📝 docs(design): RFC — session persistence via a detached daemon

[winlp4ever]

Drafts the "big gun" we discussed into a proper RFC — docs/design/SESSION_DAEMON.md — promoting ARCHITECTURE Appendix A from a sketch to a buildable plan. No code; this is the doc to argue against as we continue.

What it captures

• The two-reconnects reframe (the key clarification): app disconnect (quit/crash/auto-update) is what the daemon solves; network disconnect (sleep/wifi) needs wrapping the remote's tmux/mosh. The full ssh-survives story = daemon + remote-tmux. Machine reboot stays a non-goal (layout-restore is the fallback).
• The genuinely hard part: the daemon models the screen, not a byte pipe — a headless @xterm/headless emulator per session + @xterm/addon-serialize snapshot for a correct cold reattach, with the fidelity cases that'll eat the time (alt-screen TUIs incl. Claude's, reflow-on-resize, scrollback bound, cwd/status tracking).
• Transport/protocol + lifecycle: unix socket / named pipe, a small frozen verb set, discover-or-spawn, orphan reaping — and the rule that keeps it sane: the daemon stays dumb, all product logic stays client-side.
• Auto-update × daemon (new since Appendix A): thin frozen protocol so a new client speaks to an old daemon; new sessions on a fresh daemon, old one drains — no forced session kill on update.
• Perf budget for the extra hop (coalescer moves to the socket boundary; <1–2 ms added keystroke latency, gated on the SMTERM_PERF harness), security (0600 socket), quit-UX (keep-running-in-background default; "N background sessions" in the status bar), and WSL/hook interplay.
• A staged build plan (0–4) — pty-host child + screen model → detach it (the M5 headline) → remote ergonomics → multi-client/CLI — so we ship value early and defer the scary lifecycle work.
• §12 Open decisions for us: build-local/wrap-remote, quit default, perf gate, v1 scope, scrollback cap.

Appendix A now links here. Ready to iterate — the open questions in §12 are where I'd like your calls next.