@edryslabs/genericprovider 1.0.2 → 1.0.4

package/dist/index.js

@@ -11,6 +11,9 @@ const MESSAGE_AWARENESS = 1;
 const MESSAGE_PUBSUB = 2;
 const MESSAGE_SYNC_VERIFIED = 3; // Sync message with hash verification
 const MESSAGE_PUBSUB_TARGETED = 4; // Pub/sub message aimed at a single target
+// Sub-channel inside a MESSAGE_AWARENESS frame: [opcode 1][channel][update].
+const AWARENESS_CHANNEL_MAIN = 0;
+const AWARENESS_CHANNEL_APP = 1;
 /**
  * CRC32 lookup table for fast computation.
  * Generated once and reused for all CRC calculations.
@@ -78,16 +81,24 @@ function unwrapAndVerifyMessage(wrapped) {
     return message;
 }
 /**
- * Compute a simple hash of document state for desync detection.
- * Uses a fast non-cryptographic hash for performance.
+ * Compute a cheap hash of document state for desync detection.
  *
- * Hashes the state VECTOR, not encodeStateAsUpdate: the full update byte stream
- * is NOT canonical across CRDT-convergent replicas (client-block and tombstone
- * ordering differ per peer), so hashing it flags false divergence and triggers
- * an endless re-sync loop. The state vector (clientID -> clock) is serialized in
- * sorted clientID order by Yjs, so two convergent docs hash identically, while a
- * missed update still shows up as a differing clock — which is exactly the
- * "did we fall behind?" signal this check exists to provide.
+ * Hashes the state VECTOR, not encodeStateAsUpdate, for two independent reasons:
+ *
+ * 1. CORRECTNESS: the full update byte stream is NOT canonical across
+ *    CRDT-convergent replicas (client-block and tombstone ordering differ per
+ *    peer), so hashing it flags false divergence and triggers an endless
+ *    re-sync loop. The state vector (clientID -> clock) is serialized in sorted
+ *    clientID order by Yjs, so two convergent docs hash identically, while a
+ *    missed update still shows up as a differing clock — exactly the "did we
+ *    fall behind?" signal this check exists to provide.
+ * 2. COST: O(number of distinct clients) instead of O(document content size),
+ *    so this no longer re-serializes the entire document on every update.
+ *
+ * Two peers can only reach the same state vector by having applied the same set
+ * of operations, so real content divergence is still caught. (CRC32 already
+ * guards wire corruption, and sequence tracking guards reordering/loss — this
+ * hash is the last line of defense against logical divergence between peers.)
  */
 function computeDocHash(doc) {
     const state = Y.encodeStateVector(doc);
@@ -211,19 +222,41 @@ export class GenericProvider extends Observable {
         // BroadcastChannel state for cross-tab sync
         this._bcChannel = '';
         this._bcConnected = false;
-        // Hash verification tracking for exponential backoff
-        this._hashMismatchCount = 0;
-        this._lastHashMismatchTime = 0;
+        // Unified resync-request coordinator. Previously hash-mismatch,
+        // corrupted-message, and gap-confirmed triggers each coalesced only
+        // against themselves (three separate pending-timer fields, three
+        // separate escalation counters), so under sustained wire corruption they
+        // could each independently burn through the shared _tryReserveSyncSlot()
+        // budget in the same window - a resync storm that grew combinatorially
+        // with peer count (see test/dummy/bench-corruption-storm.ts: at 10
+        // simulated peers, 5% per-link corruption drove message volume to ~11x
+        // the corruption-free baseline). Now there is exactly ONE pending timer
+        // and ONE shared escalation counter for all three triggers - only one
+        // resync is ever in flight at a time, and any trigger that fires while
+        // one is already pending is absorbed into it instead of scheduling its
+        // own. See _requestResync().
+        this._resyncAttemptCount = 0;
+        this._lastResyncAttemptTime = 0;
         // Rate limiting for sync requests
         this._syncRequestTimes = [];
-        this._maxSyncRequestsPerWindow = 20; // max requests per 10 seconds
-        this._syncRequestWindowMs = 10000; // 10 second window
+        // SyncStep2 reply suppression (NACK-suppression style): delay a reply to
+        // a SyncStep1 request briefly, and drop it if another peer's reply is
+        // overheard first - since every reply is broadcast to the whole room
+        // anyway, this avoids every peer answering the same request redundantly.
+        // Only engages when there's genuine redundancy (see _handleIncomingMessage's
+        // MESSAGE_SYNC and MESSAGE_SYNC_VERIFIED cases) - with 0-1 other known
+        // peers there's no "someone else" to rely on, so replies go out
+        // immediately as before.
+        this._pendingSyncReply = null;
         // Sequence numbers for causal ordering
         this._localSeqNum = 0; // Our sequence number counter
-        this._remoteSeqNums = new Map(); // clientID -> last seen seqNum
-        // Message integrity tracking
-        this._corruptedMessageCount = 0; // Track rejected corrupted messages
-        this._lastCorruptedMessageTime = 0;
+        // Per-sender sequence tracking for reordering-tolerant gap detection.
+        // Applying a Yjs update is always safe even for duplicates or out-of-order
+        // arrivals (Yjs updates are idempotent/commutative) — this state exists
+        // only to detect genuine gaps (likely packet loss) without false
+        // positives from mere network reordering. See _trackRemoteSeq().
+        this._remoteSeqInfo = new Map();
+        this._gapCheckTimers = new Map();
         // Update batching/debouncing
         this._batchUpdates = 0; // milliseconds delay (0 = disabled)
         this._pendingUpdate = null;
@@ -231,6 +264,9 @@ export class GenericProvider extends Observable {
         this._awarenessInterval = 100; // ms between awareness broadcasts
         this._pendingAwarenessClients = new Set();
         this._lastAwarenessTime = 0;
+        // Independent throttle state for the app awareness channel.
