer per pass (a seed prefers its minter and the
-     * party's always-on machines, a few at most), each failing fast against an
-     * unreachable address.
-     * NOTE: if seeds ever carry many preferred peers, or the reconcile cadence tightens
-     * well below its 15 s default, add per-peer backoff here.
+     * Unbounded: the branch is already gated by "the control database has no
+     * siblings", so it stops the moment the node is actually in the party, and a node
+     * that is stranded MUST keep trying — a give-up rule would turn a transient outage
+     * into a permanent one. A failed dial earns the pass a redial wake on the ladder
+     * {@link armReconcileWake} runs (15 s, 30 s, 60 s, then every backstop, 2 minutes
+     * by default), so the steady cost is one dial per bootstrap peer per backstop (a
+     * seed prefers its minter and the party's always-on machines, a few at most), each
+     * failing fast against an unreachable address.
+     * NOTE: the ladder is per pass, not per peer; if seeds ever carry many preferred
+     * peers, add per-peer backoff here.
      *
      * Best-effort per peer, exactly like {@link dialControlSibling}: one dead
      * address never aborts the pass.
@@ -3910,14 +4133,15 @@ export class CadreNode {
         // cannot mutate what we are iterating.
         const targets = this.bootstrapPeerStore?.all();
         if (!controlNode || !targets || targets.size === 0) {
-            return [];
+            return { dialed: [], unconnected: false };
         }
         // Same skip rule as step 3 of the steady-state pass: no churn on live links.
         const connected = new Set(controlNode.getConnections().map((c) => c.remotePeer.toString()));
         const dialed = [];
+        let unconnected = false;
         for (const [peerId, entry] of targets) {
             if (!this._running || !this.controlNode) {
-                return dialed;
+                return { dialed, unconnected: false };
             }
             // NOTE: a connection this node still holds in `status: 'open'` after the remote
             // has already aborted it counts as connected here, so the cold-start retry stays
@@ -3932,9 +4156,12 @@ export class CadreNode {
             if (await this.dialBootstrapPeer(peerId, entry.addrs)) {
                 dialed.push(peerId);
             }
+            else {
+                unconnected = true;
+            }
         }
-        log('reconcileControlCohort: cold-start pass complete (bootstrap=%d, connected=%d, dialed=%d)', targets.size, connected.size, dialed.length);
-        return dialed;
+        log('reconcileControlCohort: cold-start pass complete (bootstrap=%d, connected=%d, dialed=%d, unconnected=%s)', targets.size, connected.size, dialed.length, unconnected);
+        return { dialed, unconnected };
     }
     /**
      * Bind a retained dial target's addresses (see {@link bootstrapPeerStore}) to
@@ -3951,28 +4178,32 @@ export class CadreNode {
     bootstrapDialAddrs(peerId, addrs) {
         return this.normalizeDialAddrs(this.parseMultiaddrs(addrs), peerId);
     }
-    /** Dial one cold-start bootstrap peer, best-effort. Returns whether it connected. */
-    async dialBootstrapPeer(peerId, addrs) {
+    /**
+     * Dial one peer from its retained dial target's addresses, best-effort: a cold-start bootstrap
+     * peer, or a just-claimed node ({@link redialClaimedNode}), which `context` names in the logs.
+     * Returns whether it connected.
+     */
+    async dialBootstrapPeer(peerId, addrs, context = 'reconcileControlCohort(cold-start)') {
         const controlNode = this.controlNode;
         if (!controlNode) {
             return false;
         }
         const parsed = this.bootstrapDialAddrs(peerId, addrs);
         if (parsed.length === 0) {
-            log('reconcileControlCohort(cold-start): no dialable address for bootstrap peer %s; skipping', peerId);
+            log('%s: no dialable address for bootstrap peer %s; skipping', context, peerId);
             return false;
         }
         try {
-            log('reconcileControlCohort(cold-start): dialing bootstrap peer %s (%d addr(s))', peerId, parsed.length);
-            await dialPeerAddrs(controlNode, peerId, parsed, this.controlDialBudget(), `reconcileControlCohort cold-start dial of bootstrap peer ${peerId}`);
+            log('%s: dialing bootstrap peer %s (%d addr(s))', context, peerId, parsed.length);
+            await dialPeerAddrs(controlNode, peerId, parsed, this.controlDialBudget(), `${context} dial of bootstrap peer ${peerId}`);
             return true;
         }
         catch (error) {
             if (error instanceof SelfRelayOnlyError) {
-                log('reconcileControlCohort(cold-start): bootstrap peer %s is reachable only by relaying through this node; waiting for it to reconnect', peerId);
+                log('%s: bootstrap peer %s is reachable only by relaying through this node; waiting for it to reconnect', context, peerId);
             }
             else {
-                log('reconcileControlCohort(cold-start): dial of bootstrap peer %s failed (continuing): %o', peerId, error);
+                log('%s: dial of bootstrap peer %s failed (continuing): %o', context, peerId, error);
             }
             return false;
         }
@@ -3982,31 +4213,33 @@ export class CadreNode {
     //
     // Optimystic's coordinator commits a control write WITHOUT broadcasting when the
     // block's cluster has ≤1 member. An owner that authorizes/removes a peer or
-    // (re)publishes its own record while no sibling is connected therefore writes a
+    // (re)publishes its own record while no sibling is reachable therefore writes a
     // row that exists ONLY in its local control DB — a sibling that connects later
     // never observes it (pull-on-read routes to the block's cluster, which never
     // learned of the write). reconcileControlCohort shrinks the alone window but
-    // cannot close it (a write in the instant before any sibling connects is still
-    // local-only). The remedy: detect "wrote while alone" (0 connections is a sound
-    // lower bound), queue the affected row, and RE-ISSUE the write once the cohort
-    // grows (0→≥1 connection), as an idempotent monotonic update that now broadcasts.
+    // cannot close it (a write in the instant before any sibling identifies is still
+    // local-only). The remedy: detect "wrote while alone" (no identified sibling the
+    // cohort would admit: committedAlone), queue the affected row, and RE-ISSUE the write
+    // once one is reachable, as an idempotent monotonic update that now broadcasts.
     // See docs/architecture.md (Control Network Convergence → "Writes made while alone").
     // ============================================================================
     /**
-     * Whether a control write happening now would commit local-only. A sound lower
-     * bound is "no connected control peers": 0 connections ⇒ the block's cluster is
-     * ≤1 ⇒ Optimystic commits without broadcasting. This is the pragmatic proxy for
-     * the precise signal (the block's `getClusterSize`); it over-approximates safely
-     * — a connected-but-not-in-this-block's-cluster write may still be local-only and
-     * is caught by the cohort's periodic pull-on-read instead. Re-issuing an already-
-     * replicated row is harmless (an idempotent monotonic bump), so the coarse proxy
-     * is acceptable as the agreed first cut.
+     * Whether a control write happening now would commit local-only: no sibling is reachable
+     * ({@link hasReachableControlSibling}, the public form, so the two cannot drift). A relay, a
+     * bootstrap node or a peer that has not identified yet holds a connection without being in the
+     * write's cohort, so a write made with only those connected is queued too. One made alone at
+     * the revision the cohort reached meanwhile is a fork no later consult sees, because a consult
+     * compares revision numbers only; the re-issue on the growth edge is what conflicts with the
+     * cohort's copy and pulls it in.
      *
-     * Reads {@link getControlConnectionCount} — the public, embedder-facing form of
-     * the same sample — so the two cannot drift.
+     * Still a proxy for the precise signal (the block's own cohort): a reachable sibling outside
+     * this block's cohort leaves the write local-only, caught by the cohort's pull-on-read. On the
+     * control network every block's cohort is the whole party, so that gap is a party larger than
+     * `CONTROL_REPLICATION_BREADTH`. Re-issuing a replicated row is harmless (an idempotent
+     * monotonic bump).
      */
     committedAlone() {
-        return this.getControlConnectionCount() === 0;
+        return !this.hasReachableControlSibling();
     }
     /**
      * Record (or clear) a just-committed owner membership write in the
@@ -4060,7 +4293,7 @@ export class CadreNode {
     /**
      * Drain the write-while-alone re-replication queue: re-issue the writes that
      * committed local-only now that the cohort can broadcast them. Public so the
-     * 0→≥1 growth trigger ({@link handleControlConnectionChange}) and tests can drive
+     * growth trigger ({@link updateControlSiblingReachability}) and tests can drive
      * it; concurrent calls collapse into one in-flight drain
      * ({@link drainControlReplicationInFlight}) so two growth signals never
      * double-issue. Best-effort throughout — a per-row failure leaves that entry
@@ -4098,13 +4331,15 @@ export class CadreNode {
         //    and receiving-node staleness is the pre-existing pull-on-read /
         //    periodic-refresh property.
         await this.drainPendingRevocations();
-        // 2. Self rows. On the first growth always re-touch (covers a self row carried
-        //    over from a prior process that committed local-only before this start);
-        //    afterwards only when a self-write-while-alone was recorded this session.
+        // 2. Self rows. A self write made while alone is always re-touched: the re-touch
+        //    is what carries it to the cohort. Otherwise the first growth re-touches a row
+        //    a prior process may have committed local-only, unless the row already holds a
+        //    current record this process signed (a redeemed invitation's admission).
         //    registerSelf is idempotent + single-flight, so this never races a
         //    duplicate publish with the heartbeat republish.
         if (firstGrowth || this.pendingSelfPeerWrite) {
-            await this.registerSelf().catch((error) => log('drain: registerSelf failed: %o', error));
+            const publish = this.pendingSelfPeerWrite ? this.registerSelf() : this.publishSelfRecordUnlessCurrent();
+            await publish.catch((error) => log('drain: registerSelf failed: %o', error));
             this.pendingSelfPeerWrite = false;
         }
         if (firstGrowth || this.pendingSelfDeviceWrite) {
@@ -4254,8 +4489,11 @@ export class CadreNode {
     }
     /**
      * Body of {@link reconstructAuthoredMembership}'s sweep: re-issue each candidate
-     * owner-authored row, returning how many were re-touched. A per-row failure is
-     * logged and skipped; a teardown mid-sweep stops it.
+     * owner-authored row, returning how many were re-touched. A row the sweep already holds as
+     * self-published is skipped without a call; the re-touch itself decides again over the row
+     * it reads under the lock, so a row that vanished or was self-published since the sweep's
+     * read is reported, not written. A per-row failure is logged and skipped; a teardown
+     * mid-sweep stops it.
      */
     async reissueAuthoredMembershipRows(selfPeerId, members) {
         let touched = 0;
@@ -4263,17 +4501,15 @@ export class CadreNode {
             if (!this._running || !this.controlNode || !this.controlDatabase) {
                 return touched;
             }
-            if (member.peerId === selfPeerId || this.pendingPeerWrites.has(member.peerId)) {
+            // Self, queued for the drain (step 4 re-touches it once), or peer-owned (self-Sig
+            // present) → not this sweep's to bump.
+            if (member.peerId === selfPeerId || this.pendingPeerWrites.has(member.peerId) || member.sig) {
                 continue;
             }
             try {
-                const record = await this.controlDatabase.queryPeerRecord(member.peerId);
-                // Missing (raced a delete) or peer-owned (self-Sig present) → not ours to bump.
-                if (!record || record.sig) {
-                    continue;
+                if (await this.reissuePeerAuthorize(member.peerId) === 'retouched') {
+                    touched++;
                 }
-                await this.reissuePeerAuthorize(member.peerId, record.updatedAt);
-                touched++;
             }
             catch (error) {
                 log('drain: reconstruction re-touch of %s failed (continuing): %o', member.peerId, error);
@@ -4307,9 +4543,10 @@ export class CadreNode {
         await this.deferMembershipGateRefresh('drain-reissue', () => this.reissuePendingPeerWrites());
     }
     /**
-     * Body of {@link drainPendingPeerWrites}' loop. An entry is cleared only on
-     * success (or when there is nothing left to re-issue); a failure is logged and
-     * the entry stays queued for the next cohort growth.
