@serfab/cadre-core 0.11.0 → 0.12.0

package/src/membership-connection-gater.ts

@@ -1,204 +1,533 @@
-/**
- * Control-network inbound connection gate (defense-in-depth).
- *
- * Layer 2 of the membership enforcement chain: the PRIMARY gates are per-stream
- * — every sensitive sereus control protocol rejects a peer that fails the
- * voucher-anchored membership check (wake and strand-addr via
- * `CadreNode.isAuthorizedMember`; the Optimystic control-DB protocols via the
- * materialized-snapshot gate `CadreNode.authorizeInboundControlStream`), and
- * the replicated rows an outsider *can* still write are disbelieved at read
- * time. This module adds the opportunistic connection-level layer on top: a
- * peer this node can positively determine is NOT authorized is refused at the
- * encrypted-connection checkpoint, before any protocol negotiation, so a
- * known-nothing outsider is never even in the conversation.
- *
- * ## The stranger allowlist (the ONE place it is defined)
- *
- * A libp2p connection gater decides per CONNECTION, before protocols are
- * negotiated, so "allow seed, deny repo" cannot be expressed here — instead the
- * policy admits a connection whenever a legitimate stranger interaction could be
- * riding it, and the per-stream gates take over. The complete set of protocols a
- * NOT-yet-authorized peer may legitimately speak on a control node is:
- *
- *  - `/sereus/seed/1.0.0` ({@link SEED_PROTOCOL}) — enrollment seed delivery.
- *    An owner dials a brand-new node to seed it (the new node has no members
- *    yet, so its gate is inert), and an invited phone dials the owner before it
- *    is authorized (admitted via the enrollment window `CadreNode.createInvite`
- *    opens). The handler's own trust decision is the anchored seed-trust policy.
- *  - `/sereus/formation/1.0.0` ({@link FORMATION_PROTOCOL}) — cross-party
- *    strand formation via open invitations. Stranger-facing BY DESIGN: the
- *    initiator is another party, and its token is only checkable inside the
- *    protocol. The exemption is therefore keyed on EXPECTATION of a stranger,
- *    not capability to serve one: stranger denial is suspended only while this
- *    node has at least one UNEXPIRED, NOT-FULLY-CONSUMED open invitation
- *    outstanding (`StrandSolicitationService.hasOutstandingInvitation` — the
- *    tokens this process minted or published, plus any still-redeemable
- *    `FormationInvite` row the usage recorder can see). Merely registering the
- *    responder (`CadreNode.initializeStrandSolicitation`) does NOT suspend it,
- *    so an app that registers eagerly at node bring-up keeps a live gate. The
- *    handler's own trust decision remains the per-token check, which is
- *    strictly finer than this one: a peer admitted here can still be rejected
- *    in-protocol for a bogus or spent token.
- *
- * There is one further connection-level carve-out, which is NOT a protocol
- * exemption and needs no stranger window:
- *
- *  - **Announced delegate peers** (`delegate-admission.ts`). A member's strand
- *    node runs as a separate libp2p identity whose peerId no sibling can
- *    recompute, so a relay-running control node would deny its circuit-relay
- *    reservation. Before starting a strand node, a member's control node
- *    announces that peerId over the already-authenticated strand-addr RPC, and
- *    the receiver holds a short-lived grant for it
- *    (`CadreNode.grantDelegateAdmission`). The grant admits the CONNECTION and
- *    nothing else — a reservation needs no stream — and is deliberately
- *    invisible to the per-stream gate below, so a delegate still gets refused
- *    on every members-only protocol.
- *
- * Everything else a control node handles — the Optimystic control-DB protocols
- * (`/optimystic/control-<party>/{repo,cluster,sync,block-transfer}/…`), wake
- * (`/sereus/strand-wake/1.0.0`), and strand-addr (`/sereus/strand-addr/1.0.0`)
- * — is members-only, and every one of them enforces that per-stream. Wake and
- * strand-addr check `isAuthorizedMember` inside their handlers; the four
- * control-DB protocols are gated by the fail-closed
- * `CadreNode.authorizeInboundControlStream` (wired as libp2p's
- * `authorizeInboundStream` upstream hook), which judges each inbound stream
- * against the MATERIALIZED authorized-peer snapshot — synchronous and
- * in-memory, because a live DB read from inside the gate would deadlock into
- * mutual denial. The two layers complement, not duplicate: this connection
- * gate is fail-open over a live DB read (deny only on positive proof of an
- * outsider), the stream gate is fail-closed over the snapshot and has NO
- * stranger carve-outs — an enrollment window admits a stranger's connection
- * for seed delivery, yet its repo streams are still refused.
- *
- * ## Fail-open, deliberately
- *
- * This layer only ever denies on a POSITIVE determination of "unauthorized
- * outsider while no stranger path is open". Any error, missing dependency, or
- * ambiguous state admits the connection and defers to the fail-closed stream
- * gates — a DB hiccup must not partition a legitimate cadre.