+        this._pendingAppAwarenessClients = new Set();
+        this._lastAppAwarenessTime = 0;
         // Origins whose updates are never sent to the transport (local-only txns).
         this._excludeOrigins = new Set();
         // Connect-time sync strategy: 'push-pull' (default) sends full local state
@@ -240,14 +276,22 @@ export class GenericProvider extends Observable {
         this.transport = transport;
         this.pubsub = new PubSubChannel(this);
         this.awareness = options.awareness || new awarenessProtocol.Awareness(doc);
+        this.appAwareness =
+            options.appAwareness || new awarenessProtocol.Awareness(doc);
         this._syncInterval = options.syncInterval ?? 5000;
         this._verifyUpdates = options.verifyUpdates ?? true;
-        this._batchUpdates = options.batchUpdates ?? 0;
+        this._batchUpdates =
+            options.batchUpdates ?? transport.preferredBatchMs ?? 0;
         this._disableBc = options.disableBc ?? false;
         this._awarenessInterval = options.awarenessInterval ?? 100;
         this._excludeOrigins = new Set(options.excludeOrigins ?? []);
         this._localId = options.localId;
         this._syncMode = options.syncMode ?? 'push-pull';
+        this._maxSyncRequestsPerWindow = options.maxSyncRequestsPerWindow ?? 20;
+        this._syncRequestWindowMs = options.syncRequestWindowMs ?? 10000;
+        this._syncReplySuppressionMs = options.syncReplySuppressionMs ?? 30;
+        this._gapGraceMs = options.gapGraceMs ?? 300;
+        this._seqWindowSize = options.seqWindowSize ?? 64;
         this._setupDocumentSync();
         this._setupAwarenessSync();
     }
@@ -306,18 +350,15 @@ export class GenericProvider extends Observable {
             }
             // Broadcast local awareness state
             this._broadcastAwareness([this.doc.clientID]);
+            this._broadcastAwareness([this.doc.clientID], AWARENESS_CHANNEL_APP);
             // Start periodic sync to handle packet loss
             // Just request sync without sending full state (avoid redundant broadcasts)
+            // _sendSyncStep1() already checks the shared rate limiter internally
+            // and silently drops the request if it's exceeded.
             if (this._syncInterval > 0) {
                 this._syncIntervalId = setInterval(() => {
                     if (this.transport.isConnected && !this._destroying) {
-                        // Check rate limit before syncing
-                        const now = Date.now();
-                        this._syncRequestTimes = this._syncRequestTimes.filter((t) => now - t < this._syncRequestWindowMs);
-                        if (this._syncRequestTimes.length < this._maxSyncRequestsPerWindow) {
-                            this._sendSyncStep1();
-                        }
-                        // If rate limited, skip this periodic sync - will try again next interval
+                        this._sendSyncStep1();
                     }
                 }, this._syncInterval);
             }
@@ -340,9 +381,26 @@ export class GenericProvider extends Observable {
             clearInterval(this._syncIntervalId);
             this._syncIntervalId = undefined;
         }
-        // Reset corruption tracking
-        this._corruptedMessageCount = 0;
-        this._lastCorruptedMessageTime = 0;
+        // Reset resync escalation tracking
+        this._resyncAttemptCount = 0;
+        this._lastResyncAttemptTime = 0;
+        // Cancel any pending unified resync - it would otherwise still fire
+        // syncNow() after disconnect/reconnect against a transport that may be
+        // in a completely different state by then.
+        if (this._pendingResyncTimeoutId !== undefined) {
+            clearTimeout(this._pendingResyncTimeoutId);
+            this._pendingResyncTimeoutId = undefined;
+        }
+        // Reset the sync rate-limit budget. Without this, a reconnect inherits
+        // whatever budget was left over from before the disconnect - and since
+        // syncNow()'s full-state push now shares this same limiter (see
+        // _tryReserveSyncSlot()), a rate-limited reconnect could silently skip
+        // the very push that delivers edits made while offline.
+        this._syncRequestTimes = [];
+        // Drop any pending suppressed sync reply - safe to simply discard (not
+        // flush/send like batched updates/awareness below), since a suppressed
+        // reply is by design redundant with whatever the room already has.
