@optimystic/db-p2p 1.0.0-beta.3 → 1.1.0

package/src/testing/mesh-harness.ts

@@ -1,8 +1,9 @@
 import type { PeerId, PrivateKey } from '@libp2p/interface';
-import type { IKeyNetwork, ClusterPeers, ICluster, ClusterRecord, IRepo, BlockId, ActionRev, ITransactor, ITransactionValidator, PeerId as DbPeerId } from '@optimystic/db-core';
+import type { IKeyNetwork, ClusterPeers, ICluster, ClusterRecord, IRepo, BlockId, ITransactor, ITransactionValidator, PeerId as DbPeerId, RoutingKey } from '@optimystic/db-core';
 import type { FindCoordinatorOptions } from '@optimystic/db-core';
 import type { IPeerNetwork } from '@optimystic/db-core';
-import { NetworkTransactor } from '@optimystic/db-core';
+import { NetworkTransactor, routingKeyForBlock } from '@optimystic/db-core';
+import { DigitreeStore, assembleCohort, hashKey, hashPeerId } from 'p2p-fret';
 import { peerIdFromPrivateKey } from '@libp2p/peer-id';
 import { generateKeyPair, generateKeyPairFromSeed } from '@libp2p/crypto/keys';
 import { ClusterMember, clusterMember, type ReconcileBlockCallback, type DeriveExpectedClusterCallback, type ExpectedClusterView } from '../cluster/cluster-repo.js';
@@ -15,13 +16,19 @@ import type { IRawStorage } from '../storage/i-raw-storage.js';
 import type { BlockArchive } from '../storage/struct.js';
 import { serveBlockArchive, servableProof } from '../storage/block-archive.js';
 import { coordinatorRepo, type ClusterLatestCallback, type CertifiedActionRev } from '../repo/coordinator-repo.js';
+import type { CommittedHolders } from '../cluster/rebalance-monitor.js';
 import type { CoordinatorRepo } from '../repo/coordinator-repo.js';
-import { sortPeersByDistance, type KnownPeer } from '../routing/responsibility.js';
 import { toString as u8ToString } from 'uint8arrays';
 
 export interface MeshNode {
 	peerId: PeerId;
 	privateKey: PrivateKey;
+	/**
+	 * This node's own view of the key network — the harness analogue of a real node's `keyNetwork`
+	 * attachment, and the instance its coordinator and its member's admission derivation both read. It
+	 * differs from `Mesh.keyNetwork` only under a simulated partition (`MeshFailureConfig.partitionSides`).
+	 */
+	keyNetwork: IKeyNetwork;
 	storageRepo: StorageRepo;
 	clusterMember: ClusterMember;
 	coordinatorRepo: CoordinatorRepo;
@@ -58,9 +65,9 @@ export interface MeshOptions {
 	/**
 	 * Per-node member-side cluster derivation for the membership admission gate — the harness
 	 * analogue of `libp2p-node-base`'s `deriveExpectedCluster` (findCluster + FRET confidence).
-	 * Omitted → each member gets the production-shaped derivation over its own self-including
-	 * key-network view (partition-aware, see `MeshFailureConfig.partitionSides`) with confidence
-	 * from `meshConfidence` (default 1).
+	 * Omitted → each member gets the production-shaped derivation over its own key-network view
+	 * (partition-aware, see `MeshFailureConfig.partitionSides`) with confidence from `meshConfidence`
+	 * (default 1).
 	 */
 	deriveExpectedCluster?: (node: MeshNode, blockId: BlockId) => Promise<ExpectedClusterView>;
 	/**
@@ -79,7 +86,8 @@ export interface MeshOptions {
 	 *
 	 * A mesh that arms validators can also set `clusterPolicy.unvalidatablePendPolicy: 'reject'` to
 	 * refuse pends that carry no `validation` payload (the single-collection `Collection.sync`
-	 * shape, which a validator cannot re-check); the default 'accept' admits them unchecked.
+	 * shape, and the Quereus adapter's legacy multi-tree commit — neither can be re-checked); the
+	 * default 'accept' admits them unchecked.
 	 */
 	validatorFactory?: (index: number, peerId: PeerId) => ITransactionValidator;
 	/**
@@ -104,6 +112,13 @@ export interface MeshOptions {
 	 * eventually crosses it. Omitted → random keys.
 	 */
 	keySeed?: number;
+	/**
+	 * Receives each node's committed-holders reports — from its member after a durable consensus apply
+	 * and from its coordinator on acknowledging a cohort commit — the harness analogue of the sink
+	 * `libp2p-node-base` routes to the node's rebalance monitor. The harness builds no monitor; a spec
+	 * that wants one routes reports to its own. Omitted → nothing is reported.
+	 */
+	onCommittedHolders?: (node: MeshNode, committed: CommittedHolders) => void;
 }
 
