Description
After a streaming session ends, the experience process (e.g. Mercator56-Win64-Shipping.exe) persists indefinitely, consuming 5GB+ VRAM and 96% GPU utilization with 0 active streams.
Sunshine undo prep-cmd is supposed to signal stream end, but does not kill the experience process. The hydrabody CLAUDE.md describes a planned tickSunshineAppWatchdog() that should:
- Every 30s during idle (no active streams), check for experience processes still running
- If found with no corresponding active session, kill them
- Log the cleanup event in RecentCleanups
- Verify GPU utilization drops after cleanup
This was consistently observed during testbook runs on boom-pickle-38 (RTX 4080 SUPER) — every session leaves an orphaned Mercator process requiring manual taskkill.
Nebula's reasoning: Orphaned experience processes consuming 5GB+ VRAM and 96% GPU utilization after every session is a resource leak bug, not a feature request. High priority because it directly degrades body availability — every session leaves the GPU unusable until manual intervention, which is unsustainable for production streaming. Title and description are clear and actionable, no changes needed there.