no cause confirmed

rotor
- sampled Lab's memory for five minutes instead of repeating the theory that avatar playback was leaking
- renderer heap held between 178 and 222 MB, video and array backing storage stayed near 168 to 175 MB, listener counts flat
- the 28.91 GB process that started the question was already gone before the capture ran, so nothing observed explains it
- several Python backends grew in the same window, which the avatar theory does not cover; no code changed
- repaired three worker problems together through subagents, on ownership handoff, reading a worker's history, and controlling a worker, and merged them in one commit with 262 targeted tests and all 12 repository checks passing
- live activation held back: the shared-backend restart guard does not account for background reviews, so a restart cannot be shown safe yet
- a completion record for that landing stayed inconsistent and the follow-up observation command was refused
hq
- fixed message delivery: a missing-agent dispatch path, roster-service failures, and timeouts that never ended; twenty agents then answered through HQ with 213 tests passing
- replies still take one to two minutes
- one agent has no bridge to answer through and stayed silent
selfactor
- the runner cache hook exited 1 whenever no daemon-prepared cache environment existed, which is every push-triggered job GitHub dispatches directly, and it killed releases inside their setup step
- a missing environment now logs a notice and runs the job without the local cache, while the daemon path still injects and consumes its one-use environment