 export interface MeshFailureConfig {
@@ -131,6 +146,18 @@ export interface MeshFailureConfig {
 	 * it declared, which is already its own side.
 	 */
 	partitionSides?: Set<string>[];
+	/**
+	 * Observes every REMOTE cluster delivery — promise round, commit round, commit broadcast and
+	 * scheduled commit retry alike — with the target's peer-id string and the record as sent, BEFORE
+	 * `failingPeers` is consulted for that delivery. It exists so a spec can change reachability at an
+	 * exact protocol step rather than between whole transactions: a member that promised and then
+	 * dropped before voting to commit is set unreachable here, on the delivery that would have carried
+	 * its commit vote, and that very delivery fails.
+	 *
+	 * A coordinator's OWN member is invoked in process, never through this path — as in production,
+	 * where `ClusterCoordinator.updateMember` calls it directly — so the hook never sees it.
+	 */
+	onClusterDelivery?: (targetPeerId: string, record: ClusterRecord) => void;
 }
 
 class MockPeerNetwork implements IPeerNetwork {
@@ -140,38 +167,46 @@ class MockPeerNetwork implements IPeerNetwork {
 }
 
 /**
- * Mock IKeyNetwork that returns peers based on XOR distance.
- * With responsibilityK >= nodeCount, all nodes are returned.
- * Otherwise, K-nearest by XOR distance are returned.
+ * Mock IKeyNetwork that places a block where production does, by construction: it ranks the mesh's nodes
+ * with FRET's own `assembleCohort` over a ring (`DigitreeStore`) holding every node at its real ring
+ * coordinate (`hashPeerId`), the walk `Libp2pKeyPeerNetwork` asks FRET for. Like production it hashes the
+ * routing key exactly once. Every mesh node serves, so the network-membership scoping production layers on
+ * top never removes anyone, and a cohort is simply the nearest `responsibilityK` in walk order.
+ * `test/mesh-harness-cohort-parity.spec.ts` pins the two answers equal.
+ *
+ * - `findCluster`: the nearest `responsibilityK` nodes, nearest first. Nothing is added: a node is in a
+ *   block's cohort only when it is among the nearest, as in production.
+ * - `findCoordinator`: the first non-excluded node of the whole ring's walk, which is the cohort's first
+ *   member unless excluded, else the next nearest node, cohort member or not. That mirrors production's
+ *   connected-peer fallback: a pick outside the cohort is a routing hop, not a placement.
  */
 class MockMeshKeyNetwork implements IKeyNetwork {
+	/** The ring, built on first lookup (node ids hash asynchronously, and `nodes` fills after construction)
+	 *  and rebuilt only if the node count changes; a restarted node keeps its identity and so its position. */
+	private ring: { size: number; ready: Promise<{ store: DigitreeStore; byId: Map<string, MeshNode> }> } | undefined;
+
 	constructor(
 		private readonly nodes: MeshNode[],
 		private readonly responsibilityK: number,
 		private readonly failures: MeshFailureConfig = {}
 	) {}
 
-	async findCoordinator(key: Uint8Array, options?: Partial<FindCoordinatorOptions>): Promise<PeerId> {
+	async findCoordinator(key: RoutingKey, options?: Partial<FindCoordinatorOptions>): Promise<PeerId> {
 		const excluded = new Set((options?.excludedPeers ?? []).map(p => p.toString()));
-		const sorted = this.sortedByDistance(key);
-		const pick = sorted.find(n => !excluded.has(n.peerId.toString()));
+		const pick = (await this.nearest(key, this.nodes.length)).find(n => !excluded.has(n.peerId.toString()));
 		if (!pick) {
 			throw new Error('No coordinator available for key (all candidates excluded)');
 		}
 		return pick.peerId;
 	}
 
-	async findCluster(key: Uint8Array): Promise<ClusterPeers> {
+	async findCluster(key: RoutingKey): Promise<ClusterPeers> {
 		if (this.failures.findClusterFails) {
 			return {} as ClusterPeers;
 		}
 
-		const sorted = this.sortedByDistance(key);
-		const k = Math.min(this.responsibilityK, sorted.length);
-		const selected = sorted.slice(0, k);
-
 		const peers: ClusterPeers = {};
-		for (const node of selected) {
+		for (const node of await this.nearest(key, this.responsibilityK)) {
 			peers[node.peerId.toString()] = {
 				multiaddrs: ['/ip4/127.0.0.1/tcp/8000'],
 				publicKey: u8ToString(node.peerId.publicKey!.raw, 'base64url')
@@ -180,13 +215,30 @@ class MockMeshKeyNetwork implements IKeyNetwork {
 		return peers;
 	}
 
