@fairfox/polly 0.74.0 → 0.74.1

package/dist/src/shared/lib/mesh-webrtc-adapter.d.ts

@@ -606,13 +606,54 @@ export declare class MeshWebRTCAdapter extends NetworkAdapter {
      * theirs), open an initiating slot and fire the SDP offer. Otherwise
      * do nothing — either we are not interested in this peer, we are
      * already connected, or the other side is the one expected to
-     * initiate. */
+     * initiate.
+     *
+     * A `peer-joined` for a peerId we already hold a slot for is first
+     * routed through {@link evictStaleSlotForRejoin}: the notification is
+     * authoritative proof the remote peer's signalling presence is fresh,
+     * so a slot left over from an earlier incarnation of the same peerId
+     * is torn down rather than left to wedge rediscovery. Polly issue
+     * #133. */
     handlePeerJoined(remotePeerId: string): void;
     /** Handle the signalling server's `peers-present` notification sent
      * once to a newcomer. Applies the same filter as handlePeerJoined to
      * every listed peer, so a device joining into an established lobby
-     * dials every knownPeer it is meant to initiate to in one pass. */
+     * dials every knownPeer it is meant to initiate to in one pass.
+     *
+     * Like {@link handlePeerJoined}, each listed peer is first run
+     * through {@link evictStaleSlotForRejoin} so a `peers-present` frame
+     * received on a signalling *reconnect* clears any stale slot before
+     * the dial path re-evaluates it. Polly issue #133. */
     handlePeersPresent(peerIds: string[]): void;
+    /** Tear down a slot left over from an earlier incarnation of a peer
+     * that has just (re)appeared in the signalling roster, so the dial
+     * path can rebuild it.
+     *
+     * A `peer-joined` / `peers-present` frame is authoritative proof that
+     * the remote peer's signalling presence is fresh: its
+     * {@link MeshSignalingClient} has just sent a `join`. If we are still
+     * holding a slot for that peerId, the slot belongs to a previous
+     * incarnation — *unless* the peer kept a live WebRTC session across
+     * its own signalling reconnect, in which case the data channel is
+     * still `open` on a `connected` connection and the slot is genuinely
+     * sound (peers talk directly once the channel is up; signalling is
+     * intermittent). That one shape is kept; every other shape is stale.
+     *
+     * Polly issue #133: before this, a stale slot wedged rediscovery of a
+     * peer that reuses its peerId across short-lived sessions (a CLI
+     * `chat send` re-run from the same keyring is the canonical case).
+     * The `slot-already-exists` gate in {@link evaluateInitiation}
+     * blocked the survivor from re-dialling, and {@link handleOffer}'s
+     * glare rule made the lexicographically-higher survivor ignore the
+     * newcomer's offer too — so the pair never reconnected until the
+     * {@link slotIdleTimeoutMs} watchdog reaped the corpse, far too late
+     * for the request that triggered the rejoin.
+     *
+     * A handshake that is still young and negotiating (`new`/`connecting`
+     * within {@link slotNeverConnectedTimeoutMs}) is left alone — a
+     * duplicate frame arriving mid-handshake must not nuke a connection
+     * that simply is not done yet. */
+    private evictStaleSlotForRejoin;
     /** Construct an initiating slot inside a per-peer try/catch and
      * record any throw as a `fatal-error` rejection on the per-peer
      * decision map so the snapshot surface names it. Every dial entry

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@fairfox/polly",
-  "version": "0.74.0",
+  "version": "0.74.1",
   "private": false,
   "type": "module",
   "description": "Multi-execution-context framework with reactive state and cross-context messaging for Chrome extensions, PWAs, and worker-based applications",