+     * Body of {@link drainPendingPeerWrites}' loop. An entry is cleared on every outcome the
+     * re-touch reports — re-touched, or nothing left to re-issue because the row was removed
+     * since or the peer self-published; the re-touch decides which over the row it reads under
+     * the lock. A failure is logged and the entry stays queued for the next cohort growth.
      */
     async reissuePendingPeerWrites() {
         for (const peerId of [...this.pendingPeerWrites.keys()]) {
@@ -4317,13 +4554,8 @@ export class CadreNode {
                 return;
             }
             try {
-                const record = await this.controlDatabase.queryPeerRecord(peerId);
-                if (!record || record.sig) {
-                    // Removed since, or peer self-published — nothing for the owner to re-issue.
-                    this.pendingPeerWrites.delete(peerId);
-                    continue;
-                }
-                await this.reissuePeerAuthorize(peerId, record.updatedAt);
+                const outcome = await this.reissuePeerAuthorize(peerId);
+                log('drain: re-issue of %s (authorize): %s', peerId, outcome);
                 this.pendingPeerWrites.delete(peerId);
             }
             catch (error) {
@@ -4336,57 +4568,196 @@ export class CadreNode {
      * `UpdatedAt` strictly above the stored value (and the wall clock) and re-sign via
      * the owner branch of `CadrePeer.AuthorizedUpdate`. The row already exists
      * locally (it committed there, just local-only), so this is an UPDATE, not the
-     * original INSERT.
+     * original INSERT. The bump, and whether the row is still the owner's to re-issue, are
+     * decided over the row as read inside the write ({@link SeedBootstrapService.reauthorizePeer}'s
+     * re-broadcast mode).
      */
-    async reissuePeerAuthorize(peerId, currentUpdatedAt) {
+    async reissuePeerAuthorize(peerId) {
         if (!this.seedBootstrapService) {
-            return;
+            throw new Error('Seed bootstrap service not initialized. Call initializeSeedBootstrap() first.');
         }
-        const updatedAt = Math.max(Date.now(), (currentUpdatedAt ?? 0) + 1);
-        await this.seedBootstrapService.reauthorizePeer(peerId, updatedAt);
+        return await this.seedBootstrapService.reauthorizePeer(peerId, 'rebroadcast');
     }
     /**
-     * On a control connection opening, detect the 0→≥1 transition and drain the
-     * write-while-alone re-replication queue (single-flight; fires only on the edge,
-     * not on every subsequent `connection:open`). Best-effort — a drain failure is
-     * logged, not thrown.
+     * Re-evaluate {@link hasReachableControlSibling} and act on its edges: on false→true, drain
+     * the write-while-alone re-replication queue (single-flight; best-effort, a drain failure is
+     * logged, not thrown); on true→false, re-arm, so the next reachable sibling drains again.
+     * Called when a peer identifies, when a connection closes, and after every membership refresh,
+     * where a sibling that connected before this node authorized it (the production join order)
+     * becomes reachable. A device counted in for its admission ({@link admittingDevices}) is not
+     * a trigger: the admission's own writes are in flight, and its row landing refreshes
+     * membership.
      */
-    handleControlConnectionChange() {
+    updateControlSiblingReachability() {
         if (!this._running || !this.controlNode) {
             return;
         }
-        const connected = this.controlNode.getConnections().length > 0;
-        if (!connected || this.hasControlConnection) {
+        const reachable = this.hasReachableControlSibling();
+        if (reachable === this.controlSiblingReachable) {
             return;
         }
-        this.hasControlConnection = true;
-        log('Control cohort grew (0 → ≥1 connection); draining write-while-alone re-replication queue');
-        void this.drainPendingControlReplication('connection:open')
-            .catch((error) => log('Control re-replication drain (connection:open) failed: %o', error));
+        this.controlSiblingReachable = reachable;
+        if (!reachable) {
+            return;
+        }
+        log('Control cohort grew (a sibling is reachable); draining write-while-alone re-replication queue');
+        void this.drainPendingControlReplication('sibling-reachable')
+            .catch((error) => log('Control re-replication drain (sibling-reachable) failed: %o', error));
     }
-    /** Re-arm the growth edge once all control connections drop, so a later reconnect re-drains. */
-    handleControlConnectionClose() {
-        if (!this.controlNode) {
+    /**
+     * A connected peer identified as serving the control database: count it
+     * ({@link identifiedControlPeers}), re-evaluate the growth edge, and run the sibling-open
+     * reactions, once per connected period. libp2p re-emits `peer:identify` on every
+     * identify-push from a connected peer, and those repeats start nothing.
+     */
+    onControlPeerIdentified(peerId) {
+        const id = peerId.toString();
+        // An identify that reports after its connection closed counts nobody.
+        if (this.identifiedControlPeers.has(id) || !this.controlNode || this.controlNode.getConnections(peerId).length === 0) {
             return;
         }
-        if (this.controlNode.getConnections().length === 0) {
-            this.hasControlConnection = false;
+        this.identifiedControlPeers.add(id);
+        this.updateControlSiblingReachability();
+        this.onControlSiblingConnectionChange(id, 'open');
+    }
+    /**
+     * A control connection closed: `remotePeer` stops counting once it has none left, and the
+     * growth edge re-arms if no reachable sibling remains. The sibling-close reaction runs on every
+     * close. libp2p's connection manager drops the closed connection from `getConnections` before
+     * this listener runs: libp2p dispatches each event on its internal bus, where the connection
+     * manager listens, before re-dispatching it to the node's own listeners.
+     */
+    onControlConnectionClosed(remotePeer) {
+        const id = remotePeer.toString();
+        if (!this.controlNode || this.controlNode.getConnections(remotePeer).length === 0) {
+            this.identifiedControlPeers.delete(id);
         }
+        this.updateControlSiblingReachability();
+        this.onControlSiblingConnectionChange(id, 'close');
     }
     /**
-     * Wire the control-connection growth listeners that drive write-while-alone
-     * re-replication. Wired in {@link start} (not the delayed refresh path) so no
-     * early `connection:open` is missed; torn down in {@link stopRecordRefresh}.
-     * Idempotent.
+     * Count the peers that identified before {@link wireControlConnectionListeners} ran, whose
+     * `peer:identify` has already fired, from their peer-store records. Updates the growth edge
+     * but runs no sibling-open reaction, as a connection that predates the listeners never did. A
+     * peer not yet in the peer store is still identifying, and its own event counts it.
+     */
+    async noteIdentifiedControlPeers() {
+        const node = this.controlNode;
+        if (!node) {
+            return;
+        }
+        const connected = new Map();
+        for (const connection of node.getConnections()) {
+            connected.set(connection.remotePeer.toString(), connection.remotePeer);
+        }
+        const partyId = this.config.controlNetwork.partyId;
+        await Promise.all([...connected.values()].map(async (peerId) => {
+            try {
+                const { protocols } = await node.peerStore.get(peerId);
+                // Still wired to this node: a stop while the read was out cleared the set.
+                const wired = this.controlPeerIdentifyHandler !== null && this.controlNode === node;
+                if (wired && servesControlDatabase(partyId, protocols) && node.getConnections(peerId).length > 0) {
+                    this.identifiedControlPeers.add(peerId.toString());
+                }
+            }
+            catch (error) {
+                if (error.name !== 'NotFoundError') {
+                    log('Peer store read for connected control peer %s failed — left to its identify: %o', peerId.toString(), error);
+                }
+            }
+        }));
+        this.updateControlSiblingReachability();
+    }
+    /**
+     * An AUTHORIZED sibling became reachable (its identify, once per connected period:
+     * {@link onControlPeerIdentified}) or one of its connections closed: start a reconcile
+     * pass. On open the pass asks the new sibling for strand addresses and re-arms the
+     * delegate grants; on close it re-dials. First, on open, drop the membership rows
+     * views: a machine that was offline while a sibling committed learns of the commit
+     * only when it next reads the block from storage, and dropping the views makes the
+     * pass's reads do that now rather than after the views' maximum age. Those reads can
+     * fall inside the read-repair window, so the open also arms a second drop past it
+     * ({@link armMembershipViewReload}). Any other
+     * peer — a relay, a bootstrap node, another party's machine — starts nothing.
+     *
+     * On identify rather than `connection:open` because, until the sibling identifies,
+     * Optimystic's cohort for every control block is this node alone: a pass started at the
+     * open writes (the first self-record publish) and reads on a cohort of one.
+     */
+    onControlSiblingConnectionChange(remotePeer, change) {
+        if (!this.authorizedControlPeers.has(remotePeer)) {
+            return;
+        }
+        if (change === 'open') {
+            // NOTE: a machine removed from the cadre that has not learned of it keeps redialing
+            // its siblings, each of which admits it provisionally and hangs up at the deadline
+            // (28.5 s at the default link), and every such open drops the views here and once
+            // more past the read-repair window: two storage reloads, with their cohort consults
+            // refused, per cycle until it learns. If a removed machine's churn ever shows up,
+            // skip both drops for a peer whose previous connection the remote closed inside the
+            // provisional deadline.
+            this.controlDatabase?.invalidateMembershipViews();
+            this.armMembershipViewReload();
+        }
+        // NOTE: a close starts a pass at once, and a sibling in its bring-up quiet period lets the
+        // dial resolve and then drops the connection (`createMembershipConnectionGater`'s deny
+        // timing), so a restarting sibling costs one dial-connect-close pass after another for its
+        // whole bring-up (5-6 passes in ~200 ms against a claimed node restarting locally). If a
+        // slow device's bring-up ever makes that churn show up, delay the pass a close starts when
+        // the previous connection to that sibling lived only a few milliseconds.
+        log('Control connection to sibling %s %s; reconcile pass triggered', remotePeer, change);
+        this.controlReconcileRuns?.trigger(0);
+    }
+    /**
+     * Drop the membership views again {@link MEMBERSHIP_VIEW_RECONNECT_RELOAD_MS} after a
+     * sibling connects, and start a pass whose reads reload them. This node's reads before
+     * the connection stamped Optimystic's read-repair window on the control blocks, so the
+     * reload the open itself starts is served from the local copy with no cohort consult,
+     * and the view would keep rows that predate the sibling's commits until its maximum age.
+     * The reload past the window starts the consults; one that brings a newer revision
+     * raises the table change that drops the view and refreshes the gate. Re-armed by every
+     * open, so it runs one window after the latest.
+     *
+     * NOTE: a consult Optimystic defers (16 already in flight) starts nothing and leaves the block
+     * unstamped, so the view keeps its rows until the first read past its maximum age, which on an
+     * idle node is the reconcile backstop. If `cluster-tx:read-repair-deferred` lines show at a
+     * reconnect, re-arm this reload once more instead of waiting for the backstop.
+     */
+    armMembershipViewReload() {
+        if (this.membershipViewReloadTimer) {
+            clearTimeout(this.membershipViewReloadTimer);
+        }
+        this.membershipViewReloadTimer = setTimeout(() => {
+            this.membershipViewReloadTimer = null;
+            log('Read-repair window passed since a sibling connected; reloading the membership views');
+            this.controlDatabase?.invalidateMembershipViews();
+            this.controlReconcileRuns?.trigger(0);
+        }, MEMBERSHIP_VIEW_RECONNECT_RELOAD_MS);
+        // Node-only: don't keep the event loop alive solely for this timer.
+        this.membershipViewReloadTimer?.unref?.();
+    }
+    /**
+     * Wire the control-connection listeners, which keep {@link identifiedControlPeers}: the
+     * growth edge that drives write-while-alone re-replication, and the sibling reconcile
+     * trigger ({@link onControlSiblingConnectionChange}). Wired in {@link start} (not the
+     * delayed refresh path) so no early identify is missed, and reads back the peers that
+     * identified before it ran ({@link noteIdentifiedControlPeers}); torn down in
+     * {@link stopRecordRefresh}. Idempotent.