-	private sortedByDistance(key: Uint8Array): MeshNode[] {
-		const knownPeers: KnownPeer[] = this.nodes.map(n => ({
-			id: n.peerId,
-			addrs: ['/ip4/127.0.0.1/tcp/8000']
-		}));
-		const sorted = sortPeersByDistance(knownPeers, key);
-		return sorted.map(kp => this.nodes.find(n => n.peerId.equals(kp.id))!);
+	/** The `wants` nodes nearest `key`'s ring coordinate, nearest first. */
+	private async nearest(key: RoutingKey, wants: number): Promise<MeshNode[]> {
+		const { store, byId } = await this.ringStore();
+		const ids = assembleCohort(store, await hashKey(key), Math.min(wants, byId.size));
+		return ids.map(id => byId.get(id)!);
+	}
+
+	private ringStore(): Promise<{ store: DigitreeStore; byId: Map<string, MeshNode> }> {
+		if (this.ring?.size !== this.nodes.length) {
+			const members = [...this.nodes];
+			this.ring = {
+				size: members.length,
+				ready: (async () => {
+					const store = new DigitreeStore();
+					const byId = new Map<string, MeshNode>();
+					for (const node of members) {
+						store.upsert(node.peerId.toString(), await hashPeerId(node.peerId));
+						byId.set(node.peerId.toString(), node);
+					}
+					return { store, byId };
+				})()
+			};
+		}
+		return this.ring.ready;
 	}
 }
 
@@ -194,6 +246,25 @@ export interface Mesh {
 	nodes: MeshNode[];
 	failures: MeshFailureConfig;
 	keyNetwork: IKeyNetwork;
+	/**
+	 * Restart one node the way a stopped and relaunched process comes back: same identity, same raw
+	 * storage, and nothing else. `node`'s storage repo, cluster member and coordinator are rebuilt IN
+	 * PLACE on the same `MeshNode` object (so a spec's reference to it stays valid), over the
+	 * `IRawStorage` instance it was built with — for the default `MemoryRawStorage` that is exactly
+	 * "over its own storage"; a `rawStorageFactory` store is reused as-is, never re-requested.
+	 *
+	 * Everything held only in memory goes with the old instance: the member's reservations and
+	 * executed-transaction memory, the coordinator's in-flight transactions and scheduled commit
+	 * retries, every read-repair stamp. The harness wires no `ITransactionStateStore`, so nothing is
+	 * recovered either — a production node built with `transactionStateStore` would resume its
+	 * persisted commit retries on restart (`ClusterCoordinator.recoverTransactions`). The old instance is made inert rather than merely forgotten —
+	 * its member is disposed, and every outbound call it would still make (a commit retry whose timer
+	 * fires later, a cohort consult, an archive fetch) fails the way a stopped process's dial does — so
+	 * a retry the old process owed cannot quietly deliver after the restart and pass for healing.
+	 *
+	 * Reachability (`failures`) is untouched: a node that was unreachable stays unreachable.
+	 */
+	restart(node: MeshNode): void;
 }
 
 /**
@@ -253,6 +324,12 @@ export function resolveMeshPolicy(options: MeshOptions): ResolvedClusterPolicy {
 	});
 }
 
+/** Whether one node INSTANCE is still running. Every outbound call the instance makes checks it, so an
+ *  instance {@link Mesh.restart} replaced goes quiet instead of acting alongside the new one. */
+interface NodeLifetime {
+	stopped: boolean;
+}
+
 /** The 32-byte Ed25519 seed for node `index` of a `keySeed` mesh. */
 function meshKeySeed(keySeed: number, index: number): Uint8Array {
 	const seed = new Uint8Array(32);
@@ -295,23 +372,27 @@ export async function createMesh(nodeCount: number, options: MeshOptions): Promi
 		})
 	);
 
-	// Build nodes array (partially — coordinatorRepo added after keyNetwork is ready)
 	const nodes: MeshNode[] = [];
 	const peerNetwork = new MockPeerNetwork();
 	// One real reconcile callback per node, shared between the member's commit-path `reconcileBlock`
 	// and the coordinator's read-path `acquireBlockFromCohort` — production shares one instance
-	// (`libp2p-node-base.ts`), and a spec must not be able to tell the two paths apart. Built in
-	// phase 1 (with the member), consumed again in phase 2 (by the coordinator), so it is stashed
-	// here keyed by peer id rather than widened onto the public `MeshNode` type.
+	// (`libp2p-node-base.ts`), and a spec must not be able to tell the two paths apart. Built with the
+	// member, consumed again by the coordinator, so it is stashed here keyed by peer id rather than
+	// widened onto the public `MeshNode` type.
 	const reconcileByPeer = new Map<string, ReconcileBlockCallback>();
+	/** What outlives a node instance, by peer id: the raw storage it was built over, and its index
+	 *  (`validatorFactory` is invoked with it again on restart). */
+	const durableByPeer = new Map<string, { rawStorage: IRawStorage; index: number }>();
+	/** The CURRENT instance's lifetime, by peer id — flipped to stopped by {@link Mesh.restart}. */
+	const lifetimeByPeer = new Map<string, NodeLifetime>();
 
