macOS: a density change during a resize no longer freezes the view

[wenkaifan0720]

Switching the device frame in FlutterFlow's Test Mode could freeze the preview for good. JS still answered, but nothing painted, and hide/show and reload didn't bring it back. It happens when the previewed page is slow to relayout, like a debug web build.

Cause

A device switch changes the view's size and render scale at once. CEF's OSR view keeps one resize in flight: a size or scale change "holds" until a paint arrives at the new pixel size, and a change made during the hold is queued for when it releases.

A queued density change is applied through RenderWidgetHostViewBase::UpdateScreenInfo. That sends the renderer the new scale under the surface id it already has. The renderer's next frame therefore has a different scale for the same surface id, so viz rejects it as a surface invariants violation and drops the renderer's frame sink.

On macOS the view never comes back from that, for one of two reasons:

• The renderer returns under a surface id it allocated itself. CEF 144 never embeds that id, because CefRenderWidgetHostViewOSR::DidUpdateVisualProperties is compiled out for Mac (upstream CEF has since removed that guard). The display has no damage it recognises, never draws, and the hold never releases.
• Or the external begin-frame source waits forever on the dropped sink, so no begin frames go out at all.

Found with --trace-startup: Surface invariants violation → DidLoseLayerTreeFrameSink → ChildLocalSurfaceIdAllocator::GenerateId, followed by no browser-side embed of the new id and no Display::DrawAndSwap for the rest of the run.

This also explains an earlier experiment. Dropping the NotifyScreenInfoChanged call made the freeze worse (round 1 instead of round 2): WasResized alone applies the new density through the same UpdateScreenInfo path.

Fix

• cef_host now shows CEF its geometry through cef_width / cef_height / cef_dpr, via GetViewRect and GetScreenInfo.
• Those values change only when no resize is in flight. A resize ends when CEF paints the exact pixel size it will hold for (ScaleToCeiledSize in float), and whatever the plugin asked for meanwhile is applied next. Intermediate sizes collapse into the latest one.
• A density change goes through NotifyScreenInfoChanged, which retires the surface id before applying the new scale.
• Fallback: a resize with no paint at its size gets NotifyScreenInfoChanged again, 1, 2, then 4 s apart, up to 6 times. With the geometry CEF sees unchanged, that only swaps in a fresh surface id.
• Windows is unchanged. It doesn't use external begin frames, and CEF on Windows embeds renderer-allocated ids.

Verification

New probe device_frame, added to run_probes.sh, mirrors Test Mode's widget tree: FittedBox, MediaQuery, CefWebView at devicePixelRatio × zoom. It switches devices 30 times in three ways (plain, parked-then-shown, animated). The page blocks its main thread 600 ms on every resize, and the UI thread is janked 40 ms per frame. Each round checks the page's viewport and density, then that it still presents a frame.

host (shipping framework)	result
before	froze at round 2, 3 of 3 runs
after	30/30, 3 of 3 runs; no invariants violation in a 40 s trace

Full tool/run_probes.sh: 18 of 19 passed. The one failure, host_robustness, is a pre-existing flake: a sentinel eval misses its 5 s window in a random case. It fails the same way on the baseline host (interleaved A/B: baseline 3/4, this branch 4/4). That probe creates its tiles headless and never resizes them, so none of the new code acts there.

A second commit hands CEF the latest geometry in OnAfterCreated. GetViewRect now reports what CEF was handed, so without it a resize that arrived while the browser was being created would have waited for the next resize.

The prebuilt must be republished (make publish-cef-host) before consumers repin.

🤖 Generated with Claude Code