Problem

When the host's WebRTC publish drops (flaky network / dead secondary NIC — the recurring dtls timeout ~20–90s into a stream), the room-composite egress was left orphaned: LiveKit keeps compositing the now-empty room, burning SFU CPU and pushing dead/frozen frames to every platform ("YouTube isn't getting any data") until the dead publisher finally times out server-side minutes later. On the 4-vCPU box that orphan also re-thrashed CPU and broke subsequent Go Lives.

Fix

New useAutoStopServerStream(isHosting, isServerStreaming, stopServerStream) hook stops the tracked egress on the hosting true -> false transition. The renderer stays alive during a media drop, so it can issue the stop promptly — well before the server-side participant timeout.

• Fires only on the transition (not on mount, not while hosting stays up).
• A user-initiated "stop server stream" clears the egress first (isServerStreaming -> false), so it never double-stops.
• Paired with the start route already stopping pre-existing egress (feat(egress): stop existing room egress before starting a new one #22), a fully-crashed host is also cleaned up on the next Go Live. (No LiveKit webhook infra exists, so this client + next-start combination is the coverage without standing up a reaper.)

Test

Extracted the logic into a unit-testable hook (CapturePreview is too heavy to render in tests). 4 cases: stops on host-drop while streaming; no-op when not streaming; no fire on mount; no fire while hosting stays up. 4/4 pass; full monorepo lint passes.

Ships in the Electron app (desktop release), not Railway.

🤖 Generated with Claude Code