- */
-
-import debug from 'debug';
-import type { ConnectionGater, PeerId, MultiaddrConnection } from '@libp2p/interface';
-import { SEED_PROTOCOL } from './seed-bootstrap.js';
-import { FORMATION_PROTOCOL } from './strand-formation-protocol.js';
-
-const log = debug('sereus:cadre:connection-gater');
-
-/**
- * The protocols a not-yet-authorized peer may legitimately speak on a control
- * node (see the module doc above for why each is open and where its own trust
- * decision lives). Exported as the single reference point so the allowlist
- * cannot silently drift across modules.
- */
-export const STRANGER_OPEN_PROTOCOLS: readonly string[] = [SEED_PROTOCOL, FORMATION_PROTOCOL];
-
-/**
- * How long `CadreNode.createInvite` holds the inbound gate open for strangers
- * when the invite carries no explicit expiry: the invitee must dial in before
- * it is authorized, and an expiry-less invite gives the gate no bound of its
- * own. Redeeming after this window still works once the owner re-mints (or the
- * caller re-opens the window via `CadreNode.openEnrollmentWindow`).
- */
-export const DEFAULT_ENROLLMENT_WINDOW_MS = 30 * 60 * 1000;
-
-/**
- * Deadline for one admission decision, after which the connection is ADMITTED.
- *
- * libp2p awaits `denyInboundEncryptedConnection` inside the inbound upgrade
- * WITHOUT racing its inbound-upgrade timeout signal (unlike the pre-encryption
- * `denyInboundConnection` hook), so a decision that never settles wedges that
- * upgrade forever — the connection-manager's inbound-upgrade slot is taken by
- * `acceptIncomingConnection` and released in the `finally` that never runs. The
- * real policy reads the control DB (`listAuthorizedMembers`), which can pull
- * over the network, so "never settles" is reachable. Bounding it here keeps the
- * fail-open contract honest: a slow decision admits rather than silently
- * failing closed (or not at all).
- */
-export const ADMISSION_DECISION_TIMEOUT_MS = 2_000;
-
-/**
- * The admission decision the gate defers to — implemented by
- * `CadreNode.admitInboundControlConnection` (enrollment windows, anchor state,
- * bootstrap infra, the authorized-member set). Kept injectable so the gater's
- * composition/fail-open behavior is unit-testable without a full node.
- */
-export interface InboundAdmissionPolicy {
-  /** Should an inbound encrypted connection from this peer be admitted? */
-  admitInbound(remotePeerId: string): Promise<boolean> | boolean;
-}
-
-/**
- * Build the control node's connection gater: the caller-supplied gater (if any)
- * with membership admission composed onto `denyInboundEncryptedConnection` —
- * the earliest checkpoint where the remote's authenticated PeerId is known.
- *
- * Composition semantics: every hook of `base` is preserved as-is; on the
- * inbound-encrypted hook a deny from EITHER the base gater or the admission
- * policy denies. A policy error — or a decision slower than
- * `decisionTimeoutMs` — admits (fail-open, see module doc); the base gater's
- * verdict is still honored first.
- *
- * NOTE: `base` is spread, so a gater passed as a CLASS INSTANCE would lose its
- * prototype methods; every caller in this repo (and libp2p's own default)
- * supplies a plain object. If a class-based gater ever shows up, delegate
- * per-hook instead of spreading.
- *
- * Deny timing, as observed by the denied dialer: noise negotiates the muxer in
- * the security handshake's early data, so the DIALER's upgrade may complete
- * (its `dial()` resolves) before this receiver-side hook runs. The deny then
- * aborts the receiver's upgrade — the receiver never registers the connection
- * and never creates its muxer, so no protocol can ever be negotiated — and the
- * dialer sees its "open" connection close moments later.
- *
- * Control node only: strand cohort nodes legitimately connect cross-party
- * peers, so `CadreNode` threads the raw configured gater to them unchanged.
- */
-export function createMembershipConnectionGater(
-  policy: InboundAdmissionPolicy,
-  base?: ConnectionGater,
-  decisionTimeoutMs: number = ADMISSION_DECISION_TIMEOUT_MS
-): ConnectionGater {
-  return {
-    ...base,
-    denyInboundEncryptedConnection: async (peerId: PeerId, maConn: MultiaddrConnection): Promise<boolean> => {
-      if (await base?.denyInboundEncryptedConnection?.(peerId, maConn)) {
-        return true;
-      }
-      const remotePeerId = peerId.toString();
-      try {
-        return !(await admitWithinDeadline(policy, remotePeerId, decisionTimeoutMs));
-      } catch (error) {
-        log('admitInbound threw for %s — admitting (fail-open; stream gates decide): %o', remotePeerId, error);
-        return false;
-      }
-    }
-  };
-}
-
-/**
- * Run the admission decision under {@link ADMISSION_DECISION_TIMEOUT_MS},
- * resolving to "admit" if it has not settled in time (see that constant for why
- * an unbounded await is not safe here). The timer is always cleared so a decided
- * connection never holds the event loop open.
