@edryslabs/genericprovider 1.0.4 → 1.5.2

package/dist/index.js

@@ -10,10 +10,83 @@ const MESSAGE_SYNC = 0;
 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;
+const MESSAGE_BATCH = 4; // Envelope for N independently-typed sub-messages sent as one wire message
+// Digest beacon: replaces SyncStep1 on the wire. [version][flags][sender
+// clientID][state vector][delete-set hash]. Receivers reply only when the
+// sender is behind them or the delete-set hashes differ (SyncStep2, as
+// before), or - on a JOIN-flagged beacon with equal state - with their own
+// beacon as an ack; otherwise not at all. See
+// docs/superpowers/specs/2026-09-05-digest-beacon-design.md and
+// _handleDigest(). Versions are append-only: receivers read the fields
+// they know and ignore trailing bytes.
+const MESSAGE_SYNC_DIGEST = 5;
+const DIGEST_VERSION = 1;
+const DIGEST_FLAG_JOIN = 1; // bit 0: "I just joined - send me your presence and confirm my state"
+// bit 1: this message is an ACK, not a request. It echoes the state vector
+// and delete-set hash of the JOIN beacon it answers ("I hold exactly this
+// state too"), never the sender's own state - so it is never read as a
+// request and never collects a SyncStep2 from anyone. Receivers whose state
+// equals the echoed digest mark themselves synced (the joiner it was for,
+// and everyone else in that state); receivers holding a pending ack for the
+// same digest drop it. See _handleDigest(). Before this flag existed the ack
+// was the acker's own beacon; an acker that was itself behind the room (a
+// joiner acking another joiner) made every peer ahead of it reply with a
+// SyncStep2, and acks landing after an edit burst had started did the same
+// (Task 3c in the design doc).
+const DIGEST_FLAG_ACK = 2;
+// bit 2: "ack me if our states are equal" WITHOUT the presence request of
+// DIGEST_FLAG_JOIN. Sent only by the response-wait retry
+// (_armResponseWait): a joiner that gets neither a SyncStep2 nor an ack
+// within the wait has lost one message or the other and asks again - reply
+// suppression deliberately leaves ~1 reply per request, so a single lost
+// reply would otherwise strand the joiner's `synced` until the next
+// periodic beacon (never, with syncInterval 0). Measured: the last joiner
+// of a simultaneous 5-peer burst at 10% loss failed to reach `synced` in 1
+// of 600 runs before this existed. Resync beacons do NOT request acks: in
+// an equal room every peer would answer, and at high latency the
+// suppression window cannot thin those replies (measured 4-5x more acks).
+const DIGEST_FLAG_CONFIRM = 4;
+// bit 3: the sender is CONFIRMED - it has received a SyncStep2, or an equal
+// digest from a confirmed peer, or its own join asked three times and got
+// nothing better (so it is the room). A joiner's response wait is satisfied
+// only by data (SyncStep2) or by an equal digest carrying this bit: an
+// equal ack from a fellow joiner says nothing about whether the room holds
+// more than both of us, and treating it as an answer left a joiner whose
+// SyncStep2 was lost with an empty document (measured: 1 of 150 lossy
+// 15-peer joins; research doc item 13 follow-up in the phase-1b design).
+const DIGEST_FLAG_SETTLED = 8;
+// Full-state push (connect()/syncNow()): the whole document as one update,
+// no hash, no sequence number, applied and nothing else. Before this type a
+// push was a MESSAGE_SYNC_VERIFIED update whose hash was the PUSHER's state
+// - every peer holding more data read that as divergence and scheduled a
+// resync of its own (research doc item 12), the seed of the cascade in item
+// 13. Not a SyncStep2 either: receiving one must not flip `synced` (an
+// empty joiner's push says nothing about the room's content). See
+// docs/superpowers/specs/2026-09-05-resync-cascade-design.md.
+const MESSAGE_SYNC_PUSH = 6;
+// Pub/sub message aimed at a single target: [target][topic][message]. Every
+// provider whose `localId` differs drops it. On transports with sendTo it is
+// unicast; elsewhere it is broadcast and filtered on receipt.
+//
+// NOTE: this was type 4 before the v1.5.0 merge, which upstream had taken for
+// MESSAGE_BATCH. Renumbered to 7 (the first free slot) - the two are not wire
+// compatible, so all peers of a room must run the same version, as the README
+// already requires.
+const MESSAGE_PUBSUB_TARGETED = 7;
+// A second, independent awareness channel for application/module state
+// (cursors, module presence), kept off MESSAGE_AWARENESS deliberately.
+//
+// Upstream's MESSAGE_AWARENESS receive path is now core presence machinery:
+// it feeds _knownPeers, sets _presenceCovered, and cancels pending removal
+// broadcasts. App awareness is written by untrusted third-party classroom
+// modules, so routing it through that path would let module cursor traffic
+// drive the room's peer bookkeeping - a module setting a state would mark
+// phantom peers present and suppress real removals. A separate type keeps
+// the two entirely disjoint: same throttle policy, no shared presence state.
+//
+// (Dev encoded this as a channel varint inside MESSAGE_AWARENESS, which was
+// safe when that path did nothing but applyAwarenessUpdate. It no longer is.)
+const MESSAGE_AWARENESS_APP = 8;
 /**
  * CRC32 lookup table for fast computation.
  * Generated once and reused for all CRC calculations.
@@ -80,25 +153,105 @@ function unwrapAndVerifyMessage(wrapped) {
     }
     return message;
 }
+/** Byte-wise equality of two Uint8Arrays. */
+function bytesEqual(a, b) {
+    if (a.length !== b.length)
+        return false;
+    for (let i = 0; i < a.length; i++) {
+        if (a[i] !== b[i])
+            return false;
+    }
+    return true;
+}
 /**
- * Compute a cheap hash of document state for desync detection.
- *
- * 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.
+ * Componentwise minimum of two encoded state vectors: the state a
+ * SyncStep2 must start from to serve both requesters. Clients missing from
+ * one side count as clock 0 and are omitted (omitted = 0 on the wire).
+ */
+function minStateVector(a, b) {
+    const ma = Y.decodeStateVector(a);
+    const mb = Y.decodeStateVector(b);
+    const out = new Map();
+    for (const [client, clock] of ma) {
+        const m = Math.min(clock, mb.get(client) ?? 0);
+        if (m > 0)
+            out.set(client, m);
+    }
+    return Y.encodeStateVector(out);
+}
+/**
+ * Whether this runtime has the Compression Streams API (Node 18+, all
+ * evergreen browsers). Checked once at module load; compressionThresholdBytes
+ * falls back to sending uncompressed (still flag-byte-prefixed, flag=0) if
+ * this is false, rather than throwing.
+ */
+const COMPRESSION_AVAILABLE = typeof CompressionStream !== 'undefined' &&
+    typeof DecompressionStream !== 'undefined';
+/** Drain a ReadableStream<Uint8Array> into a single concatenated Uint8Array. */
+async function _readAllChunks(readable) {
+    const chunks = [];
+    const reader = readable.getReader();
+    for (;;) {
+        const { done, value } = await reader.read();
+        if (done)
+            break;
+        chunks.push(value);
+    }
+    const total = chunks.reduce((sum, c) => sum + c.length, 0);
+    const out = new Uint8Array(total);
+    let offset = 0;
+    for (const chunk of chunks) {
+        out.set(chunk, offset);
+        offset += chunk.length;
+    }
+    return out;
+}
+/**
+ * Compress with deflate-raw (no gzip header/trailer - see
+ * compressionThresholdBytes's doc comment for why deflate-raw over gzip).
+ */
+async function compressDeflateRaw(data) {
+    const cs = new CompressionStream('deflate-raw');
+    const writer = cs.writable.getWriter();
+    // The writer promises reject too when the stream errors; the reader side
+    // already surfaces that error, and an unhandled rejection here crashes
+    // Node (seen with a corrupt deflate stream in the Nostr end-to-end test).
+    writer.write(data).catch(() => { });
+    writer.close().catch(() => { });
+    return _readAllChunks(cs.readable);
+}
+/** Inverse of compressDeflateRaw(). */
+async function decompressDeflateRaw(data) {
+    const ds = new DecompressionStream('deflate-raw');
+    const writer = ds.writable.getWriter();
+    writer.write(data).catch(() => { });
+    writer.close().catch(() => { });
+    return _readAllChunks(ds.readable);
+}
+/**
+ * Prepend a 1-byte compressed(1)/uncompressed(0) flag. Only used when
+ * compressionThresholdBytes is configured - see that option's doc comment
+ * for why this is a deliberate, opt-in wire-format change.
+ */
+function prefixCompressionFlag(flag, data) {
+    const out = new Uint8Array(1 + data.length);
+    out[0] = flag;
+    out.set(data, 1);
+    return out;
+}
+/**
+ * Compute a cheap hash of document state for verification.
  *
- * 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.)
+ * Hashes the state VECTOR (each client's clock), not the full document
+ * content — O(number of distinct clients) instead of O(document content
+ * size). `Y.encodeStateVector()` writes entries sorted by clientID, so the
+ * result is deterministic regardless of the internal Map's iteration order.
+ * Two peers can only reach the same state vector by having applied the same
+ * set of operations, so this still catches real content divergence; it just
+ * no longer re-serializes the entire document on every single update.
+ * (CRC32 already guards against wire corruption, and sequence tracking
+ * guards against reordering/loss — this hash is the last line of defense
+ * against logical divergence between peers.)
  */
 function computeDocHash(doc) {
     const state = Y.encodeStateVector(doc);
@@ -108,6 +261,109 @@ function computeDocHash(doc) {
     }
     return hash;
 }
+/**
+ * Cheap, peer-deterministic hash of the document's delete set - the half of
+ * a Yjs document's identity that the state vector does NOT cover (yjs
+ * INTERNALS.md: "deletions are tracked in the DeleteSet, and do not update
+ * the state vector"). Two docs that differ only by a lost delete-only
+ * update have identical state vectors, so `computeDocHash` can never
+ * detect that divergence; this hash can, at heartbeat granularity (see
+ * `_encodeSyncStep1()` / `_handleDigest()`).
+ *
+ * Cost: `Y.createDeleteSetFromStructStore` walks every struct (Yjs keeps no
+ * incremental delete set), so this is O(items) - fine once per heartbeat
+ * (every empty SyncStep2 already did this exact walk inside
+ * `encodeStateAsUpdate`), NOT fine per update; hence the cache in
+ * `_deleteSetHash()`. Per-client runs come out already sorted and merged;
+ * the only per-peer non-determinism is `Map` insertion order, fixed by
+ * sorting client IDs before hashing.
+ *
+ * Exported for the property check in test/dummy/bench-idle-room.ts.
+ * @internal
+ */
+export function computeDeleteSetHash(doc) {
+    const ds = Y.createDeleteSetFromStructStore(doc.store);
+    const encoder = encoding.createEncoder();
+    const clients = Array.from(ds.clients.keys()).sort((x, y) => x - y);
+    for (const client of clients) {
+        encoding.writeVarUint(encoder, client);
+        for (const item of ds.clients.get(client)) {
+            encoding.writeVarUint(encoder, item.clock);
+            encoding.writeVarUint(encoder, item.len);
+        }
+    }
+    return computeCRC32(encoding.toUint8Array(encoder));
+}
+/**
+ * The Yjs updates a CRC-wrapped frame (as handed to `Transport.send`)
+ * carries: the update of a MESSAGE_SYNC_VERIFIED or MESSAGE_SYNC
+ * Update/SyncStep2 message, the whole document of a MESSAGE_SYNC_PUSH,
+ * each such sub-message of a MESSAGE_BATCH - and nothing for awareness,
+ * pub/sub, digest beacons and SyncStep1 requests, which carry no document
+ * state. For persistence transports (`providers/indexeddb`, LiaScript's
+ * Dexie cache): store only what this returns, and only this. Storing whole
+ * frames and replaying them on the next load resurrects the previous
+ * session's clientID as a phantom peer (its presence, its beacons - which
+ * the provider then answers into the store, multiplying rows) and keeps
+ * ~10x the bytes (a keystroke's frame carries its cursor). A frame this
+ * cannot parse yields `[]`, never a partial read. Frames of a provider with
+ * `compressionThresholdBytes` set (a leading flag byte) are not supported.
+ * Round 7, item 6; measured in test/dummy/bench-persist-log.ts.
+ */
+export function extractDocUpdates(frame) {
+    const updates = [];
+    const walk = (message) => {
+        const decoder = decoding.createDecoder(message);
+        const type = decoding.readVarUint(decoder);
+        switch (type) {
+            case MESSAGE_BATCH:
+                while (decoding.hasContent(decoder))
+                    walk(decoding.readVarUint8Array(decoder));
+                break;
+            case MESSAGE_SYNC_VERIFIED:
+            case MESSAGE_SYNC: {
+                if (type === MESSAGE_SYNC_VERIFIED) {
+                    decoding.readVarUint(decoder); // sequence number
+                    decoding.readVarUint(decoder); // sender clientID
+                }
+                const sub = decoding.readVarUint(decoder);
+                if (sub === syncProtocol.messageYjsSyncStep2 ||
+                    sub === syncProtocol.messageYjsUpdate) {
+                    updates.push(decoding.readVarUint8Array(decoder));
+                }
+                break;
+            }
+            case MESSAGE_SYNC_PUSH:
+                updates.push(decoding.readVarUint8Array(decoder));
+                break;
+            default:
+                break;
+        }
+    };
+    try {
+        if (frame.length > 4)
+            walk(frame.subarray(4));
+    }
+    catch {
+        return [];
+    }
+    return updates;
+}
+/**
+ * The counterpart of `extractDocUpdates()` for the load path of a
+ * persistence transport: wraps one (merged) update as a CRC-wrapped
+ * MESSAGE_SYNC SyncStep2 frame. Handed to the `onMessage` callback, the
+ * provider applies it as the answer to its own request - `synced` fires,
+ * nothing is sent back - exactly what a local copy is: the peer that had
+ * our document.
+ */
+export function frameDocUpdate(update) {
+    const encoder = encoding.createEncoder();
+    encoding.writeVarUint(encoder, MESSAGE_SYNC);
+    encoding.writeVarUint(encoder, syncProtocol.messageYjsSyncStep2);
+    encoding.writeVarUint8Array(encoder, update);
+    return wrapMessageWithChecksum(encoding.toUint8Array(encoder));
+}
 /**
  * PubSub channel for real-time messaging alongside Yjs.
  * Allows sending ephemeral messages that don't need CRDT properties.
@@ -207,6 +463,18 @@ export class PubSubChannel extends Observable {
  * ```
  */
 export class GenericProvider extends Observable {
+    get appAwareness() {
+        if (!this._appAwareness) {
+            // Created on first use, not in the constructor: every y-protocols
+            // Awareness starts its own setInterval that only destroy() clears, so
+            // an eagerly-built second instance cost one live timer per provider
+            // for the majority of consumers that never touch this channel
+            // (bench-reload-phantoms part 2 counts exactly that).
+            this._appAwareness = new awarenessProtocol.Awareness(this.doc);
+            this._attachAppAwareness(this._appAwareness);
+        }
+        return this._appAwareness;
+    }
     /**
      * Create a new generic provider.
      *
@@ -219,6 +487,10 @@ export class GenericProvider extends Observable {
         this._status = { state: 'disconnected' };
         this._synced = false;
         this._destroying = false;
+        this._lastActivityTime = Date.now();
+        this._lastPeriodicTickTime = Date.now();
+        this._equalBeaconsHeard = 0;
+        this._beaconForced = false;
         // BroadcastChannel state for cross-tab sync
         this._bcChannel = '';
         this._bcConnected = false;
@@ -237,8 +509,15 @@ export class GenericProvider extends Observable {
         // own. See _requestResync().
         this._resyncAttemptCount = 0;
         this._lastResyncAttemptTime = 0;
-        // Rate limiting for sync requests
+        // Rate limiting for sync traffic - two budgets since phase 1b: one for
+        // what we ask for (beacons, pushes, syncNow), one for what we owe
+        // (SyncStep2 replies, acks). With a single shared budget a join burst's
+        // replies spent the slots a peer needed for its own recovery (measured:
+        // joiners arriving within 10 s of a burst converged in 13 s once, and a
+        // 50-peer room's periodic beacons ran at a third of their rate for 10 s
+        // after every join burst). Same size, same window, independent.
         this._syncRequestTimes = [];
+        this._syncReplyTimes = [];
         // 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
@@ -248,6 +527,105 @@ export class GenericProvider extends Observable {
         // peers there's no "someone else" to rely on, so replies go out
         // immediately as before.
         this._pendingSyncReply = null;
+        // Whether _pendingSyncReply is a digest-beacon ack (see _handleDigest())
+        // rather than a SyncStep2. An overheard beacon with a digest equal to ours
+        // makes a pending ACK redundant (the joiner it was for has received that
+        // same beacon and is synced by it) but says nothing about a pending
+        // SyncStep2, which carries data - so only acks are cancelled on that
+        // signal. Measured in test/dummy/bench-user-scaling.ts: without this, acks
+        // were ~94% of a 50-peer join burst's messages (Task 3b in the design doc).
+        this._pendingSyncReplyIsAck = false;
+        // The requester's state vector the pending SyncStep2 answers (null for
+        // acks and for replies to plain SyncStep1s). A later request with the
+        // same state vector is the same question: the pending reply's bytes are
+        // refreshed to the current document and the timer kept, instead of the
+        // old reply being flushed as "a different request" - measured: with one
+        // peer typing while K empty peers join, keystroke and JOIN-beacon
+        // arrivals interleave at random, every keystroke in between changed the
+        // reply bytes, and each settled peer flushed up to K replies (Task 7 in
+        // the phase-1b design doc).
+        this._pendingSyncReplyTargetSv = null;
+        this._responseWaitAttempts = 0;
+        this._responseSeen = false;
+        // Phase 1e: an equal ack or equal beacon from an UNSETTLED peer arrived
+        // during the current response wait. Not a response (a settled peer with
+        // content may still answer), but evidence that the room is a fresh one
+        // whose peers are all in our state - see _armResponseWait().
+        this._equalUnsettledSeen = false;
+        this._responseWaitFlags = 0; // flags for the retry beacon: CONFIRM after a JOIN, 0 after a resync
+        this._behindSv = null;
+        // Round-trip estimate from our own requests (JOIN/resync beacon -> first
+        // SyncStep2 or ack): the minimum of the last 8 samples, because a sample
+        // includes the responder's random suppression delay and the fastest
+        // reply had the least of it. Drives _replySuppressionMaxDelay() (a
+        // suppression window shorter than the one-way latency suppresses
+        // nothing: at 250-350 ms latency every peer ahead of a requester replied
+        // before any reply could be overheard, ~20 replies per request at N=100)
+        // and _armResponseWait()'s first delay (a fixed 1 s fired premature
+        // retries on the Matrix profile). See the phase-1b design doc, 1c.
+        this._rttSamples = [];
+        this._requestSentAt = 0;
+        // See DIGEST_FLAG_SETTLED. Reset on connect (a re-joining peer is a joiner).
