@urun-sh/core 0.4.0 → 0.4.1

package/CHANGELOG.md

@@ -1,5 +1,38 @@
 # Changelog
 
+## 0.4.1
+
+- **FIX: a re-dial no longer leaves the playout element bound to a DEAD track**
+  (urun-infra#2132; the prod qwen-voice "assistant audio drops mid-session and
+  never self-heals, reload fixes it" signature). `recvTransport.consume()` /
+  `consumeData()` are ASYNC — mediasoup-client renegotiates the recv
+  PeerConnection to add the transceiver — so a consume ISSUED on the leg that is
+  about to die can still be in flight when a dropped signaling socket re-dials
+  and `_cleanupMedia()` replaces the whole leg. That late consumer resolved onto
+  a transport the session no longer owns and was registered ANYWAY, LAST: it
+  overwrote `_tracksByName`, emitted the final `'track'` event, and leaked into
+  the live consumer set. Its track reads `readyState === 'live'` (nothing
+  stopped it — its consumer was not in `_consumers` when the cleanup ran), so
+  every liveness filter above the seam accepted it: `<Video>` / `<Audio>` swap
+  tracks into the ONE persistent `MediaStream` that stays the element's
+  `srcObject`, so the healthy, decoding track of the NEW leg was removed and a
+  track belonging to a closed PeerConnection put in its place. The rendered
+  frame froze while the downlink stayed perfectly healthy (SFU `packetsSent`
+  climbing, consumer `framesDecoded` climbing, `pliCount 0`,
+  `attachedToMediaElement: false` on every RTCPeerConnection), and the
+  components' missed-event poll could never heal it because `getTrackByName()`
+  kept answering with the stale track — only a page reload recovered.
+
+  Both consume paths now carry a LEG FENCE across their await, through
+  mediasoup's own identity rather than a timer or a poller: a consumer belongs
+  to the transport that created it, so a consumer whose transport is no longer
+  this session's recv transport (or is closed, or whose dial generation has been
+  superseded, or which landed after an intentional disconnect) is closed through
+  mediasoup-client's own per-consumer lifecycle — which stops its track — and is
+  never registered. The data twin matters just as much: a superseded
+  `consumeData` used to install a DEAD channel as a named stream's data lane,
+  silently swallowing every later payload on that name.
+
 ## 0.4.0
 
 - **BREAKING: `SessionPhaseName` loses `'ended'` and `'expired'`** (#392, #398 —