perf(desktop/compositing): Layerize is ~27.6% (3s) of desktop main-thread — the #1 component of the #4539 "Other" bucket — audit + cut composited-layer count

[koala73]

Surfaced by the #4539 desktop main-thread attribution baseline (docs/perf/desktop-mainthread-baseline-2026-07-02.md). Part of #4539 / #4487 (desktop render axis).

Signal

#4539 set out to characterize the 52% / ~11 s "Other" bucket on desktop /dashboard. The new harness (scripts/measure-desktop-mainthread.mjs) decomposed it, and the single largest previously-uncharacterized component is Layerize:

"Other" component	unthrottled (cpu 1)	throttled (cpu 4)
Layerize	27.6% (3.06 s)	15.5%
ThreadControllerImpl::RunTask (scheduler)	20.4%	16.5%
rest (IntersectionObserver, GC, mojo, …)	~7%	~10%

Layerize is stably a top-2 "Other" component across both host conditions → a real structural cost, not a host artifact. Lighthouse buckets Layerize into "Other," which is exactly why the 52% was a black box.

What Layerize is

Blink compositor layerization — assigning paint layers to compositing layers and (re)building the compositing layer tree. Its cost scales with the number of composited layers and how often the layer tree is rebuilt. ~3 s of it on desktop /dashboard means the page is creating/updating a large number of compositing layers.

Lever / investigation path

1. Count the composited layers on prod /dashboard via CDP LayerTree.enable → LayerTree.layerTreeDidChange (or DevTools "Layers" panel). Establish how many layers exist and which nodes own them.
2. Audit layer-forcing CSS for over-promotion beyond the two unavoidable map canvases (deck.gl scene-container>canvas + maplibregl-canvas): will-change, transform: translateZ(0) / 3D transforms, position: sticky/fixed, opacity/filter on large subtrees, per-panel promotion, backdrop-filter.
3. Reduce unnecessary promotions so the compositor isn't re-layerizing ~3 s worth.

Tension to respect

#4538 (convert non-composited animations to transform/opacity) promotes some elements to their own compositing layer to make paint cheap. Over-promotion is the opposite failure and inflates Layerize. Balance the two: promote what animates, demote what doesn't — don't blanket-will-change.

Acceptance

• A named cause: which elements/CSS over-promote, backed by a composited-layer count.
• Layerize self-time share drops, re-measured by scripts/measure-desktop-mainthread.mjs (before/after, both --cpu 1 and --cpu 4).
• Absolute desktop mainthread-work cross-checked against a clean PSI/Calibre run (the harness gives the relative share, not the headline absolute — KTD1).

Effort: medium (investigation-first). This is the desktop render-axis twin of the INP/map presentation work; distinct from #4536 (forced reflow / styleLayout) and #4538 (non-composited animations).