 	// The mesh key network is built BEFORE the members: each member's `deriveExpectedCluster` (the
-	// admission gate's view) needs a per-node key network in phase 1, and constructing it here beats
-	// a late-bound slot a closure could fire on before it is filled. Safe because `nodes` is captured
-	// by reference and only consulted at call time, after the array is fully populated.
+	// admission gate's view) needs a per-node key network, and constructing it here beats a late-bound
+	// slot a closure could fire on before it is filled. Safe because `nodes` is captured by reference
+	// and only consulted at call time, after the array is fully populated.
 	//
 	// `wrapKeyNetwork` is applied HERE, before `makeNodeKeyNetwork` closes over `keyNetwork` below and
-	// before phase 1 builds any `deriveExpectedCluster` closure — every one of them reads the `keyNetwork`
+	// before any `deriveExpectedCluster` closure is built — every one of them reads the `keyNetwork`
 	// binding, so a wrapper assigned to it here is what every node's coordinator, cluster member and the
 	// returned `Mesh.keyNetwork` all observe. Reassigning `mesh.keyNetwork` after this function returns
 	// only reaches whoever reads that property later (the transactor); it is too late for the rest.
@@ -320,13 +401,18 @@ export async function createMesh(nodeCount: number, options: MeshOptions): Promi
 
 	/**
 	 * One node's own view of the key network — what `Libp2pKeyPeerNetwork` gives a real node:
-	 *  - `findCluster` always includes self, so a responsible member's derived view is never empty
-	 *    (see the empty-view guard in `cluster-repo.admitMembership`);
+	 *  - `findCluster` is the shared cohort, which holds this node only when it is among the block's
+	 *    nearest `responsibilityK`, so the node's coordinator judges its own responsibility as a real
+	 *    node's does;
 	 *  - under a simulated partition (`failures.partitionSides`), a caller inside a side sees only
 	 *    its side's members of the cohort — the caller-aware filtering lives here, in the per-node
-	 *    wrapper, precisely so `IKeyNetwork` itself needs no "who is asking" parameter.
-	 * The SAME instance serves both the member's admission derivation (phase 1) and the node's
-	 * coordinator (phase 2), so the two sides of a node can never see different topologies.
+	 *    wrapper, precisely so `IKeyNetwork` itself needs no "who is asking" parameter. A side holding
+	 *    none of the cohort sees an empty view, so a node there is not responsible for the block (and
+	 *    the empty-view guard in `cluster-repo.admitMembership` treats such a view as unconfident rather
+	 *    than as a reference set).
+	 * The SAME instance serves both the member's admission derivation and the node's coordinator, so
+	 * the two sides of a node can never see different topologies. It holds no state, so a restarted
+	 * node keeps it.
 	 */
 	const makeNodeKeyNetwork = (selfPeerId: PeerId): IKeyNetwork => {
 		const selfStr = selfPeerId.toString();
@@ -340,25 +426,18 @@ export async function createMesh(nodeCount: number, options: MeshOptions): Promi
 						if (!side.has(id)) delete peers[id];
 					}
 				}
-				if (!(selfStr in peers)) {
-					peers[selfStr] = {
-						multiaddrs: ['/ip4/127.0.0.1/tcp/8000'],
-						publicKey: u8ToString(selfPeerId.publicKey!.raw, 'base64url')
-					};
-				}
 				return peers;
 			}
 		};
 	};
-	const nodeKeyNetworkByPeer = new Map<string, IKeyNetwork>();
-
-	// Phase 1: create storage + cluster members
-	let nodeIndex = 0;
-	for (const { peerId, privateKey } of keyPairs) {
-		const index = nodeIndex++;
-		const rawStorage = options.rawStorageFactory
-			? options.rawStorageFactory(index)
-			: new MemoryRawStorage();
+
+	/**
+	 * Build `meshNode`'s storage repo and cluster member over its durable raw storage, assigning both
+	 * onto the node. Run once per node at assembly, and again by {@link Mesh.restart}.
+	 */
+	const buildMember = (meshNode: MeshNode, lifetime: NodeLifetime): void => {
+		const { peerId, privateKey } = meshNode;
+		const { rawStorage, index } = durableByPeer.get(peerId.toString())!;
 		const storageRepo = new StorageRepo(
 			(blockId: BlockId) => new BlockStorage(blockId, rawStorage)
 		);
@@ -369,9 +448,12 @@ export async function createMesh(nodeCount: number, options: MeshOptions): Promi
 		// distinct peers must agree on the target `(rev, actionId)` AND on the block content, or the
 		// pass declines, persisting nothing. `reputation` is omitted — no reputation subsystem in the
 		// harness.
+		const fetchArchive = makeFetchArchive(nodes, peerId.toString(), failures);
 		const reconcileBlock = createReconcileBlock({
 			selfPeerId: peerId.toString(),
-			fetchArchive: makeFetchArchive(nodes, peerId.toString(), failures),
+			// A stopped instance fetches nothing — the production fetch swallows a failed dial into
+			// the same `undefined`.
+			fetchArchive: async (peerIdStr, blockId) => lifetime.stopped ? undefined : await fetchArchive(peerIdStr, blockId),
 			// Production shape: a proof reconcile verified against the agreed bytes is persisted so
 			// the repaired replica serves it onward.
 			saveReplicatedBlock: (blockId, block, source, verifiedProof) =>
@@ -382,33 +464,21 @@ export async function createMesh(nodeCount: number, options: MeshOptions): Promi
 		});
 		reconcileByPeer.set(peerId.toString(), reconcileBlock);
 