+        this._cancelPendingSyncReply();
         // Flush any pending batched updates before disconnecting
         if (this._batchTimeoutId !== undefined) {
             clearTimeout(this._batchTimeoutId);
@@ -365,6 +423,16 @@ export class GenericProvider extends Observable {
             }
         }
         this._pendingAwarenessClients.clear();
+        // Flush pending app awareness updates before disconnecting
+        if (this._appAwarenessTimeoutId !== undefined) {
+            clearTimeout(this._appAwarenessTimeoutId);
+            this._appAwarenessTimeoutId = undefined;
+            if (this._pendingAppAwarenessClients.size > 0 &&
+                this.transport.isConnected) {
+                this._sendAwarenessNow(Array.from(this._pendingAppAwarenessClients), AWARENESS_CHANNEL_APP);
+            }
+        }
+        this._pendingAppAwarenessClients.clear();
         // Disconnect BroadcastChannel
         this._disconnectBroadcastChannel();
         if (this._unsubscribeTransport) {
@@ -373,6 +441,7 @@ export class GenericProvider extends Observable {
         }
         // Mark local client as offline in awareness
         awarenessProtocol.removeAwarenessStates(this.awareness, [this.doc.clientID], 'disconnect');
+        awarenessProtocol.removeAwarenessStates(this.appAwareness, [this.doc.clientID], 'disconnect');
         this.transport.disconnect();
         this._synced = false;
         this._setStatus({ state: 'disconnected' });
@@ -388,6 +457,14 @@ export class GenericProvider extends Observable {
             clearInterval(this._syncIntervalId);
             this._syncIntervalId = undefined;
         }
+        // Stop any pending gap-check timers
+        for (const timer of this._gapCheckTimers.values()) {
+            clearTimeout(timer);
+        }
+        this._gapCheckTimers.clear();
+        // Drop any pending suppressed sync reply (disconnect() will also do
+        // this, but be explicit)
+        this._cancelPendingSyncReply();
         // Flush any pending batched updates before destroying
         if (this._batchTimeoutId !== undefined) {
             clearTimeout(this._batchTimeoutId);
@@ -404,17 +481,22 @@ export class GenericProvider extends Observable {
             this.doc.off('update', this._updateHandler);
             this._updateHandler = undefined;
         }
-        // Remove awareness update listener
+        // Remove awareness update listeners
         if (this._awarenessUpdateHandler) {
             this.awareness.off('update', this._awarenessUpdateHandler);
             this._awarenessUpdateHandler = undefined;
         }
+        if (this._appAwarenessUpdateHandler) {
+            this.appAwareness.off('update', this._appAwarenessUpdateHandler);
+            this._appAwarenessUpdateHandler = undefined;
+        }
         // Remove beforeunload handler
         if (this._beforeUnloadHandler && typeof window !== 'undefined') {
             window.removeEventListener('beforeunload', this._beforeUnloadHandler);
             this._beforeUnloadHandler = undefined;
         }
         this.awareness.destroy();
+        this.appAwareness.destroy();
         super.destroy();
     }
     /**
@@ -450,16 +532,32 @@ export class GenericProvider extends Observable {
             console.warn('Cannot sync: transport not connected');
             return;
         }
-        // Send our current document state to all peers
-        // This ensures any changes made while offline are transmitted
-        const update = Y.encodeStateAsUpdate(this.doc);
-        if (update.length > 0) {
-            this._sendUpdate(update);
+        // Push (full document state) and pull (SyncStep1 request) share a
+        // single rate-limit reservation. syncNow() is called from several
+        // triggers that can all fire in a short window when many peers are
+        // converging at once (hash-mismatch resyncs, gap-check confirmations,
+        // per-peer connect events on mesh transports) - without this gate the
+        // push above had NO limit at all, so each trigger broadcast the full
+        // document state to the whole room, and those broadcasts caused more
+        // reordering/mismatches elsewhere, causing more triggers. Measured in
+        // test/dummy/bench-user-scaling.ts: at 100 simulated users this drove
+        // message counts to 20-200x the theoretical linear cost. See
+        // docs/superpowers/specs/2026-07-26-dummy-benchmark-scaling-design.md.
+        if (this._tryReserveSyncSlot()) {
+            // Send our current document state to all peers
+            // This ensures any changes made while offline are transmitted
+            const update = Y.encodeStateAsUpdate(this.doc);
+            if (update.length > 0) {
+                this._sendUpdate(update);
+            }
+            // Send sync request to get updates from others
+            this._writeSyncStep1();
         }
-        // Send sync request to get updates from others
-        this._sendSyncStep1();
-        // Broadcast current awareness state
+        // Broadcast current awareness state - independently throttled and much
+        // cheaper than a full document push, so it isn't gated by the sync
+        // rate limiter above even when the sync half is skipped.
         this._broadcastAwareness([this.doc.clientID]);
+        this._broadcastAwareness([this.doc.clientID], AWARENESS_CHANNEL_APP);
     }
     /**
      * Setup automatic document synchronization.
@@ -532,10 +630,16 @@ export class GenericProvider extends Observable {
             this._broadcastAwareness(changedClients);
         };
         this.awareness.on('update', this._awarenessUpdateHandler);
+        this._appAwarenessUpdateHandler = ({ added, updated, removed, }, _origin) => {
+            const changedClients = added.concat(updated).concat(removed);
+            this._broadcastAwareness(changedClients, AWARENESS_CHANNEL_APP);
+        };
+        this.appAwareness.on('update', this._appAwarenessUpdateHandler);
         // Cleanup: mark as offline and disconnect BC when page unloads
         if (typeof window !== 'undefined') {
             this._beforeUnloadHandler = () => {
                 awarenessProtocol.removeAwarenessStates(this.awareness, [this.doc.clientID], 'window unload');
+                awarenessProtocol.removeAwarenessStates(this.appAwareness, [this.doc.clientID], 'window unload');
                 // Disconnect BroadcastChannel to notify other tabs
                 this._disconnectBroadcastChannel();
             };
@@ -552,23 +656,12 @@ export class GenericProvider extends Observable {
         const message = unwrapAndVerifyMessage(data);
         if (message === null) {
             // Message is corrupted - reject it immediately
-            this._corruptedMessageCount++;
-            const now = Date.now();
-            // Reset counter if it's been stable for 10 seconds
-            if (now - this._lastCorruptedMessageTime > 10000) {
-                this._corruptedMessageCount = 1;
-            }
-            this._lastCorruptedMessageTime = now;
-            console.warn(`[GenericProvider] 💥 Corrupted message rejected (#${this._corruptedMessageCount}): CRC32 checksum mismatch. ` +
+            console.warn(`[GenericProvider] 💥 Corrupted message rejected: CRC32 checksum mismatch. ` +
                 `This is expected if data corruption simulation is enabled.`);
-            // Request re-sync to recover any lost data
-            // Use exponential backoff: 100ms, 500ms, 2.5s, then cap at 5s
-            const delay = Math.min(5000, 100 * Math.pow(5, Math.min(this._corruptedMessageCount - 1, 3)));
-            setTimeout(() => {
-                if (this.transport.isConnected && !this._destroying) {
-                    this._sendSyncStep1();
-                }
-            }, delay);
+            // Request re-sync to recover any lost data - routed through the
+            // shared coordinator so this doesn't stack an independent timer on
+            // top of any hash-mismatch/gap-confirmed resync already pending.