      */
     wireControlConnectionListeners() {
-        if (!this.controlNode || this.controlConnectionOpenHandler) {
+        if (!this.controlNode || this.controlPeerIdentifyHandler) {
             return;
         }
-        this.controlConnectionOpenHandler = () => this.handleControlConnectionChange();
-        this.controlConnectionCloseHandler = () => this.handleControlConnectionClose();
-        this.controlNode.addEventListener('connection:open', this.controlConnectionOpenHandler);
+        const partyId = this.config.controlNetwork.partyId;
+        this.controlPeerIdentifyHandler = (event) => {
+            if (servesControlDatabase(partyId, event.detail.protocols)) {
+                this.onControlPeerIdentified(event.detail.peerId);
+            }
+        };
+        this.controlConnectionCloseHandler = (event) => this.onControlConnectionClosed(event.detail.remotePeer);
+        this.controlNode.addEventListener('peer:identify', this.controlPeerIdentifyHandler);
         this.controlNode.addEventListener('connection:close', this.controlConnectionCloseHandler);
+        void this.noteIdentifiedControlPeers();
     }
     // ============================================================================
     // Device-token registry (control-network push-token publish + resolve)
@@ -4692,6 +5063,7 @@ export class CadreNode {
             if (sAppId !== undefined) {
                 this.replicaSAppIds.set(strandId, sAppId);
                 log('Storage replica %s serves sApp %s', strandId, sAppId);
+                this.strandWatcher?.sAppIdLearned(strandId);
             }
         }
         catch (error) {
@@ -4800,6 +5172,7 @@ export class CadreNode {
         this.strandLaunchRefusals.clear();
         this.discoveredStrands.clear();
         this.replicaSAppIds.clear();
+        this.formedPartyKeys.clear();
         // Drop delegate-admission state: the grants are scoped to the session that
         // recorded them, so a stop()/start() cycle on this object must not keep
         // admitting delegates announced under the previous one, nor carry stale
@@ -4820,6 +5193,10 @@ export class CadreNode {
         // Detach the membership hub first: nothing this teardown does should drive a
         // gate refresh, and a notification arriving after this point is a no-op.
         if (this.controlDatabase) {
+            // The database's registry would dispose the gate watch too; disposing it here, before
+            // the listener goes, keeps a closing commit from scheduling a gate refresh.
+            this.membershipGateWatch?.dispose();
+            this.membershipGateWatch = null;
             this.controlDatabase.setMembershipChangeListener(null);
             this.controlDatabase.setGuardedDeleteListener(null);
             this.controlDatabase.setControlWriteAbandonedListener(null);
@@ -5157,8 +5534,9 @@ export class CadreNode {
      * registered, deliberately: both an explicit retry and the {@link StrandWatcher}'s
      * automatic relaunch need it. A failed launch here hands the strand back to that
      * relaunch — `StrandWatcher.forgetStrand` drops the id from the watcher's
-     * `knownStrands` and records the backoff, so a later poll re-attempts it on the same
-     * `pollInterval * 2^(failures-1)` ladder a watcher-driven failure gets. Because the
+     * `knownStrands`, records the backoff and wakes the watcher when it expires, so the
+     * strand is re-attempted on the same ladder a watcher-driven failure gets
+     * (`min(strandWatchInterval, 5 s) * 2^(failures-1)`, capped at 5 minutes). Because the
      * config stays registered, the retry takes `handleStrandAdded`'s auto-launch branch:
      * what the app sees is `strand:error` per failed retry and `strand:started` when one
      * succeeds, never a second `strand:discovered`. {@link detachStrand} (reached via
@@ -5196,6 +5574,7 @@ export class CadreNode {
         // A claim also overrides any earlier `stopStrand` of this id: the watcher must track
         // the strand again, or a party-wide removal would never stop it here.
         this.strandWatcher?.unsuppressStrand(strandRow.Id);
+        this.strandWatcher?.sAppIdLearned(strandRow.Id);
         log('Registered sAppConfig for strand %s (sApp: %s, founder: %s)', strandRow.Id, sAppConfig.id, founder ?? 'derived');
         // Before the launch, so a foreign strand whose launch fails, or whose app is killed
         // mid-launch, is still offered again: by the watcher's retry and by the next start.
@@ -5219,7 +5598,7 @@ export class CadreNode {
             // Hand the strand back to the watcher's retry ladder. Harmless on the founder path
             // (`foundStrand` → `publishStrand` + this call), where the watcher may never have
             // offered the strand: the `knownStrands` delete is a no-op and the recorded backoff
-            // only delays the watcher's own first attempt by one poll interval, which is what
+            // only delays the watcher's own first attempt by one ladder step, which is what
             // should happen right after an attempt that just failed.
             //
             // NOTE: the retry relaunches from the CONTROL row (or, for a strand joined from
@@ -5300,7 +5679,8 @@ export class CadreNode {
      * (the key exists nowhere else). No app-data wipe is needed for either.
      *
      * The insert is signed through this node's owner signer ({@link getOwnerSigner}), in one
-     * approval with a closed strand's party-key mint. The signer's key is persisted on the row
+     * approval and one control transaction with a closed strand's party key
+     * ({@link ControlDatabase.insertFoundingRows}). The signer's key is persisted on the row
      * as `FounderOwnerKey` (the schema pins the column to the verified signer), and this node's
      * peer id as `FounderPeerId` (bound into the signature), which is what later lets any
      * launch derive "this machine is the founder" from the row alone.
@@ -5311,7 +5691,8 @@ export class CadreNode {
      *
      * A CLOSED strand's publish also seats this party's own membership identity key —
      * the `StrandPartyKey` row {@link ensureStrandPartyKey} documents — which is what the
-     * founder bootstrap derives `Member.Key`/`Manager.MemberKey` from. The
+     * founder bootstrap derives `Member.Key`/`Manager.MemberKey` from; a key the party already
+     * holds for the id is kept. The
      * `memberPrivateKey` ARGUMENT stays the strand-wide read secret formation hands to
      * joiners; it derives nobody's identity.
      *
@@ -5331,23 +5712,8 @@ export class CadreNode {
      */
     async publishStrand(strandId, type = 'o', memberPrivateKey, sign) {
         const signer = this.requireOwnerSigner(`publish strand ${strandId}`);
-        // Trim/reject here so the id that lands matches the one unpublishStrand looks up: it
-        // trims too, and an untrimmed row would be unreachable by the same string.
-        const trimmed = requireNonBlank(strandId, 'strand id');
-        // Refuse an unusable id HERE, not only at launch. `addStrand` already asserts, but it
-        // runs after this row is written, and this row replicates: an id no node can turn into
-        // a storage scope key would otherwise reach the whole party's control database and be
-        // declined once per member. Asserted after the trim so the id checked is the id stored.
-        assertStrandScopeKey(trimmed);
-        // FounderPeerId records THIS machine as the founder, which is what later lets a relaunch
-        // derive founder-ness; ControlDatabase.insertStrand writes the same id (its own node's).
-        const desired = {
-            Id: trimmed,
-            Type: type,
-            MemberPrivateKey: memberPrivateKey ?? null,
-            FounderOwnerKey: signer.publicKey,
-            FounderPeerId: this.controlNode.peerId.toString(),
-        };
+        const desired = this.founderStrandRow(strandId, type, memberPrivateKey, signer);
+        const trimmed = desired.Id;
         // Read first rather than leaning on the insert's collision for BOTH readings. The
         // ordinary resume (an app killed after its own publish committed) then needs no error
         // matching at all: the collision path can only recognise itself by the engine's message
@@ -5367,24 +5733,57 @@ export class CadreNode {
             log('publishStrand(%s): already published (type %s) with identical content — no-op', trimmed, type);
             // A closed strand THIS machine published must hold a party key (its own membership
             // identity — see ensureStrandPartyKey). Insert-if-absent, so the ordinary repeat
-            // is a read-only no-op and a publish interrupted between the row insert and the
-            // mint heals here. A row another machine published is that machine's to mint for.
+            // is a read-only no-op and a closed row stored without its key heals here. A row
+            // another machine published is that machine's to mint for.
             if (type === 'c' && this.isSelfFoundedRow(matched)) {
                 await this.ensureStrandPartyKey(trimmed, undefined, sign);
             }
             return matched;
         }
-        return await withinOwnerApproval(signer, { partyId: this.partyId, action: `Publish the strand ${trimmed} to the cadre` }, sign, (approved) => this.insertStrandRow(desired, signer, approved));
+        return (await this.writeFoundingRows(desired, undefined, signer, sign)).strandRow;
+    }
+    /**
+     * The `Strand` row this machine publishes for `strandId`, as {@link publishStrand} and
+     * {@link foundStrand} write it: the id trimmed and checked, this node's owner key and peer id
+     * as its founder.
+     */
+    founderStrandRow(strandId, type, memberPrivateKey, signer) {
+        // Trim/reject here so the id that lands matches the one unpublishStrand looks up: it
+        // trims too, and an untrimmed row would be unreachable by the same string.
+        const trimmed = requireNonBlank(strandId, 'strand id');
+        // Refuse an unusable id HERE, not only at launch. `addStrand` already asserts, but it
+        // runs after this row is written, and this row replicates: an id no node can turn into
+        // a storage scope key would otherwise reach the whole party's control database and be
+        // declined once per member. Asserted after the trim so the id checked is the id stored.
+        assertStrandScopeKey(trimmed);
+        // FounderPeerId records THIS machine as the founder, which is what later lets a relaunch
+        // derive founder-ness; ControlDatabase.insertFoundingRows writes the same id (its own node's).
+        return {
+            Id: trimmed,
+            Type: type,
+            MemberPrivateKey: memberPrivateKey ?? null,
+            FounderOwnerKey: signer.publicKey,
+            FounderPeerId: this.controlNode.peerId.toString(),
+        };
     }
     /**
-     * The write half of {@link publishStrand}: insert `desired`, adopting a matching row a
-     * concurrent founder landed first, then mint a closed strand's party key, all under one
-     * owner approval's `sign`.
+     * The write half of {@link publishStrand} and {@link foundStrand}: under one owner approval,
+     * insert `desired`, a closed strand's party key and the requested invitation in one control
+     * transaction ({@link ControlDatabase.insertFoundingRows}), adopting a matching row a
+     * concurrent founder landed first. Reads nothing first: the caller has just read the
+     * `Strand` row, and a stored key or row is resolved from the insert's own rejection.
      */
-    async insertStrandRow(desired, signer, sign) {
-        const { Id: trimmed, Type: type } = desired;
+    async writeFoundingRows(desired, invite, signer, sign) {
+        const action = invite
+            ? `Publish the strand ${desired.Id} to the cadre, with an invitation to the app ${invite.sAppId}`
+            : `Publish the strand ${desired.Id} to the cadre`;
+        return await withinOwnerApproval(signer, { partyId: this.partyId, action }, sign, (approved) => this.insertOrAdoptFounding(desired, invite, signer, approved));
+    }
+    /** {@link writeFoundingRows} inside its approval. */
+    async insertOrAdoptFounding(desired, invite, signer, sign) {
+        let partyKey;
         try {
-            await this.controlDatabase.insertStrand(trimmed, type, signer.publicKey, sign, desired.MemberPrivateKey ?? undefined);
+            partyKey = await this.seatFoundingRows(desired, invite, signer, sign);
         }
         catch (error) {
             // Only the uniqueness collision is a candidate for the idempotent reading; every
@@ -5392,26 +5791,79 @@ export class CadreNode {
             if (!isStrandIdConflict(error, 'Strand')) {
                 throw error;
             }
-            const landed = await this.controlDatabase.queryStrand(trimmed);
-            if (!landed) {
-                // Collided, yet no row is readable: not the race this branch handles (a row
-                // deleted between the rejection and the re-read, or a stale local view). Surface
-                // the original rather than guessing.