+        this._confirmed = false;
+        // ClientIDs we have heard from (beacon and verified-update senders, and
+        // the ids a relayed presence table names), each with the time we last
+        // heard it. The reply-suppression gate ("is there someone else who could
+        // answer?") used awareness alone, and in a join burst the awareness
+        // messages trail the beacons - so the gate was still closed exactly when
+        // 49 requests arrived at once, and every one got an immediate reply
+        // (phase-1b design, item 3). Cleared on disconnect, pruned by the lease
+        // sweep after a lease of silence (round 7, item 3: on a relay transport a
+        // reload's old clientID never says goodbye, and every one of them counted
+        // in _peerCount() forever).
+        this._knownPeers = new Map();
+        // Transport address (the `from` of Transport.onMessage) per remote
+        // clientID, learned from beacons and verified updates; lets replies, acks
+        // and presence responses go to the requester alone when the transport
+        // has sendTo (phase-1c design, item B). Cleared on disconnect, pruned
+        // with _knownPeers.
+        this._peerAddress = new Map();
+        // Requesters whose JOIN presence request the pending presence-response
+        // timer covers (see _schedulePresenceResponse).
+        this._presencePending = new Set();
+        // Phase 1e: an awareness message carrying OUR state at our current clock
+        // arrived since the presence-response timer was armed - the room's
+        // relayer (see _schedulePresenceResponse) already told the joiner about
+        // us, our own response would repeat it.
+        this._presenceCovered = false;
+        // Same NACK-style suppression as _pendingSyncReply above, applied to
+        // awareness updates that are a pure timeout-triggered removal (see
+        // _scheduleAwarenessRemoval()) - every OTHER connected peer runs its own
+        // independent 30s outdatedTimeout sweep (y-protocols/awareness.js), so
+        // one peer going silent (crash/dirty drop, not a clean disconnect())
+        // causes an O(N) simultaneous "peer X is gone" broadcast burst without
+        // this. See docs/superpowers/specs/2026-09-04-sync-optimization-round-3-ideas.md
+        // item 7 and test/dummy/bench-awareness-removal-burst.ts.
+        this._pendingAwarenessRemoval = null;
+        // Peers whose channel opened since the debounce timer was armed (the
+        // `from`/sendTo address). Round 5, item 5: on a transport with sendTo
+        // each gets a plain beacon addressed to it - it answers with what we
+        // lack and its own JOIN beacon makes us answer it - instead of the
+        // full-state push + beacon broadcast to EVERY connection that
+        // `_syncNow(0)` did (the O(N^2) mesh-join burst CLAUDE.md warns about).
+        this._pendingPeerConnectIds = new Set();
+        // Round 5, item 5: the state vector and delete-set hash the room last
+        // confirmed as equal to ours (an equal digest from any peer, see
+        // _handleDigest). A reconnect's push then carries only what we produced
+        // since - our offline edits, the one thing a push exists for (round 2:
+        // one message must survive alone) - instead of the whole document, which
+        // on a chunking transport was several messages per reconnect. A room
+        // that has been replaced meanwhile shows up behind in its own JOIN
+        // beacons and is answered like any late joiner. Null until the first
+        // confirmation: the first connect still pushes everything.
+        this._confirmedSv = null;
+        this._confirmedDsHash = null;
+        // connect() is awaiting ConnectionConfig.waitFor: doc updates are the local load.
+        this._loading = false;
+        // Cached computeDeleteSetHash(doc); null = stale. Invalidated on every
+        // doc 'update' (deletes are content changes, so this is exact; inserts
+        // invalidate needlessly but cheaply). See computeDeleteSetHash's doc for
+        // why this must not be recomputed per update.
+        this._dsHashCache = null;
         // Sequence numbers for causal ordering
         this._localSeqNum = 0; // Our sequence number counter
         // Per-sender sequence tracking for reordering-tolerant gap detection.
@@ -257,41 +635,77 @@ export class GenericProvider extends Observable {
         // 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)
+        // Update batching/debouncing. `_batchUpdates` 0 (the default) no longer
+        // means a synchronous send from inside the Y.Doc 'update' event: the
+        // update is merged into `_pendingUpdate` and flushed at the end of the
+        // current task (a microtask - no timer, no measurable delay), so that
+        // (a) the several transactions an editor binding can emit for one input
+        // event leave as one message and (b) the cursor awareness the binding
+        // sets right after the text change in the same task rides along in the
+        // same wire message (round 5, item 1 - see _flushPendingUpdate()).
+        // `_batchTimeoutId` is set only for the timed (`batchUpdates > 0`) flush;
+        // `_flushScheduled` covers both.
+        this._batchUpdates = 0; // milliseconds delay (0 = end of current task)
+        // Origins whose updates are never sent to the transport (local-only txns).
+        this._excludeOrigins = new Set();
+        // 'pull' never pushes local state unasked - for read-mostly replicas that
+        // must not write into the room.
+        this._syncMode = 'push-pull';
         this._pendingUpdate = null;
+        this._flushScheduled = false;
         // Awareness throttling - prevents awareness from flooding document sync
-        this._awarenessInterval = 100; // ms between awareness broadcasts
+        this._awarenessInterval = 100; // ms between awareness broadcasts, or 'auto' (round 6, item 9)
         this._pendingAwarenessClients = new Set();
         this._lastAwarenessTime = 0;
-        // Independent throttle state for the app awareness channel.
+        // Independent throttle state for the app awareness channel, so module
+        // cursor churn never delays or coalesces with core presence.
         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
-        // then requests remote; 'pull' only requests remote state.
-        this._syncMode = 'push-pull';
         this.doc = doc;
         this.transport = transport;
         this.pubsub = new PubSubChannel(this);
         this.awareness = options.awareness || new awarenessProtocol.Awareness(doc);
-        this.appAwareness =
-            options.appAwareness || new awarenessProtocol.Awareness(doc);
+        // Only when supplied - otherwise the getter builds it on first use.
+        if (options.appAwareness) {
+            this._appAwareness = options.appAwareness;
+        }
         this._syncInterval = options.syncInterval ?? 5000;
         this._verifyUpdates = options.verifyUpdates ?? true;
         this._batchUpdates =
             options.batchUpdates ?? transport.preferredBatchMs ?? 0;
         this._disableBc = options.disableBc ?? false;
-        this._awarenessInterval = options.awarenessInterval ?? 100;
+        this._awarenessInterval =
+            options.awarenessInterval ?? transport.preferredAwarenessMs ?? 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._peerConnectDebounceMs = options.peerConnectDebounceMs ?? 50;
         this._gapGraceMs = options.gapGraceMs ?? 300;
         this._seqWindowSize = options.seqWindowSize ?? 64;
+        // Explicit 0 disables even when the transport hints a floor.
+        this._compressionThresholdBytes =
+            (options.compressionThresholdBytes ?? transport.preferredCompressMinBytes) ||
+                undefined;
+        this._idleBackoffEnabled = options.idleBackoffEnabled ?? true;
+        this._idleBackoffMaxMs = options.idleBackoffMaxMs ?? 60000;
+        this._trickleK = options.trickleK ?? 1;
+        this._ownsAwareness = !options.awareness;
+        this._awarenessTimeoutMs =
+            options.awarenessTimeoutMs ??
+                (typeof transport.onPeerDisconnect === 'function' ? 300000 : 30000);
+        // Silence y-protocols' own 3 s sweep on an awareness we created - the
+        // lease sweep that replaces it is armed in connect() and cleared in
+        // disconnect() (round 7, item 2: it used to start here and outlive
+        // disconnect(), keeping a dropped provider alive through its timer).
+        if (this._ownsAwareness) {
+            const aw = this.awareness;
+            if (aw._checkInterval !== undefined)
+                clearInterval(aw._checkInterval);
+        }
+        this._currentSyncIntervalMs = this._syncInterval;
         this._setupDocumentSync();
         this._setupAwarenessSync();
     }
@@ -317,18 +731,21 @@ export class GenericProvider extends Observable {
         try {
             // Setup BroadcastChannel for cross-tab sync (if enabled and available)
             this._setupBroadcastChannel(config);
-            // Connect the transport
-            await this.transport.connect(config);
-            // Register for incoming messages
-            this._unsubscribeTransport = this.transport.onMessage((data) => {
-                this._handleIncomingMessage(data);
+            // Register for incoming messages and new-peer notifications BEFORE
+            // connecting the transport. Some transports (e.g. PeerJS for a
+            // joining, non-coordinator peer) establish and fully open their first
+            // connection *inside* transport.connect() itself — so a peer-connect
+            // notification or an immediate reply from the other side can arrive
+            // before that promise resolves. Registering after the await left
+            // exactly that window uncovered: whatever arrived during it was
+            // silently dropped since neither callback was wired up yet.
+            this._unsubscribeTransport = this.transport.onMessage((data, from) => {
+                this._handleIncomingMessage(data, from);
             });
-            // When a new WebRTC peer channel opens, immediately push our full state
-            // so peers that reconnected after offline edits receive our changes.
             if (this.transport.onPeerConnect) {
-                const unsubPeer = this.transport.onPeerConnect((_peerId) => {
+                const unsubPeer = this.transport.onPeerConnect((peerId) => {
                     if (!this._destroying)
-                        this.syncNow();
+                        this._schedulePeerConnectSync(peerId);
                 });
                 const originalUnsub = this._unsubscribeTransport;
                 this._unsubscribeTransport = () => {
@@ -336,31 +753,125 @@ export class GenericProvider extends Observable {
                     unsubPeer();
                 };
             }
+            if (this.transport.onPeerDisconnect) {
+                const unsubLeave = this.transport.onPeerDisconnect((peerId) => {
+                    if (!this._destroying)
+                        this._handlePeerLeave(peerId);
+                });
+                const originalUnsub = this._unsubscribeTransport;
+                this._unsubscribeTransport = () => {
+                    originalUnsub?.();
+                    unsubLeave();
+                };
+            }
+            // Connect the transport
+            await this.transport.connect(config);
             this._setStatus({ state: 'connected' });
-            // Send initial sync. In 'push-pull' mode, syncNow() pushes our local
-            // state (so offline edits reach currently-connected peers) then requests
-            // remote state. In 'pull' mode, only request remote state — a relay/server
-            // holds authoritative state and we adopt it rather than pushing a
-            // competing local copy on every (re)connect.
-            if (this._syncMode === 'pull') {
-                this._sendSyncStep1();
+            this._startAwarenessSweep();
+            // Seed the round-trip estimate from the transport's hint (see
+            // Transport.expectedRttMs); the minimum-of-8 rule lets real samples
+            // take over as soon as they arrive.
+            if (this.transport.expectedRttMs) {
+                this._rttSamples = [this.transport.expectedRttMs];
             }
-            else {
-                this.syncNow();
+            // Persistence first (round 5, item 7): let a local copy load before
+            // the first beacon says what we have. While it loads, the doc
+            // updates it produces are the load, not edits - they are not
+            // broadcast - and afterwards the loaded state counts as confirmed:
+            // no full-state push. The beacon reconciles: peers behind us (our
+            // offline edits) see themselves behind and ask; peers ahead of us
+            // answer. Measured in bench-reconnect-push part 2: a 50 KB copy that
+            // loaded 100 ms after connect cost the room a 50 KB SyncStep2, and a
+            // first version of waitFor that only delayed the beacon cost 100 KB
+            // (the load's broadcast plus the push). See ConnectionConfig.waitFor.
+            if (config.waitFor) {
+                this._loading = true;
+                try {
+                    await config.waitFor;
+                }
+                catch {
+                    // The local load failed; sync from the room as if there were none.
+                }
+                this._loading = false;
+                if (this._destroying || !this.transport.isConnected)
+                    return;
+                this._confirmedSv = Y.encodeStateVector(this.doc);
+                this._confirmedDsHash = this._deleteSetHash();
             }
-            // Broadcast local awareness state
-            this._broadcastAwareness([this.doc.clientID]);
-            this._broadcastAwareness([this.doc.clientID], AWARENESS_CHANNEL_APP);
+            // Send initial sync pushing our local state plus requesting remote state.
+            // syncNow() is used instead of _sendSyncStep1() so that any offline edits
+            // made before this connect() call are pushed to currently-connected peers
+            // (e.g. same-browser tabs via BroadcastChannel).
+            this.syncNow();
+            // (Local awareness goes out inside syncNow()'s batch, or via its
+            // throttled fallback - a second broadcast here was a duplicate 100ms
+            // later.)
+            // The app channel has no equivalent path into syncNow()'s batch, so
+            // its local state is announced here when there is any.
+            this._announceAppAwareness();
             // 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.
+            //
+            // Uses a recursive setTimeout (re-jittered by ~20% each tick) rather
+            // than a plain setInterval so peers that connect() within a short
+            // window of each other - the common case: everyone joining a room
+            // near session start, or reconnecting together after a shared network
+            // blip - don't end up with near-synchronized periodic timers that all
+            // fire in the same few milliseconds every syncInterval. This doesn't
+            // reduce total periodic-sync traffic, only spreads it out so a room's
+            // background traffic is smooth instead of bursty.
             if (this._syncInterval > 0) {
-                this._syncIntervalId = setInterval(() => {
-                    if (this.transport.isConnected && !this._destroying) {
-                        this._sendSyncStep1();
-                    }
-                }, this._syncInterval);
+                // Reset idle-backoff state on every (re)connect so a reconnect
+                // always starts its first tick at the base interval, never
+                // inheriting a backed-off value left over from a previous session.
+                this._currentSyncIntervalMs = this._syncInterval;
+                this._lastActivityTime = Date.now();
+                this._lastPeriodicTickTime = Date.now();
+                this._equalBeaconsHeard = 0;
+                this._beaconForced = false;
+                const scheduleNextPeriodicSync = (delayMs) => {
+                    this._syncIntervalId = setTimeout(() => {
+                        const tickTime = Date.now();
+                        if (this._idleBackoffEnabled) {
+                            // "Activity" = anything _markActivity() call sites observed
+                            // (incoming message via transport or BroadcastChannel, local
+                            // or remote doc change, local or remote awareness change)
+                            // since the LAST tick fired - not since backoff started, so a
+                            // single quiet tick after a burst of activity still resets to
+                            // base rather than needing a full quiet cycle to catch up.
+                            const hadActivitySinceLastTick = this._lastActivityTime > this._lastPeriodicTickTime;
+                            this._currentSyncIntervalMs = hadActivitySinceLastTick
+                                ? this._syncInterval
+                                : Math.min(this._idleBackoffMaxMs, this._currentSyncIntervalMs * 2);
+                        }
+                        this._lastPeriodicTickTime = tickTime;
+                        if (this.transport.isConnected && !this._destroying) {
+                            // Beacon only. Presence is no longer re-announced per tick on
+                            // any transport: a joiner requests it via DIGEST_FLAG_JOIN
+                            // (see _handleDigest()), and y-protocols/awareness renews the
+                            // local state itself every outdatedTimeout/2 = 15s
+                            // (awareness.js _checkInterval), which the awareness update
+                            // handler broadcasts. Measured in
+                            // test/dummy/bench-idle-room.ts: the per-tick re-announce was
+                            // ~40% of an idle room's deliveries.
+                            // Trickle (round 5, item 3): silent if enough equal beacons
+                            // were overheard since the last tick - see _equalBeaconsHeard.
+                            const suppressed = this._trickleK > 0 &&
+                                !this._beaconForced &&
+                                this._equalBeaconsHeard >= this._trickleK;
+                            if (!suppressed)
+                                this._sendSyncStep1();
+                            this._equalBeaconsHeard = 0;
+                            this._beaconForced = false;
+                        }
+                        if (!this._destroying)
+                            scheduleNextPeriodicSync();
+                    }, delayMs ?? this._jitteredSyncInterval());
+                };
+                this._periodicScheduler = scheduleNextPeriodicSync;
+                scheduleNextPeriodicSync();
             }
         }
         catch (error) {
@@ -378,9 +889,15 @@ export class GenericProvider extends Observable {
     disconnect() {
         // Stop periodic sync
         if (this._syncIntervalId !== undefined) {
-            clearInterval(this._syncIntervalId);
+            clearTimeout(this._syncIntervalId);
             this._syncIntervalId = undefined;
         }
+        this._periodicScheduler = undefined;
+        // Stop the awareness lease sweep (armed in connect(); round 7, item 2).
+        if (this._awarenessSweepId !== undefined) {
+            clearTimeout(this._awarenessSweepId);
+            this._awarenessSweepId = undefined;
+        }
         // Reset resync escalation tracking
         this._resyncAttemptCount = 0;
         this._lastResyncAttemptTime = 0;
@@ -391,26 +908,73 @@ export class GenericProvider extends Observable {
             clearTimeout(this._pendingResyncTimeoutId);
             this._pendingResyncTimeoutId = undefined;
         }
+        // Stop any pending gap-check timers and forget per-sender sequence
+        // tracking. Without this, a gap-check timer armed before this
+        // disconnect() keeps running in the background and can fire
+        // _requestResync() after reconnect using sequence-number bookkeeping
+        // from the PREVIOUS connection - a spurious resync race disconnected
+        // from anything actually missing in the new session. Mirrors the
+        // equivalent cleanup in destroy().
+        for (const timer of this._gapCheckTimers.values()) {
+            clearTimeout(timer);
+        }
+        this._gapCheckTimers.clear();
+        this._remoteSeqInfo.clear();
+        // Cancel any pending debounced onPeerConnect sync - a reconnect gets a
+        // fresh burst of onPeerConnect events (if the transport supports it) and
+        // shouldn't fire a stale one left over from before this disconnect.
+        if (this._pendingPeerConnectSyncTimeoutId !== undefined) {
+            clearTimeout(this._pendingPeerConnectSyncTimeoutId);
+            this._pendingPeerConnectSyncTimeoutId = undefined;
+        }
+        this._pendingPeerConnectIds.clear();
         // 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 = [];
+        this._syncReplyTimes = [];
+        this._knownPeers.clear();
+        this._peerAddress.clear();
+        this._presencePending.clear();
+        // A pending response-wait belongs to a request on the old connection.
+        if (this._responseWaitTimer !== undefined) {
+            clearTimeout(this._responseWaitTimer);
+            this._responseWaitTimer = undefined;
+        }
+        if (this._pendingCheckTimer !== undefined) {
+            clearTimeout(this._pendingCheckTimer);
+            this._pendingCheckTimer = undefined;
+        }
+        if (this._behindCheckTimer !== undefined) {
+            clearTimeout(this._behindCheckTimer);
+            this._behindCheckTimer = undefined;
+        }
+        this._behindSv = null;
+        if (this._presenceResponseTimer !== undefined) {
+            clearTimeout(this._presenceResponseTimer);
+            this._presenceResponseTimer = undefined;
+        }
+        this._responseWaitAttempts = 0;
+        this._responseSeen = false;
+        this._equalUnsettledSeen = false;
+        this._rttSamples = [];
+        this._requestSentAt = 0;
+        this._confirmed = false;
         // 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);
-            this._batchTimeoutId = undefined;
-            // Send pending update if transport is still connected
-            if (this._pendingUpdate && this.transport.isConnected) {
-                this._sendUpdate(this._pendingUpdate);
-            }
-            this._pendingUpdate = null;
-        }
+        // Same reasoning for a pending suppressed awareness-removal broadcast -
+        // every other surviving peer is independently running the same
+        // suppression for the same departure, so dropping ours on disconnect
+        // (rather than flushing it through a transport that's about to go
+        // down) is safe.
