feat(relay): rendezvous-relay service — 7 packages + plans (contracts/transport/agent/control-plane/e2e/auth/web)

Multi-tenant reverse-tunnel service ("ngrok for Claude Code" with E2E): a
host-agent dials OUT to an operator-run relay; external devices reach the host
THROUGH the relay, routed by per-tenant subdomain, forwarding ciphertext only
(the relay never sees plaintext). Lets a customer reach their own self-hosted
web-terminal from anywhere with zero networking setup.

Packages — all tsc-strict + vitest green (656 tests), cross-package integration verified:
- relay-contracts: frozen shared contracts (mux frame codec, data model,
  capability token, E2E envelope, pairing) — the src/types.ts analog
- term-relay:   native WS mux + stateless data plane (subdomain routing, ciphertext forward)
- agent:        host-agent (pairing, per-host Ed25519 + mTLS dial-out, forwards to 127.0.0.1:3000)
- control-plane: accounts/hosts registry, pairing-code flow, routing table, provisioning
- relay-e2e:    browser<->agent E2E (X25519 ECDH through relay, AEAD, anti-replay, recoverable replay key)
- relay-auth:   Passkey/WebAuthn, capability tokens, per-host certs, deny-by-default tenant isolation
- relay-web:    browser login + Web Crypto E2E + client-side preview rendering

Security invariants INV1-15 enforced; cross-tenant isolation CI tripwire live
(.github/workflows/relay-tripwire.yml). Design + implementation-level plans in
docs/PLAN_RELAY_*.md and docs/EXPLORE_RELAY_SERVICE.md.

NOTE: generated autonomously per the reviewed plans. The security-critical
packages (relay-e2e, relay-auth) REQUIRE expert security audit before any real
deployment — passing tests prove self-consistency, not resistance to attackers.
Base app (src/, public/) unchanged; concurrent desktop work left uncommitted.

control-plane/src/routing/bus.ts

@@ -0,0 +1,49 @@
+/**
+ * RevocationBus implementations over the FROZEN `relay:revocations` channel (INDEX §4.2 FIX 4).
+ * P3 only PUBLISHES (P1 owns the subscriber that injects §4.1 CLOSE+RST / GOAWAY). Drain (T10)
+ * and revoke (T13) use the SAME named channel. The KillSignal shape + channel name are imported
+ * from relay-contracts and NEVER redefined.
+ */
+import {
+  RELAY_REVOCATIONS_CHANNEL,
+  KillSignalSchema,
+  type KillSignal,
+  type RevocationBus,
+} from 'relay-contracts'
+
+/** In-memory bus with a subscribe hook — used by tests and single-process wiring. */
+export interface TestableRevocationBus extends RevocationBus {
+  subscribe(handler: (signal: KillSignal) => void): () => void
+  readonly channel: typeof RELAY_REVOCATIONS_CHANNEL
+}
+
+export function createInMemoryRevocationBus(): TestableRevocationBus {
+  const handlers = new Set<(signal: KillSignal) => void>()
+  return {
+    channel: RELAY_REVOCATIONS_CHANNEL,
+    async publish(signal) {
+      // Validate at the boundary before it hits the wire (INV10: reason is metadata only).
+      const parsed = KillSignalSchema.parse(signal)
+      for (const h of handlers) h(parsed as KillSignal)
+    },
+    subscribe(handler) {
+      handlers.add(handler)
+      return () => handlers.delete(handler)
+    },
+  }
+}
+
+/** Minimal publisher surface of an ioredis client (integration seam; typed, not unit-tested). */
+export interface RedisPublisher {
+  publish(channel: string, message: string): Promise<number>
+}
+
+/** Redis-backed publisher (production). P1 relay nodes SUBSCRIBE to the same channel. */
+export function createRedisRevocationBus(redis: RedisPublisher): RevocationBus {
+  return {
+    async publish(signal) {
+      const parsed = KillSignalSchema.parse(signal)
+      await redis.publish(RELAY_REVOCATIONS_CHANNEL, JSON.stringify(parsed))
+    },
+  }
+}