+            const strandRow = await this.adoptConcurrentFounding(desired, error);
+            // The rejection rolled the invitation back with the row; a caller asking for one gets one.
+            if (invite) {
+                await this.publishFoundingInviteIfAbsent(strandRow.Id, invite, sign);
+            }
+            return { strandRow };
+        }
+        log('Published strand %s (type %s) to control DB under owner %s', desired.Id, desired.Type, signer.publicKey);
+        if (invite) {
+            this.registerPublishedInvitation(invite.token, invite.expiresAtMs);
+        }
+        return { strandRow: desired, partyKey };
+    }
+    /**
+     * Insert `desired`'s founding rows, minting a closed strand's party key — this party's own
+     * membership identity, which the founder bootstrap derives `Member`/`Manager` from, distinct
+     * from the row's shared `MemberPrivateKey` read secret. A key already stored for the id with
+     * no strand row (one {@link ensureStrandPartyKey} seated, which it would adopt too) refuses
+     * that transaction whole; the rows are then written again without a key.
+     *
+     * @returns The party key this seated; undefined when it seated none, so the launch reads the stored one.
+     */
+    async seatFoundingRows(desired, invite, signer, sign) {
+        const controlDatabase = this.controlDatabase;
+        const strand = { id: desired.Id, type: desired.Type, memberPrivateKey: desired.MemberPrivateKey ?? undefined };
+        if (desired.Type !== 'c') {
+            await controlDatabase.insertFoundingRows({ strand, formationInvite: invite }, signer.publicKey, sign);
+            return undefined;
+        }
+        const partyKey = await generateStrandMemberKey();
+        try {
+            await controlDatabase.insertFoundingRows({ strand, partyKey, formationInvite: invite }, signer.publicKey, sign);
+            return partyKey;
+        }
+        catch (error) {
+            if (!isStrandIdConflict(error, 'StrandPartyKey')) {
                 throw error;
             }
-            log('publishStrand(%s): lost a concurrent founding race; re-read the landed row', trimmed);
-            return requireMatchingStrandRow(landed, desired, 'landed concurrently from another founder');
+            log('publishStrand(%s): this party already holds a key for the id; publishing the row and adopting that key', desired.Id);
+            await controlDatabase.insertFoundingRows({ strand, formationInvite: invite }, signer.publicKey, sign);
+            return undefined;
         }
-        log('Published strand %s (type %s) to control DB under owner %s', trimmed, type, signer.publicKey);
-        // A closed strand's publish also mints this party's own membership identity key
-        // (StrandPartyKey) — the key the founder bootstrap derives Member/Manager from,
-        // deliberately distinct from the row's shared MemberPrivateKey read secret. A failure
-        // here surfaces (the strand row has already committed, so a retried publish takes
-        // the idempotent branch and heals the mint; so does the next founder launch).
-        if (type === 'c') {
-            await this.ensureStrandPartyKey(trimmed, undefined, sign);
+    }
+    /**
+     * Resolve a founding refused on `Strand.Id`: re-read the row that landed and adopt it when it
+     * matches `desired` — a concurrent founder, such as two machines of one party on one id —
+     * or throw naming the columns that differ.
+     */
+    async adoptConcurrentFounding(desired, error) {
+        const landed = await this.controlDatabase.queryStrand(desired.Id);
+        if (!landed) {
+            // Collided, yet no row is readable: not the race this branch handles (a row
+            // deleted between the rejection and the re-read, or a stale local view). Surface
+            // the original rather than guessing.
+            throw error;
         }
-        return desired;
+        log('publishStrand(%s): lost a concurrent founding race; re-read the landed row', desired.Id);
+        return requireMatchingStrandRow(landed, desired, 'landed concurrently from another founder');
+    }
+    /**
+     * Publish a founding's invitation on its own, bound to `strandId`, for a founding that adopted
+     * a stored row rather than writing its own — unless the invitation is already stored, bound to
+     * this strand, as a repeat of a founding whose transaction committed leaves it.
+     *
+     * @param sign - The open approval's `sign`; omitted, the publish asks for its own.
+     */
+    async publishFoundingInviteIfAbsent(strandId, invite, sign) {
+        if ((await this.controlDatabase.queryFormationInvite(invite.token))?.strandId === strandId) {
+            log('foundStrand(%s): the invitation is already published', strandId);
+            return;
+        }
+        const { token, sAppId, ...options } = invite;
+        await this.publishFormationInvite(token, sAppId, { ...options, strandId }, sign);
     }
     async foundStrand(config, options) {
         const tTotal = performance.now();
@@ -5438,17 +5890,18 @@ export class CadreNode {
      * so both forms of `foundStrand` reject on it before returning.
      */
     async publishFoundingRow(config) {
-        const { strandId, type = 'o', memberPrivateKey } = config;
+        const { strandId, type = 'o', memberPrivateKey, formationInvite, sAppConfig } = config;
         if (!this._running || !this.controlDatabase) {
             throw new Error(`CadreNode must be started before attempting to found strand ${strandId}`);
         }
         const controlDatabase = this.controlDatabase;
         const trimmed = requireNonBlank(strandId, 'strand id');
+        const invite = formationInvite && { ...formationInvite, sAppId: sAppConfig.id };
         const timed = (step, op) => timedStep('foundStrand', trimmed, step, op);
         const published = await timed('queryStrand', () => controlDatabase.queryStrand(trimmed));
-        const strandRow = published
-            ? this.adoptPublishedStrand(published, type, memberPrivateKey)
-            : await timed('publishStrand', () => this.publishStrand(trimmed, type, memberPrivateKey));
+        const { strandRow, partyKey } = published
+            ? await this.resumeFounding(published, type, memberPrivateKey, invite)
+            : await timed('publishStrand', () => this.publishFoundingStrand(trimmed, type, memberPrivateKey, invite));
         // Derived, not hardcoded `true`: adopting a row another machine published means that
         // machine runs the bootstrap — this one must attach or it would write a second Header.
         const founded = this.isSelfFoundedRow(strandRow);
@@ -5457,15 +5910,34 @@ export class CadreNode {
                 '(FounderPeerId is not this node\'s peer id) — attaching as a joiner; ' +
                 'the founder bootstrap runs on that machine', trimmed);
         }
-        return { strandRow, founded };
+        return { strandRow, founded, partyKey };
+    }
+    /** {@link publishFoundingRow} when no row is stored: write the founding rows under this node's owner. */
+    publishFoundingStrand(strandId, type, memberPrivateKey, invite) {
+        const signer = this.requireOwnerSigner(`publish strand ${strandId}`);
+        return this.writeFoundingRows(this.founderStrandRow(strandId, type, memberPrivateKey, signer), invite, signer);
+    }
+    /**
+     * {@link publishFoundingRow} over a stored row: adopt it ({@link adoptPublishedStrand}), then
+     * publish the requested invitation on its own, under its own approval.
+     */
+    async resumeFounding(published, type, memberPrivateKey, invite) {
+        const strandRow = this.adoptPublishedStrand(published, type, memberPrivateKey);
+        if (invite) {
+            await timedStep('foundStrand', strandRow.Id, 'publishFormationInvite', () => this.publishFoundingInviteIfAbsent(strandRow.Id, invite));
+        }
+        return { strandRow };
     }
     /**
      * The start half of {@link foundStrand}. Its timing lines (`addStrand: start`, then the
      * duration or `failed after`) are logged when the start settles, also in the background
      * form, so a start that hangs still shows as a start line with no end line.
+     *
+     * A party key the publish half just seated is handed to the launch, which would otherwise
+     * read back the row the same founding wrote.
      */
-    attachFoundingRow({ strandRow, founded }, { sAppConfig, awaitFirstSync }) {
-        return timedStep('foundStrand', strandRow.Id, 'addStrand', () => this.addStrand({ strandRow, sAppConfig, founder: founded, awaitFirstSync }));
+    attachFoundingRow({ strandRow, founded, partyKey }, { sAppConfig, awaitFirstSync }) {
+        return timedStep('foundStrand', strandRow.Id, 'addStrand', () => this.addStrand({ strandRow, sAppConfig, founder: founded, awaitFirstSync, partyMemberPrivateKey: partyKey }));
     }
     /**
      * Resume onto an already-published `Strand` row: keep its stored content, reject a type
@@ -5494,9 +5966,10 @@ export class CadreNode {
      * inverse of {@link publishStrand}.
      *
      * Party-wide, unlike {@link stopStrand}, which only stops the strand on THIS node: every
-     * cadre node watching the table sees the row vanish on its next poll (default 5 s) and
-     * stops its own instance. This node converges immediately — the method forces a watcher
-     * poll and stops any still-running local instance before resolving. Removing the row
+     * cadre node watching the table hears the row's removal and stops its own instance (a
+     * machine outside the `Strand` block's cohort, on a cadre larger than 16, only at its
+     * watcher's backstop read, default 5 minutes). This node converges immediately — the
+     * method forces a watcher run and stops any still-running local instance before resolving. Removing the row
      * removes OUR party's participation only: other parties in the strand keep their own
      * rows and the strand network carries on, so no cross-party sign-off is involved — this
      * is an owner-signed control-plane write like every other, not "destroy the network".
@@ -5582,16 +6055,16 @@ export class CadreNode {
         // deadline rather than dropping it (a binding cleared late is still worth more than
         // one never cleared).
         await this.strandManager.clearOwnMemberPeerBinding(trimmed);
-        // Converge locally now rather than waiting up to a poll interval. The watcher fires
-        // onStrandRemoved for a strand it tracked; the explicit teardown below covers a node
-        // whose strandFilter never admitted this strand (the watcher never knew it, so it will
-        // never fire). StrandInstanceManager.stopStrand no-ops on an unknown id, so the two
+        // Converge locally before resolving rather than when the write wakes the watcher. The
+        // watcher fires onStrandRemoved for a strand it tracked; the explicit teardown below
+        // covers a node whose strandFilter never admitted this strand (the watcher never knew
+        // it, so it will never fire). StrandInstanceManager.stopStrand no-ops on an unknown id, so the two
         // paths cannot double-stop into an error.
         //
         // detachStrand, NOT stopStrand: this is a party-wide removal, not a local abandonment
         // of a strand that still exists, so it must not SUPPRESS the id in the watcher. The
         // row is already gone — there is nothing to suppress — and an id re-published before
-        // the next poll's suppression cleanup would then never be offered again this session.
+        // the next run's suppression cleanup would then never be offered again this session.
         // (Same reasoning as handleStrandRemoved, which also avoids stopStrand's `_running`
         // guard; here the guard is already satisfied by requireOwnerSigner above.)
         await this.strandWatcher?.forcePoll();
@@ -5614,9 +6087,11 @@ export class CadreNode {
      * as `MemberPrivateKey`; docs/strands.md → "Closed-Strand Member Key Handling") and is
      * never put on the formation wire.
      *
-     * Called internally by {@link publishStrand} (closed strands mint at publish) and by a
-     * founder launch that finds no row (a publish interrupted before its mint); public so a
-     * test harness — or a joiner flow that persists a formation-issued identity — can seat
+     * A founding seats the key in the same transaction as the strand row, not through here.
+     * Called internally by {@link publishStrand}'s repeat of a published closed strand and by a
+     * founder launch that finds no row, which heal a closed strand row stored without its key
+     * (one seated by `ControlDatabase.insertStrand` directly, or whose key was removed); public
+     * so a test harness — or a joiner flow that persists a formation-issued identity — can seat
      * a specific key deliberately.
      *
      * @param strandId - The strand the key is this party's identity for.
