perf: 1.0 (installed 362 tree) idles at ~30 % CPU with every window hidden — main thread redrawing, nothing in the log

[kpmmmurphy]

Seen 2026-09-15 22:03 (launch night), while preparing a quiet baseline for the power receipt (#353)

/Applications/M1K3.app (the 362 tree, launched 14:05, 8 h uptime, Lil resident, MCP server on) at 30–38 % CPU by top over several 2–3 s samples, with visible set to false on the process (every window ordered out) and get_status reporting nothing in flight (not speaking, no mic, not answering). Resident 5.1 GB. log show --last 3m --predicate 'subsystem == "app.m1k3"' printed no lines at all — whatever this is, it is silent.

Per thread (ps -M): main thread ~16 % (priority 46T), two young worker threads ~11 % and ~10 % (each with only ~6 s of lifetime CPU — spawned recently), a fourth at ~4 %.

Hottest non-wait leaf frames (sample M1K3 5, sort by top of stack): _dispatch_timer_heap_resift (466), RB::SurfacePool::collect (RenderBox, 70), CABackingStoreCollect_ (62), _dispatch_source_invoke (72) — everything else is __psynch_cvwait / semaphore_wait_trap (the usd/tbb pools parked). Reads as SwiftUI redrawing something on a timer while nothing is visible.

Earlier the same evening, with the main window on screen, top read M1K3 at ~49 % beside WindowServer at ~51 %.

Why it matters

#293 (2026-09-12) measured the fanless idle at 0.4–1.8 % on the merged build with the same defaults (HUD on, companion set). Either 1.0 regressed between #293 and #340, or a setting on this Mac (notch HUD, menu-bar pet, screensaver poll, heartbeat) reintroduced a per-frame loop. It is also the eco story's own receipt: an idle M1K3 that burns a third of a core is not a resident, it is a fan.

First moves

1. Reproduce on a fresh launch of the installed build: top -pid for 60 s with the window hidden, then toggle notch HUD / menu-bar creature / heartbeat one at a time.
2. sample with the window hidden and match the timer against MenuBarLabel / AudioCaptureBackdrop (TimelineView capped at 30 fps in perf(avatar): unmount hidden RealityViews, pause the rest — a fanless idle (33-49% → <2% CPU) #293 — a cap is still 30 redraws a second if the view is mounted) and the notch HUD's drive loop (NotchHUDVisibility.wakeValve).
3. Compare against perf(avatar): unmount hidden RealityViews, pause the rest — a fanless idle (33-49% → <2% CPU) #293's method (sample DisplayLink count + top), not heap.

Found while measuring, not fixed here — the power receipt's idle window will carry whatever this costs until it is.