- */
-async function admitWithinDeadline(
-  policy: InboundAdmissionPolicy,
-  remotePeerId: string,
-  timeoutMs: number
-): Promise<boolean> {
-  let timer: ReturnType<typeof setTimeout> | undefined;
-  try {
-    return await Promise.race([
-      Promise.resolve(policy.admitInbound(remotePeerId)),
-      new Promise<boolean>((resolve) => {
-        timer = setTimeout(() => {
-          log('admitInbound exceeded %dms for %s — admitting (fail-open; stream gates decide)', timeoutMs, remotePeerId);
-          resolve(true);
-        }, timeoutMs);
-      })
-    ]);
-  } finally {
-    clearTimeout(timer);
-  }
-}
+/**
+ * Control-network inbound admission gates (defense-in-depth): the
+ * encrypted-connection checkpoint and the circuit-relay reservation checkpoint.
+ *
+ * Layer 2 of the membership enforcement chain: the PRIMARY gates are per-stream
+ * — every sensitive sereus control protocol rejects a peer that fails the
+ * voucher-anchored membership check (wake and strand-addr via
+ * `CadreNode.isAuthorizedMember`; the Optimystic control-DB protocols via the
+ * materialized-snapshot gate `CadreNode.authorizeInboundControlStream`), and
+ * the replicated rows an outsider *can* still write are disbelieved at read
+ * time. This module adds the opportunistic connection-level layer on top: a
+ * peer this node can positively determine is NOT authorized is refused at the
+ * encrypted-connection checkpoint, before any protocol negotiation, so a
+ * known-nothing outsider is never even in the conversation — EXCEPT on a node
+ * that runs the circuit-relay server, where the refusal moves to the
+ * relay-reservation checkpoint instead (see "The relay-reservation seam").
+ *
+ * ## The stranger allowlist (the ONE place it is defined)
+ *
+ * A libp2p connection gater decides per CONNECTION, before protocols are
+ * negotiated, so "allow seed, deny repo" cannot be expressed here — instead the
+ * policy admits a connection whenever a legitimate stranger interaction could be
+ * riding it, and the per-stream gates take over. The complete set of protocols a
+ * NOT-yet-authorized peer may legitimately speak on a control node is:
+ *
+ *  - `/sereus/seed/1.0.0` ({@link SEED_PROTOCOL}) — enrollment seed delivery.
+ *    An owner dials a brand-new node to seed it (the new node has no members
+ *    yet, so its gate is inert), and an invited phone dials the owner before it
+ *    is authorized (admitted via the enrollment window `CadreNode.createInvite`
+ *    opens). The handler's own trust decision is the anchored seed-trust policy.
+ *  - `/sereus/formation/1.0.0` ({@link FORMATION_PROTOCOL}) — cross-party
+ *    strand formation via open invitations. Stranger-facing BY DESIGN: the
+ *    initiator is another party, and its token is only checkable inside the
+ *    protocol. The exemption is therefore keyed on EXPECTATION of a stranger,
+ *    not capability to serve one: stranger denial is suspended only while this
+ *    node has at least one UNEXPIRED, NOT-FULLY-CONSUMED open invitation
+ *    outstanding (`StrandSolicitationService.hasOutstandingInvitation` — the
+ *    tokens this process minted or published, plus any still-redeemable
+ *    `FormationInvite` row the usage recorder can see). Merely registering the
+ *    responder (`CadreNode.initializeStrandSolicitation`) does NOT suspend it,
+ *    so an app that registers eagerly at node bring-up keeps a live gate. The
+ *    handler's own trust decision remains the per-token check, which is
+ *    strictly finer than this one: a peer admitted here can still be rejected
+ *    in-protocol for a bogus or spent token.
+ *
+ * There is one further connection-level carve-out, which is NOT a protocol
+ * exemption and needs no stranger window:
+ *
+ *  - **Announced delegate peers** (`delegate-admission.ts`). A member's strand
+ *    node runs as a separate libp2p identity whose peerId no sibling can
+ *    recompute. Before starting a strand node, a member's control node
+ *    announces that peerId over the already-authenticated strand-addr RPC, and
+ *    the receiver holds a short-lived grant for it
+ *    (`CadreNode.grantDelegateAdmission`). The grant admits the CONNECTION,
+ *    admits the peer's RESERVATION at the seam below (without spending the
+ *    unauthorized budget), and nothing else — it is deliberately invisible to
+ *    the per-stream gate, so a delegate still gets refused on every
+ *    members-only protocol.