-		const nodeKeyNetwork = makeNodeKeyNetwork(peerId);
-		nodeKeyNetworkByPeer.set(peerId.toString(), nodeKeyNetwork);
-
-		// The node object exists before its member so the admission derivation below can hand the
-		// finished MeshNode to spec-supplied callbacks; `clusterMember`/`coordinatorRepo` are
-		// assigned as they are built (member just below, coordinator in phase 2) and the closures
-		// only run at vote time, long after both are in place.
-		const meshNode: MeshNode = {
-			peerId,
-			privateKey,
-			storageRepo,
-			clusterMember: undefined as any,
-			coordinatorRepo: undefined as any
-		};
+		const nodeKeyNetwork = meshNode.keyNetwork;
 
 		// Member-side cluster derivation for the membership admission gate — the production shape
 		// (`libp2p-node-base.deriveExpectedCluster`): the SAME per-node key network the coordinator
-		// selects its cohort from, plus a network-size confidence. The self-including wrapper keeps a
-		// responsible member's view non-empty; `meshConfidence` is the FRET stand-in (default 1, i.e.
+		// selects its cohort from, plus a network-size confidence. A responsible member is in its own
+		// cohort, so its view is never empty; `meshConfidence` is the FRET stand-in (default 1, i.e.
 		// confident — a partition spec collapses it per side).
 		const deriveExpectedCluster: DeriveExpectedClusterCallback = options.deriveExpectedCluster
 			? (blockId) => options.deriveExpectedCluster!(meshNode, blockId)
 			: async (blockId) => ({
-				peers: await nodeKeyNetwork.findCluster(new TextEncoder().encode(blockId)) ?? {},
+				peers: await nodeKeyNetwork.findCluster(routingKeyForBlock(blockId)) ?? {},
 				confidence: options.meshConfidence?.(meshNode) ?? 1
 			});
 
+		meshNode.storageRepo = storageRepo;
 		meshNode.clusterMember = clusterMember({
 			storageRepo,
 			peerNetwork,
@@ -419,13 +489,13 @@ export async function createMesh(nodeCount: number, options: MeshOptions): Promi
 			deriveExpectedCluster,
 			// Absent by default: `undefined` here is identical to omitting the field, and
 			// `validatePendOperations` then skips the validation step entirely.
-			validator: options.validatorFactory?.(index, peerId)
+			validator: options.validatorFactory?.(index, peerId),
+			onCommittedHolders: options.onCommittedHolders && ((committed) => options.onCommittedHolders!(meshNode, committed))
 		});
+	};
 
-		nodes.push(meshNode);
-	}
-
-	// Phase 2: coordinator repos (needs all nodes for routing; key network built in phase 1)
+	// Cluster traffic to a node, resolved per call: a delivery always reaches the target's CURRENT
+	// member, so a restarted node answers with its new instance.
 	const createClusterClient = (targetPeerId: PeerId): ICluster => {
 		const target = nodes.find(n => n.peerId.equals(targetPeerId));
 		if (!target) {
@@ -433,6 +503,7 @@ export async function createMesh(nodeCount: number, options: MeshOptions): Promi
 		}
 		return {
 			async update(record: ClusterRecord): Promise<ClusterRecord> {
+				failures.onClusterDelivery?.(targetPeerId.toString(), record);
 				if (failures.failingPeers?.has(targetPeerId.toString())) {
 					throw new Error(`Peer ${targetPeerId.toString()} is unreachable`);
 				}
@@ -441,7 +512,13 @@ export async function createMesh(nodeCount: number, options: MeshOptions): Promi
 		};
 	};
 
