[TheCryptoDonkey]

When two viewers start an empty listening room together, the slower viewer can keep buffering an unpublished candidate after the other viewer has started the shared clock. Adopt the verified active room promptly, stop the empty old loader before acknowledgement, and preserve the original player and Play request. Failed acknowledgements, cancellation and browser policy rejection retain an explicit retry path.

Pin RelaySwarm source e910147 and its aligned WASM, including bounded rolling startup planning from verified load time. Cache keys keep one recording brand and SDK instance across the room/media helpers. Documentation, the AI index and the scoped candidate receipt describe the resulting behaviour and its limits.

Validation:

• 350/350 unit tests and the site build pass. Independent source/provenance review is clear.
• The room-adoption source and final import-graph integration each passed all 36 controller lifecycle checks on M4. The old controller failed the intended adoption check.
• One staged same-M4 real 4K run completed 60 seconds per viewer with zero post-frame stalls/dropped frames. Preselected receiver B accepted 62,784,300 peer bytes (25.62%); A accepted 6,222,271 (2.61%). Their actual paired observation overlap was 57.408 seconds. Fresh room comparisons matched and sampled clock offsets stayed within 0.307 seconds. One nonfatal pre-frame aborted old fragment is retained.

Uses the immutable public SDK commit e910147 from RelaySwarm PR37. Fourteen SDK CI jobs passed; its separate playback job stopped during FFmpeg installation before tests, including one bounded retry. That setup diagnostic is being improved separately. The site checks above exercised the exact pinned browser build. Deployed 330-second replay, fresh physical-phone and public TURN acceptance remain separate gates; no independent public TURN endpoint is configured.