@@ -5664,8 +6139,8 @@ export class CadreNode {
      * Resolve the party membership key a closed strand's launch threads into the founder
      * bootstrap: the explicit attach-time key when given, else the party's persisted
      * `StrandPartyKey` row, else — on the one machine the row names as founder (`FounderPeerId`)
-     * — a freshly minted-and-persisted key, which seats the identity for a publish that was
-     * interrupted before its mint. Everyone else resolves undefined: non-founding
+     * — a freshly minted-and-persisted key, which seats the identity for a closed row stored
+     * without one (see {@link ensureStrandPartyKey}). Everyone else resolves undefined: non-founding
      * machines never mint (no mint race between a party's machines), and only a founder
      * bootstrap needs the key at all.
      *
@@ -5685,6 +6160,11 @@ export class CadreNode {
         if (explicitKey !== undefined) {
             return explicitKey;
         }
+        const formed = this.formedPartyKeys.get(strand.Id);
+        if (formed !== undefined) {
+            this.formedPartyKeys.delete(strand.Id);
+            return formed;
+        }
         if (!this.controlDatabase) {
             return undefined;
         }
@@ -5708,8 +6188,8 @@ export class CadreNode {
      * Counterpart to {@link createOpenInvitation}, which only mints the
      * out-of-band {@link OpenInvitation} envelope: persisting the matching
      * `FormationInvite` row is what makes the token *redeemable* — the host's
-     * {@link ControlFormationUsageRecorder} answers `isTokenValid`/`isTokenUsed`
-     * from this row. A host minting a closed-strand invite does both (mint +
+     * {@link ControlFormationUsageRecorder} reads it (`readToken`) to check a
+     * redemption. A host minting a closed-strand invite does both (mint +
      * publish), exactly as the integration harness's `createInvitation` does.
      *
      * Signs through the same owner signer as {@link publishStrand}, whose key must be an
@@ -5726,12 +6206,19 @@ export class CadreNode {
     async publishFormationInvite(token, sAppId, options = {}, sign) {
         const signer = this.requireOwnerSigner(`publish formation invite ${token}`);
         await withinOwnerApproval(signer, { partyId: this.partyId, action: `Publish an invitation to the app ${sAppId}` }, sign, (approved) => this.controlDatabase.insertFormationInvite(token, sAppId, signer.publicKey, approved, options));
-        // A host may publish an invite whose token was minted elsewhere; registering
-        // it locally opens the connection gate's formation exemption immediately
-        // rather than waiting for the durable row to become readable.
-        this.strandSolicitationService?.registerMintedInvitation(token, options.expiresAtMs ?? Number.POSITIVE_INFINITY);
+        this.registerPublishedInvitation(token, options.expiresAtMs);
         log('Published formation invite %s (sApp %s) under owner %s', token, sAppId, signer.publicKey);
     }
+    /**
+     * Register an invitation this node just published with the strand solicitation service. A
+     * host may publish one whose token was minted elsewhere; registering it locally opens the
+     * connection gate's formation exemption immediately rather than waiting for the durable row
+     * to become readable. A token {@link createOpenInvitation} minted here is already registered,
+     * and this records the same expiry again.
+     */
+    registerPublishedInvitation(token, expiresAtMs) {
+        this.strandSolicitationService?.registerMintedInvitation(token, expiresAtMs ?? Number.POSITIVE_INFINITY);
+    }
     /**
      * Enroll an approver public key allowed to sign off on `ValidationUrl` redemptions.
      *
@@ -5839,7 +6326,7 @@ export class CadreNode {
      * fan-out, so a rediscovery costs nothing beyond the map lookup plus the (pure)
      * founder derivation. NOT a silent no-op any more when the launch resolves as a
      * FOUNDER: the tracked instance may have been launched first as a joiner (an app
-     * attach, or this node's own watcher poll winning the `resolveCohortSeed` window
+     * attach, or this node's own watcher run winning the `resolveCohortSeed` window
      * below), which used to drop the founder request and leave the strand headerless —
      * now the tracked instance is founded in place
      * ({@link StrandInstanceManager.foundExistingStrand}), waking it first if quiesced
@@ -5851,9 +6338,9 @@ export class CadreNode {
      *
      * An absent `sAppConfig` launches a storage replica ({@link CadreNodeConfig.hostUnclaimedStrands}),
      * which is ALWAYS a joiner whatever the row says: the founder bootstrap writes the sApp
-     * into `Strand.Header`, and a replica has none. So a self-founded row whose watcher poll
+     * into `Strand.Header`, and a replica has none. So a self-founded row whose watcher run
      * wins the race against its app's `addStrand` after a restart comes up as a replica — and
-     * so does a `foundStrand` whose publish a watcher poll saw before its attach. A claim
+     * so does a `foundStrand` whose publish a watcher run saw before its attach. A claim
      * that finds a tracked replica upgrades it in place, over the same node and store
      * ({@link StrandInstanceManager.attachSApp}), BEFORE any founding: the founder bootstrap
      * needs the sApp the attach supplies.
@@ -5885,8 +6372,8 @@ export class CadreNode {
         }
         // A closed strand's launch carries the party's OWN membership identity key: the
         // explicit attach-time key, else the control-layer StrandPartyKey row — minted here
-        // when this machine is the row-derived founder and no row exists yet (a publish
-        // interrupted before its mint). A joiner with no persisted key threads undefined;
+        // when this machine is the row-derived founder and no row exists yet (a closed row
+        // stored without its key). A joiner with no persisted key threads undefined;
         // only a FOUNDER bootstrap needs the key, and that path throws loudly without one
         // (StrandDatabase).
         const partyMemberPrivateKey = strand.Type === 'c'
@@ -6208,7 +6695,7 @@ export class CadreNode {
      *
      * Timestamps are recorded OPTIMISTICALLY, whatever `collectStrandAddrs` reports
      * per peer: recording only on success would re-announce to an unreachable relay
-     * on every reconcile tick, one dial timeout each. A failed INITIAL announce costs
+     * on every reconcile pass, one dial timeout each. A failed INITIAL announce costs
      * the strand supervisor its first attempt only (the relay denies the
      * reservation; every re-drive re-announces first through
      * {@link announceDelegateToRelay}); a failed REFRESH retries within
@@ -6331,9 +6818,11 @@ export class CadreNode {
      *
      * Due times are per (sibling, strand) and set from each sibling's own outcome, so
      * a sibling that connects late — a phone joining after the party's always-on
-     * machines — is asked on the next tick, and one that could not answer or refused
-     * (its view of the membership may not include us yet) is retried within
-     * {@link STRAND_PEER_ADDR_RETRY_MS} rather than {@link STRAND_PEER_ADDR_REFRESH_MS}.
+     * machines — is asked on the pass its connection starts, and one that could not
+     * answer or refused (its view of the membership may not include us yet) is retried
+     * within {@link STRAND_PEER_ADDR_RETRY_MS} rather than
+     * {@link STRAND_PEER_ADDR_REFRESH_MS}; the pass that stamped the retry arms a wake
+     * for it ({@link armReconcileWake}).
      *
      * Distinct from {@link refreshDelegateGrants}, deliberately: that pass covers
      * RELAYS on a `DELEGATE_GRANT_TTL_MS / 2` throttle to keep circuit-relay
@@ -6382,15 +6871,17 @@ export class CadreNode {
      *
      * `connectedSiblingTargets`' membership read is unbounded, and with zero
      * connections its answer is empty whatever the table holds, so decide it from the
-     * connection list instead of paying for the read once a tick.
+     * connection list instead of paying for the read once a pass.
      *
-     * NOTE: otherwise the read runs on EVERY tick with a running strand, not only when
+     * NOTE: otherwise the read runs on EVERY pass with a running strand, not only when
      * some (sibling, strand) is due: pruning a departed sibling needs the current
      * target set, and whether anyone is due cannot be told without it (a connected
-     * non-member never gets a due time to compare). One more `CadrePeer` read per 15 s
-     * tick, on top of the two `runReconcileControlCohort` already makes; if those reads
-     * get costly, share one row-set across the pass (see the NOTE there) rather than
-     * skipping ticks here.
+     * non-member never gets a due time to compare). One more `CadrePeer` read per
+     * pass, on top of the two `runReconcileControlCohort` already makes, all served
+     * from the membership rows view: an in-memory scan while the view is current,
+     * storage only after the table changed or the view's maximum age passed. Sharing
+     * one row set across the pass would save no storage read and was declined (see
+     * the NOTE there).
      *
      * NOTE: one strand-addr RPC per (running strand × due sibling) — each a tiny
      * request/response on an already-open control connection. If a node ever runs
@@ -6473,8 +6964,8 @@ export class CadreNode {
      *
      * NOTE: every pass serialises the whole FRET table (`exportTable()`), verifies one
      * signature per record and makes one `peerStore.merge` per peer, per running strand
-     * per 15 s tick. Fine at strand table sizes, and an unchanged address set writes
-     * nothing in libp2p's persistent peer store. If many strands or large tables make it
+     * per pass — every change-driven pass and the 2-minute backstop. Fine at strand table
+     * sizes, and an unchanged address set writes nothing in libp2p's persistent peer store. If many strands or large tables make it
      * show up, read only the entries whose address stamp is near the one-hour mark.
      *
      * @returns the ids of the peers the table holds a usable record for.
@@ -6757,8 +7248,8 @@ export class CadreNode {
      * The stop + emit are skipped when the strand manager holds no instance for `strandId` —
      * e.g. a party owner that published a strand's row but never ran it locally, or an
      * explicit {@link stopStrand} for an id this node never started. `hibernationManager`
-     * untrack and the `sAppConfigs` / {@link strandLaunchRefusals} / {@link discoveredStrands}
-     * deletes stay unconditional (all no-ops when there is nothing to remove), so a launch
+     * untrack and the `sAppConfigs` / {@link strandLaunchRefusals} / {@link discoveredStrands} /
+     * {@link formedPartyKeys} deletes stay unconditional (all no-ops when there is nothing to remove), so a launch
      * that failed before an instance was ever tracked still gets its stray entries cleared.
      *
      * Dropping the {@link discoveredStrands} entry here covers both callers, and both
@@ -6775,6 +7266,7 @@ export class CadreNode {
         this.sAppConfigs.delete(strandId);
         this.strandLaunchRefusals.delete(strandId);
         this.discoveredStrands.delete(strandId);
+        this.formedPartyKeys.delete(strandId);
         if (!this.strandManager.hasStrand(strandId)) {
             log('detachStrand: strand %s not tracked locally — no-op (no stop, no strand:stopped)', strandId);
             return;
@@ -7141,20 +7633,35 @@ export class CadreNode {
         return this.controlDatabase;
     }
     /**
-     * Open control-network connections right now. A lower-bound proxy for replication
-     * reach: 0 connections ⇒ a control write commits local-only. The private
-     * `committedAlone` write-while-alone seam is defined in terms of this.
-     *
-     * It approximates the precise signal (the block's cluster size), and a caller that
-     * samples it AFTER a write samples a slightly wider window than `committedAlone`
-     * does inside the write itself. It exists so an embedder can warn an owner before a
-     * control write that this machine currently sees none of its siblings, and report
-     * after one that the write may not have travelled — a warning shown unconditionally
-     * is a warning people learn to ignore.
+     * Open control-network connections right now, of every kind: siblings, relays, bootstrap
+     * nodes and peers still identifying. A report of connectivity, not of replication reach: a
+     * control write can commit local-only with connections open, and whether one would is
+     * {@link hasReachableControlSibling}.
      */
     getControlConnectionCount() {
         return this.controlNode?.getConnections().length ?? 0;
     }
+    /**
+     * Whether a control write made now reaches another machine: a connected peer has identified
+     * as serving the control database and is one this node sends control writes to (an
+     * authorized member, or a device whose admission it is writing). Optimystic counts a peer
+     * into a control block's cohort on those same terms; with none, a write commits on this node
+     * alone and is queued for re-issue once one is reachable (docs/architecture.md → "Writes made
+     * while alone").