-	for (const node of nodes) {
+	/**
+	 * Build `node`'s coordinator over its (already built) storage repo and member. It reaches the other
+	 * nodes only at call time, never at construction. Run once per node at assembly, and again by
+	 * {@link Mesh.restart}.
+	 */
+	const buildCoordinator = (node: MeshNode, lifetime: NodeLifetime): void => {
+		const stoppedError = (): Error => new Error(`${node.peerId.toString()} was restarted; this instance is stopped`);
 		// Per-node callback: reports the queried peer's latest revision, and NOTHING else. It used to
 		// also write the peer's block into local storage ("simulate data sync"), which made every
 		// read-repair assertion on this harness observe a convergence the production callback does not
@@ -449,6 +526,7 @@ export async function createMesh(nodeCount: number, options: MeshOptions): Promi
 		// existed to expose. Transfer now happens where it does in production: through
 		// `acquireBlockFromCohort` below, gated on a corroborated revision.
 		const clusterLatestCallback: ClusterLatestCallback = async (peerId: PeerId, blockId: BlockId, context?): Promise<CertifiedActionRev | undefined> => {
+			if (lifetime.stopped) throw stoppedError();
 			// Silence: the peer never answers. REJECTS, mirroring what a dial failure does to the
 			// production callback — the coordinator must count this as "did not answer", never as
 			// an absent claim (a resolved `undefined` remains the peer answering "I hold nothing").
@@ -472,19 +550,29 @@ export async function createMesh(nodeCount: number, options: MeshOptions): Promi
 			const proof = await servableProof(target.storageRepo, blockId, latest);
 			return proof ? { ...latest, proof } : latest;
 		};
-		// The node's own self-including (and partition-aware) key-network view, built in phase 1 —
-		// the SAME instance the member's admission derivation reads, matching real
-		// Libp2pKeyPeerNetwork behavior.
-		const nodeKeyNetwork = nodeKeyNetworkByPeer.get(node.peerId.toString())!;
+		// The node's own (partition-aware) key-network view — the SAME instance the member's admission
+		// derivation reads, matching real Libp2pKeyPeerNetwork behavior.
+		const nodeKeyNetwork = node.keyNetwork;
 		const factory = coordinatorRepo(
 			nodeKeyNetwork,
-			createClusterClient,
-			// The SAME resolved policy the member above was built from, spread the way
-			// `libp2p-node-base` spreads it into its coordinator factory — carrying
-			// `repairCorroborationClusterSize` (the repair floor's yardstick, DEFAULT_CLUSTER_SIZE
-			// when the mesh declared nothing), the production `minAbsoluteClusterSize` (2, not the
-			// coordinator's own fallback of 3), and the `allowUnvalidatedSmallCluster` gate —
-			// ARMED (false) unless the mesh opted out at its call site.
+			// A stopped instance's cluster traffic fails at send — checked per delivery rather than when
+			// the client is made, because `ClusterCoordinator` also sends from a commit-retry timer that
+			// can fire long after a restart.
+			(targetPeerId: PeerId): ICluster => {
+				const client = createClusterClient(targetPeerId);
+				return {
+					async update(record: ClusterRecord): Promise<ClusterRecord> {
+						if (lifetime.stopped) throw stoppedError();
+						return await client.update(record);
+					}
+				};
+			},
+			// The SAME resolved policy the member was built from, spread the way `libp2p-node-base`
+			// spreads it into its coordinator factory — carrying `repairCorroborationClusterSize` (the
+			// repair floor's yardstick, DEFAULT_CLUSTER_SIZE when the mesh declared nothing), the
+			// production `minAbsoluteClusterSize` (2, not the coordinator's own fallback of 3), and the
+			// `allowUnvalidatedSmallCluster` gate — ARMED (false) unless the mesh opted out at its call
+			// site.
 			{ ...policy }
 		);
 		node.coordinatorRepo = factory({
@@ -494,27 +582,110 @@ export async function createMesh(nodeCount: number, options: MeshOptions): Promi
 			clusterLatestCallback,
 			// The read path's transfer mechanism — the SAME instance the member uses on the commit
 			// path, mirroring how `libp2p-node-base` shares one `reconcileBlock` between both.
-			acquireBlockFromCohort: reconcileByPeer.get(node.peerId.toString())!
+			acquireBlockFromCohort: reconcileByPeer.get(node.peerId.toString())!,
+			onCommittedHolders: options.onCommittedHolders && ((committed) => options.onCommittedHolders!(node, committed))
 		});
+	};
+
+	// Phase 1: storage + cluster members, in key-pair order (the order `validatorFactory` indexes by).
+	for (const [index, { peerId, privateKey }] of keyPairs.entries()) {
+		const peerIdStr = peerId.toString();
+		durableByPeer.set(peerIdStr, {
+			rawStorage: options.rawStorageFactory ? options.rawStorageFactory(index) : new MemoryRawStorage(),
+			index
+		});
+		const lifetime: NodeLifetime = { stopped: false };
+		lifetimeByPeer.set(peerIdStr, lifetime);
+
+		// The node object exists before its member so the admission derivation can hand the finished
+		// MeshNode to spec-supplied callbacks; `storageRepo`/`clusterMember` are assigned by
+		// `buildMember` just below and `coordinatorRepo` in phase 2, and those closures only run at
+		// vote time, long after all three are in place.
+		const meshNode: MeshNode = {
+			peerId,
+			privateKey,
+			keyNetwork: makeNodeKeyNetwork(peerId),
+			storageRepo: undefined as any,
+			clusterMember: undefined as any,
+			coordinatorRepo: undefined as any
+		};
+		buildMember(meshNode, lifetime);
+		nodes.push(meshNode);
 	}
 
