Problem

AI agents using Playwright's ariaSnapshot() or CDP Accessibility.getFullAXTree() receive the entire accessibility tree, including hundreds of off-screen nodes. This wastes serialization time and consumes AI context budget on nodes the agent can't see or interact with.

Solution

Add a CefBrowserSettings.ax_viewport_collapse setting (experimental) that collapses off-screen nodes in CDP accessibility tree serialization:

• Off-screen landmarks/headings: serialized as summaries (role + name only, empty childIds)
• Off-screen interactive/structural nodes: pruned from the response
• In-viewport nodes: fully serialized with all children and properties
• position:fixed/sticky descendants: correctly detected as in-viewport even when their parent is off-screen

The setting is per-browser and can be toggled at runtime via CefBrowserHost::SetAxViewportCollapse(). Platform screen readers (NVDA, JAWS, VoiceOver) are completely unaffected — they use a separate code path. CDP nodesUpdated events are suppressed when active to maintain tree consistency. queryAXTree is left unfiltered so agents can still locate off-screen elements for scroll targeting.

Agentic workflow

With the setting enabled, an AI agent's typical interaction loop becomes:

1. Get collapsed tree — getFullAXTree returns in-viewport nodes fully serialized, off-screen landmarks/headings as name-only summaries, and everything else pruned. The agent sees a compact representation of the page.
2. Interact with visible content — buttons, inputs, links in viewport are fully described.
3. Discover off-screen sections — collapsed landmark summaries (e.g. navigation: "Footer Nav") tell the agent what exists below the fold without serializing the full subtree.
4. Scroll to a section — the agent uses the summary node's backendDOMNodeId with DOM.scrollIntoViewIfNeeded to bring it into view.
5. Re-get tree — the scrolled-to section is now fully serialized; previously visible content that scrolled out is collapsed.

No changes are required in Playwright, MCP servers, or agent code — only the CEF embedder enables the setting.

Verification

• 12 ceftests covering: default/disabled/enabled states, heading levels, childId filtering, position:fixed visibility, nested landmarks, runtime toggle, all-in-viewport no-op, queryAXTree passthrough, full agentic scroll workflow, and zoom-level coordinate correctness
• Manual verification via cef/tools/debug/ax_viewport_collapse/verify_viewport_collapse.py
• cefclient --ax-viewport-collapse enables the setting for manual testing