Launch native agent UIs through shared tmux hosts

[ScottRBK]

AgentShell can now launch each agent's actual interactive UI in tmux while exposing shared
input, event observation, and lifecycle control. Supports Codex, Claude Code, Pi, OpenCode,
Copilot CLI, Cursor, and Grok without requiring an app server or reconstructing their UIs.

• Adds open_interactive() and shared terminal/worker infrastructure; harness-specific setup
  and event parsing remain in adapters. Capabilities describe each harness's available events.
• Supports owned sessions, windows in existing sessions, and panes to the right of or below the
  caller. Numeric session targets use exact matching.
• Cleanup checks the original socket, resource IDs, and launch marker before deleting tmux
  resources, preserving replacements even after server restarts. Startup cleanup also covers
  interrupted launches and remain-on-exit.
• Shares a process guardian across native and interactive runs to terminate resistant helpers
  after the CLI exits, including when terminal job-control signals suspend the harness.
• Adds a runnable demo, integration tests, and local E2Es. Pins Codex E2Es to Luna at low effort
  and allows effort selection for health probes. No new Python dependencies.

Validation: 1,050 unit/integration tests passed, including 23 new cleanup regression cases.
Real tmux tests use explicitly isolated servers and check input, layout, focus, process
lifecycle, and preservation of borrowed resources. Real Codex/Pi split E2Es passed.
An independent Astra xhigh review approved the final cleanup changes.

Provider-wide E2Es have not all passed: previous runs encountered Claude usage limits and
Copilot account-budget and interactive trust blockers. Interactive hosts remain experimental,
support NoIsolation only, and expose differing metadata capabilities across harnesses.

[ScottRBK]

please remove the two invalid documents from the poc but aside from that it is good to go