+            this._requestResync();
             return; // Don't process corrupted message
         }
         // Message integrity verified - safe to decode
@@ -580,20 +673,35 @@ export class GenericProvider extends Observable {
                     const encoder = encoding.createEncoder();
                     encoding.writeVarUint(encoder, MESSAGE_SYNC);
                     const syncMessageType = syncProtocol.readSyncMessage(decoder, encoder, this.doc, this);
-                    // If we received SyncStep2, we're synced
-                    if (syncMessageType === syncProtocol.messageYjsSyncStep2 &&
-                        !this._synced) {
-                        this._synced = true;
-                        this.emit('synced', [true]);
+                    if (syncMessageType === syncProtocol.messageYjsSyncStep2) {
+                        // If we received SyncStep2, we're synced
+                        if (!this._synced) {
+                            this._synced = true;
+                            this.emit('synced', [true]);
+                        }
+                        // Someone else's SyncStep2 reply just arrived - our own pending
+                        // reply (if any) is now most likely redundant.
+                        this._cancelPendingSyncReply();
                     }
-                    // Send reply if needed
+                    // Send reply if needed. Suppression only engages with genuine
+                    // redundancy (>=2 other known peers via awareness) - below that,
+                    // there's no "someone else" to rely on, so reply immediately
+                    // (still rate-limited via _sendSyncReply() as a hard backstop).
                     if (encoding.length(encoder) > 1) {
-                        this._send(encoding.toUint8Array(encoder));
+                        if (this.awareness.getStates().size >= 3) {
+                            this._scheduleSyncReply(encoding.toUint8Array(encoder));
+                        }
+                        else {
+                            this._sendSyncReply(encoding.toUint8Array(encoder));
+                        }
                     }
                     break;
                 }
                 case MESSAGE_AWARENESS: {
-                    awarenessProtocol.applyAwarenessUpdate(this.awareness, decoding.readVarUint8Array(decoder), this);
+                    const channel = decoding.readVarUint(decoder);
+                    awarenessProtocol.applyAwarenessUpdate(channel === AWARENESS_CHANNEL_APP
+                        ? this.appAwareness
+                        : this.awareness, decoding.readVarUint8Array(decoder), this);
                     break;
                 }
                 case MESSAGE_PUBSUB: {
@@ -636,58 +744,61 @@ export class GenericProvider extends Observable {
                     // Read sequence number and clientID first
                     const seqNum = decoding.readVarUint(decoder);
                     const senderClientID = decoding.readVarUint(decoder);
-                    // Check for duplicate or out-of-order updates
-                    const lastSeq = this._remoteSeqNums.get(senderClientID) ?? -1;
-                    if (seqNum <= lastSeq) {
-                        console.warn(`[GenericProvider] Duplicate or out-of-order update detected from client ${senderClientID}: seqNum ${seqNum} <= lastSeen ${lastSeq}`);
-                        // Skip this update - it's a duplicate or we already have newer data
-                        break;
-                    }
-                    // Check for sequence gap (potential packet loss)
-                    if (lastSeq >= 0 && seqNum > lastSeq + 1) {
-                        const gapSize = seqNum - lastSeq - 1;
-                        console.warn(`[GenericProvider] Sequence gap detected from client ${senderClientID}: expected ${lastSeq + 1}, got ${seqNum} (gap of ${gapSize} messages)`);
-                        // Immediately request sync to recover missing updates
-                        // This is more proactive than waiting for periodic sync or hash mismatch
-                        this._sendSyncStep1();
-                    }
-                    // Update sequence tracker
-                    this._remoteSeqNums.set(senderClientID, seqNum);
-                    // Create encoder for reply with standard MESSAGE_SYNC header
-                    // (replies don't need verification since they're generated immediately)
+                    // Track for gap detection only — does NOT gate whether we apply
+                    // the update below (see _trackRemoteSeq() for why).
+                    this._trackRemoteSeq(senderClientID, seqNum);
+                    // Always apply the update. Yjs updates are idempotent/commutative,
+                    // so re-applying an already-seen update is a harmless no-op.
+                    // Under reordering, a merely-late (not actually duplicate) update
+                    // must still be applied here — the old "skip if seqNum <= last
+                    // seen" logic silently dropped such updates forever whenever a
+                    // later-numbered message happened to arrive first.
                     const encoder = encoding.createEncoder();
                     encoding.writeVarUint(encoder, MESSAGE_SYNC);
                     const syncMessageType = syncProtocol.readSyncMessage(decoder, encoder, this.doc, this);
+                    // Someone else's SyncStep2 reply just arrived - our own pending
+                    // reply (if any) is now most likely redundant. Mirrors the
+                    // MESSAGE_SYNC case: the reply encoded above is always a plain
+                    // MESSAGE_SYNC-typed message regardless of which message type
+                    // triggered it, so the same suppression scheme applies here too.
+                    if (syncMessageType === syncProtocol.messageYjsSyncStep2) {
+                        this._cancelPendingSyncReply();
+                    }
                     // Read the expected hash from sender (signed integer)
                     const expectedHash = decoding.readVarInt(decoder);
                     // Compute our local hash after applying the update
                     const localHash = computeDocHash(this.doc);
                     // Verify hash match
                     if (localHash !== expectedHash) {
-                        this._hashMismatchCount++;
-                        const now = Date.now();
-                        // Reset counter if it's been stable for 10 seconds
-                        if (now - this._lastHashMismatchTime > 10000) {
-                            this._hashMismatchCount = 1;
+                        // If we already know this sender has a suspected reordering gap
+                        // (see _trackRemoteSeq()/_scheduleGapCheck()), a hash mismatch
+                        // right now is the *expected* transient state — we're missing a
+                        // piece that's very likely still in flight, not actually
+                        // diverged. Let the pending gap-check grace period resolve it
+                        // instead of also escalating the hash-mismatch backoff: under
+                        // heavy reordering this previously caused a burst of mismatches
+                        // to rack up the exponential backoff to its 10s cap within a
+                        // single edit burst, purely from timing, not real divergence.