+        this._cancelPendingAwarenessRemoval();
+        // Flush any pending batched updates before disconnecting (sent only if
+        // the transport is still connected; dropped otherwise, as before)
+        this._flushPendingUpdate();
         // Flush pending awareness updates before disconnecting
         if (this._awarenessTimeoutId !== undefined) {
             clearTimeout(this._awarenessTimeoutId);
@@ -423,13 +987,13 @@ export class GenericProvider extends Observable {
             }
         }
         this._pendingAwarenessClients.clear();
-        // Flush pending app awareness updates before disconnecting
+        // Same for the app channel's independent throttle.
         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._sendAppAwarenessNow(Array.from(this._pendingAppAwarenessClients));
             }
         }
         this._pendingAppAwarenessClients.clear();
@@ -441,7 +1005,6 @@ 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' });
@@ -454,7 +1017,7 @@ export class GenericProvider extends Observable {
         this._destroying = true;
         // Stop periodic sync (disconnect() will also do this, but be explicit)
         if (this._syncIntervalId !== undefined) {
-            clearInterval(this._syncIntervalId);
+            clearTimeout(this._syncIntervalId);
             this._syncIntervalId = undefined;
         }
         // Stop any pending gap-check timers
@@ -465,38 +1028,42 @@ export class GenericProvider extends Observable {
         // Drop any pending suppressed sync reply (disconnect() will also do
         // this, but be explicit)
         this._cancelPendingSyncReply();
+        // Same, for a pending suppressed awareness-removal broadcast
+        // (disconnect() will also do this, but be explicit)
+        this._cancelPendingAwarenessRemoval();
         // Flush any pending batched updates before destroying
-        if (this._batchTimeoutId !== undefined) {
-            clearTimeout(this._batchTimeoutId);
-            this._batchTimeoutId = undefined;
-            // Send pending update if transport is still connected
-            if (this._pendingUpdate && this.transport.isConnected) {
-                this._sendUpdate(this._pendingUpdate);
-            }
-            this._pendingUpdate = null;
-        }
+        this._flushPendingUpdate();
         this.disconnect();
         // Remove document update listener
         if (this._updateHandler) {
             this.doc.off('update', this._updateHandler);
             this._updateHandler = undefined;
         }
-        // Remove awareness update listeners
+        // Remove awareness update listener
         if (this._awarenessUpdateHandler) {
-            this.awareness.off('update', this._awarenessUpdateHandler);
+            this.awareness.off(this._ownsAwareness ? 'change' : 'update', this._awarenessUpdateHandler);
             this._awarenessUpdateHandler = undefined;
         }
         if (this._appAwarenessUpdateHandler) {
-            this.appAwareness.off('update', this._appAwarenessUpdateHandler);
+            this._appAwareness?.off('update', this._appAwarenessUpdateHandler);
             this._appAwarenessUpdateHandler = undefined;
         }
+        if (this._appAwarenessTimeoutId !== undefined) {
+            clearTimeout(this._appAwarenessTimeoutId);
+            this._appAwarenessTimeoutId = undefined;
+        }
+        if (this._awarenessSweepId !== undefined) {
+            clearTimeout(this._awarenessSweepId);
+            this._awarenessSweepId = undefined;
+        }
         // Remove beforeunload handler
         if (this._beforeUnloadHandler && typeof window !== 'undefined') {
             window.removeEventListener('beforeunload', this._beforeUnloadHandler);
             this._beforeUnloadHandler = undefined;
         }
         this.awareness.destroy();
-        this.appAwareness.destroy();
+        // Only if one was ever built - never construct one just to destroy it.
+        this._appAwareness?.destroy();
         super.destroy();
     }
     /**
@@ -524,14 +1091,29 @@ export class GenericProvider extends Observable {
         return this._synced;
     }
     /**
-     * Force an immediate sync with remote peers.
-     * Useful after network interruptions or to manually trigger re-sync.
+     * Push local state + request remote state, gated by the shared rate
+     * limiter. Returns whether it actually reserved a slot and sent anything
+     * - `false` means the caller was rate-limited right now. Extracted out of
+     * `syncNow()` so `_requestResync()`'s scheduled retry (see below) can tell
+     * the difference between "sent" and "silently skipped" and react to it,
+     * instead of assuming a resync always succeeds once it fires.
+     *
+     * @param push - whether to also broadcast full local document state.
+     * `_requestResync()`'s retry passes `false` (pull-only is enough for a
+     * resync trigger - see its call site).
+     * @param buildExtra - optional callback, invoked ONLY once a rate-limit
+     * slot is actually reserved (so it never runs, and never mutates
+     * whatever state it touches, on a call that ends up rate-limited),
+     * returning additional already-encoded sub-messages to fold into the SAME
+     * batched wire send as the push/pull messages below - e.g. an awareness
+     * update that's ready to go out "now" anyway (see
+     * `_tryImmediateAwarenessMessage()`). Pure wire-framing: whether a caller
+     * passes this never changes whether/when the push+pull half itself sends,
+     * only how many separate `transport.send()`/`bc.publish()` calls it costs.
+     * @param flags - digest beacon flags: DIGEST_FLAG_JOIN from syncNow(), 0
+     * from the peer-connect debounce and the resync retry (see _syncNow()).
      */
-    syncNow() {
-        if (!this.transport.isConnected) {
-            console.warn('Cannot sync: transport not connected');
-            return;
-        }
+    _trySyncPushPull(push = true, buildExtra, flags = 0) {
         // 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
@@ -543,21 +1125,315 @@ export class GenericProvider extends Observable {
         // 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);
+        if (!this._tryReserveSyncSlot())
+            return false;
+        const messages = [];
+        // Send our document state to all peers - everything on the first
+        // connect, only what we produced since the room last confirmed our
+        // state afterwards (see _confirmedSv). This is what carries edits made
+        // while offline.
+        if (push) {
+            const update = Y.encodeStateAsUpdate(this.doc, this._confirmedSv ?? undefined);
+            let worthPushing = update.length > 0;
+            if (worthPushing && this._confirmedSv !== null) {
+                // A diff against a confirmed state is never byte-empty (it always
+                // carries the delete set): push it only if it holds structs, or
+                // deletes the room has not confirmed.
+                try {
+                    worthPushing =
+                        Y.parseUpdateMeta(update).to.size > 0 ||
+                            this._deleteSetHash() !== this._confirmedDsHash;
+                }
+                catch {
+                    worthPushing = true;
+                }
+            }
+            if (worthPushing) {
+                messages.push(this._encodePush(update));
+            }
+        }
+        // Send sync request to get updates from others
+        messages.push(this._encodeSyncStep1(flags));
+        if (buildExtra) {
+            messages.push(...buildExtra());
+        }
+        // Batched into one wire message (MESSAGE_BATCH) instead of one
+        // transport.send()/bc.publish() call per sub-message - see _sendBatch().
+        this._sendBatch(messages);
+        if (flags & DIGEST_FLAG_JOIN)
+            this._armResponseWait(DIGEST_FLAG_CONFIRM);
+        return true;
+    }
+    /**
+     * Force an immediate sync with remote peers.
+     * Useful after network interruptions or to manually trigger re-sync.
+     * Sends the beacon with DIGEST_FLAG_JOIN: peers answer with their
+     * presence and, if our state already matches theirs, with an ack beacon
+     * so `synced` flips without a data round trip.
+     */
+    syncNow() {
+        this._syncNow(DIGEST_FLAG_JOIN);
+    }
+    /**
+     * syncNow() body. `flags` = 0 for callers that must NOT request presence:
+     * `_schedulePeerConnectSync()` (mesh transports already re-broadcast
+     * presence to a newcomer via their own onPeerConnect -> syncNow()).
+     */
+    _syncNow(flags) {
+        if (!this.transport.isConnected) {
+            console.warn('Cannot sync: transport not connected');
+            return;
+        }
+        if (flags & DIGEST_FLAG_JOIN)
+            this._confirmed = false; // a (re-)joiner until answered
+        // Try to fold the awareness broadcast into the same wire send as the
+        // sync push+pull below. _tryImmediateAwarenessMessage() only returns
+        // non-null (and only mutates awareness-throttle state) when the
+        // throttle would have let an immediate send through anyway - so this
+        // never changes awareness throttle semantics, only whether it travels
+        // as its own message or bundled with the sync message going out "now"
+        // too. Built inside buildExtra so it's only even attempted once a sync
+        // rate-limit slot is confirmed reserved (see _trySyncPushPull's doc).
+        let awarenessBatched = false;
+        const sent = this._trySyncPushPull(
+        // 'pull': send the beacon but never the full-state push. A relay or
+        // server holds authoritative state and we adopt it rather than pushing
+        // a competing local copy on every (re)connect. Peers that are actually
+        // behind us still ask, and we still answer - the beacon reconciles.
+        this._syncMode !== 'pull', () => {
+            const msg = this._tryImmediateAwarenessMessage([this.doc.clientID]);
+            if (msg) {
+                awarenessBatched = true;
+                return [msg];
+            }
+            return [];
+        }, flags);
+        // Awareness broadcasting is independently throttled and explicitly NOT
+        // gated by the sync rate limiter above - preserve that exactly: it
+        // always ends up broadcast one way or another (batched above, or via
+        // its own throttled path here), regardless of whether the sync half
+        // above was rate-limited.
+        if (!sent || !awarenessBatched) {
+            this._broadcastAwareness([this.doc.clientID]);
+        }
+    }
+    /**
+     * Compute the next periodic-sync delay, jittered by ~+/-20% around
+     * `_currentSyncIntervalMs` (== `_syncInterval` unless `idleBackoffEnabled`
+     * has backed it off - see that option's doc comment). Re-jittered fresh
+     * each tick (not computed once per connect()) so a room's peers - which
+     * commonly all connect() within a short window of each other - drift
+     * apart over time instead of staying loosely synchronized. Extracted to
+     * its own method purely so benchmarks can shadow it to compare against
+     * the unjittered baseline.
+     */
+    _jitteredSyncInterval() {
+        const jitter = 1 + (Math.random() * 2 - 1) * 0.2; // +/-20%
+        return this._currentSyncIntervalMs * jitter;
+    }
+    /**
+     * Record that "activity" happened right now, for `idleBackoffEnabled`'s
+     * benefit. Cheap (one timestamp write) and called unconditionally
+     * regardless of whether idle backoff is enabled, so there's no behavioral
+     * branch to keep in sync - the backoff decision in connect()'s periodic
+     * tick is the only place that actually reads this.
+     *
+     * Call sites are deliberately NOT "any inbound wire message" - an earlier
+     * version of this hooked `_handleIncomingMessage()` unconditionally, which
+     * made the periodic tick's OWN SyncStep1 request and the SyncStep2 reply
+     * answering it (empty payload - nothing to sync) each count as "activity",
+     * permanently resetting the backoff on every single tick and making the
+     * whole feature a no-op (caught by this bench script's own first run: ON
+     * and OFF produced statistically indistinguishable message counts). Both
+     * Yjs's `doc.emit('update', ...)` and y-protocols' `awareness.emit('update', ...)`
+     * already only fire when something with actual content changed
+     * (`hasContent`/non-empty added+updated+removed - confirmed by reading
+     * yjs's `Transaction.js` and y-protocols' `awareness.js` directly), so
+     * hooking THOSE instead is exactly "local or remote document/awareness
+     * change" with no extra filtering needed - a no-op SyncStep2 reply, a
+     * digest beacon, or a duplicate/no-change awareness re-announce (e.g. a
+     * JOIN-triggered presence response that changed nothing) never reaches
+     * these handlers. A corrupted (CRC32
+     * mismatch) message is real evidence of wire activity that neither
+     * handler would ever see (it's rejected before decoding) - see the
+     * explicit call in `_processWrappedMessage()`'s corruption branch.
+     *
+     * Call sites: `_setupDocumentSync()`'s update handler (LOCAL document
+     * edits only, since phase 1e) and `_processWrappedMessage()`'s
+     * corrupted-message branch (wire noise, not silence).
+     */
+    _markActivity() {
+        this._lastActivityTime = Date.now();
+        // A backed-off periodic timer is re-armed at the base interval NOW, not
+        // at its next tick (which may be idleBackoffMaxMs away). Design D of
+        // the phase-1d doc rests on it: the peer that just had activity beacons
+        // within one base interval, and a peer that lost that activity's
+        // message learns from that beacon that it is behind (design A) - its
+        // own backed-off interval no longer bounds the recovery. Measured in
+        // test/dummy/bench-idle-backoff.ts. At most once per idle stretch.
+        if (this._idleBackoffEnabled &&
+            this._currentSyncIntervalMs !== this._syncInterval &&
+            this._syncIntervalId !== undefined &&
+            this._periodicScheduler !== undefined) {
+            clearTimeout(this._syncIntervalId);
+            this._currentSyncIntervalMs = this._syncInterval;
+            // Phase 1e: at a random point inside the base interval, not a full
+            // one (the phase-1d note): the loser's recovery chain starts with
+            // this beacon - so Trickle never suppresses it (round 5, item 3).
+            this._beaconForced = true;
+            this._periodicScheduler(Math.random() * this._syncInterval);
+        }
+    }
+    /**
+     * The lease sweep: y-protocols' own (awareness.js `_checkInterval`: renew
+     * at outdatedTimeout/2, remove at outdatedTimeout, every
+     * outdatedTimeout/10) replaced by the same loop at `_awarenessTimeoutMs`,
+     * the period jittered so a room that joined together does not renew in
+     * one burst (measured: all 49 listeners of a 50-peer room renewed inside
+     * the same 10 s window). The renew/remove half runs only on an awareness
+     * we created (`_ownsAwareness`); the peer-table prune (round 7, item 3)
+     * runs regardless. Armed by connect(), cleared by disconnect() - round 7,
+     * item 2: started from the constructor it outlived disconnect(), ticking
+     * ~20 times a minute and keeping the dropped provider reachable
+     * (test/dummy/bench-reload-phantoms.ts, part 2).
+     */
+    _startAwarenessSweep() {
+        if (this._awarenessSweepId !== undefined)
+            return;
+        const lease = this._awarenessTimeoutMs;
+        const arm = () => {
+            this._awarenessSweepId = setTimeout(tick, (lease / 10) * (0.8 + Math.random() * 0.4));
+        };
+        const tick = () => {
+            const now = Date.now();
+            if (this._ownsAwareness) {
+                const mine = this.awareness.meta.get(this.doc.clientID);
+                if (this.awareness.getLocalState() !== null &&
+                    mine !== undefined &&
+                    lease / 2 <= now - mine.lastUpdated) {
+                    this.awareness.setLocalState(this.awareness.getLocalState()); // renew: bumps the clock
+                    this._broadcastAwareness([this.doc.clientID]); // 'change' does not fire for an equal state (item 8)
+                }
+                const remove = [];
+                this.awareness.meta.forEach((meta, clientID) => {
+                    if (clientID !== this.doc.clientID &&
+                        lease <= now - meta.lastUpdated &&
+                        this.awareness.getStates().has(clientID)) {
+                        remove.push(clientID);
+                    }
+                });
+                if (remove.length > 0) {
+                    awarenessProtocol.removeAwarenessStates(this.awareness, remove, 'timeout');
+                }
+            }
+            // Round 7, item 3: forget peers not heard from for a lease. A live
+            // peer is heard at least every lease/2 through its presence renewal
+            // (every receiver scans it, above at MESSAGE_AWARENESS); a pruned
+            // peer that speaks again is simply learned again. Measured before
+            // this: 30 reloads in a 20-peer relay room left 49 known peers for
+            // good, and every cursor moved at a 50-peer room's 'auto' interval
+            // (test/dummy/bench-reload-phantoms.ts). awareness.meta is left to
+            // y-protocols: it holds the clock a late message is checked against.
+            for (const [id, heardAt] of this._knownPeers) {
+                if (lease <= now - heardAt) {
+                    this._knownPeers.delete(id);
+                    this._peerAddress.delete(id);
+                    this._remoteSeqInfo.delete(id);
+                }
             }
-            // Send sync request to get updates from others
-            this._writeSyncStep1();
+            // Re-arm only while connected: a disconnect() from inside a listener
+            // above has just cleared the timer, and must stay cleared.
+            if (!this._destroying && this._status.state === 'connected')
+                arm();
+            else
+                this._awarenessSweepId = undefined;
+        };
+        arm();
+    }
+    /**
+     * A digest, verified update or ack from `clientID` (or one we are about
+     * to send, for our own id) is proof of presence: refresh the lease the
+     * sweep above checks. Only for ids with a state - a departed peer's
+     * `meta` entry survives its removal (y-protocols keeps it for the clock)
+     * and must not be revived by a late message.
+     */
+    _touchPeer(clientID) {
+        const meta = this.awareness.meta.get(clientID);
+        if (meta !== undefined && this.awareness.getStates().has(clientID)) {
+            meta.lastUpdated = Date.now();
+        }
+    }
+    /**
+     * Transport.onPeerDisconnect: the peer at `peerId` is gone. Forget its
+     * address and id, drop its awareness state with origin 'peer-left': the
+     * broadcast goes through the same suppression as a timeout removal, but
+     * with a long window - every peer gets the leave signal in the same
+     * millisecond, and at the reply window (~170 ms at N=50) 28 of 49
+     * survivors broadcast before the first broadcast could be overheard
+     * (bench-awareness-removal-burst, DUMMY_PEER_EVENTS=1). One broadcast
+     * room-wide is still worth having: it corrects a joiner that received
+     * this peer in a relayed presence table but had no channel to it yet.
+     */
+    _handlePeerLeave(peerId) {
+        const gone = [];
+        for (const [clientID, address] of this._peerAddress) {
+            if (address === peerId)
+                gone.push(clientID);
+        }
+        for (const id of gone) {
+            this._peerAddress.delete(id);
+            this._knownPeers.delete(id);
+        }
+        const present = gone.filter((id) => this.awareness.getStates().has(id));
+        if (present.length > 0) {
+            awarenessProtocol.removeAwarenessStates(this.awareness, present, 'peer-left');
+        }
+    }
+    /** Cached delete-set hash - see computeDeleteSetHash(). */
+    _deleteSetHash() {
+        if (this._dsHashCache === null) {
+            this._dsHashCache = computeDeleteSetHash(this.doc);
         }
-        // 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);
+        return this._dsHashCache;
+    }
+    /**
+     * Debounce onPeerConnect-triggered syncNow() calls. A burst of connect
+     * events within `_peerConnectDebounceMs` collapses into one call instead
+     * of one per event - without this, N peers joining a mesh in a short
+     * window each independently broadcast full state to everyone already
+     * connected (O(N^2) traffic), since onPeerConnect fires once per
+     * newly-opened peer connection with no coalescing of its own.
+     */
+    _schedulePeerConnectSync(peerId) {
+        if (peerId !== undefined)
+            this._pendingPeerConnectIds.add(peerId);
+        if (this._pendingPeerConnectSyncTimeoutId !== undefined)
+            return;
+        this._pendingPeerConnectSyncTimeoutId = setTimeout(() => {
+            this._pendingPeerConnectSyncTimeoutId = undefined;
+            const ids = Array.from(this._pendingPeerConnectIds);
+            this._pendingPeerConnectIds.clear();
+            if (!this.transport.isConnected || this._destroying)
+                return;
+            // A newly-joined mesh peer has no app-channel state for us yet: the
+            // digest beacon below covers the document, and the core awareness
+            // channel re-broadcasts on its own, but nothing replays app awareness.
