Evolve experimental PiP into a multi-target Agent View

[f-trycua]

Problem

The experimental macOS PiP surface is a process-global, single-frame preview. It captures only the latest post-action native screenshot, so concurrent sessions overwrite one another and exact Chrome tabs are not represented independently.

For the shared-computer model, the visual surface should group exact targets by the existing lifecycle session without adding claim/release ownership APIs.

Scope

Evolve experimental PiP into a multi-target Agent View:

• preserve existing --experimental-pip and config compatibility;
• group cards by the existing session/runtime identity;
• automatically identify native targets by pid + window_id;
• identify browser targets by target_id + tab_id;
• capture Chrome tabs through the existing exact CDP binding, without activating the tab or foregrounding its window;
• keep multiple target cards visible instead of replacing one global frame;
• remove session cards when that session ends;
• keep target assignment automatic and internal: no claim/release MCP tools;
• retain Windows/Linux stubs, with native backends remaining tracked by PiP preview (experimental): native Windows + Linux backends #1729.

Non-goals

• virtual machines or macOS display creation;
• security ownership/lease enforcement between agents;
• continuous video streaming;
• human input forwarding from Agent View into a target;
• Windows/Linux native PiP implementation.

Acceptance evidence

• focused shared-model and routing tests;
• macOS build/compile verification;
• live macOS smoke proving two native windows and two Chrome tabs remain distinct in Agent View;
• existing CLI/config compatibility remains intact.