+ *
+ * ## The relay-reservation seam
+ *
+ * A circuit-relay reservation is established by the reserving peer DIALING the
+ * relay, so at the relay it is an inbound connection — and killing that
+ * connection kills the reservation. For a genuine member whose `CadrePeer` row
+ * has not yet replicated to this node, a connection-level deny here is NOT
+ * self-healing the way a data connection's is: an outbound reconcile dial
+ * re-establishes a data link, but no outbound dial can grant the REMOTE peer a
+ * reservation, and a relay-only peer has no address of its own to dial back —
+ * the reservation IS its address. Boot ordering makes the window ordinary (a
+ * node's circuit listener runs inside `libp2p.start()`, before its own row
+ * could have replicated anywhere), and a stalled replication makes it
+ * unbounded. So the relay would be answering a RELAY question ("may this peer
+ * use my forwarding capacity?") with a MEMBERSHIP answer, at a checkpoint where
+ * a wrong answer is unrecoverable.
+ *
+ * On a node whose relay server is enabled the two questions are separated:
+ *
+ *  - The policy returns `'admit-for-relay'` instead of `'deny'`, and the
+ *    connection is ADMITTED. The fail-closed per-stream gates still refuse such
+ *    a peer every members-only protocol, so it gains identify/ping and the
+ *    relay hop protocol, nothing else.
+ *  - The reservation itself is decided at libp2p's
+ *    `denyInboundRelayReservation` hook (the circuit-relay server consults it
+ *    per RESERVE request): the policy admits members, delegates and configured
+ *    infra outright, and admits peers it cannot place only within a bounded
+ *    budget ({@link UnauthorizedReservationBudget}) — a member whose row is in
+ *    flight always finds a slot under any sane cap, while outsiders cannot
+ *    annex the party's relay capacity.
+ *  - An `'admit-for-relay'` connection that is NOT reserving is dropped: it has
+ *    {@link RELAY_ADMISSION_RESERVE_DEADLINE_MS} to get a reservation ADMITTED
+ *    at that hook, after which the underlying connection is aborted. The
+ *    guarantee above thus weakens on relay-enabled nodes from "never in the
+ *    conversation" to "in it briefly, and can speak nothing".
+ *
+ * The deadline clears on reservation ADMISSION, not on the reservation's own
+ * success: this gate cannot observe the server-side `reserve()` outcome, so a
+ * peer whose admitted reservation is then refused for server capacity keeps its
+ * connection until either side closes it — bounded and mute, so harmless.
+ *
+ * Everything else a control node handles — the Optimystic control-DB protocols
+ * (`/optimystic/control-<party>/{repo,cluster,sync,block-transfer}/…`), wake
+ * (`/sereus/strand-wake/1.0.0`), and strand-addr (`/sereus/strand-addr/1.0.0`)
+ * — is members-only, and every one of them enforces that per-stream. Wake and
+ * strand-addr check `isAuthorizedMember` inside their handlers; the four
+ * control-DB protocols are gated by the fail-closed
+ * `CadreNode.authorizeInboundControlStream` (wired as libp2p's
+ * `authorizeInboundStream` upstream hook), which judges each inbound stream
+ * against the MATERIALIZED authorized-peer snapshot — synchronous and
+ * in-memory, because a live DB read from inside the gate would deadlock into
+ * mutual denial. The two layers complement, not duplicate: this connection
+ * gate is fail-open over a live DB read (deny only on positive proof of an
+ * outsider), the stream gate is fail-closed over the snapshot and has NO
+ * stranger carve-outs — an enrollment window admits a stranger's connection
+ * for seed delivery, yet its repo streams are still refused.
+ *
+ * ## The bring-up quiet period
+ *
+ * One state is decided BEFORE any of the above, and is not about the remote peer
+ * at all: while this node's control-database bring-up is in flight
+ * (`InboundAdmissionPolicy.bringUpInFlight`), the gate denies BOTH directions —
+ * `denyDialPeer` and `denyInboundEncryptedConnection` — and then opens.
+ *
+ * The invariant it protects is "`ControlDatabase.initialize()` runs while this
+ * node holds ZERO control connections". Building that database is a long chain
+ * of cohort-consulting block probes, and every connected same-party sibling is in
+ * the cohort those probes consult. A sibling that has not yet replicated this
+ * node's `CadrePeer` row correctly refuses them at its own fail-closed per-stream
+ * gate, so ONE connection in this window turns bring-up into
+ * `BlockUnavailableError` and `start()` rejects. Retrying cannot converge: the
+ * condition that would clear the refusal is this node's own row reaching the
+ * sibling, and writing that row needs the database the retry is building. So the
+ * only fix is to not be in the conversation yet — which is what this window is.
