Two parallel desktop-control stacks on different displays (:99 vs :1) — consolidate to canonical path #11579

Discovery (from #11506 audit)

Two parallel, both-wired desktop-control stacks exist, driving different X displays — so an agent using one path acts on a screen the human observer can't see.

Stack A — pyautogui / GUIController

• autobot-backend/worker_node.py:58 imports GUIController; instantiated at :90 as self.gui_controller
• autobot-backend/task_handlers/gui_handlers.py:53 calls ctx.worker.gui_controller.read_text_from_region(...)
• Drives Xvfb :99 — autobot-backend/gui_controller.py:58 (Xvfb :99 ...), config.display = ":99" at :62

Stack B — xdotool / vnc_manager (agent MCP tools + human observation)

• autobot-backend/api/vnc_manager.py xdotool click/type/etc.; exposed to agents via autobot-backend/api/vnc_mcp.py
• Human observes this via x11vnc/noVNC — runs on display :1 (vnc_manager.py:115 "Start VNC server on display :1")

Why this matters

• Correctness / observability: work done on :99 (Stack A) is invisible to a human watching :1 (Stack B). Breaks the "observe + take over" goal of Umbrella: agent-operated desktop sessions with safe human takeover (x11vnc) #11506.
• Canonical duplication: two implementations of "control the desktop" (pyautogui vs xdotool) with divergent humanization, capture, and audit paths.

Proposed direction

• Converge on one canonical desktop-control abstraction on one display, per the x11vnc-attach decision in Umbrella: agent-operated desktop sessions with safe human takeover (x11vnc) #11506 (agent drives the same display the human observes).
• Do NOT delete Stack A blindly — wire gui_handlers onto the canonical path (read_text_from_region has no xdotool-path equivalent yet; that capability must be preserved).

Dependency

Blocks / must be resolved as part of #11506 T1 (control-lock) — the lock and the display must be the same for both the agent and the human.

Evidence-based; file:line cited above.