control-plane/src/routing/nodes.ts

@@ -0,0 +1,67 @@
+/**
+ * T10 — relay-node registry + graceful drain (INV7). Every mutation is scoped to the caller's
+ * authenticated `NodeIdentity`: a node can only register/heartbeat/drain ITSELF. Drain marks the
+ * node 'draining', enumerates its live routes, and PUBLISHES a per-host `KillSignal` on the frozen
+ * `relay:revocations` bus (INDEX §4.2 FIX 4 — SAME channel T13 revoke uses); P1 nodes translate it
+ * into the §4.1 GOAWAY. This module does NOT encode the frame (that is P1). Drain touches NO
+ * Postgres host/session row — the running task lives on the customer machine (PTY survives).
+ *
+ * OQ4 residual: the frozen RevocationScope has no `node` kind, so a single-node operator drain is
+ * expressed as per-host KillSignals with a drain reason (flagged to the INDEX for a future scope).
+ */
+import type { RevocationBus } from 'relay-contracts'
+import type { NodeStore, RouteStore } from '../store/ports.js'
+import type { NodeIdentity } from '../node-auth/identity.js'
+import type { AuditWriter } from '../audit/log.js'
+import { noopAuditWriter } from '../audit/log.js'
+import { nowIso } from '../util/ids.js'
+
+export const DRAIN_REASON = 'operatorDrain'
+
+export interface NodeCoordinator {
+  registerNode(caller: NodeIdentity, addr: string): Promise<void>
+  nodeHeartbeat(caller: NodeIdentity): Promise<void>
+  beginDrain(caller: NodeIdentity): Promise<{ readonly hostIds: readonly string[] }>
+  completeDrain(caller: NodeIdentity): Promise<void>
+}
+
+export interface NodeCoordinatorDeps {
+  readonly nodes: NodeStore
+  readonly routes: RouteStore
+  readonly bus: RevocationBus
+  readonly audit?: AuditWriter
+}
+
+export function createNodeCoordinator(deps: NodeCoordinatorDeps): NodeCoordinator {
+  const audit = deps.audit ?? noopAuditWriter()
+  return {
+    async registerNode(caller, addr) {
+      await deps.nodes.upsert({ nodeId: caller.nodeId, addr, status: 'active', lastSeen: nowIso() })
+    },
+    async nodeHeartbeat(caller) {
+      await deps.nodes.touch(caller.nodeId, nowIso())
+    },
+    async beginDrain(caller) {
+      await deps.nodes.setStatus(caller.nodeId, 'draining')
+      const routed = await deps.routes.listByNode(caller.nodeId)
+      const at = Math.floor(Date.now() / 1000)
+      for (const { hostId } of routed) {
+        // Per-host KillSignal with a drain reason on the frozen bus (P1 emits GOAWAY).
+        // eslint-disable-next-line no-await-in-loop
+        await deps.bus.publish({ scope: { kind: 'host', hostId }, at, reason: DRAIN_REASON })
+      }
+      await audit.writeAuditEvent({
+        action: 'node.drain',
+        principalId: `node:${caller.nodeId}`,
+        accountId: 'system',
+        hostId: null,
+        ts: nowIso(),
+        meta: { nodeId: caller.nodeId, routes: String(routed.length) },
+      })
+      return { hostIds: routed.map((r) => r.hostId) }
+    },
+    async completeDrain(caller) {
+      await deps.nodes.delete(caller.nodeId) // deregister after routes re-homed
+    },
+  }
+}

control-plane/src/routing/table.ts

@@ -0,0 +1,62 @@
+/**
+ * T9 — Redis routing table (`route:{host_id}`, heartbeat-TTL, INV7). Redis holds only LOCATION,
+ * never ownership or secrets; a missing/expired key fails CLOSED (host treated offline). Every
+ * mutation is scoped to the AUTHENTICATED node identity: a node can only route hosts to ITSELF
+ * (`entry.relayNodeId === caller.nodeId`) and can only heartbeat/drop a route it currently holds.
+ */
+import type { RouteEntry } from '../model/records.js'
+import type { NodeStatus, RouteStore } from '../store/ports.js'
+import type { NodeIdentity } from '../node-auth/identity.js'
+
+export class RouteAuthError extends Error {}
+
+export interface RoutingTable {
+  upsertRoute(caller: NodeIdentity, hostId: string, entry: RouteEntry, ttlSec: number): Promise<void>
+  heartbeatRoute(caller: NodeIdentity, hostId: string, ttlSec: number): Promise<void>
+  resolveRoute(hostId: string): Promise<RouteEntry | null>
+  dropRoute(caller: NodeIdentity, hostId: string): Promise<void>
+  /** System teardown path (drain/revoke) — bypasses the holding-node check by design. */
+  systemDropRoute(hostId: string): Promise<void>
+}
+
+export interface RoutingTableDeps {
+  readonly routes: RouteStore
+  /** Optional node-status hook: a draining node stops receiving new upserts (T10). */
+  readonly nodeStatus?: (nodeId: string) => Promise<NodeStatus | null>
+}
+
+export function createRoutingTable(deps: RoutingTableDeps): RoutingTable {
+  return {
+    async upsertRoute(caller, hostId, entry, ttlSec) {
+      // A node may ONLY route hosts to itself — never inject a route for another node's id.
+      if (entry.relayNodeId !== caller.nodeId) {
+        throw new RouteAuthError('entry.relayNodeId must equal the authenticated caller nodeId')
+      }
+      if (deps.nodeStatus !== undefined) {
+        const status = await deps.nodeStatus(caller.nodeId)
+        if (status === 'draining') throw new RouteAuthError('node is draining; not accepting new routes')
+      }
+      await deps.routes.set(hostId, entry, ttlSec)
+    },
+    async heartbeatRoute(caller, hostId, ttlSec) {
+      const current = await deps.routes.get(hostId)
+      // A stale/foreign node cannot refresh (hijack) another node's live tunnel; missing ⇒ no-op.
+      if (current === null || current.relayNodeId !== caller.nodeId) return
+      await deps.routes.refreshTtl(hostId, ttlSec)
+    },
+    async resolveRoute(hostId) {
+      return deps.routes.get(hostId) // null ⇒ offline (fails closed, INV7)
+    },
+    async dropRoute(caller, hostId) {
+      const current = await deps.routes.get(hostId)
+      if (current === null) return
+      if (current.relayNodeId !== caller.nodeId) {
+        throw new RouteAuthError('only the holding node may drop its route')
+      }
+      await deps.routes.delete(hostId)
+    },
+    async systemDropRoute(hostId) {
+      await deps.routes.delete(hostId)
+    },
+  }
+}