+            // Pre-merge this rode along as a second _broadcastAwareness() on the
+            // APP channel here; the digest-beacon rewrite dropped it, which is why
+            // cursors were invisible to whichever peer joined second.
+            this._announceAppAwareness();
+            if (typeof this.transport.sendTo === 'function' && ids.length > 0) {
+                // Round 5, item 5: one plain beacon to each new peer, nothing to
+                // the rest of the mesh. Not rate-limited as a request: it answers
+                // a channel that just opened, and the peer's own JOIN beacon is
+                // the fallback if it is lost.
+                const beacon = wrapMessageWithChecksum(this._encodeSyncStep1(0));
+                for (const id of ids)
+                    this._sendToTransport(beacon, id);
+                return;
+            }
+            this._syncNow(0);
+        }, this._peerConnectDebounceMs);
     }
     /**
      * Setup automatic document synchronization.
@@ -566,27 +1442,36 @@ export class GenericProvider extends Observable {
      */
     _setupDocumentSync() {
         this._updateHandler = (update, origin) => {
-            // Don't send updates that originated from this provider
-            // This prevents infinite loops when receiving updates
-            if (origin === this)
-                return;
-            // Don't send updates from excluded origins (local-only txns)
-            if (this._excludeOrigins.has(origin))
-                return;
-            if (this._batchUpdates > 0) {
-                // Batch mode: merge updates and debounce
+            this._dsHashCache = null;
+            this._equalBeaconsHeard = 0; // our digest changed - see the Trickle fields
+            // Fires for BOTH local edits and remotely-applied updates (the latter
+            // go through doc.transact with origin=this) - see _markActivity()'s
+            // doc comment. Phase 1e: only a LOCAL edit counts as activity for
+            // idle backoff - a listener has nothing a beacon would announce, and
+            // the typist's base-interval beacon heals any listener that lost the
+            // keystroke (phase 1d design A). Before this, one typist kept all N
+            // peers at the base cadence: N*(N-1) deliveries per interval against
+            // N-1 per keystroke.
+            // Don't send updates that originated from this provider (received
+            // from the wire) or from the local load connect() is waiting for.
+            if (origin !== this && !this._loading) {
+                // Local-only transaction origins (e.g. a rollback the peer must not
+                // replicate) never reach the transport.
+                if (this._excludeOrigins.has(origin))
+                    return;
+                this._markActivity();
+                // 'pull' replicas answer requests but never push unasked.
+                if (this._syncMode === 'pull')
+                    return;
                 this._batchUpdate(update);
             }
-            else {
-                // Immediate mode: send right away
-                this._sendUpdate(update);
-            }
         };
         this.doc.on('update', this._updateHandler);
     }
     /**
-     * Batch/debounce updates to reduce network traffic.
-     * Merges multiple updates and sends after delay.
+     * Merge a local update into the pending batch and schedule its flush:
+     * after `batchUpdates` ms (debounced) when that is > 0, otherwise at the
+     * end of the current task via queueMicrotask - see `_pendingUpdate`.
      */
     _batchUpdate(update) {
         // Merge with pending update if exists
@@ -606,18 +1491,38 @@ export class GenericProvider extends Observable {
         else {
             this._pendingUpdate = update;
         }
-        // Clear existing timeout
+        if (this._batchUpdates > 0) {
+            // Debounce: restart the timer on every update.
+            if (this._batchTimeoutId !== undefined)
+                clearTimeout(this._batchTimeoutId);
+            this._batchTimeoutId = setTimeout(() => {
+                this._batchTimeoutId = undefined;
+                this._flushPendingUpdate();
+            }, this._batchUpdates);
+            this._flushScheduled = true;
+            return;
+        }
+        if (this._flushScheduled)
+            return;
+        this._flushScheduled = true;
+        queueMicrotask(() => this._flushPendingUpdate());
+    }
+    /**
+     * Send the pending update batch as one wire message, carrying any
+     * awareness change the throttle is holding (see _takePendingAwareness).
+     * Shared by the microtask flush, the timed flush, and the
+     * disconnect()/destroy() flush.
+     */
+    _flushPendingUpdate() {
+        this._flushScheduled = false;
         if (this._batchTimeoutId !== undefined) {
             clearTimeout(this._batchTimeoutId);
-        }
-        // Set new timeout to send after delay
-        this._batchTimeoutId = setTimeout(() => {
-            if (this._pendingUpdate) {
-                this._sendUpdate(this._pendingUpdate);
-                this._pendingUpdate = null;
-            }
             this._batchTimeoutId = undefined;
-        }, this._batchUpdates);
+        }
+        const update = this._pendingUpdate;
+        this._pendingUpdate = null;
+        if (update && this.transport.isConnected)
+            this._sendUpdate(update);
     }
     /**
      * Setup automatic awareness synchronization.
@@ -625,21 +1530,85 @@ export class GenericProvider extends Observable {
      */
     _setupAwarenessSync() {
         this._awarenessUpdateHandler = ({ added, updated, removed, }, origin) => {
-            // Broadcast awareness changes (unless they came from remote)
+            // Not _markActivity(): awareness changes are not something a beacon
+            // announces (phase 1e; see the idleBackoffEnabled option).
+            // Broadcast awareness changes, UNLESS they came from remote (this
+            // comment described the intent since the very first commit, but the
+            // actual origin check was never implemented until now - confirmed by
+            // `git log -p` on this handler). `origin === this` is exactly the
+            // signature the MESSAGE_AWARENESS handler stamps on an update applied
+            // from an incoming wire message (`applyAwarenessUpdate(this.awareness,
+            // ..., this)`, below). Every transport this project targets is a
+            // full-room relay (websocket/pubnub/gun/matrix/ably/supabase) or a
+            // full mesh (peerjs/simple-peer/trystero - see CLAUDE.md), so the
+            // sender's own broadcast already reached every other peer directly;
+            // re-broadcasting it here on receipt is pure redundant traffic that
+            // compounds across every OTHER receiver doing the same thing.
+            // Verified with a throwaway probe: a single awareness field change in
+            // an N-peer room cost N*(N-1) wire deliveries before this check (one
+            // echo per receiver, each reaching N-1 peers) vs. N-1 after - e.g.
+            // N=20: 380 -> 19.
+            //
+            // Carve-out: `applyAwarenessUpdate` has its own defense against a
+            // remote peer incorrectly removing OUR OWN state (a stale/racy
+            // timeout-removal from someone else's clock) - it bumps our clock
+            // instead of deleting our state, but (a quirk of that function) still
+            // reports it via `removed` including our own clientID. That specific
+            // case must still be broadcast so the room's stale belief that we're
+            // gone gets corrected promptly, instead of only self-healing on our
+            // next unrelated state change/renewal (up to ~15s later).
+            if (origin === this) {
+                // An incoming removal might be the SAME departure we have a
+                // suppressed broadcast queued for (see _scheduleAwarenessRemoval())
+                // - someone else already told the room, drop ours.
+                if (removed.length > 0) {
+                    this._cancelPendingAwarenessRemovalIfOverlaps(removed);
+                }
+                if (!removed.includes(this.awareness.clientID)) {
+                    return;
+                }
+            }
             const changedClients = added.concat(updated).concat(removed);
+            // A pure timeout-triggered removal ('timeout' is the exact origin
+            // string y-protocols/awareness.js's own _checkInterval passes to
+            // removeAwarenessStates()) is redundant across the whole room: every
+            // OTHER connected peer runs the identical 30s-timeout sweep
+            // independently, so all of them detect and would broadcast the SAME
+            // departure within the same ~3s tick - measured as O(N-1) broadcasts
+            // / O((N-1)(N-2)) deliveries for ONE departure in
+            // test/dummy/bench-awareness-removal-burst.ts (e.g. N=50: 2352
+            // deliveries). Delay + drop-if-overheard, exactly mirroring
+            // _scheduleSyncReply()/_cancelPendingSyncReply()'s suppression of
+            // redundant SyncStep2 replies. Deliberately NOT applied to
+            // 'disconnect'/'window unload' removals (a single broadcaster, not
+            // redundant) or to added/updated clients (every sender's
+            // cursor/presence data is meaningfully different and must never be
+            // suppressed).
+            if (origin === 'timeout' || origin === 'peer-left') {
+                this._scheduleAwarenessRemoval(changedClients, origin);
+                return;
+            }
             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);
+        // Round 5, item 8: broadcast on 'change' (y-protocols filters updates
+        // whose state deep-equals the previous one) rather than 'update'
+        // (every setLocalState call) - an app that re-sets unchanged state no
+        // longer costs a broadcast per call. The renewal is the one
+        // equal-state update that must go out; the provider's own sweep sends
+        // it explicitly (_startAwarenessSweep). With an app-supplied Awareness
+        // y-protocols' own sweep renews through 'update' only, so that case
+        // keeps listening on 'update'.
+        this.awareness.on(this._ownsAwareness ? 'change' : 'update', this._awarenessUpdateHandler);
+        // App channel: attached here only if one was supplied to the
+        // constructor; otherwise the getter attaches on first use. Reading
+        // `this.appAwareness` here would defeat the lazy construction.
+        if (this._appAwareness) {
+            this._attachAppAwareness(this._appAwareness);
+        }
         // 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();
             };
@@ -647,17 +1616,81 @@ export class GenericProvider extends Observable {
         }
     }
     /**
-     * Handle incoming messages from the transport.
-     * Verifies message integrity with CRC32 before processing.
-     * Corrupt messages are rejected immediately without attempting to decode.
+     * Handle incoming messages from the transport (or BroadcastChannel).
+     *
+     * When compressionThresholdBytes is disabled (the default), this is a
+     * fully synchronous fast path, byte-for-byte the same behavior as before
+     * that option existed: straight into _processWrappedMessage().
+     *
+     * When enabled, every message - from the network transport AND from
+     * BroadcastChannel (see _send()) - carries a leading compressed(1)/
+     * uncompressed(0) flag byte ahead of the usual CRC32 wrapper. Reading
+     * that flag and, if set, decompressing is inherently async (the
+     * Compression Streams API has no synchronous form), so this method
+     * dispatches to a promise chain instead of processing inline in that
+     * case. This means a large (compressed) message and a small (uncompressed
+     * or below-threshold) message that arrive back-to-back can finish
+     * processing out of arrival order - acceptable here because Yjs updates
+     * are idempotent/commutative (see MESSAGE_SYNC_VERIFIED's handling below)
+     * and because the compression threshold keeps this path almost entirely
+     * to large, full-state syncs, not the per-keystroke incremental updates
+     * that per-sender gap detection actually relies on ordering-sensitive
+     * heuristics for.
+     */
+    _handleIncomingMessage(data, from) {
+        if (!this._compressionThresholdBytes) {
+            this._processWrappedMessage(data, from);
+            return;
+        }
+        if (data.length < 1) {
+            console.warn('[GenericProvider] Dropping empty message (missing compression flag byte)');
+            return;
+        }
+        const flag = data[0];
+        const rest = data.subarray(1);
+        if (flag === 0) {
+            this._processWrappedMessage(rest, from);
+            return;
+        }
+        if (!COMPRESSION_AVAILABLE) {
+            console.warn('[GenericProvider] Received a compressed message but this runtime has no DecompressionStream - dropping it.');
+            return;
+        }
+        decompressDeflateRaw(rest)
+            .then((wrapped) => this._processWrappedMessage(wrapped, from))
+            .catch((error) => {
+            // Treat decompression failure the same as a CRC32 mismatch on the
+            // uncompressed path: request a resync rather than silently dropping.
+            console.warn('[GenericProvider] Failed to decompress incoming message, treating as corrupted:', error);
+            this._requestResync();
+        });
+    }
+    /**
+     * Verify message integrity with CRC32 and decode. Corrupt messages are
+     * rejected immediately without attempting to decode. Operates on bytes
+     * that have already had any compression flag/decompression handled by
+     * _handleIncomingMessage() - this is the pre-compression-feature
+     * implementation, unchanged.
      */
-    _handleIncomingMessage(data) {
+    _processWrappedMessage(data, from) {
         // Verify message integrity with CRC32 checksum
         const message = unwrapAndVerifyMessage(data);
         if (message === null) {
-            // Message is corrupted - reject it immediately
+            // Message is corrupted - reject it immediately. Note: if this was a
+            // MESSAGE_BATCH envelope, the CRC32 wrap covers the WHOLE batch, so a
+            // single corrupted bit here loses every sub-message it contained, not
+            // just one - a deliberate tradeoff of batching multiple logical
+            // messages behind one wire message/one checksum. See _sendBatch()'s
+            // doc comment for why this was chosen over per-sub-message checksums,
+            // and this project's benchmark suite (bench-corruption-storm.ts,
+            // bench-packet-loss.ts) for how that tradeoff was measured.
             console.warn(`[GenericProvider] 💥 Corrupted message rejected: CRC32 checksum mismatch. ` +
                 `This is expected if data corruption simulation is enabled.`);
+            // Wire noise, not silence - counts as activity for idleBackoffEnabled
+            // even though nothing here reaches the doc/awareness update handlers
+            // (the message is rejected before decoding). See _markActivity()'s
+            // doc comment.
+            this._markActivity();
             // 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.
@@ -666,168 +1699,7 @@ export class GenericProvider extends Observable {
         }
         // Message integrity verified - safe to decode
         try {
-            const decoder = decoding.createDecoder(message);
-            const messageType = decoding.readVarUint(decoder);
-            switch (messageType) {
-                case MESSAGE_SYNC: {
-                    const encoder = encoding.createEncoder();
-                    encoding.writeVarUint(encoder, MESSAGE_SYNC);
-                    const syncMessageType = syncProtocol.readSyncMessage(decoder, encoder, this.doc, this);
-                    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. 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) {
-                        if (this.awareness.getStates().size >= 3) {
-                            this._scheduleSyncReply(encoding.toUint8Array(encoder));
-                        }
-                        else {
-                            this._sendSyncReply(encoding.toUint8Array(encoder));
-                        }
-                    }
-                    break;
-                }
-                case MESSAGE_AWARENESS: {
-                    const channel = decoding.readVarUint(decoder);
-                    awarenessProtocol.applyAwarenessUpdate(channel === AWARENESS_CHANNEL_APP
-                        ? this.appAwareness
-                        : this.awareness, decoding.readVarUint8Array(decoder), this);
-                    break;
-                }
-                case MESSAGE_PUBSUB: {
-                    // Read topic
-                    const topic = decoding.readVarString(decoder);
-                    // Read message payload
-                    const payloadBytes = decoding.readVarUint8Array(decoder);
-                    try {
-                        // Decode JSON payload
-                        const decoder = new TextDecoder();
-                        const payloadStr = decoder.decode(payloadBytes);
-                        const message = JSON.parse(payloadStr);
-                        // Emit to pubsub channel
-                        this.pubsub._handleMessage(topic, message);
-                    }
-                    catch (error) {
-                        console.error('Error decoding pub/sub message:', error);
-                    }
-                    break;
-                }
-                case MESSAGE_PUBSUB_TARGETED: {
-                    const target = decoding.readVarString(decoder);
-                    const topic = decoding.readVarString(decoder);
-                    const payloadBytes = decoding.readVarUint8Array(decoder);
-                    // Drop messages aimed at someone else (broadcast-and-filter path).
-                    if (this._localId !== undefined && target !== this._localId) {
-                        break;
-                    }
-                    try {
-                        const message = JSON.parse(new TextDecoder().decode(payloadBytes));
-                        this.pubsub._handleMessage(topic, message);
-                    }
-                    catch (error) {
-                        console.error('Error decoding targeted pub/sub message:', error);
-                    }
-                    break;
-                }
-                case MESSAGE_SYNC_VERIFIED: {
-                    // Sync message with sequence number and hash verification
-                    // Read sequence number and clientID first
-                    const seqNum = decoding.readVarUint(decoder);
-                    const senderClientID = decoding.readVarUint(decoder);
-                    // 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) {
-                        // 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}`);
-                        }
-                    }
-                    // If we received SyncStep2, we're synced (unless hash mismatched)
-                    if (syncMessageType === syncProtocol.messageYjsSyncStep2 &&
-                        !this._synced &&
-                        localHash === expectedHash) {
-                        this._synced = true;
-                        this.emit('synced', [true]);
-                    }
-                    // 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) {
-                        if (this.awareness.getStates().size >= 3) {
-                            this._scheduleSyncReply(encoding.toUint8Array(encoder));
-                        }
-                        else {
-                            this._sendSyncReply(encoding.toUint8Array(encoder));
-                        }
-                    }
-                    break;
-                }
-                default:
-                    console.warn('Unknown message type:', messageType);
-            }
+            this._dispatchMessage(message, from);
         }
         catch (error) {
             // This should only happen for logic errors, not corruption
@@ -836,36 +1708,618 @@ export class GenericProvider extends Observable {
         }
     }
     /**
-     * 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.
+     * Decode and act on one already-integrity-verified, already-decompressed
+     * message. Split out of `_processWrappedMessage()` so `MESSAGE_BATCH`
+     * (see `_sendBatch()`) can recurse into this for each sub-message it
+     * unwraps, running the EXACT SAME per-message-type logic used for a
+     * top-level message rather than a parallel reimplementation. A thrown
+     * error partway through a batch's sub-messages aborts the REST of that
+     * batch (propagates up to `_processWrappedMessage()`'s catch) - same as
+     * a logic error aborting a single top-level message today, just now
+     * scoped to "the rest of this batch" instead of "this one message".
      */
-    _scheduleSyncReply(reply) {
-        if (this._pendingSyncReplyTimeoutId !== undefined) {
-            if (this._pendingSyncReply) {
-                this._sendSyncReply(this._pendingSyncReply);
+    _dispatchMessage(message, from) {
+        const decoder = decoding.createDecoder(message);
+        const messageType = decoding.readVarUint(decoder);
+        switch (messageType) {
+            case MESSAGE_BATCH: {
+                // Payload is N length-prefixed sub-messages (writeVarUint8Array
+                // per sub-message, mirroring MESSAGE_AWARENESS's own framing).
+                // Each was NOT individually CRC32-wrapped - see _sendBatch()'s doc
+                // comment - so just decode and dispatch each one directly.
+                while (decoding.hasContent(decoder)) {
+                    const subMessage = decoding.readVarUint8Array(decoder);
+                    this._dispatchMessage(subMessage, from);
+                }
+                break;
             }
-            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;
+            case MESSAGE_SYNC_DIGEST: {
+                this._handleDigest(decoder, from);
+                break;
             }
-        }, delay);
-    }
+            case MESSAGE_SYNC_PUSH: {
+                // Somebody's whole document: apply it, nothing else (see the
+                // constant's comment for why no hash check and no synced flip).
+                Y.applyUpdate(this.doc, decoding.readVarUint8Array(decoder), this);
+                this._checkPendingAfterReply();
+                break;
+            }
+            case MESSAGE_SYNC: {
+                const encoder = encoding.createEncoder();
+                encoding.writeVarUint(encoder, MESSAGE_SYNC);
+                const syncMessageType = syncProtocol.readSyncMessage(decoder, encoder, this.doc, this);
+                if (syncMessageType === syncProtocol.messageYjsSyncStep2) {
+                    // If we received SyncStep2, we're synced (and confirmed: we have
+                    // heard from a peer that had data for us)
+                    this._confirmed = true;
+                    this._noteResponse(true);
+                    this._markSynced();
+                    // Design E (phase 1d): a SyncStep2 carries its encoder's own pending
+                    // structs, so a responder missing a struct hands us the same hole.