+     *
+     * It exists so an embedder can warn an owner before a control write that this machine
+     * currently reaches none of its siblings, and report after one that the write may not have
+     * travelled; a warning shown unconditionally is a warning people learn to ignore. A caller
+     * that samples it AFTER a write samples a slightly wider window than the write did.
+     */
+    hasReachableControlSibling() {
+        for (const peerId of this.identifiedControlPeers) {
+            if (this.isControlCohortMember(peerId)) {
+                return true;
+            }
+        }
+        return false;
+    }
     /**
      * Force a poll of the strand watcher (for testing)
      */
@@ -7231,7 +7738,8 @@ export class CadreNode {
         }), this.controlNode, this.controlDatabase);
         log('Owner signer wired (owner key %s)', signer.publicKey);
         // 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 machine not yet an owner; wiring a signer writes no table it watches, so without
+        // this it would wait for its backstop.
         this.pendingJoinRunner?.kick();
     }
     /**
@@ -7455,6 +7963,10 @@ export class CadreNode {
         if (usage === undefined || invite === undefined) {
             return false;
         }
+        // NOTE: accepted tradeoff — three ed25519 verifies per invitation-admitted row on every
+        // refresh, CPU only since the rows come from the membership views; a memo of verdicts would
+        // also have to follow the trusted-owner anchor, which syncOwnerAnchor rewrites. Revisit if
+        // the gate's admission deadline ever shows it.
         return verifyInvitationAdmission(this.partyId, row, usage, invite, key => this.trustedOwnerStore?.has(key) ?? false);
     }
     /**
@@ -7477,7 +7989,7 @@ export class CadreNode {
      * builds its snapshot against that anchor, which is also what its previous snapshot was
      * judged against. Caught here rather than failing the refresh, because the membership read
      * tolerates an unreachable `Revocation` block on an isolated node
-     * (`ControlDatabase.readRevokedStampRows`) and `queryRevocations` does not: failing the
+     * (`ControlDatabase.readRevocationRowsOrNone`) and `queryRevocations` does not: failing the
      * whole refresh on that would freeze such a node's gate snapshot.
      * NOTE: on such a node the sync is skipped at every refresh until the block is held; if an
      * isolated node ever has to derive owner additions while cut off, give `queryRevocations`
@@ -7486,9 +7998,12 @@ export class CadreNode {
      * NOTE: cost per membership refresh is one `OwnerKey` read and one `Revocation` read (the
      * tombstone list, whose `OwnerKey` stamps are handed to `queryOwnerKeyRows` so it does not
      * read the retired set again, and which the hang-up of removed peers shares through
-     * `loadTombstones`), plus the owner-grant chain when some owner row is grant-seated (the
-     * NOTE in the body). Before the Revocation ledger marker is filed that one read is a cohort
-     * consult per refresh, which `control-founding-consult-budget.spec.ts` pins.
+     * `loadTombstones`), plus the owner-grant chain's two reads when some owner row is
+     * grant-seated. All are served from the control database's membership views, so they reach
+     * storage only after their table changed or the view's maximum age passed
+     * (`control-founding-consult-budget.spec.ts` pins the counts). On a party founded before
+     * genesis created every control collection, whose `Revocation` table has never been
+     * written, each `Revocation` load is a cohort consult until the first tombstone.
      *
      * @param retry - forwarded to every read; `false` on the refresh path, for the reason on
      *   {@link refreshAuthorizedControlPeers}.
@@ -7509,9 +8024,6 @@ export class CadreNode {
             const tombstones = await loadTombstones();
             const retired = new Set(tombstones.filter(tombstone => tombstone.tableName === 'OwnerKey').map(tombstone => tombstone.stampId));
             const rows = await controlDatabase.queryOwnerKeyRows(retry, retired);
-            // NOTE: the grant chain is two whole-table reads per membership refresh while any owner
-            // row is grant-seated, over tables that are append-only; the memo suggested on
-            // loadInvitationChain would cover it too.
             const grantChain = rows.some(isConsentSeated) ? await this.loadOwnerGrantChain(controlDatabase, retry) : null;
             const derivation = deriveOwnerAnchor({ partyId: this.partyId, anchor: before, rows, tombstones, grantChain });
             if (derivation.refusedRemovals.length > 0) {
@@ -7569,15 +8081,8 @@ export class CadreNode {
     /**
      * The usage and invitation rows {@link hasAnchoredProof} resolves an invitation-admitted
      * `CadrePeer` row through, keyed for lookup. Both reads honour `retry` for
-     * {@link listAuthorizedMembers}' reason.
-     *
-     * NOTE: once invitations are the only admission route (ticket
-     * `cadre-invitations-redeemable-by-any-member`), every cadre with a member not yet
-     * re-vouched pays these two whole-table reads and three ed25519 verifies per row on every
-     * membership refresh. Both tables are append-only and a verified chain never changes (expiry,
-     * use count and withdrawal are not re-checked), so if the gate's admission deadline ever
-     * shows this, memoise verified (row stamp → usage stamp) pairs per node and skip the reads
-     * when every admitted row is already memoised.
+     * {@link listAuthorizedMembers}' reason. The rows come from the membership views, so they
+     * reach storage only after the table changed or the view's maximum age passed.
      */
     async loadInvitationChain(controlDatabase, retry) {
         const [usages, invites] = await Promise.all([
@@ -7731,16 +8236,16 @@ export class CadreNode {
      * write-driven refresh never fires for it — yet applying it can ANCHOR a new
      * owner key, which flips rows already present from unauthorized to authorized.
      * Hence the explicit refresh here; without it a peer the freshly anchored owner
-     * vouched for stays denied until the next timed cohort reconcile.
+     * vouched for stays denied until the next membership change or the backstop pass.
      */
     seedEventCallbacks() {
         return {
             onSeedReceived: (partyId, peerId) => this.emit('seed:received', { partyId, peerId }),
             onSeedApplied: (partyId, peersAdded, seed) => {
-                this.emit('seed:applied', { partyId, peersAdded });
                 // A wire-delivered seed never passes through the `applySeed` wrapper, so
                 // this is where its preferred peers become cold-start bootstrap targets.
-                this.recordSeedBootstrapPeers(seed);
+                const dialTargets = this.recordSeedBootstrapPeers(seed);
+                this.emit('seed:applied', { partyId, peersAdded, dialTargets });
                 void this.refreshMembershipGate('seed-applied');
             },
             onSeedError: (partyId, error) => this.emit('seed:error', { partyId, error }),
@@ -7764,13 +8269,14 @@ export class CadreNode {
      * NO `CadrePeer` writer needs to call this: the control database notifies it
      * after every committed member-row write (`ControlDatabase.mutateCadrePeer`,
      * wired in {@link start}), which is exactly what makes the refresh automatic
-     * rather than a caller obligation. It stays public for the changes that write
-     * NO row locally and so raise no notification:
-     *
-     * - a membership row that arrived by REPLICATION (otherwise picked up on the
-     *   next timed cohort reconcile — bounded staleness by design), and
-     * - a newly anchored trusted owner key ({@link applySeed}), which flips rows
-     *   ALREADY present from unauthorized to authorized without touching them.
+     * rather than a caller obligation. A membership row that arrived by REPLICATION
+     * refreshes it too, through the gate watch {@link start} takes on the membership
+     * tables (within its 50 ms window; a change this node's storage never hears of
+     * waits for the reconcile pass's backstop). It stays public for the one change
+     * that writes no row and raises no table event: a newly anchored trusted owner
+     * key ({@link applySeed}), which flips rows ALREADY present from unauthorized to
+     * authorized without touching them. A caller may also await it to observe a
+     * refresh that began after its own change.
      *
      * Coalescing: marks the snapshot stale and resolves once a refresh that began
      * after this call has completed, so a caller that awaits it always observes
@@ -7795,13 +8301,10 @@ export class CadreNode {
      * connected peer it newly covers — the strand-side counterpart to
      * {@link refreshMembershipGate}.
      *
-     * Unlike that one, NOTHING calls this automatically: strand membership is
-     * written through the strand's own `Database` handle
-     * (`strand-membership-writer.ts`), which raises no notification this runtime
-     * can hook, and a revocation that arrives by REPLICATION raises none either.
-     * The gate therefore polls (default 30 s). An app that has just called
-     * `revokeMember` or `leaveStrand` should follow the write with this call so
-     * the cut is immediate rather than up to one poll interval late.
+     * The gate already refreshes on its own when the strand's `Member` or
+     * `MemberPeer` table changes: at once for a `revokeMember` or `leaveStrand`
+     * written on this machine, and when a replicated removal reaches this node's
+     * storage. Call this to await the cut — it resolves after the sweep.
      *
      * Quiet no-op for a strand that is not running, is quiesced, is open, or has
      * the gate disabled. Never rejects; resolves once the sweep has finished.
@@ -7852,8 +8355,8 @@ export class CadreNode {
      * CONCURRENT external write's refresh for the life of the scope: that writer's
      * promise resolves before its peer is admitted, and admission lands at scope
      * exit instead. Acceptable — scope exit is milliseconds away (bounded by the
-     * drain), versus the ~15 s reconcile interval that was the alternative before
-     * any of this existed — but it is the reason these scopes stay short and rare.
+     * drain), versus the reconcile pass that was the alternative before any of this
+     * existed — but it is the reason these scopes stay short and rare.
      */
     async deferMembershipGateRefresh(reason, body) {
         this.membershipGateDeferDepth++;
@@ -7989,7 +8492,7 @@ export class CadreNode {
             throw new Error('CadreNode must be started before planning an owner key removal');
         }
         const key = requireEd25519PublicKeyB64(ownerPublicKeyB64, 'owner key');
-        return (await this.planOwnerKeyRemovalWith(this.controlDatabase, key)).plan;
+        return await this.planOwnerKeyRemovalWith(this.controlDatabase, key);
     }
     /**
      * Remove an owner key, signed by this node's owner key (which must be a different owner): the
@@ -8022,7 +8525,7 @@ export class CadreNode {
             throw new Error('Seed bootstrap service not initialized. Call initializeSeedBootstrap() first.');
         }
         const key = requireEd25519PublicKeyB64(ownerPublicKeyB64, 'owner key');
-        const { plan, rows } = await this.planOwnerKeyRemovalWith(this.controlDatabase, key);
+        const plan = await this.planOwnerKeyRemovalWith(this.controlDatabase, key);
         if (plan.refusal !== null) {
             throw new OwnerKeyRemovalRefusedError(key, plan.refusal);
         }
@@ -8037,18 +8540,18 @@ export class CadreNode {
         }
         const action = `Remove the owner key ${namedOwnerKey(key, plan.ownerKey?.label ?? null)} from the cadre`;
         const removed = await signer.approve({ partyId: this.partyId, action }, async (sign) => {
-            await this.revouchDependents(service, revouch, rows, sign);
+            await this.revouchDependents(service, revouch, sign);
             return await service.removeOwner(key, sign);
         });
         log('removeOwner(%s): removed=%s revouched owners=%d peers=%d', key, removed, revouch.owners.length, revouch.peers.length);
         return { removed, revouched: revouch };
     }
-    /** {@link planOwnerKeyRemoval} with the rows it was judged from, which {@link removeOwner} re-vouches dependents by. */
+    /** {@link planOwnerKeyRemoval} over `controlDatabase`, for a caller that already holds it. */
     async planOwnerKeyRemovalWith(controlDatabase, key) {
         const [rows, owners, grants] = await Promise.all([
             controlDatabase.queryCadrePeers(), controlDatabase.queryOwnerKeyRows(), controlDatabase.listOwnerGrantStatuses()
         ]);
-        const plan = planOwnerKeyRemoval({
+        return planOwnerKeyRemoval({
             key,
             rows,
             owners,
@@ -8057,7 +8560,6 @@ export class CadreNode {
             selfOwnerKey: this.ownerSignerEnrolledIn(new Set(owners.map((owner) => owner.key)))?.publicKey ?? null,
             anchor: this.trustedOwnerStore?.all() ?? new Set(),
         });
-        return { plan, rows };
     }
     /**
      * The re-vouch step of {@link removeOwner}: every chosen dependent row is rebound to this
@@ -8066,12 +8568,14 @@ export class CadreNode {
      * the only rebinding `OwnerKey.NoUpdate` allows); peers through the owner re-touch
      * (`reauthorizePeer`), which rewrites `VouchOwner`/`VouchSig` and clears `VouchUsage`.