+ *
+ * The ordering is arranged so nothing SHOULD open a connection here anyway:
+ * `network.relayAddrs` resolves to a listener that dials nothing and reserves
+ * after bring-up (`relay-addrs.ts`), and the control-cohort reconcile pass is
+ * scheduled post-start. This window is what makes that a property rather than an
+ * accident of ordering — the live case it actually catches is
+ * `controlNetwork.bootstrapNodes`, where `@libp2p/bootstrap` emits its discovery
+ * events ~1 s after `libp2p.start()` and the connection manager auto-dials from
+ * there, which is a race bring-up wins only while raw-storage latency stays low.
+ *
+ * A denial here costs a retry, not a partition: libp2p's connection manager
+ * re-dials on its auto-dial cadence, the reservation supervisor re-drives, and a
+ * denied inbound peer reconnects. The window opens on bring-up FAILURE too (via
+ * `CadreNode.cleanup`), so teardown is never gated.
+ *
+ * ## Fail-open, deliberately
+ *
+ * Outside that window this layer only ever denies on a POSITIVE determination of
+ * "unauthorized outsider while no stranger path is open". Any error, missing
+ * dependency, or ambiguous state admits the connection (and the reservation) and
+ * defers to the fail-closed stream gates — a DB hiccup must not partition a
+ * legitimate cadre.
+ */
+
+import debug from 'debug';
+import type { ConnectionGater, PeerId, MultiaddrConnection } from '@libp2p/interface';
+import { SEED_PROTOCOL } from './seed-bootstrap.js';
+import { FORMATION_PROTOCOL } from './strand-formation-protocol.js';
+
+const log = debug('sereus:cadre:connection-gater');
+
+/**
+ * The protocols a not-yet-authorized peer may legitimately speak on a control
+ * node (see the module doc above for why each is open and where its own trust
+ * decision lives). Exported as the single reference point so the allowlist
+ * cannot silently drift across modules.
+ */
+export const STRANGER_OPEN_PROTOCOLS: readonly string[] = [SEED_PROTOCOL, FORMATION_PROTOCOL];
+
+/**
+ * How long `CadreNode.createInvite` holds the inbound gate open for strangers
+ * when the invite carries no explicit expiry: the invitee must dial in before
+ * it is authorized, and an expiry-less invite gives the gate no bound of its
+ * own. Redeeming after this window still works once the owner re-mints (or the
+ * caller re-opens the window via `CadreNode.openEnrollmentWindow`).
+ */
+export const DEFAULT_ENROLLMENT_WINDOW_MS = 30 * 60 * 1000;
+
+/**
+ * Deadline for one admission decision, after which the fail-open outcome is
+ * used (connection admitted / reservation admitted).
+ *
+ * libp2p awaits `denyInboundEncryptedConnection` inside the inbound upgrade
+ * WITHOUT racing its inbound-upgrade timeout signal (unlike the pre-encryption
+ * `denyInboundConnection` hook), so a decision that never settles wedges that
+ * upgrade forever — the connection-manager's inbound-upgrade slot is taken by
+ * `acceptIncomingConnection` and released in the `finally` that never runs. The
+ * real policy reads the control DB (`listAuthorizedMembers`), which can pull
+ * over the network, so "never settles" is reachable. Bounding it here keeps the
+ * fail-open contract honest: a slow decision admits rather than silently
+ * failing closed (or not at all).
+ */
+export const ADMISSION_DECISION_TIMEOUT_MS = 2_000;
+
+/**
+ * How long an `'admit-for-relay'` connection may exist without a relay
+ * reservation being ADMITTED at the `denyInboundRelayReservation` hook, after
+ * which the gate aborts the underlying connection. A reserving client asks for
+ * its slot immediately after the connection upgrades (`relay-reservation.ts`
+ * dials and requests in one drive; a strand node's configured circuit listener
+ * does the same from inside its own `listen()`), so a connection idle past this
+ * deadline is not reserving.
+ */
+export const RELAY_ADMISSION_RESERVE_DEADLINE_MS = 5_000;
+
+/**
+ * Default cap on concurrent relay reservations held by peers the membership
+ * check could not place (see {@link UnauthorizedReservationBudget}). Small on
+ * purpose: it exists for the handful of genuine members whose rows are still
+ * in flight, not as public relay capacity. Overridable per node via
+ * `network.unauthorizedRelayReservationCap` (0 refuses every unauthorized
+ * reservation — the strict pre-seam posture).
+ *
+ * NOTE: this cap shares the relay server's own reservation store, whose default
+ * size is 15 (`@libp2p/circuit-relay-v2`'s `DEFAULT_MAX_RESERVATION_STORE_SIZE`);
+ * unplaced peers may therefore occupy up to this many of those slots, and the
+ * rest are what members and delegates compete for. This module cannot read the
+ * server's size, so the two are kept apart by hand — if either is raised, keep
+ * this one well under the store's, or a fleet of unplaceable peers can crowd
+ * genuine members out of the server's own store (which the gate cannot override).
