feat(workspace): embed live GUI apps via Xvfb capture with input forwarding

feat(workspace): embed live GUI apps via Xvfb capture with input forwarding

Add a bridge/x11 package that runs GUI applications on a private,
per-app Xvfb display and captures the screen into a texture:
XvfbServer manages display lifecycle, X11Native reads the root window
into caller-owned ARGB pixels over JNA, and GuiAppSession owns the app
process (dbus-run-session isolation, generated Firefox profile that
disables system extensions) plus a 10 Hz capture thread that repaints
only on changed frames.

FirefoxPanel renders the captured screen on a textured rectangle and,
while focused, forwards input into the virtual display over XTEST:
clicks at the perspective-correct texture coordinate (with an 80 ms
button hold so Firefox synthesizes DOM clicks), keyboard events 1:1
through an AWT-to-X keysym table, scroll on both axes (X buttons
4-7), mouse back/forward (8/9, dispatched on press since AWT never
fires CLICKED for extra buttons), and pointer hover motion. Shift+ESC
or a middle click releases focus; focus loss releases server-side
modifiers so Shift never sticks.

Workspace now spawns a 4x4 grid of independent PTY terminals whose
mouse wheel maps to arrow keys honoring DECCKM. UnixPty pads the
posix_spawn structs to 512 bytes (the amd64 attribute block is 336)
and destroys them after spawning.

Main gains headless verification drivers: --guitest, --clicktest,
--keytest, --scrolltest, --hovertest. TODO.org collects ideas for
embodied agents in the world.