+                    // Our response wait just ended - make sure something still asks.
+                    this._checkPendingAfterReply();
+                    // Someone else's SyncStep2 reply just arrived - our own pending
+                    // reply (if any) is now most likely redundant.
+                    this._cancelPendingSyncReply();
+                }
+                // 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._replyToSyncRequest(encoding.toUint8Array(encoder));
+                }
+                break;
+            }
+            case MESSAGE_AWARENESS: {
+                const payload = decoding.readVarUint8Array(decoder);
+                // Someone else's removal broadcast just arrived - cancel our own
+                // suppressed removal for the same clientID(s), if pending (see
+                // _scheduleAwarenessRemoval()). This MUST be checked here, at the
+                // wire-message level, before calling applyAwarenessUpdate() below -
+                // by the time this arrives we've very likely already independently
+                // detected and applied the SAME removal ourselves (every peer's
+                // 30s outdatedTimeout sweep fires near-simultaneously, well before
+                // this message's network delay elapses - see
+                // test/dummy/bench-awareness-removal-burst.ts), so
+                // applyAwarenessUpdate() below will see a clock it already knows
+                // and emit no 'update' event at all - mirrors exactly how the
+                // MESSAGE_SYNC/MESSAGE_SYNC_VERIFIED case above cancels
+                // _pendingSyncReply on seeing a SyncStep2 message TYPE arrive, not
+                // on whether it changed anything locally.
+                const scan = this._scanAwarenessPayload(payload);
+                if (scan.removed.length > 0) {
+                    this._cancelPendingAwarenessRemovalIfOverlaps(scan.removed);
+                }
+                if (scan.coversUs)
+                    this._presenceCovered = true;
+                // Round 5, item 3: with Trickle, settled peers rarely beacon, so a
+                // joiner would learn them only from the few phase-winning beacons
+                // per interval. The relayed presence table names everyone - ids
+                // only, addresses stay beacon-learned (unicast needs a `from`).
+                const heardAt = Date.now();
+                for (const id of scan.present) {
+                    if (id !== this.doc.clientID)
+                        this._knownPeers.set(id, heardAt);
+                }
+                awarenessProtocol.applyAwarenessUpdate(this.awareness, payload, this);
+                break;
+            }
+            case MESSAGE_PUBSUB: {
+                // Read topic
+                const topic = decoding.readVarString(decoder);
+                // Read message payload
+                const payloadBytes = decoding.readVarUint8Array(decoder);
+                try {
+                    // Decode JSON payload
+                    const decoder = new TextDecoder();
+                    const payloadStr = decoder.decode(payloadBytes);
+                    const message = JSON.parse(payloadStr);
+                    // Emit to pubsub channel
+                    this.pubsub._handleMessage(topic, message);
+                }
+                catch (error) {
+                    console.error('Error decoding pub/sub message:', error);
+                }
+                break;
+            }
+            case MESSAGE_AWARENESS_APP: {
+                // Deliberately none of the presence bookkeeping the core awareness
+                // case does (_knownPeers, _presenceCovered, removal cancellation):
+                // this channel is written by untrusted modules and must not be able
+                // to influence the room's view of who is present.
+                awarenessProtocol.applyAwarenessUpdate(this.appAwareness, decoding.readVarUint8Array(decoder), this);
+                break;
+            }
+            case MESSAGE_PUBSUB_TARGETED: {
+                const target = decoding.readVarString(decoder);
+                const topic = decoding.readVarString(decoder);
+                const payloadBytes = decoding.readVarUint8Array(decoder);
+                // Drop messages aimed at someone else (broadcast-and-filter path).
+                if (this._localId !== undefined && target !== this._localId) {
+                    break;
+                }
+                try {
+                    const message = JSON.parse(new TextDecoder().decode(payloadBytes));
+                    this.pubsub._handleMessage(topic, message);
+                }
+                catch (error) {
+                    console.error('Error decoding targeted pub/sub message:', error);
+                }
+                break;
+            }
+            case MESSAGE_SYNC_VERIFIED: {
+                // Sync message with sequence number and hash verification
+                // Read sequence number and clientID first
+                const seqNum = decoding.readVarUint(decoder);
+                const senderClientID = decoding.readVarUint(decoder);
+                // Track for gap detection only — does NOT gate whether we apply
+                // the update below (see _trackRemoteSeq() for why).
+                this._trackRemoteSeq(senderClientID, seqNum);
+                this._knownPeers.set(senderClientID, Date.now());
+                if (from !== undefined)
+                    this._peerAddress.set(senderClientID, from);
+                this._touchPeer(senderClientID);
+                // 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.
+                // Peek at the sync sub-message's payload (SyncStep2/Update carry
+                // an update as a varUint8Array right after the sub-type) without
+                // consuming the decoder - _isLateUpdate() needs the bytes after
+                // readSyncMessage() has applied them.
+                const updateBytes = this._peekSyncUpdate(decoder);
+                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();
+                    this._confirmed = true;
+                    this._checkPendingAfterReply();
+                    this._noteResponse(true);
+                }
+                // 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) {
+                    // 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);
+                    // A late update - one whose content we had already been past
+                    // when it arrived (its sender's clock in the update is below
+                    // ours, because a SyncStep2 or a reordered later update got here
+                    // first) - carries a hash of a state we have legitimately moved
+                    // beyond. Its mismatch says nothing about anything we lack; it
+                    // was the other half of the item-13 cascade (a resync's reply
+                    // fast-forwards a peer, then every in-flight keystroke behind it
+                    // mismatches). Cheap to detect from the update's own metadata.
+                    const lateUpdate = this._isLateUpdate(updateBytes);
+                    // Yjs's own verdict: if the update could not be fully integrated
+                    // because a causal dependency is missing, the struct store holds
+                    // it as pending. That is the ONE mismatch that is evidence of a
+                    // gap - and under jitter it is usually a reordering that the
+                    // next few ms resolve, so it gets the same grace a sequence gap
+                    // gets before a beacon goes out (_schedulePendingCheck). It also
+                    // covers the case the sequence anchor no longer does: since the
+                    // connect push carries no sequence number (MESSAGE_SYNC_PUSH), a
+                    // peer's first keystroke can be the first numbered message we see
+                    // from it, and a reordered first burst has no earlier number to
+                    // open a gap against.
+                    const pending = this.doc.store.pendingStructs !== null ||
+                        this.doc.store.pendingDs !== null;
+                    if (pending) {
+                        this._schedulePendingCheck();
+                    }
+                    else if (!reorderingSuspected && !lateUpdate) {
+                        // 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}`);
+                    }
+                }
+                // If we received SyncStep2, we're synced (unless hash mismatched)
+                if (syncMessageType === syncProtocol.messageYjsSyncStep2 &&
+                    localHash === expectedHash) {
+                    this._markSynced();
+                }
+                // 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._replyToSyncRequest(encoding.toUint8Array(encoder));
+                }
+                break;
+            }
+            default:
+                console.warn('Unknown message type:', messageType);
+        }
+    }
+    /**
+     * Handle a digest beacon (MESSAGE_SYNC_DIGEST). Reply rule (design doc
+     * §3): SyncStep2 if the sender is behind us or its delete-set hash
+     * differs from ours (the SyncStep2 always carries our full delete set, so
+     * it also heals a lost delete on their side - and their beacon does the
+     * same for us, symmetrically, within one interval); our own beacon as an
+     * ack if the beacon is JOIN-flagged and states are equal; nothing
+     * otherwise - which is what removes the ~5-12 empty replies per heartbeat
+     * measured at N=50 in test/dummy/bench-idle-room.ts. "Sender is ahead of
+     * us" triggers no reply: our own next beacon fetches it. Nothing here
+     * removes a recovery path (the round-2 lesson in
+     * 2026-09-04-resync-message-reduction-design.md's addendum), only
+     * replies that carry no information.
+     *
+     * `synced`: a beacon we are not behind, with equal delete-set hash, is a
+     * stronger statement than the empty SyncStep2 it replaces ("you lack
+     * nothing I have"), so it marks us synced too - this is what keeps two
+     * fresh peers, or a whole concurrent join burst, converging to `synced`
+     * with no acks needing to survive the rate limiter.
+     */
+    _handleDigest(decoder, from) {
+        decoding.readVarUint(decoder); // DIGEST_VERSION - append-only, nothing to branch on yet
+        const flags = decoding.readVarUint(decoder);
+        const senderClientID = decoding.readVarUint(decoder);
+        this._knownPeers.set(senderClientID, Date.now());
+        if (from !== undefined)
+            this._peerAddress.set(senderClientID, from);
+        this._touchPeer(senderClientID);
+        const remoteSv = decoding.readVarUint8Array(decoder);
+        const remoteDsHash = decoding.readVarUint(decoder);
+        // Any trailing bytes belong to a newer version; ignored by design.
+        const remote = Y.decodeStateVector(remoteSv);
+        const local = Y.decodeStateVector(Y.encodeStateVector(this.doc));
+        let senderBehind = false;
+        for (const [client, clock] of local) {
+            if ((remote.get(client) ?? 0) < clock) {
+                senderBehind = true;
+                break;
+            }
+        }
+        let weBehind = false;
+        for (const [client, clock] of remote) {
+            if ((local.get(client) ?? 0) < clock) {
+                weBehind = true;
+                break;
+            }
+        }
+        const dsEqual = remoteDsHash === this._deleteSetHash();
+        const equal = !senderBehind && !weBehind && dsEqual;
+        if (equal) {
+            // Any peer holding exactly our state has confirmed it - see _confirmedSv.
+            this._confirmedSv = remoteSv;
+            this._confirmedDsHash = remoteDsHash;
+        }
+        if (flags & DIGEST_FLAG_ACK) {
+            // Somebody confirmed the echoed state (see DIGEST_FLAG_ACK). If it is
+            // ours, we're synced; if we were about to confirm the same state, we
+            // no longer need to. Never a request: no reply, no presence.
+            if (equal) {
+                if (flags & DIGEST_FLAG_SETTLED) {
+                    this._confirmed = true;
+                    this._noteResponse(true);
+                }
+                else {
+                    this._equalUnsettledSeen = true;
+                }
+                this._markSynced();
+                this._cancelPendingAck();
+            }
+            return;
+        }
+        // An equal-digest beacon from a settled peer's periodic tick has just
+        // told the room (including whoever our pending ack was for) that this
+        // state is confirmed. Our ack would say the same thing again. Task 3b
+        // in the design doc: without this, acks were ~94% of a 50-peer join
+        // burst's messages, throttled only by the rate limiter. An equal JOIN
+        // beacon (another joiner in the same burst) asks for an ack too; our
+        // pending ack - identical bytes, since it echoes that same state -
+        // answers it as well, and _scheduleSyncReply() dedupes it below.
+        if (equal && !(flags & (DIGEST_FLAG_JOIN | DIGEST_FLAG_CONFIRM))) {
+            // A peer's periodic/resync beacon in our state: it makes our pending
+            // ack redundant, and - if that peer is confirmed - it answers our
+            // own outstanding join as well as any ack would. It also counts for
+            // Trickle: the room has just compared itself against our digest.
+            this._equalBeaconsHeard++;
+            if (flags & DIGEST_FLAG_SETTLED) {
+                this._confirmed = true;
+                this._noteResponse(false);
+            }
+            else {
+                this._equalUnsettledSeen = true;
+            }
+            this._cancelPendingAck();
+        }
+        if (senderBehind || !dsEqual) {
+            const encoder = encoding.createEncoder();
+            encoding.writeVarUint(encoder, MESSAGE_SYNC);
+            syncProtocol.writeSyncStep2(encoder, this.doc, remoteSv);
+            let reply = encoding.toUint8Array(encoder);
+            if (!dsEqual) {
+                // Round 5: a delete moves no clock, so the hash cannot say which of
+                // us lacks a delete - and a SyncStep2 heals only its receiver. The
+                // loser of a delete-only update was healed only by its OWN next
+                // beacon, which idle backoff parks for up to 60 s (bench-idle-room
+                // part b failed its 11 s cap at a 5 s interval). So the reply asks
+                // back: our beacon rides in the same batch, and if the sender is
+                // the one ahead it answers with its delete set. Terminates: a peer
+                // healed by the SyncStep2 half sees an equal digest in the beacon
+                // half and stays silent. (A behind-check timer on the mismatch was
+                // tried first: every peer that overheard the loser's beacon armed
+                // one and asked - ~2x the deliveries of a lossy edit burst on the
+                // Matrix profile.)
+                reply = this._encodeBatch([reply, this._encodeSyncStep1(0)]);
+            }
+            this._replyToSyncRequest(reply, false, remoteSv, senderClientID);
+        }
+        else if (flags & (DIGEST_FLAG_JOIN | DIGEST_FLAG_CONFIRM)) {
+            this._replyToSyncRequest(this._encodeAck(remoteSv, remoteDsHash), true, null, senderClientID);
+        }
+        if (!weBehind && dsEqual) {
+            this._markSynced();
+        }
+        if (weBehind) {
+            // The sender has something we lack. Not a request yet - the update
+            // may still be in flight (see _scheduleBehindCheck).
+            this._scheduleBehindCheck(remoteSv);
+        }
+        if (flags & DIGEST_FLAG_JOIN && this.awareness.getLocalState() !== null) {
+            // Presence on demand: the joiner asked. One broadcast per burst of
+            // joiners (see _schedulePresenceResponse), never suppressed (each
+            // responder's state is distinct). Skipped when we have no state to
+            // announce.
+            this._schedulePresenceResponse(senderClientID);
+        }
+    }
+    /**
+     * Answer a JOIN beacon's presence request once for all JOIN beacons that
+     * arrive within `clamp(2 * minRTT, 100, 500)` ms of the first - long
+     * enough to cover a join burst spread by latency, short enough that a
+     * lone joiner sees the room's presence within a few round trips.
+     */
+    _schedulePresenceResponse(requester) {
+        this._presencePending.add(requester);
+        if (this._presenceResponseTimer !== undefined)
+            return;
+        this._presenceCovered = false;
+        const rtt = this._rttMinMs();
+        // Phase 1e, relay path: one peer per 2 s bucket - the first-ranked
+        // for a constant requester - relays the whole awareness table at once
+        // (the way y-websocket's server does; clocks travel with the states,
+        // a peer's own echoed state at an equal clock is ignored by
+        // applyAwarenessUpdate). Everyone else waits the usual window, and
+        // stays silent if that table carried their state (_presenceCovered).
+        // Presence per late join: ~N deliveries instead of (N-1)^2 - it was
+        // 82% of a late join into a 100-peer room (bench-join-census). Bytes
+        // are unchanged (the table goes to everyone). Peers that cannot yet
+        // tell (no RTT estimate on a slow link, so the table may arrive after
+        // their window) fall back to the broadcast of their own state, as
+        // before. The unicast path below is untouched.
+        // Needs a view of the room: in a fresh burst the first JOIN arrives
+        // before any peer is known, everyone would rank first and relay a
+        // table each - the broadcast fallback is right for that case.
+        const relayer = !this._canUnicast(requester) &&
+            this._knownPeers.size >= 3 &&
+            this._responderRank(0, 1) === 0;
+        const delay = relayer ? 0 : Math.min(500, Math.max(100, rtt === null ? 0 : 2 * rtt));
+        this._presenceResponseTimer = setTimeout(() => {
+            this._presenceResponseTimer = undefined;
+            const requesters = Array.from(this._presencePending);
+            this._presencePending.clear();
+            if (this._destroying || !this.transport.isConnected)
+                return;
+            // The app channel has no presence machinery of its own (deliberately -
+            // it must not drive _knownPeers), so it has nothing that would tell a
+            // joiner about existing cursors. A JOIN is the one signal that someone
+            // needs our state, so answer it on both channels. Sent on its own wire
+            // type and its own throttle; the relay/unicast fan-out optimisations
+            // below apply to core presence only, since the app table is typically
+            // a few cursors rather than the room's full census.
+            this._announceAppAwareness();
+            if (this.awareness.getLocalState() === null)
+                return;
+            // Every joiner covered by this timer is addressable: one unicast
+            // each ((N-1) deliveries per joiner room-wide) instead of one
+            // broadcast ((N-1)^2). Otherwise the broadcast, as before.
+            if (requesters.every((id) => this._canUnicast(id))) {
+                const msg = this._encodeAwareness([this.doc.clientID]);
+                for (const id of requesters)
+                    this._sendDirect(id, msg);
+            }
+            else if (relayer) {
+                this._sendAwarenessNow(Array.from(this.awareness.getStates().keys()));
+            }
+            else if (!this._presenceCovered) {
+                this._broadcastAwareness([this.doc.clientID]);
+            }
+        }, delay);
+    }
+    /**
+     * Max random delay (ms) before replying to a SyncStep1 request, scaled by
+     * a room-size signal already available (`this.awareness.getStates().size`
+     * - the same signal read at the `>= 3` suppression gate). A fixed window
+     * (the pre-fix behavior: always `_syncReplySuppressionMs`) doesn't scale
+     * with room size, so a larger room has more independent repliers racing
+     * to answer the same request within the same window - more of them lose
+     * the race and get silently dropped by the `_sendSyncReply()` rate-limit
+     * backstop instead of never sending in the first place. Measured in
+     * test/dummy/bench-corruption-storm.ts: the SyncStep2/SyncStep1 ratio (
+     * ideally ~1 if suppression alone were sufficient) grew from ~1.1-1.3 at
+     * N=2 to ~4.5-5.9 at N=10 with the fixed 30ms window.
+     *
+     * `min(cap, base * log2(peerCount))` - log2 growth spreads replies over a
+     * wider window as the room grows without the delay exploding at very high
+     * N. Capped at 200ms: the slowest-profile round trip this project
+     * benchmarks against (Matrix, ~350ms one-way) already tolerates hundreds
+     * of ms of latency, so 200ms of extra requester-perceived delay stays
+     * well inside that budget while still giving a 100-peer room roughly
+     * 6-7x the base window instead of an unbounded one.
+     */
+    /**
+     * How many peers we believe are in the room: awareness states (includes
+     * ourselves) or, if larger, the distinct beacon/update senders we have
+     * heard plus ourselves. See `_knownPeers`.
+     */
+    _peerCount() {
+        return Math.max(this.awareness.getStates().size, this._knownPeers.size + 1);
+    }
+    /**
+     * Resolves `_awarenessInterval` to a concrete ms value: the configured
+     * fixed number, or (round 6, item 9) `max(transport hint ?? 100,
+     * AWARENESS_AUTO_MS_PER_PEER * peerCount)` when set to `'auto'`.
+     */
+    _effectiveAwarenessInterval() {
+        if (this._awarenessInterval !== 'auto')
+            return this._awarenessInterval;
+        const hint = this.transport.preferredAwarenessMs ?? 100;
+        return Math.max(hint, GenericProvider.AWARENESS_AUTO_MS_PER_PEER * this._peerCount());
+    }
+    _replySuppressionMaxDelay() {
+        const peerCount = this._peerCount();
+        const byRoomSize = Math.min(200, this._syncReplySuppressionMs * Math.log2(Math.max(2, peerCount)));
+        // Phase 1b: the window must exceed the one-way latency or nobody
+        // overhears anybody in time (see _rttSamples). 1.5x the smallest
+        // observed round trip, capped at 2 s - on a 350 ms profile that is
+        // ~1 s of extra requester-perceived delay in exchange for ~1 reply
+        // instead of ~20.