+                        // A hash mismatch with NO pending gap (in-order, but still
+                        // wrong) is not explained by reordering and still escalates
+                        // normally below.
+                        const reorderingSuspected = this._gapCheckTimers.has(senderClientID);
+                        if (!reorderingSuspected) {
+                            // Push our full state AND request theirs (syncNow() does
+                            // both). A hash mismatch means the two peers have diverged -
+                            // one side may have edits the other lacks. Routed through the
+                            // shared coordinator so this doesn't stack an independent
+                            // timer on top of any corrupted-message/gap-confirmed resync
+                            // already pending.
+                            this._requestResync();
+                            // Logged with the shared attempt counter (kept as "#N" for
+                            // compatibility with existing tooling/benchmarks that grep
+                            // for this exact "Hash mismatch #" pattern) - it now reflects
+                            // the unified resync-attempt count rather than a
+                            // hash-mismatch-specific one, since the two escalation
+                            // counters were merged.
+                            console.warn(`[GenericProvider] Hash mismatch #${this._resyncAttemptCount} detected! Local: ${localHash}, Expected: ${expectedHash}`);
                         }
-                        this._lastHashMismatchTime = now;
-                        // Exponential backoff: 10ms, 50ms, 250ms, 1.25s, 6.25s, then cap at 10s
-                        const delay = Math.min(10000, 10 * Math.pow(5, this._hashMismatchCount - 1));
-                        console.warn(`[GenericProvider] Hash mismatch #${this._hashMismatchCount} detected! Local: ${localHash}, Expected: ${expectedHash}`);
-                        console.warn(`[GenericProvider] Re-sync scheduled in ${delay}ms...`);
-                        // Push our full state AND request theirs.
-                        // A hash mismatch means the two peers have diverged — one side may
-                        // have edits the other lacks.  Calling only _sendSyncStep1() (pull)
-                        // never delivers our own surplus edits to the other side.
-                        setTimeout(() => {
-                            if (this.transport.isConnected && !this._destroying) {
-                                this.syncNow();
-                            }
-                        }, delay);
-                    }
-                    else {
-                        // Hash matched - reset failure counter
-                        this._hashMismatchCount = 0;
                     }
                     // If we received SyncStep2, we're synced (unless hash mismatched)
                     if (syncMessageType === syncProtocol.messageYjsSyncStep2 &&
@@ -696,9 +807,21 @@ export class GenericProvider extends Observable {
                         this._synced = true;
                         this.emit('synced', [true]);
                     }
-                    // Send reply if needed (as standard MESSAGE_SYNC)
+                    // Send reply if needed (as standard MESSAGE_SYNC). Suppression
+                    // only engages with genuine redundancy (>=2 other known peers via
+                    // awareness) - below that, reply immediately (still rate-limited
+                    // via _sendSyncReply() as a hard backstop). Matches the
+                    // MESSAGE_SYNC case's gate exactly; without this, a hash-mismatch
+                    // resync burst under packet loss bypassed suppression entirely,
+                    // since every peer answering a post-mismatch SyncStep1 replied
+                    // immediately via this path.
                     if (encoding.length(encoder) > 1) {
-                        this._send(encoding.toUint8Array(encoder));
+                        if (this.awareness.getStates().size >= 3) {
+                            this._scheduleSyncReply(encoding.toUint8Array(encoder));
+                        }
+                        else {
+                            this._sendSyncReply(encoding.toUint8Array(encoder));
+                        }
                     }
                     break;
                 }
@@ -713,22 +836,210 @@ export class GenericProvider extends Observable {
         }
     }
     /**
-     * Send SyncStep1 message to request missing updates.
-     * This is sent when first connecting to sync with remote peers.
-     * Note: SyncStep1 is just a request and doesn't include hash verification.
-     * Rate limited to prevent spam.
+     * Schedule a SyncStep2 reply after a short random delay instead of
+     * sending immediately. If another peer's reply is overheard in the
+     * meantime (`_cancelPendingSyncReply`), this reply is dropped as
+     * redundant - the requester likely already got what it needed.
+     *
+     * A reply that is already pending when this is called answers a
+     * *different* SyncStep1 request (e.g. peer A's request, followed 5ms
+     * later by peer B's) - it must not be silently overwritten by the new
+     * one. Flush it immediately, then schedule the new reply fresh. The only
+     * sanctioned way a reply gets dropped is `_cancelPendingSyncReply()`,
+     * because we overheard someone else's SyncStep2 for the SAME request.
      */
-    _sendSyncStep1() {
+    _scheduleSyncReply(reply) {
+        if (this._pendingSyncReplyTimeoutId !== undefined) {
+            if (this._pendingSyncReply) {
+                this._sendSyncReply(this._pendingSyncReply);
+            }
+            clearTimeout(this._pendingSyncReplyTimeoutId);
+            this._pendingSyncReplyTimeoutId = undefined;
+        }
+        this._pendingSyncReply = reply;
+        const delay = Math.random() * this._syncReplySuppressionMs;
+        this._pendingSyncReplyTimeoutId = setTimeout(() => {
+            this._pendingSyncReplyTimeoutId = undefined;
+            if (this._pendingSyncReply) {
+                this._sendSyncReply(this._pendingSyncReply);
+                this._pendingSyncReply = null;
+            }
+        }, delay);
+    }
+    /** Cancel a pending suppressed reply, if any. */
+    _cancelPendingSyncReply() {
+        if (this._pendingSyncReplyTimeoutId !== undefined) {
+            clearTimeout(this._pendingSyncReplyTimeoutId);
+            this._pendingSyncReplyTimeoutId = undefined;
+        }
+        this._pendingSyncReply = null;
+    }
+    /**
+     * Send a SyncStep2 reply, gated by the same shared per-peer budget as
+     * SyncStep1 requests/syncNow() pushes (`_tryReserveSyncSlot()`).
