Multiplexer interoperability: behavior matrix and contract tests

[raiseCatError]

Closes #175

#175 is a research issue: "deliver a concise behavior matrix and prioritized implementation follow-ups … NMSh is not becoming a pane/window multiplexer; the outcome is reliable coexistence." This PR delivers that. No product behavior changes.

Deliverables

• docs/architecture/multiplexer-interop.md: a behavior matrix for NMSh inside tmux and GNU screen and for each running inside NMSh, the persistence-overlap analysis, the detection decision, and prioritized follow-ups.
• tests/muxInterop.test.ts: pins the verified contract with real tmux and screen, skipping when they're not installed, so CI never depends on a multiplexer. Plus environment-only host-detection cases for tmux, screen and Zellij.

Verified (tmux 3.7c, GNU screen 4.00, isolated sandboxes)

• NMSh inside tmux:
  • the shell sees TERM=tmux-256color, TERM_PROGRAM=tmux and TMUX;
  • resize propagates (the pane minus the composer rows);
  • Ctrl+Z and jobs work; less passes through and returns;
  • a multi-line paste-buffer -p stays a paste;
  • closing the pane detaches the live session, which stays restorable.
• NMSh inside screen: rendering, resize, Ctrl+Z, and less with its return all work.
• screen and tmux inside NMSh: they pass through via the alternate screen and return, and tmux repaints after reattach (existing test).
• A live session keeps its creation environment: made in tmux and reattached outside it, the shell still has TERM=tmux-256color, TERM_PROGRAM=tmux and a stale TMUX. This is the main interop wrinkle and is documented, not changed.

Not verified (physical QA or follow-ups)

• Zellij and TUIOS: not installed, so marked unverified.
• Mouse under tmux: needs mouse on.
• Shift+Enter under tmux: tmux send-keys can't produce real extended keys. NMSh decodes both extended encodings, but tmux's default settings likely deliver Enter.
• /copy via pbcopy inside tmux.

Detection

Baseline coexistence needs no multiplexer detection; nothing was added. #197's window launcher keeps its current behavior: inside tmux started from Ghostty, it detects Ghostty through the inherited GHOSTTY_RESOURCES_DIR, so "Open all" opens Ghostty windows outside tmux. That's documented and tested, and follow-up 3 decides whether to change it.

Prioritized follow-ups (in the doc)

1. A reattach notice when the attaching environment differs from the session's (P1).
2. Request modifyOtherKeys alongside kitty, so Shift+Enter works under tmux extended-keys on (P1).
3. "Open all" from inside a multiplexer (P2).
4. A Zellij pass (P2).
5. Recommended tmux settings in the README (P3).

Also

• Fixed a race in Agent-aware live session status in /resume #174's end-to-end status test: it typed /resume before a background command's prompt had returned, and failed once under full-suite load. Test-only change.

Tests

• tests/muxInterop.test.ts passed 4/4 repeatedly, including alongside liveHardening and ctrlZJobControl, with no tmux, screen or nmshd processes left behind.
• Build, typecheck and git diff --check are clean. The full suite passed 595/595 locally.

Review cleanup (bf51095)

• Every matrix row now states its evidence: automated test, manual observation or expected. The test file is described as coverage for selected tmux/screen paths.
• New tests:
  • closing only the tmux pane (server still running) detaches the live session;
  • NMSh inside GNU screen renders, sees STY, follows a resize and suspends a job;
  • an alternate-screen program gets the full terminal in passthrough;
  • the stale-environment test proves the tmux server is gone.
• tmux runs with multiplexer variables stripped from the inherited environment.
• The screen-inside-NMSh test now quits its screen session, which detaches instead of exiting when its terminal closes.
• Corrected claim: screen 4.00 keeps the window size it started with, so inside NMSh it keeps the composer-reduced height. It had been wrongly described as corrected by SIGWINCH. Added follow-up 6: hand known multiplexers the full terminal before they start.
• vim and fg under tmux, and less under screen, are marked manual. Stale TMUX/TERM is documented as affecting later child commands, while the attaching frontend's host detection uses its own environment.
• CI: the first two runs hit two pre-existing flakes in tests this PR doesn't touch (Live sessions: attach, detach and reattach #128's reattach-repaint test and the Ctrl+Z fg then Ctrl+C step). They'll be fixed in a stacked reliability PR. The rerun is green on Node 22 and 26.