+        const rtt = this._rttMinMs();
+        return rtt === null ? byRoomSize : Math.min(2000, Math.max(byRoomSize, 1.5 * rtt));
+    }
+    /**
+     * 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* 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
+     * ways a reply gets dropped are `_cancelPendingSyncReply()` (we overheard
+     * someone else's SyncStep2 for the SAME request), `_cancelPendingAck()`,
+     * and the identical-bytes case below.
+     *
+     * Identical-bytes case (Task 3c in the design doc): K peers with the same
+     * state asking at once (K empty joiners in a burst) get K byte-identical
+     * SyncStep2s from us - the same full document K times, one flushed
+     * immediately per arriving request, each burning a rate-limit slot. If
+     * the new reply's bytes equal the pending reply's bytes, the pending one
+     * already answers this request too: keep it (same delay, same
+     * suppression) and drop the new one. Measured in
+     * test/dummy/bench-join-after-burst.ts.
+     */
+    _scheduleSyncReply(reply, isAck = false, targetSv = null, requester = null) {
+        if (this._pendingSyncReplyTimeoutId !== undefined && this._pendingSyncReply !== null) {
+            if (bytesEqual(this._pendingSyncReply, reply)) {
+                return; // identical answer already scheduled
+            }
+            if (targetSv !== null &&
+                this._pendingSyncReplyTargetSv !== null &&
+                bytesEqual(this._pendingSyncReplyTargetSv, targetSv)) {
+                // Same question (same requester state), newer document: refresh
+                // the answer, keep the timer. See _pendingSyncReplyTargetSv.
+                this._pendingSyncReply = reply;
+                return;
+            }
+            if (!isAck &&
+                !this._pendingSyncReplyIsAck &&
+                targetSv !== null &&
+                this._pendingSyncReplyTargetSv !== null) {
+                // Two requesters, both behind, different states: one SyncStep2
+                // from the componentwise minimum of both state vectors contains
+                // everything either of them lacks. Keep the pending reply's timer
+                // and widen its content instead of flushing it. The flush (the
+                // rule below, kept for acks and legacy SyncStep1s) sent an
+                // unsuppressed broadcast for every second request that arrived
+                // inside the suppression window; with the window at 1.5x RTT on a
+                // 350 ms link and several peers behind after a lossy edit burst
+                // that was ~4-5 broadcast replies per beacon (phase-1c results).
+                const merged = minStateVector(this._pendingSyncReplyTargetSv, targetSv);
+                const encoder = encoding.createEncoder();
+                encoding.writeVarUint(encoder, MESSAGE_SYNC);
+                syncProtocol.writeSyncStep2(encoder, this.doc, merged);
+                this._pendingSyncReply = encoding.toUint8Array(encoder);
+                this._pendingSyncReplyTargetSv = merged;
+                return;
+            }
+            // Phase 1e: an ack and a SyncStep2 never flush each other. A room
+            // whose JOIN waits expire together sends N CONFIRMs in the same
+            // millisecond; while a lossy edit burst is still healing, some of
+            // them find us equal (ack) and some behind (SyncStep2), and the flush
+            // rule below turned every type change into an immediate broadcast -
+            // ~650 replies in 200 ms at N=50, 5 % loss (probe timeline in the
+            // phase-1e design doc). An ack adds nothing to a pending reply of
+            // either kind (the requester's wait retries, or an equal peer's
+            // SETTLED ack confirms it); a SyncStep2 replaces a pending ack.
+            if (isAck)
+                return;
+            if (this._pendingSyncReplyIsAck) {
+                clearTimeout(this._pendingSyncReplyTimeoutId);
+                this._pendingSyncReplyTimeoutId = undefined;
+            }
+        }
+        if (this._pendingSyncReplyTimeoutId !== undefined) {
+            // Only a legacy plain SyncStep1 (no target state vector) still
+            // flushes a pending SyncStep2.
+            if (this._pendingSyncReply) {
+                this._sendSyncReply(this._pendingSyncReply);
+            }
+            clearTimeout(this._pendingSyncReplyTimeoutId);
+            this._pendingSyncReplyTimeoutId = undefined;
+        }
+        this._pendingSyncReply = reply;
+        this._pendingSyncReplyIsAck = isAck;
+        this._pendingSyncReplyTargetSv = targetSv;
+        // Acks keep the uniform window: they carry no data, so their delay
+        // costs nothing but a few ms on a joiner's `synced` flip, and in a
+        // join burst one pending ack answers every equal JOIN that arrives
+        // inside that window (identical bytes, deduped above). Ranked, rank 0
+        // fired at once for every requester - measured: fresh-burst acks
+        // 6,039 -> 15,147 at Gun N=100, join-after-burst Matrix 147 -> 686.
+        const delay = isAck
+            ? Math.random() * this._replySuppressionMaxDelay()
+            : this._replyDelay(requester);
+        this._pendingSyncReplyTimeoutId = setTimeout(() => {
+            this._pendingSyncReplyTimeoutId = undefined;
+            if (this._pendingSyncReply) {
+                this._sendSyncReply(this._pendingSyncReply);
+                this._pendingSyncReply = null;
+                this._pendingSyncReplyTargetSv = null;
+            }
+        }, delay);
+    }
     /** Cancel a pending suppressed reply, if any. */
     _cancelPendingSyncReply() {
         if (this._pendingSyncReplyTimeoutId !== undefined) {
@@ -873,6 +2327,497 @@ export class GenericProvider extends Observable {
             this._pendingSyncReplyTimeoutId = undefined;
         }
         this._pendingSyncReply = null;
+        this._pendingSyncReplyIsAck = false;
+        this._pendingSyncReplyTargetSv = null;
+    }
+    /**
+     * Cancel a pending reply only if it is a digest ack - see
+     * `_pendingSyncReplyIsAck`. Called from `_handleDigest()` on every
+     * overheard beacon whose digest equals ours.
+     */
+    _cancelPendingAck() {
+        if (this._pendingSyncReplyIsAck) {
+            this._cancelPendingSyncReply();
+        }
+    }
+    /**
+     * Route a SyncStep2 (or digest-ack) reply through the redundancy
+     * suppression when there's genuine redundancy (>= 2 other known peers via
+     * awareness - below that there's no "someone else" to rely on), else send
+     * immediately. Both paths are rate-limited by `_sendSyncReply()`. Shared
+     * by the MESSAGE_SYNC, MESSAGE_SYNC_VERIFIED and MESSAGE_SYNC_DIGEST cases.
+     */
+    _replyToSyncRequest(reply, isAck = false, targetSv = null, toClientID) {
+        // A peer that knows it is incomplete does not answer. A SyncStep2 is
+        // encoded from integrated structs only, so with structs (or a delete
+        // set) still pending ours would be provably partial - and the
+        // requester's response wait ends on the first SyncStep2 it gets, so a
+        // partial answer strands it until its next trigger (phase-1c gates:
+        // 5 s resync backoff, or a stall with syncInterval 0). In relay mode a
+        // partial broadcast also cancels the complete replies other peers had
+        // pending. Let them answer; the requester retries if nobody does, and
+        // in unicast mode the rank bucket rotates the responders every 2 s.
+        // An ack from us would likewise confirm a state we do not trust.
+        if (this.doc.store.pendingStructs !== null ||
+            this.doc.store.pendingDs !== null) {
+            return;
+        }
+        // Unicast path (transport has sendTo and we know the requester's
+        // address): nobody overhears a unicast, so the delay-and-cancel
+        // suppression below cannot thin the replies. Instead each candidate
+        // responder decides for itself whether it is one of ~3 that answer
+        // (_selectedResponder), and answers at once - no suppression delay,
+        // one delivery. The requester's response wait retries if all ~3 are
+        // lost. Phase-1c design, item B.
+        if (toClientID !== undefined && this._canUnicast(toClientID)) {
+            if (!this._selectedResponder(toClientID))
+                return;
+            if (!this._tryReserveReplySlot())
+                return;
+            this._sendDirect(toClientID, reply);
+            return;
+        }
+        // Acks ALWAYS take the delayed/suppressible path: they carry no data,
+        // so the only cost of delaying one is a few ms on the joiner's `synced`
+        // flip (measured before this rule: 37,240 of a 50-peer join burst's
+        // 40,915 deliveries were immediate acks). Everything else goes through
+        // suppression once there is someone else who could answer - counted
+        // from beacon/update senders as well as awareness, see _peerCount().
+        if (isAck || this._peerCount() >= 3) {
+            this._scheduleSyncReply(reply, isAck, targetSv, toClientID ?? null);
+        }
+        else {
+            this._sendSyncReply(reply);
+        }
+    }
+    /** Whether a reply to `clientID` can go over Transport.sendTo. */
+    _canUnicast(clientID) {
+        return (typeof this.transport.sendTo === 'function' &&
+            this._peerAddress.has(clientID));
+    }
+    /**
+     * Responder self-selection for unicast replies: the three peers whose
+     * hash for this requester ranks lowest among the peers we know answer
+     * it. Every candidate ranks itself against the same known set, so the
+     * sets agree wherever the views agree, and the peer that ranks first in
+     * the true order always ranks first in its own view - the selection is
+     * never empty. A 2 s time bucket in the hash rotates the ranking, so
+     * three departed peers at the top only delay a reply until the
+     * requester's next attempt. Everyone answers in rooms of four or fewer.
+     * (A first cut chose each responder independently with probability 3/N;
+     * ~5 % of requests then selected nobody and waited for the 1 s retry.)
+     */
+    _selectedResponder(requester) {
+        if (this._peerCount() < 4)
+            return true;
+        return this._responderRank(requester, 3) < 3;
+    }
+    /**
+     * How many known peers rank below us for `requester` in the current 2 s
+     * bucket (counting stops at `cap`). Shared by unicast self-selection
+     * (rank < 3 answers) and, since phase 1e, the relay-mode reply delay
+     * (rank r waits r slots, see _replyDelay()).
+     */
+    _responderRank(requester, cap) {
+        const bucket = Math.floor(Date.now() / 2000);
+        const rank = (id) => (Math.imul(requester ^ bucket, 0x9e3779b1) ^ Math.imul(id, 0x85ebca6b)) >>> 0;
+        const mine = rank(this.doc.clientID);
+        let better = 0;
+        for (const id of this._knownPeers.keys()) {
+            if (id === requester || id === this.doc.clientID)
+                continue;
+            if (rank(id) < mine && ++better >= cap)
+                break;
+        }
+        return better;
+    }
+    /**
+     * Delay before a suppressible reply goes out (relay path). Phase 1e:
+     * ranked, not uniform. A uniform draw from [0, W] lets ~N * L / W
+     * repliers fire before the first reply is overheard (L = one-way
+     * latency): 10-27 SyncStep2 sends per request at N=100 in
+     * test/dummy/bench-join-census.ts, and the WebRTC join-burst cell's
+     * 16-34k spread. With the responder rank (the same hash the unicast
+     * self-selection uses) rank 0 answers at once and rank r waits r
+     * windows (W = _replySuppressionMaxDelay(), 1.5x the minimum round
+     * trip: with request arrival spread 2jL and reply flight L(1+j), rank 1
+     * has overheard rank 0 iff the slot is >= L(1+3j), which 3L(1-j) covers
+     * up to j~0.33). Ranks >= 8 add a random window on top so a room whose
+     * first eight ranked peers are all gone does not answer in one
+     * avalanche. Without an RTT sample or a requester id (legacy SyncStep1)
+     * the uniform window stays.
+     */
+    _replyDelay(requester) {
+        const window = this._replySuppressionMaxDelay();
+        const rtt = this._rttMinMs();
+        if (requester === null || rtt === null)
+            return Math.random() * window;
+        // Half a window per rank (0.75 x the minimum round trip): a rank that
+        // stays silent (pending structs during a lossy burst, phase 1d B)
+        // costs the requester half a window, not a whole one - at a full
+        // window the Matrix 5 % loss fan-out's median convergence doubled
+        // (1.1 -> 2.2 s, worst 7.6 s); the price is an occasional second
+        // reply where the jitter exceeds ~1/3 (probe numbers in the design doc).
+        const slot = Math.max(this._syncReplySuppressionMs, 0.75 * rtt);
+        const rank = this._responderRank(requester, 8);
+        return rank * slot + (rank >= 8 ? Math.random() * window : 0);
+    }
+    /**
+     * Send one already-encoded message to a single peer over
+     * Transport.sendTo, with the same CRC32 wrapping and optional compression
+     * as a broadcast. Not mirrored to BroadcastChannel (a same-browser tab
+     * never appears as an addressable peer). Returns false if the peer's
+     * address is unknown or the transport cannot unicast.
+     */
+    _sendDirect(clientID, data) {
+        const address = this._peerAddress.get(clientID);
+        if (address === undefined || typeof this.transport.sendTo !== 'function') {
+            return false;
+        }
+        if (!this.transport.isConnected)
+            return false;
+        this._sendToTransport(wrapMessageWithChecksum(data), address);
+        return true;
+    }
+    /**
+     * Re-check Yjs's pending-struct store after the gap grace period and
+     * request a resync (a beacon, see _requestResync) only if something is
+     * still missing. One timer; a check scheduled while one is pending is
+     * absorbed. Cleared on disconnect/destroy.
+     */
+    /**
+     * A beacon (a peer's periodic tick, or its request) just showed its
+     * sender ahead of us. Until phase 1d nothing happened with that: a peer
+     * whose last update was lost (no later message to open a sequence gap
+     * against), or whose request was answered by a responder that was
+     * itself behind, waited for its OWN next periodic beacon - up to
+     * syncInterval, up to idleBackoffMaxMs with idle backoff on. Now we
+     * check again after a grace and, if still behind that state, ask through
+     * the resync coordinator (coalesced, backed off, rate-limited).
+     *
+     * The grace is what keeps this quiet during typing: at Matrix latency
+     * almost every receiver of a periodic beacon is "behind" by a keystroke
+     * that is still in flight (jitter +-140 ms); max(gapGraceMs, 2 x minRTT)
+     * later it has arrived and the check finds nothing to do. A lost
+     * keystroke that opened a sequence gap is already being requested by the
+     * gap check - the outstanding response wait tells us so, and we stay
+     * quiet. One timer, the newest state vector: a later beacon that shows us
+     * behind by more replaces the reference, the timer keeps running.
+     */
+    _scheduleBehindCheck(remoteSv) {
+        this._behindSv = remoteSv;
+        if (this._behindCheckTimer !== undefined)
+            return;
+        const rtt = this._rttMinMs();
+        const delay = Math.max(this._gapGraceMs, rtt === null ? 0 : 2 * rtt);
+        this._behindCheckTimer = setTimeout(() => {
+            this._behindCheckTimer = undefined;
+            const sv = this._behindSv;
+            this._behindSv = null;
+            if (sv === null || this._destroying || !this.transport.isConnected)
+                return;
+            // A request of ours sent within the grace is still being answered.
+            // An OLDER outstanding wait is not a reason to stay behind: a new
+            // room's first peers keep their JOIN wait parked for seconds (three
+            // retries, nobody SETTLED yet), and measured against it the check
+            // never fired - bench-idle-backoff's recovery stayed at one
+            // backed-off interval (1,846 ms) with this line reading
+            // `_responseWaitTimer !== undefined`.
+            if (this._requestSentAt > 0 && Date.now() - this._requestSentAt < delay)
+                return;
+            const remote = Y.decodeStateVector(sv);
+            const local = Y.decodeStateVector(Y.encodeStateVector(this.doc));
+            for (const [client, clock] of remote) {
+                if ((local.get(client) ?? 0) < clock) {
+                    this._requestResync();
+                    return;
+                }
+            }
+        }, delay);
+    }
+    /** Design E: after a reply or push, pending structs mean the sender had the same hole - arm the grace check. */
+    _checkPendingAfterReply() {
+        if (this.doc.store.pendingStructs !== null ||
+            this.doc.store.pendingDs !== null) {
+            this._schedulePendingCheck();
+        }
+    }
+    _schedulePendingCheck() {
+        if (this._pendingCheckTimer !== undefined)
+            return;
+        this._pendingCheckTimer = setTimeout(() => {
+            this._pendingCheckTimer = undefined;
+            if (this._destroying || !this.transport.isConnected)
+                return;
+            // A request of ours may already be outstanding (JOIN or resync beacon
+            // with its response wait running): whatever is pending is what that
+            // request is fetching, and the response wait retries if it is lost.
+            // Asking again here just raced the responders' suppression window
+            // (measured: a joiner's keystrokes-before-content check fired at
+            // 300 ms while the settled peers' replies were still delayed by a
+            // window of up to ~300 ms, and every such re-beacon collected
+            // another round of replies).
+            // ... but only a request that went out within the last grace, whose
+            // answer may still be on its way. Deferring to ANY outstanding
+            // response wait (the rule until phase 1d) parked every pending check
+            // behind a new room's JOIN wait - three retries with nobody SETTLED
+            // yet, 19.6 s on a 700 ms link - so a peer that lost a keystroke in
+            // such a room asked only once an overheard reply happened to clear
+            // that wait (phase-1d design doc, "After Task 4", the fresh-budget
+            // Matrix fan-out). Look again once the recent request is answered or
+            // given up - dropping the check here left a joiner whose SyncStep2
+            // predated the keystrokes it had received with nine pending updates
+            // and no trigger (measured: 2 of 150 lossy 15-peer joins).
+            const rtt = this._rttMinMs();
+            const recent = Math.max(this._gapGraceMs, rtt === null ? 0 : 2 * rtt);
+            if (this._requestSentAt > 0 && Date.now() - this._requestSentAt < recent) {
+                this._schedulePendingCheck();
+                return;
+            }
+            if (this.doc.store.pendingStructs !== null ||
+                this.doc.store.pendingDs !== null) {
+                this._requestResync();
+            }
+        }, this._gapGraceMs);
+    }
+    /**
+     * Return the update payload of a SyncStep2/Update sync sub-message
+     * without advancing `decoder` (null for SyncStep1 or malformed input).
+     * y-protocols frames both as [subType varUint][update varUint8Array].
+     */
+    _peekSyncUpdate(decoder) {
+        const peek = decoding.clone(decoder);
+        try {
+            const subType = decoding.readVarUint(peek);
+            if (subType === syncProtocol.messageYjsSyncStep1)
+                return null;
+            return decoding.readVarUint8Array(peek);
+        }
+        catch {
+            return null;
+        }
+    }
+    /**
+     * Whether an update we just applied was already superseded here: every
+     * client it touches ends at a clock we were at or beyond BEFORE this
+     * update (i.e. it added nothing). Uses the update's own metadata
+     * (`Y.parseUpdateMeta`), O(clients in the update).
+     */
+    _isLateUpdate(updateBytes) {
+        if (!updateBytes)
+            return false;
+        try {
+            const { to } = Y.parseUpdateMeta(updateBytes);
+            if (to.size === 0)
+                return false;
+            const local = Y.decodeStateVector(Y.encodeStateVector(this.doc));
+            for (const [client, clock] of to) {
+                // `to` is the exclusive end clock of the update's range for that
+                // client; our state vector is exclusive too. Equal means the update
+                // brought us exactly here (not late); greater means we were past it.