+     *
+     * Previously SyncStep2 replies were completely unrated - the only
+     * defense against redundant replies was the best-effort NACK-style
+     * suppression in `_scheduleSyncReply()`/`_cancelPendingSyncReply()`,
+     * which itself is just an ordinary broadcast message subject to the same
+     * wire corruption as everything else. Under sustained corruption, more
+     * competing repliers independently miss the "someone already answered"
+     * signal as peer count grows, and none of that traffic was bounded.
+     * Measured in test/dummy/bench-corruption-storm.ts: SyncStep2/SyncStep1
+     * ratio grew from ~1.1-1.3 at N=2 to ~4.5-5.9 at N=10 (should stay near
+     * 1 if suppression alone were sufficient). This is a hard backstop on
+     * top of that suppression, not a replacement for it - a rate-limited
+     * reply is dropped silently (no warn) since under normal, uncorrupted
+     * operation this path is rarely exercised and logging every drop here
+     * would itself become log spam exactly when things are already noisy.
+     */
+    _sendSyncReply(reply) {
+        if (!this._tryReserveSyncSlot()) {
+            return; // Rate limited - drop the reply silently
+        }
+        this._send(reply);
+    }
+    /**
+     * Track a received sequence number for reordering-tolerant gap detection.
+     * Does not gate whether the update gets applied — only decides whether a
+     * gap looks suspicious enough to (eventually) request a resync.
+     */
+    _trackRemoteSeq(senderClientID, seqNum) {
+        let info = this._remoteSeqInfo.get(senderClientID);
+        if (!info) {
+            info = { highest: -1, seen: new Set() };
+            this._remoteSeqInfo.set(senderClientID, info);
+        }
+        if (info.seen.has(seqNum)) {
+            return; // genuine duplicate - nothing new to track
+        }
+        info.seen.add(seqNum);
+        if (seqNum > info.highest) {
+            if (info.highest >= 0 && seqNum > info.highest + 1) {
+                this._scheduleGapCheck(senderClientID, info.highest + 1, seqNum - 1);
+            }
+            info.highest = seqNum;
+        }
+        // Bound memory: forget seqNums far behind the current high-water mark.
+        const floor = info.highest - this._seqWindowSize;
+        if (floor > 0) {
+            for (const s of info.seen) {
+                if (s < floor)
+                    info.seen.delete(s);
+            }
+        }
+    }
+    /**
+     * Re-check a suspected sequence gap after a short grace period instead of
+     * requesting a resync immediately. Pure network reordering (a message
+     * that's merely late, not lost) typically resolves itself within the
+     * grace window, so this avoids the resync storms that immediate gap
+     * detection caused under jitter. Real packet loss still gets caught —
+     * just `_gapGraceMs` later — and the periodic sync interval / hash
+     * verification remain as further safety nets regardless.
+     */
+    _scheduleGapCheck(clientID, gapStart, gapEnd) {
+        // Only one pending check per sender; a newly-opened gap while a check
+        // is already scheduled will still be caught by periodic sync / hash
+        // verification even if not by this specific check.
+        if (this._gapCheckTimers.has(clientID))
+            return;
+        const timer = setTimeout(() => {
+            this._gapCheckTimers.delete(clientID);
+            const info = this._remoteSeqInfo.get(clientID);
+            if (!info || this._destroying)
+                return;
+            let stillMissing = 0;
+            for (let s = gapStart; s <= gapEnd; s++) {
+                if (!info.seen.has(s))
+                    stillMissing++;
+            }
+            if (stillMissing > 0 && this.transport.isConnected) {
+                console.warn(`[GenericProvider] Sequence gap confirmed from client ${clientID}: ` +
+                    `${stillMissing} message(s) still missing after ${this._gapGraceMs}ms grace period`);
+                // Routed through the shared coordinator (previously called
+                // _sendSyncStep1() directly with NO coalescing at all - the one
+                // remaining gap that let this trigger steal rate-limit slots
+                // independently of the hash-mismatch/corrupted-message triggers).
+                this._requestResync();
+            }
+        }, this._gapGraceMs);
+        this._gapCheckTimers.set(clientID, timer);
+    }
+    /**
+     * Unified entry point for ALL resync triggers (hash mismatch, corrupted
+     * message, confirmed sequence gap). Coalesces them behind a single
+     * pending timer and a single shared escalation counter, so a burst of
+     * triggers from different causes in a short window schedules exactly one
+     * resync instead of three independent ones each able to draw on the
+     * shared `_tryReserveSyncSlot()` budget on their own.
+     *
+     * Always resolves to `syncNow()` (push + pull) rather than distinguishing
+     * a push-only/pull-only variant per trigger. `syncNow()`'s push half is
+     * already a no-op when there's nothing to send (it only calls
+     * `_sendUpdate()` when `update.length > 0`), so unifying on push+pull is
+     * strictly simpler than threading a `push` flag through a *shared*
+     * coordinator (where the "right" answer for an absorbed trigger is
+     * ambiguous anyway - was it push-worthy or not?). It also closes a latent
+     * gap where the corrupted-message and gap-confirmed triggers previously
+     * called pull-only `_sendSyncStep1()` and could never deliver this peer's
+     * own surplus edits made during a divergence window.
