@urun-sh/core 0.2.44 → 0.2.45

package/CHANGELOG.md

@@ -1,5 +1,37 @@
 # Changelog
 
+## 0.2.44
+
+- **A flat inbound RTP counter no longer tears down a healthy media transport.** The
+  media-liveness watchdog treated any inbound-bytes counter that stayed flat past
+  `stallTimeoutMs` (default 6s) as media death and answered it with a full
+  teardown + re-dial. A cumulative receive counter cannot tell a dead path from a
+  SENDER WITH NOTHING TO SEND — both look identical at the receiver — so every
+  on-demand app blacked itself out: zimage emits a handful of H.264 frames per
+  generation and then legitimately goes quiet, so each idle gap destroyed the
+  transport, the SFU scored the teardown as consumer-abandoned and ended the
+  session, the runtime tore down its encoder, and the next dial had no keyframe to
+  decode. Self-sustaining reconnect loop; same class as the 3724x reconnect storm.
+  The ICE agent already answers "is this path alive?" natively — consent freshness
+  (RFC 7675) re-probes the selected pair every ~5s independently of media, and
+  `transport.connectionState` IS that verdict. The stall window now only
+  CORROBORATES a transport the ICE agent has already given up on. While ICE says
+  `connected`, a quiet stream emits a new `media-idle` (`level: 'info'`) diagnostic
+  and the transport, its consumers and the last decoded frame are left alone.
+  Genuine media death keeps both authoritative library-native detectors —
+  `connectionstatechange` → `restartIce()` in place, and a higher control-doc
+  incarnation → renegotiate — neither of which needs the counter. Detection stays
+  level-triggered, so a path that dies AFTER a long idle gap is still recovered.
+  Locked by the `no-teardown-on-idle-inbound-counter` ast-grep guard.
+- **The idle gap no longer leaks into the DECODE rule.** `_checkDecodeLiveness` is only
+  reachable while inbound bytes are ADVANCING, so nothing refreshed `_lastDecodeAdvanceAt`
+  across a quiet stretch. On the first resume tick the new generation's RTP can land before
+  the decoder has assembled a frame from it — bytes advance, `framesDecoded` has not ticked
+  yet — and the decode rule then measured the WHOLE idle gap and declared a `decode-wall`,
+  firing exactly the `_renegotiate()` teardown the fix above removes. The `media-idle` branch
+  now re-arms the decode clock too: a decoder with nothing to decode is not a stalled decoder.
+  A genuine decode wall after a resume is still caught.
+
 ## 0.2.24
 
 - Docs-only release: README accuracy pass against the 0.2.23+ surface (the READMEs ship in