+                if ((local.get(client) ?? 0) <= clock)
+                    return false;
+            }
+            return true;
+        }
+        catch {
+            return false;
+        }
+    }
+    /** Flip `synced` once and emit; idempotent. */
+    _markSynced() {
+        if (!this._synced) {
+            this._synced = true;
+            this.emit('synced', [true]);
+        }
+    }
+    /**
+     * Wait for a response to the JOIN or resync beacon we just sent. If
+     * neither a SyncStep2 nor an equal ack/beacon arrives within 1s (then
+     * 2s, 4s), ask again - with a CONFIRM beacon after a JOIN (so an equal
+     * room acks), with a plain beacon after a resync (only peers ahead of us
+     * need to answer; an equal room's silence is the correct answer and its
+     * periodic beacons end the wait) - three times at most; after that the
+     * periodic beacon is the fallback, as before. Requester-side retry is how the protocol
+     * stays loss-tolerant now that reply suppression leaves ~1 reply per
+     * request; N-fold redundant replies were the old (accidental) way.
+     */
+    _armResponseWait(retryFlags) {
+        if (this._responseWaitTimer !== undefined)
+            return;
+        this._responseSeen = false;
+        if (this._responseWaitAttempts === 0)
+            this._equalUnsettledSeen = false;
+        this._responseWaitFlags = retryFlags;
+        this._requestSentAt = Date.now();
+        const rtt = this._rttMinMs();
+        const delay = Math.max(1000, rtt === null ? 0 : 4 * rtt) *
+            Math.pow(2, this._responseWaitAttempts);
+        this._responseWaitTimer = setTimeout(() => {
+            this._responseWaitTimer = undefined;
+            // Phase 1e: a fresh room - equal but unsettled peers answered both
+            // the JOIN and one CONFIRM retry, nobody settled did. Two rounds
+            // instead of three; the first confirmed peers' acks then carry
+            // SETTLED for everyone after them. Measured in
+            // test/dummy/bench-join-census.ts (b): the third round was ~a third
+            // of a fresh N=100 room's 92k-delivery join burst.
+            const freshRoomDone = this._equalUnsettledSeen && this._responseWaitAttempts >= 1;
+            if (this._responseSeen ||
+                this._responseWaitAttempts >= 3 ||
+                freshRoomDone ||
+                !this.transport.isConnected ||
+                this._destroying) {
+                // Asked three times, nobody had more for us: we are the room's
+                // state (or its first peer). Bootstraps DIGEST_FLAG_SETTLED in a
+                // brand-new room so later joiners are confirmed by our acks.
+                if (this._responseWaitAttempts >= 3 || freshRoomDone)
+                    this._confirmed = true;
+                this._responseWaitAttempts = 0;
+                return;
+            }
+            this._responseWaitAttempts++;
+            this._sendSyncStep1(this._responseWaitFlags); // rate-limited; a dropped attempt is simply retried next round
+            this._armResponseWait(this._responseWaitFlags);
+        }, delay);
+    }
+    /**
+     * A SyncStep2 or an equal ack/beacon arrived - whatever we asked for is
+     * answered. `sample` = it was a direct reply (SyncStep2/ack), so its
+     * timing is a round-trip sample; an equal periodic beacon from a settled
+     * peer also ends the wait but says nothing about latency.
+     */
+    _noteResponse(sample) {
+        if (sample && this._requestSentAt > 0) {
+            this._rttSamples.push(Date.now() - this._requestSentAt);
+            if (this._rttSamples.length > 8)
+                this._rttSamples.shift();
+        }
+        this._requestSentAt = 0;
+        this._responseSeen = true;
+        this._responseWaitAttempts = 0;
+        if (this._responseWaitTimer !== undefined) {
+            clearTimeout(this._responseWaitTimer);
+            this._responseWaitTimer = undefined;
+        }
+    }
+    /** Minimum of the recent round-trip samples, or null before the first reply. */
+    _rttMinMs() {
+        return this._rttSamples.length === 0 ? null : Math.min(...this._rttSamples);
+    }
+    /**
+     * Delay a pure timeout-removal awareness broadcast and drop it if
+     * another peer's broadcast of the SAME removal is overheard first (see
+     * the `origin === this` branch in `_setupAwarenessSync()`'s handler,
+     * which calls `_cancelPendingAwarenessRemovalIfOverlaps()`) - the exact
+     * same NACK-style suppression `_scheduleSyncReply()` already applies to
+     * SyncStep2 replies, reusing the same room-size-scaled delay
+     * (`_replySuppressionMaxDelay()`).
+     *
+     * A pending removal already queued when this is called is for a
+     * DIFFERENT departure (two peers timing out, or leaving, within the same
+     * window): since round 5 the ids are merged into the pending set and its
+     * timer kept - one broadcast carries both - instead of flushing the
+     * first as an unsuppressed broadcast (with the long leave window below a
+     * burst of departures would have flushed on every peer). An overheard
+     * broadcast trims only the ids it covers from the pending set
+     * (`_cancelPendingAwarenessRemovalIfOverlaps()`).
+     *
+     * Window: the reply-suppression window for timeouts (sweeps are spread
+     * over seconds anyway); for leaves reported by the transport - all
+     * survivors learn of them in the same millisecond - ten times that,
+     * at least a second, so the first broadcast is overheard before the
+     * rest fire. A departure is not urgent: every peer already dropped the
+     * state locally.
+     */
+    _scheduleAwarenessRemoval(clients, origin = 'timeout') {
+        if (this._pendingAwarenessRemovalTimeoutId !== undefined) {
+            const pending = this._pendingAwarenessRemoval ?? [];
+            for (const id of clients)
+                if (!pending.includes(id))
+                    pending.push(id);
+            this._pendingAwarenessRemoval = pending;
+            return;
+        }
+        this._pendingAwarenessRemoval = clients.slice();
+        const window = this._replySuppressionMaxDelay();
+        const delay = Math.random() * (origin === 'peer-left' ? Math.max(1000, 10 * window) : window);
+        this._pendingAwarenessRemovalTimeoutId = setTimeout(() => {
+            this._pendingAwarenessRemovalTimeoutId = undefined;
+            if (this._pendingAwarenessRemoval) {
+                this._broadcastAwareness(this._pendingAwarenessRemoval);
+                this._pendingAwarenessRemoval = null;
+            }
+        }, delay);
+    }
+    /**
+     * Peek at an awareness-update payload (still in
+     * `awarenessProtocol.encodeAwarenessUpdate()`'s wire encoding) for
+     * clientIDs whose state is `null` (a removal), without applying it.
+     * y-protocols/awareness.js doesn't export a standalone decoder for this,
+     * only `applyAwarenessUpdate()` (which also mutates state) and
+     * `modifyAwarenessUpdate()` (which re-encodes) - so this mirrors the
+     * format by hand: varUint length, then per entry
+     * [varUint clientID][varUint clock][varString JSON state]. Used to cancel
+     * a pending suppressed removal (see `_scheduleAwarenessRemoval()`) at the
+     * wire-message level, before `applyAwarenessUpdate()` runs - see the
+     * `MESSAGE_AWARENESS` case's comment for why timing matters here.
+     */
+    _scanAwarenessPayload(payload) {
+        const removed = [];
+        const present = [];
+        let coversUs = false;
+        try {
+            const d = decoding.createDecoder(payload);
+            const len = decoding.readVarUint(d);
+            const ourClock = this.awareness.meta.get(this.awareness.clientID)?.clock ?? 0;
+            for (let i = 0; i < len; i++) {
+                const clientID = decoding.readVarUint(d);
+                const clock = decoding.readVarUint(d);
+                const state = JSON.parse(decoding.readVarString(d));
+                if (state === null)
+                    removed.push(clientID);
+                else {
+                    present.push(clientID);
+                    if (clientID === this.awareness.clientID && clock >= ourClock)
+                        coversUs = true;
+                }
+            }
+        }
+        catch {
+            // Malformed payload - let applyAwarenessUpdate() below be the one
+            // that deals with it (or throws); suppression is a pure optimization,
+            // never worth failing the actual message handling over.
+        }
+        return { removed, present, coversUs };
+    }
+    /**
+     * Trim a pending suppressed removal broadcast by `removedClientIds` -
+     * someone else already broadcast those departures; what they did not
+     * cover stays queued.
+     */
+    _cancelPendingAwarenessRemovalIfOverlaps(removedClientIds) {
+        if (!this._pendingAwarenessRemoval)
+            return;
+        const rest = this._pendingAwarenessRemoval.filter((id) => !removedClientIds.includes(id));
+        if (rest.length === this._pendingAwarenessRemoval.length)
+            return;
+        if (rest.length > 0) {
+            this._pendingAwarenessRemoval = rest; // someone covered part of it; the rest stays queued
+            return;
+        }
+        if (this._pendingAwarenessRemovalTimeoutId !== undefined) {
+            clearTimeout(this._pendingAwarenessRemovalTimeoutId);
+            this._pendingAwarenessRemovalTimeoutId = undefined;
+        }
+        this._pendingAwarenessRemoval = null;
+    }
+    /** Cancel a pending suppressed awareness-removal broadcast, if any. */
+    _cancelPendingAwarenessRemoval() {
+        if (this._pendingAwarenessRemovalTimeoutId !== undefined) {
+            clearTimeout(this._pendingAwarenessRemovalTimeoutId);
+            this._pendingAwarenessRemovalTimeoutId = undefined;
+        }
+        this._pendingAwarenessRemoval = null;
     }
     /**
      * Send a SyncStep2 reply, gated by the same shared per-peer budget as
@@ -894,7 +2839,7 @@ export class GenericProvider extends Observable {
      * would itself become log spam exactly when things are already noisy.
      */
     _sendSyncReply(reply) {
-        if (!this._tryReserveSyncSlot()) {
+        if (!this._tryReserveReplySlot()) {
             return; // Rate limited - drop the reply silently
         }
         this._send(reply);
@@ -1011,8 +2956,27 @@ export class GenericProvider extends Observable {
         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();
+            if (!this.transport.isConnected || this._destroying)
+                return;
+            // A resync trigger means "I may be missing something" - never "the
+            // room is missing my data" (my updates travel on their own, and the
+            // connect-time push covers offline edits). So: ask with a 12-byte
+            // beacon; the peers ahead of me reply with exactly the diff (a
+            // SyncStep2 against my state vector, suppressed as any reply), the
+            // rest stay silent. Until phase 1b this pushed the WHOLE document to
+            // the room on every trigger, and that push - a hashed update of my
+            // state - failed the hash check at every peer ahead of me, which
+            // scheduled a resync of its own: the cascade of research doc item 13
+            // (198,990 deliveries for 10 keystrokes at N=100 on the Gun profile,
+            // the rate limiter's ceiling). If the beacon or its reply is lost,
+            // _armResponseWait() re-beacons (1s/2s/4s); the periodic beacon is
+            // the fallback after that. A rate-limited attempt re-arms through
+            // this same coordinator so a stranded peer keeps retrying.
+            if (this._sendSyncStep1(0)) {
+                this._armResponseWait(0);
+            }
+            else {
+                this._requestResync();
             }
         }, delay);
     }
@@ -1025,48 +2989,90 @@ export class GenericProvider extends Observable {
      * its own uncapped or separately-capped allowance.
      */
     _tryReserveSyncSlot() {
+        return this._tryReserveSlot(this._syncRequestTimes);
+    }
+    /** Same limiter, separate budget, for SyncStep2 replies and acks. */
+    _tryReserveReplySlot() {
+        return this._tryReserveSlot(this._syncReplyTimes);
+    }
+    _tryReserveSlot(times) {
         const now = Date.now();
-        // Clean up old entries outside the rate limit window
-        this._syncRequestTimes = this._syncRequestTimes.filter((t) => now - t < this._syncRequestWindowMs);
-        if (this._syncRequestTimes.length >= this._maxSyncRequestsPerWindow) {
+        // Drop entries outside the rolling window (in place: the arrays are
+        // referenced from two fields)
+        let keep = 0;
+        for (const t of times) {
+            if (now - t < this._syncRequestWindowMs)
+                times[keep++] = t;
+        }
+        times.length = keep;
+        if (times.length >= this._maxSyncRequestsPerWindow) {
             return false;
         }
-        this._syncRequestTimes.push(now);
+        times.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.
+     * Encode the digest beacon that replaces SyncStep1 (see
+     * MESSAGE_SYNC_DIGEST). Still the one place every "request sync" path
+     * goes through (connect()'s syncNow(), the periodic tick,
+     * _requestResync()'s retry), so they all switched together.
      */
-    _writeSyncStep1() {
+    _encodeSyncStep1(flags = 0) {
+        if (this._confirmed)
+            flags |= DIGEST_FLAG_SETTLED;
+        this._touchPeer(this.doc.clientID);
         const encoder = encoding.createEncoder();
-        // SyncStep1 is always sent as standard MESSAGE_SYNC (no verification)
-        // It's just a request, not an assertion of state
-        encoding.writeVarUint(encoder, MESSAGE_SYNC);
-        syncProtocol.writeSyncStep1(encoder, this.doc);
-        this._send(encoding.toUint8Array(encoder));
+        encoding.writeVarUint(encoder, MESSAGE_SYNC_DIGEST);
+        encoding.writeVarUint(encoder, DIGEST_VERSION);
+        encoding.writeVarUint(encoder, flags);
+        encoding.writeVarUint(encoder, this.doc.clientID);
+        encoding.writeVarUint8Array(encoder, Y.encodeStateVector(this.doc));
+        encoding.writeVarUint(encoder, this._deleteSetHash());
+        return encoding.toUint8Array(encoder);
+    }
+    /**
+     * Encode an ack for a JOIN beacon: same framing as a beacon, DIGEST_FLAG_ACK
+     * set, and the JOINER's state vector + delete-set hash echoed back instead
+     * of ours (see DIGEST_FLAG_ACK for why it must never carry our own state).
+     */
+    _encodeAck(ackedSv, ackedDsHash) {
+        this._touchPeer(this.doc.clientID);
+        const encoder = encoding.createEncoder();
+        encoding.writeVarUint(encoder, MESSAGE_SYNC_DIGEST);
+        encoding.writeVarUint(encoder, DIGEST_VERSION);
+        encoding.writeVarUint(encoder, DIGEST_FLAG_ACK | (this._confirmed ? DIGEST_FLAG_SETTLED : 0));
+        encoding.writeVarUint(encoder, this.doc.clientID);
+        encoding.writeVarUint8Array(encoder, ackedSv);
+        encoding.writeVarUint(encoder, ackedDsHash);
+        return 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.
+     * Send the periodic digest beacon. Rate limited to prevent spam. Returns
+     * whether it actually sent (false means rate-limited).
      */
-    _sendSyncStep1() {
+    _sendSyncStep1(flags = 0) {
         if (!this._tryReserveSyncSlot()) {
             console.warn(`[GenericProvider] Sync rate limit exceeded (${this._maxSyncRequestsPerWindow} requests per ${this._syncRequestWindowMs / 1000}s), throttling...`);
-            return; // Drop the request
+            return false; // Drop the request
         }
-        this._writeSyncStep1();
+        this._send(this._encodeSyncStep1(flags));
+        return true;
     }
     /**
-     * Send a document update to the transport.
-     * If verifyUpdates is enabled, includes sequence number and document hash for ordering and desync detection.
+     * Encode a document update, without sending it. Extracted from the old
+     * `_sendUpdate()` so `_trySyncPushPull()` can fold the push half into a
+     * batched wire send (see `_sendBatch()`); `_sendUpdate()` below is the
+     * send-immediately form still used by every other update-emitting path
+     * (the doc-update handler, batch-flush, disconnect/destroy flush) since
+     * those aren't part of this batching effort's scope.
+     *
+     * NOTE: has a side effect (`_localSeqNum++`) - call exactly once per
+     * logical update, same as before.
      */
-    _sendUpdate(update) {
+    _encodeUpdate(update) {
         const encoder = encoding.createEncoder();
         if (this._verifyUpdates) {
+            this._touchPeer(this.doc.clientID);
             // Use verified sync protocol with sequence number and hash
             encoding.writeVarUint(encoder, MESSAGE_SYNC_VERIFIED);
             // Include sequence number and clientID for causal ordering
@@ -1082,6 +3088,35 @@ export class GenericProvider extends Observable {
             encoding.writeVarUint(encoder, MESSAGE_SYNC);
             syncProtocol.writeUpdate(encoder, update);
         }
+        return encoding.toUint8Array(encoder);
+    }
+    /**
+     * Encode a full-state push (see MESSAGE_SYNC_PUSH): the document as one
+     * update, deliberately without the hash and sequence number that
+     * `_encodeUpdate()` adds to incremental updates.
+     */
+    _encodePush(update) {
+        const encoder = encoding.createEncoder();
+        encoding.writeVarUint(encoder, MESSAGE_SYNC_PUSH);
+        encoding.writeVarUint8Array(encoder, update);
+        return encoding.toUint8Array(encoder);
+    }
+    /**
+     * Send a document update to the transport, with whatever awareness change
+     * the throttle is holding folded into the same wire message (round 5,
+     * item 1). If verifyUpdates is enabled, the update carries a sequence
+     * number and document hash for ordering and desync detection.
+     */
+    _sendUpdate(update) {
+        this._sendBatch([this._encodeUpdate(update), ...this._takePendingAwareness()]);
+    }
+    /**
+     * 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));
     }
     /**
@@ -1112,10 +3147,11 @@ export class GenericProvider extends Observable {
         }
     }
     /**
-     * Send a targeted pub/sub message.
-     * Uses transport.sendTo when available (direct delivery), otherwise
-     * broadcasts a targeted frame that non-target providers drop.
-     * Internal method called by PubSubChannel.
+     * Send a pub/sub message to a single target.
+     *
+     * With `Transport.sendTo` the frame is unicast to that peer; without it the
+     * frame is broadcast with the target embedded and dropped on receipt by
+     * every provider whose `localId` differs.
      */
     _sendPubSubTo(target, topic, message) {
         if (!this.transport.isConnected) {
@@ -1154,63 +3190,215 @@ export class GenericProvider extends Observable {
      * Throttled to prevent awareness updates from flooding document sync.
      * Multiple rapid updates are batched together.
      */
-    _broadcastAwareness(clients, channel = AWARENESS_CHANNEL_MAIN) {
+    _broadcastAwareness(clients) {
         if (clients.length === 0)
             return;
-        const isApp = channel === AWARENESS_CHANNEL_APP;
-        const pending = isApp
-            ? this._pendingAppAwarenessClients
-            : this._pendingAwarenessClients;
+        const interval = this._effectiveAwarenessInterval();
         // If throttling is disabled, send immediately
-        if (this._awarenessInterval <= 0) {
-            this._sendAwarenessNow(clients, channel);
+        if (interval <= 0) {
+            this._sendAwarenessNow(clients);
             return;
         }
         // Add clients to pending set
         for (const client of clients) {
-            pending.add(client);
+            this._pendingAwarenessClients.add(client);
         }
         // If we already have a scheduled broadcast, let it handle the batched clients
-        if ((isApp ? this._appAwarenessTimeoutId : this._awarenessTimeoutId) !==
-            undefined) {
+        if (this._awarenessTimeoutId !== undefined) {
             return;
         }
         // Calculate delay - respect minimum interval since last broadcast
         const now = Date.now();
-        const lastTime = isApp ? this._lastAppAwarenessTime : this._lastAwarenessTime;
-        const delay = Math.max(0, this._awarenessInterval - (now - lastTime));
+        const timeSinceLastBroadcast = now - this._lastAwarenessTime;
+        const delay = Math.max(0, interval - timeSinceLastBroadcast);
         // Schedule the batched broadcast
-        const timeoutId = setTimeout(() => {
-            if (isApp) {
-                this._appAwarenessTimeoutId = undefined;
-                this._lastAppAwarenessTime = Date.now();
-            }
-            else {
-                this._awarenessTimeoutId = undefined;
-                this._lastAwarenessTime = Date.now();
+        this._awarenessTimeoutId = setTimeout(() => {
+            this._awarenessTimeoutId = undefined;
+            this._lastAwarenessTime = Date.now();
+            // Send all pending clients in one message - and a pending timed
+            // update batch (`batchUpdates > 0`) with them, update first.