+ */
+export const MAX_UNAUTHORIZED_RELAY_RESERVATIONS = 8;
+
+/**
+ * How long one {@link UnauthorizedReservationBudget} entry occupies a slot
+ * without a refresh. Mirrors the relay server's own default reservation TTL
+ * (`@libp2p/circuit-relay-v2`'s `DEFAULT_MAX_RESERVATION_TTL`, 2 h): the server
+ * holds an unrefreshed reservation exactly that long, and a live reserver
+ * re-requests (re-hitting the admission hook, refreshing its entry) well before
+ * expiry — so the live entry count tracks the server's own occupancy without
+ * this module reaching into the server's reservation store.
+ *
+ * NOTE: a mirror, not a live coupling — if the relay server is ever configured
+ * with a non-default `reservationTtl` (db-p2p's `relayServerInit`), update this
+ * constant alongside it or the budget's occupancy drifts from the server's.
+ */
+export const UNAUTHORIZED_RESERVATION_TTL_MS = 2 * 60 * 60 * 1000;
+
+/**
+ * The connection-level outcome of the admission policy:
+ *  - `'admit'` — a peer with a legitimate claim on the connection (member,
+ *    delegate, infra, open stranger window, or any fail-open state).
+ *  - `'deny'` — positively an outsider, and this node runs no relay server, so
+ *    the connection can carry nothing legitimate.
+ *  - `'admit-for-relay'` — positively an outsider (or an unreplicated member —
+ *    indistinguishable), but this node runs the relay server: admit the
+ *    connection so a reservation can be asked for, decide at the reservation
+ *    seam, and drop the connection if no reservation is admitted in time.
+ */
+export type InboundConnectionVerdict = 'admit' | 'deny' | 'admit-for-relay';
+
+/**
+ * The admission decisions the gate defers to — implemented by
+ * `CadreNode.admitInboundControlConnection` /
+ * `CadreNode.admitControlRelayReservation` (enrollment windows, anchor state,
+ * bootstrap infra, delegate grants, the authorized-member set, the
+ * unauthorized-reservation budget). Kept injectable so the gater's
+ * composition/fail-open behavior is unit-testable without a full node.
+ */
+export interface InboundAdmissionPolicy {
+  /** Verdict on an inbound encrypted connection from this peer. */
+  admitInbound(remotePeerId: string): Promise<InboundConnectionVerdict> | InboundConnectionVerdict;
+  /** Should this peer be granted a circuit-relay reservation slot? */
+  admitRelayReservation(remotePeerId: string): Promise<boolean> | boolean;
+  /**
+   * True while this node's control-database bring-up is in flight — the
+   * {@link createMembershipConnectionGater} quiet period (see the module doc).
+   * Peer-independent, unlike the two decisions above, and SYNCHRONOUS: it is
+   * read on the outbound-dial path, where an await would be a new stall.
+   *
+   * Optional — a policy that omits it is never quiet, which is the right default
+   * for every caller that is not a booting `CadreNode`.
+   */
+  bringUpInFlight?(): boolean;
+}
+
+/**
+ * Bounded budget of concurrent relay reservations for peers the membership
+ * check could not place. `tryAdmit` is the whole surface: an already-admitted
+ * peer refreshes its entry (a reservation refresh never double-counts), a new
+ * peer takes a free slot or is refused. Entries expire after `ttlMs` (see
+ * {@link UNAUTHORIZED_RESERVATION_TTL_MS} for why that mirrors the relay
+ * server's own hold). Injectable `now` keeps expiry testable without fake
+ * timers. Authorized members and delegates are never run through this — the
+ * policy admits them before consulting the budget, and `release`s the slot such
+ * a peer took while it was still unplaceable, so the boot-ordering window a
+ * member passes through costs the budget nothing once its row lands.
+ */
+export class UnauthorizedReservationBudget {
+  private readonly admitted = new Map<string, number>();
+
+  constructor(
+    private readonly cap: number = MAX_UNAUTHORIZED_RELAY_RESERVATIONS,
+    private readonly ttlMs: number = UNAUTHORIZED_RESERVATION_TTL_MS
+  ) {}
+
+  /** Number of live (unexpired at last prune) entries — test/diagnostic surface. */
+  get size(): number {
+    return this.admitted.size;
+  }
+
+  /** Admit (or refresh) `remotePeerId` if a slot is free; false when the cap is spent. */
+  tryAdmit(remotePeerId: string, now: number = Date.now()): boolean {
+    this.prune(now);
+    if (!this.admitted.has(remotePeerId) && this.admitted.size >= this.cap) {
+      log('Unauthorized-reservation budget spent (%d/%d) — refusing %s', this.admitted.size, this.cap, remotePeerId);
+      return false;
+    }
+    this.admitted.set(remotePeerId, now + this.ttlMs);
+    return true;
+  }
+
+  /**
+   * Give back the slot `remotePeerId` holds, if any — called when the peer has
+   * become admissible on its own merits (its membership row replicated, or a
+   * delegate grant landed), so the slot it took during the boot-ordering window
+   * does not stay spent for the remaining TTL on a peer that no longer needs it.