+     */
+    _requestResync() {
+        // Coalesced: if a resync is already pending (regardless of which
+        // trigger scheduled it), this trigger is absorbed into it instead of
+        // stacking another independent timer/broadcast. Escalation only
+        // advances when we actually schedule a NEW timer below - incrementing
+        // unconditionally here (once per absorbed trigger too) would let a
+        // burst of many corrupted/mismatched messages while one resync is
+        // already pending ratchet the counter straight to its cap, so the
+        // *next* resync (after this one fires) always schedules at the max
+        // backoff instead of escalating gradually.
+        if (this._pendingResyncTimeoutId !== undefined) {
+            return;
+        }
+        this._resyncAttemptCount++;
+        const now = Date.now();
+        // Reset the escalation counter if it's been stable for 10 seconds -
+        // same quiet-period reset the old per-trigger counters used.
+        if (now - this._lastResyncAttemptTime > 10000) {
+            this._resyncAttemptCount = 1;
+        }
+        this._lastResyncAttemptTime = now;
+        // Exponential backoff: 100ms, 500ms, 2.5s, then cap at 5s.
+        const delay = Math.min(5000, 100 * Math.pow(5, Math.min(this._resyncAttemptCount - 1, 3)));
+        console.warn(`[GenericProvider] Resync scheduled in ${delay}ms (attempt #${this._resyncAttemptCount})...`);
+        this._pendingResyncTimeoutId = setTimeout(() => {
+            this._pendingResyncTimeoutId = undefined;
+            if (this.transport.isConnected && !this._destroying) {
+                this.syncNow();
+            }
+        }, delay);
+    }
+    /**
+     * Reserve a slot in the sync rate limiter (max `_maxSyncRequestsPerWindow`
+     * per `_syncRequestWindowMs`), recording the request if there's room.
+     * Shared by `_sendSyncStep1()` and `syncNow()` so a burst of triggers from
+     * different sources (periodic sync, hash-mismatch resyncs, gap-check
+     * confirmations) draws from one combined budget instead of each having
+     * its own uncapped or separately-capped allowance.
+     */
+    _tryReserveSyncSlot() {
         const now = Date.now();
         // Clean up old entries outside the rate limit window
         this._syncRequestTimes = this._syncRequestTimes.filter((t) => now - t < this._syncRequestWindowMs);
-        // Check rate limit
         if (this._syncRequestTimes.length >= this._maxSyncRequestsPerWindow) {
-            console.warn(`[GenericProvider] Sync rate limit exceeded (${this._maxSyncRequestsPerWindow} requests per ${this._syncRequestWindowMs / 1000}s), throttling...`);
-            return; // Drop the request
+            return false;
         }
-        // Record this request
         this._syncRequestTimes.push(now);
+        return true;
+    }
+    /**
+     * Encode and send a SyncStep1 message requesting missing updates.
+     * Does not check the rate limiter itself - callers must reserve a slot
+     * via `_tryReserveSyncSlot()` first.
+     */
+    _writeSyncStep1() {
         const encoder = encoding.createEncoder();
         // SyncStep1 is always sent as standard MESSAGE_SYNC (no verification)
         // It's just a request, not an assertion of state
@@ -736,6 +1047,19 @@ export class GenericProvider extends Observable {
         syncProtocol.writeSyncStep1(encoder, this.doc);
         this._send(encoding.toUint8Array(encoder));
     }
+    /**
+     * Send SyncStep1 message to request missing updates.
+     * This is sent when first connecting to sync with remote peers.
+     * Note: SyncStep1 is just a request and doesn't include hash verification.
+     * Rate limited to prevent spam.
+     */
+    _sendSyncStep1() {
+        if (!this._tryReserveSyncSlot()) {
+            console.warn(`[GenericProvider] Sync rate limit exceeded (${this._maxSyncRequestsPerWindow} requests per ${this._syncRequestWindowMs / 1000}s), throttling...`);
+            return; // Drop the request
+        }
+        this._writeSyncStep1();
+    }
     /**
      * Send a document update to the transport.
      * If verifyUpdates is enabled, includes sequence number and document hash for ordering and desync detection.
@@ -760,15 +1084,6 @@ export class GenericProvider extends Observable {
         }
         this._send(encoding.toUint8Array(encoder));
     }
-    /**
-     * Send awareness update to the transport.
-     */
-    _sendAwarenessUpdate(changedClients) {
-        const encoder = encoding.createEncoder();
-        encoding.writeVarUint(encoder, MESSAGE_AWARENESS);
-        encoding.writeVarUint8Array(encoder, awarenessProtocol.encodeAwarenessUpdate(this.awareness, changedClients));
-        this._send(encoding.toUint8Array(encoder));
-    }
     /**
      * Send a pub/sub message.
      * Internal method called by PubSubChannel.
@@ -839,45 +1154,62 @@ export class GenericProvider extends Observable {
      * Throttled to prevent awareness updates from flooding document sync.
      * Multiple rapid updates are batched together.