      *
-     * The peer re-touch keeps the row's `UpdatedAt` where it is. The owner branch of
-     * `CadrePeer.AuthorizedUpdate` does not require a bump, and bumping a self-published row
-     * would leave its `Sig` unverifiable, so unresolvable and unlisted as self-published, until
-     * the machine's next heartbeat (the NOTE on that constraint); the write-while-alone drain
-     * bumps because it re-issues an otherwise unchanged row, whereas here the voucher itself
-     * changes. A row never self-published carries no stamp to keep and gets the clock.
+     * The peer re-touch is the re-vouch mode of `reauthorizePeer`, which leaves the row's
+     * `UpdatedAt` untouched and writes whether or not the machine has self-published. The owner
+     * branch of `CadrePeer.AuthorizedUpdate` does not require a bump, and bumping a self-published
+     * row would leave its `Sig` unverifiable, so unresolvable and unlisted as self-published, until
+     * the machine's next heartbeat (the NOTE on that constraint); the write-while-alone drain's
+     * re-broadcast mode bumps because it re-issues an otherwise unchanged row, whereas here the
+     * voucher itself changes. A dependent whose row vanished since the plan is reported absent and
+     * skipped.
      *
      * A failure rethrows and stops the call before the removal; re-vouches already made stay
      * (the row is only re-signed by a current owner) and the caller retries. An owner re-vouch
@@ -8082,7 +8586,7 @@ export class CadreNode {
      * rows anyway, so the other machines may get the removal's tombstone without the new voucher
      * (`tickets/backlog/bug-owner-key-removal-revouch-made-alone-is-not-reissued`).
      */
-    async revouchDependents(service, revouch, rows, sign) {
+    async revouchDependents(service, revouch, sign) {
         for (const key of revouch.owners) {
             await service.revouchOwner(key, sign);
             if (this.committedAlone()) {
@@ -8091,9 +8595,7 @@ export class CadreNode {
             }
         }
         for (const dependent of revouch.peers) {
-            const row = rows.find((candidate) => candidate.peerId === dependent);
-            const updatedAt = row !== undefined && row.updatedAt > 0 ? row.updatedAt : Date.now();
-            await service.reauthorizePeer(dependent, updatedAt, sign);
+            await service.reauthorizePeer(dependent, 'revouch', sign);
         }
     }
     /**
@@ -8118,6 +8620,10 @@ export class CadreNode {
             throw new Error('CadreNode must be started before planning a machine removal');
         }
         const machines = await this.readMachines(this.controlDatabase);
+        // NOTE: the invitation and usage lists come from the membership views, so on a node that
+        // heard of a row with no storage event (outside the block's cohort) an invitation issued
+        // elsewhere in the last MEMBERSHIP_VIEW_MAX_AGE_MS can be missing and left unwithdrawn. If
+        // removal must close that window, read both tables fresh here.
         const [invites, usages, owners] = await Promise.all([
             this.controlDatabase.listCadreInviteStatuses(),
             this.controlDatabase.queryCadreInviteUsages(),
@@ -8328,12 +8834,13 @@ export class CadreNode {
      * that connection from the addresses it was handed, on this launch or a later
      * one.
      *
-     * Nothing is dialed here: the drone has not received the seed yet. After
-     * delivering it, call {@link reconcileControlCohort} to dial straight away;
-     * otherwise the next timed reconcile pass does. A pass already in flight is
-     * joined rather than restarted, and one that listed siblings before this add
-     * does not dial the drone, so a caller waiting for the connection should allow
-     * for one more timed pass.
+     * Nothing is dialed here: the drone has not received the seed yet. Its `CadrePeer`
+     * insert is a membership change, so a reconcile pass dials the retained addresses
+     * within about 50 ms of this call returning, whether or not the caller runs one, and
+     * a failed dial is retried on the redial ladder (15 s, 30 s, 60 s, then every
+     * backstop). A caller that wants to await the connection calls
+     * {@link reconcileControlCohort} itself, twice in a row: a call made while a pass is
+     * in flight joins it, and that pass may have listed siblings before this add.
      */
     async addDrone(options) {
         if (!this.seedBootstrapService) {
@@ -8378,12 +8885,19 @@ export class CadreNode {
      * by the node (its anchored branch), the row insert is a no-op on an existing row
      * (`ControlDatabase.insertCadrePeer`), and the dial target is replaced, not duplicated.
      *
-     * The first dial is kicked (not awaited) so it happens now rather than on the next
-     * timed reconcile pass; a caller that wants to wait for the connection awaits
-     * {@link reconcileControlCohort} itself, with `addDrone`'s caveat that a pass already
-     * in flight is joined rather than restarted. The claimed node dials this one back too,
-     * as the seed's minter, when this node has a listen or relay address; a phone
-     * with neither is reached by its own reconcile pass alone.
+     * Once it has acknowledged, the node restarts into this party (cadre-cli's
+     * `restartIntoClaimedParty`), closing the connection the claim used. The dial target puts
+     * the address that connection was dialed on first, so every later dial tries the address
+     * that just worked before the others (a QR code lists public addresses before LAN ones, and
+     * a home router often drops a connection to its own public address from inside). A
+     * reconcile pass is kicked (not awaited) now, and the vouch's `CadrePeer` insert starts
+     * another about 50 ms later; for the first {@link CONTROL_COHORT_REDIAL_MS} after the
+     * acknowledgement this node also retries the node every second while it restarts
+     * ({@link redialClaimedNode}). A caller that wants to wait for the connection awaits
+     * {@link reconcileControlCohort} itself, with `addDrone`'s caveat that a pass already in
+     * flight is joined rather than restarted. The claimed node dials this one back too, as the
+     * seed's minter, when this node has a listen or relay address; a phone with neither is
+     * reached by its own dials alone.
      *
      * @throws {ClaimRefusedError} when the node was reached and refused the claim.
      * @throws {PeerUnreachableError} when the node was never reached.
@@ -8398,6 +8912,7 @@ export class CadreNode {
         if (!signer) {
             throw new Error('Owner key required to create seeds');
         }
+        let dialAddrs = target.multiaddrs;
         await signer.approve({ partyId: this.partyId, action: `Claim the node ${target.peerId} for the cadre` }, async (sign) => {
             const seed = await service.createSeed(sign);
             const proof = claimProof(secret, target.peerId, seed.signerKey, seedDigest(seed));
@@ -8405,13 +8920,87 @@ export class CadreNode {
             if (!ack.accepted) {
                 throw new ClaimRefusedError(target.peerId, ack.reason ?? 'no reason given', ack.code);
             }
+            // Read before anything else is awaited: the node's restart closes the connection.
+            dialAddrs = this.connectedAddrFirst(target.peerId, target.multiaddrs);
             log('claimNode: %s accepted the claim; vouching it into the cadre', target.peerId);
             // The wrapper queues the write for re-replication: a phone claiming its first
             // always-on node has no control connection yet, so the insert commits local-only.
+            // NOTE: accepted tradeoff — the node's own signed record is not carried in the
+            // acknowledgement and written with this vouch: it would be signed before the node
+            // restarts, so its addresses may not be the restarted node's (a new relay reservation),
+            // and the write-while-alone drain re-sends only rows whose `Sig` is null
+            // (drainPendingPeerWrites, reconstructAuthoredMembership), so a vouch carrying it would
+            // never reach this cadre's other members. The node publishes its record once it has
+            // restarted. Revisit if the drain ever re-sends signed rows.
             await this.vouchPeer(target.peerId, target.multiaddrs, sign);
         });
-        this.retainDialTarget(target.peerId, target.multiaddrs, 'claimNode');
+        this.retainDialTarget(target.peerId, dialAddrs, 'claimNode');
         void this.reconcileControlCohort().catch((error) => log('claimNode: reconcile kick after claiming %s failed: %o', target.peerId, error));
+        void this.redialClaimedNode(target.peerId).catch((error) => log('claimNode: redialing %s while it restarts failed: %o', target.peerId, error));
+    }
+    /**
+     * `addrs` with the address this node's open connection to `peerId` was dialed on moved to
+     * the front, when it is one of them; `addrs` unchanged otherwise. Compared without a
+     * trailing `/p2p/<peerId>`, which a QR code's address may or may not carry ({@link addrKey}).
+     */
+    connectedAddrFirst(peerId, addrs) {
+        const controlNode = this.controlNode;
+        if (!controlNode) {
+            return addrs;
+        }
+        const remote = peerIdFromString(peerId);
+        const connected = new Set(controlNode.getConnections(remote).map((c) => addrKey(c.remoteAddr.toString(), remote)));
+        const index = addrs.findIndex((addr) => connected.has(addrKey(addr, remote)));
+        return index <= 0 ? addrs : [addrs[index], ...addrs.slice(0, index), ...addrs.slice(index + 1)];
+    }
+    /**
+     * Retry a node this node has just claimed ({@link claimNode}) once every
+     * {@link CLAIM_RESTART_REDIAL_STEP_MS} for {@link CONTROL_COHORT_REDIAL_MS} after it accepted.
+     * The node restarts into this party right after it acknowledges, and while it does, a
+     * reconcile pass's dial either fails or opens a connection the restarting node closes at
+     * once. The first failure puts the pass's next redial 15 s out ({@link armReconcileWake}),
+     * and an owner that cannot be dialed has no other way back to the node.
+     *
+     * Each step dials once from the retained dial target, unless a connection to the node is
+     * open: the one the claim used until the restart closes it, or one to the restarted node
+     * that has not identified yet. A dial that connects does not end the loop, because the
+     * restarting node may still close it. The loop ends once the node has identified as serving
+     * this party's control database ({@link identifiedControlPeers}, which the placeholder node
+     * it ran before the claim never joins), at the window's end, when this node stops, or when
+     * the target is forgotten ({@link removePeer}). Steps run one after another, so a dial
+     * finishes on its own budget first. The redial streak is left alone: the ladder carries on
+     * after the window. One loop per peer: a repeated claim while one runs starts none.
+     */
+    async redialClaimedNode(peerId) {
+        if (this.claimRedialPeers.has(peerId)) {
+            return;
+        }
+        this.claimRedialPeers.add(peerId);
+        try {
+            const endsAt = Date.now() + CONTROL_COHORT_REDIAL_MS;
+            while (Date.now() < endsAt && await this.redialClaimedNodeStep(peerId)) {
+                await new Promise((resolve) => {
+                    const timer = setTimeout(resolve, CLAIM_RESTART_REDIAL_STEP_MS);
+                    // Node-only: a stop mid-window must not keep the event loop alive for the step.
+                    timer.unref?.();
+                });
+            }
+        }
+        finally {
+            this.claimRedialPeers.delete(peerId);
+        }
+    }
+    /** One step of {@link redialClaimedNode}; false ends the loop. */
+    async redialClaimedNodeStep(peerId) {
+        const controlNode = this.controlNode;
+        const target = this.bootstrapPeerStore?.all().get(peerId);
+        if (!this._running || !controlNode || !target || this.identifiedControlPeers.has(peerId)) {
+            return false;
+        }
+        if (controlNode.getConnections(peerIdFromString(peerId)).length === 0) {
+            await this.dialBootstrapPeer(peerId, target.addrs, 'claimNode(redial)');
+        }
+        return true;
     }
     /**
      * Add a phone to the cadre with relay support.
