feat(agent): add a vim-native responsive agent workspace

[fcoury-oai]

Why

The Codex agent should feel like a native part of Red: prompts and follow-ups need real Vim editing, the conversation needs to fit any editor layout, and native Codex tools and approvals should remain available without weakening the isolated review workflow.

What changed

• Add one real, buffer-backed modal composer for floating prompts and persistent agent conversations, including multiline Vim editing, undo and history, Unicode-safe cursor movement, mode-specific cursor shapes, and immediate Ctrl+Enter submission.
• Make the conversation dock responsive and movable to the left, right, top, or bottom with Ctrl+W H/J/K/L and :AgentLeft, :AgentRight, :AgentTop, and :AgentBottom. Preserve drafts, history, reading position, focus, editor space, and theme-correct rendering across moves and terminal resizes.
• Support native Agent, Plan, and isolated Review modes; persistent and resumable workspace-scoped threads; model and reasoning selection; live turn steering; streaming tool activity; and explicit, fail-closed Codex approvals.
• Synchronize native agent edits into clean open buffers while preserving unsaved buffers and reporting conflicts. Keep automatic context and staged review overlays bounded to safe, non-sensitive workspace files.
• Introduce and document the version 0.4.1 agent conversation plugin APIs, update [agent] defaults and workflow documentation, and add targeted modal composer, panel, editor, plugin, and app-server regressions.

How to Test

1. Run the focused automated coverage:

   cargo test --lib modal_composer
   cargo test --test editing
   cargo test --test codex_app_server

   Expected: modal Vim editing, modified-Enter submission, four-sided dock layout, native session and approval handling, resumable threads, and isolated review regressions pass.
2. Start Red with an authenticated codex on PATH, press Space A, compose a multiline prompt, switch between Insert and Normal modes, and submit with Ctrl+Enter. Expected: the prompt opens a streaming docked conversation; plain Enter still inserts a newline in Insert mode, and Ctrl+S remains an editor save shortcut. When using tmux, first enable extended-keys on and extended-keys-format csi-u.
3. Focus the conversation, enter an unsent follow-up, and move the dock with Ctrl+W H, Ctrl+W J, Ctrl+W K, and Ctrl+W L; resize to a narrow terminal and back. Expected: all four positions work and the draft, conversation, cursor, focus, and usable editor viewport survive each move and resize.
4. Set [agent] mode = "review", request a file edit, and inspect it with :AgentReview. Expected: the workspace remains unchanged until the proposed hunk is explicitly accepted. In native Agent mode, verify a completed change reloads a clean open buffer while an unsaved buffer is preserved and reported as a conflict; dismissing an approval must deny it.