      */
-    _broadcastAwareness(clients) {
+    _broadcastAwareness(clients, channel = AWARENESS_CHANNEL_MAIN) {
         if (clients.length === 0)
             return;
+        const isApp = channel === AWARENESS_CHANNEL_APP;
+        const pending = isApp
+            ? this._pendingAppAwarenessClients
+            : this._pendingAwarenessClients;
         // If throttling is disabled, send immediately
         if (this._awarenessInterval <= 0) {
-            this._sendAwarenessNow(clients);
+            this._sendAwarenessNow(clients, channel);
             return;
         }
         // Add clients to pending set
         for (const client of clients) {
-            this._pendingAwarenessClients.add(client);
+            pending.add(client);
         }
         // If we already have a scheduled broadcast, let it handle the batched clients
-        if (this._awarenessTimeoutId !== undefined) {
+        if ((isApp ? this._appAwarenessTimeoutId : this._awarenessTimeoutId) !==
+            undefined) {
             return;
         }
         // Calculate delay - respect minimum interval since last broadcast
         const now = Date.now();
-        const timeSinceLastBroadcast = now - this._lastAwarenessTime;
-        const delay = Math.max(0, this._awarenessInterval - timeSinceLastBroadcast);
+        const lastTime = isApp ? this._lastAppAwarenessTime : this._lastAwarenessTime;
+        const delay = Math.max(0, this._awarenessInterval - (now - lastTime));
         // Schedule the batched broadcast
-        this._awarenessTimeoutId = setTimeout(() => {
-            this._awarenessTimeoutId = undefined;
-            this._lastAwarenessTime = Date.now();
+        const timeoutId = setTimeout(() => {
+            if (isApp) {
+                this._appAwarenessTimeoutId = undefined;
+                this._lastAppAwarenessTime = Date.now();
+            }
+            else {
+                this._awarenessTimeoutId = undefined;
+                this._lastAwarenessTime = Date.now();
+            }
             // Send all pending clients in one message
-            const clientsToSend = Array.from(this._pendingAwarenessClients);
-            this._pendingAwarenessClients.clear();
+            const clientsToSend = Array.from(pending);
+            pending.clear();
             if (clientsToSend.length > 0) {
-                this._sendAwarenessNow(clientsToSend);
+                this._sendAwarenessNow(clientsToSend, channel);
             }
         }, delay);
+        if (isApp)
+            this._appAwarenessTimeoutId = timeoutId;
+        else
+            this._awarenessTimeoutId = timeoutId;
     }
     /**
      * Send awareness update immediately without throttling.
      */
-    _sendAwarenessNow(clients) {
+    _sendAwarenessNow(clients, channel = AWARENESS_CHANNEL_MAIN) {
+        const aw = channel === AWARENESS_CHANNEL_APP ? this.appAwareness : this.awareness;
         const encoder = encoding.createEncoder();
         encoding.writeVarUint(encoder, MESSAGE_AWARENESS);
-        encoding.writeVarUint8Array(encoder, awarenessProtocol.encodeAwarenessUpdate(this.awareness, clients));
+        encoding.writeVarUint(encoder, channel);
+        encoding.writeVarUint8Array(encoder, awarenessProtocol.encodeAwarenessUpdate(aw, clients));
         this._send(encoding.toUint8Array(encoder));
     }
     /**
@@ -920,13 +1252,21 @@ export class GenericProvider extends Observable {
         bc.publish(this._bcChannel, wrapMessageWithChecksum(encoding.toUint8Array(encoderState)), this);
         // Broadcast local awareness state via BroadcastChannel (wrapped with CRC32)
         if (this.awareness.getLocalState() !== null) {
-            const encoderAwareness = encoding.createEncoder();
-            encoding.writeVarUint(encoderAwareness, MESSAGE_AWARENESS);
-            encoding.writeVarUint8Array(encoderAwareness, awarenessProtocol.encodeAwarenessUpdate(this.awareness, [
-                this.doc.clientID,
-            ]));
-            bc.publish(this._bcChannel, wrapMessageWithChecksum(encoding.toUint8Array(encoderAwareness)), this);
+            this._publishAwarenessToBroadcastChannel(this.awareness, AWARENESS_CHANNEL_MAIN);
         }
+        if (this.appAwareness.getLocalState() !== null) {
+            this._publishAwarenessToBroadcastChannel(this.appAwareness, AWARENESS_CHANNEL_APP);
+        }
+    }
+    /**
+     * Encode and publish an awareness update for the local client to the BroadcastChannel.
+     */
+    _publishAwarenessToBroadcastChannel(awareness, channel) {
+        const encoder = encoding.createEncoder();
+        encoding.writeVarUint(encoder, MESSAGE_AWARENESS);
+        encoding.writeVarUint(encoder, channel);
+        encoding.writeVarUint8Array(encoder, awarenessProtocol.encodeAwarenessUpdate(awareness, [this.doc.clientID]));
+        bc.publish(this._bcChannel, wrapMessageWithChecksum(encoding.toUint8Array(encoder)), this);
     }
     /**
      * Disconnect from BroadcastChannel and mark local client as offline.
@@ -936,10 +1276,16 @@ export class GenericProvider extends Observable {
             return;
         }
         // Broadcast awareness state with null (indicating disconnect) - wrapped with CRC32
-        const encoder = encoding.createEncoder();
-        encoding.writeVarUint(encoder, MESSAGE_AWARENESS);
-        encoding.writeVarUint8Array(encoder, awarenessProtocol.encodeAwarenessUpdate(this.awareness, [this.doc.clientID], new Map()));
-        bc.publish(this._bcChannel, wrapMessageWithChecksum(encoding.toUint8Array(encoder)), this);
+        for (const [awareness, channel] of [
+            [this.awareness, AWARENESS_CHANNEL_MAIN],
+            [this.appAwareness, AWARENESS_CHANNEL_APP],
+        ]) {
+            const encoder = encoding.createEncoder();
+            encoding.writeVarUint(encoder, MESSAGE_AWARENESS);
+            encoding.writeVarUint(encoder, channel);
+            encoding.writeVarUint8Array(encoder, awarenessProtocol.encodeAwarenessUpdate(awareness, [this.doc.clientID], new Map()));
+            bc.publish(this._bcChannel, wrapMessageWithChecksum(encoding.toUint8Array(encoder)), this);
+        }
         // Unsubscribe from channel
         bc.unsubscribe(this._bcChannel, this._bcSubscriber);
         this._bcConnected = false;