@@ -8569,11 +9158,13 @@ export class CadreNode {
      * Steps: refuse an invitation for another party; pin its `ownerKeys` into the node-local
      * anchor (`'invite'` provenance — the pin sticks even if the redemption then fails, as it
      * does for a seed invite, so a later attempt or seed from the same owner is anchored);
-     * sign the request; dial the listed members in order (`redeemAtMembers`); on
-     * acceptance verify the reply against the pinned keys, merge its dial hints into the peer
-     * store, retain the answering member as a cold-start dial target, refresh the membership
-     * gate and start a cohort reconcile pass. This node then syncs the control database over
-     * the connection it already holds, the same unified behaviour as after a seed.
+     * sign the request, which carries this node's own signed address record; dial the listed
+     * members in order (`redeemAtMembers`); on acceptance verify the reply against the pinned
+     * keys, merge its dial hints into the peer store, retain the answering member as a
+     * cold-start dial target and start a cohort reconcile pass. This node then syncs the control
+     * database over the connection it already holds, the same unified behaviour as after a seed.
+     * The member writes the record on this node's `CadrePeer` row when its signature verifies, so
+     * the row resolves to this node's addresses from its admission.
      *
      * A device that is already a member (a restart with the invitation still configured) is
      * answered as accepted by the member's idempotent redemption, with no second usage row.
@@ -8617,7 +9208,15 @@ export class CadreNode {
         }
         await this.trustOwnerKeys(invitation.ownerKeys, 'invite');
         const pinned = new Set(invitation.ownerKeys);
-        const request = signRedeemRequest(invitation, { peerKey: signer.publicKeyB64, peerPrivateKey: signer.privateKeyB64 }, generateStampId(controlNode.peerId.toString()), await this.resolveInviteAddresses());
+        const selfPeerId = controlNode.peerId.toString();
+        // The addresses a publish would carry, sanitized before signing: the member stores the
+        // signature only over the exact list it stores.
+        const record = this.signSelfRecord(selfPeerId, signer, sanitizeAddrs(await this.collectSelfAddrs()), null);
+        const request = signRedeemRequest(invitation, { peerKey: signer.publicKeyB64, peerPrivateKey: signer.privateKeyB64 }, generateStampId(selfPeerId), record);
+        // Before dialing, not on acceptance: the member's commit can reach this node before its
+        // reply does, and a reconcile pass that reads the row first must find it current. A record
+        // no member wrote matches no row, so a failed redemption leaves nothing to undo.
+        this.lastSignedSelfRecord = record;
         const { reply, memberPeerId, memberAddr } = await redeemInPasses(controlNode, {
             invitation,
             request,
@@ -8628,7 +9227,6 @@ export class CadreNode {
             ...(options.onProgress ? { onProgress: options.onProgress } : {}),
         });
         log('Cadre invitation %s redeemed at %s', invitation.invite.key, memberAddr);
-        const selfPeerId = controlNode.peerId.toString();
         await mergeSeedPeers(controlNode, reply.peers.filter((peer) => peer.peerId !== selfPeerId));
         if (memberPeerId !== null) {
             // Every address the bundle gave the answering member, so a restart has the same choices.
@@ -8637,9 +9235,8 @@ export class CadreNode {
                 .map((addr) => addr.toString());
             this.retainDialTarget(memberPeerId, memberAddrs, 'redeemCadreInvitation');
         }
-        // The pins above flip nothing yet (this node holds no rows), but the rows about to
-        // replicate in are judged against them; refresh so the snapshot is not stale when they land.
-        await this.refreshMembershipGate('invitation-redeemed');
+        // No gate refresh here: the rows about to replicate in are judged against the pins above
+        // when they land, by the gate watch's refresh and the head refresh of this pass.
         void this.reconcileControlCohort().catch((error) => log('redeemCadreInvitation: the post-redemption reconcile pass failed: %o', error));
         return { peerId: memberPeerId, redeemedAt: new Date().toISOString() };
     }
@@ -8811,25 +9408,31 @@ export class CadreNode {
         return { ownerKey: signer.publicKey, alreadyOwner: reply.alreadyOwner, memberPeerId };
     }
     /**
-     * {@link redeemOwnerGrant}'s local path: the alone refusal, the seat, the liveness refusal,
-     * then one approval around the consent signature and the write. Every refusal comes before
-     * the approval, so a prompting signer never asks for a redemption this node would refuse.
+     * {@link redeemOwnerGrant}'s local path: the alone refusal, the seat (only when this node holds
+     * no grant row), the liveness refusal, then one approval around the consent signature and the
+     * write. Every refusal comes before the approval, so a prompting signer never asks for a
+     * redemption this node would refuse. The held row is read once and carried into the liveness
+     * check and the write, as the member handler carries it (`MemberRedemptionHandler`).
      */
     async redeemOwnerGrantLocally(controlDatabase, bundle, signer) {
         if (await this.ownerKeyWriteAlone()) {
             throw new OwnerKeyAddRefusedError(signer.publicKey, 'alone');
         }
-        await refusingUnsyncedGrant(() => controlDatabase.seatOwnerGrant(bundle.grant));
+        const held = await controlDatabase.queryOwnerGrant(bundle.grant.key);
+        if (held === null) {
+            await refusingUnsyncedGrant(() => controlDatabase.seatOwnerGrant(bundle.grant));
+        }
         // The write answers a key that is already an owner without reading the grant, so only a
         // key that is not one needs the grant live (the same permissive pre-check the handler's
         // admission uses; the write enforces every condition).
-        if (!(await controlDatabase.getOwnerKeys()).has(signer.publicKey) && !await controlDatabase.isOwnerGrantLive(bundle.grant)) {
+        if (!(await controlDatabase.getOwnerKeys()).has(signer.publicKey) && !await controlDatabase.isOwnerGrantLive(bundle.grant, held)) {
             throw new OwnerGrantRejectedError('grant-spent', 'The grant is expired, withdrawn, exhausted, or its issuer is no longer an owner');
         }
         const { alreadyOwner } = await signer.approve(ownerGrantApproval(bundle), async (sign) => {
             const request = await this.signOwnerGrantRequest(bundle, signer, sign);
             return await refusingUnsyncedGrant(() => controlDatabase.redeemOwnerGrant({
                 grantKey: bundle.grant.key,
+                ...(held ? { held } : {}),
                 seatedKey: signer.publicKey,
                 usageStampId: request.usageStampId,
                 grantSig: request.grantSig,
@@ -9015,7 +9618,7 @@ export class CadreNode {
             // Overridden on the same grounds: only this node holds the running strand
             // instance's database and this party's `StrandPartyKey` identity, and a wrong
             // issuer here would admit joiners under someone else's authority.
-            issueMembershipInvite: (strandId, signal) => this.issueStrandMembershipInvite(strandId, signal)
+            issueMembershipInvite: (host, signal) => this.issueStrandMembershipInvite(host, signal)
         });
         const previous = this.strandSolicitationService;
         if (previous) {
@@ -9280,10 +9883,17 @@ export class CadreNode {
             selfId: this.controlNode.peerId.toString(),
             linkRoundTripMs: this.config.network?.linkRoundTripMs,
             isOwner: () => this.isOwner(),
-            // NOTE: one owner-key read and one pending-join read (three tables) per pass (30 s); for a
-            // party that never asked for a join the second reads never-written blocks, which consults
-            // the cohort. If it shows up in a device profile, poll more slowly while the tables read empty.
             readRows: () => database().queryPendingJoins(),
+            // What `isOwner` and `queryPendingJoins` read: the owner keys, the joins and their
+            // outcomes, and the tombstones both filter by. A dismissal writes its tombstone with the
+            // request's delete, but the two can replicate separately.
+            watchChanges: (onChange, options) => database().watchTables([
+                'CadreControl.JoinRequest',
+                'CadreControl.JoinSuccess',
+                'CadreControl.JoinFailure',
+                'CadreControl.OwnerKey',
+                'CadreControl.Revocation',
+            ], onChange, options),
             readRow: (id) => database().queryPendingJoin(id),
             attempt: (join) => this.formStrand(this.pendingJoinInvitation(join), parseStoredDisclosure(join)),
             record: (expected, outcome) => ownerWrite(`Record the outcome of the pending join ${expected.Id}`, (ownerKey, sign) => database().recordJoinOutcome(expected, outcome, ownerKey, sign)),
@@ -9344,6 +9954,13 @@ export class CadreNode {
      * token is spent, and a join no store names would vanish at the next restart.
      */
     async rememberFormedStrand(result) {
+        // NOTE: accepted tradeoff — the party-wide `JoinedStrand` row is not written here, in
+        // formStrand's own approval, but later by an owner machine's connected reconcile pass
+        // (`JoinedStrandSession.syncWithParty`). That publish is gated on a control connection so a
+        // machine cut off from its own party never commits it alone, and during formStrand the
+        // count always includes the connection to the other party's responder, so the gate would
+        // always pass here. Non-owner and signer-less machines need this local path regardless.
+        // Revisit if the gate can tell a connection to this party's own machines from any other.
         try {
             await this.joinedStrands.remember({
                 Id: result.strandId,
@@ -9383,7 +10000,7 @@ export class CadreNode {
      */
     async adoptFormationMembershipInvite(strandId, invite) {
         try {
-            await this.ensureStrandPartyKey(strandId);
+            this.formedPartyKeys.set(strandId, await this.ensureStrandPartyKey(strandId));
         }
         catch (error) {
             throw new FormationPostApprovalError(strandId, `Formation for strand ${strandId} was approved (its one-time token is spent), but ` +
@@ -9428,6 +10045,10 @@ export class CadreNode {
      * implements): mint a single-use `Strand.Invite` against the LIVE host strand so a
      * validated joiner can seat its own `Strand.Member` row.
      *
+     * `host` is the strand as the formation session resolved it from its `Strand` row, read
+     * once there; this method reads no `Strand` row of its own. A row removed in between is
+     * refused when the consent row is written (the `StrandExists` CHECK).
+     *
      * - Open host strand → `null` (no members, nothing to invite into).
      * - Closed host strand whose founder launch was refused as pre-split
      *   ({@link strandLaunchRefusals}) → rethrow the recorded `PreSplitStrandIdentityError`;
@@ -9475,17 +10096,12 @@ export class CadreNode {
      * `MEMBERSHIP_INVITE_UNAVAILABLE_REASON` forever; that flow needs manager delegation,
      * not a retry.
      */
-    async issueStrandMembershipInvite(strandId, signal) {
+    async issueStrandMembershipInvite(host, signal) {
+        const { strandId } = host;
         if (!this.controlDatabase) {
             throw new Error(`Cannot issue a membership invitation for strand ${strandId}: control database unavailable`);
         }
-        const row = await this.controlDatabase.queryStrand(strandId);
-        if (!row) {
-            // The bound resolution saw this row moments ago; a vanished row is a concurrent
-            // unpublish — reject rather than invite into a strand this party just removed.
-            throw new Error(`Cannot issue a membership invitation for strand ${strandId}: its Strand row is gone`);
-        }
-        if (row.Type !== 'c') {
+        if (host.strandType !== 'c') {
             return null;
         }
         const refusal = this.strandLaunchRefusals.get(strandId);
@@ -9503,8 +10119,8 @@ export class CadreNode {
             throw new Error(`Cannot issue a membership invitation for closed strand ${strandId}: its runtime is ` +
                 'not live on this responder (not launched, or still starting)');
         }
-        if (row.MemberPrivateKey) {
-            await assertNotPreSplitStrand(db, strandId, strandMemberKeyPair(row.MemberPrivateKey).publicKeyB64);
+        if (host.memberPrivateKey) {
+            await assertNotPreSplitStrand(db, strandId, strandMemberKeyPair(host.memberPrivateKey).publicKeyB64);
         }
         if (signal?.aborted) {
             throw new Error(`Cannot issue a membership invitation for closed strand ${strandId}: the formation ` +