feat(ui): run the turn on a worker thread, marshal the UI to the loop

[lavindeep]

Branch 4/9 of the UI v2 full-screen TUI rework — the threading keystone, and the first live-test milestone.

What

In the full-screen app the prompt_toolkit event loop is the main thread, so driving a model turn synchronously there would freeze the UI for the whole turn. Run the one synchronous turn on a worker thread and marshal every UI callback back onto the loop thread, so AppUI state + pane repaint only ever touch the loop thread.

• ThreadedUI — wraps AppUI + an injected schedule; each fire-and-forget call is enqueued via functools.partial. Blocking approvals return a value, so they delegate to the inner (which still raises until branch 7, surfaced as a pane error).
• TurnRunner — one worker per turn; busy flag touched only on the loop thread (set in start, cleared via a scheduled _mark_done) — no lock. The worker only calls run_turn + schedule; any failure surfaces as a pane error and the finally clears busy, so a crashed turn never wedges the app. schedule reads app.loop lazily and call_soon_threadsafe's a callback that runs the call then invalidate().
• build_app gains on_submit; the opt-in run_app (SHELLPILOT_UI=app) wires it and app.run()s it. The default REPL is byte-identical — app mode is off unless the env opt-in is set.

Verification

• Phase gate green: ruff + mypy --strict clean, 1334 passed (20 new test_app_turn.py tests).
• Marshaling/threading proven via a real worker thread + FakeLLM; every load-bearing test proven red-first (marshal-not-inline, busy-guard, exception-surfaced, busy-cleared, on_submit).
• Independent skeptical review (SHIP) + second-model cross-thread review against the prompt_toolkit source (concerns only) — call_soon_threadsafe + lazy app.loop + invalidate-from-callback all confirmed safe; the conversation-guard moved inside the worker try and the shutdown-drop documented per their feedback.
• DESIGN §31.13 in the same commit.

Scope

No cancellation (branch 6), no approval focus-swap (branch 7), no queue (branch 9), no thinking indicator (branch 5). Live behaviors (real app.run() loop, chrome-persists-through-a-turn, resize/scroll) are live-test items, run by the maintainer — not CI.