-	return { nodes, failures, keyNetwork };
+	// Phase 2: coordinator repos, once every member exists.
+	for (const node of nodes) {
+		buildCoordinator(node, lifetimeByPeer.get(node.peerId.toString())!);
+	}
+
+	const restart = (node: MeshNode): void => {
+		const peerIdStr = node.peerId.toString();
+		if (!nodes.includes(node)) {
+			throw new Error(`restart: ${peerIdStr} is not a node of this mesh`);
+		}
+		// NOTE: this severs the old instance's traffic to OTHER nodes, not its in-process path to its own
+		// member and storage repo, which still write the shared raw storage if called. After a restart
+		// between transactions only one thing could reach it: a scheduled commit retry aimed at the
+		// coordinator's OWN member, which exists only when that member threw while applying the broadcast —
+		// a local fault no spec injects today. If a spec ever does, or restarts a node MID-transaction,
+		// sever the old coordinator's local member too.
+		lifetimeByPeer.get(peerIdStr)!.stopped = true;
+		node.clusterMember.dispose();
+		const lifetime: NodeLifetime = { stopped: false };
+		lifetimeByPeer.set(peerIdStr, lifetime);
+		buildMember(node, lifetime);
+		buildCoordinator(node, lifetime);
+	};
+
+	return { nodes, failures, keyNetwork, restart };
 }
 
 /**
  * The nodes the key network keeps OUT of `blockId`'s cohort — peers that receive none of the
  * block's cluster traffic, and so hold none of its content until something repairs them.
  *
- * Peer ids are generated fresh per mesh, so which node is responsible for a given block is random
- * from run to run: in a 3-node `responsibilityK: 1` mesh, `nodes[1]` is the block's sole responsible
- * peer about a third of the time, and then it receives the writer's commit directly. A test that
- * needs a genuinely non-responsible node has to ask the routing layer rather than assume an index.
+ * Unless the mesh is seeded (`keySeed`), peer ids are generated fresh per mesh, so which node is
+ * responsible for a given block is random from run to run: in a 3-node `responsibilityK: 1` mesh,
+ * `nodes[1]` is the block's sole responsible peer about a third of the time, and then it receives the
+ * writer's commit directly. A test that needs a genuinely non-responsible node has to ask the routing
+ * layer rather than assume an index.
  */
 export async function nonResponsibleNodes(mesh: Mesh, blockId: string): Promise<MeshNode[]> {
-	const cohort = await mesh.keyNetwork.findCluster(new TextEncoder().encode(blockId));
+	const cohort = await mesh.keyNetwork.findCluster(routingKeyForBlock(blockId));
 	return mesh.nodes.filter(node => !(node.peerId.toString() in cohort));
 }
 