+   */
+  release(remotePeerId: string): void {
+    this.admitted.delete(remotePeerId);
+  }
+
+  private prune(now: number): void {
+    for (const [peerId, expiresAt] of this.admitted) {
+      if (expiresAt <= now) {
+        this.admitted.delete(peerId);
+      }
+    }
+  }
+}
+
+/**
+ * Build the control node's connection gater: the caller-supplied gater (if any)
+ * with membership admission composed onto `denyInboundEncryptedConnection` —
+ * the earliest checkpoint where the remote's authenticated PeerId is known —
+ * and reservation admission composed onto `denyInboundRelayReservation` — the
+ * hook the circuit-relay server consults per RESERVE request (inert on a node
+ * whose relay server is off; libp2p never calls it there).
+ *
+ * `denyDialPeer` is composed too, but ONLY for the bring-up quiet period (see
+ * the module doc): outside that window outbound dials are never gated by
+ * membership — this node decides for itself who to talk to.
+ *
+ * Composition semantics: every hook of `base` is preserved as-is; on the three
+ * composed hooks a deny from EITHER the base gater or the admission policy
+ * denies. A policy error — or a decision slower than `decisionTimeoutMs` —
+ * takes the fail-open outcome (connection admitted / reservation admitted, see
+ * module doc); the base gater's verdict is still honored first.
+ *
+ * An `'admit-for-relay'` verdict admits the connection and arms a
+ * `reserveDeadlineMs` timer against it; the timer is disarmed when a
+ * reservation for that peer is ADMITTED at the reservation hook, and aborts the
+ * underlying `MultiaddrConnection` when it fires first.
+ *
+ * NOTE: `base` is spread, so a gater passed as a CLASS INSTANCE would lose its
+ * prototype methods; every caller in this repo (and libp2p's own default)
+ * supplies a plain object. If a class-based gater ever shows up, delegate
+ * per-hook instead of spreading.
+ *
+ * Deny timing, as observed by the denied dialer: noise negotiates the muxer in
+ * the security handshake's early data, so the DIALER's upgrade may complete
+ * (its `dial()` resolves) before this receiver-side hook runs. The deny then
+ * aborts the receiver's upgrade — the receiver never registers the connection
+ * and never creates its muxer, so no protocol can ever be negotiated — and the
+ * dialer sees its "open" connection close moments later.
+ *
+ * Control node only: strand cohort nodes legitimately connect cross-party
+ * peers, so `CadreNode` threads the raw configured gater to them unchanged.
+ */
+export function createMembershipConnectionGater(
+  policy: InboundAdmissionPolicy,
+  base?: ConnectionGater,
+  decisionTimeoutMs: number = ADMISSION_DECISION_TIMEOUT_MS,
+  reserveDeadlineMs: number = RELAY_ADMISSION_RESERVE_DEADLINE_MS
+): ConnectionGater {
+  const pendingReservations = new PendingReserveDeadlines(reserveDeadlineMs);
+  const quiet = (): boolean => policy.bringUpInFlight?.() ?? false;
+  return {
+    ...base,
+    denyDialPeer: async (peerId: PeerId): Promise<boolean> => {
+      if (await base?.denyDialPeer?.(peerId)) {
+        return true;
+      }
+      if (quiet()) {
+        log('Control-database bring-up in flight — refusing the outbound dial to %s', peerId.toString());
+        return true;
+      }
+      return false;
+    },
+    denyInboundEncryptedConnection: async (peerId: PeerId, maConn: MultiaddrConnection): Promise<boolean> => {
+      if (await base?.denyInboundEncryptedConnection?.(peerId, maConn)) {
+        return true;
+      }
+      if (quiet()) {
+        log('Control-database bring-up in flight — refusing the inbound connection from %s', peerId.toString());
+        return true;
+      }
+      const remotePeerId = peerId.toString();
+      let verdict: InboundConnectionVerdict;
+      try {
+        verdict = await decideWithinDeadline(
+          () => policy.admitInbound(remotePeerId), 'admit', decisionTimeoutMs, `admitInbound(${remotePeerId})`
+        );
+      } catch (error) {
+        log('admitInbound threw for %s — admitting (fail-open; stream gates decide): %o', remotePeerId, error);
+        return false;
+      }
+      if (verdict === 'admit-for-relay') {
+        pendingReservations.arm(remotePeerId, maConn);
+        return false;
+      }
+      return verdict === 'deny';
+    },
+    denyInboundRelayReservation: async (peerId: PeerId): Promise<boolean> => {
+      if (await base?.denyInboundRelayReservation?.(peerId)) {
+        return true;
+      }
+      const remotePeerId = peerId.toString();
+      let admitted: boolean;
+      try {
+        admitted = await decideWithinDeadline(
+          () => policy.admitRelayReservation(remotePeerId), true, decisionTimeoutMs, `admitRelayReservation(${remotePeerId})`
+        );
+      } catch (error) {
+        log('admitRelayReservation threw for %s — admitting (fail-open): %o', remotePeerId, error);
+        admitted = true;
+      }
+      if (admitted) {
+        pendingReservations.disarm(remotePeerId);
+      }
+      return !admitted;
+    }
+  };
+}
+
+/**
+ * Run one admission decision under `timeoutMs`, resolving to `fallback` (the
+ * fail-open outcome) if it has not settled in time — see
+ * {@link ADMISSION_DECISION_TIMEOUT_MS} for why an unbounded await is not safe
+ * on the connection hook. The timer is always cleared so a decided call never
+ * holds the event loop open.
