docs: live-session findings 2026-09-13 (metrics + prioritized issues)

[iamfatness]

Metrics captured from a ~5-hour real test-meeting session (10→16 participants, 8-member Tiles wall at 1080p, streaming to YouTube), with the prioritized list of what to address.

Issues filed from this sweep:

• P0 (in flight): fix(rtmp): read the live video pipe greedily so the stream drops to live instead of falling behind #515 (streaming falls behind live — -re fix), fix(tiles): pin the Tiles wall at 1080p so preview no longer drops on-air members #514 (tiles preview-drop — 1080p wall)
• P1 (load): perf: a command handler held coreMutex for 143ms, freezing the render thread #516 (143 ms coreMutex command freeze), perf: Program render hits the one-frame budget at 16 participants with the 1080p wall #517 (Program render at the frame budget at 16 participants), Operator stutter: what flips the participant structure signature, and which rebuild step costs 197ms #509 (shell structural rebuild 17–21 ms — now measured on the owner machine)
• P2 (verify/cleanup): verify: Zoom audio path on air — FADER LAW drops 'zoom-mix' as unrouted #518 (verify Zoom audio path / FADER LAW), RTMP sender bytesSent and latencyMs are fake (estimated/hard-coded) and mislead diagnosis #519 (fake bytesSent/latencyMs counters)

Also notes the cap-vs-9-camera-on decision (relates to #514) and the two sender counters that lie.

🤖 Generated with Claude Code

https://claude.ai/code/session_014yuH7EMvWCWdkhvevrtMyJ