+/**
+ * The nodes the key network places in `blockId`'s cohort, nearest first — the complement of
+ * {@link nonResponsibleNodes}. A node outside this list refuses a write for the block through its own
+ * coordinator (`CoordinatorRepo` checks responsibility), so a spec that writes through one node's
+ * coordinator directly picks that node from here rather than by index.
+ */
+export async function responsibleNodes(mesh: Mesh, blockId: string): Promise<MeshNode[]> {
+	const cohort = Object.keys(await mesh.keyNetwork.findCluster(routingKeyForBlock(blockId)));
+	return cohort.map(id => mesh.nodes.find(node => node.peerId.toString() === id)!);
+}
+
+/**
+ * The first `count` ids of the form `${prefix}-${i}` whose cohort includes `node` — for a spec that needs
+ * several blocks one node is responsible for (a multi-block pend through that node's coordinator, or
+ * sequential writes it coordinates alone). In a `responsibilityK: 1` mesh these are blocks the node is
+ * the SOLE responsible peer for. Throws after `maxCandidates` ids rather than looping on a node whose ring
+ * arc no id lands in.
+ */
+export async function blockIdsInCohortOf(mesh: Mesh, node: MeshNode, count: number, prefix: string, maxCandidates = 10_000): Promise<BlockId[]> {
+	const nodeId = node.peerId.toString();
+	const ids: BlockId[] = [];
+	for (let i = 0; ids.length < count && i < maxCandidates; i++) {
+		const id = `${prefix}-${i}` as BlockId;
+		if (nodeId in await mesh.keyNetwork.findCluster(routingKeyForBlock(id))) ids.push(id);
+	}
+	if (ids.length < count) {
+		throw new Error(`blockIdsInCohortOf: only ${ids.length} of ${count} ids with prefix ${prefix} place ${nodeId} in their cohort`);
+	}
+	return ids;
+}
+
 export interface BuildTransactorOptions {
 	timeoutMs?: number;
 	abortOrCancelTimeoutMs?: number;
@@ -529,10 +700,30 @@ export interface BuildTransactorOptions {
 /**
  * Builds a NetworkTransactor over a mesh. All nodes share the same mock
  * infrastructure so a single transactor routes to every peer via `getRepo`.
+ * It runs on no node (no `localPeerId`), so a coverage tie between cohort
+ * members goes to the nearest — a client-only writer's shape.
  * Suitable for solo-mesh tests; for multi-node tests prefer
  * `buildNetworkTransactors` to label "which node is driving".
  */
-export const buildNetworkTransactor = (mesh: Mesh, options: BuildTransactorOptions = {}): ITransactor => {
+export const buildNetworkTransactor = (mesh: Mesh, options: BuildTransactorOptions = {}): ITransactor =>
+	meshTransactor(mesh, options, undefined);
+
+/**
+ * Builds one NetworkTransactor per mesh node, keyed by peer-id string. Each
+ * transactor shares the mesh's key network and peer→repo map, and runs on its
+ * node the way a production node's transactor does (`localPeerId`): when that
+ * node is in a block's cohort and ties another member on coverage, it
+ * coordinates the write itself. Reads and retries still route by proximity.
+ */
+export const buildNetworkTransactors = (mesh: Mesh, options: BuildTransactorOptions = {}): Map<string, ITransactor> => {
+	const transactors = new Map<string, ITransactor>();
+	for (const node of mesh.nodes) {
+		transactors.set(node.peerId.toString(), meshTransactor(mesh, options, node));
+	}
+	return transactors;
+};
+
+function meshTransactor(mesh: Mesh, options: BuildTransactorOptions, localNode: MeshNode | undefined): ITransactor {
 	const repoByPeer = new Map<string, IRepo>();
 	for (const node of mesh.nodes) {
 		const repo = node.coordinatorRepo as unknown as IRepo;
@@ -541,24 +732,14 @@ export const buildNetworkTransactor = (mesh: Mesh, options: BuildTransactorOptio
 	return new NetworkTransactor({
 		timeoutMs: options.timeoutMs ?? 5_000,
 		abortOrCancelTimeoutMs: options.abortOrCancelTimeoutMs ?? 5_000,
+		// NOTE: every transactor, `localNode`'s included, reads the unpartitioned shared view, not `localNode.keyNetwork`;
+		// no spec drives `partitionSides` through a node's transactor today. If one does, pass the node's own view.
 		keyNetwork: mesh.keyNetwork,
 		getRepo: (peerId: DbPeerId) => {
 			const repo = repoByPeer.get(peerId.toString());
 			if (!repo) throw new Error(`Unknown peer ${peerId.toString()}`);
 			return repo;
-		}
+		},
+		localPeerId: localNode?.peerId
 	});
-};
-
-/**
- * Builds one NetworkTransactor per mesh node, keyed by peer-id string. Each
- * transactor shares the mesh's key network and peer→repo map — the separate
- * instances exist so tests can semantically say "driven by node A".
- */
-export const buildNetworkTransactors = (mesh: Mesh, options: BuildTransactorOptions = {}): Map<string, ITransactor> => {
-	const transactors = new Map<string, ITransactor>();
-	for (const node of mesh.nodes) {
-		transactors.set(node.peerId.toString(), buildNetworkTransactor(mesh, options));
-	}
-	return transactors;
-};
+}