+            const clientsToSend = Array.from(this._pendingAwarenessClients);
+            this._pendingAwarenessClients.clear();
+            if (clientsToSend.length > 0) {
+                this._sendBatch([...this._takePendingUpdate(), this._encodeAwareness(clientsToSend)]);
             }
-            // Send all pending clients in one message
-            const clientsToSend = Array.from(pending);
-            pending.clear();
+        }, delay);
+    }
+    /**
+     * Register the app channel's update listener. Same echo suppression as the
+     * core channel (never re-broadcast what came off the wire), but none of its
+     * presence/removal handling - this channel has no bearing on who the room
+     * thinks is present.
+     */
+    /**
+     * Announce local app-channel state, if the channel is in use and has any.
+     *
+     * Reads the private field, never the getter: an announce must never be the
+     * thing that constructs the instance (that would re-introduce the eager
+     * timer the lazy getter exists to avoid). A provider whose modules never
+     * touch the channel therefore announces nothing, which is correct - there
+     * is no local state to announce.
+     */
+    _announceAppAwareness() {
+        if (this._appAwareness?.getLocalState() != null) {
+            this._broadcastAppAwareness([this.doc.clientID]);
+        }
+    }
+    _attachAppAwareness(aw) {
+        if (this._appAwarenessUpdateHandler)
+            return;
+        this._appAwarenessUpdateHandler = ({ added, updated, removed, }, origin) => {
+            if (origin === this)
+                return;
+            const changedClients = [...added, ...updated, ...removed];
+            if (changedClients.length === 0)
+                return;
+            this._broadcastAppAwareness(changedClients);
+        };
+        aw.on('update', this._appAwarenessUpdateHandler);
+    }
+    /**
+     * Broadcast app-channel awareness. Mirrors `_broadcastAwareness()`'s
+     * throttle, against its own pending set and timer, and deliberately does
+     * NOT piggyback on the core sync batch: module cursor churn must not pull
+     * document or presence traffic onto its cadence (or vice versa).
+     */
+    _broadcastAppAwareness(clients) {
+        if (clients.length === 0)
+            return;
+        const interval = this._effectiveAwarenessInterval();
+        if (interval <= 0) {
+            this._sendAppAwarenessNow(clients);
+            return;
+        }
+        for (const client of clients) {
+            this._pendingAppAwarenessClients.add(client);
+        }
+        if (this._appAwarenessTimeoutId !== undefined) {
+            return;
+        }
+        const delay = Math.max(0, interval - (Date.now() - this._lastAppAwarenessTime));
+        this._appAwarenessTimeoutId = setTimeout(() => {
+            this._appAwarenessTimeoutId = undefined;
+            this._lastAppAwarenessTime = Date.now();
+            const clientsToSend = Array.from(this._pendingAppAwarenessClients);
+            this._pendingAppAwarenessClients.clear();
             if (clientsToSend.length > 0) {
-                this._sendAwarenessNow(clientsToSend, channel);
+                this._sendAppAwarenessNow(clientsToSend);
             }
         }, delay);
-        if (isApp)
-            this._appAwarenessTimeoutId = timeoutId;
-        else
-            this._awarenessTimeoutId = timeoutId;
+    }
+    /** Encode and send an app-channel awareness update immediately. */
+    _sendAppAwarenessNow(clients) {
+        const encoder = encoding.createEncoder();
+        encoding.writeVarUint(encoder, MESSAGE_AWARENESS_APP);
+        encoding.writeVarUint8Array(encoder, awarenessProtocol.encodeAwarenessUpdate(this.appAwareness, clients));
+        this._send(encoding.toUint8Array(encoder));
     }
     /**
-     * Send awareness update immediately without throttling.
+     * Round 5, item 1: the awareness change the throttle is holding rides
+     * along with a wire message that is leaving anyway. Returns the encoded
+     * awareness sub-message (or nothing) and commits the throttle state
+     * exactly as the timer's own flush would. A piggybacked broadcast costs
+     * no message, only its payload bytes, so it goes out early instead of as
+     * its own message up to `_awarenessInterval` later. Measured in
+     * test/dummy/bench-typing-census.ts: a keystroke in an editor binding is
+     * a text insert plus a cursor update - two broadcasts per keystroke
+     * before this, one after. Broadcast paths only: `_sendDirect` and the
+     * BroadcastChannel-only publishes never call this.
      */
-    _sendAwarenessNow(clients, channel = AWARENESS_CHANNEL_MAIN) {
-        const aw = channel === AWARENESS_CHANNEL_APP ? this.appAwareness : this.awareness;
+    _takePendingAwareness() {
+        if (this._awarenessTimeoutId === undefined)
+            return [];
+        clearTimeout(this._awarenessTimeoutId);
+        this._awarenessTimeoutId = undefined;
+        const clients = Array.from(this._pendingAwarenessClients);
+        this._pendingAwarenessClients.clear();
+        if (clients.length === 0)
+            return [];
+        this._lastAwarenessTime = Date.now();
+        return [this._encodeAwareness(clients)];
+    }
+    /**
+     * The counterpart for the timed batch: a `batchUpdates > 0` batch that is
+     * still waiting rides along with an awareness flush (Matrix: both
+     * default to 2 s, so a typist's cursor and text leave as one PUT).
+     */
+    _takePendingUpdate() {
+        if (this._batchTimeoutId === undefined || !this._pendingUpdate)
+            return [];
+        clearTimeout(this._batchTimeoutId);
+        this._batchTimeoutId = undefined;
+        this._flushScheduled = false;
+        const update = this._pendingUpdate;
+        this._pendingUpdate = null;
+        return [this._encodeUpdate(update)];
+    }
+    /**
+     * Encode an awareness update, without sending it. Extracted from the old
+     * `_sendAwarenessNow()` so `_tryImmediateAwarenessMessage()` can fold it
+     * into a batched wire send instead of always sending it as its own
+     * message.
+     */
+    _encodeAwareness(clients) {
         const encoder = encoding.createEncoder();
         encoding.writeVarUint(encoder, MESSAGE_AWARENESS);
-        encoding.writeVarUint(encoder, channel);
-        encoding.writeVarUint8Array(encoder, awarenessProtocol.encodeAwarenessUpdate(aw, clients));
-        this._send(encoding.toUint8Array(encoder));
+        encoding.writeVarUint8Array(encoder, awarenessProtocol.encodeAwarenessUpdate(this.awareness, clients));
+        return encoding.toUint8Array(encoder);
+    }
+    /**
+     * Send awareness update immediately without throttling.
+     */
+    _sendAwarenessNow(clients) {
+        this._send(this._encodeAwareness(clients));
+    }
+    /**
+     * Attempt to build an awareness broadcast message for immediate
+     * inclusion in the same wire send as a sync message a caller is about to
+     * send anyway (see `_trySyncPushPull`'s `buildExtra` parameter), instead
+     * of going through
+     * `_broadcastAwareness()`'s independent debounce.
+     *
+     * Only returns non-null when the throttle would have let an immediate
+     * send through anyway - i.e. no debounced broadcast is already pending
+     * AND (throttling is disabled, or at least `_awarenessInterval` ms have
+     * passed since the last broadcast) - so this never changes awareness
+     * throttle semantics, only whether the resulting message travels as its
+     * own wire send or bundled with a sync message that happens to be going
+     * out "now" too.
+     *
+     * Mutates the same state `_broadcastAwareness()`'s own immediate-send
+     * branches mutate (`_pendingAwarenessClients`, `_lastAwarenessTime`) -
+     * once this returns non-null, the state is already committed as "sent
+     * now", so the caller MUST actually send the returned message (bundled
+     * or standalone) rather than discarding it.
+     */
+    _tryImmediateAwarenessMessage(clients) {
+        if (clients.length === 0)
+            return null;
+        // A debounced broadcast is already scheduled - let it handle these
+        // clients via the normal pending-set path below, don't race it with an
+        // immediate send here.
+        if (this._awarenessTimeoutId !== undefined)
+            return null;
+        const interval = this._effectiveAwarenessInterval();
+        if (interval > 0) {
+            const timeSinceLastBroadcast = Date.now() - this._lastAwarenessTime;
+            if (timeSinceLastBroadcast < interval)
+                return null;
+        }
+        // Merge with anything already pending (normally empty here since no
+        // timer is scheduled, but merge for safety) and commit to sending now -
+        // mirrors exactly what _broadcastAwareness()'s own immediate
+        // (_awarenessInterval <= 0) and debounced-timer-fired branches do.
+        for (const c of clients)
+            this._pendingAwarenessClients.add(c);
+        const clientsToSend = Array.from(this._pendingAwarenessClients);
+        this._pendingAwarenessClients.clear();
+        if (clientsToSend.length === 0)
+            return null;
+        this._lastAwarenessTime = Date.now();
+        return this._encodeAwareness(clientsToSend);
     }
     /**
      * Setup BroadcastChannel for cross-tab communication.
@@ -1244,30 +3432,35 @@ export class GenericProvider extends Observable {
         const encoderSync = encoding.createEncoder();
         encoding.writeVarUint(encoderSync, MESSAGE_SYNC);
         syncProtocol.writeSyncStep1(encoderSync, this.doc);
-        bc.publish(this._bcChannel, wrapMessageWithChecksum(encoding.toUint8Array(encoderSync)), this);
+        this._bcPublish(wrapMessageWithChecksum(encoding.toUint8Array(encoderSync)));
         // Broadcast local state via BroadcastChannel (wrapped with CRC32)
         const encoderState = encoding.createEncoder();
         encoding.writeVarUint(encoderState, MESSAGE_SYNC);
         syncProtocol.writeSyncStep2(encoderState, this.doc);
-        bc.publish(this._bcChannel, wrapMessageWithChecksum(encoding.toUint8Array(encoderState)), this);
+        this._bcPublish(wrapMessageWithChecksum(encoding.toUint8Array(encoderState)));
         // Broadcast local awareness state via BroadcastChannel (wrapped with CRC32)
         if (this.awareness.getLocalState() !== null) {
-            this._publishAwarenessToBroadcastChannel(this.awareness, AWARENESS_CHANNEL_MAIN);
+            const encoderAwareness = encoding.createEncoder();
+            encoding.writeVarUint(encoderAwareness, MESSAGE_AWARENESS);
+            encoding.writeVarUint8Array(encoderAwareness, awarenessProtocol.encodeAwarenessUpdate(this.awareness, [
+                this.doc.clientID,
+            ]));
+            this._bcPublish(wrapMessageWithChecksum(encoding.toUint8Array(encoderAwareness)));
         }
-        if (this.appAwareness.getLocalState() !== null) {
-            this._publishAwarenessToBroadcastChannel(this.appAwareness, AWARENESS_CHANNEL_APP);
+        // Same for the app channel: ongoing changes reach other tabs for free
+        // (_sendAppAwarenessNow goes through _send, which publishes to BC), but
+        // the connect-time announce does not - so without this a second tab saw
+        // the first tab's presence and never its cursors. Guarded on the private
+        // field, like every other announce: never construct the instance here.
+        if (this._appAwareness?.getLocalState() != null) {
+            const encoderApp = encoding.createEncoder();
+            encoding.writeVarUint(encoderApp, MESSAGE_AWARENESS_APP);
+            encoding.writeVarUint8Array(encoderApp, awarenessProtocol.encodeAwarenessUpdate(this._appAwareness, [
+                this.doc.clientID,
+            ]));
+            this._bcPublish(wrapMessageWithChecksum(encoding.toUint8Array(encoderApp)));
         }
     }
-    /**
-     * 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.
      */
@@ -1276,21 +3469,71 @@ export class GenericProvider extends Observable {
             return;
         }
         // Broadcast awareness state with null (indicating disconnect) - wrapped with CRC32
-        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);
-        }
+        const encoder = encoding.createEncoder();
+        encoding.writeVarUint(encoder, MESSAGE_AWARENESS);
+        encoding.writeVarUint8Array(encoder, awarenessProtocol.encodeAwarenessUpdate(this.awareness, [this.doc.clientID], new Map()));
+        this._bcPublish(wrapMessageWithChecksum(encoding.toUint8Array(encoder)));
         // Unsubscribe from channel
         bc.unsubscribe(this._bcChannel, this._bcSubscriber);
         this._bcConnected = false;
         this._bcSubscriber = undefined;
     }
+    /**
+     * Send N already-encoded, already-typed sub-messages as ONE wire message
+     * instead of N separate `transport.send()`/`bc.publish()` calls, when
+     * there's more than one to send. Used at trigger points that
+     * conceptually produce a single event but historically sent multiple
+     * independent messages for it (sync push, sync pull, awareness) - see
+     * `_trySyncPushPull()` (connect-time push + digest beacon + awareness in
+     * one wire message).
+     *
+     * Design (see the task's framing requirements):
+     * - Each sub-message is length-prefixed with `writeVarUint8Array`,
+     *   consistent with how this codebase already frames variable-length
+     *   payloads elsewhere (e.g. MESSAGE_AWARENESS). On receipt,
+     *   `_dispatchMessage()`'s `MESSAGE_BATCH` case unwraps and re-dispatches
+     *   each one through the EXACT SAME per-message-type logic used for a
+     *   top-level message - no parallel reimplementation.
+     * - Sub-messages are NOT individually CRC32-wrapped here - the whole
+     *   batch envelope goes through the normal, single `_send()` pipeline
+     *   below, which wraps the WHOLE envelope in exactly one CRC32 checksum
+     *   (and, if `compressionThresholdBytes` is configured, one compression
+     *   pass) - built and computed exactly like any other outgoing message,
+     *   so this composes with the existing compression pipeline for free
+     *   rather than fighting it with a second, nested wrap/compress step.
+     *   The tradeoff: a single corrupted bit anywhere in a batched wire
+     *   message now invalidates every sub-message it carried, not just one -
+     *   per-sub-message CRC32s would avoid that, at the cost of ~4 extra
+     *   bytes per sub-message for a benefit that only matters under active
+     *   corruption. This tradeoff is exactly what
+     *   test/dummy/bench-corruption-storm.ts and bench-packet-loss.ts exist
+     *   to measure empirically, per this task's validation requirements,
+     *   rather than deciding it by design argument alone.
+     * - BroadcastChannel (cross-tab) traffic is NOT specially batched beyond
+     *   whatever `_send()` already does per call - same-tab-group cross-tab
+     *   traffic is local/cheap, and `_send()` already only issues one
+     *   `bc.publish()` per call regardless, so a batch of N sub-messages
+     *   already becomes exactly one `bc.publish()` call for free once routed
+     *   through here - no separate BC-specific batching logic needed.
+     */
+    _sendBatch(messages) {
+        if (messages.length === 0)
+            return;
+        if (messages.length === 1) {
+            this._send(messages[0]);
+            return;
+        }
+        this._send(this._encodeBatch(messages));
+    }
+    /** The MESSAGE_BATCH envelope of `_sendBatch`, without sending it. */
+    _encodeBatch(messages) {
+        const encoder = encoding.createEncoder();
+        encoding.writeVarUint(encoder, MESSAGE_BATCH);
+        for (const message of messages) {
+            encoding.writeVarUint8Array(encoder, message);
+        }
+        return encoding.toUint8Array(encoder);
+    }
     /**
      * Send data through both BroadcastChannel (if connected) and transport.
      * All messages are wrapped with CRC32 checksum for integrity verification.
@@ -1299,16 +3542,66 @@ export class GenericProvider extends Observable {
     _send(data) {
         // Wrap message with CRC32 checksum
         const wrappedData = wrapMessageWithChecksum(data);
-        // Send via BroadcastChannel to other tabs first
-        if (this._bcConnected) {
-            bc.publish(this._bcChannel, wrappedData, this);
-        }
+        // Send via BroadcastChannel to other tabs first.
+        if (this._bcConnected)
+            this._bcPublish(wrappedData);
         // Send via network transport
         if (!this.transport.isConnected) {
             return;
         }
+        this._sendToTransport(wrappedData);
+    }
+    /**
+     * Publish already-CRC32-wrapped bytes to the other tabs. BroadcastChannel
+     * is same-process - never worth compressing - but when
+     * compressionThresholdBytes is enabled every message still needs the
+     * leading flag byte _handleIncomingMessage() expects regardless of
+     * source, so this sends flag=0 in that case. The ONE place for every BC
+     * publish: the connect-time burst in _setupBroadcastChannel() used to
+     * publish without the flag, and with compression on (the transport
+     * hints made that a default) the other tab read a CRC byte as the flag
+     * and failed to inflate three messages per join (Nostr playground,
+     * 2026-09-06).
+     */
+    _bcPublish(wrappedData) {
+        bc.publish(this._bcChannel, this._compressionThresholdBytes
+            ? prefixCompressionFlag(0, wrappedData)
+            : wrappedData, this);
+    }
+    /**
+     * Send already-CRC32-wrapped bytes to the network transport, compressing
+     * first if compressionThresholdBytes is configured and this payload
+     * clears it. See that option's doc comment for the size threshold
+     * reasoning and the wire-format compatibility tradeoff of enabling it.
+     */
+    _sendToTransport(wrappedData, to) {
+        const threshold = this._compressionThresholdBytes;
+        if (!threshold) {
+            this._dispatchToTransport(wrappedData, to);
+            return;
+        }
+        if (!COMPRESSION_AVAILABLE || wrappedData.length < threshold) {
+            this._dispatchToTransport(prefixCompressionFlag(0, wrappedData), to);
+            return;
+        }
+        compressDeflateRaw(wrappedData)
+            .then((compressed) => {
+            this._dispatchToTransport(prefixCompressionFlag(1, compressed), to);
+        })
+            .catch((error) => {
+            console.error('[GenericProvider] Compression failed, sending uncompressed:', error);
+            this._dispatchToTransport(prefixCompressionFlag(0, wrappedData), to);
+        });
+    }
+    /**
+     * Hand fully-framed bytes to transport.send() - or transport.sendTo() when
+     * a peer address is given - tolerating a sync or async result.
+     */
+    _dispatchToTransport(data, to) {
         try {
-            const result = this.transport.send(wrappedData);
+            const result = to !== undefined && typeof this.transport.sendTo === 'function'
+                ? this.transport.sendTo(to, data)
+                : this.transport.send(data);
             // Handle async send
             if (result instanceof Promise) {
                 result.catch((error) => {
@@ -1344,4 +3637,6 @@ export class GenericProvider extends Observable {
         return this._localSeqNum;
     }
 }
+// ms of throttle added per known peer under `awarenessInterval: 'auto'`
+GenericProvider.AWARENESS_AUTO_MS_PER_PEER = 20;
 //# sourceMappingURL=index.js.map