+ */
+async function decideWithinDeadline<T>(
+  decide: () => Promise<T> | T,
+  fallback: T,
+  timeoutMs: number,
+  what: string
+): Promise<T> {
+  let timer: ReturnType<typeof setTimeout> | undefined;
+  try {
+    return await Promise.race([
+      Promise.resolve().then(decide),
+      new Promise<T>((resolve) => {
+        timer = setTimeout(() => {
+          log('%s exceeded %dms — taking the fail-open outcome (stream gates decide)', what, timeoutMs);
+          resolve(fallback);
+        }, timeoutMs);
+      })
+    ]);
+  } finally {
+    clearTimeout(timer);
+  }
+}
+
+/** One armed not-reserving deadline: the connection it will abort, and its timer. */
+interface PendingReserveDeadline {
+  maConn: MultiaddrConnection;
+  timer: ReturnType<typeof setTimeout>;
+}
+
+/**
+ * The not-reserving deadlines for `'admit-for-relay'` connections, keyed by
+ * remote peerId (a Set per peer — one peer can hold several in-flight
+ * connections). `arm` starts a timer that aborts the connection; `disarm`
+ * (called when a reservation for that peer is admitted) cancels every pending
+ * timer for the peer. Timers are unref'd so an armed deadline never holds a
+ * process open, and a timer that fires against an already-closed connection is
+ * a swallowed no-op.
+ *
+ * Disarming is final for the connections it cancelled: a peer that reserved
+ * once and then lets its reservation lapse keeps a mute connection (the
+ * fail-closed stream gates refuse it every members-only protocol) until either
+ * side closes it. Bounded — only a peer the reservation policy already admitted
+ * can reach that state, so unplaceable peers are bounded by the budget cap.
+ */
+class PendingReserveDeadlines {
+  private readonly byPeer = new Map<string, Set<PendingReserveDeadline>>();
+
+  constructor(private readonly deadlineMs: number) {}
+
+  arm(remotePeerId: string, maConn: MultiaddrConnection): void {
+    const entry: PendingReserveDeadline = {
+      maConn,
+      timer: setTimeout(() => this.expire(remotePeerId, entry), this.deadlineMs)
+    };
+    (entry.timer as { unref?: () => void }).unref?.();
+    let entries = this.byPeer.get(remotePeerId);
+    if (!entries) {
+      entries = new Set();
+      this.byPeer.set(remotePeerId, entries);
+    }
+    entries.add(entry);
+    log('Admitted %s for relay only — dropping the connection unless a reservation is admitted within %dms', remotePeerId, this.deadlineMs);
+  }
+
+  disarm(remotePeerId: string): void {
+    const entries = this.byPeer.get(remotePeerId);
+    if (!entries) {
+      return;
+    }
+    this.byPeer.delete(remotePeerId);
+    for (const entry of entries) {
+      clearTimeout(entry.timer);
+    }
+  }
+
+  private expire(remotePeerId: string, entry: PendingReserveDeadline): void {
+    const entries = this.byPeer.get(remotePeerId);
+    entries?.delete(entry);
+    if (entries?.size === 0) {
+      this.byPeer.delete(remotePeerId);
+    }
+    log('Relay-only admission expired for %s — no reservation admitted within %dms, aborting the connection', remotePeerId, this.deadlineMs);
+    try {
+      entry.maConn.abort(new Error(`relay-only admission expired: no relay reservation admitted within ${this.deadlineMs}ms`));
+    } catch (error) {
+      // Aborting a connection that already closed on its own is not a failure.
+      log('Aborting the expired relay-only connection from %s threw: %o', remotePeerId, error);
+    }
+  }
+}